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>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_max_reordering __read_mostly
= 300;
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
= 1000;
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;
96 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
98 int sysctl_tcp_thin_dupack __read_mostly
;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
101 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 #define REXMIT_NONE 0 /* no loss recovery to do */
128 #define REXMIT_LOST 1 /* retransmit packets marked lost */
129 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
131 /* Adapt the MSS value used to make delayed ack decision to the
134 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
136 struct inet_connection_sock
*icsk
= inet_csk(sk
);
137 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
140 icsk
->icsk_ack
.last_seg_size
= 0;
142 /* skb->len may jitter because of SACKs, even if peer
143 * sends good full-sized frames.
145 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
146 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
147 icsk
->icsk_ack
.rcv_mss
= len
;
149 /* Otherwise, we make more careful check taking into account,
150 * that SACKs block is variable.
152 * "len" is invariant segment length, including TCP header.
154 len
+= skb
->data
- skb_transport_header(skb
);
155 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
156 /* If PSH is not set, packet should be
157 * full sized, provided peer TCP is not badly broken.
158 * This observation (if it is correct 8)) allows
159 * to handle super-low mtu links fairly.
161 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
162 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
163 /* Subtract also invariant (if peer is RFC compliant),
164 * tcp header plus fixed timestamp option length.
165 * Resulting "len" is MSS free of SACK jitter.
167 len
-= tcp_sk(sk
)->tcp_header_len
;
168 icsk
->icsk_ack
.last_seg_size
= len
;
170 icsk
->icsk_ack
.rcv_mss
= len
;
174 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
175 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
176 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
180 static void tcp_incr_quickack(struct sock
*sk
)
182 struct inet_connection_sock
*icsk
= inet_csk(sk
);
183 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
187 if (quickacks
> icsk
->icsk_ack
.quick
)
188 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
191 static void tcp_enter_quickack_mode(struct sock
*sk
)
193 struct inet_connection_sock
*icsk
= inet_csk(sk
);
194 tcp_incr_quickack(sk
);
195 icsk
->icsk_ack
.pingpong
= 0;
196 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
199 /* Send ACKs quickly, if "quick" count is not exhausted
200 * and the session is not interactive.
203 static bool tcp_in_quickack_mode(struct sock
*sk
)
205 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
206 const struct dst_entry
*dst
= __sk_dst_get(sk
);
208 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
209 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
212 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
214 if (tp
->ecn_flags
& TCP_ECN_OK
)
215 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
218 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
220 if (tcp_hdr(skb
)->cwr
)
221 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
224 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
226 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
229 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
231 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
232 case INET_ECN_NOT_ECT
:
233 /* Funny extension: if ECT is not set on a segment,
234 * and we already seen ECT on a previous segment,
235 * it is probably a retransmit.
237 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
238 tcp_enter_quickack_mode((struct sock
*)tp
);
241 if (tcp_ca_needs_ecn((struct sock
*)tp
))
242 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
244 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
245 /* Better not delay acks, sender can have a very low cwnd */
246 tcp_enter_quickack_mode((struct sock
*)tp
);
247 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
249 tp
->ecn_flags
|= TCP_ECN_SEEN
;
252 if (tcp_ca_needs_ecn((struct sock
*)tp
))
253 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
254 tp
->ecn_flags
|= TCP_ECN_SEEN
;
259 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
261 if (tp
->ecn_flags
& TCP_ECN_OK
)
262 __tcp_ecn_check_ce(tp
, skb
);
265 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
267 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
268 tp
->ecn_flags
&= ~TCP_ECN_OK
;
271 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
273 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
274 tp
->ecn_flags
&= ~TCP_ECN_OK
;
277 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
279 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
284 /* Buffer size and advertised window tuning.
286 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
289 static void tcp_sndbuf_expand(struct sock
*sk
)
291 const struct tcp_sock
*tp
= tcp_sk(sk
);
295 /* Worst case is non GSO/TSO : each frame consumes one skb
296 * and skb->head is kmalloced using power of two area of memory
298 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
300 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
302 per_mss
= roundup_pow_of_two(per_mss
) +
303 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
305 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
306 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
308 /* Fast Recovery (RFC 5681 3.2) :
309 * Cubic needs 1.7 factor, rounded to 2 to include
310 * extra cushion (application might react slowly to POLLOUT)
312 sndmem
= 2 * nr_segs
* per_mss
;
314 if (sk
->sk_sndbuf
< sndmem
)
315 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
318 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
320 * All tcp_full_space() is split to two parts: "network" buffer, allocated
321 * forward and advertised in receiver window (tp->rcv_wnd) and
322 * "application buffer", required to isolate scheduling/application
323 * latencies from network.
324 * window_clamp is maximal advertised window. It can be less than
325 * tcp_full_space(), in this case tcp_full_space() - window_clamp
326 * is reserved for "application" buffer. The less window_clamp is
327 * the smoother our behaviour from viewpoint of network, but the lower
328 * throughput and the higher sensitivity of the connection to losses. 8)
330 * rcv_ssthresh is more strict window_clamp used at "slow start"
331 * phase to predict further behaviour of this connection.
332 * It is used for two goals:
333 * - to enforce header prediction at sender, even when application
334 * requires some significant "application buffer". It is check #1.
335 * - to prevent pruning of receive queue because of misprediction
336 * of receiver window. Check #2.
338 * The scheme does not work when sender sends good segments opening
339 * window and then starts to feed us spaghetti. But it should work
340 * in common situations. Otherwise, we have to rely on queue collapsing.
343 /* Slow part of check#2. */
344 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
346 struct tcp_sock
*tp
= tcp_sk(sk
);
348 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
349 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
351 while (tp
->rcv_ssthresh
<= window
) {
352 if (truesize
<= skb
->len
)
353 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
361 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
363 struct tcp_sock
*tp
= tcp_sk(sk
);
366 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
367 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
368 !tcp_under_memory_pressure(sk
)) {
371 /* Check #2. Increase window, if skb with such overhead
372 * will fit to rcvbuf in future.
374 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
375 incr
= 2 * tp
->advmss
;
377 incr
= __tcp_grow_window(sk
, skb
);
380 incr
= max_t(int, incr
, 2 * skb
->len
);
381 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
383 inet_csk(sk
)->icsk_ack
.quick
|= 1;
388 /* 3. Tuning rcvbuf, when connection enters established state. */
389 static void tcp_fixup_rcvbuf(struct sock
*sk
)
391 u32 mss
= tcp_sk(sk
)->advmss
;
394 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
395 tcp_default_init_rwnd(mss
);
397 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
398 * Allow enough cushion so that sender is not limited by our window
400 if (sysctl_tcp_moderate_rcvbuf
)
403 if (sk
->sk_rcvbuf
< rcvmem
)
404 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
407 /* 4. Try to fixup all. It is made immediately after connection enters
410 void tcp_init_buffer_space(struct sock
*sk
)
412 struct tcp_sock
*tp
= tcp_sk(sk
);
415 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
416 tcp_fixup_rcvbuf(sk
);
417 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
418 tcp_sndbuf_expand(sk
);
420 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
421 tp
->rcvq_space
.time
= tcp_time_stamp
;
422 tp
->rcvq_space
.seq
= tp
->copied_seq
;
424 maxwin
= tcp_full_space(sk
);
426 if (tp
->window_clamp
>= maxwin
) {
427 tp
->window_clamp
= maxwin
;
429 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
430 tp
->window_clamp
= max(maxwin
-
431 (maxwin
>> sysctl_tcp_app_win
),
435 /* Force reservation of one segment. */
436 if (sysctl_tcp_app_win
&&
437 tp
->window_clamp
> 2 * tp
->advmss
&&
438 tp
->window_clamp
+ tp
->advmss
> maxwin
)
439 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
441 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
442 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
445 /* 5. Recalculate window clamp after socket hit its memory bounds. */
446 static void tcp_clamp_window(struct sock
*sk
)
448 struct tcp_sock
*tp
= tcp_sk(sk
);
449 struct inet_connection_sock
*icsk
= inet_csk(sk
);
451 icsk
->icsk_ack
.quick
= 0;
453 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
454 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
455 !tcp_under_memory_pressure(sk
) &&
456 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
457 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
460 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
461 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
464 /* Initialize RCV_MSS value.
465 * RCV_MSS is an our guess about MSS used by the peer.
466 * We haven't any direct information about the MSS.
467 * It's better to underestimate the RCV_MSS rather than overestimate.
468 * Overestimations make us ACKing less frequently than needed.
469 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
471 void tcp_initialize_rcv_mss(struct sock
*sk
)
473 const struct tcp_sock
*tp
= tcp_sk(sk
);
474 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
476 hint
= min(hint
, tp
->rcv_wnd
/ 2);
477 hint
= min(hint
, TCP_MSS_DEFAULT
);
478 hint
= max(hint
, TCP_MIN_MSS
);
480 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
482 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
484 /* Receiver "autotuning" code.
486 * The algorithm for RTT estimation w/o timestamps is based on
487 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
488 * <http://public.lanl.gov/radiant/pubs.html#DRS>
490 * More detail on this code can be found at
491 * <http://staff.psc.edu/jheffner/>,
492 * though this reference is out of date. A new paper
495 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
497 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
503 if (new_sample
!= 0) {
504 /* If we sample in larger samples in the non-timestamp
505 * case, we could grossly overestimate the RTT especially
506 * with chatty applications or bulk transfer apps which
507 * are stalled on filesystem I/O.
509 * Also, since we are only going for a minimum in the
510 * non-timestamp case, we do not smooth things out
511 * else with timestamps disabled convergence takes too
515 m
-= (new_sample
>> 3);
523 /* No previous measure. */
527 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
528 tp
->rcv_rtt_est
.rtt
= new_sample
;
531 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
533 if (tp
->rcv_rtt_est
.time
== 0)
535 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
537 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
540 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
541 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
544 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
545 const struct sk_buff
*skb
)
547 struct tcp_sock
*tp
= tcp_sk(sk
);
548 if (tp
->rx_opt
.rcv_tsecr
&&
549 (TCP_SKB_CB(skb
)->end_seq
-
550 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
551 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
555 * This function should be called every time data is copied to user space.
556 * It calculates the appropriate TCP receive buffer space.
558 void tcp_rcv_space_adjust(struct sock
*sk
)
560 struct tcp_sock
*tp
= tcp_sk(sk
);
564 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
565 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
568 /* Number of bytes copied to user in last RTT */
569 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
570 if (copied
<= tp
->rcvq_space
.space
)
574 * copied = bytes received in previous RTT, our base window
575 * To cope with packet losses, we need a 2x factor
576 * To cope with slow start, and sender growing its cwin by 100 %
577 * every RTT, we need a 4x factor, because the ACK we are sending
578 * now is for the next RTT, not the current one :
579 * <prev RTT . ><current RTT .. ><next RTT .... >
582 if (sysctl_tcp_moderate_rcvbuf
&&
583 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
584 int rcvwin
, rcvmem
, rcvbuf
;
586 /* minimal window to cope with packet losses, assuming
587 * steady state. Add some cushion because of small variations.
589 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
591 /* If rate increased by 25%,
592 * assume slow start, rcvwin = 3 * copied
593 * If rate increased by 50%,
594 * assume sender can use 2x growth, rcvwin = 4 * copied
597 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
599 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
602 rcvwin
+= (rcvwin
>> 1);
605 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
606 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
609 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
610 if (rcvbuf
> sk
->sk_rcvbuf
) {
611 sk
->sk_rcvbuf
= rcvbuf
;
613 /* Make the window clamp follow along. */
614 tp
->window_clamp
= rcvwin
;
617 tp
->rcvq_space
.space
= copied
;
620 tp
->rcvq_space
.seq
= tp
->copied_seq
;
621 tp
->rcvq_space
.time
= tcp_time_stamp
;
624 /* There is something which you must keep in mind when you analyze the
625 * behavior of the tp->ato delayed ack timeout interval. When a
626 * connection starts up, we want to ack as quickly as possible. The
627 * problem is that "good" TCP's do slow start at the beginning of data
628 * transmission. The means that until we send the first few ACK's the
629 * sender will sit on his end and only queue most of his data, because
630 * he can only send snd_cwnd unacked packets at any given time. For
631 * each ACK we send, he increments snd_cwnd and transmits more of his
634 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
636 struct tcp_sock
*tp
= tcp_sk(sk
);
637 struct inet_connection_sock
*icsk
= inet_csk(sk
);
640 inet_csk_schedule_ack(sk
);
642 tcp_measure_rcv_mss(sk
, skb
);
644 tcp_rcv_rtt_measure(tp
);
646 now
= tcp_time_stamp
;
648 if (!icsk
->icsk_ack
.ato
) {
649 /* The _first_ data packet received, initialize
650 * delayed ACK engine.
652 tcp_incr_quickack(sk
);
653 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
655 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
657 if (m
<= TCP_ATO_MIN
/ 2) {
658 /* The fastest case is the first. */
659 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
660 } else if (m
< icsk
->icsk_ack
.ato
) {
661 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
662 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
663 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
664 } else if (m
> icsk
->icsk_rto
) {
665 /* Too long gap. Apparently sender failed to
666 * restart window, so that we send ACKs quickly.
668 tcp_incr_quickack(sk
);
672 icsk
->icsk_ack
.lrcvtime
= now
;
674 tcp_ecn_check_ce(tp
, skb
);
677 tcp_grow_window(sk
, skb
);
680 /* Called to compute a smoothed rtt estimate. The data fed to this
681 * routine either comes from timestamps, or from segments that were
682 * known _not_ to have been retransmitted [see Karn/Partridge
683 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
684 * piece by Van Jacobson.
685 * NOTE: the next three routines used to be one big routine.
686 * To save cycles in the RFC 1323 implementation it was better to break
687 * it up into three procedures. -- erics
689 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
691 struct tcp_sock
*tp
= tcp_sk(sk
);
692 long m
= mrtt_us
; /* RTT */
693 u32 srtt
= tp
->srtt_us
;
695 /* The following amusing code comes from Jacobson's
696 * article in SIGCOMM '88. Note that rtt and mdev
697 * are scaled versions of rtt and mean deviation.
698 * This is designed to be as fast as possible
699 * m stands for "measurement".
701 * On a 1990 paper the rto value is changed to:
702 * RTO = rtt + 4 * mdev
704 * Funny. This algorithm seems to be very broken.
705 * These formulae increase RTO, when it should be decreased, increase
706 * too slowly, when it should be increased quickly, decrease too quickly
707 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
708 * does not matter how to _calculate_ it. Seems, it was trap
709 * that VJ failed to avoid. 8)
712 m
-= (srtt
>> 3); /* m is now error in rtt est */
713 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
715 m
= -m
; /* m is now abs(error) */
716 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
717 /* This is similar to one of Eifel findings.
718 * Eifel blocks mdev updates when rtt decreases.
719 * This solution is a bit different: we use finer gain
720 * for mdev in this case (alpha*beta).
721 * Like Eifel it also prevents growth of rto,
722 * but also it limits too fast rto decreases,
723 * happening in pure Eifel.
728 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
730 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
731 if (tp
->mdev_us
> tp
->mdev_max_us
) {
732 tp
->mdev_max_us
= tp
->mdev_us
;
733 if (tp
->mdev_max_us
> tp
->rttvar_us
)
734 tp
->rttvar_us
= tp
->mdev_max_us
;
736 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
737 if (tp
->mdev_max_us
< tp
->rttvar_us
)
738 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
739 tp
->rtt_seq
= tp
->snd_nxt
;
740 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
743 /* no previous measure. */
744 srtt
= m
<< 3; /* take the measured time to be rtt */
745 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
746 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
747 tp
->mdev_max_us
= tp
->rttvar_us
;
748 tp
->rtt_seq
= tp
->snd_nxt
;
750 tp
->srtt_us
= max(1U, srtt
);
753 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
754 * Note: TCP stack does not yet implement pacing.
755 * FQ packet scheduler can be used to implement cheap but effective
756 * TCP pacing, to smooth the burst on large writes when packets
757 * in flight is significantly lower than cwnd (or rwin)
759 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
760 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
762 static void tcp_update_pacing_rate(struct sock
*sk
)
764 const struct tcp_sock
*tp
= tcp_sk(sk
);
767 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
768 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
770 /* current rate is (cwnd * mss) / srtt
771 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
772 * In Congestion Avoidance phase, set it to 120 % the current rate.
774 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
775 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
776 * end of slow start and should slow down.
778 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
779 rate
*= sysctl_tcp_pacing_ss_ratio
;
781 rate
*= sysctl_tcp_pacing_ca_ratio
;
783 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
785 if (likely(tp
->srtt_us
))
786 do_div(rate
, tp
->srtt_us
);
788 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
789 * without any lock. We want to make sure compiler wont store
790 * intermediate values in this location.
792 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
793 sk
->sk_max_pacing_rate
);
796 /* Calculate rto without backoff. This is the second half of Van Jacobson's
797 * routine referred to above.
799 static void tcp_set_rto(struct sock
*sk
)
801 const struct tcp_sock
*tp
= tcp_sk(sk
);
802 /* Old crap is replaced with new one. 8)
805 * 1. If rtt variance happened to be less 50msec, it is hallucination.
806 * It cannot be less due to utterly erratic ACK generation made
807 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
808 * to do with delayed acks, because at cwnd>2 true delack timeout
809 * is invisible. Actually, Linux-2.4 also generates erratic
810 * ACKs in some circumstances.
812 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
814 /* 2. Fixups made earlier cannot be right.
815 * If we do not estimate RTO correctly without them,
816 * all the algo is pure shit and should be replaced
817 * with correct one. It is exactly, which we pretend to do.
820 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
821 * guarantees that rto is higher.
