1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
80 int sysctl_tcp_fack __read_mostly
;
81 int sysctl_tcp_max_reordering __read_mostly
= 300;
82 int sysctl_tcp_dsack __read_mostly
= 1;
83 int sysctl_tcp_app_win __read_mostly
= 31;
84 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
85 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
87 /* rfc5961 challenge ack rate limiting */
88 int sysctl_tcp_challenge_ack_limit
= 1000;
90 int sysctl_tcp_stdurg __read_mostly
;
91 int sysctl_tcp_rfc1337 __read_mostly
;
92 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
93 int sysctl_tcp_frto __read_mostly
= 2;
94 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
95 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
96 int sysctl_tcp_early_retrans __read_mostly
= 3;
97 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
99 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
100 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
101 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
102 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
103 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
104 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
105 #define FLAG_ECE 0x40 /* ECE in this ACK */
106 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
107 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
108 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
109 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
110 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
111 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
112 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
113 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
114 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 #define REXMIT_NONE 0 /* no loss recovery to do */
125 #define REXMIT_LOST 1 /* retransmit packets marked lost */
126 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
128 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
131 static bool __once __read_mostly
;
134 struct net_device
*dev
;
139 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
140 if (!dev
|| len
>= dev
->mtu
)
141 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
142 dev
? dev
->name
: "Unknown driver");
147 /* Adapt the MSS value used to make delayed ack decision to the
150 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
152 struct inet_connection_sock
*icsk
= inet_csk(sk
);
153 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
156 icsk
->icsk_ack
.last_seg_size
= 0;
158 /* skb->len may jitter because of SACKs, even if peer
159 * sends good full-sized frames.
161 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
162 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
163 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
165 /* Account for possibly-removed options */
166 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
167 MAX_TCP_OPTION_SPACE
))
168 tcp_gro_dev_warn(sk
, skb
, len
);
170 /* Otherwise, we make more careful check taking into account,
171 * that SACKs block is variable.
173 * "len" is invariant segment length, including TCP header.
175 len
+= skb
->data
- skb_transport_header(skb
);
176 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
177 /* If PSH is not set, packet should be
178 * full sized, provided peer TCP is not badly broken.
179 * This observation (if it is correct 8)) allows
180 * to handle super-low mtu links fairly.
182 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
183 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
184 /* Subtract also invariant (if peer is RFC compliant),
185 * tcp header plus fixed timestamp option length.
186 * Resulting "len" is MSS free of SACK jitter.
188 len
-= tcp_sk(sk
)->tcp_header_len
;
189 icsk
->icsk_ack
.last_seg_size
= len
;
191 icsk
->icsk_ack
.rcv_mss
= len
;
195 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
196 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
197 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
201 static void tcp_incr_quickack(struct sock
*sk
)
203 struct inet_connection_sock
*icsk
= inet_csk(sk
);
204 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
208 if (quickacks
> icsk
->icsk_ack
.quick
)
209 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
212 static void tcp_enter_quickack_mode(struct sock
*sk
)
214 struct inet_connection_sock
*icsk
= inet_csk(sk
);
215 tcp_incr_quickack(sk
);
216 icsk
->icsk_ack
.pingpong
= 0;
217 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
220 /* Send ACKs quickly, if "quick" count is not exhausted
221 * and the session is not interactive.
224 static bool tcp_in_quickack_mode(struct sock
*sk
)
226 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
227 const struct dst_entry
*dst
= __sk_dst_get(sk
);
229 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
230 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
233 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
235 if (tp
->ecn_flags
& TCP_ECN_OK
)
236 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
239 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
241 if (tcp_hdr(skb
)->cwr
)
242 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
245 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
247 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
250 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
252 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
253 case INET_ECN_NOT_ECT
:
254 /* Funny extension: if ECT is not set on a segment,
255 * and we already seen ECT on a previous segment,
256 * it is probably a retransmit.
258 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
259 tcp_enter_quickack_mode((struct sock
*)tp
);
262 if (tcp_ca_needs_ecn((struct sock
*)tp
))
263 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
265 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
266 /* Better not delay acks, sender can have a very low cwnd */
267 tcp_enter_quickack_mode((struct sock
*)tp
);
268 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
270 tp
->ecn_flags
|= TCP_ECN_SEEN
;
273 if (tcp_ca_needs_ecn((struct sock
*)tp
))
274 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
275 tp
->ecn_flags
|= TCP_ECN_SEEN
;
280 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
282 if (tp
->ecn_flags
& TCP_ECN_OK
)
283 __tcp_ecn_check_ce(tp
, skb
);
286 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
288 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
289 tp
->ecn_flags
&= ~TCP_ECN_OK
;
292 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
294 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
295 tp
->ecn_flags
&= ~TCP_ECN_OK
;
298 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
300 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
305 /* Buffer size and advertised window tuning.
307 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
310 static void tcp_sndbuf_expand(struct sock
*sk
)
312 const struct tcp_sock
*tp
= tcp_sk(sk
);
313 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
317 /* Worst case is non GSO/TSO : each frame consumes one skb
318 * and skb->head is kmalloced using power of two area of memory
320 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
322 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
324 per_mss
= roundup_pow_of_two(per_mss
) +
325 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
327 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
328 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
330 /* Fast Recovery (RFC 5681 3.2) :
331 * Cubic needs 1.7 factor, rounded to 2 to include
332 * extra cushion (application might react slowly to POLLOUT)
334 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
335 sndmem
*= nr_segs
* per_mss
;
337 if (sk
->sk_sndbuf
< sndmem
)
338 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
341 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
343 * All tcp_full_space() is split to two parts: "network" buffer, allocated
344 * forward and advertised in receiver window (tp->rcv_wnd) and
345 * "application buffer", required to isolate scheduling/application
346 * latencies from network.
347 * window_clamp is maximal advertised window. It can be less than
348 * tcp_full_space(), in this case tcp_full_space() - window_clamp
349 * is reserved for "application" buffer. The less window_clamp is
350 * the smoother our behaviour from viewpoint of network, but the lower
351 * throughput and the higher sensitivity of the connection to losses. 8)
353 * rcv_ssthresh is more strict window_clamp used at "slow start"
354 * phase to predict further behaviour of this connection.
355 * It is used for two goals:
356 * - to enforce header prediction at sender, even when application
357 * requires some significant "application buffer". It is check #1.
358 * - to prevent pruning of receive queue because of misprediction
359 * of receiver window. Check #2.
361 * The scheme does not work when sender sends good segments opening
362 * window and then starts to feed us spaghetti. But it should work
363 * in common situations. Otherwise, we have to rely on queue collapsing.
366 /* Slow part of check#2. */
367 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
369 struct tcp_sock
*tp
= tcp_sk(sk
);
371 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
372 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
374 while (tp
->rcv_ssthresh
<= window
) {
375 if (truesize
<= skb
->len
)
376 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
384 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
386 struct tcp_sock
*tp
= tcp_sk(sk
);
389 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
390 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
391 !tcp_under_memory_pressure(sk
)) {
394 /* Check #2. Increase window, if skb with such overhead
395 * will fit to rcvbuf in future.
397 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
398 incr
= 2 * tp
->advmss
;
400 incr
= __tcp_grow_window(sk
, skb
);
403 incr
= max_t(int, incr
, 2 * skb
->len
);
404 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
406 inet_csk(sk
)->icsk_ack
.quick
|= 1;
411 /* 3. Tuning rcvbuf, when connection enters established state. */
412 static void tcp_fixup_rcvbuf(struct sock
*sk
)
414 u32 mss
= tcp_sk(sk
)->advmss
;
417 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
418 tcp_default_init_rwnd(mss
);
420 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
421 * Allow enough cushion so that sender is not limited by our window
423 if (sysctl_tcp_moderate_rcvbuf
)
426 if (sk
->sk_rcvbuf
< rcvmem
)
427 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
430 /* 4. Try to fixup all. It is made immediately after connection enters
433 void tcp_init_buffer_space(struct sock
*sk
)
435 struct tcp_sock
*tp
= tcp_sk(sk
);
438 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
439 tcp_fixup_rcvbuf(sk
);
440 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
441 tcp_sndbuf_expand(sk
);
443 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
444 tcp_mstamp_refresh(tp
);
445 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
446 tp
->rcvq_space
.seq
= tp
->copied_seq
;
448 maxwin
= tcp_full_space(sk
);
450 if (tp
->window_clamp
>= maxwin
) {
451 tp
->window_clamp
= maxwin
;
453 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
454 tp
->window_clamp
= max(maxwin
-
455 (maxwin
>> sysctl_tcp_app_win
),
459 /* Force reservation of one segment. */
460 if (sysctl_tcp_app_win
&&
461 tp
->window_clamp
> 2 * tp
->advmss
&&
462 tp
->window_clamp
+ tp
->advmss
> maxwin
)
463 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
465 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
466 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
469 /* 5. Recalculate window clamp after socket hit its memory bounds. */
470 static void tcp_clamp_window(struct sock
*sk
)
472 struct tcp_sock
*tp
= tcp_sk(sk
);
473 struct inet_connection_sock
*icsk
= inet_csk(sk
);
475 icsk
->icsk_ack
.quick
= 0;
477 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
478 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
479 !tcp_under_memory_pressure(sk
) &&
480 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
481 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
484 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
485 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
488 /* Initialize RCV_MSS value.
489 * RCV_MSS is an our guess about MSS used by the peer.
490 * We haven't any direct information about the MSS.
491 * It's better to underestimate the RCV_MSS rather than overestimate.
492 * Overestimations make us ACKing less frequently than needed.
493 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
495 void tcp_initialize_rcv_mss(struct sock
*sk
)
497 const struct tcp_sock
*tp
= tcp_sk(sk
);
498 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
500 hint
= min(hint
, tp
->rcv_wnd
/ 2);
501 hint
= min(hint
, TCP_MSS_DEFAULT
);
502 hint
= max(hint
, TCP_MIN_MSS
);
504 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
506 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
508 /* Receiver "autotuning" code.
510 * The algorithm for RTT estimation w/o timestamps is based on
511 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
512 * <http://public.lanl.gov/radiant/pubs.html#DRS>
514 * More detail on this code can be found at
515 * <http://staff.psc.edu/jheffner/>,
516 * though this reference is out of date. A new paper
519 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
521 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
527 if (new_sample
!= 0) {
528 /* If we sample in larger samples in the non-timestamp
529 * case, we could grossly overestimate the RTT especially
530 * with chatty applications or bulk transfer apps which
531 * are stalled on filesystem I/O.
533 * Also, since we are only going for a minimum in the
534 * non-timestamp case, we do not smooth things out
535 * else with timestamps disabled convergence takes too
539 m
-= (new_sample
>> 3);
547 /* No previous measure. */
551 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
554 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
558 if (tp
->rcv_rtt_est
.time
== 0)
560 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
562 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
563 tcp_rcv_rtt_update(tp
, delta_us
, 1);
566 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
567 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
570 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
571 const struct sk_buff
*skb
)
573 struct tcp_sock
*tp
= tcp_sk(sk
);
575 if (tp
->rx_opt
.rcv_tsecr
&&
576 (TCP_SKB_CB(skb
)->end_seq
-
577 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
)) {
578 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
579 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
581 tcp_rcv_rtt_update(tp
, delta_us
, 0);
586 * This function should be called every time data is copied to user space.
587 * It calculates the appropriate TCP receive buffer space.
589 void tcp_rcv_space_adjust(struct sock
*sk
)
591 struct tcp_sock
*tp
= tcp_sk(sk
);
595 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
596 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
599 /* Number of bytes copied to user in last RTT */
600 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
601 if (copied
<= tp
->rcvq_space
.space
)
605 * copied = bytes received in previous RTT, our base window
606 * To cope with packet losses, we need a 2x factor
607 * To cope with slow start, and sender growing its cwin by 100 %
608 * every RTT, we need a 4x factor, because the ACK we are sending
609 * now is for the next RTT, not the current one :
610 * <prev RTT . ><current RTT .. ><next RTT .... >
613 if (sysctl_tcp_moderate_rcvbuf
&&
614 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
615 int rcvwin
, rcvmem
, rcvbuf
;
617 /* minimal window to cope with packet losses, assuming
618 * steady state. Add some cushion because of small variations.
620 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
622 /* If rate increased by 25%,
623 * assume slow start, rcvwin = 3 * copied
624 * If rate increased by 50%,
625 * assume sender can use 2x growth, rcvwin = 4 * copied
628 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
630 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
633 rcvwin
+= (rcvwin
>> 1);
636 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
637 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
640 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
641 if (rcvbuf
> sk
->sk_rcvbuf
) {
642 sk
->sk_rcvbuf
= rcvbuf
;
644 /* Make the window clamp follow along. */
645 tp
->window_clamp
= rcvwin
;
648 tp
->rcvq_space
.space
= copied
;
651 tp
->rcvq_space
.seq
= tp
->copied_seq
;
652 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
655 /* There is something which you must keep in mind when you analyze the
656 * behavior of the tp->ato delayed ack timeout interval. When a
657 * connection starts up, we want to ack as quickly as possible. The
658 * problem is that "good" TCP's do slow start at the beginning of data
659 * transmission. The means that until we send the first few ACK's the
660 * sender will sit on his end and only queue most of his data, because
661 * he can only send snd_cwnd unacked packets at any given time. For
662 * each ACK we send, he increments snd_cwnd and transmits more of his
665 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
667 struct tcp_sock
*tp
= tcp_sk(sk
);
668 struct inet_connection_sock
*icsk
= inet_csk(sk
);
671 inet_csk_schedule_ack(sk
);
673 tcp_measure_rcv_mss(sk
, skb
);
675 tcp_rcv_rtt_measure(tp
);
679 if (!icsk
->icsk_ack
.ato
) {
680 /* The _first_ data packet received, initialize
681 * delayed ACK engine.
683 tcp_incr_quickack(sk
);
684 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
686 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
688 if (m
<= TCP_ATO_MIN
/ 2) {
689 /* The fastest case is the first. */
690 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
691 } else if (m
< icsk
->icsk_ack
.ato
) {
692 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
693 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
694 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
695 } else if (m
> icsk
->icsk_rto
) {
696 /* Too long gap. Apparently sender failed to
697 * restart window, so that we send ACKs quickly.
699 tcp_incr_quickack(sk
);
703 icsk
->icsk_ack
.lrcvtime
= now
;
705 tcp_ecn_check_ce(tp
, skb
);
708 tcp_grow_window(sk
, skb
);
711 /* Called to compute a smoothed rtt estimate. The data fed to this
712 * routine either comes from timestamps, or from segments that were
713 * known _not_ to have been retransmitted [see Karn/Partridge
714 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
715 * piece by Van Jacobson.
716 * NOTE: the next three routines used to be one big routine.
717 * To save cycles in the RFC 1323 implementation it was better to break
718 * it up into three procedures. -- erics
720 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
722 struct tcp_sock
*tp
= tcp_sk(sk
);
723 long m
= mrtt_us
; /* RTT */
724 u32 srtt
= tp
->srtt_us
;
726 /* The following amusing code comes from Jacobson's
727 * article in SIGCOMM '88. Note that rtt and mdev
728 * are scaled versions of rtt and mean deviation.
729 * This is designed to be as fast as possible
730 * m stands for "measurement".
732 * On a 1990 paper the rto value is changed to:
733 * RTO = rtt + 4 * mdev
735 * Funny. This algorithm seems to be very broken.
736 * These formulae increase RTO, when it should be decreased, increase
737 * too slowly, when it should be increased quickly, decrease too quickly
738 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
739 * does not matter how to _calculate_ it. Seems, it was trap
740 * that VJ failed to avoid. 8)
743 m
-= (srtt
>> 3); /* m is now error in rtt est */
744 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
746 m
= -m
; /* m is now abs(error) */
747 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
748 /* This is similar to one of Eifel findings.
749 * Eifel blocks mdev updates when rtt decreases.
750 * This solution is a bit different: we use finer gain
751 * for mdev in this case (alpha*beta).
752 * Like Eifel it also prevents growth of rto,
753 * but also it limits too fast rto decreases,
754 * happening in pure Eifel.
759 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
761 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
762 if (tp
->mdev_us
> tp
->mdev_max_us
) {
763 tp
->mdev_max_us
= tp
->mdev_us
;
764 if (tp
->mdev_max_us
> tp
->rttvar_us
)
765 tp
->rttvar_us
= tp
->mdev_max_us
;
767 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
768 if (tp
->mdev_max_us
< tp
->rttvar_us
)
769 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
770 tp
->rtt_seq
= tp
->snd_nxt
;
771 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
774 /* no previous measure. */
775 srtt
= m
<< 3; /* take the measured time to be rtt */
776 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
777 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
778 tp
->mdev_max_us
= tp
->rttvar_us
;
779 tp
->rtt_seq
= tp
->snd_nxt
;
781 tp
->srtt_us
= max(1U, srtt
);
784 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
785 * Note: TCP stack does not yet implement pacing.
786 * FQ packet scheduler can be used to implement cheap but effective
787 * TCP pacing, to smooth the burst on large writes when packets
788 * in flight is significantly lower than cwnd (or rwin)
790 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
791 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
793 static void tcp_update_pacing_rate(struct sock
*sk
)
795 const struct tcp_sock
*tp
= tcp_sk(sk
);
798 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
799 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
801 /* current rate is (cwnd * mss) / srtt
802 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
803 * In Congestion Avoidance phase, set it to 120 % the current rate.
805 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
806 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
807 * end of slow start and should slow down.
809 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
810 rate
*= sysctl_tcp_pacing_ss_ratio
;
812 rate
*= sysctl_tcp_pacing_ca_ratio
;
814 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
816 if (likely(tp
->srtt_us
))
817 do_div(rate
, tp
->srtt_us
);
819 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
820 * without any lock. We want to make sure compiler wont store
821 * intermediate values in this location.
823 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
824 sk
->sk_max_pacing_rate
);
827 /* Calculate rto without backoff. This is the second half of Van Jacobson's
828 * routine referred to above.
830 static void tcp_set_rto(struct sock
*sk
)
832 const struct tcp_sock
*tp
= tcp_sk(sk
);
833 /* Old crap is replaced with new one. 8)
836 * 1. If rtt variance happened to be less 50msec, it is hallucination.
