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;
98 int sysctl_tcp_default_init_rwnd __read_mostly
= TCP_INIT_CWND
* 2;
100 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
101 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
102 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
103 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
104 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
105 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
106 #define FLAG_ECE 0x40 /* ECE in this ACK */
107 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
108 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
109 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
110 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
111 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
112 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
115 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 #define REXMIT_NONE 0 /* no loss recovery to do */
126 #define REXMIT_LOST 1 /* retransmit packets marked lost */
127 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
129 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
132 static bool __once __read_mostly
;
135 struct net_device
*dev
;
140 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
141 if (!dev
|| len
>= dev
->mtu
)
142 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
143 dev
? dev
->name
: "Unknown driver");
148 /* Adapt the MSS value used to make delayed ack decision to the
151 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
154 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
157 icsk
->icsk_ack
.last_seg_size
= 0;
159 /* skb->len may jitter because of SACKs, even if peer
160 * sends good full-sized frames.
162 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
163 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
164 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
166 /* Account for possibly-removed options */
167 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
168 MAX_TCP_OPTION_SPACE
))
169 tcp_gro_dev_warn(sk
, skb
, len
);
171 /* Otherwise, we make more careful check taking into account,
172 * that SACKs block is variable.
174 * "len" is invariant segment length, including TCP header.
176 len
+= skb
->data
- skb_transport_header(skb
);
177 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
178 /* If PSH is not set, packet should be
179 * full sized, provided peer TCP is not badly broken.
180 * This observation (if it is correct 8)) allows
181 * to handle super-low mtu links fairly.
183 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
184 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
185 /* Subtract also invariant (if peer is RFC compliant),
186 * tcp header plus fixed timestamp option length.
187 * Resulting "len" is MSS free of SACK jitter.
189 len
-= tcp_sk(sk
)->tcp_header_len
;
190 icsk
->icsk_ack
.last_seg_size
= len
;
192 icsk
->icsk_ack
.rcv_mss
= len
;
196 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
197 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
198 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
202 static void tcp_incr_quickack(struct sock
*sk
, unsigned int max_quickacks
)
204 struct inet_connection_sock
*icsk
= inet_csk(sk
);
205 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
209 quickacks
= min(quickacks
, max_quickacks
);
210 if (quickacks
> icsk
->icsk_ack
.quick
)
211 icsk
->icsk_ack
.quick
= quickacks
;
214 void tcp_enter_quickack_mode(struct sock
*sk
, unsigned int max_quickacks
)
216 struct inet_connection_sock
*icsk
= inet_csk(sk
);
218 tcp_incr_quickack(sk
, max_quickacks
);
219 icsk
->icsk_ack
.pingpong
= 0;
220 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
222 EXPORT_SYMBOL(tcp_enter_quickack_mode
);
224 /* Send ACKs quickly, if "quick" count is not exhausted
225 * and the session is not interactive.
228 static bool tcp_in_quickack_mode(struct sock
*sk
)
230 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
231 const struct dst_entry
*dst
= __sk_dst_get(sk
);
233 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
234 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
237 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
239 if (tp
->ecn_flags
& TCP_ECN_OK
)
240 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
243 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
245 if (tcp_hdr(skb
)->cwr
)
246 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
249 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
251 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
254 static void __tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
256 struct tcp_sock
*tp
= tcp_sk(sk
);
258 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
259 case INET_ECN_NOT_ECT
:
260 /* Funny extension: if ECT is not set on a segment,
261 * and we already seen ECT on a previous segment,
262 * it is probably a retransmit.
264 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
265 tcp_enter_quickack_mode(sk
, 2);
268 if (tcp_ca_needs_ecn(sk
))
269 tcp_ca_event(sk
, CA_EVENT_ECN_IS_CE
);
271 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
272 /* Better not delay acks, sender can have a very low cwnd */
273 tcp_enter_quickack_mode(sk
, 2);
274 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
276 tp
->ecn_flags
|= TCP_ECN_SEEN
;
279 if (tcp_ca_needs_ecn(sk
))
280 tcp_ca_event(sk
, CA_EVENT_ECN_NO_CE
);
281 tp
->ecn_flags
|= TCP_ECN_SEEN
;
286 static void tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
288 if (tcp_sk(sk
)->ecn_flags
& TCP_ECN_OK
)
289 __tcp_ecn_check_ce(sk
, skb
);
292 static void tcp_ecn_rcv_synack(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 void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
300 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
301 tp
->ecn_flags
&= ~TCP_ECN_OK
;
304 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
306 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
311 /* Buffer size and advertised window tuning.
313 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
316 static void tcp_sndbuf_expand(struct sock
*sk
)
318 const struct tcp_sock
*tp
= tcp_sk(sk
);
319 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
323 /* Worst case is non GSO/TSO : each frame consumes one skb
324 * and skb->head is kmalloced using power of two area of memory
326 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
328 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
330 per_mss
= roundup_pow_of_two(per_mss
) +
331 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
333 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
334 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
336 /* Fast Recovery (RFC 5681 3.2) :
337 * Cubic needs 1.7 factor, rounded to 2 to include
338 * extra cushion (application might react slowly to POLLOUT)
340 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
341 sndmem
*= nr_segs
* per_mss
;
343 if (sk
->sk_sndbuf
< sndmem
)
344 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
347 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
349 * All tcp_full_space() is split to two parts: "network" buffer, allocated
350 * forward and advertised in receiver window (tp->rcv_wnd) and
351 * "application buffer", required to isolate scheduling/application
352 * latencies from network.
353 * window_clamp is maximal advertised window. It can be less than
354 * tcp_full_space(), in this case tcp_full_space() - window_clamp
355 * is reserved for "application" buffer. The less window_clamp is
356 * the smoother our behaviour from viewpoint of network, but the lower
357 * throughput and the higher sensitivity of the connection to losses. 8)
359 * rcv_ssthresh is more strict window_clamp used at "slow start"
360 * phase to predict further behaviour of this connection.
361 * It is used for two goals:
362 * - to enforce header prediction at sender, even when application
363 * requires some significant "application buffer". It is check #1.
364 * - to prevent pruning of receive queue because of misprediction
365 * of receiver window. Check #2.
367 * The scheme does not work when sender sends good segments opening
368 * window and then starts to feed us spaghetti. But it should work
369 * in common situations. Otherwise, we have to rely on queue collapsing.
372 /* Slow part of check#2. */
373 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
375 struct tcp_sock
*tp
= tcp_sk(sk
);
377 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
378 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
380 while (tp
->rcv_ssthresh
<= window
) {
381 if (truesize
<= skb
->len
)
382 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
390 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
392 struct tcp_sock
*tp
= tcp_sk(sk
);
395 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
396 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
397 !tcp_under_memory_pressure(sk
)) {
400 /* Check #2. Increase window, if skb with such overhead
401 * will fit to rcvbuf in future.
403 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
404 incr
= 2 * tp
->advmss
;
406 incr
= __tcp_grow_window(sk
, skb
);
409 incr
= max_t(int, incr
, 2 * skb
->len
);
410 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
412 inet_csk(sk
)->icsk_ack
.quick
|= 1;
417 /* 3. Tuning rcvbuf, when connection enters established state. */
418 static void tcp_fixup_rcvbuf(struct sock
*sk
)
420 u32 mss
= tcp_sk(sk
)->advmss
;
423 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
424 tcp_default_init_rwnd(mss
);
426 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
427 * Allow enough cushion so that sender is not limited by our window
429 if (sysctl_tcp_moderate_rcvbuf
)
432 if (sk
->sk_rcvbuf
< rcvmem
)
433 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
436 /* 4. Try to fixup all. It is made immediately after connection enters
439 void tcp_init_buffer_space(struct sock
*sk
)
441 struct tcp_sock
*tp
= tcp_sk(sk
);
444 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
445 tcp_fixup_rcvbuf(sk
);
446 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
447 tcp_sndbuf_expand(sk
);
449 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
450 tcp_mstamp_refresh(tp
);
451 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
452 tp
->rcvq_space
.seq
= tp
->copied_seq
;
454 maxwin
= tcp_full_space(sk
);
456 if (tp
->window_clamp
>= maxwin
) {
457 tp
->window_clamp
= maxwin
;
459 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
460 tp
->window_clamp
= max(maxwin
-
461 (maxwin
>> sysctl_tcp_app_win
),
465 /* Force reservation of one segment. */
466 if (sysctl_tcp_app_win
&&
467 tp
->window_clamp
> 2 * tp
->advmss
&&
468 tp
->window_clamp
+ tp
->advmss
> maxwin
)
469 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
471 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
472 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
475 /* 5. Recalculate window clamp after socket hit its memory bounds. */
476 static void tcp_clamp_window(struct sock
*sk
)
478 struct tcp_sock
*tp
= tcp_sk(sk
);
479 struct inet_connection_sock
*icsk
= inet_csk(sk
);
481 icsk
->icsk_ack
.quick
= 0;
483 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
484 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
485 !tcp_under_memory_pressure(sk
) &&
486 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
487 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
490 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
491 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
494 /* Initialize RCV_MSS value.
495 * RCV_MSS is an our guess about MSS used by the peer.
496 * We haven't any direct information about the MSS.
497 * It's better to underestimate the RCV_MSS rather than overestimate.
498 * Overestimations make us ACKing less frequently than needed.
499 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
501 void tcp_initialize_rcv_mss(struct sock
*sk
)
503 const struct tcp_sock
*tp
= tcp_sk(sk
);
504 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
506 hint
= min(hint
, tp
->rcv_wnd
/ 2);
507 hint
= min(hint
, TCP_MSS_DEFAULT
);
508 hint
= max(hint
, TCP_MIN_MSS
);
510 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
512 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
514 /* Receiver "autotuning" code.
516 * The algorithm for RTT estimation w/o timestamps is based on
517 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
518 * <http://public.lanl.gov/radiant/pubs.html#DRS>
520 * More detail on this code can be found at
521 * <http://staff.psc.edu/jheffner/>,
522 * though this reference is out of date. A new paper
525 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
527 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
530 if (new_sample
!= 0) {
531 /* If we sample in larger samples in the non-timestamp
532 * case, we could grossly overestimate the RTT especially
533 * with chatty applications or bulk transfer apps which
534 * are stalled on filesystem I/O.
536 * Also, since we are only going for a minimum in the
537 * non-timestamp case, we do not smooth things out
538 * else with timestamps disabled convergence takes too
542 m
-= (new_sample
>> 3);
550 /* No previous measure. */
554 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
557 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
561 if (tp
->rcv_rtt_est
.time
== 0)
563 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
565 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
568 tcp_rcv_rtt_update(tp
, delta_us
, 1);
571 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
572 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
575 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
576 const struct sk_buff
*skb
)
578 struct tcp_sock
*tp
= tcp_sk(sk
);
580 if (tp
->rx_opt
.rcv_tsecr
&&
581 (TCP_SKB_CB(skb
)->end_seq
-
582 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
)) {
583 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
588 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
589 tcp_rcv_rtt_update(tp
, delta_us
, 0);
594 * This function should be called every time data is copied to user space.
595 * It calculates the appropriate TCP receive buffer space.
597 void tcp_rcv_space_adjust(struct sock
*sk
)
599 struct tcp_sock
*tp
= tcp_sk(sk
);
603 tcp_mstamp_refresh(tp
);
604 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
605 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
608 /* Number of bytes copied to user in last RTT */
609 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
610 if (copied
<= tp
->rcvq_space
.space
)
614 * copied = bytes received in previous RTT, our base window
615 * To cope with packet losses, we need a 2x factor
616 * To cope with slow start, and sender growing its cwin by 100 %
617 * every RTT, we need a 4x factor, because the ACK we are sending
618 * now is for the next RTT, not the current one :
619 * <prev RTT . ><current RTT .. ><next RTT .... >
622 if (sysctl_tcp_moderate_rcvbuf
&&
623 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
627 /* minimal window to cope with packet losses, assuming
628 * steady state. Add some cushion because of small variations.
630 rcvwin
= ((u64
)copied
<< 1) + 16 * tp
->advmss
;
632 /* If rate increased by 25%,
633 * assume slow start, rcvwin = 3 * copied
634 * If rate increased by 50%,
635 * assume sender can use 2x growth, rcvwin = 4 * copied
638 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
640 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
643 rcvwin
+= (rcvwin
>> 1);
646 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
647 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
650 do_div(rcvwin
, tp
->advmss
);
651 rcvbuf
= min_t(u64
, rcvwin
* rcvmem
, sysctl_tcp_rmem
[2]);
652 if (rcvbuf
> sk
->sk_rcvbuf
) {
653 sk
->sk_rcvbuf
= rcvbuf
;
655 /* Make the window clamp follow along. */
656 tp
->window_clamp
= tcp_win_from_space(rcvbuf
);
659 tp
->rcvq_space
.space
= copied
;
662 tp
->rcvq_space
.seq
= tp
->copied_seq
;
663 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
666 /* There is something which you must keep in mind when you analyze the
667 * behavior of the tp->ato delayed ack timeout interval. When a
668 * connection starts up, we want to ack as quickly as possible. The
669 * problem is that "good" TCP's do slow start at the beginning of data
670 * transmission. The means that until we send the first few ACK's the
671 * sender will sit on his end and only queue most of his data, because
672 * he can only send snd_cwnd unacked packets at any given time. For
673 * each ACK we send, he increments snd_cwnd and transmits more of his
676 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
678 struct tcp_sock
*tp
= tcp_sk(sk
);
679 struct inet_connection_sock
*icsk
= inet_csk(sk
);
682 inet_csk_schedule_ack(sk
);
684 tcp_measure_rcv_mss(sk
, skb
);
686 tcp_rcv_rtt_measure(tp
);
690 if (!icsk
->icsk_ack
.ato
) {
691 /* The _first_ data packet received, initialize
692 * delayed ACK engine.
694 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
695 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
697 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
699 if (m
<= TCP_ATO_MIN
/ 2) {
700 /* The fastest case is the first. */
701 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
702 } else if (m
< icsk
->icsk_ack
.ato
) {
703 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
704 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
705 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
706 } else if (m
> icsk
->icsk_rto
) {
707 /* Too long gap. Apparently sender failed to
708 * restart window, so that we send ACKs quickly.
710 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
714 icsk
->icsk_ack
.lrcvtime
= now
;
716 tcp_ecn_check_ce(sk
, skb
);
719 tcp_grow_window(sk
, skb
);
722 /* Called to compute a smoothed rtt estimate. The data fed to this
723 * routine either comes from timestamps, or from segments that were
724 * known _not_ to have been retransmitted [see Karn/Partridge
725 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
726 * piece by Van Jacobson.
727 * NOTE: the next three routines used to be one big routine.
728 * To save cycles in the RFC 1323 implementation it was better to break
729 * it up into three procedures. -- erics
731 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
733 struct tcp_sock
*tp
= tcp_sk(sk
);
734 long m
= mrtt_us
; /* RTT */
735 u32 srtt
= tp
->srtt_us
;
737 /* The following amusing code comes from Jacobson's
738 * article in SIGCOMM '88. Note that rtt and mdev
739 * are scaled versions of rtt and mean deviation.
740 * This is designed to be as fast as possible
741 * m stands for "measurement".
743 * On a 1990 paper the rto value is changed to:
744 * RTO = rtt + 4 * mdev
746 * Funny. This algorithm seems to be very broken.
747 * These formulae increase RTO, when it should be decreased, increase
748 * too slowly, when it should be increased quickly, decrease too quickly
749 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
750 * does not matter how to _calculate_ it. Seems, it was trap
751 * that VJ failed to avoid. 8)
754 m
-= (srtt
>> 3); /* m is now error in rtt est */
755 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
757 m
= -m
; /* m is now abs(error) */
758 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
759 /* This is similar to one of Eifel findings.
760 * Eifel blocks mdev updates when rtt decreases.
761 * This solution is a bit different: we use finer gain
762 * for mdev in this case (alpha*beta).
763 * Like Eifel it also prevents growth of rto,
764 * but also it limits too fast rto decreases,
765 * happening in pure Eifel.
770 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
772 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
773 if (tp
->mdev_us
> tp
->mdev_max_us
) {
774 tp
->mdev_max_us
= tp
->mdev_us
;
775 if (tp
->mdev_max_us
> tp
->rttvar_us
)
776 tp
->rttvar_us
= tp
->mdev_max_us
;
778 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
779 if (tp
->mdev_max_us
< tp
->rttvar_us
)
780 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
781 tp
->rtt_seq
= tp
->snd_nxt
;
782 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
785 /* no previous measure. */
786 srtt
= m
<< 3; /* take the measured time to be rtt */
787 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
788 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
789 tp
->mdev_max_us
= tp
->rttvar_us
;
790 tp
->rtt_seq
= tp
->snd_nxt
;
792 tp
->srtt_us
= max(1U, srtt
);
795 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
796 * Note: TCP stack does not yet implement pacing.