826 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
828 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
831 cwnd
= TCP_INIT_CWND
;
832 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
836 * Packet counting of FACK is based on in-order assumptions, therefore TCP
837 * disables it when reordering is detected
839 void tcp_disable_fack(struct tcp_sock
*tp
)
841 /* RFC3517 uses different metric in lost marker => reset on change */
843 tp
->lost_skb_hint
= NULL
;
844 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
847 /* Take a notice that peer is sending D-SACKs */
848 static void tcp_dsack_seen(struct tcp_sock
*tp
)
850 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
853 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
856 struct tcp_sock
*tp
= tcp_sk(sk
);
857 if (metric
> tp
->reordering
) {
860 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
862 /* This exciting event is worth to be remembered. 8) */
864 mib_idx
= LINUX_MIB_TCPTSREORDER
;
865 else if (tcp_is_reno(tp
))
866 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
867 else if (tcp_is_fack(tp
))
868 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
870 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
872 NET_INC_STATS(sock_net(sk
), mib_idx
);
873 #if FASTRETRANS_DEBUG > 1
874 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
875 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
879 tp
->undo_marker
? tp
->undo_retrans
: 0);
881 tcp_disable_fack(tp
);
885 tcp_disable_early_retrans(tp
);
889 /* This must be called before lost_out is incremented */
890 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
892 if (!tp
->retransmit_skb_hint
||
893 before(TCP_SKB_CB(skb
)->seq
,
894 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
895 tp
->retransmit_skb_hint
= skb
;
898 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
899 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
902 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
904 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
905 tcp_verify_retransmit_hint(tp
, skb
);
907 tp
->lost_out
+= tcp_skb_pcount(skb
);
908 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
912 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
914 tcp_verify_retransmit_hint(tp
, skb
);
916 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
917 tp
->lost_out
+= tcp_skb_pcount(skb
);
918 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
922 /* This procedure tags the retransmission queue when SACKs arrive.
924 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
925 * Packets in queue with these bits set are counted in variables
926 * sacked_out, retrans_out and lost_out, correspondingly.
928 * Valid combinations are:
929 * Tag InFlight Description
930 * 0 1 - orig segment is in flight.
931 * S 0 - nothing flies, orig reached receiver.
932 * L 0 - nothing flies, orig lost by net.
933 * R 2 - both orig and retransmit are in flight.
934 * L|R 1 - orig is lost, retransmit is in flight.
935 * S|R 1 - orig reached receiver, retrans is still in flight.
936 * (L|S|R is logically valid, it could occur when L|R is sacked,
937 * but it is equivalent to plain S and code short-curcuits it to S.
938 * L|S is logically invalid, it would mean -1 packet in flight 8))
940 * These 6 states form finite state machine, controlled by the following events:
941 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
942 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
943 * 3. Loss detection event of two flavors:
944 * A. Scoreboard estimator decided the packet is lost.
945 * A'. Reno "three dupacks" marks head of queue lost.
946 * A''. Its FACK modification, head until snd.fack is lost.
947 * B. SACK arrives sacking SND.NXT at the moment, when the
948 * segment was retransmitted.
949 * 4. D-SACK added new rule: D-SACK changes any tag to S.
951 * It is pleasant to note, that state diagram turns out to be commutative,
952 * so that we are allowed not to be bothered by order of our actions,
953 * when multiple events arrive simultaneously. (see the function below).
955 * Reordering detection.
956 * --------------------
957 * Reordering metric is maximal distance, which a packet can be displaced
958 * in packet stream. With SACKs we can estimate it:
960 * 1. SACK fills old hole and the corresponding segment was not
961 * ever retransmitted -> reordering. Alas, we cannot use it
962 * when segment was retransmitted.
963 * 2. The last flaw is solved with D-SACK. D-SACK arrives
964 * for retransmitted and already SACKed segment -> reordering..
965 * Both of these heuristics are not used in Loss state, when we cannot
966 * account for retransmits accurately.
968 * SACK block validation.
969 * ----------------------
971 * SACK block range validation checks that the received SACK block fits to
972 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
973 * Note that SND.UNA is not included to the range though being valid because
974 * it means that the receiver is rather inconsistent with itself reporting
975 * SACK reneging when it should advance SND.UNA. Such SACK block this is
976 * perfectly valid, however, in light of RFC2018 which explicitly states
977 * that "SACK block MUST reflect the newest segment. Even if the newest
978 * segment is going to be discarded ...", not that it looks very clever
979 * in case of head skb. Due to potentional receiver driven attacks, we
980 * choose to avoid immediate execution of a walk in write queue due to
981 * reneging and defer head skb's loss recovery to standard loss recovery
982 * procedure that will eventually trigger (nothing forbids us doing this).
984 * Implements also blockage to start_seq wrap-around. Problem lies in the
985 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
986 * there's no guarantee that it will be before snd_nxt (n). The problem
987 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
990 * <- outs wnd -> <- wrapzone ->
991 * u e n u_w e_w s n_w
993 * |<------------+------+----- TCP seqno space --------------+---------->|
994 * ...-- <2^31 ->| |<--------...
995 * ...---- >2^31 ------>| |<--------...
997 * Current code wouldn't be vulnerable but it's better still to discard such
998 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
999 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1000 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1001 * equal to the ideal case (infinite seqno space without wrap caused issues).
1003 * With D-SACK the lower bound is extended to cover sequence space below
1004 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1005 * again, D-SACK block must not to go across snd_una (for the same reason as
1006 * for the normal SACK blocks, explained above). But there all simplicity
1007 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1008 * fully below undo_marker they do not affect behavior in anyway and can
1009 * therefore be safely ignored. In rare cases (which are more or less
1010 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1011 * fragmentation and packet reordering past skb's retransmission. To consider
1012 * them correctly, the acceptable range must be extended even more though
1013 * the exact amount is rather hard to quantify. However, tp->max_window can
1014 * be used as an exaggerated estimate.
1016 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1017 u32 start_seq
, u32 end_seq
)
1019 /* Too far in future, or reversed (interpretation is ambiguous) */
1020 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1023 /* Nasty start_seq wrap-around check (see comments above) */
1024 if (!before(start_seq
, tp
->snd_nxt
))
1027 /* In outstanding window? ...This is valid exit for D-SACKs too.
1028 * start_seq == snd_una is non-sensical (see comments above)
1030 if (after(start_seq
, tp
->snd_una
))
1033 if (!is_dsack
|| !tp
->undo_marker
)
1036 /* ...Then it's D-SACK, and must reside below snd_una completely */
1037 if (after(end_seq
, tp
->snd_una
))
1040 if (!before(start_seq
, tp
->undo_marker
))
1044 if (!after(end_seq
, tp
->undo_marker
))
1047 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1048 * start_seq < undo_marker and end_seq >= undo_marker.
1050 return !before(start_seq
, end_seq
- tp
->max_window
);
1053 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1054 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1057 struct tcp_sock
*tp
= tcp_sk(sk
);
1058 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1059 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1060 bool dup_sack
= false;
1062 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1065 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1066 } else if (num_sacks
> 1) {
1067 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1068 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1070 if (!after(end_seq_0
, end_seq_1
) &&
1071 !before(start_seq_0
, start_seq_1
)) {
1074 NET_INC_STATS(sock_net(sk
),
1075 LINUX_MIB_TCPDSACKOFORECV
);
1079 /* D-SACK for already forgotten data... Do dumb counting. */
1080 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1081 !after(end_seq_0
, prior_snd_una
) &&
1082 after(end_seq_0
, tp
->undo_marker
))
1088 struct tcp_sacktag_state
{
1091 /* Timestamps for earliest and latest never-retransmitted segment
1092 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1093 * but congestion control should still get an accurate delay signal.
1095 struct skb_mstamp first_sackt
;
1096 struct skb_mstamp last_sackt
;
1100 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1101 * the incoming SACK may not exactly match but we can find smaller MSS
1102 * aligned portion of it that matches. Therefore we might need to fragment
1103 * which may fail and creates some hassle (caller must handle error case
1106 * FIXME: this could be merged to shift decision code
1108 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1109 u32 start_seq
, u32 end_seq
)
1113 unsigned int pkt_len
;
1116 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1117 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1119 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1120 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1121 mss
= tcp_skb_mss(skb
);
1122 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1125 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1129 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1134 /* Round if necessary so that SACKs cover only full MSSes
1135 * and/or the remaining small portion (if present)
1137 if (pkt_len
> mss
) {
1138 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1139 if (!in_sack
&& new_len
< pkt_len
) {
1141 if (new_len
>= skb
->len
)
1146 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1154 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1155 static u8
tcp_sacktag_one(struct sock
*sk
,
1156 struct tcp_sacktag_state
*state
, u8 sacked
,
1157 u32 start_seq
, u32 end_seq
,
1158 int dup_sack
, int pcount
,
1159 const struct skb_mstamp
*xmit_time
)
1161 struct tcp_sock
*tp
= tcp_sk(sk
);
1162 int fack_count
= state
->fack_count
;
1164 /* Account D-SACK for retransmitted packet. */
1165 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1166 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1167 after(end_seq
, tp
->undo_marker
))
1169 if (sacked
& TCPCB_SACKED_ACKED
)
1170 state
->reord
= min(fack_count
, state
->reord
);
1173 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1174 if (!after(end_seq
, tp
->snd_una
))
1177 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1178 tcp_rack_advance(tp
, xmit_time
, sacked
);
1180 if (sacked
& TCPCB_SACKED_RETRANS
) {
1181 /* If the segment is not tagged as lost,
1182 * we do not clear RETRANS, believing
1183 * that retransmission is still in flight.
1185 if (sacked
& TCPCB_LOST
) {
1186 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1187 tp
->lost_out
-= pcount
;
1188 tp
->retrans_out
-= pcount
;
1191 if (!(sacked
& TCPCB_RETRANS
)) {
1192 /* New sack for not retransmitted frame,
1193 * which was in hole. It is reordering.
1195 if (before(start_seq
,
1196 tcp_highest_sack_seq(tp
)))
1197 state
->reord
= min(fack_count
,
1199 if (!after(end_seq
, tp
->high_seq
))
1200 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1201 if (state
->first_sackt
.v64
== 0)
1202 state
->first_sackt
= *xmit_time
;
1203 state
->last_sackt
= *xmit_time
;
1206 if (sacked
& TCPCB_LOST
) {
1207 sacked
&= ~TCPCB_LOST
;
1208 tp
->lost_out
-= pcount
;
1212 sacked
|= TCPCB_SACKED_ACKED
;
1213 state
->flag
|= FLAG_DATA_SACKED
;
1214 tp
->sacked_out
+= pcount
;
1215 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1217 fack_count
+= pcount
;
1219 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1220 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1221 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1222 tp
->lost_cnt_hint
+= pcount
;
1224 if (fack_count
> tp
->fackets_out
)
1225 tp
->fackets_out
= fack_count
;
1228 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1229 * frames and clear it. undo_retrans is decreased above, L|R frames
1230 * are accounted above as well.
1232 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1233 sacked
&= ~TCPCB_SACKED_RETRANS
;
1234 tp
->retrans_out
-= pcount
;
1240 /* Shift newly-SACKed bytes from this skb to the immediately previous
1241 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1243 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1244 struct tcp_sacktag_state
*state
,
1245 unsigned int pcount
, int shifted
, int mss
,
1248 struct tcp_sock
*tp
= tcp_sk(sk
);
1249 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1250 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1251 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1255 /* Adjust counters and hints for the newly sacked sequence
1256 * range but discard the return value since prev is already
1257 * marked. We must tag the range first because the seq
1258 * advancement below implicitly advances
1259 * tcp_highest_sack_seq() when skb is highest_sack.
1261 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1262 start_seq
, end_seq
, dup_sack
, pcount
,
1265 if (skb
== tp
->lost_skb_hint
)
1266 tp
->lost_cnt_hint
+= pcount
;
1268 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1269 TCP_SKB_CB(skb
)->seq
+= shifted
;
1271 tcp_skb_pcount_add(prev
, pcount
);
1272 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1273 tcp_skb_pcount_add(skb
, -pcount
);
1275 /* When we're adding to gso_segs == 1, gso_size will be zero,
1276 * in theory this shouldn't be necessary but as long as DSACK
1277 * code can come after this skb later on it's better to keep
1278 * setting gso_size to something.
1280 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1281 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1283 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1284 if (tcp_skb_pcount(skb
) <= 1)
1285 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1287 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1288 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1291 BUG_ON(!tcp_skb_pcount(skb
));
1292 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1296 /* Whole SKB was eaten :-) */
1298 if (skb
== tp
->retransmit_skb_hint
)
1299 tp
->retransmit_skb_hint
= prev
;
1300 if (skb
== tp
->lost_skb_hint
) {
1301 tp
->lost_skb_hint
= prev
;
1302 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1305 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1306 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1307 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1308 TCP_SKB_CB(prev
)->end_seq
++;
1310 if (skb
== tcp_highest_sack(sk
))
1311 tcp_advance_highest_sack(sk
, skb
);
1313 tcp_skb_collapse_tstamp(prev
, skb
);
1314 tcp_unlink_write_queue(skb
, sk
);
1315 sk_wmem_free_skb(sk
, skb
);
1317 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1322 /* I wish gso_size would have a bit more sane initialization than
1323 * something-or-zero which complicates things
1325 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1327 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1330 /* Shifting pages past head area doesn't work */
1331 static int skb_can_shift(const struct sk_buff
*skb
)
1333 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1336 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1339 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1340 struct tcp_sacktag_state
*state
,
1341 u32 start_seq
, u32 end_seq
,
1344 struct tcp_sock
*tp
= tcp_sk(sk
);
1345 struct sk_buff
*prev
;
1351 if (!sk_can_gso(sk
))
1354 /* Normally R but no L won't result in plain S */
1356 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1358 if (!skb_can_shift(skb
))
1360 /* This frame is about to be dropped (was ACKed). */
1361 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1364 /* Can only happen with delayed DSACK + discard craziness */
1365 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1367 prev
= tcp_write_queue_prev(sk
, skb
);
1369 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1372 if (!tcp_skb_can_collapse_to(prev
))
1375 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1376 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1380 pcount
= tcp_skb_pcount(skb
);
1381 mss
= tcp_skb_seglen(skb
);
1383 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1384 * drop this restriction as unnecessary
1386 if (mss
!= tcp_skb_seglen(prev
))
1389 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1391 /* CHECKME: This is non-MSS split case only?, this will
1392 * cause skipped skbs due to advancing loop btw, original
1393 * has that feature too
1395 if (tcp_skb_pcount(skb
) <= 1)
1398 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1400 /* TODO: head merge to next could be attempted here
1401 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1402 * though it might not be worth of the additional hassle
1404 * ...we can probably just fallback to what was done
1405 * previously. We could try merging non-SACKed ones
1406 * as well but it probably isn't going to buy off
1407 * because later SACKs might again split them, and
1408 * it would make skb timestamp tracking considerably
1414 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1416 BUG_ON(len
> skb
->len
);
1418 /* MSS boundaries should be honoured or else pcount will
1419 * severely break even though it makes things bit trickier.
1420 * Optimize common case to avoid most of the divides
1422 mss
= tcp_skb_mss(skb
);
1424 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1425 * drop this restriction as unnecessary
1427 if (mss
!= tcp_skb_seglen(prev
))
1432 } else if (len
< mss
) {
1440 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1441 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1444 if (!skb_shift(prev
, skb
, len
))
1446 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1449 /* Hole filled allows collapsing with the next as well, this is very
1450 * useful when hole on every nth skb pattern happens
1452 if (prev
== tcp_write_queue_tail(sk
))
1454 skb
= tcp_write_queue_next(sk
, prev
);
1456 if (!skb_can_shift(skb
) ||
1457 (skb
== tcp_send_head(sk
)) ||
1458 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1459 (mss
!= tcp_skb_seglen(skb
)))
1463 if (skb_shift(prev
, skb
, len
)) {
1464 pcount
+= tcp_skb_pcount(skb
);
1465 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1469 state
->fack_count
+= pcount
;
1476 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1480 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1481 struct tcp_sack_block
*next_dup
,
1482 struct tcp_sacktag_state
*state
,
1483 u32 start_seq
, u32 end_seq
,
1486 struct tcp_sock
*tp
= tcp_sk(sk
);
1487 struct sk_buff
*tmp
;
1489 tcp_for_write_queue_from(skb
, sk
) {
1491 bool dup_sack
= dup_sack_in
;
1493 if (skb
== tcp_send_head(sk
))
1496 /* queue is in-order => we can short-circuit the walk early */
1497 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1501 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1502 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1503 next_dup
->start_seq
,
1509 /* skb reference here is a bit tricky to get right, since
1510 * shifting can eat and free both this skb and the next,
1511 * so not even _safe variant of the loop is enough.
1514 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1515 start_seq
, end_seq
, dup_sack
);
1524 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1530 if (unlikely(in_sack
< 0))
1534 TCP_SKB_CB(skb
)->sacked
=
1537 TCP_SKB_CB(skb
)->sacked
,
1538 TCP_SKB_CB(skb
)->seq
,
1539 TCP_SKB_CB(skb
)->end_seq
,
1541 tcp_skb_pcount(skb
),
1544 if (!before(TCP_SKB_CB(skb
)->seq
,
1545 tcp_highest_sack_seq(tp
)))
1546 tcp_advance_highest_sack(sk
, skb
);
1549 state
->fack_count
+= tcp_skb_pcount(skb
);
1554 /* Avoid all extra work that is being done by sacktag while walking in
1557 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1558 struct tcp_sacktag_state
*state
,
1561 tcp_for_write_queue_from(skb
, sk
) {
1562 if (skb
== tcp_send_head(sk
))
1565 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1568 state
->fack_count
+= tcp_skb_pcount(skb
);
1573 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1575 struct tcp_sack_block
*next_dup
,
1576 struct tcp_sacktag_state
*state
,
1582 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1583 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1584 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1585 next_dup
->start_seq
, next_dup
->end_seq
,
1592 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1594 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1598 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1599 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1601 struct tcp_sock
*tp
= tcp_sk(sk
);
1602 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1603 TCP_SKB_CB(ack_skb
)->sacked
);
1604 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1605 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1606 struct tcp_sack_block
*cache
;
1607 struct sk_buff
*skb
;
1608 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1610 bool found_dup_sack
= false;
1612 int first_sack_index
;
1615 state
->reord
= tp
->packets_out
;
1617 if (!tp
->sacked_out
) {
1618 if (WARN_ON(tp
->fackets_out
))
1619 tp
->fackets_out
= 0;
1620 tcp_highest_sack_reset(sk
);
1623 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1624 num_sacks
, prior_snd_una
);
1626 state
->flag
|= FLAG_DSACKING_ACK
;
1628 /* Eliminate too old ACKs, but take into
1629 * account more or less fresh ones, they can
1630 * contain valid SACK info.