837 * It cannot be less due to utterly erratic ACK generation made
838 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
839 * to do with delayed acks, because at cwnd>2 true delack timeout
840 * is invisible. Actually, Linux-2.4 also generates erratic
841 * ACKs in some circumstances.
843 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
845 /* 2. Fixups made earlier cannot be right.
846 * If we do not estimate RTO correctly without them,
847 * all the algo is pure shit and should be replaced
848 * with correct one. It is exactly, which we pretend to do.
851 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
852 * guarantees that rto is higher.
857 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
859 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
862 cwnd
= TCP_INIT_CWND
;
863 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
867 * Packet counting of FACK is based on in-order assumptions, therefore TCP
868 * disables it when reordering is detected
870 void tcp_disable_fack(struct tcp_sock
*tp
)
872 /* RFC3517 uses different metric in lost marker => reset on change */
874 tp
->lost_skb_hint
= NULL
;
875 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
878 /* Take a notice that peer is sending D-SACKs */
879 static void tcp_dsack_seen(struct tcp_sock
*tp
)
881 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
884 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
887 struct tcp_sock
*tp
= tcp_sk(sk
);
890 if (WARN_ON_ONCE(metric
< 0))
893 if (metric
> tp
->reordering
) {
894 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
896 #if FASTRETRANS_DEBUG > 1
897 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
898 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
902 tp
->undo_marker
? tp
->undo_retrans
: 0);
904 tcp_disable_fack(tp
);
909 /* This exciting event is worth to be remembered. 8) */
911 mib_idx
= LINUX_MIB_TCPTSREORDER
;
912 else if (tcp_is_reno(tp
))
913 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
914 else if (tcp_is_fack(tp
))
915 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
917 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
919 NET_INC_STATS(sock_net(sk
), mib_idx
);
922 /* This must be called before lost_out is incremented */
923 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
925 if (!tp
->retransmit_skb_hint
||
926 before(TCP_SKB_CB(skb
)->seq
,
927 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
928 tp
->retransmit_skb_hint
= skb
;
931 /* Sum the number of packets on the wire we have marked as lost.
932 * There are two cases we care about here:
933 * a) Packet hasn't been marked lost (nor retransmitted),
934 * and this is the first loss.
935 * b) Packet has been marked both lost and retransmitted,
936 * and this means we think it was lost again.
938 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
940 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
942 if (!(sacked
& TCPCB_LOST
) ||
943 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
944 tp
->lost
+= tcp_skb_pcount(skb
);
947 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
949 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
950 tcp_verify_retransmit_hint(tp
, skb
);
952 tp
->lost_out
+= tcp_skb_pcount(skb
);
953 tcp_sum_lost(tp
, skb
);
954 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
958 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
960 tcp_verify_retransmit_hint(tp
, skb
);
962 tcp_sum_lost(tp
, skb
);
963 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
964 tp
->lost_out
+= tcp_skb_pcount(skb
);
965 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
969 /* This procedure tags the retransmission queue when SACKs arrive.
971 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
972 * Packets in queue with these bits set are counted in variables
973 * sacked_out, retrans_out and lost_out, correspondingly.
975 * Valid combinations are:
976 * Tag InFlight Description
977 * 0 1 - orig segment is in flight.
978 * S 0 - nothing flies, orig reached receiver.
979 * L 0 - nothing flies, orig lost by net.
980 * R 2 - both orig and retransmit are in flight.
981 * L|R 1 - orig is lost, retransmit is in flight.
982 * S|R 1 - orig reached receiver, retrans is still in flight.
983 * (L|S|R is logically valid, it could occur when L|R is sacked,
984 * but it is equivalent to plain S and code short-curcuits it to S.
985 * L|S is logically invalid, it would mean -1 packet in flight 8))
987 * These 6 states form finite state machine, controlled by the following events:
988 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
989 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
990 * 3. Loss detection event of two flavors:
991 * A. Scoreboard estimator decided the packet is lost.
992 * A'. Reno "three dupacks" marks head of queue lost.
993 * A''. Its FACK modification, head until snd.fack is lost.
994 * B. SACK arrives sacking SND.NXT at the moment, when the
995 * segment was retransmitted.
996 * 4. D-SACK added new rule: D-SACK changes any tag to S.
998 * It is pleasant to note, that state diagram turns out to be commutative,
999 * so that we are allowed not to be bothered by order of our actions,
1000 * when multiple events arrive simultaneously. (see the function below).
1002 * Reordering detection.
1003 * --------------------
1004 * Reordering metric is maximal distance, which a packet can be displaced
1005 * in packet stream. With SACKs we can estimate it:
1007 * 1. SACK fills old hole and the corresponding segment was not
1008 * ever retransmitted -> reordering. Alas, we cannot use it
1009 * when segment was retransmitted.
1010 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1011 * for retransmitted and already SACKed segment -> reordering..
1012 * Both of these heuristics are not used in Loss state, when we cannot
1013 * account for retransmits accurately.
1015 * SACK block validation.
1016 * ----------------------
1018 * SACK block range validation checks that the received SACK block fits to
1019 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1020 * Note that SND.UNA is not included to the range though being valid because
1021 * it means that the receiver is rather inconsistent with itself reporting
1022 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1023 * perfectly valid, however, in light of RFC2018 which explicitly states
1024 * that "SACK block MUST reflect the newest segment. Even if the newest
1025 * segment is going to be discarded ...", not that it looks very clever
1026 * in case of head skb. Due to potentional receiver driven attacks, we
1027 * choose to avoid immediate execution of a walk in write queue due to
1028 * reneging and defer head skb's loss recovery to standard loss recovery
1029 * procedure that will eventually trigger (nothing forbids us doing this).
1031 * Implements also blockage to start_seq wrap-around. Problem lies in the
1032 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1033 * there's no guarantee that it will be before snd_nxt (n). The problem
1034 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1037 * <- outs wnd -> <- wrapzone ->
1038 * u e n u_w e_w s n_w
1040 * |<------------+------+----- TCP seqno space --------------+---------->|
1041 * ...-- <2^31 ->| |<--------...
1042 * ...---- >2^31 ------>| |<--------...
1044 * Current code wouldn't be vulnerable but it's better still to discard such
1045 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1046 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1047 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1048 * equal to the ideal case (infinite seqno space without wrap caused issues).
1050 * With D-SACK the lower bound is extended to cover sequence space below
1051 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1052 * again, D-SACK block must not to go across snd_una (for the same reason as
1053 * for the normal SACK blocks, explained above). But there all simplicity
1054 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1055 * fully below undo_marker they do not affect behavior in anyway and can
1056 * therefore be safely ignored. In rare cases (which are more or less
1057 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1058 * fragmentation and packet reordering past skb's retransmission. To consider
1059 * them correctly, the acceptable range must be extended even more though
1060 * the exact amount is rather hard to quantify. However, tp->max_window can
1061 * be used as an exaggerated estimate.
1063 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1064 u32 start_seq
, u32 end_seq
)
1066 /* Too far in future, or reversed (interpretation is ambiguous) */
1067 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1070 /* Nasty start_seq wrap-around check (see comments above) */
1071 if (!before(start_seq
, tp
->snd_nxt
))
1074 /* In outstanding window? ...This is valid exit for D-SACKs too.
1075 * start_seq == snd_una is non-sensical (see comments above)
1077 if (after(start_seq
, tp
->snd_una
))
1080 if (!is_dsack
|| !tp
->undo_marker
)
1083 /* ...Then it's D-SACK, and must reside below snd_una completely */
1084 if (after(end_seq
, tp
->snd_una
))
1087 if (!before(start_seq
, tp
->undo_marker
))
1091 if (!after(end_seq
, tp
->undo_marker
))
1094 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1095 * start_seq < undo_marker and end_seq >= undo_marker.
1097 return !before(start_seq
, end_seq
- tp
->max_window
);
1100 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1101 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1104 struct tcp_sock
*tp
= tcp_sk(sk
);
1105 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1106 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1107 bool dup_sack
= false;
1109 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1112 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1113 } else if (num_sacks
> 1) {
1114 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1115 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1117 if (!after(end_seq_0
, end_seq_1
) &&
1118 !before(start_seq_0
, start_seq_1
)) {
1121 NET_INC_STATS(sock_net(sk
),
1122 LINUX_MIB_TCPDSACKOFORECV
);
1126 /* D-SACK for already forgotten data... Do dumb counting. */
1127 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1128 !after(end_seq_0
, prior_snd_una
) &&
1129 after(end_seq_0
, tp
->undo_marker
))
1135 struct tcp_sacktag_state
{
1138 /* Timestamps for earliest and latest never-retransmitted segment
1139 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1140 * but congestion control should still get an accurate delay signal.
1144 struct rate_sample
*rate
;
1148 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1149 * the incoming SACK may not exactly match but we can find smaller MSS
1150 * aligned portion of it that matches. Therefore we might need to fragment
1151 * which may fail and creates some hassle (caller must handle error case
1154 * FIXME: this could be merged to shift decision code
1156 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1157 u32 start_seq
, u32 end_seq
)
1161 unsigned int pkt_len
;
1164 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1165 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1167 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1168 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1169 mss
= tcp_skb_mss(skb
);
1170 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1173 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1177 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1182 /* Round if necessary so that SACKs cover only full MSSes
1183 * and/or the remaining small portion (if present)
1185 if (pkt_len
> mss
) {
1186 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1187 if (!in_sack
&& new_len
< pkt_len
)
1192 if (pkt_len
>= skb
->len
&& !in_sack
)
1195 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1203 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1204 static u8
tcp_sacktag_one(struct sock
*sk
,
1205 struct tcp_sacktag_state
*state
, u8 sacked
,
1206 u32 start_seq
, u32 end_seq
,
1207 int dup_sack
, int pcount
,
1210 struct tcp_sock
*tp
= tcp_sk(sk
);
1211 int fack_count
= state
->fack_count
;
1213 /* Account D-SACK for retransmitted packet. */
1214 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1215 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1216 after(end_seq
, tp
->undo_marker
))
1218 if (sacked
& TCPCB_SACKED_ACKED
)
1219 state
->reord
= min(fack_count
, state
->reord
);
1222 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1223 if (!after(end_seq
, tp
->snd_una
))
1226 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1227 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1229 if (sacked
& TCPCB_SACKED_RETRANS
) {
1230 /* If the segment is not tagged as lost,
1231 * we do not clear RETRANS, believing
1232 * that retransmission is still in flight.
1234 if (sacked
& TCPCB_LOST
) {
1235 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1236 tp
->lost_out
-= pcount
;
1237 tp
->retrans_out
-= pcount
;
1240 if (!(sacked
& TCPCB_RETRANS
)) {
1241 /* New sack for not retransmitted frame,
1242 * which was in hole. It is reordering.
1244 if (before(start_seq
,
1245 tcp_highest_sack_seq(tp
)))
1246 state
->reord
= min(fack_count
,
1248 if (!after(end_seq
, tp
->high_seq
))
1249 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1250 if (state
->first_sackt
== 0)
1251 state
->first_sackt
= xmit_time
;
1252 state
->last_sackt
= xmit_time
;
1255 if (sacked
& TCPCB_LOST
) {
1256 sacked
&= ~TCPCB_LOST
;
1257 tp
->lost_out
-= pcount
;
1261 sacked
|= TCPCB_SACKED_ACKED
;
1262 state
->flag
|= FLAG_DATA_SACKED
;
1263 tp
->sacked_out
+= pcount
;
1264 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1266 fack_count
+= pcount
;
1268 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1269 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1270 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1271 tp
->lost_cnt_hint
+= pcount
;
1273 if (fack_count
> tp
->fackets_out
)
1274 tp
->fackets_out
= fack_count
;
1277 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1278 * frames and clear it. undo_retrans is decreased above, L|R frames
1279 * are accounted above as well.
1281 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1282 sacked
&= ~TCPCB_SACKED_RETRANS
;
1283 tp
->retrans_out
-= pcount
;
1289 /* Shift newly-SACKed bytes from this skb to the immediately previous
1290 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1292 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1293 struct tcp_sacktag_state
*state
,
1294 unsigned int pcount
, int shifted
, int mss
,
1297 struct tcp_sock
*tp
= tcp_sk(sk
);
1298 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1299 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1300 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1304 /* Adjust counters and hints for the newly sacked sequence
1305 * range but discard the return value since prev is already
1306 * marked. We must tag the range first because the seq
1307 * advancement below implicitly advances
1308 * tcp_highest_sack_seq() when skb is highest_sack.
1310 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1311 start_seq
, end_seq
, dup_sack
, pcount
,
1313 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1315 if (skb
== tp
->lost_skb_hint
)
1316 tp
->lost_cnt_hint
+= pcount
;
1318 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1319 TCP_SKB_CB(skb
)->seq
+= shifted
;
1321 tcp_skb_pcount_add(prev
, pcount
);
1322 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1323 tcp_skb_pcount_add(skb
, -pcount
);
1325 /* When we're adding to gso_segs == 1, gso_size will be zero,
1326 * in theory this shouldn't be necessary but as long as DSACK
1327 * code can come after this skb later on it's better to keep
1328 * setting gso_size to something.
1330 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1331 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1333 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1334 if (tcp_skb_pcount(skb
) <= 1)
1335 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1337 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1338 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1341 BUG_ON(!tcp_skb_pcount(skb
));
1342 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1346 /* Whole SKB was eaten :-) */
1348 if (skb
== tp
->retransmit_skb_hint
)
1349 tp
->retransmit_skb_hint
= prev
;
1350 if (skb
== tp
->lost_skb_hint
) {
1351 tp
->lost_skb_hint
= prev
;
1352 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1355 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1356 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1357 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1358 TCP_SKB_CB(prev
)->end_seq
++;
1360 if (skb
== tcp_highest_sack(sk
))
1361 tcp_advance_highest_sack(sk
, skb
);
1363 tcp_skb_collapse_tstamp(prev
, skb
);
1364 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1365 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1367 tcp_unlink_write_queue(skb
, sk
);
1368 sk_wmem_free_skb(sk
, skb
);
1370 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1375 /* I wish gso_size would have a bit more sane initialization than
1376 * something-or-zero which complicates things
1378 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1380 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1383 /* Shifting pages past head area doesn't work */
1384 static int skb_can_shift(const struct sk_buff
*skb
)
1386 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1389 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1392 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1393 struct tcp_sacktag_state
*state
,
1394 u32 start_seq
, u32 end_seq
,
1397 struct tcp_sock
*tp
= tcp_sk(sk
);
1398 struct sk_buff
*prev
;
1404 if (!sk_can_gso(sk
))
1407 /* Normally R but no L won't result in plain S */
1409 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1411 if (!skb_can_shift(skb
))
1413 /* This frame is about to be dropped (was ACKed). */
1414 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1417 /* Can only happen with delayed DSACK + discard craziness */
1418 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1420 prev
= tcp_write_queue_prev(sk
, skb
);
1422 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1425 if (!tcp_skb_can_collapse_to(prev
))
1428 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1429 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1433 pcount
= tcp_skb_pcount(skb
);
1434 mss
= tcp_skb_seglen(skb
);
1436 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1437 * drop this restriction as unnecessary
1439 if (mss
!= tcp_skb_seglen(prev
))
1442 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1444 /* CHECKME: This is non-MSS split case only?, this will
1445 * cause skipped skbs due to advancing loop btw, original
1446 * has that feature too
1448 if (tcp_skb_pcount(skb
) <= 1)
1451 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1453 /* TODO: head merge to next could be attempted here
1454 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1455 * though it might not be worth of the additional hassle
1457 * ...we can probably just fallback to what was done
1458 * previously. We could try merging non-SACKed ones
1459 * as well but it probably isn't going to buy off
1460 * because later SACKs might again split them, and
1461 * it would make skb timestamp tracking considerably
1467 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1469 BUG_ON(len
> skb
->len
);
1471 /* MSS boundaries should be honoured or else pcount will
1472 * severely break even though it makes things bit trickier.
1473 * Optimize common case to avoid most of the divides
1475 mss
= tcp_skb_mss(skb
);
1477 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1478 * drop this restriction as unnecessary
1480 if (mss
!= tcp_skb_seglen(prev
))
1485 } else if (len
< mss
) {
1493 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1494 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1497 if (!skb_shift(prev
, skb
, len
))
1499 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1502 /* Hole filled allows collapsing with the next as well, this is very
1503 * useful when hole on every nth skb pattern happens
1505 if (prev
== tcp_write_queue_tail(sk
))
1507 skb
= tcp_write_queue_next(sk
, prev
);
1509 if (!skb_can_shift(skb
) ||
1510 (skb
== tcp_send_head(sk
)) ||
1511 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1512 (mss
!= tcp_skb_seglen(skb
)))
1516 if (skb_shift(prev
, skb
, len
)) {
1517 pcount
+= tcp_skb_pcount(skb
);
1518 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1522 state
->fack_count
+= pcount
;
1529 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1533 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1534 struct tcp_sack_block
*next_dup
,
1535 struct tcp_sacktag_state
*state
,
1536 u32 start_seq
, u32 end_seq
,
1539 struct tcp_sock
*tp
= tcp_sk(sk
);
1540 struct sk_buff
*tmp
;
1542 tcp_for_write_queue_from(skb
, sk
) {
1544 bool dup_sack
= dup_sack_in
;
1546 if (skb
== tcp_send_head(sk
))
1549 /* queue is in-order => we can short-circuit the walk early */
1550 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1554 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1555 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1556 next_dup
->start_seq
,
1562 /* skb reference here is a bit tricky to get right, since
1563 * shifting can eat and free both this skb and the next,
1564 * so not even _safe variant of the loop is enough.