797 * FQ packet scheduler can be used to implement cheap but effective
798 * TCP pacing, to smooth the burst on large writes when packets
799 * in flight is significantly lower than cwnd (or rwin)
801 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
802 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
804 static void tcp_update_pacing_rate(struct sock
*sk
)
806 const struct tcp_sock
*tp
= tcp_sk(sk
);
809 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
810 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
812 /* current rate is (cwnd * mss) / srtt
813 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
814 * In Congestion Avoidance phase, set it to 120 % the current rate.
816 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
817 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
818 * end of slow start and should slow down.
820 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
821 rate
*= sysctl_tcp_pacing_ss_ratio
;
823 rate
*= sysctl_tcp_pacing_ca_ratio
;
825 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
827 if (likely(tp
->srtt_us
))
828 do_div(rate
, tp
->srtt_us
);
830 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
831 * without any lock. We want to make sure compiler wont store
832 * intermediate values in this location.
834 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
835 sk
->sk_max_pacing_rate
);
838 /* Calculate rto without backoff. This is the second half of Van Jacobson's
839 * routine referred to above.
841 static void tcp_set_rto(struct sock
*sk
)
843 const struct tcp_sock
*tp
= tcp_sk(sk
);
844 /* Old crap is replaced with new one. 8)
847 * 1. If rtt variance happened to be less 50msec, it is hallucination.
848 * It cannot be less due to utterly erratic ACK generation made
849 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
850 * to do with delayed acks, because at cwnd>2 true delack timeout
851 * is invisible. Actually, Linux-2.4 also generates erratic
852 * ACKs in some circumstances.
854 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
856 /* 2. Fixups made earlier cannot be right.
857 * If we do not estimate RTO correctly without them,
858 * all the algo is pure shit and should be replaced
859 * with correct one. It is exactly, which we pretend to do.
862 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
863 * guarantees that rto is higher.
868 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
870 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
873 cwnd
= TCP_INIT_CWND
;
874 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
878 * Packet counting of FACK is based on in-order assumptions, therefore TCP
879 * disables it when reordering is detected
881 void tcp_disable_fack(struct tcp_sock
*tp
)
883 /* RFC3517 uses different metric in lost marker => reset on change */
885 tp
->lost_skb_hint
= NULL
;
886 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
889 /* Take a notice that peer is sending D-SACKs */
890 static void tcp_dsack_seen(struct tcp_sock
*tp
)
892 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
895 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
898 struct tcp_sock
*tp
= tcp_sk(sk
);
901 if (WARN_ON_ONCE(metric
< 0))
904 if (metric
> tp
->reordering
) {
905 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
907 #if FASTRETRANS_DEBUG > 1
908 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
909 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
913 tp
->undo_marker
? tp
->undo_retrans
: 0);
915 tcp_disable_fack(tp
);
920 /* This exciting event is worth to be remembered. 8) */
922 mib_idx
= LINUX_MIB_TCPTSREORDER
;
923 else if (tcp_is_reno(tp
))
924 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
925 else if (tcp_is_fack(tp
))
926 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
928 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
930 NET_INC_STATS(sock_net(sk
), mib_idx
);
933 /* This must be called before lost_out is incremented */
934 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
936 if (!tp
->retransmit_skb_hint
||
937 before(TCP_SKB_CB(skb
)->seq
,
938 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
939 tp
->retransmit_skb_hint
= skb
;
942 /* Sum the number of packets on the wire we have marked as lost.
943 * There are two cases we care about here:
944 * a) Packet hasn't been marked lost (nor retransmitted),
945 * and this is the first loss.
946 * b) Packet has been marked both lost and retransmitted,
947 * and this means we think it was lost again.
949 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
951 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
953 if (!(sacked
& TCPCB_LOST
) ||
954 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
955 tp
->lost
+= tcp_skb_pcount(skb
);
958 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
960 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
961 tcp_verify_retransmit_hint(tp
, skb
);
963 tp
->lost_out
+= tcp_skb_pcount(skb
);
964 tcp_sum_lost(tp
, skb
);
965 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
969 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
971 tcp_verify_retransmit_hint(tp
, skb
);
973 tcp_sum_lost(tp
, skb
);
974 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
975 tp
->lost_out
+= tcp_skb_pcount(skb
);
976 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
980 /* This procedure tags the retransmission queue when SACKs arrive.
982 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
983 * Packets in queue with these bits set are counted in variables
984 * sacked_out, retrans_out and lost_out, correspondingly.
986 * Valid combinations are:
987 * Tag InFlight Description
988 * 0 1 - orig segment is in flight.
989 * S 0 - nothing flies, orig reached receiver.
990 * L 0 - nothing flies, orig lost by net.
991 * R 2 - both orig and retransmit are in flight.
992 * L|R 1 - orig is lost, retransmit is in flight.
993 * S|R 1 - orig reached receiver, retrans is still in flight.
994 * (L|S|R is logically valid, it could occur when L|R is sacked,
995 * but it is equivalent to plain S and code short-curcuits it to S.
996 * L|S is logically invalid, it would mean -1 packet in flight 8))
998 * These 6 states form finite state machine, controlled by the following events:
999 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1000 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1001 * 3. Loss detection event of two flavors:
1002 * A. Scoreboard estimator decided the packet is lost.
1003 * A'. Reno "three dupacks" marks head of queue lost.
1004 * A''. Its FACK modification, head until snd.fack is lost.
1005 * B. SACK arrives sacking SND.NXT at the moment, when the
1006 * segment was retransmitted.
1007 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1009 * It is pleasant to note, that state diagram turns out to be commutative,
1010 * so that we are allowed not to be bothered by order of our actions,
1011 * when multiple events arrive simultaneously. (see the function below).
1013 * Reordering detection.
1014 * --------------------
1015 * Reordering metric is maximal distance, which a packet can be displaced
1016 * in packet stream. With SACKs we can estimate it:
1018 * 1. SACK fills old hole and the corresponding segment was not
1019 * ever retransmitted -> reordering. Alas, we cannot use it
1020 * when segment was retransmitted.
1021 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1022 * for retransmitted and already SACKed segment -> reordering..
1023 * Both of these heuristics are not used in Loss state, when we cannot
1024 * account for retransmits accurately.
1026 * SACK block validation.
1027 * ----------------------
1029 * SACK block range validation checks that the received SACK block fits to
1030 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1031 * Note that SND.UNA is not included to the range though being valid because
1032 * it means that the receiver is rather inconsistent with itself reporting
1033 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1034 * perfectly valid, however, in light of RFC2018 which explicitly states
1035 * that "SACK block MUST reflect the newest segment. Even if the newest
1036 * segment is going to be discarded ...", not that it looks very clever
1037 * in case of head skb. Due to potentional receiver driven attacks, we
1038 * choose to avoid immediate execution of a walk in write queue due to
1039 * reneging and defer head skb's loss recovery to standard loss recovery
1040 * procedure that will eventually trigger (nothing forbids us doing this).
1042 * Implements also blockage to start_seq wrap-around. Problem lies in the
1043 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1044 * there's no guarantee that it will be before snd_nxt (n). The problem
1045 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1048 * <- outs wnd -> <- wrapzone ->
1049 * u e n u_w e_w s n_w
1051 * |<------------+------+----- TCP seqno space --------------+---------->|
1052 * ...-- <2^31 ->| |<--------...
1053 * ...---- >2^31 ------>| |<--------...
1055 * Current code wouldn't be vulnerable but it's better still to discard such
1056 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1057 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1058 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1059 * equal to the ideal case (infinite seqno space without wrap caused issues).
1061 * With D-SACK the lower bound is extended to cover sequence space below
1062 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1063 * again, D-SACK block must not to go across snd_una (for the same reason as
1064 * for the normal SACK blocks, explained above). But there all simplicity
1065 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1066 * fully below undo_marker they do not affect behavior in anyway and can
1067 * therefore be safely ignored. In rare cases (which are more or less
1068 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1069 * fragmentation and packet reordering past skb's retransmission. To consider
1070 * them correctly, the acceptable range must be extended even more though
1071 * the exact amount is rather hard to quantify. However, tp->max_window can
1072 * be used as an exaggerated estimate.
1074 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1075 u32 start_seq
, u32 end_seq
)
1077 /* Too far in future, or reversed (interpretation is ambiguous) */
1078 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1081 /* Nasty start_seq wrap-around check (see comments above) */
1082 if (!before(start_seq
, tp
->snd_nxt
))
1085 /* In outstanding window? ...This is valid exit for D-SACKs too.
1086 * start_seq == snd_una is non-sensical (see comments above)
1088 if (after(start_seq
, tp
->snd_una
))
1091 if (!is_dsack
|| !tp
->undo_marker
)
1094 /* ...Then it's D-SACK, and must reside below snd_una completely */
1095 if (after(end_seq
, tp
->snd_una
))
1098 if (!before(start_seq
, tp
->undo_marker
))
1102 if (!after(end_seq
, tp
->undo_marker
))
1105 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1106 * start_seq < undo_marker and end_seq >= undo_marker.
1108 return !before(start_seq
, end_seq
- tp
->max_window
);
1111 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1112 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1115 struct tcp_sock
*tp
= tcp_sk(sk
);
1116 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1117 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1118 bool dup_sack
= false;
1120 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1123 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1124 } else if (num_sacks
> 1) {
1125 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1126 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1128 if (!after(end_seq_0
, end_seq_1
) &&
1129 !before(start_seq_0
, start_seq_1
)) {
1132 NET_INC_STATS(sock_net(sk
),
1133 LINUX_MIB_TCPDSACKOFORECV
);
1137 /* D-SACK for already forgotten data... Do dumb counting. */
1138 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1139 !after(end_seq_0
, prior_snd_una
) &&
1140 after(end_seq_0
, tp
->undo_marker
))
1146 struct tcp_sacktag_state
{
1149 /* Timestamps for earliest and latest never-retransmitted segment
1150 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1151 * but congestion control should still get an accurate delay signal.
1155 struct rate_sample
*rate
;
1159 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1160 * the incoming SACK may not exactly match but we can find smaller MSS
1161 * aligned portion of it that matches. Therefore we might need to fragment
1162 * which may fail and creates some hassle (caller must handle error case
1165 * FIXME: this could be merged to shift decision code
1167 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1168 u32 start_seq
, u32 end_seq
)
1172 unsigned int pkt_len
;
1175 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1176 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1178 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1179 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1180 mss
= tcp_skb_mss(skb
);
1181 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1184 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1188 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1193 /* Round if necessary so that SACKs cover only full MSSes
1194 * and/or the remaining small portion (if present)
1196 if (pkt_len
> mss
) {
1197 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1198 if (!in_sack
&& new_len
< pkt_len
)
1203 if (pkt_len
>= skb
->len
&& !in_sack
)
1206 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1214 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1215 static u8
tcp_sacktag_one(struct sock
*sk
,
1216 struct tcp_sacktag_state
*state
, u8 sacked
,
1217 u32 start_seq
, u32 end_seq
,
1218 int dup_sack
, int pcount
,
1221 struct tcp_sock
*tp
= tcp_sk(sk
);
1222 int fack_count
= state
->fack_count
;
1224 /* Account D-SACK for retransmitted packet. */
1225 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1226 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1227 after(end_seq
, tp
->undo_marker
))
1229 if (sacked
& TCPCB_SACKED_ACKED
)
1230 state
->reord
= min(fack_count
, state
->reord
);
1233 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1234 if (!after(end_seq
, tp
->snd_una
))
1237 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1238 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1240 if (sacked
& TCPCB_SACKED_RETRANS
) {
1241 /* If the segment is not tagged as lost,
1242 * we do not clear RETRANS, believing
1243 * that retransmission is still in flight.
1245 if (sacked
& TCPCB_LOST
) {
1246 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1247 tp
->lost_out
-= pcount
;
1248 tp
->retrans_out
-= pcount
;
1251 if (!(sacked
& TCPCB_RETRANS
)) {
1252 /* New sack for not retransmitted frame,
1253 * which was in hole. It is reordering.
1255 if (before(start_seq
,
1256 tcp_highest_sack_seq(tp
)))
1257 state
->reord
= min(fack_count
,
1259 if (!after(end_seq
, tp
->high_seq
))
1260 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1261 if (state
->first_sackt
== 0)
1262 state
->first_sackt
= xmit_time
;
1263 state
->last_sackt
= xmit_time
;
1266 if (sacked
& TCPCB_LOST
) {
1267 sacked
&= ~TCPCB_LOST
;
1268 tp
->lost_out
-= pcount
;
1272 sacked
|= TCPCB_SACKED_ACKED
;
1273 state
->flag
|= FLAG_DATA_SACKED
;
1274 tp
->sacked_out
+= pcount
;
1275 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1277 fack_count
+= pcount
;
1279 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1280 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1281 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1282 tp
->lost_cnt_hint
+= pcount
;
1284 if (fack_count
> tp
->fackets_out
)
1285 tp
->fackets_out
= fack_count
;
1288 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1289 * frames and clear it. undo_retrans is decreased above, L|R frames
1290 * are accounted above as well.
1292 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1293 sacked
&= ~TCPCB_SACKED_RETRANS
;
1294 tp
->retrans_out
-= pcount
;
1300 /* Shift newly-SACKed bytes from this skb to the immediately previous
1301 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1303 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1304 struct tcp_sacktag_state
*state
,
1305 unsigned int pcount
, int shifted
, int mss
,
1308 struct tcp_sock
*tp
= tcp_sk(sk
);
1309 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1310 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1311 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1315 /* Adjust counters and hints for the newly sacked sequence
1316 * range but discard the return value since prev is already
1317 * marked. We must tag the range first because the seq
1318 * advancement below implicitly advances
1319 * tcp_highest_sack_seq() when skb is highest_sack.
1321 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1322 start_seq
, end_seq
, dup_sack
, pcount
,
1324 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1326 if (skb
== tp
->lost_skb_hint
)
1327 tp
->lost_cnt_hint
+= pcount
;
1329 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1330 TCP_SKB_CB(skb
)->seq
+= shifted
;
1332 tcp_skb_pcount_add(prev
, pcount
);
1333 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1334 tcp_skb_pcount_add(skb
, -pcount
);
1336 /* When we're adding to gso_segs == 1, gso_size will be zero,
1337 * in theory this shouldn't be necessary but as long as DSACK
1338 * code can come after this skb later on it's better to keep
1339 * setting gso_size to something.
1341 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1342 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1344 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1345 if (tcp_skb_pcount(skb
) <= 1)
1346 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1348 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1349 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1352 BUG_ON(!tcp_skb_pcount(skb
));
1353 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1357 /* Whole SKB was eaten :-) */
1359 if (skb
== tp
->retransmit_skb_hint
)
1360 tp
->retransmit_skb_hint
= prev
;
1361 if (skb
== tp
->lost_skb_hint
) {
1362 tp
->lost_skb_hint
= prev
;
1363 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1366 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1367 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1368 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1369 TCP_SKB_CB(prev
)->end_seq
++;
1371 if (skb
== tcp_highest_sack(sk
))
1372 tcp_advance_highest_sack(sk
, skb
);
1374 tcp_skb_collapse_tstamp(prev
, skb
);
1375 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1376 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1378 tcp_unlink_write_queue(skb
, sk
);
1379 sk_wmem_free_skb(sk
, skb
);
1381 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1386 /* I wish gso_size would have a bit more sane initialization than
1387 * something-or-zero which complicates things
1389 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1391 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1394 /* Shifting pages past head area doesn't work */
1395 static int skb_can_shift(const struct sk_buff
*skb
)
1397 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1400 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1403 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1404 struct tcp_sacktag_state
*state
,
1405 u32 start_seq
, u32 end_seq
,
1408 struct tcp_sock
*tp
= tcp_sk(sk
);
1409 struct sk_buff
*prev
;
1415 if (!sk_can_gso(sk
))
1418 /* Normally R but no L won't result in plain S */
1420 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1422 if (!skb_can_shift(skb
))
1424 /* This frame is about to be dropped (was ACKed). */
1425 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1428 /* Can only happen with delayed DSACK + discard craziness */
1429 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1431 prev
= tcp_write_queue_prev(sk
, skb
);
1433 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1436 if (!tcp_skb_can_collapse_to(prev
))
1439 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1440 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1444 pcount
= tcp_skb_pcount(skb
);
1445 mss
= tcp_skb_seglen(skb
);
1447 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1448 * drop this restriction as unnecessary
1450 if (mss
!= tcp_skb_seglen(prev
))
1453 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1455 /* CHECKME: This is non-MSS split case only?, this will
1456 * cause skipped skbs due to advancing loop btw, original
1457 * has that feature too
1459 if (tcp_skb_pcount(skb
) <= 1)
1462 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1464 /* TODO: head merge to next could be attempted here
1465 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1466 * though it might not be worth of the additional hassle
1468 * ...we can probably just fallback to what was done
1469 * previously. We could try merging non-SACKed ones
1470 * as well but it probably isn't going to buy off
1471 * because later SACKs might again split them, and
1472 * it would make skb timestamp tracking considerably
1478 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1480 BUG_ON(len
> skb
->len
);
1482 /* MSS boundaries should be honoured or else pcount will
1483 * severely break even though it makes things bit trickier.