1632 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1635 if (!tp
->packets_out
)
1639 first_sack_index
= 0;
1640 for (i
= 0; i
< num_sacks
; i
++) {
1641 bool dup_sack
= !i
&& found_dup_sack
;
1643 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1644 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1646 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1647 sp
[used_sacks
].start_seq
,
1648 sp
[used_sacks
].end_seq
)) {
1652 if (!tp
->undo_marker
)
1653 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1655 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1657 /* Don't count olds caused by ACK reordering */
1658 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1659 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1661 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1664 NET_INC_STATS(sock_net(sk
), mib_idx
);
1666 first_sack_index
= -1;
1670 /* Ignore very old stuff early */
1671 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1677 /* order SACK blocks to allow in order walk of the retrans queue */
1678 for (i
= used_sacks
- 1; i
> 0; i
--) {
1679 for (j
= 0; j
< i
; j
++) {
1680 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1681 swap(sp
[j
], sp
[j
+ 1]);
1683 /* Track where the first SACK block goes to */
1684 if (j
== first_sack_index
)
1685 first_sack_index
= j
+ 1;
1690 skb
= tcp_write_queue_head(sk
);
1691 state
->fack_count
= 0;
1694 if (!tp
->sacked_out
) {
1695 /* It's already past, so skip checking against it */
1696 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1698 cache
= tp
->recv_sack_cache
;
1699 /* Skip empty blocks in at head of the cache */
1700 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1705 while (i
< used_sacks
) {
1706 u32 start_seq
= sp
[i
].start_seq
;
1707 u32 end_seq
= sp
[i
].end_seq
;
1708 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1709 struct tcp_sack_block
*next_dup
= NULL
;
1711 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1712 next_dup
= &sp
[i
+ 1];
1714 /* Skip too early cached blocks */
1715 while (tcp_sack_cache_ok(tp
, cache
) &&
1716 !before(start_seq
, cache
->end_seq
))
1719 /* Can skip some work by looking recv_sack_cache? */
1720 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1721 after(end_seq
, cache
->start_seq
)) {
1724 if (before(start_seq
, cache
->start_seq
)) {
1725 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1727 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1734 /* Rest of the block already fully processed? */
1735 if (!after(end_seq
, cache
->end_seq
))
1738 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1742 /* ...tail remains todo... */
1743 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1744 /* ...but better entrypoint exists! */
1745 skb
= tcp_highest_sack(sk
);
1748 state
->fack_count
= tp
->fackets_out
;
1753 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1754 /* Check overlap against next cached too (past this one already) */
1759 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1760 skb
= tcp_highest_sack(sk
);
1763 state
->fack_count
= tp
->fackets_out
;
1765 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1768 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1769 start_seq
, end_seq
, dup_sack
);
1775 /* Clear the head of the cache sack blocks so we can skip it next time */
1776 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1777 tp
->recv_sack_cache
[i
].start_seq
= 0;
1778 tp
->recv_sack_cache
[i
].end_seq
= 0;
1780 for (j
= 0; j
< used_sacks
; j
++)
1781 tp
->recv_sack_cache
[i
++] = sp
[j
];
1783 if ((state
->reord
< tp
->fackets_out
) &&
1784 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1785 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1787 tcp_verify_left_out(tp
);
1790 #if FASTRETRANS_DEBUG > 0
1791 WARN_ON((int)tp
->sacked_out
< 0);
1792 WARN_ON((int)tp
->lost_out
< 0);
1793 WARN_ON((int)tp
->retrans_out
< 0);
1794 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1799 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1800 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1802 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1806 holes
= max(tp
->lost_out
, 1U);
1807 holes
= min(holes
, tp
->packets_out
);
1809 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1810 tp
->sacked_out
= tp
->packets_out
- holes
;
1816 /* If we receive more dupacks than we expected counting segments
1817 * in assumption of absent reordering, interpret this as reordering.
1818 * The only another reason could be bug in receiver TCP.
1820 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1822 struct tcp_sock
*tp
= tcp_sk(sk
);
1823 if (tcp_limit_reno_sacked(tp
))
1824 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1827 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1829 static void tcp_add_reno_sack(struct sock
*sk
)
1831 struct tcp_sock
*tp
= tcp_sk(sk
);
1832 u32 prior_sacked
= tp
->sacked_out
;
1835 tcp_check_reno_reordering(sk
, 0);
1836 if (tp
->sacked_out
> prior_sacked
)
1837 tp
->delivered
++; /* Some out-of-order packet is delivered */
1838 tcp_verify_left_out(tp
);
1841 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1843 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1845 struct tcp_sock
*tp
= tcp_sk(sk
);
1848 /* One ACK acked hole. The rest eat duplicate ACKs. */
1849 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1850 if (acked
- 1 >= tp
->sacked_out
)
1853 tp
->sacked_out
-= acked
- 1;
1855 tcp_check_reno_reordering(sk
, acked
);
1856 tcp_verify_left_out(tp
);
1859 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1864 void tcp_clear_retrans(struct tcp_sock
*tp
)
1866 tp
->retrans_out
= 0;
1868 tp
->undo_marker
= 0;
1869 tp
->undo_retrans
= -1;
1870 tp
->fackets_out
= 0;
1874 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1876 tp
->undo_marker
= tp
->snd_una
;
1877 /* Retransmission still in flight may cause DSACKs later. */
1878 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1881 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1882 * and reset tags completely, otherwise preserve SACKs. If receiver
1883 * dropped its ofo queue, we will know this due to reneging detection.
1885 void tcp_enter_loss(struct sock
*sk
)
1887 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1888 struct tcp_sock
*tp
= tcp_sk(sk
);
1889 struct net
*net
= sock_net(sk
);
1890 struct sk_buff
*skb
;
1891 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1892 bool is_reneg
; /* is receiver reneging on SACKs? */
1894 /* Reduce ssthresh if it has not yet been made inside this window. */
1895 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1896 !after(tp
->high_seq
, tp
->snd_una
) ||
1897 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1898 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1899 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1900 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1904 tp
->snd_cwnd_cnt
= 0;
1905 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1907 tp
->retrans_out
= 0;
1910 if (tcp_is_reno(tp
))
1911 tcp_reset_reno_sack(tp
);
1913 skb
= tcp_write_queue_head(sk
);
1914 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1916 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1918 tp
->fackets_out
= 0;
1920 tcp_clear_all_retrans_hints(tp
);
1922 tcp_for_write_queue(skb
, sk
) {
1923 if (skb
== tcp_send_head(sk
))
1926 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1927 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1928 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1929 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1930 tp
->lost_out
+= tcp_skb_pcount(skb
);
1931 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1934 tcp_verify_left_out(tp
);
1936 /* Timeout in disordered state after receiving substantial DUPACKs
1937 * suggests that the degree of reordering is over-estimated.
1939 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1940 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1941 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1942 net
->ipv4
.sysctl_tcp_reordering
);
1943 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1944 tp
->high_seq
= tp
->snd_nxt
;
1945 tcp_ecn_queue_cwr(tp
);
1947 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1948 * loss recovery is underway except recurring timeout(s) on
1949 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1951 tp
->frto
= sysctl_tcp_frto
&&
1952 (new_recovery
|| icsk
->icsk_retransmits
) &&
1953 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1956 /* If ACK arrived pointing to a remembered SACK, it means that our
1957 * remembered SACKs do not reflect real state of receiver i.e.
1958 * receiver _host_ is heavily congested (or buggy).
1960 * To avoid big spurious retransmission bursts due to transient SACK
1961 * scoreboard oddities that look like reneging, we give the receiver a
1962 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1963 * restore sanity to the SACK scoreboard. If the apparent reneging
1964 * persists until this RTO then we'll clear the SACK scoreboard.
1966 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1968 if (flag
& FLAG_SACK_RENEGING
) {
1969 struct tcp_sock
*tp
= tcp_sk(sk
);
1970 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
1971 msecs_to_jiffies(10));
1973 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1974 delay
, TCP_RTO_MAX
);
1980 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1982 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1985 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1986 * counter when SACK is enabled (without SACK, sacked_out is used for
1989 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1990 * segments up to the highest received SACK block so far and holes in
1993 * With reordering, holes may still be in flight, so RFC3517 recovery
1994 * uses pure sacked_out (total number of SACKed segments) even though
1995 * it violates the RFC that uses duplicate ACKs, often these are equal
1996 * but when e.g. out-of-window ACKs or packet duplication occurs,
1997 * they differ. Since neither occurs due to loss, TCP should really
2000 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2002 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2005 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2007 struct tcp_sock
*tp
= tcp_sk(sk
);
2008 unsigned long delay
;
2010 /* Delay early retransmit and entering fast recovery for
2011 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2012 * available, or RTO is scheduled to fire first.
2014 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2015 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2018 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2019 msecs_to_jiffies(2));
2021 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2024 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2029 /* Linux NewReno/SACK/FACK/ECN state machine.
2030 * --------------------------------------
2032 * "Open" Normal state, no dubious events, fast path.
2033 * "Disorder" In all the respects it is "Open",
2034 * but requires a bit more attention. It is entered when
2035 * we see some SACKs or dupacks. It is split of "Open"
2036 * mainly to move some processing from fast path to slow one.
2037 * "CWR" CWND was reduced due to some Congestion Notification event.
2038 * It can be ECN, ICMP source quench, local device congestion.
2039 * "Recovery" CWND was reduced, we are fast-retransmitting.
2040 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2042 * tcp_fastretrans_alert() is entered:
2043 * - each incoming ACK, if state is not "Open"
2044 * - when arrived ACK is unusual, namely:
2049 * Counting packets in flight is pretty simple.
2051 * in_flight = packets_out - left_out + retrans_out
2053 * packets_out is SND.NXT-SND.UNA counted in packets.
2055 * retrans_out is number of retransmitted segments.
2057 * left_out is number of segments left network, but not ACKed yet.
2059 * left_out = sacked_out + lost_out
2061 * sacked_out: Packets, which arrived to receiver out of order
2062 * and hence not ACKed. With SACKs this number is simply
2063 * amount of SACKed data. Even without SACKs
2064 * it is easy to give pretty reliable estimate of this number,
2065 * counting duplicate ACKs.
2067 * lost_out: Packets lost by network. TCP has no explicit
2068 * "loss notification" feedback from network (for now).
2069 * It means that this number can be only _guessed_.
2070 * Actually, it is the heuristics to predict lossage that
2071 * distinguishes different algorithms.
2073 * F.e. after RTO, when all the queue is considered as lost,
2074 * lost_out = packets_out and in_flight = retrans_out.
2076 * Essentially, we have now two algorithms counting
2079 * FACK: It is the simplest heuristics. As soon as we decided
2080 * that something is lost, we decide that _all_ not SACKed
2081 * packets until the most forward SACK are lost. I.e.
2082 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2083 * It is absolutely correct estimate, if network does not reorder
2084 * packets. And it loses any connection to reality when reordering
2085 * takes place. We use FACK by default until reordering
2086 * is suspected on the path to this destination.
2088 * NewReno: when Recovery is entered, we assume that one segment
2089 * is lost (classic Reno). While we are in Recovery and
2090 * a partial ACK arrives, we assume that one more packet
2091 * is lost (NewReno). This heuristics are the same in NewReno
2094 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2095 * deflation etc. CWND is real congestion window, never inflated, changes
2096 * only according to classic VJ rules.
2098 * Really tricky (and requiring careful tuning) part of algorithm
2099 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2100 * The first determines the moment _when_ we should reduce CWND and,
2101 * hence, slow down forward transmission. In fact, it determines the moment
2102 * when we decide that hole is caused by loss, rather than by a reorder.
2104 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2105 * holes, caused by lost packets.
2107 * And the most logically complicated part of algorithm is undo
2108 * heuristics. We detect false retransmits due to both too early
2109 * fast retransmit (reordering) and underestimated RTO, analyzing
2110 * timestamps and D-SACKs. When we detect that some segments were
2111 * retransmitted by mistake and CWND reduction was wrong, we undo
2112 * window reduction and abort recovery phase. This logic is hidden
2113 * inside several functions named tcp_try_undo_<something>.
2116 /* This function decides, when we should leave Disordered state
2117 * and enter Recovery phase, reducing congestion window.
2119 * Main question: may we further continue forward transmission
2120 * with the same cwnd?
2122 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2124 struct tcp_sock
*tp
= tcp_sk(sk
);
2126 int tcp_reordering
= sock_net(sk
)->ipv4
.sysctl_tcp_reordering
;
2128 /* Trick#1: The loss is proven. */
2132 /* Not-A-Trick#2 : Classic rule... */
2133 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2136 /* Trick#4: It is still not OK... But will it be useful to delay
2139 packets_out
= tp
->packets_out
;
2140 if (packets_out
<= tp
->reordering
&&
2141 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, tcp_reordering
) &&
2142 !tcp_may_send_now(sk
)) {
2143 /* We have nothing to send. This connection is limited
2144 * either by receiver window or by application.
2149 /* If a thin stream is detected, retransmit after first
2150 * received dupack. Employ only if SACK is supported in order
2151 * to avoid possible corner-case series of spurious retransmissions
2152 * Use only if there are no unsent data.
2154 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2155 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2156 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2159 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2160 * retransmissions due to small network reorderings, we implement
2161 * Mitigation A.3 in the RFC and delay the retransmission for a short
2162 * interval if appropriate.
2164 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2165 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2166 !tcp_may_send_now(sk
))
2167 return !tcp_pause_early_retransmit(sk
, flag
);
2172 /* Detect loss in event "A" above by marking head of queue up as lost.
2173 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2174 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2175 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2176 * the maximum SACKed segments to pass before reaching this limit.
2178 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2180 struct tcp_sock
*tp
= tcp_sk(sk
);
2181 struct sk_buff
*skb
;
2182 int cnt
, oldcnt
, lost
;
2184 /* Use SACK to deduce losses of new sequences sent during recovery */
2185 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2187 WARN_ON(packets
> tp
->packets_out
);
2188 if (tp
->lost_skb_hint
) {
2189 skb
= tp
->lost_skb_hint
;
2190 cnt
= tp
->lost_cnt_hint
;
2191 /* Head already handled? */
2192 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2195 skb
= tcp_write_queue_head(sk
);
2199 tcp_for_write_queue_from(skb
, sk
) {
2200 if (skb
== tcp_send_head(sk
))
2202 /* TODO: do this better */
2203 /* this is not the most efficient way to do this... */
2204 tp
->lost_skb_hint
= skb
;
2205 tp
->lost_cnt_hint
= cnt
;
2207 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2211 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2212 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2213 cnt
+= tcp_skb_pcount(skb
);
2215 if (cnt
> packets
) {
2216 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2217 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2218 (oldcnt
>= packets
))
2221 mss
= tcp_skb_mss(skb
);
2222 /* If needed, chop off the prefix to mark as lost. */
2223 lost
= (packets
- oldcnt
) * mss
;
2224 if (lost
< skb
->len
&&
2225 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2230 tcp_skb_mark_lost(tp
, skb
);
2235 tcp_verify_left_out(tp
);
2238 /* Account newly detected lost packet(s) */
2240 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2242 struct tcp_sock
*tp
= tcp_sk(sk
);
2244 if (tcp_is_reno(tp
)) {
2245 tcp_mark_head_lost(sk
, 1, 1);
2246 } else if (tcp_is_fack(tp
)) {
2247 int lost
= tp
->fackets_out
- tp
->reordering
;
2250 tcp_mark_head_lost(sk
, lost
, 0);
2252 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2253 if (sacked_upto
>= 0)
2254 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2255 else if (fast_rexmit
)
2256 tcp_mark_head_lost(sk
, 1, 1);
2260 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2262 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2263 before(tp
->rx_opt
.rcv_tsecr
, when
);
2266 /* skb is spurious retransmitted if the returned timestamp echo
2267 * reply is prior to the skb transmission time
2269 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2270 const struct sk_buff
*skb
)
2272 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2273 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2276 /* Nothing was retransmitted or returned timestamp is less
2277 * than timestamp of the first retransmission.
2279 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2281 return !tp
->retrans_stamp
||
2282 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2285 /* Undo procedures. */
2287 /* We can clear retrans_stamp when there are no retransmissions in the
2288 * window. It would seem that it is trivially available for us in
2289 * tp->retrans_out, however, that kind of assumptions doesn't consider
2290 * what will happen if errors occur when sending retransmission for the
2291 * second time. ...It could the that such segment has only
2292 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2293 * the head skb is enough except for some reneging corner cases that
2294 * are not worth the effort.
2296 * Main reason for all this complexity is the fact that connection dying
2297 * time now depends on the validity of the retrans_stamp, in particular,
2298 * that successive retransmissions of a segment must not advance
2299 * retrans_stamp under any conditions.