1567 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1568 start_seq
, end_seq
, dup_sack
);
1577 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1583 if (unlikely(in_sack
< 0))
1587 TCP_SKB_CB(skb
)->sacked
=
1590 TCP_SKB_CB(skb
)->sacked
,
1591 TCP_SKB_CB(skb
)->seq
,
1592 TCP_SKB_CB(skb
)->end_seq
,
1594 tcp_skb_pcount(skb
),
1596 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1598 if (!before(TCP_SKB_CB(skb
)->seq
,
1599 tcp_highest_sack_seq(tp
)))
1600 tcp_advance_highest_sack(sk
, skb
);
1603 state
->fack_count
+= tcp_skb_pcount(skb
);
1608 /* Avoid all extra work that is being done by sacktag while walking in
1611 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1612 struct tcp_sacktag_state
*state
,
1615 tcp_for_write_queue_from(skb
, sk
) {
1616 if (skb
== tcp_send_head(sk
))
1619 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1622 state
->fack_count
+= tcp_skb_pcount(skb
);
1627 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1629 struct tcp_sack_block
*next_dup
,
1630 struct tcp_sacktag_state
*state
,
1636 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1637 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1638 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1639 next_dup
->start_seq
, next_dup
->end_seq
,
1646 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1648 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1652 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1653 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1655 struct tcp_sock
*tp
= tcp_sk(sk
);
1656 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1657 TCP_SKB_CB(ack_skb
)->sacked
);
1658 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1659 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1660 struct tcp_sack_block
*cache
;
1661 struct sk_buff
*skb
;
1662 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1664 bool found_dup_sack
= false;
1666 int first_sack_index
;
1669 state
->reord
= tp
->packets_out
;
1671 if (!tp
->sacked_out
) {
1672 if (WARN_ON(tp
->fackets_out
))
1673 tp
->fackets_out
= 0;
1674 tcp_highest_sack_reset(sk
);
1677 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1678 num_sacks
, prior_snd_una
);
1679 if (found_dup_sack
) {
1680 state
->flag
|= FLAG_DSACKING_ACK
;
1681 tp
->delivered
++; /* A spurious retransmission is delivered */
1684 /* Eliminate too old ACKs, but take into
1685 * account more or less fresh ones, they can
1686 * contain valid SACK info.
1688 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1691 if (!tp
->packets_out
)
1695 first_sack_index
= 0;
1696 for (i
= 0; i
< num_sacks
; i
++) {
1697 bool dup_sack
= !i
&& found_dup_sack
;
1699 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1700 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1702 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1703 sp
[used_sacks
].start_seq
,
1704 sp
[used_sacks
].end_seq
)) {
1708 if (!tp
->undo_marker
)
1709 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1711 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1713 /* Don't count olds caused by ACK reordering */
1714 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1715 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1717 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1720 NET_INC_STATS(sock_net(sk
), mib_idx
);
1722 first_sack_index
= -1;
1726 /* Ignore very old stuff early */
1727 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1733 /* order SACK blocks to allow in order walk of the retrans queue */
1734 for (i
= used_sacks
- 1; i
> 0; i
--) {
1735 for (j
= 0; j
< i
; j
++) {
1736 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1737 swap(sp
[j
], sp
[j
+ 1]);
1739 /* Track where the first SACK block goes to */
1740 if (j
== first_sack_index
)
1741 first_sack_index
= j
+ 1;
1746 skb
= tcp_write_queue_head(sk
);
1747 state
->fack_count
= 0;
1750 if (!tp
->sacked_out
) {
1751 /* It's already past, so skip checking against it */
1752 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1754 cache
= tp
->recv_sack_cache
;
1755 /* Skip empty blocks in at head of the cache */
1756 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1761 while (i
< used_sacks
) {
1762 u32 start_seq
= sp
[i
].start_seq
;
1763 u32 end_seq
= sp
[i
].end_seq
;
1764 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1765 struct tcp_sack_block
*next_dup
= NULL
;
1767 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1768 next_dup
= &sp
[i
+ 1];
1770 /* Skip too early cached blocks */
1771 while (tcp_sack_cache_ok(tp
, cache
) &&
1772 !before(start_seq
, cache
->end_seq
))
1775 /* Can skip some work by looking recv_sack_cache? */
1776 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1777 after(end_seq
, cache
->start_seq
)) {
1780 if (before(start_seq
, cache
->start_seq
)) {
1781 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1783 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1790 /* Rest of the block already fully processed? */
1791 if (!after(end_seq
, cache
->end_seq
))
1794 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1798 /* ...tail remains todo... */
1799 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1800 /* ...but better entrypoint exists! */
1801 skb
= tcp_highest_sack(sk
);
1804 state
->fack_count
= tp
->fackets_out
;
1809 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1810 /* Check overlap against next cached too (past this one already) */
1815 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1816 skb
= tcp_highest_sack(sk
);
1819 state
->fack_count
= tp
->fackets_out
;
1821 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1824 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1825 start_seq
, end_seq
, dup_sack
);
1831 /* Clear the head of the cache sack blocks so we can skip it next time */
1832 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1833 tp
->recv_sack_cache
[i
].start_seq
= 0;
1834 tp
->recv_sack_cache
[i
].end_seq
= 0;
1836 for (j
= 0; j
< used_sacks
; j
++)
1837 tp
->recv_sack_cache
[i
++] = sp
[j
];
1839 if ((state
->reord
< tp
->fackets_out
) &&
1840 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1841 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1843 tcp_verify_left_out(tp
);
1846 #if FASTRETRANS_DEBUG > 0
1847 WARN_ON((int)tp
->sacked_out
< 0);
1848 WARN_ON((int)tp
->lost_out
< 0);
1849 WARN_ON((int)tp
->retrans_out
< 0);
1850 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1855 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1856 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1858 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1862 holes
= max(tp
->lost_out
, 1U);
1863 holes
= min(holes
, tp
->packets_out
);
1865 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1866 tp
->sacked_out
= tp
->packets_out
- holes
;
1872 /* If we receive more dupacks than we expected counting segments
1873 * in assumption of absent reordering, interpret this as reordering.
1874 * The only another reason could be bug in receiver TCP.
1876 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1878 struct tcp_sock
*tp
= tcp_sk(sk
);
1879 if (tcp_limit_reno_sacked(tp
))
1880 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1883 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1885 static void tcp_add_reno_sack(struct sock
*sk
)
1887 struct tcp_sock
*tp
= tcp_sk(sk
);
1888 u32 prior_sacked
= tp
->sacked_out
;
1891 tcp_check_reno_reordering(sk
, 0);
1892 if (tp
->sacked_out
> prior_sacked
)
1893 tp
->delivered
++; /* Some out-of-order packet is delivered */
1894 tcp_verify_left_out(tp
);
1897 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1899 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1901 struct tcp_sock
*tp
= tcp_sk(sk
);
1904 /* One ACK acked hole. The rest eat duplicate ACKs. */
1905 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1906 if (acked
- 1 >= tp
->sacked_out
)
1909 tp
->sacked_out
-= acked
- 1;
1911 tcp_check_reno_reordering(sk
, acked
);
1912 tcp_verify_left_out(tp
);
1915 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1920 void tcp_clear_retrans(struct tcp_sock
*tp
)
1922 tp
->retrans_out
= 0;
1924 tp
->undo_marker
= 0;
1925 tp
->undo_retrans
= -1;
1926 tp
->fackets_out
= 0;
1930 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1932 tp
->undo_marker
= tp
->snd_una
;
1933 /* Retransmission still in flight may cause DSACKs later. */
1934 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1937 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1938 * and reset tags completely, otherwise preserve SACKs. If receiver
1939 * dropped its ofo queue, we will know this due to reneging detection.
1941 void tcp_enter_loss(struct sock
*sk
)
1943 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1944 struct tcp_sock
*tp
= tcp_sk(sk
);
1945 struct net
*net
= sock_net(sk
);
1946 struct sk_buff
*skb
;
1947 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1948 bool is_reneg
; /* is receiver reneging on SACKs? */
1951 /* Reduce ssthresh if it has not yet been made inside this window. */
1952 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1953 !after(tp
->high_seq
, tp
->snd_una
) ||
1954 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1955 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1956 tp
->prior_cwnd
= tp
->snd_cwnd
;
1957 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1958 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1962 tp
->snd_cwnd_cnt
= 0;
1963 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1965 tp
->retrans_out
= 0;
1968 if (tcp_is_reno(tp
))
1969 tcp_reset_reno_sack(tp
);
1971 skb
= tcp_write_queue_head(sk
);
1972 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1974 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1976 tp
->fackets_out
= 0;
1978 tcp_clear_all_retrans_hints(tp
);
1980 tcp_for_write_queue(skb
, sk
) {
1981 if (skb
== tcp_send_head(sk
))
1984 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1987 tcp_sum_lost(tp
, skb
);
1988 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1990 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1991 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1992 tp
->lost_out
+= tcp_skb_pcount(skb
);
1995 tcp_verify_left_out(tp
);
1997 /* Timeout in disordered state after receiving substantial DUPACKs
1998 * suggests that the degree of reordering is over-estimated.
2000 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
2001 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
2002 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2003 net
->ipv4
.sysctl_tcp_reordering
);
2004 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2005 tp
->high_seq
= tp
->snd_nxt
;
2006 tcp_ecn_queue_cwr(tp
);
2008 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2009 * loss recovery is underway except recurring timeout(s) on
2010 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2012 * In theory F-RTO can be used repeatedly during loss recovery.
2013 * In practice this interacts badly with broken middle-boxes that
2014 * falsely raise the receive window, which results in repeated
2015 * timeouts and stop-and-go behavior.
2017 tp
->frto
= sysctl_tcp_frto
&&
2018 (new_recovery
|| icsk
->icsk_retransmits
) &&
2019 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2022 /* If ACK arrived pointing to a remembered SACK, it means that our
2023 * remembered SACKs do not reflect real state of receiver i.e.
2024 * receiver _host_ is heavily congested (or buggy).
2026 * To avoid big spurious retransmission bursts due to transient SACK
2027 * scoreboard oddities that look like reneging, we give the receiver a
2028 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2029 * restore sanity to the SACK scoreboard. If the apparent reneging
2030 * persists until this RTO then we'll clear the SACK scoreboard.
2032 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2034 if (flag
& FLAG_SACK_RENEGING
) {
2035 struct tcp_sock
*tp
= tcp_sk(sk
);
2036 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2037 msecs_to_jiffies(10));
2039 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2040 delay
, TCP_RTO_MAX
);
2046 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2048 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2051 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2052 * counter when SACK is enabled (without SACK, sacked_out is used for
2055 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2056 * segments up to the highest received SACK block so far and holes in
2059 * With reordering, holes may still be in flight, so RFC3517 recovery
2060 * uses pure sacked_out (total number of SACKed segments) even though
2061 * it violates the RFC that uses duplicate ACKs, often these are equal
2062 * but when e.g. out-of-window ACKs or packet duplication occurs,
2063 * they differ. Since neither occurs due to loss, TCP should really
2066 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2068 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2071 /* Linux NewReno/SACK/FACK/ECN state machine.
2072 * --------------------------------------
2074 * "Open" Normal state, no dubious events, fast path.
2075 * "Disorder" In all the respects it is "Open",
2076 * but requires a bit more attention. It is entered when
2077 * we see some SACKs or dupacks. It is split of "Open"
2078 * mainly to move some processing from fast path to slow one.
2079 * "CWR" CWND was reduced due to some Congestion Notification event.
2080 * It can be ECN, ICMP source quench, local device congestion.
2081 * "Recovery" CWND was reduced, we are fast-retransmitting.
2082 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2084 * tcp_fastretrans_alert() is entered:
2085 * - each incoming ACK, if state is not "Open"
2086 * - when arrived ACK is unusual, namely:
2091 * Counting packets in flight is pretty simple.
2093 * in_flight = packets_out - left_out + retrans_out
2095 * packets_out is SND.NXT-SND.UNA counted in packets.
2097 * retrans_out is number of retransmitted segments.
2099 * left_out is number of segments left network, but not ACKed yet.
2101 * left_out = sacked_out + lost_out
2103 * sacked_out: Packets, which arrived to receiver out of order
2104 * and hence not ACKed. With SACKs this number is simply
2105 * amount of SACKed data. Even without SACKs
2106 * it is easy to give pretty reliable estimate of this number,
2107 * counting duplicate ACKs.
2109 * lost_out: Packets lost by network. TCP has no explicit
2110 * "loss notification" feedback from network (for now).
2111 * It means that this number can be only _guessed_.
2112 * Actually, it is the heuristics to predict lossage that
2113 * distinguishes different algorithms.
2115 * F.e. after RTO, when all the queue is considered as lost,
2116 * lost_out = packets_out and in_flight = retrans_out.
2118 * Essentially, we have now a few algorithms detecting
2121 * If the receiver supports SACK:
2123 * RFC6675/3517: It is the conventional algorithm. A packet is
2124 * considered lost if the number of higher sequence packets
2125 * SACKed is greater than or equal the DUPACK thoreshold
2126 * (reordering). This is implemented in tcp_mark_head_lost and
2127 * tcp_update_scoreboard.
2129 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2130 * (2017-) that checks timing instead of counting DUPACKs.
2131 * Essentially a packet is considered lost if it's not S/ACKed
2132 * after RTT + reordering_window, where both metrics are
2133 * dynamically measured and adjusted. This is implemented in
2134 * tcp_rack_mark_lost.
2136 * FACK (Disabled by default. Subsumbed by RACK):
2137 * It is the simplest heuristics. As soon as we decided
2138 * that something is lost, we decide that _all_ not SACKed
2139 * packets until the most forward SACK are lost. I.e.
2140 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2141 * It is absolutely correct estimate, if network does not reorder
2142 * packets. And it loses any connection to reality when reordering
2143 * takes place. We use FACK by default until reordering
2144 * is suspected on the path to this destination.
2146 * If the receiver does not support SACK:
2148 * NewReno (RFC6582): in Recovery we assume that one segment
2149 * is lost (classic Reno). While we are in Recovery and
2150 * a partial ACK arrives, we assume that one more packet
2151 * is lost (NewReno). This heuristics are the same in NewReno
2154 * Really tricky (and requiring careful tuning) part of algorithm
2155 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2156 * The first determines the moment _when_ we should reduce CWND and,
2157 * hence, slow down forward transmission. In fact, it determines the moment
2158 * when we decide that hole is caused by loss, rather than by a reorder.
2160 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2161 * holes, caused by lost packets.
2163 * And the most logically complicated part of algorithm is undo
2164 * heuristics. We detect false retransmits due to both too early
2165 * fast retransmit (reordering) and underestimated RTO, analyzing
2166 * timestamps and D-SACKs. When we detect that some segments were
2167 * retransmitted by mistake and CWND reduction was wrong, we undo
2168 * window reduction and abort recovery phase. This logic is hidden
2169 * inside several functions named tcp_try_undo_<something>.
2172 /* This function decides, when we should leave Disordered state
2173 * and enter Recovery phase, reducing congestion window.
2175 * Main question: may we further continue forward transmission
2176 * with the same cwnd?
2178 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2180 struct tcp_sock
*tp
= tcp_sk(sk
);
2182 /* Trick#1: The loss is proven. */
2186 /* Not-A-Trick#2 : Classic rule... */
2187 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2193 /* Detect loss in event "A" above by marking head of queue up as lost.
2194 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2195 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2196 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2197 * the maximum SACKed segments to pass before reaching this limit.
2199 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2201 struct tcp_sock
*tp
= tcp_sk(sk
);
2202 struct sk_buff
*skb
;
2203 int cnt
, oldcnt
, lost
;
2205 /* Use SACK to deduce losses of new sequences sent during recovery */
2206 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2208 WARN_ON(packets
> tp
->packets_out
);
2209 if (tp
->lost_skb_hint
) {
2210 skb
= tp
->lost_skb_hint
;
2211 cnt
= tp
->lost_cnt_hint
;
2212 /* Head already handled? */
2213 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2216 skb
= tcp_write_queue_head(sk
);
2220 tcp_for_write_queue_from(skb
, sk
) {
2221 if (skb
== tcp_send_head(sk
))
2223 /* TODO: do this better */
2224 /* this is not the most efficient way to do this... */
2225 tp
->lost_skb_hint
= skb
;
2226 tp
->lost_cnt_hint
= cnt
;
2228 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2232 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2233 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2234 cnt
+= tcp_skb_pcount(skb
);
2236 if (cnt
> packets
) {
2237 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2238 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2239 (oldcnt
>= packets
))
2242 mss
= tcp_skb_mss(skb
);
2243 /* If needed, chop off the prefix to mark as lost. */
2244 lost
= (packets
- oldcnt
) * mss
;
2245 if (lost
< skb
->len
&&
2246 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2251 tcp_skb_mark_lost(tp
, skb
);
2256 tcp_verify_left_out(tp
);
2259 /* Account newly detected lost packet(s) */
2261 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2263 struct tcp_sock
*tp
= tcp_sk(sk
);
2265 if (tcp_is_reno(tp
)) {
2266 tcp_mark_head_lost(sk
, 1, 1);
2267 } else if (tcp_is_fack(tp
)) {
2268 int lost
= tp
->fackets_out
- tp
->reordering
;
2271 tcp_mark_head_lost(sk
, lost
, 0);
2273 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2274 if (sacked_upto
>= 0)
2275 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2276 else if (fast_rexmit
)
2277 tcp_mark_head_lost(sk
, 1, 1);
2281 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2283 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2284 before(tp
->rx_opt
.rcv_tsecr
, when
);
2287 /* skb is spurious retransmitted if the returned timestamp echo
2288 * reply is prior to the skb transmission time
2290 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2291 const struct sk_buff
*skb
)
2293 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2294 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2297 /* Nothing was retransmitted or returned timestamp is less
2298 * than timestamp of the first retransmission.
2300 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2302 return !tp
->retrans_stamp
||
2303 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2306 /* Undo procedures. */
2308 /* We can clear retrans_stamp when there are no retransmissions in the
2309 * window. It would seem that it is trivially available for us in
2310 * tp->retrans_out, however, that kind of assumptions doesn't consider
2311 * what will happen if errors occur when sending retransmission for the
2312 * second time. ...It could the that such segment has only
2313 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2314 * the head skb is enough except for some reneging corner cases that
2315 * are not worth the effort.
2317 * Main reason for all this complexity is the fact that connection dying
2318 * time now depends on the validity of the retrans_stamp, in particular,
2319 * that successive retransmissions of a segment must not advance
2320 * retrans_stamp under any conditions.