1484 * Optimize common case to avoid most of the divides
1486 mss
= tcp_skb_mss(skb
);
1488 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1489 * drop this restriction as unnecessary
1491 if (mss
!= tcp_skb_seglen(prev
))
1496 } else if (len
< mss
) {
1504 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1505 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1508 if (!skb_shift(prev
, skb
, len
))
1510 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1513 /* Hole filled allows collapsing with the next as well, this is very
1514 * useful when hole on every nth skb pattern happens
1516 if (prev
== tcp_write_queue_tail(sk
))
1518 skb
= tcp_write_queue_next(sk
, prev
);
1520 if (!skb_can_shift(skb
) ||
1521 (skb
== tcp_send_head(sk
)) ||
1522 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1523 (mss
!= tcp_skb_seglen(skb
)))
1527 if (skb_shift(prev
, skb
, len
)) {
1528 pcount
+= tcp_skb_pcount(skb
);
1529 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1533 state
->fack_count
+= pcount
;
1540 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1544 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1545 struct tcp_sack_block
*next_dup
,
1546 struct tcp_sacktag_state
*state
,
1547 u32 start_seq
, u32 end_seq
,
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1551 struct sk_buff
*tmp
;
1553 tcp_for_write_queue_from(skb
, sk
) {
1555 bool dup_sack
= dup_sack_in
;
1557 if (skb
== tcp_send_head(sk
))
1560 /* queue is in-order => we can short-circuit the walk early */
1561 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1565 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1566 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1567 next_dup
->start_seq
,
1573 /* skb reference here is a bit tricky to get right, since
1574 * shifting can eat and free both this skb and the next,
1575 * so not even _safe variant of the loop is enough.
1578 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1579 start_seq
, end_seq
, dup_sack
);
1588 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1594 if (unlikely(in_sack
< 0))
1598 TCP_SKB_CB(skb
)->sacked
=
1601 TCP_SKB_CB(skb
)->sacked
,
1602 TCP_SKB_CB(skb
)->seq
,
1603 TCP_SKB_CB(skb
)->end_seq
,
1605 tcp_skb_pcount(skb
),
1607 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1609 if (!before(TCP_SKB_CB(skb
)->seq
,
1610 tcp_highest_sack_seq(tp
)))
1611 tcp_advance_highest_sack(sk
, skb
);
1614 state
->fack_count
+= tcp_skb_pcount(skb
);
1619 /* Avoid all extra work that is being done by sacktag while walking in
1622 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1623 struct tcp_sacktag_state
*state
,
1626 tcp_for_write_queue_from(skb
, sk
) {
1627 if (skb
== tcp_send_head(sk
))
1630 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1633 state
->fack_count
+= tcp_skb_pcount(skb
);
1638 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1640 struct tcp_sack_block
*next_dup
,
1641 struct tcp_sacktag_state
*state
,
1647 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1648 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1649 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1650 next_dup
->start_seq
, next_dup
->end_seq
,
1657 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1659 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1663 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1664 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1666 struct tcp_sock
*tp
= tcp_sk(sk
);
1667 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1668 TCP_SKB_CB(ack_skb
)->sacked
);
1669 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1670 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1671 struct tcp_sack_block
*cache
;
1672 struct sk_buff
*skb
;
1673 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1675 bool found_dup_sack
= false;
1677 int first_sack_index
;
1680 state
->reord
= tp
->packets_out
;
1682 if (!tp
->sacked_out
) {
1683 if (WARN_ON(tp
->fackets_out
))
1684 tp
->fackets_out
= 0;
1685 tcp_highest_sack_reset(sk
);
1688 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1689 num_sacks
, prior_snd_una
);
1690 if (found_dup_sack
) {
1691 state
->flag
|= FLAG_DSACKING_ACK
;
1692 tp
->delivered
++; /* A spurious retransmission is delivered */
1695 /* Eliminate too old ACKs, but take into
1696 * account more or less fresh ones, they can
1697 * contain valid SACK info.
1699 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1702 if (!tp
->packets_out
)
1706 first_sack_index
= 0;
1707 for (i
= 0; i
< num_sacks
; i
++) {
1708 bool dup_sack
= !i
&& found_dup_sack
;
1710 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1711 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1713 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1714 sp
[used_sacks
].start_seq
,
1715 sp
[used_sacks
].end_seq
)) {
1719 if (!tp
->undo_marker
)
1720 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1722 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1724 /* Don't count olds caused by ACK reordering */
1725 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1726 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1728 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1731 NET_INC_STATS(sock_net(sk
), mib_idx
);
1733 first_sack_index
= -1;
1737 /* Ignore very old stuff early */
1738 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1744 /* order SACK blocks to allow in order walk of the retrans queue */
1745 for (i
= used_sacks
- 1; i
> 0; i
--) {
1746 for (j
= 0; j
< i
; j
++) {
1747 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1748 swap(sp
[j
], sp
[j
+ 1]);
1750 /* Track where the first SACK block goes to */
1751 if (j
== first_sack_index
)
1752 first_sack_index
= j
+ 1;
1757 skb
= tcp_write_queue_head(sk
);
1758 state
->fack_count
= 0;
1761 if (!tp
->sacked_out
) {
1762 /* It's already past, so skip checking against it */
1763 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1765 cache
= tp
->recv_sack_cache
;
1766 /* Skip empty blocks in at head of the cache */
1767 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1772 while (i
< used_sacks
) {
1773 u32 start_seq
= sp
[i
].start_seq
;
1774 u32 end_seq
= sp
[i
].end_seq
;
1775 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1776 struct tcp_sack_block
*next_dup
= NULL
;
1778 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1779 next_dup
= &sp
[i
+ 1];
1781 /* Skip too early cached blocks */
1782 while (tcp_sack_cache_ok(tp
, cache
) &&
1783 !before(start_seq
, cache
->end_seq
))
1786 /* Can skip some work by looking recv_sack_cache? */
1787 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1788 after(end_seq
, cache
->start_seq
)) {
1791 if (before(start_seq
, cache
->start_seq
)) {
1792 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1794 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1801 /* Rest of the block already fully processed? */
1802 if (!after(end_seq
, cache
->end_seq
))
1805 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1809 /* ...tail remains todo... */
1810 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1811 /* ...but better entrypoint exists! */
1812 skb
= tcp_highest_sack(sk
);
1815 state
->fack_count
= tp
->fackets_out
;
1820 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1821 /* Check overlap against next cached too (past this one already) */
1826 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1827 skb
= tcp_highest_sack(sk
);
1830 state
->fack_count
= tp
->fackets_out
;
1832 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1835 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1836 start_seq
, end_seq
, dup_sack
);
1842 /* Clear the head of the cache sack blocks so we can skip it next time */
1843 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1844 tp
->recv_sack_cache
[i
].start_seq
= 0;
1845 tp
->recv_sack_cache
[i
].end_seq
= 0;
1847 for (j
= 0; j
< used_sacks
; j
++)
1848 tp
->recv_sack_cache
[i
++] = sp
[j
];
1850 if ((state
->reord
< tp
->fackets_out
) &&
1851 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1852 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1854 tcp_verify_left_out(tp
);
1857 #if FASTRETRANS_DEBUG > 0
1858 WARN_ON((int)tp
->sacked_out
< 0);
1859 WARN_ON((int)tp
->lost_out
< 0);
1860 WARN_ON((int)tp
->retrans_out
< 0);
1861 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1866 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1867 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1869 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1873 holes
= max(tp
->lost_out
, 1U);
1874 holes
= min(holes
, tp
->packets_out
);
1876 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1877 tp
->sacked_out
= tp
->packets_out
- holes
;
1883 /* If we receive more dupacks than we expected counting segments
1884 * in assumption of absent reordering, interpret this as reordering.
1885 * The only another reason could be bug in receiver TCP.
1887 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1889 struct tcp_sock
*tp
= tcp_sk(sk
);
1890 if (tcp_limit_reno_sacked(tp
))
1891 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1894 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1896 static void tcp_add_reno_sack(struct sock
*sk
)
1898 struct tcp_sock
*tp
= tcp_sk(sk
);
1899 u32 prior_sacked
= tp
->sacked_out
;
1902 tcp_check_reno_reordering(sk
, 0);
1903 if (tp
->sacked_out
> prior_sacked
)
1904 tp
->delivered
++; /* Some out-of-order packet is delivered */
1905 tcp_verify_left_out(tp
);
1908 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1910 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1912 struct tcp_sock
*tp
= tcp_sk(sk
);
1915 /* One ACK acked hole. The rest eat duplicate ACKs. */
1916 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1917 if (acked
- 1 >= tp
->sacked_out
)
1920 tp
->sacked_out
-= acked
- 1;
1922 tcp_check_reno_reordering(sk
, acked
);
1923 tcp_verify_left_out(tp
);
1926 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1931 void tcp_clear_retrans(struct tcp_sock
*tp
)
1933 tp
->retrans_out
= 0;
1935 tp
->undo_marker
= 0;
1936 tp
->undo_retrans
= -1;
1937 tp
->fackets_out
= 0;
1941 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1943 tp
->undo_marker
= tp
->snd_una
;
1944 /* Retransmission still in flight may cause DSACKs later. */
1945 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1948 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1949 * and reset tags completely, otherwise preserve SACKs. If receiver
1950 * dropped its ofo queue, we will know this due to reneging detection.
1952 void tcp_enter_loss(struct sock
*sk
)
1954 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1955 struct tcp_sock
*tp
= tcp_sk(sk
);
1956 struct net
*net
= sock_net(sk
);
1957 struct sk_buff
*skb
;
1958 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1959 bool is_reneg
; /* is receiver reneging on SACKs? */
1962 /* Reduce ssthresh if it has not yet been made inside this window. */
1963 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1964 !after(tp
->high_seq
, tp
->snd_una
) ||
1965 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1966 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1967 tp
->prior_cwnd
= tp
->snd_cwnd
;
1968 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1969 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1973 tp
->snd_cwnd_cnt
= 0;
1974 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1976 tp
->retrans_out
= 0;
1979 if (tcp_is_reno(tp
))
1980 tcp_reset_reno_sack(tp
);
1982 skb
= tcp_write_queue_head(sk
);
1983 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1985 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1987 tp
->fackets_out
= 0;
1988 /* Mark SACK reneging until we recover from this loss event. */
1989 tp
->is_sack_reneg
= 1;
1991 tcp_clear_all_retrans_hints(tp
);
1993 tcp_for_write_queue(skb
, sk
) {
1994 if (skb
== tcp_send_head(sk
))
1997 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2000 tcp_sum_lost(tp
, skb
);
2001 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2003 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2004 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2005 tp
->lost_out
+= tcp_skb_pcount(skb
);
2008 tcp_verify_left_out(tp
);
2010 /* Timeout in disordered state after receiving substantial DUPACKs
2011 * suggests that the degree of reordering is over-estimated.
2013 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
2014 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
2015 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2016 net
->ipv4
.sysctl_tcp_reordering
);
2017 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2018 tp
->high_seq
= tp
->snd_nxt
;
2019 tcp_ecn_queue_cwr(tp
);
2021 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2022 * loss recovery is underway except recurring timeout(s) on
2023 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2025 tp
->frto
= sysctl_tcp_frto
&&
2026 (new_recovery
|| icsk
->icsk_retransmits
) &&
2027 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2030 /* If ACK arrived pointing to a remembered SACK, it means that our
2031 * remembered SACKs do not reflect real state of receiver i.e.
2032 * receiver _host_ is heavily congested (or buggy).
2034 * To avoid big spurious retransmission bursts due to transient SACK
2035 * scoreboard oddities that look like reneging, we give the receiver a
2036 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2037 * restore sanity to the SACK scoreboard. If the apparent reneging
2038 * persists until this RTO then we'll clear the SACK scoreboard.
2040 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2042 if (flag
& FLAG_SACK_RENEGING
) {
2043 struct tcp_sock
*tp
= tcp_sk(sk
);
2044 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2045 msecs_to_jiffies(10));
2047 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2048 delay
, TCP_RTO_MAX
);
2054 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2056 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2059 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2060 * counter when SACK is enabled (without SACK, sacked_out is used for
2063 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2064 * segments up to the highest received SACK block so far and holes in
2067 * With reordering, holes may still be in flight, so RFC3517 recovery
2068 * uses pure sacked_out (total number of SACKed segments) even though
2069 * it violates the RFC that uses duplicate ACKs, often these are equal
2070 * but when e.g. out-of-window ACKs or packet duplication occurs,
2071 * they differ. Since neither occurs due to loss, TCP should really
2074 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2076 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2079 /* Linux NewReno/SACK/FACK/ECN state machine.
2080 * --------------------------------------
2082 * "Open" Normal state, no dubious events, fast path.
2083 * "Disorder" In all the respects it is "Open",
2084 * but requires a bit more attention. It is entered when
2085 * we see some SACKs or dupacks. It is split of "Open"
2086 * mainly to move some processing from fast path to slow one.
2087 * "CWR" CWND was reduced due to some Congestion Notification event.
2088 * It can be ECN, ICMP source quench, local device congestion.
2089 * "Recovery" CWND was reduced, we are fast-retransmitting.
2090 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2092 * tcp_fastretrans_alert() is entered:
2093 * - each incoming ACK, if state is not "Open"
2094 * - when arrived ACK is unusual, namely:
2099 * Counting packets in flight is pretty simple.
2101 * in_flight = packets_out - left_out + retrans_out
2103 * packets_out is SND.NXT-SND.UNA counted in packets.
2105 * retrans_out is number of retransmitted segments.
2107 * left_out is number of segments left network, but not ACKed yet.
2109 * left_out = sacked_out + lost_out
2111 * sacked_out: Packets, which arrived to receiver out of order
2112 * and hence not ACKed. With SACKs this number is simply
2113 * amount of SACKed data. Even without SACKs
2114 * it is easy to give pretty reliable estimate of this number,
2115 * counting duplicate ACKs.
2117 * lost_out: Packets lost by network. TCP has no explicit
2118 * "loss notification" feedback from network (for now).
2119 * It means that this number can be only _guessed_.
2120 * Actually, it is the heuristics to predict lossage that
2121 * distinguishes different algorithms.
2123 * F.e. after RTO, when all the queue is considered as lost,
2124 * lost_out = packets_out and in_flight = retrans_out.
2126 * Essentially, we have now a few algorithms detecting
2129 * If the receiver supports SACK:
2131 * RFC6675/3517: It is the conventional algorithm. A packet is
2132 * considered lost if the number of higher sequence packets
2133 * SACKed is greater than or equal the DUPACK thoreshold
2134 * (reordering). This is implemented in tcp_mark_head_lost and
2135 * tcp_update_scoreboard.
2137 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2138 * (2017-) that checks timing instead of counting DUPACKs.
2139 * Essentially a packet is considered lost if it's not S/ACKed
2140 * after RTT + reordering_window, where both metrics are
2141 * dynamically measured and adjusted. This is implemented in
2142 * tcp_rack_mark_lost.
2144 * FACK (Disabled by default. Subsumbed by RACK):
2145 * It is the simplest heuristics. As soon as we decided
2146 * that something is lost, we decide that _all_ not SACKed
2147 * packets until the most forward SACK are lost. I.e.
2148 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2149 * It is absolutely correct estimate, if network does not reorder
2150 * packets. And it loses any connection to reality when reordering
2151 * takes place. We use FACK by default until reordering
2152 * is suspected on the path to this destination.