2301 static bool tcp_any_retrans_done(const struct sock
*sk
)
2303 const struct tcp_sock
*tp
= tcp_sk(sk
);
2304 struct sk_buff
*skb
;
2306 if (tp
->retrans_out
)
2309 skb
= tcp_write_queue_head(sk
);
2310 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2316 #if FASTRETRANS_DEBUG > 1
2317 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2319 struct tcp_sock
*tp
= tcp_sk(sk
);
2320 struct inet_sock
*inet
= inet_sk(sk
);
2322 if (sk
->sk_family
== AF_INET
) {
2323 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2325 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2326 tp
->snd_cwnd
, tcp_left_out(tp
),
2327 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2330 #if IS_ENABLED(CONFIG_IPV6)
2331 else if (sk
->sk_family
== AF_INET6
) {
2332 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2333 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2335 &np
->daddr
, ntohs(inet
->inet_dport
),
2336 tp
->snd_cwnd
, tcp_left_out(tp
),
2337 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2343 #define DBGUNDO(x...) do { } while (0)
2346 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2348 struct tcp_sock
*tp
= tcp_sk(sk
);
2351 struct sk_buff
*skb
;
2353 tcp_for_write_queue(skb
, sk
) {
2354 if (skb
== tcp_send_head(sk
))
2356 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2359 tcp_clear_all_retrans_hints(tp
);
2362 if (tp
->prior_ssthresh
) {
2363 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2365 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2366 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2368 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2370 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2371 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2372 tcp_ecn_withdraw_cwr(tp
);
2375 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2376 tp
->undo_marker
= 0;
2379 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2381 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2384 /* People celebrate: "We love our President!" */
2385 static bool tcp_try_undo_recovery(struct sock
*sk
)
2387 struct tcp_sock
*tp
= tcp_sk(sk
);
2389 if (tcp_may_undo(tp
)) {
2392 /* Happy end! We did not retransmit anything
2393 * or our original transmission succeeded.
2395 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2396 tcp_undo_cwnd_reduction(sk
, false);
2397 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2398 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2400 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2402 NET_INC_STATS(sock_net(sk
), mib_idx
);
2404 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2405 /* Hold old state until something *above* high_seq
2406 * is ACKed. For Reno it is MUST to prevent false
2407 * fast retransmits (RFC2582). SACK TCP is safe. */
2408 if (!tcp_any_retrans_done(sk
))
2409 tp
->retrans_stamp
= 0;
2412 tcp_set_ca_state(sk
, TCP_CA_Open
);
2416 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2417 static bool tcp_try_undo_dsack(struct sock
*sk
)
2419 struct tcp_sock
*tp
= tcp_sk(sk
);
2421 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2422 DBGUNDO(sk
, "D-SACK");
2423 tcp_undo_cwnd_reduction(sk
, false);
2424 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2430 /* Undo during loss recovery after partial ACK or using F-RTO. */
2431 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2433 struct tcp_sock
*tp
= tcp_sk(sk
);
2435 if (frto_undo
|| tcp_may_undo(tp
)) {
2436 tcp_undo_cwnd_reduction(sk
, true);
2438 DBGUNDO(sk
, "partial loss");
2439 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2441 NET_INC_STATS(sock_net(sk
),
2442 LINUX_MIB_TCPSPURIOUSRTOS
);
2443 inet_csk(sk
)->icsk_retransmits
= 0;
2444 if (frto_undo
|| tcp_is_sack(tp
))
2445 tcp_set_ca_state(sk
, TCP_CA_Open
);
2451 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2452 * It computes the number of packets to send (sndcnt) based on packets newly
2454 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2455 * cwnd reductions across a full RTT.
2456 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2457 * But when the retransmits are acked without further losses, PRR
2458 * slow starts cwnd up to ssthresh to speed up the recovery.
2460 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2462 struct tcp_sock
*tp
= tcp_sk(sk
);
2464 tp
->high_seq
= tp
->snd_nxt
;
2465 tp
->tlp_high_seq
= 0;
2466 tp
->snd_cwnd_cnt
= 0;
2467 tp
->prior_cwnd
= tp
->snd_cwnd
;
2468 tp
->prr_delivered
= 0;
2470 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2471 tcp_ecn_queue_cwr(tp
);
2474 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2477 struct tcp_sock
*tp
= tcp_sk(sk
);
2479 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2481 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2484 tp
->prr_delivered
+= newly_acked_sacked
;
2486 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2488 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2489 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2490 !(flag
& FLAG_LOST_RETRANS
)) {
2491 sndcnt
= min_t(int, delta
,
2492 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2493 newly_acked_sacked
) + 1);
2495 sndcnt
= min(delta
, newly_acked_sacked
);
2497 /* Force a fast retransmit upon entering fast recovery */
2498 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2499 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2502 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2504 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2507 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2508 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2509 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2510 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2512 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2515 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2516 void tcp_enter_cwr(struct sock
*sk
)
2518 struct tcp_sock
*tp
= tcp_sk(sk
);
2520 tp
->prior_ssthresh
= 0;
2521 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2522 tp
->undo_marker
= 0;
2523 tcp_init_cwnd_reduction(sk
);
2524 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2527 EXPORT_SYMBOL(tcp_enter_cwr
);
2529 static void tcp_try_keep_open(struct sock
*sk
)
2531 struct tcp_sock
*tp
= tcp_sk(sk
);
2532 int state
= TCP_CA_Open
;
2534 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2535 state
= TCP_CA_Disorder
;
2537 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2538 tcp_set_ca_state(sk
, state
);
2539 tp
->high_seq
= tp
->snd_nxt
;
2543 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2545 struct tcp_sock
*tp
= tcp_sk(sk
);
2547 tcp_verify_left_out(tp
);
2549 if (!tcp_any_retrans_done(sk
))
2550 tp
->retrans_stamp
= 0;
2552 if (flag
& FLAG_ECE
)
2555 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2556 tcp_try_keep_open(sk
);
2560 static void tcp_mtup_probe_failed(struct sock
*sk
)
2562 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2564 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2565 icsk
->icsk_mtup
.probe_size
= 0;
2566 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2569 static void tcp_mtup_probe_success(struct sock
*sk
)
2571 struct tcp_sock
*tp
= tcp_sk(sk
);
2572 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2574 /* FIXME: breaks with very large cwnd */
2575 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2576 tp
->snd_cwnd
= tp
->snd_cwnd
*
2577 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2578 icsk
->icsk_mtup
.probe_size
;
2579 tp
->snd_cwnd_cnt
= 0;
2580 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2581 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2583 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2584 icsk
->icsk_mtup
.probe_size
= 0;
2585 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2586 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2589 /* Do a simple retransmit without using the backoff mechanisms in
2590 * tcp_timer. This is used for path mtu discovery.
2591 * The socket is already locked here.
2593 void tcp_simple_retransmit(struct sock
*sk
)
2595 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 struct tcp_sock
*tp
= tcp_sk(sk
);
2597 struct sk_buff
*skb
;
2598 unsigned int mss
= tcp_current_mss(sk
);
2599 u32 prior_lost
= tp
->lost_out
;
2601 tcp_for_write_queue(skb
, sk
) {
2602 if (skb
== tcp_send_head(sk
))
2604 if (tcp_skb_seglen(skb
) > mss
&&
2605 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2606 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2607 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2608 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2610 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2614 tcp_clear_retrans_hints_partial(tp
);
2616 if (prior_lost
== tp
->lost_out
)
2619 if (tcp_is_reno(tp
))
2620 tcp_limit_reno_sacked(tp
);
2622 tcp_verify_left_out(tp
);
2624 /* Don't muck with the congestion window here.
2625 * Reason is that we do not increase amount of _data_
2626 * in network, but units changed and effective
2627 * cwnd/ssthresh really reduced now.
2629 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2630 tp
->high_seq
= tp
->snd_nxt
;
2631 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2632 tp
->prior_ssthresh
= 0;
2633 tp
->undo_marker
= 0;
2634 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2636 tcp_xmit_retransmit_queue(sk
);
2638 EXPORT_SYMBOL(tcp_simple_retransmit
);
2640 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2642 struct tcp_sock
*tp
= tcp_sk(sk
);
2645 if (tcp_is_reno(tp
))
2646 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2648 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2650 NET_INC_STATS(sock_net(sk
), mib_idx
);
2652 tp
->prior_ssthresh
= 0;
2655 if (!tcp_in_cwnd_reduction(sk
)) {
2657 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2658 tcp_init_cwnd_reduction(sk
);
2660 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2663 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2664 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2666 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2669 struct tcp_sock
*tp
= tcp_sk(sk
);
2670 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2672 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2673 tcp_try_undo_loss(sk
, false))
2676 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2677 /* Step 3.b. A timeout is spurious if not all data are
2678 * lost, i.e., never-retransmitted data are (s)acked.
2680 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2681 tcp_try_undo_loss(sk
, true))
2684 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2685 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2686 tp
->frto
= 0; /* Step 3.a. loss was real */
2687 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2688 tp
->high_seq
= tp
->snd_nxt
;
2689 /* Step 2.b. Try send new data (but deferred until cwnd
2690 * is updated in tcp_ack()). Otherwise fall back to
2691 * the conventional recovery.
2693 if (tcp_send_head(sk
) &&
2694 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2695 *rexmit
= REXMIT_NEW
;
2703 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2704 tcp_try_undo_recovery(sk
);
2707 if (tcp_is_reno(tp
)) {
2708 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2709 * delivered. Lower inflight to clock out (re)tranmissions.
2711 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2712 tcp_add_reno_sack(sk
);
2713 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2714 tcp_reset_reno_sack(tp
);
2716 *rexmit
= REXMIT_LOST
;
2719 /* Undo during fast recovery after partial ACK. */
2720 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2722 struct tcp_sock
*tp
= tcp_sk(sk
);
2724 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2725 /* Plain luck! Hole if filled with delayed
2726 * packet, rather than with a retransmit.
2728 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2730 /* We are getting evidence that the reordering degree is higher
2731 * than we realized. If there are no retransmits out then we
2732 * can undo. Otherwise we clock out new packets but do not
2733 * mark more packets lost or retransmit more.
2735 if (tp
->retrans_out
)
2738 if (!tcp_any_retrans_done(sk
))
2739 tp
->retrans_stamp
= 0;
2741 DBGUNDO(sk
, "partial recovery");
2742 tcp_undo_cwnd_reduction(sk
, true);
2743 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2744 tcp_try_keep_open(sk
);
2750 /* Process an event, which can update packets-in-flight not trivially.
2751 * Main goal of this function is to calculate new estimate for left_out,
2752 * taking into account both packets sitting in receiver's buffer and
2753 * packets lost by network.
2755 * Besides that it updates the congestion state when packet loss or ECN
2756 * is detected. But it does not reduce the cwnd, it is done by the
2757 * congestion control later.
2759 * It does _not_ decide what to send, it is made in function
2760 * tcp_xmit_retransmit_queue().
2762 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2763 bool is_dupack
, int *ack_flag
, int *rexmit
)
2765 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2766 struct tcp_sock
*tp
= tcp_sk(sk
);
2767 int fast_rexmit
= 0, flag
= *ack_flag
;
2768 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2769 (tcp_fackets_out(tp
) > tp
->reordering
));
2771 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2773 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2774 tp
->fackets_out
= 0;
2776 /* Now state machine starts.
2777 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2778 if (flag
& FLAG_ECE
)
2779 tp
->prior_ssthresh
= 0;
2781 /* B. In all the states check for reneging SACKs. */
2782 if (tcp_check_sack_reneging(sk
, flag
))
2785 /* C. Check consistency of the current state. */
2786 tcp_verify_left_out(tp
);
2788 /* D. Check state exit conditions. State can be terminated
2789 * when high_seq is ACKed. */
2790 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2791 WARN_ON(tp
->retrans_out
!= 0);
2792 tp
->retrans_stamp
= 0;
2793 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2794 switch (icsk
->icsk_ca_state
) {
2796 /* CWR is to be held something *above* high_seq
2797 * is ACKed for CWR bit to reach receiver. */
2798 if (tp
->snd_una
!= tp
->high_seq
) {
2799 tcp_end_cwnd_reduction(sk
);
2800 tcp_set_ca_state(sk
, TCP_CA_Open
);
2804 case TCP_CA_Recovery
:
2805 if (tcp_is_reno(tp
))
2806 tcp_reset_reno_sack(tp
);
2807 if (tcp_try_undo_recovery(sk
))
2809 tcp_end_cwnd_reduction(sk
);
2814 /* Use RACK to detect loss */
2815 if (sysctl_tcp_recovery
& TCP_RACK_LOST_RETRANS
&&
2816 tcp_rack_mark_lost(sk
)) {
2817 flag
|= FLAG_LOST_RETRANS
;
2818 *ack_flag
|= FLAG_LOST_RETRANS
;
2821 /* E. Process state. */
2822 switch (icsk
->icsk_ca_state
) {
2823 case TCP_CA_Recovery
:
2824 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2825 if (tcp_is_reno(tp
) && is_dupack
)
2826 tcp_add_reno_sack(sk
);
2828 if (tcp_try_undo_partial(sk
, acked
))
2830 /* Partial ACK arrived. Force fast retransmit. */
2831 do_lost
= tcp_is_reno(tp
) ||
2832 tcp_fackets_out(tp
) > tp
->reordering
;
2834 if (tcp_try_undo_dsack(sk
)) {
2835 tcp_try_keep_open(sk
);
2840 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2841 if (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2842 !(flag
& FLAG_LOST_RETRANS
))
2844 /* Change state if cwnd is undone or retransmits are lost */
2846 if (tcp_is_reno(tp
)) {
2847 if (flag
& FLAG_SND_UNA_ADVANCED
)
2848 tcp_reset_reno_sack(tp
);
2850 tcp_add_reno_sack(sk
);
2853 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2854 tcp_try_undo_dsack(sk
);
2856 if (!tcp_time_to_recover(sk
, flag
)) {
2857 tcp_try_to_open(sk
, flag
);
2861 /* MTU probe failure: don't reduce cwnd */
2862 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2863 icsk
->icsk_mtup
.probe_size
&&
2864 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2865 tcp_mtup_probe_failed(sk
);
2866 /* Restores the reduction we did in tcp_mtup_probe() */
2868 tcp_simple_retransmit(sk
);
2872 /* Otherwise enter Recovery state */
2873 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2878 tcp_update_scoreboard(sk
, fast_rexmit
);
2879 *rexmit
= REXMIT_LOST
;
2882 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2884 struct tcp_sock
*tp
= tcp_sk(sk
);
2885 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2887 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_time_stamp
,
2888 rtt_us
? : jiffies_to_usecs(1));
2891 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2892 long seq_rtt_us
, long sack_rtt_us
,
2895 const struct tcp_sock
*tp
= tcp_sk(sk
);
2897 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2898 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2899 * Karn's algorithm forbids taking RTT if some retransmitted data
2900 * is acked (RFC6298).
2903 seq_rtt_us
= sack_rtt_us
;
2905 /* RTTM Rule: A TSecr value received in a segment is used to
2906 * update the averaged RTT measurement only if the segment
2907 * acknowledges some new data, i.e., only if it advances the
2908 * left edge of the send window.
2909 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2911 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2913 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2914 tp
->rx_opt
.rcv_tsecr
);
2918 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2919 * always taken together with ACK, SACK, or TS-opts. Any negative
2920 * values will be skipped with the seq_rtt_us < 0 check above.
2922 tcp_update_rtt_min(sk
, ca_rtt_us
);
2923 tcp_rtt_estimator(sk
, seq_rtt_us
);
2926 /* RFC6298: only reset backoff on valid RTT measurement. */
2927 inet_csk(sk
)->icsk_backoff
= 0;
2931 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2932 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2936 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2937 struct skb_mstamp now
;
2939 skb_mstamp_get(&now
);
2940 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
2943 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
2947 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2949 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2951 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2952 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2955 /* Restart timer after forward progress on connection.
2956 * RFC2988 recommends to restart timer to now+rto.
2958 void tcp_rearm_rto(struct sock
*sk
)
2960 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2961 struct tcp_sock
*tp
= tcp_sk(sk
);
2963 /* If the retrans timer is currently being used by Fast Open
2964 * for SYN-ACK retrans purpose, stay put.
2966 if (tp
->fastopen_rsk
)
2969 if (!tp
->packets_out
) {
2970 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2972 u32 rto
= inet_csk(sk
)->icsk_rto
;
2973 /* Offset the time elapsed after installing regular RTO */
2974 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2975 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2976 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2977 const u32 rto_time_stamp
=
2978 tcp_skb_timestamp(skb
) + rto
;
2979 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2980 /* delta may not be positive if the socket is locked
2981 * when the retrans timer fires and is rescheduled.
2986 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2991 /* This function is called when the delayed ER timer fires. TCP enters
2992 * fast recovery and performs fast-retransmit.
2994 void tcp_resume_early_retransmit(struct sock
*sk
)
2996 struct tcp_sock
*tp
= tcp_sk(sk
);
3000 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3001 if (!tp
->do_early_retrans
)
3004 tcp_enter_recovery(sk
, false);
3005 tcp_update_scoreboard(sk
, 1);
3006 tcp_xmit_retransmit_queue(sk
);
3009 /* If we get here, the whole TSO packet has not been acked. */
3010 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3012 struct tcp_sock
*tp
= tcp_sk(sk
);
3015 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3017 packets_acked
= tcp_skb_pcount(skb
);
3018 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3020 packets_acked
-= tcp_skb_pcount(skb
);
3022 if (packets_acked
) {
3023 BUG_ON(tcp_skb_pcount(skb
) == 0);
3024 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3027 return packets_acked
;
3030 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3033 const struct skb_shared_info
*shinfo
;
3035 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3036 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3039 shinfo
= skb_shinfo(skb
);
3040 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3041 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3042 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3045 /* Remove acknowledged frames from the retransmission queue. If our packet
3046 * is before the ack sequence we can discard it as it's confirmed to have
3047 * arrived at the other end.