2322 static bool tcp_any_retrans_done(const struct sock
*sk
)
2324 const struct tcp_sock
*tp
= tcp_sk(sk
);
2325 struct sk_buff
*skb
;
2327 if (tp
->retrans_out
)
2330 skb
= tcp_write_queue_head(sk
);
2331 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2337 #if FASTRETRANS_DEBUG > 1
2338 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2340 struct tcp_sock
*tp
= tcp_sk(sk
);
2341 struct inet_sock
*inet
= inet_sk(sk
);
2343 if (sk
->sk_family
== AF_INET
) {
2344 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2346 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2347 tp
->snd_cwnd
, tcp_left_out(tp
),
2348 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2351 #if IS_ENABLED(CONFIG_IPV6)
2352 else if (sk
->sk_family
== AF_INET6
) {
2353 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2355 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2356 tp
->snd_cwnd
, tcp_left_out(tp
),
2357 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2363 #define DBGUNDO(x...) do { } while (0)
2366 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2368 struct tcp_sock
*tp
= tcp_sk(sk
);
2371 struct sk_buff
*skb
;
2373 tcp_for_write_queue(skb
, sk
) {
2374 if (skb
== tcp_send_head(sk
))
2376 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2379 tcp_clear_all_retrans_hints(tp
);
2382 if (tp
->prior_ssthresh
) {
2383 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2385 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2387 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2388 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2389 tcp_ecn_withdraw_cwr(tp
);
2392 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2393 tp
->undo_marker
= 0;
2396 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2398 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2401 /* People celebrate: "We love our President!" */
2402 static bool tcp_try_undo_recovery(struct sock
*sk
)
2404 struct tcp_sock
*tp
= tcp_sk(sk
);
2406 if (tcp_may_undo(tp
)) {
2409 /* Happy end! We did not retransmit anything
2410 * or our original transmission succeeded.
2412 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2413 tcp_undo_cwnd_reduction(sk
, false);
2414 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2415 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2417 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2419 NET_INC_STATS(sock_net(sk
), mib_idx
);
2421 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2422 /* Hold old state until something *above* high_seq
2423 * is ACKed. For Reno it is MUST to prevent false
2424 * fast retransmits (RFC2582). SACK TCP is safe. */
2425 if (!tcp_any_retrans_done(sk
))
2426 tp
->retrans_stamp
= 0;
2429 tcp_set_ca_state(sk
, TCP_CA_Open
);
2433 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2434 static bool tcp_try_undo_dsack(struct sock
*sk
)
2436 struct tcp_sock
*tp
= tcp_sk(sk
);
2438 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2439 DBGUNDO(sk
, "D-SACK");
2440 tcp_undo_cwnd_reduction(sk
, false);
2441 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2447 /* Undo during loss recovery after partial ACK or using F-RTO. */
2448 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2450 struct tcp_sock
*tp
= tcp_sk(sk
);
2452 if (frto_undo
|| tcp_may_undo(tp
)) {
2453 tcp_undo_cwnd_reduction(sk
, true);
2455 DBGUNDO(sk
, "partial loss");
2456 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2458 NET_INC_STATS(sock_net(sk
),
2459 LINUX_MIB_TCPSPURIOUSRTOS
);
2460 inet_csk(sk
)->icsk_retransmits
= 0;
2461 if (frto_undo
|| tcp_is_sack(tp
))
2462 tcp_set_ca_state(sk
, TCP_CA_Open
);
2468 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2469 * It computes the number of packets to send (sndcnt) based on packets newly
2471 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2472 * cwnd reductions across a full RTT.
2473 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2474 * But when the retransmits are acked without further losses, PRR
2475 * slow starts cwnd up to ssthresh to speed up the recovery.
2477 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2479 struct tcp_sock
*tp
= tcp_sk(sk
);
2481 tp
->high_seq
= tp
->snd_nxt
;
2482 tp
->tlp_high_seq
= 0;
2483 tp
->snd_cwnd_cnt
= 0;
2484 tp
->prior_cwnd
= tp
->snd_cwnd
;
2485 tp
->prr_delivered
= 0;
2487 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2488 tcp_ecn_queue_cwr(tp
);
2491 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2493 struct tcp_sock
*tp
= tcp_sk(sk
);
2495 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2497 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2500 tp
->prr_delivered
+= newly_acked_sacked
;
2502 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2504 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2505 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2506 !(flag
& FLAG_LOST_RETRANS
)) {
2507 sndcnt
= min_t(int, delta
,
2508 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2509 newly_acked_sacked
) + 1);
2511 sndcnt
= min(delta
, newly_acked_sacked
);
2513 /* Force a fast retransmit upon entering fast recovery */
2514 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2515 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2518 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2520 struct tcp_sock
*tp
= tcp_sk(sk
);
2522 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2525 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2526 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2527 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2528 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2529 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2531 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2534 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2535 void tcp_enter_cwr(struct sock
*sk
)
2537 struct tcp_sock
*tp
= tcp_sk(sk
);
2539 tp
->prior_ssthresh
= 0;
2540 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2541 tp
->undo_marker
= 0;
2542 tcp_init_cwnd_reduction(sk
);
2543 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2546 EXPORT_SYMBOL(tcp_enter_cwr
);
2548 static void tcp_try_keep_open(struct sock
*sk
)
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 int state
= TCP_CA_Open
;
2553 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2554 state
= TCP_CA_Disorder
;
2556 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2557 tcp_set_ca_state(sk
, state
);
2558 tp
->high_seq
= tp
->snd_nxt
;
2562 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2564 struct tcp_sock
*tp
= tcp_sk(sk
);
2566 tcp_verify_left_out(tp
);
2568 if (!tcp_any_retrans_done(sk
))
2569 tp
->retrans_stamp
= 0;
2571 if (flag
& FLAG_ECE
)
2574 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2575 tcp_try_keep_open(sk
);
2579 static void tcp_mtup_probe_failed(struct sock
*sk
)
2581 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2583 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2584 icsk
->icsk_mtup
.probe_size
= 0;
2585 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2588 static void tcp_mtup_probe_success(struct sock
*sk
)
2590 struct tcp_sock
*tp
= tcp_sk(sk
);
2591 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2593 /* FIXME: breaks with very large cwnd */
2594 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2595 tp
->snd_cwnd
= tp
->snd_cwnd
*
2596 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2597 icsk
->icsk_mtup
.probe_size
;
2598 tp
->snd_cwnd_cnt
= 0;
2599 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2600 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2602 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2603 icsk
->icsk_mtup
.probe_size
= 0;
2604 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2605 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2608 /* Do a simple retransmit without using the backoff mechanisms in
2609 * tcp_timer. This is used for path mtu discovery.
2610 * The socket is already locked here.
2612 void tcp_simple_retransmit(struct sock
*sk
)
2614 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2615 struct tcp_sock
*tp
= tcp_sk(sk
);
2616 struct sk_buff
*skb
;
2617 unsigned int mss
= tcp_current_mss(sk
);
2618 u32 prior_lost
= tp
->lost_out
;
2620 tcp_for_write_queue(skb
, sk
) {
2621 if (skb
== tcp_send_head(sk
))
2623 if (tcp_skb_seglen(skb
) > mss
&&
2624 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2625 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2626 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2627 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2629 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2633 tcp_clear_retrans_hints_partial(tp
);
2635 if (prior_lost
== tp
->lost_out
)
2638 if (tcp_is_reno(tp
))
2639 tcp_limit_reno_sacked(tp
);
2641 tcp_verify_left_out(tp
);
2643 /* Don't muck with the congestion window here.
2644 * Reason is that we do not increase amount of _data_
2645 * in network, but units changed and effective
2646 * cwnd/ssthresh really reduced now.
2648 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2649 tp
->high_seq
= tp
->snd_nxt
;
2650 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2651 tp
->prior_ssthresh
= 0;
2652 tp
->undo_marker
= 0;
2653 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2655 tcp_xmit_retransmit_queue(sk
);
2657 EXPORT_SYMBOL(tcp_simple_retransmit
);
2659 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2661 struct tcp_sock
*tp
= tcp_sk(sk
);
2664 if (tcp_is_reno(tp
))
2665 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2667 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2669 NET_INC_STATS(sock_net(sk
), mib_idx
);
2671 tp
->prior_ssthresh
= 0;
2674 if (!tcp_in_cwnd_reduction(sk
)) {
2676 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2677 tcp_init_cwnd_reduction(sk
);
2679 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2682 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2683 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2685 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2688 struct tcp_sock
*tp
= tcp_sk(sk
);
2689 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2691 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2692 tcp_try_undo_loss(sk
, false))
2695 /* The ACK (s)acks some never-retransmitted data meaning not all
2696 * the data packets before the timeout were lost. Therefore we
2697 * undo the congestion window and state. This is essentially
2698 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2699 * a retransmitted skb is permantly marked, we can apply such an
2700 * operation even if F-RTO was not used.
2702 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2703 tcp_try_undo_loss(sk
, tp
->undo_marker
))
2706 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2707 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2708 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2709 tp
->frto
= 0; /* Step 3.a. loss was real */
2710 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2711 tp
->high_seq
= tp
->snd_nxt
;
2712 /* Step 2.b. Try send new data (but deferred until cwnd
2713 * is updated in tcp_ack()). Otherwise fall back to
2714 * the conventional recovery.
2716 if (tcp_send_head(sk
) &&
2717 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2718 *rexmit
= REXMIT_NEW
;
2726 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2727 tcp_try_undo_recovery(sk
);
2730 if (tcp_is_reno(tp
)) {
2731 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2732 * delivered. Lower inflight to clock out (re)tranmissions.
2734 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2735 tcp_add_reno_sack(sk
);
2736 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2737 tcp_reset_reno_sack(tp
);
2739 *rexmit
= REXMIT_LOST
;
2742 /* Undo during fast recovery after partial ACK. */
2743 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2745 struct tcp_sock
*tp
= tcp_sk(sk
);
2747 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2748 /* Plain luck! Hole if filled with delayed
2749 * packet, rather than with a retransmit.
2751 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2753 /* We are getting evidence that the reordering degree is higher
2754 * than we realized. If there are no retransmits out then we
2755 * can undo. Otherwise we clock out new packets but do not
2756 * mark more packets lost or retransmit more.
2758 if (tp
->retrans_out
)
2761 if (!tcp_any_retrans_done(sk
))
2762 tp
->retrans_stamp
= 0;
2764 DBGUNDO(sk
, "partial recovery");
2765 tcp_undo_cwnd_reduction(sk
, true);
2766 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2767 tcp_try_keep_open(sk
);
2773 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
)
2775 struct tcp_sock
*tp
= tcp_sk(sk
);
2777 /* Use RACK to detect loss */
2778 if (sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2779 u32 prior_retrans
= tp
->retrans_out
;
2781 tcp_rack_mark_lost(sk
);
2782 if (prior_retrans
> tp
->retrans_out
)
2783 *ack_flag
|= FLAG_LOST_RETRANS
;
2787 /* Process an event, which can update packets-in-flight not trivially.
2788 * Main goal of this function is to calculate new estimate for left_out,
2789 * taking into account both packets sitting in receiver's buffer and
2790 * packets lost by network.
2792 * Besides that it updates the congestion state when packet loss or ECN
2793 * is detected. But it does not reduce the cwnd, it is done by the
2794 * congestion control later.
2796 * It does _not_ decide what to send, it is made in function
2797 * tcp_xmit_retransmit_queue().
2799 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2800 bool is_dupack
, int *ack_flag
, int *rexmit
)
2802 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2803 struct tcp_sock
*tp
= tcp_sk(sk
);
2804 int fast_rexmit
= 0, flag
= *ack_flag
;
2805 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2806 (tcp_fackets_out(tp
) > tp
->reordering
));
2808 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2810 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2811 tp
->fackets_out
= 0;
2813 /* Now state machine starts.
2814 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2815 if (flag
& FLAG_ECE
)
2816 tp
->prior_ssthresh
= 0;
2818 /* B. In all the states check for reneging SACKs. */
2819 if (tcp_check_sack_reneging(sk
, flag
))
2822 /* C. Check consistency of the current state. */
2823 tcp_verify_left_out(tp
);
2825 /* D. Check state exit conditions. State can be terminated
2826 * when high_seq is ACKed. */
2827 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2828 WARN_ON(tp
->retrans_out
!= 0);
2829 tp
->retrans_stamp
= 0;
2830 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2831 switch (icsk
->icsk_ca_state
) {
2833 /* CWR is to be held something *above* high_seq
2834 * is ACKed for CWR bit to reach receiver. */
2835 if (tp
->snd_una
!= tp
->high_seq
) {
2836 tcp_end_cwnd_reduction(sk
);
2837 tcp_set_ca_state(sk
, TCP_CA_Open
);
2841 case TCP_CA_Recovery
:
2842 if (tcp_is_reno(tp
))
2843 tcp_reset_reno_sack(tp
);
2844 if (tcp_try_undo_recovery(sk
))
2846 tcp_end_cwnd_reduction(sk
);
2851 /* E. Process state. */
2852 switch (icsk
->icsk_ca_state
) {
2853 case TCP_CA_Recovery
:
2854 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2855 if (tcp_is_reno(tp
) && is_dupack
)
2856 tcp_add_reno_sack(sk
);
2858 if (tcp_try_undo_partial(sk
, acked
))
2860 /* Partial ACK arrived. Force fast retransmit. */
2861 do_lost
= tcp_is_reno(tp
) ||
2862 tcp_fackets_out(tp
) > tp
->reordering
;
2864 if (tcp_try_undo_dsack(sk
)) {
2865 tcp_try_keep_open(sk
);
2868 tcp_rack_identify_loss(sk
, ack_flag
);
2871 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2872 tcp_rack_identify_loss(sk
, ack_flag
);
2873 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2874 (*ack_flag
& FLAG_LOST_RETRANS
)))
2876 /* Change state if cwnd is undone or retransmits are lost */
2878 if (tcp_is_reno(tp
)) {
2879 if (flag
& FLAG_SND_UNA_ADVANCED
)
2880 tcp_reset_reno_sack(tp
);
2882 tcp_add_reno_sack(sk
);
2885 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2886 tcp_try_undo_dsack(sk
);
2888 tcp_rack_identify_loss(sk
, ack_flag
);
2889 if (!tcp_time_to_recover(sk
, flag
)) {
2890 tcp_try_to_open(sk
, flag
);
2894 /* MTU probe failure: don't reduce cwnd */
2895 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2896 icsk
->icsk_mtup
.probe_size
&&
2897 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2898 tcp_mtup_probe_failed(sk
);
2899 /* Restores the reduction we did in tcp_mtup_probe() */
2901 tcp_simple_retransmit(sk
);
2905 /* Otherwise enter Recovery state */
2906 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2911 tcp_update_scoreboard(sk
, fast_rexmit
);
2912 *rexmit
= REXMIT_LOST
;
2915 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2917 struct tcp_sock
*tp
= tcp_sk(sk
);
2918 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2920 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2921 rtt_us
? : jiffies_to_usecs(1));
2924 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2925 long seq_rtt_us
, long sack_rtt_us
,
2926 long ca_rtt_us
, struct rate_sample
*rs
)
2928 const struct tcp_sock
*tp
= tcp_sk(sk
);
2930 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2931 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2932 * Karn's algorithm forbids taking RTT if some retransmitted data
2933 * is acked (RFC6298).
2936 seq_rtt_us
= sack_rtt_us
;
2938 /* RTTM Rule: A TSecr value received in a segment is used to
2939 * update the averaged RTT measurement only if the segment
2940 * acknowledges some new data, i.e., only if it advances the
2941 * left edge of the send window.
2942 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2944 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2945 flag
& FLAG_ACKED
) {
2946 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2947 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2949 seq_rtt_us
= ca_rtt_us
= delta_us
;
2951 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2955 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2956 * always taken together with ACK, SACK, or TS-opts. Any negative
2957 * values will be skipped with the seq_rtt_us < 0 check above.
2959 tcp_update_rtt_min(sk
, ca_rtt_us
);
2960 tcp_rtt_estimator(sk
, seq_rtt_us
);
2963 /* RFC6298: only reset backoff on valid RTT measurement. */
2964 inet_csk(sk
)->icsk_backoff
= 0;
2968 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2969 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2971 struct rate_sample rs
;
2974 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2975 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2977 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2981 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2983 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2985 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2986 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2989 /* Restart timer after forward progress on connection.
2990 * RFC2988 recommends to restart timer to now+rto.
2992 void tcp_rearm_rto(struct sock
*sk
)
2994 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2995 struct tcp_sock
*tp
= tcp_sk(sk
);
2997 /* If the retrans timer is currently being used by Fast Open
2998 * for SYN-ACK retrans purpose, stay put.
3000 if (tp
->fastopen_rsk
)
3003 if (!tp
->packets_out
) {
3004 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3006 u32 rto
= inet_csk(sk
)->icsk_rto
;
3007 /* Offset the time elapsed after installing regular RTO */
3008 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
3009 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3010 s64 delta_us
= tcp_rto_delta_us(sk
);
3011 /* delta_us may not be positive if the socket is locked
3012 * when the retrans timer fires and is rescheduled.
3014 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
3016 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3021 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3022 static void tcp_set_xmit_timer(struct sock
*sk
)
3024 if (!tcp_schedule_loss_probe(sk
))
3028 /* If we get here, the whole TSO packet has not been acked. */
3029 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3031 struct tcp_sock
*tp
= tcp_sk(sk
);
3034 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3036 packets_acked
= tcp_skb_pcount(skb
);
3037 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3039 packets_acked
-= tcp_skb_pcount(skb
);
3041 if (packets_acked
) {
3042 BUG_ON(tcp_skb_pcount(skb
) == 0);
3043 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3046 return packets_acked
;
3049 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3052 const struct skb_shared_info
*shinfo
;
3054 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3055 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3058 shinfo
= skb_shinfo(skb
);
3059 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3060 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3061 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3064 /* Remove acknowledged frames from the retransmission queue. If our packet
3065 * is before the ack sequence we can discard it as it's confirmed to have
3066 * arrived at the other end.