2154 * If the receiver does not support SACK:
2156 * NewReno (RFC6582): in Recovery we assume that one segment
2157 * is lost (classic Reno). While we are in Recovery and
2158 * a partial ACK arrives, we assume that one more packet
2159 * is lost (NewReno). This heuristics are the same in NewReno
2162 * Really tricky (and requiring careful tuning) part of algorithm
2163 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2164 * The first determines the moment _when_ we should reduce CWND and,
2165 * hence, slow down forward transmission. In fact, it determines the moment
2166 * when we decide that hole is caused by loss, rather than by a reorder.
2168 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2169 * holes, caused by lost packets.
2171 * And the most logically complicated part of algorithm is undo
2172 * heuristics. We detect false retransmits due to both too early
2173 * fast retransmit (reordering) and underestimated RTO, analyzing
2174 * timestamps and D-SACKs. When we detect that some segments were
2175 * retransmitted by mistake and CWND reduction was wrong, we undo
2176 * window reduction and abort recovery phase. This logic is hidden
2177 * inside several functions named tcp_try_undo_<something>.
2180 /* This function decides, when we should leave Disordered state
2181 * and enter Recovery phase, reducing congestion window.
2183 * Main question: may we further continue forward transmission
2184 * with the same cwnd?
2186 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2188 struct tcp_sock
*tp
= tcp_sk(sk
);
2190 /* Trick#1: The loss is proven. */
2194 /* Not-A-Trick#2 : Classic rule... */
2195 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2201 /* Detect loss in event "A" above by marking head of queue up as lost.
2202 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2203 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2204 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2205 * the maximum SACKed segments to pass before reaching this limit.
2207 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2209 struct tcp_sock
*tp
= tcp_sk(sk
);
2210 struct sk_buff
*skb
;
2211 int cnt
, oldcnt
, lost
;
2213 /* Use SACK to deduce losses of new sequences sent during recovery */
2214 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2216 WARN_ON(packets
> tp
->packets_out
);
2217 if (tp
->lost_skb_hint
) {
2218 skb
= tp
->lost_skb_hint
;
2219 cnt
= tp
->lost_cnt_hint
;
2220 /* Head already handled? */
2221 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2224 skb
= tcp_write_queue_head(sk
);
2228 tcp_for_write_queue_from(skb
, sk
) {
2229 if (skb
== tcp_send_head(sk
))
2231 /* TODO: do this better */
2232 /* this is not the most efficient way to do this... */
2233 tp
->lost_skb_hint
= skb
;
2234 tp
->lost_cnt_hint
= cnt
;
2236 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2240 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2241 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2242 cnt
+= tcp_skb_pcount(skb
);
2244 if (cnt
> packets
) {
2245 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2246 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2247 (oldcnt
>= packets
))
2250 mss
= tcp_skb_mss(skb
);
2251 /* If needed, chop off the prefix to mark as lost. */
2252 lost
= (packets
- oldcnt
) * mss
;
2253 if (lost
< skb
->len
&&
2254 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2259 tcp_skb_mark_lost(tp
, skb
);
2264 tcp_verify_left_out(tp
);
2267 /* Account newly detected lost packet(s) */
2269 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2271 struct tcp_sock
*tp
= tcp_sk(sk
);
2273 if (tcp_is_reno(tp
)) {
2274 tcp_mark_head_lost(sk
, 1, 1);
2275 } else if (tcp_is_fack(tp
)) {
2276 int lost
= tp
->fackets_out
- tp
->reordering
;
2279 tcp_mark_head_lost(sk
, lost
, 0);
2281 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2282 if (sacked_upto
>= 0)
2283 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2284 else if (fast_rexmit
)
2285 tcp_mark_head_lost(sk
, 1, 1);
2289 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2291 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2292 before(tp
->rx_opt
.rcv_tsecr
, when
);
2295 /* skb is spurious retransmitted if the returned timestamp echo
2296 * reply is prior to the skb transmission time
2298 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2299 const struct sk_buff
*skb
)
2301 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2302 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2305 /* Nothing was retransmitted or returned timestamp is less
2306 * than timestamp of the first retransmission.
2308 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2310 return !tp
->retrans_stamp
||
2311 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2314 /* Undo procedures. */
2316 /* We can clear retrans_stamp when there are no retransmissions in the
2317 * window. It would seem that it is trivially available for us in
2318 * tp->retrans_out, however, that kind of assumptions doesn't consider
2319 * what will happen if errors occur when sending retransmission for the
2320 * second time. ...It could the that such segment has only
2321 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2322 * the head skb is enough except for some reneging corner cases that
2323 * are not worth the effort.
2325 * Main reason for all this complexity is the fact that connection dying
2326 * time now depends on the validity of the retrans_stamp, in particular,
2327 * that successive retransmissions of a segment must not advance
2328 * retrans_stamp under any conditions.
2330 static bool tcp_any_retrans_done(const struct sock
*sk
)
2332 const struct tcp_sock
*tp
= tcp_sk(sk
);
2333 struct sk_buff
*skb
;
2335 if (tp
->retrans_out
)
2338 skb
= tcp_write_queue_head(sk
);
2339 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2345 #if FASTRETRANS_DEBUG > 1
2346 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2348 struct tcp_sock
*tp
= tcp_sk(sk
);
2349 struct inet_sock
*inet
= inet_sk(sk
);
2351 if (sk
->sk_family
== AF_INET
) {
2352 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2354 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2355 tp
->snd_cwnd
, tcp_left_out(tp
),
2356 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2359 #if IS_ENABLED(CONFIG_IPV6)
2360 else if (sk
->sk_family
== AF_INET6
) {
2361 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2363 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2364 tp
->snd_cwnd
, tcp_left_out(tp
),
2365 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2371 #define DBGUNDO(x...) do { } while (0)
2374 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2376 struct tcp_sock
*tp
= tcp_sk(sk
);
2379 struct sk_buff
*skb
;
2381 tcp_for_write_queue(skb
, sk
) {
2382 if (skb
== tcp_send_head(sk
))
2384 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2387 tcp_clear_all_retrans_hints(tp
);
2390 if (tp
->prior_ssthresh
) {
2391 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2393 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2395 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2396 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2397 tcp_ecn_withdraw_cwr(tp
);
2400 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2401 tp
->undo_marker
= 0;
2404 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2406 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2409 /* People celebrate: "We love our President!" */
2410 static bool tcp_try_undo_recovery(struct sock
*sk
)
2412 struct tcp_sock
*tp
= tcp_sk(sk
);
2414 if (tcp_may_undo(tp
)) {
2417 /* Happy end! We did not retransmit anything
2418 * or our original transmission succeeded.
2420 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2421 tcp_undo_cwnd_reduction(sk
, false);
2422 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2423 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2425 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2427 NET_INC_STATS(sock_net(sk
), mib_idx
);
2429 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2430 /* Hold old state until something *above* high_seq
2431 * is ACKed. For Reno it is MUST to prevent false
2432 * fast retransmits (RFC2582). SACK TCP is safe. */
2433 if (!tcp_any_retrans_done(sk
))
2434 tp
->retrans_stamp
= 0;
2437 tcp_set_ca_state(sk
, TCP_CA_Open
);
2438 tp
->is_sack_reneg
= 0;
2442 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2443 static bool tcp_try_undo_dsack(struct sock
*sk
)
2445 struct tcp_sock
*tp
= tcp_sk(sk
);
2447 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2448 DBGUNDO(sk
, "D-SACK");
2449 tcp_undo_cwnd_reduction(sk
, false);
2450 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2456 /* Undo during loss recovery after partial ACK or using F-RTO. */
2457 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2459 struct tcp_sock
*tp
= tcp_sk(sk
);
2461 if (frto_undo
|| tcp_may_undo(tp
)) {
2462 tcp_undo_cwnd_reduction(sk
, true);
2464 DBGUNDO(sk
, "partial loss");
2465 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2467 NET_INC_STATS(sock_net(sk
),
2468 LINUX_MIB_TCPSPURIOUSRTOS
);
2469 inet_csk(sk
)->icsk_retransmits
= 0;
2470 if (frto_undo
|| tcp_is_sack(tp
)) {
2471 tcp_set_ca_state(sk
, TCP_CA_Open
);
2472 tp
->is_sack_reneg
= 0;
2479 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2480 * It computes the number of packets to send (sndcnt) based on packets newly
2482 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2483 * cwnd reductions across a full RTT.
2484 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2485 * But when the retransmits are acked without further losses, PRR
2486 * slow starts cwnd up to ssthresh to speed up the recovery.
2488 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2490 struct tcp_sock
*tp
= tcp_sk(sk
);
2492 tp
->high_seq
= tp
->snd_nxt
;
2493 tp
->tlp_high_seq
= 0;
2494 tp
->snd_cwnd_cnt
= 0;
2495 tp
->prior_cwnd
= tp
->snd_cwnd
;
2496 tp
->prr_delivered
= 0;
2498 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2499 tcp_ecn_queue_cwr(tp
);
2502 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2504 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2508 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2511 tp
->prr_delivered
+= newly_acked_sacked
;
2513 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2515 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2516 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2517 !(flag
& FLAG_LOST_RETRANS
)) {
2518 sndcnt
= min_t(int, delta
,
2519 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2520 newly_acked_sacked
) + 1);
2522 sndcnt
= min(delta
, newly_acked_sacked
);
2524 /* Force a fast retransmit upon entering fast recovery */
2525 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2526 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2529 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2531 struct tcp_sock
*tp
= tcp_sk(sk
);
2533 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2536 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2537 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2538 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2539 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2540 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2542 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2545 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2546 void tcp_enter_cwr(struct sock
*sk
)
2548 struct tcp_sock
*tp
= tcp_sk(sk
);
2550 tp
->prior_ssthresh
= 0;
2551 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2552 tp
->undo_marker
= 0;
2553 tcp_init_cwnd_reduction(sk
);
2554 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2557 EXPORT_SYMBOL(tcp_enter_cwr
);
2559 static void tcp_try_keep_open(struct sock
*sk
)
2561 struct tcp_sock
*tp
= tcp_sk(sk
);
2562 int state
= TCP_CA_Open
;
2564 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2565 state
= TCP_CA_Disorder
;
2567 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2568 tcp_set_ca_state(sk
, state
);
2569 tp
->high_seq
= tp
->snd_nxt
;
2573 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2575 struct tcp_sock
*tp
= tcp_sk(sk
);
2577 tcp_verify_left_out(tp
);
2579 if (!tcp_any_retrans_done(sk
))
2580 tp
->retrans_stamp
= 0;
2582 if (flag
& FLAG_ECE
)
2585 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2586 tcp_try_keep_open(sk
);
2590 static void tcp_mtup_probe_failed(struct sock
*sk
)
2592 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2594 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2595 icsk
->icsk_mtup
.probe_size
= 0;
2596 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2599 static void tcp_mtup_probe_success(struct sock
*sk
)
2601 struct tcp_sock
*tp
= tcp_sk(sk
);
2602 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2604 /* FIXME: breaks with very large cwnd */
2605 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2606 tp
->snd_cwnd
= tp
->snd_cwnd
*
2607 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2608 icsk
->icsk_mtup
.probe_size
;
2609 tp
->snd_cwnd_cnt
= 0;
2610 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2611 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2613 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2614 icsk
->icsk_mtup
.probe_size
= 0;
2615 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2616 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2619 /* Do a simple retransmit without using the backoff mechanisms in
2620 * tcp_timer. This is used for path mtu discovery.
2621 * The socket is already locked here.
2623 void tcp_simple_retransmit(struct sock
*sk
)
2625 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2626 struct tcp_sock
*tp
= tcp_sk(sk
);
2627 struct sk_buff
*skb
;
2628 unsigned int mss
= tcp_current_mss(sk
);
2630 tcp_for_write_queue(skb
, sk
) {
2631 if (skb
== tcp_send_head(sk
))
2633 if (tcp_skb_seglen(skb
) > mss
&&
2634 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2635 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2636 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2637 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2639 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2643 tcp_clear_retrans_hints_partial(tp
);
2648 if (tcp_is_reno(tp
))
2649 tcp_limit_reno_sacked(tp
);
2651 tcp_verify_left_out(tp
);
2653 /* Don't muck with the congestion window here.
2654 * Reason is that we do not increase amount of _data_
2655 * in network, but units changed and effective
2656 * cwnd/ssthresh really reduced now.
2658 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2659 tp
->high_seq
= tp
->snd_nxt
;
2660 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2661 tp
->prior_ssthresh
= 0;
2662 tp
->undo_marker
= 0;
2663 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2665 tcp_xmit_retransmit_queue(sk
);
2667 EXPORT_SYMBOL(tcp_simple_retransmit
);
2669 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2671 struct tcp_sock
*tp
= tcp_sk(sk
);
2674 if (tcp_is_reno(tp
))
2675 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2677 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2679 NET_INC_STATS(sock_net(sk
), mib_idx
);
2681 tp
->prior_ssthresh
= 0;
2684 if (!tcp_in_cwnd_reduction(sk
)) {
2686 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2687 tcp_init_cwnd_reduction(sk
);
2689 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2692 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2693 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2695 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2698 struct tcp_sock
*tp
= tcp_sk(sk
);
2699 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2701 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2702 tcp_try_undo_loss(sk
, false))
2705 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2706 /* Step 3.b. A timeout is spurious if not all data are
2707 * lost, i.e., never-retransmitted data are (s)acked.
2709 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2710 tcp_try_undo_loss(sk
, true))
2713 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2714 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2715 tp
->frto
= 0; /* Step 3.a. loss was real */
2716 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2717 tp
->high_seq
= tp
->snd_nxt
;
2718 /* Step 2.b. Try send new data (but deferred until cwnd
2719 * is updated in tcp_ack()). Otherwise fall back to
2720 * the conventional recovery.
2722 if (tcp_send_head(sk
) &&
2723 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2724 *rexmit
= REXMIT_NEW
;
2732 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2733 tcp_try_undo_recovery(sk
);
2736 if (tcp_is_reno(tp
)) {
2737 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2738 * delivered. Lower inflight to clock out (re)tranmissions.
2740 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2741 tcp_add_reno_sack(sk
);
2742 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2743 tcp_reset_reno_sack(tp
);
2745 *rexmit
= REXMIT_LOST
;
2748 /* Undo during fast recovery after partial ACK. */
2749 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2751 struct tcp_sock
*tp
= tcp_sk(sk
);
2753 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2754 /* Plain luck! Hole if filled with delayed
2755 * packet, rather than with a retransmit.
2757 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2759 /* We are getting evidence that the reordering degree is higher
2760 * than we realized. If there are no retransmits out then we
2761 * can undo. Otherwise we clock out new packets but do not
2762 * mark more packets lost or retransmit more.
2764 if (tp
->retrans_out
)
2767 if (!tcp_any_retrans_done(sk
))
2768 tp
->retrans_stamp
= 0;
2770 DBGUNDO(sk
, "partial recovery");
2771 tcp_undo_cwnd_reduction(sk
, true);
2772 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2773 tcp_try_keep_open(sk
);
2779 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
)
2781 struct tcp_sock
*tp
= tcp_sk(sk
);
2783 /* Use RACK to detect loss */
2784 if (sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2785 u32 prior_retrans
= tp
->retrans_out
;
2787 tcp_rack_mark_lost(sk
);
2788 if (prior_retrans
> tp
->retrans_out
)
2789 *ack_flag
|= FLAG_LOST_RETRANS
;
2793 /* Process an event, which can update packets-in-flight not trivially.
2794 * Main goal of this function is to calculate new estimate for left_out,
2795 * taking into account both packets sitting in receiver's buffer and
2796 * packets lost by network.
2798 * Besides that it updates the congestion state when packet loss or ECN
2799 * is detected. But it does not reduce the cwnd, it is done by the
2800 * congestion control later.
2802 * It does _not_ decide what to send, it is made in function
2803 * tcp_xmit_retransmit_queue().
2805 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2806 bool is_dupack
, int *ack_flag
, int *rexmit
)
2808 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2809 struct tcp_sock
*tp
= tcp_sk(sk
);
2810 int fast_rexmit
= 0, flag
= *ack_flag
;
2811 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2812 (tcp_fackets_out(tp
) > tp
->reordering
));
2814 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2816 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2817 tp
->fackets_out
= 0;
2819 /* Now state machine starts.