3049 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3050 u32 prior_snd_una
, int *acked
,
3051 struct tcp_sacktag_state
*sack
)
3053 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3054 struct skb_mstamp first_ackt
, last_ackt
, now
;
3055 struct tcp_sock
*tp
= tcp_sk(sk
);
3056 u32 prior_sacked
= tp
->sacked_out
;
3057 u32 reord
= tp
->packets_out
;
3058 bool fully_acked
= true;
3059 long sack_rtt_us
= -1L;
3060 long seq_rtt_us
= -1L;
3061 long ca_rtt_us
= -1L;
3062 struct sk_buff
*skb
;
3064 u32 last_in_flight
= 0;
3070 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3071 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3072 u8 sacked
= scb
->sacked
;
3075 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3077 /* Determine how many packets and what bytes were acked, tso and else */
3078 if (after(scb
->end_seq
, tp
->snd_una
)) {
3079 if (tcp_skb_pcount(skb
) == 1 ||
3080 !after(tp
->snd_una
, scb
->seq
))
3083 acked_pcount
= tcp_tso_acked(sk
, skb
);
3087 fully_acked
= false;
3089 /* Speedup tcp_unlink_write_queue() and next loop */
3090 prefetchw(skb
->next
);
3091 acked_pcount
= tcp_skb_pcount(skb
);
3094 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3095 if (sacked
& TCPCB_SACKED_RETRANS
)
3096 tp
->retrans_out
-= acked_pcount
;
3097 flag
|= FLAG_RETRANS_DATA_ACKED
;
3098 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3099 last_ackt
= skb
->skb_mstamp
;
3100 WARN_ON_ONCE(last_ackt
.v64
== 0);
3101 if (!first_ackt
.v64
)
3102 first_ackt
= last_ackt
;
3104 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3105 reord
= min(pkts_acked
, reord
);
3106 if (!after(scb
->end_seq
, tp
->high_seq
))
3107 flag
|= FLAG_ORIG_SACK_ACKED
;
3110 if (sacked
& TCPCB_SACKED_ACKED
) {
3111 tp
->sacked_out
-= acked_pcount
;
3112 } else if (tcp_is_sack(tp
)) {
3113 tp
->delivered
+= acked_pcount
;
3114 if (!tcp_skb_spurious_retrans(tp
, skb
))
3115 tcp_rack_advance(tp
, &skb
->skb_mstamp
, sacked
);
3117 if (sacked
& TCPCB_LOST
)
3118 tp
->lost_out
-= acked_pcount
;
3120 tp
->packets_out
-= acked_pcount
;
3121 pkts_acked
+= acked_pcount
;
3123 /* Initial outgoing SYN's get put onto the write_queue
3124 * just like anything else we transmit. It is not
3125 * true data, and if we misinform our callers that
3126 * this ACK acks real data, we will erroneously exit
3127 * connection startup slow start one packet too
3128 * quickly. This is severely frowned upon behavior.
3130 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3131 flag
|= FLAG_DATA_ACKED
;
3133 flag
|= FLAG_SYN_ACKED
;
3134 tp
->retrans_stamp
= 0;
3140 tcp_unlink_write_queue(skb
, sk
);
3141 sk_wmem_free_skb(sk
, skb
);
3142 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3143 tp
->retransmit_skb_hint
= NULL
;
3144 if (unlikely(skb
== tp
->lost_skb_hint
))
3145 tp
->lost_skb_hint
= NULL
;
3148 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3149 tp
->snd_up
= tp
->snd_una
;
3151 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3152 flag
|= FLAG_SACK_RENEGING
;
3154 skb_mstamp_get(&now
);
3155 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3156 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3157 ca_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3159 if (sack
->first_sackt
.v64
) {
3160 sack_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->first_sackt
);
3161 ca_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->last_sackt
);
3164 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3167 if (flag
& FLAG_ACKED
) {
3169 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3170 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3171 tcp_mtup_probe_success(sk
);
3174 if (tcp_is_reno(tp
)) {
3175 tcp_remove_reno_sacks(sk
, pkts_acked
);
3179 /* Non-retransmitted hole got filled? That's reordering */
3180 if (reord
< prior_fackets
)
3181 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3183 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3184 prior_sacked
- tp
->sacked_out
;
3185 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3188 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3190 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3191 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3192 /* Do not re-arm RTO if the sack RTT is measured from data sent
3193 * after when the head was last (re)transmitted. Otherwise the
3194 * timeout may continue to extend in loss recovery.
3199 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3200 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3201 .rtt_us
= ca_rtt_us
,
3202 .in_flight
= last_in_flight
};
3204 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3207 #if FASTRETRANS_DEBUG > 0
3208 WARN_ON((int)tp
->sacked_out
< 0);
3209 WARN_ON((int)tp
->lost_out
< 0);
3210 WARN_ON((int)tp
->retrans_out
< 0);
3211 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3212 icsk
= inet_csk(sk
);
3214 pr_debug("Leak l=%u %d\n",
3215 tp
->lost_out
, icsk
->icsk_ca_state
);
3218 if (tp
->sacked_out
) {
3219 pr_debug("Leak s=%u %d\n",
3220 tp
->sacked_out
, icsk
->icsk_ca_state
);
3223 if (tp
->retrans_out
) {
3224 pr_debug("Leak r=%u %d\n",
3225 tp
->retrans_out
, icsk
->icsk_ca_state
);
3226 tp
->retrans_out
= 0;
3230 *acked
= pkts_acked
;
3234 static void tcp_ack_probe(struct sock
*sk
)
3236 const struct tcp_sock
*tp
= tcp_sk(sk
);
3237 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3239 /* Was it a usable window open? */
3241 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3242 icsk
->icsk_backoff
= 0;
3243 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3244 /* Socket must be waked up by subsequent tcp_data_snd_check().
3245 * This function is not for random using!
3248 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3250 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3255 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3257 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3258 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3261 /* Decide wheather to run the increase function of congestion control. */
3262 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3264 /* If reordering is high then always grow cwnd whenever data is
3265 * delivered regardless of its ordering. Otherwise stay conservative
3266 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3267 * new SACK or ECE mark may first advance cwnd here and later reduce
3268 * cwnd in tcp_fastretrans_alert() based on more states.
3270 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3271 return flag
& FLAG_FORWARD_PROGRESS
;
3273 return flag
& FLAG_DATA_ACKED
;
3276 /* The "ultimate" congestion control function that aims to replace the rigid
3277 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3278 * It's called toward the end of processing an ACK with precise rate
3279 * information. All transmission or retransmission are delayed afterwards.
3281 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3284 if (tcp_in_cwnd_reduction(sk
)) {
3285 /* Reduce cwnd if state mandates */
3286 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3287 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3288 /* Advance cwnd if state allows */
3289 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3291 tcp_update_pacing_rate(sk
);
3294 /* Check that window update is acceptable.
3295 * The function assumes that snd_una<=ack<=snd_next.
3297 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3298 const u32 ack
, const u32 ack_seq
,
3301 return after(ack
, tp
->snd_una
) ||
3302 after(ack_seq
, tp
->snd_wl1
) ||
3303 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3306 /* If we update tp->snd_una, also update tp->bytes_acked */
3307 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3309 u32 delta
= ack
- tp
->snd_una
;
3311 sock_owned_by_me((struct sock
*)tp
);
3312 u64_stats_update_begin_raw(&tp
->syncp
);
3313 tp
->bytes_acked
+= delta
;
3314 u64_stats_update_end_raw(&tp
->syncp
);
3318 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3319 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3321 u32 delta
= seq
- tp
->rcv_nxt
;
3323 sock_owned_by_me((struct sock
*)tp
);
3324 u64_stats_update_begin_raw(&tp
->syncp
);
3325 tp
->bytes_received
+= delta
;
3326 u64_stats_update_end_raw(&tp
->syncp
);
3330 /* Update our send window.
3332 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3333 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3335 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3338 struct tcp_sock
*tp
= tcp_sk(sk
);
3340 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3342 if (likely(!tcp_hdr(skb
)->syn
))
3343 nwin
<<= tp
->rx_opt
.snd_wscale
;
3345 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3346 flag
|= FLAG_WIN_UPDATE
;
3347 tcp_update_wl(tp
, ack_seq
);
3349 if (tp
->snd_wnd
!= nwin
) {
3352 /* Note, it is the only place, where
3353 * fast path is recovered for sending TCP.
3356 tcp_fast_path_check(sk
);
3358 if (tcp_send_head(sk
))
3359 tcp_slow_start_after_idle_check(sk
);
3361 if (nwin
> tp
->max_window
) {
3362 tp
->max_window
= nwin
;
3363 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3368 tcp_snd_una_update(tp
, ack
);
3373 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3374 u32
*last_oow_ack_time
)
3376 if (*last_oow_ack_time
) {
3377 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3379 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3380 NET_INC_STATS(net
, mib_idx
);
3381 return true; /* rate-limited: don't send yet! */
3385 *last_oow_ack_time
= tcp_time_stamp
;
3387 return false; /* not rate-limited: go ahead, send dupack now! */
3390 /* Return true if we're currently rate-limiting out-of-window ACKs and
3391 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3392 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3393 * attacks that send repeated SYNs or ACKs for the same connection. To
3394 * do this, we do not send a duplicate SYNACK or ACK if the remote
3395 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3397 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3398 int mib_idx
, u32
*last_oow_ack_time
)
3400 /* Data packets without SYNs are not likely part of an ACK loop. */
3401 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3405 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3408 /* RFC 5961 7 [ACK Throttling] */
3409 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3411 /* unprotected vars, we dont care of overwrites */
3412 static u32 challenge_timestamp
;
3413 static unsigned int challenge_count
;
3414 struct tcp_sock
*tp
= tcp_sk(sk
);
3417 /* First check our per-socket dupack rate limit. */
3418 if (__tcp_oow_rate_limited(sock_net(sk
),
3419 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3420 &tp
->last_oow_ack_time
))
3423 /* Then check host-wide RFC 5961 rate limit. */
3425 if (now
!= challenge_timestamp
) {
3426 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3428 challenge_timestamp
= now
;
3429 WRITE_ONCE(challenge_count
, half
+
3430 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3432 count
= READ_ONCE(challenge_count
);
3434 WRITE_ONCE(challenge_count
, count
- 1);
3435 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3440 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3442 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3443 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3446 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3448 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3449 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3450 * extra check below makes sure this can only happen
3451 * for pure ACK frames. -DaveM
3453 * Not only, also it occurs for expired timestamps.
3456 if (tcp_paws_check(&tp
->rx_opt
, 0))
3457 tcp_store_ts_recent(tp
);
3461 /* This routine deals with acks during a TLP episode.
3462 * We mark the end of a TLP episode on receiving TLP dupack or when
3463 * ack is after tlp_high_seq.
3464 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3466 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3468 struct tcp_sock
*tp
= tcp_sk(sk
);
3470 if (before(ack
, tp
->tlp_high_seq
))
3473 if (flag
& FLAG_DSACKING_ACK
) {
3474 /* This DSACK means original and TLP probe arrived; no loss */
3475 tp
->tlp_high_seq
= 0;
3476 } else if (after(ack
, tp
->tlp_high_seq
)) {
3477 /* ACK advances: there was a loss, so reduce cwnd. Reset
3478 * tlp_high_seq in tcp_init_cwnd_reduction()
3480 tcp_init_cwnd_reduction(sk
);
3481 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3482 tcp_end_cwnd_reduction(sk
);
3483 tcp_try_keep_open(sk
);
3484 NET_INC_STATS(sock_net(sk
),
3485 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3486 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3487 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3488 /* Pure dupack: original and TLP probe arrived; no loss */
3489 tp
->tlp_high_seq
= 0;
3493 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3495 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3497 if (icsk
->icsk_ca_ops
->in_ack_event
)
3498 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3501 /* Congestion control has updated the cwnd already. So if we're in
3502 * loss recovery then now we do any new sends (for FRTO) or
3503 * retransmits (for CA_Loss or CA_recovery) that make sense.
3505 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3507 struct tcp_sock
*tp
= tcp_sk(sk
);
3509 if (rexmit
== REXMIT_NONE
)
3512 if (unlikely(rexmit
== 2)) {
3513 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3515 if (after(tp
->snd_nxt
, tp
->high_seq
))
3519 tcp_xmit_retransmit_queue(sk
);
3522 /* This routine deals with incoming acks, but not outgoing ones. */
3523 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3525 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3526 struct tcp_sock
*tp
= tcp_sk(sk
);
3527 struct tcp_sacktag_state sack_state
;
3528 u32 prior_snd_una
= tp
->snd_una
;
3529 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3530 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3531 bool is_dupack
= false;
3533 int prior_packets
= tp
->packets_out
;
3534 u32 prior_delivered
= tp
->delivered
;
3535 int acked
= 0; /* Number of packets newly acked */
3536 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3538 sack_state
.first_sackt
.v64
= 0;
3540 /* We very likely will need to access write queue head. */
3541 prefetchw(sk
->sk_write_queue
.next
);
3543 /* If the ack is older than previous acks
3544 * then we can probably ignore it.
3546 if (before(ack
, prior_snd_una
)) {
3547 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3548 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3549 tcp_send_challenge_ack(sk
, skb
);
3555 /* If the ack includes data we haven't sent yet, discard
3556 * this segment (RFC793 Section 3.9).
3558 if (after(ack
, tp
->snd_nxt
))
3561 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3562 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3565 if (after(ack
, prior_snd_una
)) {
3566 flag
|= FLAG_SND_UNA_ADVANCED
;
3567 icsk
->icsk_retransmits
= 0;
3570 prior_fackets
= tp
->fackets_out
;
3572 /* ts_recent update must be made after we are sure that the packet
3575 if (flag
& FLAG_UPDATE_TS_RECENT
)
3576 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3578 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3579 /* Window is constant, pure forward advance.
3580 * No more checks are required.
3581 * Note, we use the fact that SND.UNA>=SND.WL2.
3583 tcp_update_wl(tp
, ack_seq
);
3584 tcp_snd_una_update(tp
, ack
);
3585 flag
|= FLAG_WIN_UPDATE
;
3587 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3589 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3591 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3593 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3596 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3598 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3600 if (TCP_SKB_CB(skb
)->sacked
)
3601 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3604 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3606 ack_ev_flags
|= CA_ACK_ECE
;
3609 if (flag
& FLAG_WIN_UPDATE
)
3610 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3612 tcp_in_ack_event(sk
, ack_ev_flags
);
3615 /* We passed data and got it acked, remove any soft error
3616 * log. Something worked...
3618 sk
->sk_err_soft
= 0;
3619 icsk
->icsk_probes_out
= 0;
3620 tp
->rcv_tstamp
= tcp_time_stamp
;
3624 /* See if we can take anything off of the retransmit queue. */
3625 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3628 if (tcp_ack_is_dubious(sk
, flag
)) {
3629 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3630 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3632 if (tp
->tlp_high_seq
)
3633 tcp_process_tlp_ack(sk
, ack
, flag
);
3635 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3636 struct dst_entry
*dst
= __sk_dst_get(sk
);
3641 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3642 tcp_schedule_loss_probe(sk
);
3643 tcp_cong_control(sk
, ack
, tp
->delivered
- prior_delivered
, flag
);
3644 tcp_xmit_recovery(sk
, rexmit
);
3648 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3649 if (flag
& FLAG_DSACKING_ACK
)
3650 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3651 /* If this ack opens up a zero window, clear backoff. It was
3652 * being used to time the probes, and is probably far higher than
3653 * it needs to be for normal retransmission.
3655 if (tcp_send_head(sk
))
3658 if (tp
->tlp_high_seq
)
3659 tcp_process_tlp_ack(sk
, ack
, flag
);
3663 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3667 /* If data was SACKed, tag it and see if we should send more data.
3668 * If data was DSACKed, see if we can undo a cwnd reduction.
3670 if (TCP_SKB_CB(skb
)->sacked
) {
3671 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3673 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3674 tcp_xmit_recovery(sk
, rexmit
);
3677 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3681 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3682 bool syn
, struct tcp_fastopen_cookie
*foc
,
3685 /* Valid only in SYN or SYN-ACK with an even length. */
3686 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3689 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3690 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3691 memcpy(foc
->val
, cookie
, len
);
3698 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3699 * But, this can also be called on packets in the established flow when
3700 * the fast version below fails.
3702 void tcp_parse_options(const struct sk_buff
*skb
,
3703 struct tcp_options_received
*opt_rx
, int estab
,
3704 struct tcp_fastopen_cookie
*foc
)
3706 const unsigned char *ptr
;
3707 const struct tcphdr
*th
= tcp_hdr(skb
);
3708 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3710 ptr
= (const unsigned char *)(th
+ 1);
3711 opt_rx
->saw_tstamp
= 0;
3713 while (length
> 0) {
3714 int opcode
= *ptr
++;
3720 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3725 if (opsize
< 2) /* "silly options" */
3727 if (opsize
> length
)
3728 return; /* don't parse partial options */
3731 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3732 u16 in_mss
= get_unaligned_be16(ptr
);
3734 if (opt_rx
->user_mss
&&
3735 opt_rx
->user_mss
< in_mss
)
3736 in_mss
= opt_rx
->user_mss
;
3737 opt_rx
->mss_clamp
= in_mss
;
3742 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3743 !estab
&& sysctl_tcp_window_scaling
) {
3744 __u8 snd_wscale
= *(__u8
*)ptr
;
3745 opt_rx
->wscale_ok
= 1;
3746 if (snd_wscale
> 14) {
3747 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3752 opt_rx
->snd_wscale
= snd_wscale
;
3755 case TCPOPT_TIMESTAMP
:
3756 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3757 ((estab
&& opt_rx
->tstamp_ok
) ||
3758 (!estab
&& sysctl_tcp_timestamps
))) {
3759 opt_rx
->saw_tstamp
= 1;
3760 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3761 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3764 case TCPOPT_SACK_PERM
:
3765 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3766 !estab
&& sysctl_tcp_sack
) {
3767 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3768 tcp_sack_reset(opt_rx
);
3773 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3774 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3776 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3779 #ifdef CONFIG_TCP_MD5SIG
3782 * The MD5 Hash has already been
3783 * checked (see tcp_v{4,6}_do_rcv()).
3787 case TCPOPT_FASTOPEN
:
3788 tcp_parse_fastopen_option(
3789 opsize
- TCPOLEN_FASTOPEN_BASE
,
3790 ptr
, th
->syn
, foc
, false);
3794 /* Fast Open option shares code 254 using a
3795 * 16 bits magic number.
3797 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3798 get_unaligned_be16(ptr
) ==
3799 TCPOPT_FASTOPEN_MAGIC
)
3800 tcp_parse_fastopen_option(opsize
-
3801 TCPOLEN_EXP_FASTOPEN_BASE
,
3802 ptr
+ 2, th
->syn
, foc
, true);
3811 EXPORT_SYMBOL(tcp_parse_options
);
3813 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3815 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3817 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3818 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3819 tp
->rx_opt
.saw_tstamp
= 1;
3821 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3824 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3826 tp
->rx_opt
.rcv_tsecr
= 0;
3832 /* Fast parse options. This hopes to only see timestamps.