3068 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3069 u32 prior_snd_una
, int *acked
,
3070 struct tcp_sacktag_state
*sack
)
3072 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3073 u64 first_ackt
, last_ackt
;
3074 struct tcp_sock
*tp
= tcp_sk(sk
);
3075 u32 prior_sacked
= tp
->sacked_out
;
3076 u32 reord
= tp
->packets_out
;
3077 bool fully_acked
= true;
3078 long sack_rtt_us
= -1L;
3079 long seq_rtt_us
= -1L;
3080 long ca_rtt_us
= -1L;
3081 struct sk_buff
*skb
;
3083 u32 last_in_flight
= 0;
3089 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3090 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3091 u8 sacked
= scb
->sacked
;
3094 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3096 /* Determine how many packets and what bytes were acked, tso and else */
3097 if (after(scb
->end_seq
, tp
->snd_una
)) {
3098 if (tcp_skb_pcount(skb
) == 1 ||
3099 !after(tp
->snd_una
, scb
->seq
))
3102 acked_pcount
= tcp_tso_acked(sk
, skb
);
3105 fully_acked
= false;
3107 /* Speedup tcp_unlink_write_queue() and next loop */
3108 prefetchw(skb
->next
);
3109 acked_pcount
= tcp_skb_pcount(skb
);
3112 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3113 if (sacked
& TCPCB_SACKED_RETRANS
)
3114 tp
->retrans_out
-= acked_pcount
;
3115 flag
|= FLAG_RETRANS_DATA_ACKED
;
3116 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3117 last_ackt
= skb
->skb_mstamp
;
3118 WARN_ON_ONCE(last_ackt
== 0);
3120 first_ackt
= last_ackt
;
3122 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3123 reord
= min(pkts_acked
, reord
);
3124 if (!after(scb
->end_seq
, tp
->high_seq
))
3125 flag
|= FLAG_ORIG_SACK_ACKED
;
3128 if (sacked
& TCPCB_SACKED_ACKED
) {
3129 tp
->sacked_out
-= acked_pcount
;
3130 } else if (tcp_is_sack(tp
)) {
3131 tp
->delivered
+= acked_pcount
;
3132 if (!tcp_skb_spurious_retrans(tp
, skb
))
3133 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3136 if (sacked
& TCPCB_LOST
)
3137 tp
->lost_out
-= acked_pcount
;
3139 tp
->packets_out
-= acked_pcount
;
3140 pkts_acked
+= acked_pcount
;
3141 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3143 /* Initial outgoing SYN's get put onto the write_queue
3144 * just like anything else we transmit. It is not
3145 * true data, and if we misinform our callers that
3146 * this ACK acks real data, we will erroneously exit
3147 * connection startup slow start one packet too
3148 * quickly. This is severely frowned upon behavior.
3150 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3151 flag
|= FLAG_DATA_ACKED
;
3153 flag
|= FLAG_SYN_ACKED
;
3154 tp
->retrans_stamp
= 0;
3160 tcp_unlink_write_queue(skb
, sk
);
3161 sk_wmem_free_skb(sk
, skb
);
3162 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3163 tp
->retransmit_skb_hint
= NULL
;
3164 if (unlikely(skb
== tp
->lost_skb_hint
))
3165 tp
->lost_skb_hint
= NULL
;
3169 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3171 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3172 tp
->snd_up
= tp
->snd_una
;
3174 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3175 flag
|= FLAG_SACK_RENEGING
;
3177 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3178 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3179 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3181 if (sack
->first_sackt
) {
3182 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3183 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3185 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3186 ca_rtt_us
, sack
->rate
);
3188 if (flag
& FLAG_ACKED
) {
3189 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3190 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3191 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3192 tcp_mtup_probe_success(sk
);
3195 if (tcp_is_reno(tp
)) {
3196 tcp_remove_reno_sacks(sk
, pkts_acked
);
3200 /* Non-retransmitted hole got filled? That's reordering */
3201 if (reord
< prior_fackets
&& reord
<= tp
->fackets_out
)
3202 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3204 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3205 prior_sacked
- tp
->sacked_out
;
3206 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3209 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3211 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3212 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
, skb
->skb_mstamp
)) {
3213 /* Do not re-arm RTO if the sack RTT is measured from data sent
3214 * after when the head was last (re)transmitted. Otherwise the
3215 * timeout may continue to extend in loss recovery.
3217 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3220 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3221 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3222 .rtt_us
= sack
->rate
->rtt_us
,
3223 .in_flight
= last_in_flight
};
3225 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3228 #if FASTRETRANS_DEBUG > 0
3229 WARN_ON((int)tp
->sacked_out
< 0);
3230 WARN_ON((int)tp
->lost_out
< 0);
3231 WARN_ON((int)tp
->retrans_out
< 0);
3232 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3233 icsk
= inet_csk(sk
);
3235 pr_debug("Leak l=%u %d\n",
3236 tp
->lost_out
, icsk
->icsk_ca_state
);
3239 if (tp
->sacked_out
) {
3240 pr_debug("Leak s=%u %d\n",
3241 tp
->sacked_out
, icsk
->icsk_ca_state
);
3244 if (tp
->retrans_out
) {
3245 pr_debug("Leak r=%u %d\n",
3246 tp
->retrans_out
, icsk
->icsk_ca_state
);
3247 tp
->retrans_out
= 0;
3251 *acked
= pkts_acked
;
3255 static void tcp_ack_probe(struct sock
*sk
)
3257 const struct tcp_sock
*tp
= tcp_sk(sk
);
3258 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3260 /* Was it a usable window open? */
3262 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3263 icsk
->icsk_backoff
= 0;
3264 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3265 /* Socket must be waked up by subsequent tcp_data_snd_check().
3266 * This function is not for random using!
3269 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3271 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3276 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3278 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3279 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3282 /* Decide wheather to run the increase function of congestion control. */
3283 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3285 /* If reordering is high then always grow cwnd whenever data is
3286 * delivered regardless of its ordering. Otherwise stay conservative
3287 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3288 * new SACK or ECE mark may first advance cwnd here and later reduce
3289 * cwnd in tcp_fastretrans_alert() based on more states.
3291 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3292 return flag
& FLAG_FORWARD_PROGRESS
;
3294 return flag
& FLAG_DATA_ACKED
;
3297 /* The "ultimate" congestion control function that aims to replace the rigid
3298 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3299 * It's called toward the end of processing an ACK with precise rate
3300 * information. All transmission or retransmission are delayed afterwards.
3302 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3303 int flag
, const struct rate_sample
*rs
)
3305 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3307 if (icsk
->icsk_ca_ops
->cong_control
) {
3308 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3312 if (tcp_in_cwnd_reduction(sk
)) {
3313 /* Reduce cwnd if state mandates */
3314 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3315 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3316 /* Advance cwnd if state allows */
3317 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3319 tcp_update_pacing_rate(sk
);
3322 /* Check that window update is acceptable.
3323 * The function assumes that snd_una<=ack<=snd_next.
3325 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3326 const u32 ack
, const u32 ack_seq
,
3329 return after(ack
, tp
->snd_una
) ||
3330 after(ack_seq
, tp
->snd_wl1
) ||
3331 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3334 /* If we update tp->snd_una, also update tp->bytes_acked */
3335 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3337 u32 delta
= ack
- tp
->snd_una
;
3339 sock_owned_by_me((struct sock
*)tp
);
3340 tp
->bytes_acked
+= delta
;
3344 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3345 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3347 u32 delta
= seq
- tp
->rcv_nxt
;
3349 sock_owned_by_me((struct sock
*)tp
);
3350 tp
->bytes_received
+= delta
;
3354 /* Update our send window.
3356 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3357 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3359 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3362 struct tcp_sock
*tp
= tcp_sk(sk
);
3364 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3366 if (likely(!tcp_hdr(skb
)->syn
))
3367 nwin
<<= tp
->rx_opt
.snd_wscale
;
3369 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3370 flag
|= FLAG_WIN_UPDATE
;
3371 tcp_update_wl(tp
, ack_seq
);
3373 if (tp
->snd_wnd
!= nwin
) {
3376 /* Note, it is the only place, where
3377 * fast path is recovered for sending TCP.
3380 tcp_fast_path_check(sk
);
3382 if (tcp_send_head(sk
))
3383 tcp_slow_start_after_idle_check(sk
);
3385 if (nwin
> tp
->max_window
) {
3386 tp
->max_window
= nwin
;
3387 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3392 tcp_snd_una_update(tp
, ack
);
3397 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3398 u32
*last_oow_ack_time
)
3400 if (*last_oow_ack_time
) {
3401 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3403 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3404 NET_INC_STATS(net
, mib_idx
);
3405 return true; /* rate-limited: don't send yet! */
3409 *last_oow_ack_time
= tcp_jiffies32
;
3411 return false; /* not rate-limited: go ahead, send dupack now! */
3414 /* Return true if we're currently rate-limiting out-of-window ACKs and
3415 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3416 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3417 * attacks that send repeated SYNs or ACKs for the same connection. To
3418 * do this, we do not send a duplicate SYNACK or ACK if the remote
3419 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3421 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3422 int mib_idx
, u32
*last_oow_ack_time
)
3424 /* Data packets without SYNs are not likely part of an ACK loop. */
3425 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3429 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3432 /* RFC 5961 7 [ACK Throttling] */
3433 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3435 /* unprotected vars, we dont care of overwrites */
3436 static u32 challenge_timestamp
;
3437 static unsigned int challenge_count
;
3438 struct tcp_sock
*tp
= tcp_sk(sk
);
3441 /* First check our per-socket dupack rate limit. */
3442 if (__tcp_oow_rate_limited(sock_net(sk
),
3443 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3444 &tp
->last_oow_ack_time
))
3447 /* Then check host-wide RFC 5961 rate limit. */
3449 if (now
!= challenge_timestamp
) {
3450 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3452 challenge_timestamp
= now
;
3453 WRITE_ONCE(challenge_count
, half
+
3454 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3456 count
= READ_ONCE(challenge_count
);
3458 WRITE_ONCE(challenge_count
, count
- 1);
3459 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3464 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3466 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3467 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3470 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3472 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3473 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3474 * extra check below makes sure this can only happen
3475 * for pure ACK frames. -DaveM
3477 * Not only, also it occurs for expired timestamps.
3480 if (tcp_paws_check(&tp
->rx_opt
, 0))
3481 tcp_store_ts_recent(tp
);
3485 /* This routine deals with acks during a TLP episode.
3486 * We mark the end of a TLP episode on receiving TLP dupack or when
3487 * ack is after tlp_high_seq.
3488 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3490 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3492 struct tcp_sock
*tp
= tcp_sk(sk
);
3494 if (before(ack
, tp
->tlp_high_seq
))
3497 if (flag
& FLAG_DSACKING_ACK
) {
3498 /* This DSACK means original and TLP probe arrived; no loss */
3499 tp
->tlp_high_seq
= 0;
3500 } else if (after(ack
, tp
->tlp_high_seq
)) {
3501 /* ACK advances: there was a loss, so reduce cwnd. Reset
3502 * tlp_high_seq in tcp_init_cwnd_reduction()
3504 tcp_init_cwnd_reduction(sk
);
3505 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3506 tcp_end_cwnd_reduction(sk
);
3507 tcp_try_keep_open(sk
);
3508 NET_INC_STATS(sock_net(sk
),
3509 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3510 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3511 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3512 /* Pure dupack: original and TLP probe arrived; no loss */
3513 tp
->tlp_high_seq
= 0;
3517 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3519 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3521 if (icsk
->icsk_ca_ops
->in_ack_event
)
3522 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3525 /* Congestion control has updated the cwnd already. So if we're in
3526 * loss recovery then now we do any new sends (for FRTO) or
3527 * retransmits (for CA_Loss or CA_recovery) that make sense.
3529 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3531 struct tcp_sock
*tp
= tcp_sk(sk
);
3533 if (rexmit
== REXMIT_NONE
)
3536 if (unlikely(rexmit
== 2)) {
3537 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3539 if (after(tp
->snd_nxt
, tp
->high_seq
))
3543 tcp_xmit_retransmit_queue(sk
);
3546 /* This routine deals with incoming acks, but not outgoing ones. */
3547 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3549 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3550 struct tcp_sock
*tp
= tcp_sk(sk
);
3551 struct tcp_sacktag_state sack_state
;
3552 struct rate_sample rs
= { .prior_delivered
= 0 };
3553 u32 prior_snd_una
= tp
->snd_una
;
3554 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3555 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3556 bool is_dupack
= false;
3558 int prior_packets
= tp
->packets_out
;
3559 u32 delivered
= tp
->delivered
;
3560 u32 lost
= tp
->lost
;
3561 int acked
= 0; /* Number of packets newly acked */
3562 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3564 sack_state
.first_sackt
= 0;
3565 sack_state
.rate
= &rs
;
3567 /* We very likely will need to access write queue head. */
3568 prefetchw(sk
->sk_write_queue
.next
);
3570 /* If the ack is older than previous acks
3571 * then we can probably ignore it.
3573 if (before(ack
, prior_snd_una
)) {
3574 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3575 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3576 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3577 tcp_send_challenge_ack(sk
, skb
);
3583 /* If the ack includes data we haven't sent yet, discard
3584 * this segment (RFC793 Section 3.9).
3586 if (after(ack
, tp
->snd_nxt
))
3589 if (after(ack
, prior_snd_una
)) {
3590 flag
|= FLAG_SND_UNA_ADVANCED
;
3591 icsk
->icsk_retransmits
= 0;
3594 prior_fackets
= tp
->fackets_out
;
3595 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3597 /* ts_recent update must be made after we are sure that the packet
3600 if (flag
& FLAG_UPDATE_TS_RECENT
)
3601 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3603 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3604 /* Window is constant, pure forward advance.
3605 * No more checks are required.
3606 * Note, we use the fact that SND.UNA>=SND.WL2.
3608 tcp_update_wl(tp
, ack_seq
);
3609 tcp_snd_una_update(tp
, ack
);
3610 flag
|= FLAG_WIN_UPDATE
;
3612 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3614 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3616 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3618 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3621 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3623 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3625 if (TCP_SKB_CB(skb
)->sacked
)
3626 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3629 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3631 ack_ev_flags
|= CA_ACK_ECE
;
3634 if (flag
& FLAG_WIN_UPDATE
)
3635 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3637 tcp_in_ack_event(sk
, ack_ev_flags
);
3640 /* We passed data and got it acked, remove any soft error
3641 * log. Something worked...
3643 sk
->sk_err_soft
= 0;
3644 icsk
->icsk_probes_out
= 0;
3645 tp
->rcv_tstamp
= tcp_jiffies32
;
3649 /* See if we can take anything off of the retransmit queue. */
3650 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3653 if (tp
->tlp_high_seq
)
3654 tcp_process_tlp_ack(sk
, ack
, flag
);
3655 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3656 if (flag
& FLAG_SET_XMIT_TIMER
)
3657 tcp_set_xmit_timer(sk
);
3659 if (tcp_ack_is_dubious(sk
, flag
)) {
3660 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3661 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3664 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3667 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3668 lost
= tp
->lost
- lost
; /* freshly marked lost */
3669 tcp_rate_gen(sk
, delivered
, lost
, sack_state
.rate
);
3670 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3671 tcp_xmit_recovery(sk
, rexmit
);
3675 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3676 if (flag
& FLAG_DSACKING_ACK
)
3677 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3678 /* If this ack opens up a zero window, clear backoff. It was
3679 * being used to time the probes, and is probably far higher than
3680 * it needs to be for normal retransmission.
3682 if (tcp_send_head(sk
))
3685 if (tp
->tlp_high_seq
)
3686 tcp_process_tlp_ack(sk
, ack
, flag
);
3690 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3694 /* If data was SACKed, tag it and see if we should send more data.
3695 * If data was DSACKed, see if we can undo a cwnd reduction.
3697 if (TCP_SKB_CB(skb
)->sacked
) {
3698 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3700 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3701 tcp_xmit_recovery(sk
, rexmit
);
3704 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3708 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3709 bool syn
, struct tcp_fastopen_cookie
*foc
,
3712 /* Valid only in SYN or SYN-ACK with an even length. */
3713 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3716 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3717 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3718 memcpy(foc
->val
, cookie
, len
);
3725 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3726 * But, this can also be called on packets in the established flow when
3727 * the fast version below fails.
3729 void tcp_parse_options(const struct net
*net
,
3730 const struct sk_buff
*skb
,
3731 struct tcp_options_received
*opt_rx
, int estab
,
3732 struct tcp_fastopen_cookie
*foc
)
3734 const unsigned char *ptr
;
3735 const struct tcphdr
*th
= tcp_hdr(skb
);
3736 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3738 ptr
= (const unsigned char *)(th
+ 1);
3739 opt_rx
->saw_tstamp
= 0;
3741 while (length
> 0) {
3742 int opcode
= *ptr
++;
3748 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3753 if (opsize
< 2) /* "silly options" */
3755 if (opsize
> length
)
3756 return; /* don't parse partial options */
3759 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3760 u16 in_mss
= get_unaligned_be16(ptr
);
3762 if (opt_rx
->user_mss
&&
3763 opt_rx
->user_mss
< in_mss
)
3764 in_mss
= opt_rx
->user_mss
;
3765 opt_rx
->mss_clamp
= in_mss
;
3770 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3771 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3772 __u8 snd_wscale
= *(__u8
*)ptr
;
3773 opt_rx
->wscale_ok
= 1;
3774 if (snd_wscale
> TCP_MAX_WSCALE
) {
3775 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3779 snd_wscale
= TCP_MAX_WSCALE
;
3781 opt_rx
->snd_wscale
= snd_wscale
;
3784 case TCPOPT_TIMESTAMP
:
3785 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3786 ((estab
&& opt_rx
->tstamp_ok
) ||
3787 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3788 opt_rx
->saw_tstamp
= 1;
3789 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3790 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3793 case TCPOPT_SACK_PERM
:
3794 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3795 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3796 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3797 tcp_sack_reset(opt_rx
);
3802 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3803 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3805 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3808 #ifdef CONFIG_TCP_MD5SIG
3811 * The MD5 Hash has already been
3812 * checked (see tcp_v{4,6}_do_rcv()).
3816 case TCPOPT_FASTOPEN
:
3817 tcp_parse_fastopen_option(
3818 opsize
- TCPOLEN_FASTOPEN_BASE
,
3819 ptr
, th
->syn
, foc
, false);
3823 /* Fast Open option shares code 254 using a
3824 * 16 bits magic number.