2820 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2821 if (flag
& FLAG_ECE
)
2822 tp
->prior_ssthresh
= 0;
2824 /* B. In all the states check for reneging SACKs. */
2825 if (tcp_check_sack_reneging(sk
, flag
))
2828 /* C. Check consistency of the current state. */
2829 tcp_verify_left_out(tp
);
2831 /* D. Check state exit conditions. State can be terminated
2832 * when high_seq is ACKed. */
2833 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2834 WARN_ON(tp
->retrans_out
!= 0);
2835 tp
->retrans_stamp
= 0;
2836 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2837 switch (icsk
->icsk_ca_state
) {
2839 /* CWR is to be held something *above* high_seq
2840 * is ACKed for CWR bit to reach receiver. */
2841 if (tp
->snd_una
!= tp
->high_seq
) {
2842 tcp_end_cwnd_reduction(sk
);
2843 tcp_set_ca_state(sk
, TCP_CA_Open
);
2847 case TCP_CA_Recovery
:
2848 if (tcp_is_reno(tp
))
2849 tcp_reset_reno_sack(tp
);
2850 if (tcp_try_undo_recovery(sk
))
2852 tcp_end_cwnd_reduction(sk
);
2857 /* E. Process state. */
2858 switch (icsk
->icsk_ca_state
) {
2859 case TCP_CA_Recovery
:
2860 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2861 if (tcp_is_reno(tp
) && is_dupack
)
2862 tcp_add_reno_sack(sk
);
2864 if (tcp_try_undo_partial(sk
, acked
))
2866 /* Partial ACK arrived. Force fast retransmit. */
2867 do_lost
= tcp_is_reno(tp
) ||
2868 tcp_fackets_out(tp
) > tp
->reordering
;
2870 if (tcp_try_undo_dsack(sk
)) {
2871 tcp_try_keep_open(sk
);
2874 tcp_rack_identify_loss(sk
, ack_flag
);
2877 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2878 tcp_rack_identify_loss(sk
, ack_flag
);
2879 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2880 (*ack_flag
& FLAG_LOST_RETRANS
)))
2882 /* Change state if cwnd is undone or retransmits are lost */
2884 if (tcp_is_reno(tp
)) {
2885 if (flag
& FLAG_SND_UNA_ADVANCED
)
2886 tcp_reset_reno_sack(tp
);
2888 tcp_add_reno_sack(sk
);
2891 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2892 tcp_try_undo_dsack(sk
);
2894 tcp_rack_identify_loss(sk
, ack_flag
);
2895 if (!tcp_time_to_recover(sk
, flag
)) {
2896 tcp_try_to_open(sk
, flag
);
2900 /* MTU probe failure: don't reduce cwnd */
2901 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2902 icsk
->icsk_mtup
.probe_size
&&
2903 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2904 tcp_mtup_probe_failed(sk
);
2905 /* Restores the reduction we did in tcp_mtup_probe() */
2907 tcp_simple_retransmit(sk
);
2911 /* Otherwise enter Recovery state */
2912 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2917 tcp_update_scoreboard(sk
, fast_rexmit
);
2918 *rexmit
= REXMIT_LOST
;
2921 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2923 struct tcp_sock
*tp
= tcp_sk(sk
);
2924 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2926 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2927 rtt_us
? : jiffies_to_usecs(1));
2930 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2931 long seq_rtt_us
, long sack_rtt_us
,
2932 long ca_rtt_us
, struct rate_sample
*rs
)
2934 const struct tcp_sock
*tp
= tcp_sk(sk
);
2936 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2937 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2938 * Karn's algorithm forbids taking RTT if some retransmitted data
2939 * is acked (RFC6298).
2942 seq_rtt_us
= sack_rtt_us
;
2944 /* RTTM Rule: A TSecr value received in a segment is used to
2945 * update the averaged RTT measurement only if the segment
2946 * acknowledges some new data, i.e., only if it advances the
2947 * left edge of the send window.
2948 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2950 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2951 flag
& FLAG_ACKED
) {
2952 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2953 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2955 seq_rtt_us
= ca_rtt_us
= delta_us
;
2957 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2961 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2962 * always taken together with ACK, SACK, or TS-opts. Any negative
2963 * values will be skipped with the seq_rtt_us < 0 check above.
2965 tcp_update_rtt_min(sk
, ca_rtt_us
);
2966 tcp_rtt_estimator(sk
, seq_rtt_us
);
2969 /* RFC6298: only reset backoff on valid RTT measurement. */
2970 inet_csk(sk
)->icsk_backoff
= 0;
2974 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2975 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2977 struct rate_sample rs
;
2980 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2981 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2983 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2987 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2989 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2991 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2992 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2995 /* Restart timer after forward progress on connection.
2996 * RFC2988 recommends to restart timer to now+rto.
2998 void tcp_rearm_rto(struct sock
*sk
)
3000 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3001 struct tcp_sock
*tp
= tcp_sk(sk
);
3003 /* If the retrans timer is currently being used by Fast Open
3004 * for SYN-ACK retrans purpose, stay put.
3006 if (tp
->fastopen_rsk
)
3009 if (!tp
->packets_out
) {
3010 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3012 u32 rto
= inet_csk(sk
)->icsk_rto
;
3013 /* Offset the time elapsed after installing regular RTO */
3014 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
3015 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3016 s64 delta_us
= tcp_rto_delta_us(sk
);
3017 /* delta_us may not be positive if the socket is locked
3018 * when the retrans timer fires and is rescheduled.
3020 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
3022 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3027 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3028 static void tcp_set_xmit_timer(struct sock
*sk
)
3030 if (!tcp_schedule_loss_probe(sk
, true))
3034 /* If we get here, the whole TSO packet has not been acked. */
3035 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3037 struct tcp_sock
*tp
= tcp_sk(sk
);
3040 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3042 packets_acked
= tcp_skb_pcount(skb
);
3043 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3045 packets_acked
-= tcp_skb_pcount(skb
);
3047 if (packets_acked
) {
3048 BUG_ON(tcp_skb_pcount(skb
) == 0);
3049 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3052 return packets_acked
;
3055 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3058 const struct skb_shared_info
*shinfo
;
3060 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3061 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3064 shinfo
= skb_shinfo(skb
);
3065 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3066 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3067 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3070 /* Remove acknowledged frames from the retransmission queue. If our packet
3071 * is before the ack sequence we can discard it as it's confirmed to have
3072 * arrived at the other end.
3074 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3075 u32 prior_snd_una
, int *acked
,
3076 struct tcp_sacktag_state
*sack
)
3078 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3079 u64 first_ackt
, last_ackt
;
3080 struct tcp_sock
*tp
= tcp_sk(sk
);
3081 u32 prior_sacked
= tp
->sacked_out
;
3082 u32 reord
= tp
->packets_out
;
3083 bool fully_acked
= true;
3084 long sack_rtt_us
= -1L;
3085 long seq_rtt_us
= -1L;
3086 long ca_rtt_us
= -1L;
3087 struct sk_buff
*skb
;
3089 u32 last_in_flight
= 0;
3095 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3096 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3097 u8 sacked
= scb
->sacked
;
3100 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3102 /* Determine how many packets and what bytes were acked, tso and else */
3103 if (after(scb
->end_seq
, tp
->snd_una
)) {
3104 if (tcp_skb_pcount(skb
) == 1 ||
3105 !after(tp
->snd_una
, scb
->seq
))
3108 acked_pcount
= tcp_tso_acked(sk
, skb
);
3111 fully_acked
= false;
3113 /* Speedup tcp_unlink_write_queue() and next loop */
3114 prefetchw(skb
->next
);
3115 acked_pcount
= tcp_skb_pcount(skb
);
3118 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3119 if (sacked
& TCPCB_SACKED_RETRANS
)
3120 tp
->retrans_out
-= acked_pcount
;
3121 flag
|= FLAG_RETRANS_DATA_ACKED
;
3122 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3123 last_ackt
= skb
->skb_mstamp
;
3124 WARN_ON_ONCE(last_ackt
== 0);
3126 first_ackt
= last_ackt
;
3128 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3129 reord
= min(pkts_acked
, reord
);
3130 if (!after(scb
->end_seq
, tp
->high_seq
))
3131 flag
|= FLAG_ORIG_SACK_ACKED
;
3134 if (sacked
& TCPCB_SACKED_ACKED
) {
3135 tp
->sacked_out
-= acked_pcount
;
3136 } else if (tcp_is_sack(tp
)) {
3137 tp
->delivered
+= acked_pcount
;
3138 if (!tcp_skb_spurious_retrans(tp
, skb
))
3139 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3142 if (sacked
& TCPCB_LOST
)
3143 tp
->lost_out
-= acked_pcount
;
3145 tp
->packets_out
-= acked_pcount
;
3146 pkts_acked
+= acked_pcount
;
3147 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3149 /* Initial outgoing SYN's get put onto the write_queue
3150 * just like anything else we transmit. It is not
3151 * true data, and if we misinform our callers that
3152 * this ACK acks real data, we will erroneously exit
3153 * connection startup slow start one packet too
3154 * quickly. This is severely frowned upon behavior.
3156 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3157 flag
|= FLAG_DATA_ACKED
;
3159 flag
|= FLAG_SYN_ACKED
;
3160 tp
->retrans_stamp
= 0;
3166 tcp_unlink_write_queue(skb
, sk
);
3167 sk_wmem_free_skb(sk
, skb
);
3168 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3169 tp
->retransmit_skb_hint
= NULL
;
3170 if (unlikely(skb
== tp
->lost_skb_hint
))
3171 tp
->lost_skb_hint
= NULL
;
3175 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3177 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3178 tp
->snd_up
= tp
->snd_una
;
3180 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3181 flag
|= FLAG_SACK_RENEGING
;
3183 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3184 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3185 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3187 if (sack
->first_sackt
) {
3188 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3189 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3191 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3192 ca_rtt_us
, sack
->rate
);
3194 if (flag
& FLAG_ACKED
) {
3195 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3196 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3197 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3198 tcp_mtup_probe_success(sk
);
3201 if (tcp_is_reno(tp
)) {
3202 tcp_remove_reno_sacks(sk
, pkts_acked
);
3204 /* If any of the cumulatively ACKed segments was
3205 * retransmitted, non-SACK case cannot confirm that
3206 * progress was due to original transmission due to
3207 * lack of TCPCB_SACKED_ACKED bits even if some of
3208 * the packets may have been never retransmitted.
3210 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3211 flag
&= ~FLAG_ORIG_SACK_ACKED
;
3215 /* Non-retransmitted hole got filled? That's reordering */
3216 if (reord
< prior_fackets
&& reord
<= tp
->fackets_out
)
3217 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3219 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3220 prior_sacked
- tp
->sacked_out
;
3221 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3224 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3226 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3227 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
, skb
->skb_mstamp
)) {
3228 /* Do not re-arm RTO if the sack RTT is measured from data sent
3229 * after when the head was last (re)transmitted. Otherwise the
3230 * timeout may continue to extend in loss recovery.
3232 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3235 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3236 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3237 .rtt_us
= sack
->rate
->rtt_us
,
3238 .in_flight
= last_in_flight
};
3240 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3243 #if FASTRETRANS_DEBUG > 0
3244 WARN_ON((int)tp
->sacked_out
< 0);
3245 WARN_ON((int)tp
->lost_out
< 0);
3246 WARN_ON((int)tp
->retrans_out
< 0);
3247 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3248 icsk
= inet_csk(sk
);
3250 pr_debug("Leak l=%u %d\n",
3251 tp
->lost_out
, icsk
->icsk_ca_state
);
3254 if (tp
->sacked_out
) {
3255 pr_debug("Leak s=%u %d\n",
3256 tp
->sacked_out
, icsk
->icsk_ca_state
);
3259 if (tp
->retrans_out
) {
3260 pr_debug("Leak r=%u %d\n",
3261 tp
->retrans_out
, icsk
->icsk_ca_state
);
3262 tp
->retrans_out
= 0;
3266 *acked
= pkts_acked
;
3270 static void tcp_ack_probe(struct sock
*sk
)
3272 const struct tcp_sock
*tp
= tcp_sk(sk
);
3273 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3275 /* Was it a usable window open? */
3277 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3278 icsk
->icsk_backoff
= 0;
3279 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3280 /* Socket must be waked up by subsequent tcp_data_snd_check().
3281 * This function is not for random using!
3284 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3286 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3291 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3293 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3294 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3297 /* Decide wheather to run the increase function of congestion control. */
3298 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3300 /* If reordering is high then always grow cwnd whenever data is
3301 * delivered regardless of its ordering. Otherwise stay conservative
3302 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3303 * new SACK or ECE mark may first advance cwnd here and later reduce
3304 * cwnd in tcp_fastretrans_alert() based on more states.
3306 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3307 return flag
& FLAG_FORWARD_PROGRESS
;
3309 return flag
& FLAG_DATA_ACKED
;
3312 /* The "ultimate" congestion control function that aims to replace the rigid
3313 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3314 * It's called toward the end of processing an ACK with precise rate
3315 * information. All transmission or retransmission are delayed afterwards.
3317 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3318 int flag
, const struct rate_sample
*rs
)
3320 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3322 if (icsk
->icsk_ca_ops
->cong_control
) {
3323 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3327 if (tcp_in_cwnd_reduction(sk
)) {
3328 /* Reduce cwnd if state mandates */
3329 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3330 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3331 /* Advance cwnd if state allows */
3332 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3334 tcp_update_pacing_rate(sk
);
3337 /* Check that window update is acceptable.
3338 * The function assumes that snd_una<=ack<=snd_next.
3340 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3341 const u32 ack
, const u32 ack_seq
,
3344 return after(ack
, tp
->snd_una
) ||
3345 after(ack_seq
, tp
->snd_wl1
) ||
3346 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3349 /* If we update tp->snd_una, also update tp->bytes_acked */
3350 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3352 u32 delta
= ack
- tp
->snd_una
;
3354 sock_owned_by_me((struct sock
*)tp
);
3355 tp
->bytes_acked
+= delta
;
3359 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3360 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3362 u32 delta
= seq
- tp
->rcv_nxt
;
3364 sock_owned_by_me((struct sock
*)tp
);
3365 tp
->bytes_received
+= delta
;
3369 /* Update our send window.
3371 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3372 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3374 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3377 struct tcp_sock
*tp
= tcp_sk(sk
);
3379 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3381 if (likely(!tcp_hdr(skb
)->syn
))
3382 nwin
<<= tp
->rx_opt
.snd_wscale
;
3384 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3385 flag
|= FLAG_WIN_UPDATE
;
3386 tcp_update_wl(tp
, ack_seq
);
3388 if (tp
->snd_wnd
!= nwin
) {
3391 /* Note, it is the only place, where
3392 * fast path is recovered for sending TCP.
3395 tcp_fast_path_check(sk
);
3397 if (tcp_send_head(sk
))
3398 tcp_slow_start_after_idle_check(sk
);
3400 if (nwin
> tp
->max_window
) {
3401 tp
->max_window
= nwin
;
3402 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3407 tcp_snd_una_update(tp
, ack
);
3412 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3413 u32
*last_oow_ack_time
)
3415 if (*last_oow_ack_time
) {
3416 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3418 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3419 NET_INC_STATS(net
, mib_idx
);
3420 return true; /* rate-limited: don't send yet! */
3424 *last_oow_ack_time
= tcp_jiffies32
;
3426 return false; /* not rate-limited: go ahead, send dupack now! */
3429 /* Return true if we're currently rate-limiting out-of-window ACKs and
3430 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3431 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3432 * attacks that send repeated SYNs or ACKs for the same connection. To
3433 * do this, we do not send a duplicate SYNACK or ACK if the remote
3434 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3436 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3437 int mib_idx
, u32
*last_oow_ack_time
)
3439 /* Data packets without SYNs are not likely part of an ACK loop. */
3440 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3444 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3447 /* RFC 5961 7 [ACK Throttling] */
3448 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3450 /* unprotected vars, we dont care of overwrites */
3451 static u32 challenge_timestamp
;
3452 static unsigned int challenge_count
;
3453 struct tcp_sock
*tp
= tcp_sk(sk
);
3456 /* First check our per-socket dupack rate limit. */
3457 if (__tcp_oow_rate_limited(sock_net(sk
),
3458 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3459 &tp
->last_oow_ack_time
))
3462 /* Then check host-wide RFC 5961 rate limit. */
3464 if (now
!= challenge_timestamp
) {
3465 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3467 challenge_timestamp
= now
;
3468 WRITE_ONCE(challenge_count
, half
+
3469 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3471 count
= READ_ONCE(challenge_count
);
3473 WRITE_ONCE(challenge_count
, count
- 1);
3474 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3479 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3481 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3482 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3485 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3487 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3488 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3489 * extra check below makes sure this can only happen
3490 * for pure ACK frames. -DaveM
3492 * Not only, also it occurs for expired timestamps.
3495 if (tcp_paws_check(&tp
->rx_opt
, 0))
3496 tcp_store_ts_recent(tp
);
3500 /* This routine deals with acks during a TLP episode.
3501 * We mark the end of a TLP episode on receiving TLP dupack or when
3502 * ack is after tlp_high_seq.