3833 * If it is wrong it falls back on tcp_parse_options().
3835 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3836 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3838 /* In the spirit of fast parsing, compare doff directly to constant
3839 * values. Because equality is used, short doff can be ignored here.
3841 if (th
->doff
== (sizeof(*th
) / 4)) {
3842 tp
->rx_opt
.saw_tstamp
= 0;
3844 } else if (tp
->rx_opt
.tstamp_ok
&&
3845 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3846 if (tcp_parse_aligned_timestamp(tp
, th
))
3850 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3851 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3852 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3857 #ifdef CONFIG_TCP_MD5SIG
3859 * Parse MD5 Signature option
3861 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3863 int length
= (th
->doff
<< 2) - sizeof(*th
);
3864 const u8
*ptr
= (const u8
*)(th
+ 1);
3866 /* If the TCP option is too short, we can short cut */
3867 if (length
< TCPOLEN_MD5SIG
)
3870 while (length
> 0) {
3871 int opcode
= *ptr
++;
3882 if (opsize
< 2 || opsize
> length
)
3884 if (opcode
== TCPOPT_MD5SIG
)
3885 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3892 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3895 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3897 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3898 * it can pass through stack. So, the following predicate verifies that
3899 * this segment is not used for anything but congestion avoidance or
3900 * fast retransmit. Moreover, we even are able to eliminate most of such
3901 * second order effects, if we apply some small "replay" window (~RTO)
3902 * to timestamp space.
3904 * All these measures still do not guarantee that we reject wrapped ACKs
3905 * on networks with high bandwidth, when sequence space is recycled fastly,
3906 * but it guarantees that such events will be very rare and do not affect
3907 * connection seriously. This doesn't look nice, but alas, PAWS is really
3910 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3911 * states that events when retransmit arrives after original data are rare.
3912 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3913 * the biggest problem on large power networks even with minor reordering.
3914 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3915 * up to bandwidth of 18Gigabit/sec. 8) ]
3918 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3920 const struct tcp_sock
*tp
= tcp_sk(sk
);
3921 const struct tcphdr
*th
= tcp_hdr(skb
);
3922 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3923 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3925 return (/* 1. Pure ACK with correct sequence number. */
3926 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3928 /* 2. ... and duplicate ACK. */
3929 ack
== tp
->snd_una
&&
3931 /* 3. ... and does not update window. */
3932 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3934 /* 4. ... and sits in replay window. */
3935 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3938 static inline bool tcp_paws_discard(const struct sock
*sk
,
3939 const struct sk_buff
*skb
)
3941 const struct tcp_sock
*tp
= tcp_sk(sk
);
3943 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3944 !tcp_disordered_ack(sk
, skb
);
3947 /* Check segment sequence number for validity.
3949 * Segment controls are considered valid, if the segment
3950 * fits to the window after truncation to the window. Acceptability
3951 * of data (and SYN, FIN, of course) is checked separately.
3952 * See tcp_data_queue(), for example.
3954 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3955 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3956 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3957 * (borrowed from freebsd)
3960 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3962 return !before(end_seq
, tp
->rcv_wup
) &&
3963 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3966 /* When we get a reset we do this. */
3967 void tcp_reset(struct sock
*sk
)
3969 /* We want the right error as BSD sees it (and indeed as we do). */
3970 switch (sk
->sk_state
) {
3972 sk
->sk_err
= ECONNREFUSED
;
3974 case TCP_CLOSE_WAIT
:
3980 sk
->sk_err
= ECONNRESET
;
3982 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3985 if (!sock_flag(sk
, SOCK_DEAD
))
3986 sk
->sk_error_report(sk
);
3992 * Process the FIN bit. This now behaves as it is supposed to work
3993 * and the FIN takes effect when it is validly part of sequence
3994 * space. Not before when we get holes.
3996 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3997 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4000 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4001 * close and we go into CLOSING (and later onto TIME-WAIT)
4003 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4005 void tcp_fin(struct sock
*sk
)
4007 struct tcp_sock
*tp
= tcp_sk(sk
);
4009 inet_csk_schedule_ack(sk
);
4011 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4012 sock_set_flag(sk
, SOCK_DONE
);
4014 switch (sk
->sk_state
) {
4016 case TCP_ESTABLISHED
:
4017 /* Move to CLOSE_WAIT */
4018 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4019 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4022 case TCP_CLOSE_WAIT
:
4024 /* Received a retransmission of the FIN, do
4029 /* RFC793: Remain in the LAST-ACK state. */
4033 /* This case occurs when a simultaneous close
4034 * happens, we must ack the received FIN and
4035 * enter the CLOSING state.
4038 tcp_set_state(sk
, TCP_CLOSING
);
4041 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4043 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4046 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4047 * cases we should never reach this piece of code.
4049 pr_err("%s: Impossible, sk->sk_state=%d\n",
4050 __func__
, sk
->sk_state
);
4054 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4055 * Probably, we should reset in this case. For now drop them.
4057 skb_rbtree_purge(&tp
->out_of_order_queue
);
4058 if (tcp_is_sack(tp
))
4059 tcp_sack_reset(&tp
->rx_opt
);
4062 if (!sock_flag(sk
, SOCK_DEAD
)) {
4063 sk
->sk_state_change(sk
);
4065 /* Do not send POLL_HUP for half duplex close. */
4066 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4067 sk
->sk_state
== TCP_CLOSE
)
4068 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4070 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4074 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4077 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4078 if (before(seq
, sp
->start_seq
))
4079 sp
->start_seq
= seq
;
4080 if (after(end_seq
, sp
->end_seq
))
4081 sp
->end_seq
= end_seq
;
4087 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4089 struct tcp_sock
*tp
= tcp_sk(sk
);
4091 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4094 if (before(seq
, tp
->rcv_nxt
))
4095 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4097 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4099 NET_INC_STATS(sock_net(sk
), mib_idx
);
4101 tp
->rx_opt
.dsack
= 1;
4102 tp
->duplicate_sack
[0].start_seq
= seq
;
4103 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4107 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4109 struct tcp_sock
*tp
= tcp_sk(sk
);
4111 if (!tp
->rx_opt
.dsack
)
4112 tcp_dsack_set(sk
, seq
, end_seq
);
4114 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4117 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4119 struct tcp_sock
*tp
= tcp_sk(sk
);
4121 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4122 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4123 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4124 tcp_enter_quickack_mode(sk
);
4126 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4127 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4129 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4130 end_seq
= tp
->rcv_nxt
;
4131 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4138 /* These routines update the SACK block as out-of-order packets arrive or
4139 * in-order packets close up the sequence space.
4141 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4144 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4145 struct tcp_sack_block
*swalk
= sp
+ 1;
4147 /* See if the recent change to the first SACK eats into
4148 * or hits the sequence space of other SACK blocks, if so coalesce.
4150 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4151 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4154 /* Zap SWALK, by moving every further SACK up by one slot.
4155 * Decrease num_sacks.
4157 tp
->rx_opt
.num_sacks
--;
4158 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4162 this_sack
++, swalk
++;
4166 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4168 struct tcp_sock
*tp
= tcp_sk(sk
);
4169 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4170 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4176 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4177 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4178 /* Rotate this_sack to the first one. */
4179 for (; this_sack
> 0; this_sack
--, sp
--)
4180 swap(*sp
, *(sp
- 1));
4182 tcp_sack_maybe_coalesce(tp
);
4187 /* Could not find an adjacent existing SACK, build a new one,
4188 * put it at the front, and shift everyone else down. We
4189 * always know there is at least one SACK present already here.
4191 * If the sack array is full, forget about the last one.
4193 if (this_sack
>= TCP_NUM_SACKS
) {
4195 tp
->rx_opt
.num_sacks
--;
4198 for (; this_sack
> 0; this_sack
--, sp
--)
4202 /* Build the new head SACK, and we're done. */
4203 sp
->start_seq
= seq
;
4204 sp
->end_seq
= end_seq
;
4205 tp
->rx_opt
.num_sacks
++;
4208 /* RCV.NXT advances, some SACKs should be eaten. */
4210 static void tcp_sack_remove(struct tcp_sock
*tp
)
4212 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4213 int num_sacks
= tp
->rx_opt
.num_sacks
;
4216 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4217 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4218 tp
->rx_opt
.num_sacks
= 0;
4222 for (this_sack
= 0; this_sack
< num_sacks
;) {
4223 /* Check if the start of the sack is covered by RCV.NXT. */
4224 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4227 /* RCV.NXT must cover all the block! */
4228 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4230 /* Zap this SACK, by moving forward any other SACKS. */
4231 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4232 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4239 tp
->rx_opt
.num_sacks
= num_sacks
;
4243 * tcp_try_coalesce - try to merge skb to prior one
4246 * @from: buffer to add in queue
4247 * @fragstolen: pointer to boolean
4249 * Before queueing skb @from after @to, try to merge them
4250 * to reduce overall memory use and queue lengths, if cost is small.
4251 * Packets in ofo or receive queues can stay a long time.
4252 * Better try to coalesce them right now to avoid future collapses.
4253 * Returns true if caller should free @from instead of queueing it
4255 static bool tcp_try_coalesce(struct sock
*sk
,
4257 struct sk_buff
*from
,
4262 *fragstolen
= false;
4264 /* Its possible this segment overlaps with prior segment in queue */
4265 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4268 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4271 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4272 sk_mem_charge(sk
, delta
);
4273 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4274 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4275 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4276 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4280 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4282 sk_drops_add(sk
, skb
);
4286 /* This one checks to see if we can put data from the
4287 * out_of_order queue into the receive_queue.
4289 static void tcp_ofo_queue(struct sock
*sk
)
4291 struct tcp_sock
*tp
= tcp_sk(sk
);
4292 __u32 dsack_high
= tp
->rcv_nxt
;
4293 bool fin
, fragstolen
, eaten
;
4294 struct sk_buff
*skb
, *tail
;
4297 p
= rb_first(&tp
->out_of_order_queue
);
4299 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4300 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4303 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4304 __u32 dsack
= dsack_high
;
4305 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4306 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4307 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4310 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4312 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4313 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4317 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4318 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4319 TCP_SKB_CB(skb
)->end_seq
);
4321 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4322 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4323 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4324 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4326 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4328 kfree_skb_partial(skb
, fragstolen
);
4330 if (unlikely(fin
)) {
4332 /* tcp_fin() purges tp->out_of_order_queue,
4333 * so we must end this loop right now.
4340 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4341 static int tcp_prune_queue(struct sock
*sk
);
4343 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4346 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4347 !sk_rmem_schedule(sk
, skb
, size
)) {
4349 if (tcp_prune_queue(sk
) < 0)
4352 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4353 if (!tcp_prune_ofo_queue(sk
))
4360 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4362 struct tcp_sock
*tp
= tcp_sk(sk
);
4363 struct rb_node
**p
, *q
, *parent
;
4364 struct sk_buff
*skb1
;
4368 tcp_ecn_check_ce(tp
, skb
);
4370 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4371 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4376 /* Disable header prediction. */
4378 inet_csk_schedule_ack(sk
);
4380 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4381 seq
= TCP_SKB_CB(skb
)->seq
;
4382 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4383 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4384 tp
->rcv_nxt
, seq
, end_seq
);
4386 p
= &tp
->out_of_order_queue
.rb_node
;
4387 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4388 /* Initial out of order segment, build 1 SACK. */
4389 if (tcp_is_sack(tp
)) {
4390 tp
->rx_opt
.num_sacks
= 1;
4391 tp
->selective_acks
[0].start_seq
= seq
;
4392 tp
->selective_acks
[0].end_seq
= end_seq
;
4394 rb_link_node(&skb
->rbnode
, NULL
, p
);
4395 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4396 tp
->ooo_last_skb
= skb
;
4400 /* In the typical case, we are adding an skb to the end of the list.
4401 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4403 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
, skb
, &fragstolen
)) {
4405 tcp_grow_window(sk
, skb
);
4406 kfree_skb_partial(skb
, fragstolen
);
4410 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4411 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4412 parent
= &tp
->ooo_last_skb
->rbnode
;
4413 p
= &parent
->rb_right
;
4417 /* Find place to insert this segment. Handle overlaps on the way. */
4421 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4422 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4423 p
= &parent
->rb_left
;
4426 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4427 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4428 /* All the bits are present. Drop. */
4429 NET_INC_STATS(sock_net(sk
),
4430 LINUX_MIB_TCPOFOMERGE
);
4433 tcp_dsack_set(sk
, seq
, end_seq
);
4436 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4437 /* Partial overlap. */
4438 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4440 /* skb's seq == skb1's seq and skb covers skb1.
4441 * Replace skb1 with skb.
4443 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4444 &tp
->out_of_order_queue
);
4445 tcp_dsack_extend(sk
,
4446 TCP_SKB_CB(skb1
)->seq
,
4447 TCP_SKB_CB(skb1
)->end_seq
);
4448 NET_INC_STATS(sock_net(sk
),
4449 LINUX_MIB_TCPOFOMERGE
);
4453 } else if (tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4456 p
= &parent
->rb_right
;
4459 /* Insert segment into RB tree. */
4460 rb_link_node(&skb
->rbnode
, parent
, p
);
4461 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4464 /* Remove other segments covered by skb. */
4465 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4466 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4468 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4470 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4471 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4475 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4476 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4477 TCP_SKB_CB(skb1
)->end_seq
);
4478 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4481 /* If there is no skb after us, we are the last_skb ! */
4483 tp
->ooo_last_skb
= skb
;
4486 if (tcp_is_sack(tp
))
4487 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4490 tcp_grow_window(sk
, skb
);
4491 skb_set_owner_r(skb
, sk
);
4495 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4499 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4501 __skb_pull(skb
, hdrlen
);
4503 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4504 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4506 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4507 skb_set_owner_r(skb
, sk
);
4512 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4514 struct sk_buff
*skb
;
4522 if (size
> PAGE_SIZE
) {
4523 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4525 data_len
= npages
<< PAGE_SHIFT
;
4526 size
= data_len
+ (size
& ~PAGE_MASK
);
4528 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4529 PAGE_ALLOC_COSTLY_ORDER
,
4530 &err
, sk
->sk_allocation
);
4534 skb_put(skb
, size
- data_len
);
4535 skb
->data_len
= data_len
;
4538 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4541 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4545 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4546 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4547 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4549 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4550 WARN_ON_ONCE(fragstolen
); /* should not happen */
4562 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4564 struct tcp_sock
*tp
= tcp_sk(sk
);
4565 bool fragstolen
= false;
4568 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4573 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4575 tcp_ecn_accept_cwr(tp
, skb
);
4577 tp
->rx_opt
.dsack
= 0;
4579 /* Queue data for delivery to the user.
4580 * Packets in sequence go to the receive queue.
4581 * Out of sequence packets to the out_of_order_queue.
4583 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4584 if (tcp_receive_window(tp
) == 0)
4587 /* Ok. In sequence. In window. */
4588 if (tp
->ucopy
.task
== current
&&
4589 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4590 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4591 int chunk
= min_t(unsigned int, skb
->len
,
4594 __set_current_state(TASK_RUNNING
);
4596 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4597 tp
->ucopy
.len
-= chunk
;
4598 tp
->copied_seq
+= chunk
;
4599 eaten
= (chunk
== skb
->len
);
4600 tcp_rcv_space_adjust(sk
);
4607 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4608 sk_forced_mem_schedule(sk
, skb
->truesize
);
4609 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4612 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4614 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4616 tcp_event_data_recv(sk
, skb
);
4617 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4620 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4623 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4624 * gap in queue is filled.
4626 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4627 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4630 if (tp
->rx_opt
.num_sacks
)
4631 tcp_sack_remove(tp
);
4633 tcp_fast_path_check(sk
);
4636 kfree_skb_partial(skb
, fragstolen
);
4637 if (!sock_flag(sk
, SOCK_DEAD
))
4638 sk
->sk_data_ready(sk
);
4642 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4643 /* A retransmit, 2nd most common case. Force an immediate ack. */
4644 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4645 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4648 tcp_enter_quickack_mode(sk
);
4649 inet_csk_schedule_ack(sk
);
4655 /* Out of window. F.e. zero window probe. */
4656 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4659 tcp_enter_quickack_mode(sk
);
4661 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4662 /* Partial packet, seq < rcv_next < end_seq */
4663 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4664 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4665 TCP_SKB_CB(skb
)->end_seq
);
4667 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4669 /* If window is closed, drop tail of packet. But after
4670 * remembering D-SACK for its head made in previous line.
4672 if (!tcp_receive_window(tp
))
4677 tcp_data_queue_ofo(sk
, skb
);
4680 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4683 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4685 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4688 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4689 struct sk_buff_head
*list
,
4690 struct rb_root
*root
)
4692 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4695 __skb_unlink(skb
, list
);
4697 rb_erase(&skb
->rbnode
, root
);
4700 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4705 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4706 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4708 struct rb_node
**p
= &root
->rb_node
;
4709 struct rb_node
*parent
= NULL
;
4710 struct sk_buff
*skb1
;
4714 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4715 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4716 p
= &parent
->rb_left
;
4718 p
= &parent
->rb_right
;
4720 rb_link_node(&skb
->rbnode
, parent
, p
);
4721 rb_insert_color(&skb
->rbnode
, root
);
4724 /* Collapse contiguous sequence of skbs head..tail with
4725 * sequence numbers start..end.
4727 * If tail is NULL, this means until the end of the queue.
4729 * Segments with FIN/SYN are not collapsed (only because this
4733 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4734 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4736 struct sk_buff
*skb
= head
, *n
;
4737 struct sk_buff_head tmp
;
4740 /* First, check that queue is collapsible and find
4741 * the point where collapsing can be useful.