3826 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3827 get_unaligned_be16(ptr
) ==
3828 TCPOPT_FASTOPEN_MAGIC
)
3829 tcp_parse_fastopen_option(opsize
-
3830 TCPOLEN_EXP_FASTOPEN_BASE
,
3831 ptr
+ 2, th
->syn
, foc
, true);
3840 EXPORT_SYMBOL(tcp_parse_options
);
3842 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3844 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3846 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3847 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3848 tp
->rx_opt
.saw_tstamp
= 1;
3850 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3853 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3855 tp
->rx_opt
.rcv_tsecr
= 0;
3861 /* Fast parse options. This hopes to only see timestamps.
3862 * If it is wrong it falls back on tcp_parse_options().
3864 static bool tcp_fast_parse_options(const struct net
*net
,
3865 const struct sk_buff
*skb
,
3866 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3868 /* In the spirit of fast parsing, compare doff directly to constant
3869 * values. Because equality is used, short doff can be ignored here.
3871 if (th
->doff
== (sizeof(*th
) / 4)) {
3872 tp
->rx_opt
.saw_tstamp
= 0;
3874 } else if (tp
->rx_opt
.tstamp_ok
&&
3875 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3876 if (tcp_parse_aligned_timestamp(tp
, th
))
3880 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3881 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3882 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3887 #ifdef CONFIG_TCP_MD5SIG
3889 * Parse MD5 Signature option
3891 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3893 int length
= (th
->doff
<< 2) - sizeof(*th
);
3894 const u8
*ptr
= (const u8
*)(th
+ 1);
3896 /* If the TCP option is too short, we can short cut */
3897 if (length
< TCPOLEN_MD5SIG
)
3900 while (length
> 0) {
3901 int opcode
= *ptr
++;
3912 if (opsize
< 2 || opsize
> length
)
3914 if (opcode
== TCPOPT_MD5SIG
)
3915 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3922 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3925 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3927 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3928 * it can pass through stack. So, the following predicate verifies that
3929 * this segment is not used for anything but congestion avoidance or
3930 * fast retransmit. Moreover, we even are able to eliminate most of such
3931 * second order effects, if we apply some small "replay" window (~RTO)
3932 * to timestamp space.
3934 * All these measures still do not guarantee that we reject wrapped ACKs
3935 * on networks with high bandwidth, when sequence space is recycled fastly,
3936 * but it guarantees that such events will be very rare and do not affect
3937 * connection seriously. This doesn't look nice, but alas, PAWS is really
3940 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3941 * states that events when retransmit arrives after original data are rare.
3942 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3943 * the biggest problem on large power networks even with minor reordering.
3944 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3945 * up to bandwidth of 18Gigabit/sec. 8) ]
3948 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3950 const struct tcp_sock
*tp
= tcp_sk(sk
);
3951 const struct tcphdr
*th
= tcp_hdr(skb
);
3952 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3953 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3955 return (/* 1. Pure ACK with correct sequence number. */
3956 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3958 /* 2. ... and duplicate ACK. */
3959 ack
== tp
->snd_una
&&
3961 /* 3. ... and does not update window. */
3962 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3964 /* 4. ... and sits in replay window. */
3965 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3968 static inline bool tcp_paws_discard(const struct sock
*sk
,
3969 const struct sk_buff
*skb
)
3971 const struct tcp_sock
*tp
= tcp_sk(sk
);
3973 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3974 !tcp_disordered_ack(sk
, skb
);
3977 /* Check segment sequence number for validity.
3979 * Segment controls are considered valid, if the segment
3980 * fits to the window after truncation to the window. Acceptability
3981 * of data (and SYN, FIN, of course) is checked separately.
3982 * See tcp_data_queue(), for example.
3984 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3985 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3986 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3987 * (borrowed from freebsd)
3990 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3992 return !before(end_seq
, tp
->rcv_wup
) &&
3993 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3996 /* When we get a reset we do this. */
3997 void tcp_reset(struct sock
*sk
)
3999 /* We want the right error as BSD sees it (and indeed as we do). */
4000 switch (sk
->sk_state
) {
4002 sk
->sk_err
= ECONNREFUSED
;
4004 case TCP_CLOSE_WAIT
:
4010 sk
->sk_err
= ECONNRESET
;
4012 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4017 if (!sock_flag(sk
, SOCK_DEAD
))
4018 sk
->sk_error_report(sk
);
4022 * Process the FIN bit. This now behaves as it is supposed to work
4023 * and the FIN takes effect when it is validly part of sequence
4024 * space. Not before when we get holes.
4026 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4027 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4030 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4031 * close and we go into CLOSING (and later onto TIME-WAIT)
4033 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4035 void tcp_fin(struct sock
*sk
)
4037 struct tcp_sock
*tp
= tcp_sk(sk
);
4039 inet_csk_schedule_ack(sk
);
4041 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4042 sock_set_flag(sk
, SOCK_DONE
);
4044 switch (sk
->sk_state
) {
4046 case TCP_ESTABLISHED
:
4047 /* Move to CLOSE_WAIT */
4048 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4049 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4052 case TCP_CLOSE_WAIT
:
4054 /* Received a retransmission of the FIN, do
4059 /* RFC793: Remain in the LAST-ACK state. */
4063 /* This case occurs when a simultaneous close
4064 * happens, we must ack the received FIN and
4065 * enter the CLOSING state.
4068 tcp_set_state(sk
, TCP_CLOSING
);
4071 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4073 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4076 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4077 * cases we should never reach this piece of code.
4079 pr_err("%s: Impossible, sk->sk_state=%d\n",
4080 __func__
, sk
->sk_state
);
4084 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4085 * Probably, we should reset in this case. For now drop them.
4087 skb_rbtree_purge(&tp
->out_of_order_queue
);
4088 if (tcp_is_sack(tp
))
4089 tcp_sack_reset(&tp
->rx_opt
);
4092 if (!sock_flag(sk
, SOCK_DEAD
)) {
4093 sk
->sk_state_change(sk
);
4095 /* Do not send POLL_HUP for half duplex close. */
4096 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4097 sk
->sk_state
== TCP_CLOSE
)
4098 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4100 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4104 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4107 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4108 if (before(seq
, sp
->start_seq
))
4109 sp
->start_seq
= seq
;
4110 if (after(end_seq
, sp
->end_seq
))
4111 sp
->end_seq
= end_seq
;
4117 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4119 struct tcp_sock
*tp
= tcp_sk(sk
);
4121 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4124 if (before(seq
, tp
->rcv_nxt
))
4125 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4127 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4129 NET_INC_STATS(sock_net(sk
), mib_idx
);
4131 tp
->rx_opt
.dsack
= 1;
4132 tp
->duplicate_sack
[0].start_seq
= seq
;
4133 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4137 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4139 struct tcp_sock
*tp
= tcp_sk(sk
);
4141 if (!tp
->rx_opt
.dsack
)
4142 tcp_dsack_set(sk
, seq
, end_seq
);
4144 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4147 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4149 struct tcp_sock
*tp
= tcp_sk(sk
);
4151 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4152 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4153 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4154 tcp_enter_quickack_mode(sk
);
4156 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4157 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4159 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4160 end_seq
= tp
->rcv_nxt
;
4161 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4168 /* These routines update the SACK block as out-of-order packets arrive or
4169 * in-order packets close up the sequence space.
4171 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4174 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4175 struct tcp_sack_block
*swalk
= sp
+ 1;
4177 /* See if the recent change to the first SACK eats into
4178 * or hits the sequence space of other SACK blocks, if so coalesce.
4180 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4181 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4184 /* Zap SWALK, by moving every further SACK up by one slot.
4185 * Decrease num_sacks.
4187 tp
->rx_opt
.num_sacks
--;
4188 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4192 this_sack
++, swalk
++;
4196 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4198 struct tcp_sock
*tp
= tcp_sk(sk
);
4199 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4200 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4206 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4207 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4208 /* Rotate this_sack to the first one. */
4209 for (; this_sack
> 0; this_sack
--, sp
--)
4210 swap(*sp
, *(sp
- 1));
4212 tcp_sack_maybe_coalesce(tp
);
4217 /* Could not find an adjacent existing SACK, build a new one,
4218 * put it at the front, and shift everyone else down. We
4219 * always know there is at least one SACK present already here.
4221 * If the sack array is full, forget about the last one.
4223 if (this_sack
>= TCP_NUM_SACKS
) {
4225 tp
->rx_opt
.num_sacks
--;
4228 for (; this_sack
> 0; this_sack
--, sp
--)
4232 /* Build the new head SACK, and we're done. */
4233 sp
->start_seq
= seq
;
4234 sp
->end_seq
= end_seq
;
4235 tp
->rx_opt
.num_sacks
++;
4238 /* RCV.NXT advances, some SACKs should be eaten. */
4240 static void tcp_sack_remove(struct tcp_sock
*tp
)
4242 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4243 int num_sacks
= tp
->rx_opt
.num_sacks
;
4246 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4247 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4248 tp
->rx_opt
.num_sacks
= 0;
4252 for (this_sack
= 0; this_sack
< num_sacks
;) {
4253 /* Check if the start of the sack is covered by RCV.NXT. */
4254 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4257 /* RCV.NXT must cover all the block! */
4258 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4260 /* Zap this SACK, by moving forward any other SACKS. */
4261 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4262 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4269 tp
->rx_opt
.num_sacks
= num_sacks
;
4278 * tcp_try_coalesce - try to merge skb to prior one
4280 * @dest: destination queue
4282 * @from: buffer to add in queue
4283 * @fragstolen: pointer to boolean
4285 * Before queueing skb @from after @to, try to merge them
4286 * to reduce overall memory use and queue lengths, if cost is small.
4287 * Packets in ofo or receive queues can stay a long time.
4288 * Better try to coalesce them right now to avoid future collapses.
4289 * Returns true if caller should free @from instead of queueing it
4291 static bool tcp_try_coalesce(struct sock
*sk
,
4292 enum tcp_queue dest
,
4294 struct sk_buff
*from
,
4299 *fragstolen
= false;
4301 /* Its possible this segment overlaps with prior segment in queue */
4302 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4305 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4308 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4309 sk_mem_charge(sk
, delta
);
4310 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4311 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4312 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4313 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4315 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4316 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4317 if (dest
== OOO_QUEUE
)
4318 TCP_SKB_CB(to
)->swtstamp
= TCP_SKB_CB(from
)->swtstamp
;
4320 to
->tstamp
= from
->tstamp
;
4326 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4328 sk_drops_add(sk
, skb
);
4332 /* This one checks to see if we can put data from the
4333 * out_of_order queue into the receive_queue.
4335 static void tcp_ofo_queue(struct sock
*sk
)
4337 struct tcp_sock
*tp
= tcp_sk(sk
);
4338 __u32 dsack_high
= tp
->rcv_nxt
;
4339 bool fin
, fragstolen
, eaten
;
4340 struct sk_buff
*skb
, *tail
;
4343 p
= rb_first(&tp
->out_of_order_queue
);
4345 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4346 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4349 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4350 __u32 dsack
= dsack_high
;
4351 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4352 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4353 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4356 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4357 /* Replace tstamp which was stomped by rbnode */
4358 if (TCP_SKB_CB(skb
)->has_rxtstamp
)
4359 skb
->tstamp
= TCP_SKB_CB(skb
)->swtstamp
;
4361 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4362 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4366 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4367 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4368 TCP_SKB_CB(skb
)->end_seq
);
4370 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4371 eaten
= tail
&& tcp_try_coalesce(sk
, RCV_QUEUE
,
4372 tail
, skb
, &fragstolen
);
4373 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4374 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4376 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4378 kfree_skb_partial(skb
, fragstolen
);
4380 if (unlikely(fin
)) {
4382 /* tcp_fin() purges tp->out_of_order_queue,
4383 * so we must end this loop right now.
4390 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4391 static int tcp_prune_queue(struct sock
*sk
);
4393 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4396 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4397 !sk_rmem_schedule(sk
, skb
, size
)) {
4399 if (tcp_prune_queue(sk
) < 0)
4402 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4403 if (!tcp_prune_ofo_queue(sk
))
4410 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4412 struct tcp_sock
*tp
= tcp_sk(sk
);
4413 struct rb_node
**p
, *q
, *parent
;
4414 struct sk_buff
*skb1
;
4418 tcp_ecn_check_ce(tp
, skb
);
4420 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4421 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4426 /* Stash tstamp to avoid being stomped on by rbnode */
4427 if (TCP_SKB_CB(skb
)->has_rxtstamp
)
4428 TCP_SKB_CB(skb
)->swtstamp
= skb
->tstamp
;
4430 /* Disable header prediction. */
4432 inet_csk_schedule_ack(sk
);
4434 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4435 seq
= TCP_SKB_CB(skb
)->seq
;
4436 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4437 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4438 tp
->rcv_nxt
, seq
, end_seq
);
4440 p
= &tp
->out_of_order_queue
.rb_node
;
4441 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4442 /* Initial out of order segment, build 1 SACK. */
4443 if (tcp_is_sack(tp
)) {
4444 tp
->rx_opt
.num_sacks
= 1;
4445 tp
->selective_acks
[0].start_seq
= seq
;
4446 tp
->selective_acks
[0].end_seq
= end_seq
;
4448 rb_link_node(&skb
->rbnode
, NULL
, p
);
4449 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4450 tp
->ooo_last_skb
= skb
;
4454 /* In the typical case, we are adding an skb to the end of the list.
4455 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4457 if (tcp_try_coalesce(sk
, OOO_QUEUE
, tp
->ooo_last_skb
,
4458 skb
, &fragstolen
)) {
4460 tcp_grow_window(sk
, skb
);
4461 kfree_skb_partial(skb
, fragstolen
);
4465 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4466 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4467 parent
= &tp
->ooo_last_skb
->rbnode
;
4468 p
= &parent
->rb_right
;
4472 /* Find place to insert this segment. Handle overlaps on the way. */
4476 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4477 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4478 p
= &parent
->rb_left
;
4481 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4482 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4483 /* All the bits are present. Drop. */
4484 NET_INC_STATS(sock_net(sk
),
4485 LINUX_MIB_TCPOFOMERGE
);
4488 tcp_dsack_set(sk
, seq
, end_seq
);
4491 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4492 /* Partial overlap. */
4493 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4495 /* skb's seq == skb1's seq and skb covers skb1.
4496 * Replace skb1 with skb.
4498 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4499 &tp
->out_of_order_queue
);
4500 tcp_dsack_extend(sk
,
4501 TCP_SKB_CB(skb1
)->seq
,
4502 TCP_SKB_CB(skb1
)->end_seq
);
4503 NET_INC_STATS(sock_net(sk
),
4504 LINUX_MIB_TCPOFOMERGE
);
4508 } else if (tcp_try_coalesce(sk
, OOO_QUEUE
, skb1
,
4509 skb
, &fragstolen
)) {
4512 p
= &parent
->rb_right
;
4515 /* Insert segment into RB tree. */
4516 rb_link_node(&skb
->rbnode
, parent
, p
);
4517 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4520 /* Remove other segments covered by skb. */
4521 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4522 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4524 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4526 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4527 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4531 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4532 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4533 TCP_SKB_CB(skb1
)->end_seq
);
4534 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4537 /* If there is no skb after us, we are the last_skb ! */
4539 tp
->ooo_last_skb
= skb
;
4542 if (tcp_is_sack(tp
))
4543 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4546 tcp_grow_window(sk
, skb
);
4548 skb_set_owner_r(skb
, sk
);
4552 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4556 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4558 __skb_pull(skb
, hdrlen
);
4560 tcp_try_coalesce(sk
, RCV_QUEUE
, tail
,
4561 skb
, fragstolen
)) ? 1 : 0;
4562 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4564 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4565 skb_set_owner_r(skb
, sk
);
4570 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4572 struct sk_buff
*skb
;
4580 if (size
> PAGE_SIZE
) {
4581 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4583 data_len
= npages
<< PAGE_SHIFT
;
4584 size
= data_len
+ (size
& ~PAGE_MASK
);
4586 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4587 PAGE_ALLOC_COSTLY_ORDER
,
4588 &err
, sk
->sk_allocation
);
4592 skb_put(skb
, size
- data_len
);
4593 skb
->data_len
= data_len
;
4596 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4599 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4603 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4604 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4605 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4607 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4608 WARN_ON_ONCE(fragstolen
); /* should not happen */
4620 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4622 struct tcp_sock
*tp
= tcp_sk(sk
);
4626 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4631 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4633 tcp_ecn_accept_cwr(tp
, skb
);
4635 tp
->rx_opt
.dsack
= 0;
4637 /* Queue data for delivery to the user.
4638 * Packets in sequence go to the receive queue.
4639 * Out of sequence packets to the out_of_order_queue.
4641 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4642 if (tcp_receive_window(tp
) == 0)
4645 /* Ok. In sequence. In window. */
4647 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4648 sk_forced_mem_schedule(sk
, skb
->truesize
);
4649 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4652 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4653 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4655 tcp_event_data_recv(sk
, skb
);
4656 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4659 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4662 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4663 * gap in queue is filled.
4665 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4666 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4669 if (tp
->rx_opt
.num_sacks
)
4670 tcp_sack_remove(tp
);
4672 tcp_fast_path_check(sk
);
4675 kfree_skb_partial(skb
, fragstolen
);
4676 if (!sock_flag(sk
, SOCK_DEAD
))
4677 sk
->sk_data_ready(sk
);
4681 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4682 /* A retransmit, 2nd most common case. Force an immediate ack. */
4683 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4684 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4687 tcp_enter_quickack_mode(sk
);
4688 inet_csk_schedule_ack(sk
);
4694 /* Out of window. F.e. zero window probe. */
4695 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4698 tcp_enter_quickack_mode(sk
);
4700 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4701 /* Partial packet, seq < rcv_next < end_seq */
4702 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4703 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4704 TCP_SKB_CB(skb
)->end_seq
);
4706 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4708 /* If window is closed, drop tail of packet. But after
4709 * remembering D-SACK for its head made in previous line.
4711 if (!tcp_receive_window(tp
))
4716 tcp_data_queue_ofo(sk
, skb
);
4719 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4722 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4724 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4727 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4728 struct sk_buff_head
*list
,
4729 struct rb_root
*root
)
4731 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4734 __skb_unlink(skb
, list
);
4736 rb_erase(&skb
->rbnode
, root
);
4739 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4744 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4745 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4747 struct rb_node
**p
= &root
->rb_node
;
4748 struct rb_node
*parent
= NULL
;
4749 struct sk_buff
*skb1
;
4753 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4754 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4755 p
= &parent
->rb_left
;
4757 p
= &parent
->rb_right
;
4759 rb_link_node(&skb
->rbnode
, parent
, p
);
4760 rb_insert_color(&skb
->rbnode
, root
);
4763 /* Collapse contiguous sequence of skbs head..tail with
4764 * sequence numbers start..end.