3503 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3505 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3507 struct tcp_sock
*tp
= tcp_sk(sk
);
3509 if (before(ack
, tp
->tlp_high_seq
))
3512 if (flag
& FLAG_DSACKING_ACK
) {
3513 /* This DSACK means original and TLP probe arrived; no loss */
3514 tp
->tlp_high_seq
= 0;
3515 } else if (after(ack
, tp
->tlp_high_seq
)) {
3516 /* ACK advances: there was a loss, so reduce cwnd. Reset
3517 * tlp_high_seq in tcp_init_cwnd_reduction()
3519 tcp_init_cwnd_reduction(sk
);
3520 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3521 tcp_end_cwnd_reduction(sk
);
3522 tcp_try_keep_open(sk
);
3523 NET_INC_STATS(sock_net(sk
),
3524 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3525 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3526 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3527 /* Pure dupack: original and TLP probe arrived; no loss */
3528 tp
->tlp_high_seq
= 0;
3532 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3534 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3536 if (icsk
->icsk_ca_ops
->in_ack_event
)
3537 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3540 /* Congestion control has updated the cwnd already. So if we're in
3541 * loss recovery then now we do any new sends (for FRTO) or
3542 * retransmits (for CA_Loss or CA_recovery) that make sense.
3544 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3546 struct tcp_sock
*tp
= tcp_sk(sk
);
3548 if (rexmit
== REXMIT_NONE
)
3551 if (unlikely(rexmit
== 2)) {
3552 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3554 if (after(tp
->snd_nxt
, tp
->high_seq
))
3558 tcp_xmit_retransmit_queue(sk
);
3561 /* This routine deals with incoming acks, but not outgoing ones. */
3562 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3564 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3565 struct tcp_sock
*tp
= tcp_sk(sk
);
3566 struct tcp_sacktag_state sack_state
;
3567 struct rate_sample rs
= { .prior_delivered
= 0 };
3568 u32 prior_snd_una
= tp
->snd_una
;
3569 bool is_sack_reneg
= tp
->is_sack_reneg
;
3570 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3571 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3572 bool is_dupack
= false;
3574 int prior_packets
= tp
->packets_out
;
3575 u32 delivered
= tp
->delivered
;
3576 u32 lost
= tp
->lost
;
3577 int acked
= 0; /* Number of packets newly acked */
3578 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3580 sack_state
.first_sackt
= 0;
3581 sack_state
.rate
= &rs
;
3583 /* We very likely will need to access write queue head. */
3584 prefetchw(sk
->sk_write_queue
.next
);
3586 /* If the ack is older than previous acks
3587 * then we can probably ignore it.
3589 if (before(ack
, prior_snd_una
)) {
3590 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3591 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3592 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3593 tcp_send_challenge_ack(sk
, skb
);
3599 /* If the ack includes data we haven't sent yet, discard
3600 * this segment (RFC793 Section 3.9).
3602 if (after(ack
, tp
->snd_nxt
))
3605 if (after(ack
, prior_snd_una
)) {
3606 flag
|= FLAG_SND_UNA_ADVANCED
;
3607 icsk
->icsk_retransmits
= 0;
3610 prior_fackets
= tp
->fackets_out
;
3611 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3613 /* ts_recent update must be made after we are sure that the packet
3616 if (flag
& FLAG_UPDATE_TS_RECENT
)
3617 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3619 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3620 /* Window is constant, pure forward advance.
3621 * No more checks are required.
3622 * Note, we use the fact that SND.UNA>=SND.WL2.
3624 tcp_update_wl(tp
, ack_seq
);
3625 tcp_snd_una_update(tp
, ack
);
3626 flag
|= FLAG_WIN_UPDATE
;
3628 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3630 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3632 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3634 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3637 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3639 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3641 if (TCP_SKB_CB(skb
)->sacked
)
3642 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3645 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3647 ack_ev_flags
|= CA_ACK_ECE
;
3650 if (flag
& FLAG_WIN_UPDATE
)
3651 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3653 tcp_in_ack_event(sk
, ack_ev_flags
);
3656 /* We passed data and got it acked, remove any soft error
3657 * log. Something worked...
3659 sk
->sk_err_soft
= 0;
3660 icsk
->icsk_probes_out
= 0;
3661 tp
->rcv_tstamp
= tcp_jiffies32
;
3665 /* See if we can take anything off of the retransmit queue. */
3666 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3669 if (tp
->tlp_high_seq
)
3670 tcp_process_tlp_ack(sk
, ack
, flag
);
3671 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3672 if (flag
& FLAG_SET_XMIT_TIMER
)
3673 tcp_set_xmit_timer(sk
);
3675 if (tcp_ack_is_dubious(sk
, flag
)) {
3676 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3677 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3680 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3683 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3684 lost
= tp
->lost
- lost
; /* freshly marked lost */
3685 tcp_rate_gen(sk
, delivered
, lost
, is_sack_reneg
, sack_state
.rate
);
3686 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3687 tcp_xmit_recovery(sk
, rexmit
);
3691 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3692 if (flag
& FLAG_DSACKING_ACK
)
3693 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3694 /* If this ack opens up a zero window, clear backoff. It was
3695 * being used to time the probes, and is probably far higher than
3696 * it needs to be for normal retransmission.
3698 if (tcp_send_head(sk
))
3701 if (tp
->tlp_high_seq
)
3702 tcp_process_tlp_ack(sk
, ack
, flag
);
3706 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3710 /* If data was SACKed, tag it and see if we should send more data.
3711 * If data was DSACKed, see if we can undo a cwnd reduction.
3713 if (TCP_SKB_CB(skb
)->sacked
) {
3714 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3716 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3717 tcp_xmit_recovery(sk
, rexmit
);
3720 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3724 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3725 bool syn
, struct tcp_fastopen_cookie
*foc
,
3728 /* Valid only in SYN or SYN-ACK with an even length. */
3729 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3732 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3733 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3734 memcpy(foc
->val
, cookie
, len
);
3741 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3742 * But, this can also be called on packets in the established flow when
3743 * the fast version below fails.
3745 void tcp_parse_options(const struct net
*net
,
3746 const struct sk_buff
*skb
,
3747 struct tcp_options_received
*opt_rx
, int estab
,
3748 struct tcp_fastopen_cookie
*foc
)
3750 const unsigned char *ptr
;
3751 const struct tcphdr
*th
= tcp_hdr(skb
);
3752 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3754 ptr
= (const unsigned char *)(th
+ 1);
3755 opt_rx
->saw_tstamp
= 0;
3757 while (length
> 0) {
3758 int opcode
= *ptr
++;
3764 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3769 if (opsize
< 2) /* "silly options" */
3771 if (opsize
> length
)
3772 return; /* don't parse partial options */
3775 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3776 u16 in_mss
= get_unaligned_be16(ptr
);
3778 if (opt_rx
->user_mss
&&
3779 opt_rx
->user_mss
< in_mss
)
3780 in_mss
= opt_rx
->user_mss
;
3781 opt_rx
->mss_clamp
= in_mss
;
3786 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3787 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3788 __u8 snd_wscale
= *(__u8
*)ptr
;
3789 opt_rx
->wscale_ok
= 1;
3790 if (snd_wscale
> TCP_MAX_WSCALE
) {
3791 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3795 snd_wscale
= TCP_MAX_WSCALE
;
3797 opt_rx
->snd_wscale
= snd_wscale
;
3800 case TCPOPT_TIMESTAMP
:
3801 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3802 ((estab
&& opt_rx
->tstamp_ok
) ||
3803 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3804 opt_rx
->saw_tstamp
= 1;
3805 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3806 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3809 case TCPOPT_SACK_PERM
:
3810 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3811 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3812 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3813 tcp_sack_reset(opt_rx
);
3818 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3819 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3821 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3824 #ifdef CONFIG_TCP_MD5SIG
3827 * The MD5 Hash has already been
3828 * checked (see tcp_v{4,6}_do_rcv()).
3832 case TCPOPT_FASTOPEN
:
3833 tcp_parse_fastopen_option(
3834 opsize
- TCPOLEN_FASTOPEN_BASE
,
3835 ptr
, th
->syn
, foc
, false);
3839 /* Fast Open option shares code 254 using a
3840 * 16 bits magic number.
3842 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3843 get_unaligned_be16(ptr
) ==
3844 TCPOPT_FASTOPEN_MAGIC
)
3845 tcp_parse_fastopen_option(opsize
-
3846 TCPOLEN_EXP_FASTOPEN_BASE
,
3847 ptr
+ 2, th
->syn
, foc
, true);
3856 EXPORT_SYMBOL(tcp_parse_options
);
3858 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3860 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3862 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3863 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3864 tp
->rx_opt
.saw_tstamp
= 1;
3866 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3869 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3871 tp
->rx_opt
.rcv_tsecr
= 0;
3877 /* Fast parse options. This hopes to only see timestamps.
3878 * If it is wrong it falls back on tcp_parse_options().
3880 static bool tcp_fast_parse_options(const struct net
*net
,
3881 const struct sk_buff
*skb
,
3882 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3884 /* In the spirit of fast parsing, compare doff directly to constant
3885 * values. Because equality is used, short doff can be ignored here.
3887 if (th
->doff
== (sizeof(*th
) / 4)) {
3888 tp
->rx_opt
.saw_tstamp
= 0;
3890 } else if (tp
->rx_opt
.tstamp_ok
&&
3891 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3892 if (tcp_parse_aligned_timestamp(tp
, th
))
3896 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3897 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3898 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3903 #ifdef CONFIG_TCP_MD5SIG
3905 * Parse MD5 Signature option
3907 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3909 int length
= (th
->doff
<< 2) - sizeof(*th
);
3910 const u8
*ptr
= (const u8
*)(th
+ 1);
3912 /* If not enough data remaining, we can short cut */
3913 while (length
>= TCPOLEN_MD5SIG
) {
3914 int opcode
= *ptr
++;
3925 if (opsize
< 2 || opsize
> length
)
3927 if (opcode
== TCPOPT_MD5SIG
)
3928 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3935 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3938 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3940 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3941 * it can pass through stack. So, the following predicate verifies that
3942 * this segment is not used for anything but congestion avoidance or
3943 * fast retransmit. Moreover, we even are able to eliminate most of such
3944 * second order effects, if we apply some small "replay" window (~RTO)
3945 * to timestamp space.
3947 * All these measures still do not guarantee that we reject wrapped ACKs
3948 * on networks with high bandwidth, when sequence space is recycled fastly,
3949 * but it guarantees that such events will be very rare and do not affect
3950 * connection seriously. This doesn't look nice, but alas, PAWS is really
3953 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3954 * states that events when retransmit arrives after original data are rare.
3955 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3956 * the biggest problem on large power networks even with minor reordering.
3957 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3958 * up to bandwidth of 18Gigabit/sec. 8) ]
3961 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3963 const struct tcp_sock
*tp
= tcp_sk(sk
);
3964 const struct tcphdr
*th
= tcp_hdr(skb
);
3965 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3966 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3968 return (/* 1. Pure ACK with correct sequence number. */
3969 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3971 /* 2. ... and duplicate ACK. */
3972 ack
== tp
->snd_una
&&
3974 /* 3. ... and does not update window. */
3975 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3977 /* 4. ... and sits in replay window. */
3978 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3981 static inline bool tcp_paws_discard(const struct sock
*sk
,
3982 const struct sk_buff
*skb
)
3984 const struct tcp_sock
*tp
= tcp_sk(sk
);
3986 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3987 !tcp_disordered_ack(sk
, skb
);
3990 /* Check segment sequence number for validity.
3992 * Segment controls are considered valid, if the segment
3993 * fits to the window after truncation to the window. Acceptability
3994 * of data (and SYN, FIN, of course) is checked separately.
3995 * See tcp_data_queue(), for example.
3997 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3998 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3999 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4000 * (borrowed from freebsd)
4003 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4005 return !before(end_seq
, tp
->rcv_wup
) &&
4006 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4009 /* When we get a reset we do this. */
4010 void tcp_reset(struct sock
*sk
)
4012 /* We want the right error as BSD sees it (and indeed as we do). */
4013 switch (sk
->sk_state
) {
4015 sk
->sk_err
= ECONNREFUSED
;
4017 case TCP_CLOSE_WAIT
:
4023 sk
->sk_err
= ECONNRESET
;
4025 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4028 tcp_write_queue_purge(sk
);
4031 if (!sock_flag(sk
, SOCK_DEAD
))
4032 sk
->sk_error_report(sk
);
4036 * Process the FIN bit. This now behaves as it is supposed to work
4037 * and the FIN takes effect when it is validly part of sequence
4038 * space. Not before when we get holes.
4040 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4041 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4044 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4045 * close and we go into CLOSING (and later onto TIME-WAIT)
4047 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4049 void tcp_fin(struct sock
*sk
)
4051 struct tcp_sock
*tp
= tcp_sk(sk
);
4053 inet_csk_schedule_ack(sk
);
4055 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4056 sock_set_flag(sk
, SOCK_DONE
);
4058 switch (sk
->sk_state
) {
4060 case TCP_ESTABLISHED
:
4061 /* Move to CLOSE_WAIT */
4062 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4063 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4066 case TCP_CLOSE_WAIT
:
4068 /* Received a retransmission of the FIN, do
4073 /* RFC793: Remain in the LAST-ACK state. */
4077 /* This case occurs when a simultaneous close
4078 * happens, we must ack the received FIN and
4079 * enter the CLOSING state.
4082 tcp_set_state(sk
, TCP_CLOSING
);
4085 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4087 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4090 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4091 * cases we should never reach this piece of code.
4093 pr_err("%s: Impossible, sk->sk_state=%d\n",
4094 __func__
, sk
->sk_state
);
4098 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4099 * Probably, we should reset in this case. For now drop them.
4101 skb_rbtree_purge(&tp
->out_of_order_queue
);
4102 if (tcp_is_sack(tp
))
4103 tcp_sack_reset(&tp
->rx_opt
);
4106 if (!sock_flag(sk
, SOCK_DEAD
)) {
4107 sk
->sk_state_change(sk
);
4109 /* Do not send POLL_HUP for half duplex close. */
4110 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4111 sk
->sk_state
== TCP_CLOSE
)
4112 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4114 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4118 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4121 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4122 if (before(seq
, sp
->start_seq
))
4123 sp
->start_seq
= seq
;
4124 if (after(end_seq
, sp
->end_seq
))
4125 sp
->end_seq
= end_seq
;
4131 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4133 struct tcp_sock
*tp
= tcp_sk(sk
);
4135 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4138 if (before(seq
, tp
->rcv_nxt
))
4139 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4141 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4143 NET_INC_STATS(sock_net(sk
), mib_idx
);
4145 tp
->rx_opt
.dsack
= 1;
4146 tp
->duplicate_sack
[0].start_seq
= seq
;
4147 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4151 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4153 struct tcp_sock
*tp
= tcp_sk(sk
);
4155 if (!tp
->rx_opt
.dsack
)
4156 tcp_dsack_set(sk
, seq
, end_seq
);
4158 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4161 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4163 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4166 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4167 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4168 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4170 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4171 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4173 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4174 end_seq
= tp
->rcv_nxt
;
4175 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4182 /* These routines update the SACK block as out-of-order packets arrive or
4183 * in-order packets close up the sequence space.
4185 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4188 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4189 struct tcp_sack_block
*swalk
= sp
+ 1;
4191 /* See if the recent change to the first SACK eats into
4192 * or hits the sequence space of other SACK blocks, if so coalesce.
4194 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4195 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4198 /* Zap SWALK, by moving every further SACK up by one slot.
4199 * Decrease num_sacks.
4201 tp
->rx_opt
.num_sacks
--;
4202 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4206 this_sack
++, swalk
++;
4210 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4212 struct tcp_sock
*tp
= tcp_sk(sk
);
4213 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4214 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4220 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4221 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4222 /* Rotate this_sack to the first one. */
4223 for (; this_sack
> 0; this_sack
--, sp
--)
4224 swap(*sp
, *(sp
- 1));
4226 tcp_sack_maybe_coalesce(tp
);
4231 /* Could not find an adjacent existing SACK, build a new one,
4232 * put it at the front, and shift everyone else down. We
4233 * always know there is at least one SACK present already here.
4235 * If the sack array is full, forget about the last one.