4744 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4745 n
= tcp_skb_next(skb
, list
);
4747 /* No new bits? It is possible on ofo queue. */
4748 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4749 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4755 /* The first skb to collapse is:
4757 * - bloated or contains data before "start" or
4758 * overlaps to the next one.
4760 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4761 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4762 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4763 end_of_skbs
= false;
4767 if (n
&& n
!= tail
&&
4768 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4769 end_of_skbs
= false;
4773 /* Decided to skip this, advance start seq. */
4774 start
= TCP_SKB_CB(skb
)->end_seq
;
4777 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4780 __skb_queue_head_init(&tmp
);
4782 while (before(start
, end
)) {
4783 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4784 struct sk_buff
*nskb
;
4786 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4790 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4791 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4793 __skb_queue_before(list
, skb
, nskb
);
4795 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4796 skb_set_owner_r(nskb
, sk
);
4798 /* Copy data, releasing collapsed skbs. */
4800 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4801 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4805 size
= min(copy
, size
);
4806 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4808 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4812 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4813 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4816 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4822 skb_queue_walk_safe(&tmp
, skb
, n
)
4823 tcp_rbtree_insert(root
, skb
);
4826 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4827 * and tcp_collapse() them until all the queue is collapsed.
4829 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4831 struct tcp_sock
*tp
= tcp_sk(sk
);
4832 struct sk_buff
*skb
, *head
;
4836 p
= rb_first(&tp
->out_of_order_queue
);
4837 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4840 p
= rb_last(&tp
->out_of_order_queue
);
4841 /* Note: This is possible p is NULL here. We do not
4842 * use rb_entry_safe(), as ooo_last_skb is valid only
4843 * if rbtree is not empty.
4845 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4848 start
= TCP_SKB_CB(skb
)->seq
;
4849 end
= TCP_SKB_CB(skb
)->end_seq
;
4851 for (head
= skb
;;) {
4852 skb
= tcp_skb_next(skb
, NULL
);
4854 /* Range is terminated when we see a gap or when
4855 * we are at the queue end.
4858 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4859 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4860 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4861 head
, skb
, start
, end
);
4865 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4866 start
= TCP_SKB_CB(skb
)->seq
;
4867 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4868 end
= TCP_SKB_CB(skb
)->end_seq
;
4873 * Clean the out-of-order queue to make room.
4874 * We drop high sequences packets to :
4875 * 1) Let a chance for holes to be filled.
4876 * 2) not add too big latencies if thousands of packets sit there.
4877 * (But if application shrinks SO_RCVBUF, we could still end up
4878 * freeing whole queue here)
4880 * Return true if queue has shrunk.
4882 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4884 struct tcp_sock
*tp
= tcp_sk(sk
);
4885 struct rb_node
*node
, *prev
;
4887 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4890 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4891 node
= &tp
->ooo_last_skb
->rbnode
;
4893 prev
= rb_prev(node
);
4894 rb_erase(node
, &tp
->out_of_order_queue
);
4895 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4897 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4898 !tcp_under_memory_pressure(sk
))
4902 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4904 /* Reset SACK state. A conforming SACK implementation will
4905 * do the same at a timeout based retransmit. When a connection
4906 * is in a sad state like this, we care only about integrity
4907 * of the connection not performance.
4909 if (tp
->rx_opt
.sack_ok
)
4910 tcp_sack_reset(&tp
->rx_opt
);
4914 /* Reduce allocated memory if we can, trying to get
4915 * the socket within its memory limits again.
4917 * Return less than zero if we should start dropping frames
4918 * until the socket owning process reads some of the data
4919 * to stabilize the situation.
4921 static int tcp_prune_queue(struct sock
*sk
)
4923 struct tcp_sock
*tp
= tcp_sk(sk
);
4925 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4927 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4929 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4930 tcp_clamp_window(sk
);
4931 else if (tcp_under_memory_pressure(sk
))
4932 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4934 tcp_collapse_ofo_queue(sk
);
4935 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4936 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4937 skb_peek(&sk
->sk_receive_queue
),
4939 tp
->copied_seq
, tp
->rcv_nxt
);
4942 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4945 /* Collapsing did not help, destructive actions follow.
4946 * This must not ever occur. */
4948 tcp_prune_ofo_queue(sk
);
4950 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4953 /* If we are really being abused, tell the caller to silently
4954 * drop receive data on the floor. It will get retransmitted
4955 * and hopefully then we'll have sufficient space.
4957 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4959 /* Massive buffer overcommit. */
4964 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4966 const struct tcp_sock
*tp
= tcp_sk(sk
);
4968 /* If the user specified a specific send buffer setting, do
4971 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4974 /* If we are under global TCP memory pressure, do not expand. */
4975 if (tcp_under_memory_pressure(sk
))
4978 /* If we are under soft global TCP memory pressure, do not expand. */
4979 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4982 /* If we filled the congestion window, do not expand. */
4983 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4989 /* When incoming ACK allowed to free some skb from write_queue,
4990 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4991 * on the exit from tcp input handler.
4993 * PROBLEM: sndbuf expansion does not work well with largesend.
4995 static void tcp_new_space(struct sock
*sk
)
4997 struct tcp_sock
*tp
= tcp_sk(sk
);
4999 if (tcp_should_expand_sndbuf(sk
)) {
5000 tcp_sndbuf_expand(sk
);
5001 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
5004 sk
->sk_write_space(sk
);
5007 static void tcp_check_space(struct sock
*sk
)
5009 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5010 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5011 /* pairs with tcp_poll() */
5012 smp_mb__after_atomic();
5013 if (sk
->sk_socket
&&
5014 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5019 static inline void tcp_data_snd_check(struct sock
*sk
)
5021 tcp_push_pending_frames(sk
);
5022 tcp_check_space(sk
);
5026 * Check if sending an ack is needed.
5028 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5030 struct tcp_sock
*tp
= tcp_sk(sk
);
5032 /* More than one full frame received... */
5033 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5034 /* ... and right edge of window advances far enough.
5035 * (tcp_recvmsg() will send ACK otherwise). Or...
5037 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5038 /* We ACK each frame or... */
5039 tcp_in_quickack_mode(sk
) ||
5040 /* We have out of order data. */
5041 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5042 /* Then ack it now */
5045 /* Else, send delayed ack. */
5046 tcp_send_delayed_ack(sk
);
5050 static inline void tcp_ack_snd_check(struct sock
*sk
)
5052 if (!inet_csk_ack_scheduled(sk
)) {
5053 /* We sent a data segment already. */
5056 __tcp_ack_snd_check(sk
, 1);
5060 * This routine is only called when we have urgent data
5061 * signaled. Its the 'slow' part of tcp_urg. It could be
5062 * moved inline now as tcp_urg is only called from one
5063 * place. We handle URGent data wrong. We have to - as
5064 * BSD still doesn't use the correction from RFC961.
5065 * For 1003.1g we should support a new option TCP_STDURG to permit
5066 * either form (or just set the sysctl tcp_stdurg).
5069 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5071 struct tcp_sock
*tp
= tcp_sk(sk
);
5072 u32 ptr
= ntohs(th
->urg_ptr
);
5074 if (ptr
&& !sysctl_tcp_stdurg
)
5076 ptr
+= ntohl(th
->seq
);
5078 /* Ignore urgent data that we've already seen and read. */
5079 if (after(tp
->copied_seq
, ptr
))
5082 /* Do not replay urg ptr.
5084 * NOTE: interesting situation not covered by specs.
5085 * Misbehaving sender may send urg ptr, pointing to segment,
5086 * which we already have in ofo queue. We are not able to fetch
5087 * such data and will stay in TCP_URG_NOTYET until will be eaten
5088 * by recvmsg(). Seems, we are not obliged to handle such wicked
5089 * situations. But it is worth to think about possibility of some
5090 * DoSes using some hypothetical application level deadlock.
5092 if (before(ptr
, tp
->rcv_nxt
))
5095 /* Do we already have a newer (or duplicate) urgent pointer? */
5096 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5099 /* Tell the world about our new urgent pointer. */
5102 /* We may be adding urgent data when the last byte read was
5103 * urgent. To do this requires some care. We cannot just ignore
5104 * tp->copied_seq since we would read the last urgent byte again
5105 * as data, nor can we alter copied_seq until this data arrives
5106 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5108 * NOTE. Double Dutch. Rendering to plain English: author of comment
5109 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5110 * and expect that both A and B disappear from stream. This is _wrong_.
5111 * Though this happens in BSD with high probability, this is occasional.
5112 * Any application relying on this is buggy. Note also, that fix "works"
5113 * only in this artificial test. Insert some normal data between A and B and we will
5114 * decline of BSD again. Verdict: it is better to remove to trap
5117 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5118 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5119 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5121 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5122 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5127 tp
->urg_data
= TCP_URG_NOTYET
;
5130 /* Disable header prediction. */
5134 /* This is the 'fast' part of urgent handling. */
5135 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5137 struct tcp_sock
*tp
= tcp_sk(sk
);
5139 /* Check if we get a new urgent pointer - normally not. */
5141 tcp_check_urg(sk
, th
);
5143 /* Do we wait for any urgent data? - normally not... */
5144 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5145 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5148 /* Is the urgent pointer pointing into this packet? */
5149 if (ptr
< skb
->len
) {
5151 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5153 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5154 if (!sock_flag(sk
, SOCK_DEAD
))
5155 sk
->sk_data_ready(sk
);
5160 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5162 struct tcp_sock
*tp
= tcp_sk(sk
);
5163 int chunk
= skb
->len
- hlen
;
5166 if (skb_csum_unnecessary(skb
))
5167 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5169 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5172 tp
->ucopy
.len
-= chunk
;
5173 tp
->copied_seq
+= chunk
;
5174 tcp_rcv_space_adjust(sk
);
5180 /* Does PAWS and seqno based validation of an incoming segment, flags will
5181 * play significant role here.
5183 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5184 const struct tcphdr
*th
, int syn_inerr
)
5186 struct tcp_sock
*tp
= tcp_sk(sk
);
5187 bool rst_seq_match
= false;
5189 /* RFC1323: H1. Apply PAWS check first. */
5190 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5191 tcp_paws_discard(sk
, skb
)) {
5193 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5194 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5195 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5196 &tp
->last_oow_ack_time
))
5197 tcp_send_dupack(sk
, skb
);
5200 /* Reset is accepted even if it did not pass PAWS. */
5203 /* Step 1: check sequence number */
5204 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5205 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5206 * (RST) segments are validated by checking their SEQ-fields."
5207 * And page 69: "If an incoming segment is not acceptable,
5208 * an acknowledgment should be sent in reply (unless the RST
5209 * bit is set, if so drop the segment and return)".
5214 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5215 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5216 &tp
->last_oow_ack_time
))
5217 tcp_send_dupack(sk
, skb
);
5222 /* Step 2: check RST bit */
5224 /* RFC 5961 3.2 (extend to match against SACK too if available):
5225 * If seq num matches RCV.NXT or the right-most SACK block,
5227 * RESET the connection
5229 * Send a challenge ACK
5231 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5232 rst_seq_match
= true;
5233 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5234 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5235 int max_sack
= sp
[0].end_seq
;
5238 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5240 max_sack
= after(sp
[this_sack
].end_seq
,
5242 sp
[this_sack
].end_seq
: max_sack
;
5245 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5246 rst_seq_match
= true;
5252 tcp_send_challenge_ack(sk
, skb
);
5256 /* step 3: check security and precedence [ignored] */
5258 /* step 4: Check for a SYN
5259 * RFC 5961 4.2 : Send a challenge ack
5264 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5265 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5266 tcp_send_challenge_ack(sk
, skb
);
5278 * TCP receive function for the ESTABLISHED state.
5280 * It is split into a fast path and a slow path. The fast path is
5282 * - A zero window was announced from us - zero window probing
5283 * is only handled properly in the slow path.
5284 * - Out of order segments arrived.
5285 * - Urgent data is expected.
5286 * - There is no buffer space left
5287 * - Unexpected TCP flags/window values/header lengths are received
5288 * (detected by checking the TCP header against pred_flags)
5289 * - Data is sent in both directions. Fast path only supports pure senders
5290 * or pure receivers (this means either the sequence number or the ack
5291 * value must stay constant)
5292 * - Unexpected TCP option.
5294 * When these conditions are not satisfied it drops into a standard
5295 * receive procedure patterned after RFC793 to handle all cases.
5296 * The first three cases are guaranteed by proper pred_flags setting,
5297 * the rest is checked inline. Fast processing is turned on in
5298 * tcp_data_queue when everything is OK.
5300 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5301 const struct tcphdr
*th
, unsigned int len
)
5303 struct tcp_sock
*tp
= tcp_sk(sk
);
5305 if (unlikely(!sk
->sk_rx_dst
))
5306 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5308 * Header prediction.
5309 * The code loosely follows the one in the famous
5310 * "30 instruction TCP receive" Van Jacobson mail.
5312 * Van's trick is to deposit buffers into socket queue
5313 * on a device interrupt, to call tcp_recv function
5314 * on the receive process context and checksum and copy
5315 * the buffer to user space. smart...
5317 * Our current scheme is not silly either but we take the
5318 * extra cost of the net_bh soft interrupt processing...
5319 * We do checksum and copy also but from device to kernel.
5322 tp
->rx_opt
.saw_tstamp
= 0;
5324 /* pred_flags is 0xS?10 << 16 + snd_wnd
5325 * if header_prediction is to be made
5326 * 'S' will always be tp->tcp_header_len >> 2
5327 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5328 * turn it off (when there are holes in the receive
5329 * space for instance)
5330 * PSH flag is ignored.
5333 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5334 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5335 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5336 int tcp_header_len
= tp
->tcp_header_len
;
5338 /* Timestamp header prediction: tcp_header_len
5339 * is automatically equal to th->doff*4 due to pred_flags
5343 /* Check timestamp */
5344 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5345 /* No? Slow path! */
5346 if (!tcp_parse_aligned_timestamp(tp
, th
))
5349 /* If PAWS failed, check it more carefully in slow path */
5350 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5353 /* DO NOT update ts_recent here, if checksum fails
5354 * and timestamp was corrupted part, it will result
5355 * in a hung connection since we will drop all
5356 * future packets due to the PAWS test.
5360 if (len
<= tcp_header_len
) {
5361 /* Bulk data transfer: sender */
5362 if (len
== tcp_header_len
) {
5363 /* Predicted packet is in window by definition.
5364 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5365 * Hence, check seq<=rcv_wup reduces to:
5367 if (tcp_header_len
==
5368 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5369 tp
->rcv_nxt
== tp
->rcv_wup
)
5370 tcp_store_ts_recent(tp
);
5372 /* We know that such packets are checksummed
5375 tcp_ack(sk
, skb
, 0);
5377 tcp_data_snd_check(sk
);
5379 } else { /* Header too small */
5380 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5385 bool fragstolen
= false;
5387 if (tp
->ucopy
.task
== current
&&
5388 tp
->copied_seq
== tp
->rcv_nxt
&&
5389 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5390 sock_owned_by_user(sk
)) {
5391 __set_current_state(TASK_RUNNING
);
5393 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5394 /* Predicted packet is in window by definition.
5395 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5396 * Hence, check seq<=rcv_wup reduces to:
5398 if (tcp_header_len
==
5399 (sizeof(struct tcphdr
) +
5400 TCPOLEN_TSTAMP_ALIGNED
) &&
5401 tp
->rcv_nxt
== tp
->rcv_wup
)
5402 tcp_store_ts_recent(tp
);
5404 tcp_rcv_rtt_measure_ts(sk
, skb
);
5406 __skb_pull(skb
, tcp_header_len
);
5407 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5408 NET_INC_STATS(sock_net(sk
),
5409 LINUX_MIB_TCPHPHITSTOUSER
);
5414 if (tcp_checksum_complete(skb
))
5417 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5420 /* Predicted packet is in window by definition.
5421 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5422 * Hence, check seq<=rcv_wup reduces to:
5424 if (tcp_header_len
==
5425 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5426 tp
->rcv_nxt
== tp
->rcv_wup
)
5427 tcp_store_ts_recent(tp
);
5429 tcp_rcv_rtt_measure_ts(sk
, skb
);
5431 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5433 /* Bulk data transfer: receiver */
5434 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5438 tcp_event_data_recv(sk
, skb
);
5440 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5441 /* Well, only one small jumplet in fast path... */
5442 tcp_ack(sk
, skb
, FLAG_DATA
);
5443 tcp_data_snd_check(sk
);
5444 if (!inet_csk_ack_scheduled(sk
))
5448 __tcp_ack_snd_check(sk
, 0);
5451 kfree_skb_partial(skb
, fragstolen
);
5452 sk
->sk_data_ready(sk
);
5458 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5461 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5465 * Standard slow path.
5468 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5472 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5475 tcp_rcv_rtt_measure_ts(sk
, skb
);
5477 /* Process urgent data. */
5478 tcp_urg(sk
, skb
, th
);
5480 /* step 7: process the segment text */
5481 tcp_data_queue(sk
, skb
);
5483 tcp_data_snd_check(sk
);
5484 tcp_ack_snd_check(sk
);
5488 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5489 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5494 EXPORT_SYMBOL(tcp_rcv_established
);
5496 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5498 struct tcp_sock
*tp
= tcp_sk(sk
);
5499 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5501 tcp_set_state(sk
, TCP_ESTABLISHED
);
5504 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5505 security_inet_conn_established(sk
, skb
);
5508 /* Make sure socket is routed, for correct metrics. */
5509 icsk
->icsk_af_ops
->rebuild_header(sk
);
5511 tcp_init_metrics(sk
);
5513 tcp_init_congestion_control(sk
);
5515 /* Prevent spurious tcp_cwnd_restart() on first data
5518 tp
->lsndtime
= tcp_time_stamp
;
5520 tcp_init_buffer_space(sk
);
5522 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5523 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5525 if (!tp
->rx_opt
.snd_wscale
)
5526 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5530 if (!sock_flag(sk
, SOCK_DEAD
)) {
5531 sk
->sk_state_change(sk
);
5532 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5536 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5537 struct tcp_fastopen_cookie
*cookie
)
5539 struct tcp_sock
*tp
= tcp_sk(sk
);
5540 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5541 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5542 bool syn_drop
= false;
5544 if (mss
== tp
->rx_opt
.user_mss
) {
5545 struct tcp_options_received opt
;
5547 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5548 tcp_clear_options(&opt
);
5549 opt
.user_mss
= opt
.mss_clamp
= 0;
5550 tcp_parse_options(synack
, &opt
, 0, NULL
);
5551 mss
= opt
.mss_clamp
;
5554 if (!tp
->syn_fastopen
) {
5555 /* Ignore an unsolicited cookie */
5557 } else if (tp
->total_retrans
) {
5558 /* SYN timed out and the SYN-ACK neither has a cookie nor
5559 * acknowledges data. Presumably the remote received only
5560 * the retransmitted (regular) SYNs: either the original
5561 * SYN-data or the corresponding SYN-ACK was dropped.