4766 * If tail is NULL, this means until the end of the queue.
4768 * Segments with FIN/SYN are not collapsed (only because this
4772 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4773 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4775 struct sk_buff
*skb
= head
, *n
;
4776 struct sk_buff_head tmp
;
4779 /* First, check that queue is collapsible and find
4780 * the point where collapsing can be useful.
4783 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4784 n
= tcp_skb_next(skb
, list
);
4786 /* No new bits? It is possible on ofo queue. */
4787 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4788 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4794 /* The first skb to collapse is:
4796 * - bloated or contains data before "start" or
4797 * overlaps to the next one.
4799 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4800 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4801 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4802 end_of_skbs
= false;
4806 if (n
&& n
!= tail
&&
4807 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4808 end_of_skbs
= false;
4812 /* Decided to skip this, advance start seq. */
4813 start
= TCP_SKB_CB(skb
)->end_seq
;
4816 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4819 __skb_queue_head_init(&tmp
);
4821 while (before(start
, end
)) {
4822 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4823 struct sk_buff
*nskb
;
4825 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4829 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4830 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4832 __skb_queue_before(list
, skb
, nskb
);
4834 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4835 skb_set_owner_r(nskb
, sk
);
4837 /* Copy data, releasing collapsed skbs. */
4839 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4840 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4844 size
= min(copy
, size
);
4845 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4847 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4851 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4852 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4855 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4861 skb_queue_walk_safe(&tmp
, skb
, n
)
4862 tcp_rbtree_insert(root
, skb
);
4865 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4866 * and tcp_collapse() them until all the queue is collapsed.
4868 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4870 struct tcp_sock
*tp
= tcp_sk(sk
);
4871 struct sk_buff
*skb
, *head
;
4875 p
= rb_first(&tp
->out_of_order_queue
);
4876 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4879 p
= rb_last(&tp
->out_of_order_queue
);
4880 /* Note: This is possible p is NULL here. We do not
4881 * use rb_entry_safe(), as ooo_last_skb is valid only
4882 * if rbtree is not empty.
4884 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4887 start
= TCP_SKB_CB(skb
)->seq
;
4888 end
= TCP_SKB_CB(skb
)->end_seq
;
4890 for (head
= skb
;;) {
4891 skb
= tcp_skb_next(skb
, NULL
);
4893 /* Range is terminated when we see a gap or when
4894 * we are at the queue end.
4897 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4898 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4899 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4900 head
, skb
, start
, end
);
4904 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4905 start
= TCP_SKB_CB(skb
)->seq
;
4906 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4907 end
= TCP_SKB_CB(skb
)->end_seq
;
4912 * Clean the out-of-order queue to make room.
4913 * We drop high sequences packets to :
4914 * 1) Let a chance for holes to be filled.
4915 * 2) not add too big latencies if thousands of packets sit there.
4916 * (But if application shrinks SO_RCVBUF, we could still end up
4917 * freeing whole queue here)
4919 * Return true if queue has shrunk.
4921 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4923 struct tcp_sock
*tp
= tcp_sk(sk
);
4924 struct rb_node
*node
, *prev
;
4926 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4929 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4930 node
= &tp
->ooo_last_skb
->rbnode
;
4932 prev
= rb_prev(node
);
4933 rb_erase(node
, &tp
->out_of_order_queue
);
4934 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4936 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4937 !tcp_under_memory_pressure(sk
))
4941 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4943 /* Reset SACK state. A conforming SACK implementation will
4944 * do the same at a timeout based retransmit. When a connection
4945 * is in a sad state like this, we care only about integrity
4946 * of the connection not performance.
4948 if (tp
->rx_opt
.sack_ok
)
4949 tcp_sack_reset(&tp
->rx_opt
);
4953 /* Reduce allocated memory if we can, trying to get
4954 * the socket within its memory limits again.
4956 * Return less than zero if we should start dropping frames
4957 * until the socket owning process reads some of the data
4958 * to stabilize the situation.
4960 static int tcp_prune_queue(struct sock
*sk
)
4962 struct tcp_sock
*tp
= tcp_sk(sk
);
4964 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4966 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4968 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4969 tcp_clamp_window(sk
);
4970 else if (tcp_under_memory_pressure(sk
))
4971 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4973 tcp_collapse_ofo_queue(sk
);
4974 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4975 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4976 skb_peek(&sk
->sk_receive_queue
),
4978 tp
->copied_seq
, tp
->rcv_nxt
);
4981 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4984 /* Collapsing did not help, destructive actions follow.
4985 * This must not ever occur. */
4987 tcp_prune_ofo_queue(sk
);
4989 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4992 /* If we are really being abused, tell the caller to silently
4993 * drop receive data on the floor. It will get retransmitted
4994 * and hopefully then we'll have sufficient space.
4996 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4998 /* Massive buffer overcommit. */
5003 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5005 const struct tcp_sock
*tp
= tcp_sk(sk
);
5007 /* If the user specified a specific send buffer setting, do
5010 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5013 /* If we are under global TCP memory pressure, do not expand. */
5014 if (tcp_under_memory_pressure(sk
))
5017 /* If we are under soft global TCP memory pressure, do not expand. */
5018 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5021 /* If we filled the congestion window, do not expand. */
5022 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5028 /* When incoming ACK allowed to free some skb from write_queue,
5029 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5030 * on the exit from tcp input handler.
5032 * PROBLEM: sndbuf expansion does not work well with largesend.
5034 static void tcp_new_space(struct sock
*sk
)
5036 struct tcp_sock
*tp
= tcp_sk(sk
);
5038 if (tcp_should_expand_sndbuf(sk
)) {
5039 tcp_sndbuf_expand(sk
);
5040 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5043 sk
->sk_write_space(sk
);
5046 static void tcp_check_space(struct sock
*sk
)
5048 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5049 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5050 /* pairs with tcp_poll() */
5052 if (sk
->sk_socket
&&
5053 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5055 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5056 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5061 static inline void tcp_data_snd_check(struct sock
*sk
)
5063 tcp_push_pending_frames(sk
);
5064 tcp_check_space(sk
);
5068 * Check if sending an ack is needed.
5070 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5072 struct tcp_sock
*tp
= tcp_sk(sk
);
5074 /* More than one full frame received... */
5075 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5076 /* ... and right edge of window advances far enough.
5077 * (tcp_recvmsg() will send ACK otherwise). Or...
5079 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5080 /* We ACK each frame or... */
5081 tcp_in_quickack_mode(sk
) ||
5082 /* We have out of order data. */
5083 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5084 /* Then ack it now */
5087 /* Else, send delayed ack. */
5088 tcp_send_delayed_ack(sk
);
5092 static inline void tcp_ack_snd_check(struct sock
*sk
)
5094 if (!inet_csk_ack_scheduled(sk
)) {
5095 /* We sent a data segment already. */
5098 __tcp_ack_snd_check(sk
, 1);
5102 * This routine is only called when we have urgent data
5103 * signaled. Its the 'slow' part of tcp_urg. It could be
5104 * moved inline now as tcp_urg is only called from one
5105 * place. We handle URGent data wrong. We have to - as
5106 * BSD still doesn't use the correction from RFC961.
5107 * For 1003.1g we should support a new option TCP_STDURG to permit
5108 * either form (or just set the sysctl tcp_stdurg).
5111 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5113 struct tcp_sock
*tp
= tcp_sk(sk
);
5114 u32 ptr
= ntohs(th
->urg_ptr
);
5116 if (ptr
&& !sysctl_tcp_stdurg
)
5118 ptr
+= ntohl(th
->seq
);
5120 /* Ignore urgent data that we've already seen and read. */
5121 if (after(tp
->copied_seq
, ptr
))
5124 /* Do not replay urg ptr.
5126 * NOTE: interesting situation not covered by specs.
5127 * Misbehaving sender may send urg ptr, pointing to segment,
5128 * which we already have in ofo queue. We are not able to fetch
5129 * such data and will stay in TCP_URG_NOTYET until will be eaten
5130 * by recvmsg(). Seems, we are not obliged to handle such wicked
5131 * situations. But it is worth to think about possibility of some
5132 * DoSes using some hypothetical application level deadlock.
5134 if (before(ptr
, tp
->rcv_nxt
))
5137 /* Do we already have a newer (or duplicate) urgent pointer? */
5138 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5141 /* Tell the world about our new urgent pointer. */
5144 /* We may be adding urgent data when the last byte read was
5145 * urgent. To do this requires some care. We cannot just ignore
5146 * tp->copied_seq since we would read the last urgent byte again
5147 * as data, nor can we alter copied_seq until this data arrives
5148 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5150 * NOTE. Double Dutch. Rendering to plain English: author of comment
5151 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5152 * and expect that both A and B disappear from stream. This is _wrong_.
5153 * Though this happens in BSD with high probability, this is occasional.
5154 * Any application relying on this is buggy. Note also, that fix "works"
5155 * only in this artificial test. Insert some normal data between A and B and we will
5156 * decline of BSD again. Verdict: it is better to remove to trap
5159 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5160 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5161 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5163 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5164 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5169 tp
->urg_data
= TCP_URG_NOTYET
;
5172 /* Disable header prediction. */
5176 /* This is the 'fast' part of urgent handling. */
5177 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5179 struct tcp_sock
*tp
= tcp_sk(sk
);
5181 /* Check if we get a new urgent pointer - normally not. */
5183 tcp_check_urg(sk
, th
);
5185 /* Do we wait for any urgent data? - normally not... */
5186 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5187 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5190 /* Is the urgent pointer pointing into this packet? */
5191 if (ptr
< skb
->len
) {
5193 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5195 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5196 if (!sock_flag(sk
, SOCK_DEAD
))
5197 sk
->sk_data_ready(sk
);
5202 /* Accept RST for rcv_nxt - 1 after a FIN.
5203 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5204 * FIN is sent followed by a RST packet. The RST is sent with the same
5205 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5206 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5207 * ACKs on the closed socket. In addition middleboxes can drop either the
5208 * challenge ACK or a subsequent RST.
5210 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5212 struct tcp_sock
*tp
= tcp_sk(sk
);
5214 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5215 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5219 /* Does PAWS and seqno based validation of an incoming segment, flags will
5220 * play significant role here.
5222 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5223 const struct tcphdr
*th
, int syn_inerr
)
5225 struct tcp_sock
*tp
= tcp_sk(sk
);
5226 bool rst_seq_match
= false;
5228 /* RFC1323: H1. Apply PAWS check first. */
5229 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5230 tp
->rx_opt
.saw_tstamp
&&
5231 tcp_paws_discard(sk
, skb
)) {
5233 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5234 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5235 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5236 &tp
->last_oow_ack_time
))
5237 tcp_send_dupack(sk
, skb
);
5240 /* Reset is accepted even if it did not pass PAWS. */
5243 /* Step 1: check sequence number */
5244 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5245 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5246 * (RST) segments are validated by checking their SEQ-fields."
5247 * And page 69: "If an incoming segment is not acceptable,
5248 * an acknowledgment should be sent in reply (unless the RST
5249 * bit is set, if so drop the segment and return)".
5254 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5255 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5256 &tp
->last_oow_ack_time
))
5257 tcp_send_dupack(sk
, skb
);
5258 } else if (tcp_reset_check(sk
, skb
)) {
5264 /* Step 2: check RST bit */
5266 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5267 * FIN and SACK too if available):
5268 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5269 * the right-most SACK block,
5271 * RESET the connection
5273 * Send a challenge ACK
5275 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5276 tcp_reset_check(sk
, skb
)) {
5277 rst_seq_match
= true;
5278 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5279 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5280 int max_sack
= sp
[0].end_seq
;
5283 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5285 max_sack
= after(sp
[this_sack
].end_seq
,
5287 sp
[this_sack
].end_seq
: max_sack
;
5290 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5291 rst_seq_match
= true;
5297 /* Disable TFO if RST is out-of-order
5298 * and no data has been received
5299 * for current active TFO socket
5301 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5302 sk
->sk_state
== TCP_ESTABLISHED
)
5303 tcp_fastopen_active_disable(sk
);
5304 tcp_send_challenge_ack(sk
, skb
);
5309 /* step 3: check security and precedence [ignored] */
5311 /* step 4: Check for a SYN
5312 * RFC 5961 4.2 : Send a challenge ack
5317 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5318 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5319 tcp_send_challenge_ack(sk
, skb
);
5331 * TCP receive function for the ESTABLISHED state.
5333 * It is split into a fast path and a slow path. The fast path is
5335 * - A zero window was announced from us - zero window probing
5336 * is only handled properly in the slow path.
5337 * - Out of order segments arrived.
5338 * - Urgent data is expected.
5339 * - There is no buffer space left
5340 * - Unexpected TCP flags/window values/header lengths are received
5341 * (detected by checking the TCP header against pred_flags)
5342 * - Data is sent in both directions. Fast path only supports pure senders
5343 * or pure receivers (this means either the sequence number or the ack
5344 * value must stay constant)
5345 * - Unexpected TCP option.
5347 * When these conditions are not satisfied it drops into a standard
5348 * receive procedure patterned after RFC793 to handle all cases.
5349 * The first three cases are guaranteed by proper pred_flags setting,
5350 * the rest is checked inline. Fast processing is turned on in
5351 * tcp_data_queue when everything is OK.
5353 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5354 const struct tcphdr
*th
)
5356 unsigned int len
= skb
->len
;
5357 struct tcp_sock
*tp
= tcp_sk(sk
);
5359 tcp_mstamp_refresh(tp
);
5360 if (unlikely(!sk
->sk_rx_dst
))
5361 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5363 * Header prediction.
5364 * The code loosely follows the one in the famous
5365 * "30 instruction TCP receive" Van Jacobson mail.
5367 * Van's trick is to deposit buffers into socket queue
5368 * on a device interrupt, to call tcp_recv function
5369 * on the receive process context and checksum and copy
5370 * the buffer to user space. smart...
5372 * Our current scheme is not silly either but we take the
5373 * extra cost of the net_bh soft interrupt processing...
5374 * We do checksum and copy also but from device to kernel.
5377 tp
->rx_opt
.saw_tstamp
= 0;
5379 /* pred_flags is 0xS?10 << 16 + snd_wnd
5380 * if header_prediction is to be made
5381 * 'S' will always be tp->tcp_header_len >> 2
5382 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5383 * turn it off (when there are holes in the receive
5384 * space for instance)
5385 * PSH flag is ignored.
5388 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5389 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5390 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5391 int tcp_header_len
= tp
->tcp_header_len
;
5393 /* Timestamp header prediction: tcp_header_len
5394 * is automatically equal to th->doff*4 due to pred_flags
5398 /* Check timestamp */
5399 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5400 /* No? Slow path! */
5401 if (!tcp_parse_aligned_timestamp(tp
, th
))
5404 /* If PAWS failed, check it more carefully in slow path */
5405 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5408 /* DO NOT update ts_recent here, if checksum fails
5409 * and timestamp was corrupted part, it will result
5410 * in a hung connection since we will drop all
5411 * future packets due to the PAWS test.
5415 if (len
<= tcp_header_len
) {
5416 /* Bulk data transfer: sender */
5417 if (len
== tcp_header_len
) {
5418 /* Predicted packet is in window by definition.
5419 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5420 * Hence, check seq<=rcv_wup reduces to:
5422 if (tcp_header_len
==
5423 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5424 tp
->rcv_nxt
== tp
->rcv_wup
)
5425 tcp_store_ts_recent(tp
);
5427 /* We know that such packets are checksummed
5430 tcp_ack(sk
, skb
, 0);
5432 tcp_data_snd_check(sk
);
5434 } else { /* Header too small */
5435 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5440 bool fragstolen
= false;
5442 if (tcp_checksum_complete(skb
))
5445 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5448 /* Predicted packet is in window by definition.
5449 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5450 * Hence, check seq<=rcv_wup reduces to:
5452 if (tcp_header_len
==
5453 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5454 tp
->rcv_nxt
== tp
->rcv_wup
)
5455 tcp_store_ts_recent(tp
);
5457 tcp_rcv_rtt_measure_ts(sk
, skb
);
5459 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5461 /* Bulk data transfer: receiver */
5462 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5465 tcp_event_data_recv(sk
, skb
);
5467 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5468 /* Well, only one small jumplet in fast path... */
5469 tcp_ack(sk
, skb
, FLAG_DATA
);
5470 tcp_data_snd_check(sk
);
5471 if (!inet_csk_ack_scheduled(sk
))
5475 __tcp_ack_snd_check(sk
, 0);
5478 kfree_skb_partial(skb
, fragstolen
);
5479 sk
->sk_data_ready(sk
);
5485 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5488 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5492 * Standard slow path.
5495 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5499 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5502 tcp_rcv_rtt_measure_ts(sk
, skb
);
5504 /* Process urgent data. */
5505 tcp_urg(sk
, skb
, th
);
5507 /* step 7: process the segment text */
5508 tcp_data_queue(sk
, skb
);
5510 tcp_data_snd_check(sk
);
5511 tcp_ack_snd_check(sk
);
5515 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5516 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5521 EXPORT_SYMBOL(tcp_rcv_established
);
5523 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5525 struct tcp_sock
*tp
= tcp_sk(sk
);
5526 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5528 tcp_set_state(sk
, TCP_ESTABLISHED
);
5529 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5532 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5533 security_inet_conn_established(sk
, skb
);
5536 /* Make sure socket is routed, for correct metrics. */
5537 icsk
->icsk_af_ops
->rebuild_header(sk
);
5539 tcp_init_metrics(sk
);
5540 tcp_call_bpf(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5541 tcp_init_congestion_control(sk
);
5543 /* Prevent spurious tcp_cwnd_restart() on first data
5546 tp
->lsndtime
= tcp_jiffies32
;
5548 tcp_init_buffer_space(sk
);
5550 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5551 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5553 if (!tp
->rx_opt
.snd_wscale
)
5554 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5559 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5560 struct tcp_fastopen_cookie
*cookie
)
5562 struct tcp_sock
*tp
= tcp_sk(sk
);
5563 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5564 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5565 bool syn_drop
= false;
5567 if (mss
== tp
->rx_opt
.user_mss
) {
5568 struct tcp_options_received opt
;
5570 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5571 tcp_clear_options(&opt
);
5572 opt
.user_mss
= opt
.mss_clamp
= 0;
5573 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5574 mss
= opt
.mss_clamp
;
5577 if (!tp
->syn_fastopen
) {
5578 /* Ignore an unsolicited cookie */
5580 } else if (tp
->total_retrans
) {
5581 /* SYN timed out and the SYN-ACK neither has a cookie nor
5582 * acknowledges data. Presumably the remote received only
5583 * the retransmitted (regular) SYNs: either the original
5584 * SYN-data or the corresponding SYN-ACK was dropped.