4237 if (this_sack
>= TCP_NUM_SACKS
) {
4239 tp
->rx_opt
.num_sacks
--;
4242 for (; this_sack
> 0; this_sack
--, sp
--)
4246 /* Build the new head SACK, and we're done. */
4247 sp
->start_seq
= seq
;
4248 sp
->end_seq
= end_seq
;
4249 tp
->rx_opt
.num_sacks
++;
4252 /* RCV.NXT advances, some SACKs should be eaten. */
4254 static void tcp_sack_remove(struct tcp_sock
*tp
)
4256 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4257 int num_sacks
= tp
->rx_opt
.num_sacks
;
4260 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4261 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4262 tp
->rx_opt
.num_sacks
= 0;
4266 for (this_sack
= 0; this_sack
< num_sacks
;) {
4267 /* Check if the start of the sack is covered by RCV.NXT. */
4268 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4271 /* RCV.NXT must cover all the block! */
4272 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4274 /* Zap this SACK, by moving forward any other SACKS. */
4275 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4276 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4283 tp
->rx_opt
.num_sacks
= num_sacks
;
4292 * tcp_try_coalesce - try to merge skb to prior one
4294 * @dest: destination queue
4296 * @from: buffer to add in queue
4297 * @fragstolen: pointer to boolean
4299 * Before queueing skb @from after @to, try to merge them
4300 * to reduce overall memory use and queue lengths, if cost is small.
4301 * Packets in ofo or receive queues can stay a long time.
4302 * Better try to coalesce them right now to avoid future collapses.
4303 * Returns true if caller should free @from instead of queueing it
4305 static bool tcp_try_coalesce(struct sock
*sk
,
4306 enum tcp_queue dest
,
4308 struct sk_buff
*from
,
4313 *fragstolen
= false;
4315 /* Its possible this segment overlaps with prior segment in queue */
4316 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4319 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4322 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4323 sk_mem_charge(sk
, delta
);
4324 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4325 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4326 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4327 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4329 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4330 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4331 if (dest
== OOO_QUEUE
)
4332 TCP_SKB_CB(to
)->swtstamp
= TCP_SKB_CB(from
)->swtstamp
;
4334 to
->tstamp
= from
->tstamp
;
4340 static bool tcp_ooo_try_coalesce(struct sock
*sk
,
4342 struct sk_buff
*from
,
4345 bool res
= tcp_try_coalesce(sk
, OOO_QUEUE
, to
, from
, fragstolen
);
4347 /* In case tcp_drop() is called later, update to->gso_segs */
4349 u32 gso_segs
= max_t(u16
, 1, skb_shinfo(to
)->gso_segs
) +
4350 max_t(u16
, 1, skb_shinfo(from
)->gso_segs
);
4352 skb_shinfo(to
)->gso_segs
= min_t(u32
, gso_segs
, 0xFFFF);
4357 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4359 sk_drops_add(sk
, skb
);
4363 /* This one checks to see if we can put data from the
4364 * out_of_order queue into the receive_queue.
4366 static void tcp_ofo_queue(struct sock
*sk
)
4368 struct tcp_sock
*tp
= tcp_sk(sk
);
4369 __u32 dsack_high
= tp
->rcv_nxt
;
4370 bool fin
, fragstolen
, eaten
;
4371 struct sk_buff
*skb
, *tail
;
4374 p
= rb_first(&tp
->out_of_order_queue
);
4377 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4380 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4381 __u32 dsack
= dsack_high
;
4382 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4383 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4384 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4387 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4388 /* Replace tstamp which was stomped by rbnode */
4389 if (TCP_SKB_CB(skb
)->has_rxtstamp
)
4390 skb
->tstamp
= TCP_SKB_CB(skb
)->swtstamp
;
4392 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4393 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4397 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4398 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4399 TCP_SKB_CB(skb
)->end_seq
);
4401 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4402 eaten
= tail
&& tcp_try_coalesce(sk
, RCV_QUEUE
,
4403 tail
, skb
, &fragstolen
);
4404 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4405 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4407 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4409 kfree_skb_partial(skb
, fragstolen
);
4411 if (unlikely(fin
)) {
4413 /* tcp_fin() purges tp->out_of_order_queue,
4414 * so we must end this loop right now.
4421 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4422 static int tcp_prune_queue(struct sock
*sk
);
4424 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4427 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4428 !sk_rmem_schedule(sk
, skb
, size
)) {
4430 if (tcp_prune_queue(sk
) < 0)
4433 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4434 if (!tcp_prune_ofo_queue(sk
))
4441 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4443 struct tcp_sock
*tp
= tcp_sk(sk
);
4444 struct rb_node
**p
, *parent
;
4445 struct sk_buff
*skb1
;
4449 tcp_ecn_check_ce(sk
, skb
);
4451 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4452 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4457 /* Stash tstamp to avoid being stomped on by rbnode */
4458 if (TCP_SKB_CB(skb
)->has_rxtstamp
)
4459 TCP_SKB_CB(skb
)->swtstamp
= skb
->tstamp
;
4461 /* Disable header prediction. */
4463 inet_csk_schedule_ack(sk
);
4465 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4466 seq
= TCP_SKB_CB(skb
)->seq
;
4467 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4468 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4469 tp
->rcv_nxt
, seq
, end_seq
);
4471 p
= &tp
->out_of_order_queue
.rb_node
;
4472 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4473 /* Initial out of order segment, build 1 SACK. */
4474 if (tcp_is_sack(tp
)) {
4475 tp
->rx_opt
.num_sacks
= 1;
4476 tp
->selective_acks
[0].start_seq
= seq
;
4477 tp
->selective_acks
[0].end_seq
= end_seq
;
4479 rb_link_node(&skb
->rbnode
, NULL
, p
);
4480 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4481 tp
->ooo_last_skb
= skb
;
4485 /* In the typical case, we are adding an skb to the end of the list.
4486 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4488 if (tcp_ooo_try_coalesce(sk
, tp
->ooo_last_skb
,
4489 skb
, &fragstolen
)) {
4491 tcp_grow_window(sk
, skb
);
4492 kfree_skb_partial(skb
, fragstolen
);
4496 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4497 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4498 parent
= &tp
->ooo_last_skb
->rbnode
;
4499 p
= &parent
->rb_right
;
4503 /* Find place to insert this segment. Handle overlaps on the way. */
4507 skb1
= rb_to_skb(parent
);
4508 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4509 p
= &parent
->rb_left
;
4512 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4513 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4514 /* All the bits are present. Drop. */
4515 NET_INC_STATS(sock_net(sk
),
4516 LINUX_MIB_TCPOFOMERGE
);
4519 tcp_dsack_set(sk
, seq
, end_seq
);
4522 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4523 /* Partial overlap. */
4524 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4526 /* skb's seq == skb1's seq and skb covers skb1.
4527 * Replace skb1 with skb.
4529 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4530 &tp
->out_of_order_queue
);
4531 tcp_dsack_extend(sk
,
4532 TCP_SKB_CB(skb1
)->seq
,
4533 TCP_SKB_CB(skb1
)->end_seq
);
4534 NET_INC_STATS(sock_net(sk
),
4535 LINUX_MIB_TCPOFOMERGE
);
4539 } else if (tcp_ooo_try_coalesce(sk
, skb1
,
4540 skb
, &fragstolen
)) {
4543 p
= &parent
->rb_right
;
4546 /* Insert segment into RB tree. */
4547 rb_link_node(&skb
->rbnode
, parent
, p
);
4548 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4551 /* Remove other segments covered by skb. */
4552 while ((skb1
= skb_rb_next(skb
)) != NULL
) {
4553 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4555 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4556 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4560 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4561 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4562 TCP_SKB_CB(skb1
)->end_seq
);
4563 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4566 /* If there is no skb after us, we are the last_skb ! */
4568 tp
->ooo_last_skb
= skb
;
4571 if (tcp_is_sack(tp
))
4572 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4575 tcp_grow_window(sk
, skb
);
4577 skb_set_owner_r(skb
, sk
);
4581 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4585 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4587 __skb_pull(skb
, hdrlen
);
4589 tcp_try_coalesce(sk
, RCV_QUEUE
, tail
,
4590 skb
, fragstolen
)) ? 1 : 0;
4591 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4593 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4594 skb_set_owner_r(skb
, sk
);
4599 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4601 struct sk_buff
*skb
;
4609 if (size
> PAGE_SIZE
) {
4610 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4612 data_len
= npages
<< PAGE_SHIFT
;
4613 size
= data_len
+ (size
& ~PAGE_MASK
);
4615 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4616 PAGE_ALLOC_COSTLY_ORDER
,
4617 &err
, sk
->sk_allocation
);
4621 skb_put(skb
, size
- data_len
);
4622 skb
->data_len
= data_len
;
4625 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4628 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4632 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4633 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4634 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4636 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4637 WARN_ON_ONCE(fragstolen
); /* should not happen */
4649 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4651 struct tcp_sock
*tp
= tcp_sk(sk
);
4655 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4660 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4662 tcp_ecn_accept_cwr(tp
, skb
);
4664 tp
->rx_opt
.dsack
= 0;
4666 /* Queue data for delivery to the user.
4667 * Packets in sequence go to the receive queue.
4668 * Out of sequence packets to the out_of_order_queue.
4670 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4671 if (tcp_receive_window(tp
) == 0)
4674 /* Ok. In sequence. In window. */
4676 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4677 sk_forced_mem_schedule(sk
, skb
->truesize
);
4678 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4681 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4682 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4684 tcp_event_data_recv(sk
, skb
);
4685 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4688 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4691 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4692 * gap in queue is filled.
4694 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4695 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4698 if (tp
->rx_opt
.num_sacks
)
4699 tcp_sack_remove(tp
);
4701 tcp_fast_path_check(sk
);
4704 kfree_skb_partial(skb
, fragstolen
);
4705 if (!sock_flag(sk
, SOCK_DEAD
))
4706 sk
->sk_data_ready(sk
);
4710 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4711 /* A retransmit, 2nd most common case. Force an immediate ack. */
4712 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4713 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4716 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4717 inet_csk_schedule_ack(sk
);
4723 /* Out of window. F.e. zero window probe. */
4724 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4727 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4728 /* Partial packet, seq < rcv_next < end_seq */
4729 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4730 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4731 TCP_SKB_CB(skb
)->end_seq
);
4733 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4735 /* If window is closed, drop tail of packet. But after
4736 * remembering D-SACK for its head made in previous line.
4738 if (!tcp_receive_window(tp
))
4743 tcp_data_queue_ofo(sk
, skb
);
4746 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4749 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4751 return skb_rb_next(skb
);
4754 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4755 struct sk_buff_head
*list
,
4756 struct rb_root
*root
)
4758 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4761 __skb_unlink(skb
, list
);
4763 rb_erase(&skb
->rbnode
, root
);
4766 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4771 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4772 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4774 struct rb_node
**p
= &root
->rb_node
;
4775 struct rb_node
*parent
= NULL
;
4776 struct sk_buff
*skb1
;
4780 skb1
= rb_to_skb(parent
);
4781 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4782 p
= &parent
->rb_left
;
4784 p
= &parent
->rb_right
;
4786 rb_link_node(&skb
->rbnode
, parent
, p
);
4787 rb_insert_color(&skb
->rbnode
, root
);
4790 /* Collapse contiguous sequence of skbs head..tail with
4791 * sequence numbers start..end.
4793 * If tail is NULL, this means until the end of the queue.
4795 * Segments with FIN/SYN are not collapsed (only because this
4799 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4800 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4802 struct sk_buff
*skb
= head
, *n
;
4803 struct sk_buff_head tmp
;
4806 /* First, check that queue is collapsible and find
4807 * the point where collapsing can be useful.
4810 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4811 n
= tcp_skb_next(skb
, list
);
4813 /* No new bits? It is possible on ofo queue. */
4814 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4815 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4821 /* The first skb to collapse is:
4823 * - bloated or contains data before "start" or
4824 * overlaps to the next one.
4826 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4827 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4828 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4829 end_of_skbs
= false;
4833 if (n
&& n
!= tail
&&
4834 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4835 end_of_skbs
= false;
4839 /* Decided to skip this, advance start seq. */
4840 start
= TCP_SKB_CB(skb
)->end_seq
;
4843 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4846 __skb_queue_head_init(&tmp
);
4848 while (before(start
, end
)) {
4849 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4850 struct sk_buff
*nskb
;
4852 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4856 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4857 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4859 __skb_queue_before(list
, skb
, nskb
);
4861 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4862 skb_set_owner_r(nskb
, sk
);
4864 /* Copy data, releasing collapsed skbs. */
4866 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4867 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4871 size
= min(copy
, size
);
4872 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4874 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4878 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4879 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4882 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4888 skb_queue_walk_safe(&tmp
, skb
, n
)
4889 tcp_rbtree_insert(root
, skb
);
4892 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4893 * and tcp_collapse() them until all the queue is collapsed.
4895 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4897 struct tcp_sock
*tp
= tcp_sk(sk
);
4898 u32 range_truesize
, sum_tiny
= 0;
4899 struct sk_buff
*skb
, *head
;
4902 skb
= skb_rb_first(&tp
->out_of_order_queue
);
4905 tp
->ooo_last_skb
= skb_rb_last(&tp
->out_of_order_queue
);
4908 start
= TCP_SKB_CB(skb
)->seq
;
4909 end
= TCP_SKB_CB(skb
)->end_seq
;
4910 range_truesize
= skb
->truesize
;
4912 for (head
= skb
;;) {
4913 skb
= skb_rb_next(skb
);
4915 /* Range is terminated when we see a gap or when
4916 * we are at the queue end.
4919 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4920 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4921 /* Do not attempt collapsing tiny skbs */
4922 if (range_truesize
!= head
->truesize
||
4923 end
- start
>= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM
)) {
4924 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4925 head
, skb
, start
, end
);
4927 sum_tiny
+= range_truesize
;
4928 if (sum_tiny
> sk
->sk_rcvbuf
>> 3)
4934 range_truesize
+= skb
->truesize
;
4935 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4936 start
= TCP_SKB_CB(skb
)->seq
;
4937 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4938 end
= TCP_SKB_CB(skb
)->end_seq
;
4943 * Clean the out-of-order queue to make room.
4944 * We drop high sequences packets to :
4945 * 1) Let a chance for holes to be filled.
4946 * 2) not add too big latencies if thousands of packets sit there.
4947 * (But if application shrinks SO_RCVBUF, we could still end up
4948 * freeing whole queue here)
4949 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
4951 * Return true if queue has shrunk.
4953 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4955 struct tcp_sock
*tp
= tcp_sk(sk
);
4956 struct rb_node
*node
, *prev
;
4959 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4962 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4963 goal
= sk
->sk_rcvbuf
>> 3;
4964 node
= &tp
->ooo_last_skb
->rbnode
;
4966 prev
= rb_prev(node
);
4967 rb_erase(node
, &tp
->out_of_order_queue
);
4968 goal
-= rb_to_skb(node
)->truesize
;
4969 tcp_drop(sk
, rb_to_skb(node
));
4970 if (!prev
|| goal
<= 0) {
4972 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4973 !tcp_under_memory_pressure(sk
))
4975 goal
= sk
->sk_rcvbuf
>> 3;
4979 tp
->ooo_last_skb
= rb_to_skb(prev
);
4981 /* Reset SACK state. A conforming SACK implementation will
4982 * do the same at a timeout based retransmit. When a connection
4983 * is in a sad state like this, we care only about integrity
4984 * of the connection not performance.
4986 if (tp
->rx_opt
.sack_ok
)
4987 tcp_sack_reset(&tp
->rx_opt
);
4991 /* Reduce allocated memory if we can, trying to get
4992 * the socket within its memory limits again.
4994 * Return less than zero if we should start dropping frames
4995 * until the socket owning process reads some of the data
4996 * to stabilize the situation.
4998 static int tcp_prune_queue(struct sock
*sk
)
5000 struct tcp_sock
*tp
= tcp_sk(sk
);
5002 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
5004 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
5006 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
5007 tcp_clamp_window(sk
);
5008 else if (tcp_under_memory_pressure(sk
))
5009 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
5011 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5014 tcp_collapse_ofo_queue(sk
);
5015 if (!skb_queue_empty(&sk
->sk_receive_queue
))
5016 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
5017 skb_peek(&sk
->sk_receive_queue
),
5019 tp
->copied_seq
, tp
->rcv_nxt
);
5022 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5025 /* Collapsing did not help, destructive actions follow.
5026 * This must not ever occur. */
5028 tcp_prune_ofo_queue(sk
);
5030 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5033 /* If we are really being abused, tell the caller to silently
5034 * drop receive data on the floor. It will get retransmitted
5035 * and hopefully then we'll have sufficient space.