5563 syn_drop
= (cookie
->len
< 0 && data
);
5564 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5565 /* We requested a cookie but didn't get it. If we did not use
5566 * the (old) exp opt format then try so next time (try_exp=1).
5567 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5569 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5572 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5574 if (data
) { /* Retransmit unacked data in SYN */
5575 tcp_for_write_queue_from(data
, sk
) {
5576 if (data
== tcp_send_head(sk
) ||
5577 __tcp_retransmit_skb(sk
, data
, 1))
5581 NET_INC_STATS(sock_net(sk
),
5582 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5585 tp
->syn_data_acked
= tp
->syn_data
;
5586 if (tp
->syn_data_acked
)
5587 NET_INC_STATS(sock_net(sk
),
5588 LINUX_MIB_TCPFASTOPENACTIVE
);
5590 tcp_fastopen_add_skb(sk
, synack
);
5595 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5596 const struct tcphdr
*th
)
5598 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5599 struct tcp_sock
*tp
= tcp_sk(sk
);
5600 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5601 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5603 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5604 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5605 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5609 * "If the state is SYN-SENT then
5610 * first check the ACK bit
5611 * If the ACK bit is set
5612 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5613 * a reset (unless the RST bit is set, if so drop
5614 * the segment and return)"
5616 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5617 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5618 goto reset_and_undo
;
5620 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5621 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5623 NET_INC_STATS(sock_net(sk
),
5624 LINUX_MIB_PAWSACTIVEREJECTED
);
5625 goto reset_and_undo
;
5628 /* Now ACK is acceptable.
5630 * "If the RST bit is set
5631 * If the ACK was acceptable then signal the user "error:
5632 * connection reset", drop the segment, enter CLOSED state,
5633 * delete TCB, and return."
5642 * "fifth, if neither of the SYN or RST bits is set then
5643 * drop the segment and return."
5649 goto discard_and_undo
;
5652 * "If the SYN bit is on ...
5653 * are acceptable then ...
5654 * (our SYN has been ACKed), change the connection
5655 * state to ESTABLISHED..."
5658 tcp_ecn_rcv_synack(tp
, th
);
5660 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5661 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5663 /* Ok.. it's good. Set up sequence numbers and
5664 * move to established.
5666 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5667 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5669 /* RFC1323: The window in SYN & SYN/ACK segments is
5672 tp
->snd_wnd
= ntohs(th
->window
);
5674 if (!tp
->rx_opt
.wscale_ok
) {
5675 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5676 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5679 if (tp
->rx_opt
.saw_tstamp
) {
5680 tp
->rx_opt
.tstamp_ok
= 1;
5681 tp
->tcp_header_len
=
5682 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5683 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5684 tcp_store_ts_recent(tp
);
5686 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5689 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5690 tcp_enable_fack(tp
);
5693 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5694 tcp_initialize_rcv_mss(sk
);
5696 /* Remember, tcp_poll() does not lock socket!
5697 * Change state from SYN-SENT only after copied_seq
5698 * is initialized. */
5699 tp
->copied_seq
= tp
->rcv_nxt
;
5703 tcp_finish_connect(sk
, skb
);
5705 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5706 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5709 if (sk
->sk_write_pending
||
5710 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5711 icsk
->icsk_ack
.pingpong
) {
5712 /* Save one ACK. Data will be ready after
5713 * several ticks, if write_pending is set.
5715 * It may be deleted, but with this feature tcpdumps
5716 * look so _wonderfully_ clever, that I was not able
5717 * to stand against the temptation 8) --ANK
5719 inet_csk_schedule_ack(sk
);
5720 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5721 tcp_enter_quickack_mode(sk
);
5722 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5723 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5734 /* No ACK in the segment */
5738 * "If the RST bit is set
5740 * Otherwise (no ACK) drop the segment and return."
5743 goto discard_and_undo
;
5747 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5748 tcp_paws_reject(&tp
->rx_opt
, 0))
5749 goto discard_and_undo
;
5752 /* We see SYN without ACK. It is attempt of
5753 * simultaneous connect with crossed SYNs.
5754 * Particularly, it can be connect to self.
5756 tcp_set_state(sk
, TCP_SYN_RECV
);
5758 if (tp
->rx_opt
.saw_tstamp
) {
5759 tp
->rx_opt
.tstamp_ok
= 1;
5760 tcp_store_ts_recent(tp
);
5761 tp
->tcp_header_len
=
5762 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5764 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5767 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5768 tp
->copied_seq
= tp
->rcv_nxt
;
5769 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5771 /* RFC1323: The window in SYN & SYN/ACK segments is
5774 tp
->snd_wnd
= ntohs(th
->window
);
5775 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5776 tp
->max_window
= tp
->snd_wnd
;
5778 tcp_ecn_rcv_syn(tp
, th
);
5781 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5782 tcp_initialize_rcv_mss(sk
);
5784 tcp_send_synack(sk
);
5786 /* Note, we could accept data and URG from this segment.
5787 * There are no obstacles to make this (except that we must
5788 * either change tcp_recvmsg() to prevent it from returning data
5789 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5791 * However, if we ignore data in ACKless segments sometimes,
5792 * we have no reasons to accept it sometimes.
5793 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5794 * is not flawless. So, discard packet for sanity.
5795 * Uncomment this return to process the data.
5802 /* "fifth, if neither of the SYN or RST bits is set then
5803 * drop the segment and return."
5807 tcp_clear_options(&tp
->rx_opt
);
5808 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5812 tcp_clear_options(&tp
->rx_opt
);
5813 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5818 * This function implements the receiving procedure of RFC 793 for
5819 * all states except ESTABLISHED and TIME_WAIT.
5820 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5821 * address independent.
5824 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5826 struct tcp_sock
*tp
= tcp_sk(sk
);
5827 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5828 const struct tcphdr
*th
= tcp_hdr(skb
);
5829 struct request_sock
*req
;
5833 switch (sk
->sk_state
) {
5847 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5856 tp
->rx_opt
.saw_tstamp
= 0;
5857 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5861 /* Do step6 onward by hand. */
5862 tcp_urg(sk
, skb
, th
);
5864 tcp_data_snd_check(sk
);
5868 tp
->rx_opt
.saw_tstamp
= 0;
5869 req
= tp
->fastopen_rsk
;
5871 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5872 sk
->sk_state
!= TCP_FIN_WAIT1
);
5874 if (!tcp_check_req(sk
, skb
, req
, true))
5878 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5881 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5884 /* step 5: check the ACK field */
5885 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5886 FLAG_UPDATE_TS_RECENT
) > 0;
5888 switch (sk
->sk_state
) {
5894 tcp_synack_rtt_meas(sk
, req
);
5896 /* Once we leave TCP_SYN_RECV, we no longer need req
5900 tp
->total_retrans
= req
->num_retrans
;
5901 reqsk_fastopen_remove(sk
, req
, false);
5903 /* Make sure socket is routed, for correct metrics. */
5904 icsk
->icsk_af_ops
->rebuild_header(sk
);
5905 tcp_init_congestion_control(sk
);
5908 tp
->copied_seq
= tp
->rcv_nxt
;
5909 tcp_init_buffer_space(sk
);
5912 tcp_set_state(sk
, TCP_ESTABLISHED
);
5913 sk
->sk_state_change(sk
);
5915 /* Note, that this wakeup is only for marginal crossed SYN case.
5916 * Passively open sockets are not waked up, because
5917 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5920 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5922 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5923 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5924 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5926 if (tp
->rx_opt
.tstamp_ok
)
5927 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5930 /* Re-arm the timer because data may have been sent out.
5931 * This is similar to the regular data transmission case
5932 * when new data has just been ack'ed.
5934 * (TFO) - we could try to be more aggressive and
5935 * retransmitting any data sooner based on when they
5940 tcp_init_metrics(sk
);
5942 tcp_update_pacing_rate(sk
);
5944 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5945 tp
->lsndtime
= tcp_time_stamp
;
5947 tcp_initialize_rcv_mss(sk
);
5948 tcp_fast_path_on(tp
);
5951 case TCP_FIN_WAIT1
: {
5952 struct dst_entry
*dst
;
5955 /* If we enter the TCP_FIN_WAIT1 state and we are a
5956 * Fast Open socket and this is the first acceptable
5957 * ACK we have received, this would have acknowledged
5958 * our SYNACK so stop the SYNACK timer.
5961 /* Return RST if ack_seq is invalid.
5962 * Note that RFC793 only says to generate a
5963 * DUPACK for it but for TCP Fast Open it seems
5964 * better to treat this case like TCP_SYN_RECV
5969 /* We no longer need the request sock. */
5970 reqsk_fastopen_remove(sk
, req
, false);
5973 if (tp
->snd_una
!= tp
->write_seq
)
5976 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5977 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5979 dst
= __sk_dst_get(sk
);
5983 if (!sock_flag(sk
, SOCK_DEAD
)) {
5984 /* Wake up lingering close() */
5985 sk
->sk_state_change(sk
);
5989 if (tp
->linger2
< 0 ||
5990 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5991 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5993 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5997 tmo
= tcp_fin_time(sk
);
5998 if (tmo
> TCP_TIMEWAIT_LEN
) {
5999 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6000 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6001 /* Bad case. We could lose such FIN otherwise.
6002 * It is not a big problem, but it looks confusing
6003 * and not so rare event. We still can lose it now,
6004 * if it spins in bh_lock_sock(), but it is really
6007 inet_csk_reset_keepalive_timer(sk
, tmo
);
6009 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6016 if (tp
->snd_una
== tp
->write_seq
) {
6017 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6023 if (tp
->snd_una
== tp
->write_seq
) {
6024 tcp_update_metrics(sk
);
6031 /* step 6: check the URG bit */
6032 tcp_urg(sk
, skb
, th
);
6034 /* step 7: process the segment text */
6035 switch (sk
->sk_state
) {
6036 case TCP_CLOSE_WAIT
:
6039 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6043 /* RFC 793 says to queue data in these states,
6044 * RFC 1122 says we MUST send a reset.
6045 * BSD 4.4 also does reset.
6047 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6048 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6049 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6050 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6056 case TCP_ESTABLISHED
:
6057 tcp_data_queue(sk
, skb
);
6062 /* tcp_data could move socket to TIME-WAIT */
6063 if (sk
->sk_state
!= TCP_CLOSE
) {
6064 tcp_data_snd_check(sk
);
6065 tcp_ack_snd_check(sk
);
6074 EXPORT_SYMBOL(tcp_rcv_state_process
);
6076 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6078 struct inet_request_sock
*ireq
= inet_rsk(req
);
6080 if (family
== AF_INET
)
6081 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6082 &ireq
->ir_rmt_addr
, port
);
6083 #if IS_ENABLED(CONFIG_IPV6)
6084 else if (family
== AF_INET6
)
6085 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6086 &ireq
->ir_v6_rmt_addr
, port
);
6090 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6092 * If we receive a SYN packet with these bits set, it means a
6093 * network is playing bad games with TOS bits. In order to
6094 * avoid possible false congestion notifications, we disable
6095 * TCP ECN negotiation.
6097 * Exception: tcp_ca wants ECN. This is required for DCTCP
6098 * congestion control: Linux DCTCP asserts ECT on all packets,
6099 * including SYN, which is most optimal solution; however,
6100 * others, such as FreeBSD do not.
6102 static void tcp_ecn_create_request(struct request_sock
*req
,
6103 const struct sk_buff
*skb
,
6104 const struct sock
*listen_sk
,
6105 const struct dst_entry
*dst
)
6107 const struct tcphdr
*th
= tcp_hdr(skb
);
6108 const struct net
*net
= sock_net(listen_sk
);
6109 bool th_ecn
= th
->ece
&& th
->cwr
;
6116 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6117 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6118 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6120 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6121 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6122 inet_rsk(req
)->ecn_ok
= 1;
6125 static void tcp_openreq_init(struct request_sock
*req
,
6126 const struct tcp_options_received
*rx_opt
,
6127 struct sk_buff
*skb
, const struct sock
*sk
)
6129 struct inet_request_sock
*ireq
= inet_rsk(req
);
6131 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6133 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6134 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6135 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6136 tcp_rsk(req
)->last_oow_ack_time
= 0;
6137 req
->mss
= rx_opt
->mss_clamp
;
6138 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6139 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6140 ireq
->sack_ok
= rx_opt
->sack_ok
;
6141 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6142 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6145 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6146 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6147 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6150 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6151 struct sock
*sk_listener
,
6152 bool attach_listener
)
6154 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6158 struct inet_request_sock
*ireq
= inet_rsk(req
);
6160 kmemcheck_annotate_bitfield(ireq
, flags
);
6162 #if IS_ENABLED(CONFIG_IPV6)
6163 ireq
->pktopts
= NULL
;
6165 atomic64_set(&ireq
->ir_cookie
, 0);
6166 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6167 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6168 ireq
->ireq_family
= sk_listener
->sk_family
;
6173 EXPORT_SYMBOL(inet_reqsk_alloc
);
6176 * Return true if a syncookie should be sent
6178 static bool tcp_syn_flood_action(const struct sock
*sk
,
6179 const struct sk_buff
*skb
,
6182 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6183 const char *msg
= "Dropping request";
6184 bool want_cookie
= false;
6185 struct net
*net
= sock_net(sk
);
6187 #ifdef CONFIG_SYN_COOKIES
6188 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6189 msg
= "Sending cookies";
6191 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6194 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6196 if (!queue
->synflood_warned
&&
6197 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6198 xchg(&queue
->synflood_warned
, 1) == 0)
6199 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6200 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6205 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6206 struct request_sock
*req
,
6207 const struct sk_buff
*skb
)
6209 if (tcp_sk(sk
)->save_syn
) {
6210 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6213 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6216 memcpy(©
[1], skb_network_header(skb
), len
);
6217 req
->saved_syn
= copy
;
6222 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6223 const struct tcp_request_sock_ops
*af_ops
,
6224 struct sock
*sk
, struct sk_buff
*skb
)
6226 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6227 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6228 struct tcp_options_received tmp_opt
;
6229 struct tcp_sock
*tp
= tcp_sk(sk
);
6230 struct net
*net
= sock_net(sk
);
6231 struct sock
*fastopen_sk
= NULL
;
6232 struct dst_entry
*dst
= NULL
;
6233 struct request_sock
*req
;
6234 bool want_cookie
= false;
6237 /* TW buckets are converted to open requests without
6238 * limitations, they conserve resources and peer is
6239 * evidently real one.
6241 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6242 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6243 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6249 /* Accept backlog is full. If we have already queued enough
6250 * of warm entries in syn queue, drop request. It is better than
6251 * clogging syn queue with openreqs with exponentially increasing
6254 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
6255 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6259 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6263 tcp_rsk(req
)->af_specific
= af_ops
;
6265 tcp_clear_options(&tmp_opt
);
6266 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6267 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6268 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6270 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6271 tcp_clear_options(&tmp_opt
);
6273 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6274 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6276 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6277 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6279 af_ops
->init_req(req
, sk
, skb
);
6281 if (security_inet_conn_request(sk
, skb
, req
))
6284 if (!want_cookie
&& !isn
) {
6285 /* VJ's idea. We save last timestamp seen
6286 * from the destination in peer table, when entering
6287 * state TIME-WAIT, and check against it before
6288 * accepting new connection request.
6290 * If "isn" is not zero, this request hit alive
6291 * timewait bucket, so that all the necessary checks
6292 * are made in the function processing timewait state.
6294 if (tcp_death_row
.sysctl_tw_recycle
) {
6297 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6299 if (dst
&& strict
&&
6300 !tcp_peer_is_proven(req
, dst
, true,
6301 tmp_opt
.saw_tstamp
)) {
6302 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6303 goto drop_and_release
;
6306 /* Kill the following clause, if you dislike this way. */
6307 else if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6308 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6309 (sysctl_max_syn_backlog
>> 2)) &&
6310 !tcp_peer_is_proven(req
, dst
, false,
6311 tmp_opt
.saw_tstamp
)) {
6312 /* Without syncookies last quarter of
6313 * backlog is filled with destinations,
6314 * proven to be alive.
6315 * It means that we continue to communicate
6316 * to destinations, already remembered
6317 * to the moment of synflood.
6319 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6321 goto drop_and_release
;
6324 isn
= af_ops
->init_seq(skb
);
6327 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6332 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6335 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6336 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6337 if (!tmp_opt
.tstamp_ok
)
6338 inet_rsk(req
)->ecn_ok
= 0;
6341 tcp_rsk(req
)->snt_isn
= isn
;
6342 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6343 tcp_openreq_init_rwin(req
, sk
, dst
);
6345 tcp_reqsk_record_syn(sk
, req
, skb
);
6346 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6349 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6350 &foc
, TCP_SYNACK_FASTOPEN
);
6351 /* Add the child socket directly into the accept queue */
6352 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6353 sk
->sk_data_ready(sk
);
6354 bh_unlock_sock(fastopen_sk
);
6355 sock_put(fastopen_sk
);
6357 tcp_rsk(req
)->tfo_listener
= false;
6359 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6360 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6361 !want_cookie
? TCP_SYNACK_NORMAL
:
6379 EXPORT_SYMBOL(tcp_conn_request
);