5586 syn_drop
= (cookie
->len
< 0 && data
);
5587 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5588 /* We requested a cookie but didn't get it. If we did not use
5589 * the (old) exp opt format then try so next time (try_exp=1).
5590 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5592 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5595 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5597 if (data
) { /* Retransmit unacked data in SYN */
5598 tcp_for_write_queue_from(data
, sk
) {
5599 if (data
== tcp_send_head(sk
) ||
5600 __tcp_retransmit_skb(sk
, data
, 1))
5604 NET_INC_STATS(sock_net(sk
),
5605 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5608 tp
->syn_data_acked
= tp
->syn_data
;
5609 if (tp
->syn_data_acked
)
5610 NET_INC_STATS(sock_net(sk
),
5611 LINUX_MIB_TCPFASTOPENACTIVE
);
5613 tcp_fastopen_add_skb(sk
, synack
);
5618 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5619 const struct tcphdr
*th
)
5621 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5622 struct tcp_sock
*tp
= tcp_sk(sk
);
5623 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5624 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5627 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5628 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5629 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5633 * "If the state is SYN-SENT then
5634 * first check the ACK bit
5635 * If the ACK bit is set
5636 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5637 * a reset (unless the RST bit is set, if so drop
5638 * the segment and return)"
5640 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5641 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5642 goto reset_and_undo
;
5644 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5645 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5646 tcp_time_stamp(tp
))) {
5647 NET_INC_STATS(sock_net(sk
),
5648 LINUX_MIB_PAWSACTIVEREJECTED
);
5649 goto reset_and_undo
;
5652 /* Now ACK is acceptable.
5654 * "If the RST bit is set
5655 * If the ACK was acceptable then signal the user "error:
5656 * connection reset", drop the segment, enter CLOSED state,
5657 * delete TCB, and return."
5666 * "fifth, if neither of the SYN or RST bits is set then
5667 * drop the segment and return."
5673 goto discard_and_undo
;
5676 * "If the SYN bit is on ...
5677 * are acceptable then ...
5678 * (our SYN has been ACKed), change the connection
5679 * state to ESTABLISHED..."
5682 tcp_ecn_rcv_synack(tp
, th
);
5684 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5685 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5687 /* Ok.. it's good. Set up sequence numbers and
5688 * move to established.
5690 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5691 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5693 /* RFC1323: The window in SYN & SYN/ACK segments is
5696 tp
->snd_wnd
= ntohs(th
->window
);
5698 if (!tp
->rx_opt
.wscale_ok
) {
5699 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5700 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5703 if (tp
->rx_opt
.saw_tstamp
) {
5704 tp
->rx_opt
.tstamp_ok
= 1;
5705 tp
->tcp_header_len
=
5706 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5707 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5708 tcp_store_ts_recent(tp
);
5710 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5713 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5714 tcp_enable_fack(tp
);
5717 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5718 tcp_initialize_rcv_mss(sk
);
5720 /* Remember, tcp_poll() does not lock socket!
5721 * Change state from SYN-SENT only after copied_seq
5722 * is initialized. */
5723 tp
->copied_seq
= tp
->rcv_nxt
;
5727 tcp_finish_connect(sk
, skb
);
5729 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5730 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5732 if (!sock_flag(sk
, SOCK_DEAD
)) {
5733 sk
->sk_state_change(sk
);
5734 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5738 if (sk
->sk_write_pending
||
5739 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5740 icsk
->icsk_ack
.pingpong
) {
5741 /* Save one ACK. Data will be ready after
5742 * several ticks, if write_pending is set.
5744 * It may be deleted, but with this feature tcpdumps
5745 * look so _wonderfully_ clever, that I was not able
5746 * to stand against the temptation 8) --ANK
5748 inet_csk_schedule_ack(sk
);
5749 tcp_enter_quickack_mode(sk
);
5750 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5751 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5762 /* No ACK in the segment */
5766 * "If the RST bit is set
5768 * Otherwise (no ACK) drop the segment and return."
5771 goto discard_and_undo
;
5775 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5776 tcp_paws_reject(&tp
->rx_opt
, 0))
5777 goto discard_and_undo
;
5780 /* We see SYN without ACK. It is attempt of
5781 * simultaneous connect with crossed SYNs.
5782 * Particularly, it can be connect to self.
5784 tcp_set_state(sk
, TCP_SYN_RECV
);
5786 if (tp
->rx_opt
.saw_tstamp
) {
5787 tp
->rx_opt
.tstamp_ok
= 1;
5788 tcp_store_ts_recent(tp
);
5789 tp
->tcp_header_len
=
5790 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5792 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5795 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5796 tp
->copied_seq
= tp
->rcv_nxt
;
5797 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5799 /* RFC1323: The window in SYN & SYN/ACK segments is
5802 tp
->snd_wnd
= ntohs(th
->window
);
5803 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5804 tp
->max_window
= tp
->snd_wnd
;
5806 tcp_ecn_rcv_syn(tp
, th
);
5809 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5810 tcp_initialize_rcv_mss(sk
);
5812 tcp_send_synack(sk
);
5814 /* Note, we could accept data and URG from this segment.
5815 * There are no obstacles to make this (except that we must
5816 * either change tcp_recvmsg() to prevent it from returning data
5817 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5819 * However, if we ignore data in ACKless segments sometimes,
5820 * we have no reasons to accept it sometimes.
5821 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5822 * is not flawless. So, discard packet for sanity.
5823 * Uncomment this return to process the data.
5830 /* "fifth, if neither of the SYN or RST bits is set then
5831 * drop the segment and return."
5835 tcp_clear_options(&tp
->rx_opt
);
5836 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5840 tcp_clear_options(&tp
->rx_opt
);
5841 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5846 * This function implements the receiving procedure of RFC 793 for
5847 * all states except ESTABLISHED and TIME_WAIT.
5848 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5849 * address independent.
5852 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5854 struct tcp_sock
*tp
= tcp_sk(sk
);
5855 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5856 const struct tcphdr
*th
= tcp_hdr(skb
);
5857 struct request_sock
*req
;
5861 switch (sk
->sk_state
) {
5875 /* It is possible that we process SYN packets from backlog,
5876 * so we need to make sure to disable BH right there.
5879 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5890 tp
->rx_opt
.saw_tstamp
= 0;
5891 tcp_mstamp_refresh(tp
);
5892 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5896 /* Do step6 onward by hand. */
5897 tcp_urg(sk
, skb
, th
);
5899 tcp_data_snd_check(sk
);
5903 tcp_mstamp_refresh(tp
);
5904 tp
->rx_opt
.saw_tstamp
= 0;
5905 req
= tp
->fastopen_rsk
;
5907 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5908 sk
->sk_state
!= TCP_FIN_WAIT1
);
5910 if (!tcp_check_req(sk
, skb
, req
, true))
5914 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5917 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5920 /* step 5: check the ACK field */
5921 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5922 FLAG_UPDATE_TS_RECENT
|
5923 FLAG_NO_CHALLENGE_ACK
) > 0;
5926 if (sk
->sk_state
== TCP_SYN_RECV
)
5927 return 1; /* send one RST */
5928 tcp_send_challenge_ack(sk
, skb
);
5931 switch (sk
->sk_state
) {
5934 tcp_synack_rtt_meas(sk
, req
);
5936 /* Once we leave TCP_SYN_RECV, we no longer need req
5940 inet_csk(sk
)->icsk_retransmits
= 0;
5941 reqsk_fastopen_remove(sk
, req
, false);
5943 /* Make sure socket is routed, for correct metrics. */
5944 icsk
->icsk_af_ops
->rebuild_header(sk
);
5945 tcp_call_bpf(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
5946 tcp_init_congestion_control(sk
);
5949 tp
->copied_seq
= tp
->rcv_nxt
;
5950 tcp_init_buffer_space(sk
);
5953 tcp_set_state(sk
, TCP_ESTABLISHED
);
5954 sk
->sk_state_change(sk
);
5956 /* Note, that this wakeup is only for marginal crossed SYN case.
5957 * Passively open sockets are not waked up, because
5958 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5961 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5963 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5964 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5965 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5967 if (tp
->rx_opt
.tstamp_ok
)
5968 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5971 /* Re-arm the timer because data may have been sent out.
5972 * This is similar to the regular data transmission case
5973 * when new data has just been ack'ed.
5975 * (TFO) - we could try to be more aggressive and
5976 * retransmitting any data sooner based on when they
5981 tcp_init_metrics(sk
);
5983 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
5984 tcp_update_pacing_rate(sk
);
5986 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5987 tp
->lsndtime
= tcp_jiffies32
;
5989 tcp_initialize_rcv_mss(sk
);
5990 tcp_fast_path_on(tp
);
5993 case TCP_FIN_WAIT1
: {
5996 /* If we enter the TCP_FIN_WAIT1 state and we are a
5997 * Fast Open socket and this is the first acceptable
5998 * ACK we have received, this would have acknowledged
5999 * our SYNACK so stop the SYNACK timer.
6002 /* We no longer need the request sock. */
6003 reqsk_fastopen_remove(sk
, req
, false);
6006 if (tp
->snd_una
!= tp
->write_seq
)
6009 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6010 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6014 if (!sock_flag(sk
, SOCK_DEAD
)) {
6015 /* Wake up lingering close() */
6016 sk
->sk_state_change(sk
);
6020 if (tp
->linger2
< 0) {
6022 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6025 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6026 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6027 /* Receive out of order FIN after close() */
6028 if (tp
->syn_fastopen
&& th
->fin
)
6029 tcp_fastopen_active_disable(sk
);
6031 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6035 tmo
= tcp_fin_time(sk
);
6036 if (tmo
> TCP_TIMEWAIT_LEN
) {
6037 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6038 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6039 /* Bad case. We could lose such FIN otherwise.
6040 * It is not a big problem, but it looks confusing
6041 * and not so rare event. We still can lose it now,
6042 * if it spins in bh_lock_sock(), but it is really
6045 inet_csk_reset_keepalive_timer(sk
, tmo
);
6047 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6054 if (tp
->snd_una
== tp
->write_seq
) {
6055 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6061 if (tp
->snd_una
== tp
->write_seq
) {
6062 tcp_update_metrics(sk
);
6069 /* step 6: check the URG bit */
6070 tcp_urg(sk
, skb
, th
);
6072 /* step 7: process the segment text */
6073 switch (sk
->sk_state
) {
6074 case TCP_CLOSE_WAIT
:
6077 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6081 /* RFC 793 says to queue data in these states,
6082 * RFC 1122 says we MUST send a reset.
6083 * BSD 4.4 also does reset.
6085 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6086 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6087 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6088 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6094 case TCP_ESTABLISHED
:
6095 tcp_data_queue(sk
, skb
);
6100 /* tcp_data could move socket to TIME-WAIT */
6101 if (sk
->sk_state
!= TCP_CLOSE
) {
6102 tcp_data_snd_check(sk
);
6103 tcp_ack_snd_check(sk
);
6112 EXPORT_SYMBOL(tcp_rcv_state_process
);
6114 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6116 struct inet_request_sock
*ireq
= inet_rsk(req
);
6118 if (family
== AF_INET
)
6119 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6120 &ireq
->ir_rmt_addr
, port
);
6121 #if IS_ENABLED(CONFIG_IPV6)
6122 else if (family
== AF_INET6
)
6123 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6124 &ireq
->ir_v6_rmt_addr
, port
);
6128 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6130 * If we receive a SYN packet with these bits set, it means a
6131 * network is playing bad games with TOS bits. In order to
6132 * avoid possible false congestion notifications, we disable
6133 * TCP ECN negotiation.
6135 * Exception: tcp_ca wants ECN. This is required for DCTCP
6136 * congestion control: Linux DCTCP asserts ECT on all packets,
6137 * including SYN, which is most optimal solution; however,
6138 * others, such as FreeBSD do not.
6140 static void tcp_ecn_create_request(struct request_sock
*req
,
6141 const struct sk_buff
*skb
,
6142 const struct sock
*listen_sk
,
6143 const struct dst_entry
*dst
)
6145 const struct tcphdr
*th
= tcp_hdr(skb
);
6146 const struct net
*net
= sock_net(listen_sk
);
6147 bool th_ecn
= th
->ece
&& th
->cwr
;
6154 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6155 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6156 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6158 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6159 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6160 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6161 inet_rsk(req
)->ecn_ok
= 1;
6164 static void tcp_openreq_init(struct request_sock
*req
,
6165 const struct tcp_options_received
*rx_opt
,
6166 struct sk_buff
*skb
, const struct sock
*sk
)
6168 struct inet_request_sock
*ireq
= inet_rsk(req
);
6170 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6172 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6173 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6174 tcp_rsk(req
)->snt_synack
= tcp_clock_us();
6175 tcp_rsk(req
)->last_oow_ack_time
= 0;
6176 req
->mss
= rx_opt
->mss_clamp
;
6177 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6178 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6179 ireq
->sack_ok
= rx_opt
->sack_ok
;
6180 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6181 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6184 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6185 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6186 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6189 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6190 struct sock
*sk_listener
,
6191 bool attach_listener
)
6193 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6197 struct inet_request_sock
*ireq
= inet_rsk(req
);
6199 kmemcheck_annotate_bitfield(ireq
, flags
);
6200 ireq
->ireq_opt
= NULL
;
6201 #if IS_ENABLED(CONFIG_IPV6)
6202 ireq
->pktopts
= NULL
;
6204 atomic64_set(&ireq
->ir_cookie
, 0);
6205 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6206 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6207 ireq
->ireq_family
= sk_listener
->sk_family
;
6212 EXPORT_SYMBOL(inet_reqsk_alloc
);
6215 * Return true if a syncookie should be sent
6217 static bool tcp_syn_flood_action(const struct sock
*sk
,
6218 const struct sk_buff
*skb
,
6221 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6222 const char *msg
= "Dropping request";
6223 bool want_cookie
= false;
6224 struct net
*net
= sock_net(sk
);
6226 #ifdef CONFIG_SYN_COOKIES
6227 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6228 msg
= "Sending cookies";
6230 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6233 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6235 if (!queue
->synflood_warned
&&
6236 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6237 xchg(&queue
->synflood_warned
, 1) == 0)
6238 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6239 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6244 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6245 struct request_sock
*req
,
6246 const struct sk_buff
*skb
)
6248 if (tcp_sk(sk
)->save_syn
) {
6249 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6252 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6255 memcpy(©
[1], skb_network_header(skb
), len
);
6256 req
->saved_syn
= copy
;
6261 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6262 const struct tcp_request_sock_ops
*af_ops
,
6263 struct sock
*sk
, struct sk_buff
*skb
)
6265 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6266 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6267 struct tcp_options_received tmp_opt
;
6268 struct tcp_sock
*tp
= tcp_sk(sk
);
6269 struct net
*net
= sock_net(sk
);
6270 struct sock
*fastopen_sk
= NULL
;
6271 struct request_sock
*req
;
6272 bool want_cookie
= false;
6273 struct dst_entry
*dst
;
6276 /* TW buckets are converted to open requests without
6277 * limitations, they conserve resources and peer is
6278 * evidently real one.
6280 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6281 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6282 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6287 if (sk_acceptq_is_full(sk
)) {
6288 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6292 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6296 tcp_rsk(req
)->af_specific
= af_ops
;
6297 tcp_rsk(req
)->ts_off
= 0;
6299 tcp_clear_options(&tmp_opt
);
6300 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6301 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6302 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6303 want_cookie
? NULL
: &foc
);
6305 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6306 tcp_clear_options(&tmp_opt
);
6308 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6309 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6310 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6312 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6313 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6315 af_ops
->init_req(req
, sk
, skb
);
6317 if (security_inet_conn_request(sk
, skb
, req
))
6320 if (tmp_opt
.tstamp_ok
)
6321 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6323 dst
= af_ops
->route_req(sk
, &fl
, req
);
6327 if (!want_cookie
&& !isn
) {
6328 /* Kill the following clause, if you dislike this way. */
6329 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6330 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6331 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6332 !tcp_peer_is_proven(req
, dst
)) {
6333 /* Without syncookies last quarter of
6334 * backlog is filled with destinations,
6335 * proven to be alive.
6336 * It means that we continue to communicate
6337 * to destinations, already remembered
6338 * to the moment of synflood.
6340 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6342 goto drop_and_release
;
6345 isn
= af_ops
->init_seq(skb
);
6348 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6351 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6352 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6353 if (!tmp_opt
.tstamp_ok
)
6354 inet_rsk(req
)->ecn_ok
= 0;
6357 tcp_rsk(req
)->snt_isn
= isn
;
6358 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6359 tcp_openreq_init_rwin(req
, sk
, dst
);
6361 tcp_reqsk_record_syn(sk
, req
, skb
);
6362 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
);
6365 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6366 &foc
, TCP_SYNACK_FASTOPEN
);
6367 /* Add the child socket directly into the accept queue */
6368 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6369 sk
->sk_data_ready(sk
);
6370 bh_unlock_sock(fastopen_sk
);
6371 sock_put(fastopen_sk
);
6373 tcp_rsk(req
)->tfo_listener
= false;
6375 inet_csk_reqsk_queue_hash_add(sk
, req
,
6376 tcp_timeout_init((struct sock
*)req
));
6377 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6378 !want_cookie
? TCP_SYNACK_NORMAL
:
6396 EXPORT_SYMBOL(tcp_conn_request
);