5037 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
5039 /* Massive buffer overcommit. */
5044 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5046 const struct tcp_sock
*tp
= tcp_sk(sk
);
5048 /* If the user specified a specific send buffer setting, do
5051 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5054 /* If we are under global TCP memory pressure, do not expand. */
5055 if (tcp_under_memory_pressure(sk
))
5058 /* If we are under soft global TCP memory pressure, do not expand. */
5059 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5062 /* If we filled the congestion window, do not expand. */
5063 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5069 /* When incoming ACK allowed to free some skb from write_queue,
5070 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5071 * on the exit from tcp input handler.
5073 * PROBLEM: sndbuf expansion does not work well with largesend.
5075 static void tcp_new_space(struct sock
*sk
)
5077 struct tcp_sock
*tp
= tcp_sk(sk
);
5079 if (tcp_should_expand_sndbuf(sk
)) {
5080 tcp_sndbuf_expand(sk
);
5081 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5084 sk
->sk_write_space(sk
);
5087 static void tcp_check_space(struct sock
*sk
)
5089 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5090 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5091 /* pairs with tcp_poll() */
5093 if (sk
->sk_socket
&&
5094 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5096 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5097 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5102 static inline void tcp_data_snd_check(struct sock
*sk
)
5104 tcp_push_pending_frames(sk
);
5105 tcp_check_space(sk
);
5109 * Check if sending an ack is needed.
5111 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5113 struct tcp_sock
*tp
= tcp_sk(sk
);
5115 /* More than one full frame received... */
5116 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5117 /* ... and right edge of window advances far enough.
5118 * (tcp_recvmsg() will send ACK otherwise). Or...
5120 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5121 /* We ACK each frame or... */
5122 tcp_in_quickack_mode(sk
) ||
5123 /* We have out of order data. */
5124 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5125 /* Then ack it now */
5128 /* Else, send delayed ack. */
5129 tcp_send_delayed_ack(sk
);
5133 static inline void tcp_ack_snd_check(struct sock
*sk
)
5135 if (!inet_csk_ack_scheduled(sk
)) {
5136 /* We sent a data segment already. */
5139 __tcp_ack_snd_check(sk
, 1);
5143 * This routine is only called when we have urgent data
5144 * signaled. Its the 'slow' part of tcp_urg. It could be
5145 * moved inline now as tcp_urg is only called from one
5146 * place. We handle URGent data wrong. We have to - as
5147 * BSD still doesn't use the correction from RFC961.
5148 * For 1003.1g we should support a new option TCP_STDURG to permit
5149 * either form (or just set the sysctl tcp_stdurg).
5152 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5154 struct tcp_sock
*tp
= tcp_sk(sk
);
5155 u32 ptr
= ntohs(th
->urg_ptr
);
5157 if (ptr
&& !sysctl_tcp_stdurg
)
5159 ptr
+= ntohl(th
->seq
);
5161 /* Ignore urgent data that we've already seen and read. */
5162 if (after(tp
->copied_seq
, ptr
))
5165 /* Do not replay urg ptr.
5167 * NOTE: interesting situation not covered by specs.
5168 * Misbehaving sender may send urg ptr, pointing to segment,
5169 * which we already have in ofo queue. We are not able to fetch
5170 * such data and will stay in TCP_URG_NOTYET until will be eaten
5171 * by recvmsg(). Seems, we are not obliged to handle such wicked
5172 * situations. But it is worth to think about possibility of some
5173 * DoSes using some hypothetical application level deadlock.
5175 if (before(ptr
, tp
->rcv_nxt
))
5178 /* Do we already have a newer (or duplicate) urgent pointer? */
5179 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5182 /* Tell the world about our new urgent pointer. */
5185 /* We may be adding urgent data when the last byte read was
5186 * urgent. To do this requires some care. We cannot just ignore
5187 * tp->copied_seq since we would read the last urgent byte again
5188 * as data, nor can we alter copied_seq until this data arrives
5189 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5191 * NOTE. Double Dutch. Rendering to plain English: author of comment
5192 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5193 * and expect that both A and B disappear from stream. This is _wrong_.
5194 * Though this happens in BSD with high probability, this is occasional.
5195 * Any application relying on this is buggy. Note also, that fix "works"
5196 * only in this artificial test. Insert some normal data between A and B and we will
5197 * decline of BSD again. Verdict: it is better to remove to trap
5200 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5201 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5202 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5204 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5205 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5210 tp
->urg_data
= TCP_URG_NOTYET
;
5213 /* Disable header prediction. */
5217 /* This is the 'fast' part of urgent handling. */
5218 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5220 struct tcp_sock
*tp
= tcp_sk(sk
);
5222 /* Check if we get a new urgent pointer - normally not. */
5224 tcp_check_urg(sk
, th
);
5226 /* Do we wait for any urgent data? - normally not... */
5227 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5228 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5231 /* Is the urgent pointer pointing into this packet? */
5232 if (ptr
< skb
->len
) {
5234 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5236 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5237 if (!sock_flag(sk
, SOCK_DEAD
))
5238 sk
->sk_data_ready(sk
);
5243 /* Accept RST for rcv_nxt - 1 after a FIN.
5244 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5245 * FIN is sent followed by a RST packet. The RST is sent with the same
5246 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5247 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5248 * ACKs on the closed socket. In addition middleboxes can drop either the
5249 * challenge ACK or a subsequent RST.
5251 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5253 struct tcp_sock
*tp
= tcp_sk(sk
);
5255 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5256 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5260 /* Does PAWS and seqno based validation of an incoming segment, flags will
5261 * play significant role here.
5263 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5264 const struct tcphdr
*th
, int syn_inerr
)
5266 struct tcp_sock
*tp
= tcp_sk(sk
);
5267 bool rst_seq_match
= false;
5269 /* RFC1323: H1. Apply PAWS check first. */
5270 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5271 tp
->rx_opt
.saw_tstamp
&&
5272 tcp_paws_discard(sk
, skb
)) {
5274 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5275 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5276 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5277 &tp
->last_oow_ack_time
))
5278 tcp_send_dupack(sk
, skb
);
5281 /* Reset is accepted even if it did not pass PAWS. */
5284 /* Step 1: check sequence number */
5285 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5286 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5287 * (RST) segments are validated by checking their SEQ-fields."
5288 * And page 69: "If an incoming segment is not acceptable,
5289 * an acknowledgment should be sent in reply (unless the RST
5290 * bit is set, if so drop the segment and return)".
5295 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5296 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5297 &tp
->last_oow_ack_time
))
5298 tcp_send_dupack(sk
, skb
);
5299 } else if (tcp_reset_check(sk
, skb
)) {
5305 /* Step 2: check RST bit */
5307 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5308 * FIN and SACK too if available):
5309 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5310 * the right-most SACK block,
5312 * RESET the connection
5314 * Send a challenge ACK
5316 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5317 tcp_reset_check(sk
, skb
)) {
5318 rst_seq_match
= true;
5319 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5320 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5321 int max_sack
= sp
[0].end_seq
;
5324 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5326 max_sack
= after(sp
[this_sack
].end_seq
,
5328 sp
[this_sack
].end_seq
: max_sack
;
5331 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5332 rst_seq_match
= true;
5338 /* Disable TFO if RST is out-of-order
5339 * and no data has been received
5340 * for current active TFO socket
5342 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5343 sk
->sk_state
== TCP_ESTABLISHED
)
5344 tcp_fastopen_active_disable(sk
);
5345 tcp_send_challenge_ack(sk
, skb
);
5350 /* step 3: check security and precedence [ignored] */
5352 /* step 4: Check for a SYN
5353 * RFC 5961 4.2 : Send a challenge ack
5358 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5359 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5360 tcp_send_challenge_ack(sk
, skb
);
5372 * TCP receive function for the ESTABLISHED state.
5374 * It is split into a fast path and a slow path. The fast path is
5376 * - A zero window was announced from us - zero window probing
5377 * is only handled properly in the slow path.
5378 * - Out of order segments arrived.
5379 * - Urgent data is expected.
5380 * - There is no buffer space left
5381 * - Unexpected TCP flags/window values/header lengths are received
5382 * (detected by checking the TCP header against pred_flags)
5383 * - Data is sent in both directions. Fast path only supports pure senders
5384 * or pure receivers (this means either the sequence number or the ack
5385 * value must stay constant)
5386 * - Unexpected TCP option.
5388 * When these conditions are not satisfied it drops into a standard
5389 * receive procedure patterned after RFC793 to handle all cases.
5390 * The first three cases are guaranteed by proper pred_flags setting,
5391 * the rest is checked inline. Fast processing is turned on in
5392 * tcp_data_queue when everything is OK.
5394 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5395 const struct tcphdr
*th
)
5397 unsigned int len
= skb
->len
;
5398 struct tcp_sock
*tp
= tcp_sk(sk
);
5400 tcp_mstamp_refresh(tp
);
5401 if (unlikely(!sk
->sk_rx_dst
))
5402 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5404 * Header prediction.
5405 * The code loosely follows the one in the famous
5406 * "30 instruction TCP receive" Van Jacobson mail.
5408 * Van's trick is to deposit buffers into socket queue
5409 * on a device interrupt, to call tcp_recv function
5410 * on the receive process context and checksum and copy
5411 * the buffer to user space. smart...
5413 * Our current scheme is not silly either but we take the
5414 * extra cost of the net_bh soft interrupt processing...
5415 * We do checksum and copy also but from device to kernel.
5418 tp
->rx_opt
.saw_tstamp
= 0;
5420 /* pred_flags is 0xS?10 << 16 + snd_wnd
5421 * if header_prediction is to be made
5422 * 'S' will always be tp->tcp_header_len >> 2
5423 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5424 * turn it off (when there are holes in the receive
5425 * space for instance)
5426 * PSH flag is ignored.
5429 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5430 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5431 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5432 int tcp_header_len
= tp
->tcp_header_len
;
5434 /* Timestamp header prediction: tcp_header_len
5435 * is automatically equal to th->doff*4 due to pred_flags
5439 /* Check timestamp */
5440 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5441 /* No? Slow path! */
5442 if (!tcp_parse_aligned_timestamp(tp
, th
))
5445 /* If PAWS failed, check it more carefully in slow path */
5446 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5449 /* DO NOT update ts_recent here, if checksum fails
5450 * and timestamp was corrupted part, it will result
5451 * in a hung connection since we will drop all
5452 * future packets due to the PAWS test.
5456 if (len
<= tcp_header_len
) {
5457 /* Bulk data transfer: sender */
5458 if (len
== tcp_header_len
) {
5459 /* Predicted packet is in window by definition.
5460 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5461 * Hence, check seq<=rcv_wup reduces to:
5463 if (tcp_header_len
==
5464 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5465 tp
->rcv_nxt
== tp
->rcv_wup
)
5466 tcp_store_ts_recent(tp
);
5468 /* We know that such packets are checksummed
5471 tcp_ack(sk
, skb
, 0);
5473 tcp_data_snd_check(sk
);
5475 } else { /* Header too small */
5476 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5481 bool fragstolen
= false;
5483 if (tcp_checksum_complete(skb
))
5486 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5489 /* Predicted packet is in window by definition.
5490 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5491 * Hence, check seq<=rcv_wup reduces to:
5493 if (tcp_header_len
==
5494 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5495 tp
->rcv_nxt
== tp
->rcv_wup
)
5496 tcp_store_ts_recent(tp
);
5498 tcp_rcv_rtt_measure_ts(sk
, skb
);
5500 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5502 /* Bulk data transfer: receiver */
5503 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5506 tcp_event_data_recv(sk
, skb
);
5508 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5509 /* Well, only one small jumplet in fast path... */
5510 tcp_ack(sk
, skb
, FLAG_DATA
);
5511 tcp_data_snd_check(sk
);
5512 if (!inet_csk_ack_scheduled(sk
))
5516 __tcp_ack_snd_check(sk
, 0);
5519 kfree_skb_partial(skb
, fragstolen
);
5520 sk
->sk_data_ready(sk
);
5526 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5529 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5533 * Standard slow path.
5536 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5540 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5543 tcp_rcv_rtt_measure_ts(sk
, skb
);
5545 /* Process urgent data. */
5546 tcp_urg(sk
, skb
, th
);
5548 /* step 7: process the segment text */
5549 tcp_data_queue(sk
, skb
);
5551 tcp_data_snd_check(sk
);
5552 tcp_ack_snd_check(sk
);
5556 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5557 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5562 EXPORT_SYMBOL(tcp_rcv_established
);
5564 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5566 struct tcp_sock
*tp
= tcp_sk(sk
);
5567 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5569 tcp_set_state(sk
, TCP_ESTABLISHED
);
5570 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5573 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5574 security_inet_conn_established(sk
, skb
);
5577 /* Make sure socket is routed, for correct metrics. */
5578 icsk
->icsk_af_ops
->rebuild_header(sk
);
5580 tcp_init_metrics(sk
);
5581 tcp_call_bpf(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5582 tcp_init_congestion_control(sk
);
5584 /* Prevent spurious tcp_cwnd_restart() on first data
5587 tp
->lsndtime
= tcp_jiffies32
;
5589 tcp_init_buffer_space(sk
);
5591 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5592 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5594 if (!tp
->rx_opt
.snd_wscale
)
5595 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5600 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5601 struct tcp_fastopen_cookie
*cookie
)
5603 struct tcp_sock
*tp
= tcp_sk(sk
);
5604 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5605 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5606 bool syn_drop
= false;
5608 if (mss
== tp
->rx_opt
.user_mss
) {
5609 struct tcp_options_received opt
;
5611 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5612 tcp_clear_options(&opt
);
5613 opt
.user_mss
= opt
.mss_clamp
= 0;
5614 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5615 mss
= opt
.mss_clamp
;
5618 if (!tp
->syn_fastopen
) {
5619 /* Ignore an unsolicited cookie */
5621 } else if (tp
->total_retrans
) {
5622 /* SYN timed out and the SYN-ACK neither has a cookie nor
5623 * acknowledges data. Presumably the remote received only
5624 * the retransmitted (regular) SYNs: either the original
5625 * SYN-data or the corresponding SYN-ACK was dropped.
5627 syn_drop
= (cookie
->len
< 0 && data
);
5628 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5629 /* We requested a cookie but didn't get it. If we did not use
5630 * the (old) exp opt format then try so next time (try_exp=1).
5631 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5633 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5636 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5638 if (data
) { /* Retransmit unacked data in SYN */
5639 tcp_for_write_queue_from(data
, sk
) {
5640 if (data
== tcp_send_head(sk
) ||
5641 __tcp_retransmit_skb(sk
, data
, 1))
5645 NET_INC_STATS(sock_net(sk
),
5646 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5649 tp
->syn_data_acked
= tp
->syn_data
;
5650 if (tp
->syn_data_acked
)
5651 NET_INC_STATS(sock_net(sk
),
5652 LINUX_MIB_TCPFASTOPENACTIVE
);
5654 tcp_fastopen_add_skb(sk
, synack
);
5659 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5660 const struct tcphdr
*th
)
5662 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5663 struct tcp_sock
*tp
= tcp_sk(sk
);
5664 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5665 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5668 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5669 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5670 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5674 * "If the state is SYN-SENT then
5675 * first check the ACK bit
5676 * If the ACK bit is set
5677 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5678 * a reset (unless the RST bit is set, if so drop
5679 * the segment and return)"
5681 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5682 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5683 goto reset_and_undo
;
5685 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5686 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5687 tcp_time_stamp(tp
))) {
5688 NET_INC_STATS(sock_net(sk
),
5689 LINUX_MIB_PAWSACTIVEREJECTED
);
5690 goto reset_and_undo
;
5693 /* Now ACK is acceptable.
5695 * "If the RST bit is set
5696 * If the ACK was acceptable then signal the user "error:
5697 * connection reset", drop the segment, enter CLOSED state,
5698 * delete TCB, and return."
5707 * "fifth, if neither of the SYN or RST bits is set then
5708 * drop the segment and return."
5714 goto discard_and_undo
;
5717 * "If the SYN bit is on ...
5718 * are acceptable then ...
5719 * (our SYN has been ACKed), change the connection
5720 * state to ESTABLISHED..."
5723 tcp_ecn_rcv_synack(tp
, th
);
5725 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5726 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5728 /* Ok.. it's good. Set up sequence numbers and
5729 * move to established.
5731 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5732 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5734 /* RFC1323: The window in SYN & SYN/ACK segments is
5737 tp
->snd_wnd
= ntohs(th
->window
);
5739 if (!tp
->rx_opt
.wscale_ok
) {
5740 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5741 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5744 if (tp
->rx_opt
.saw_tstamp
) {
5745 tp
->rx_opt
.tstamp_ok
= 1;
5746 tp
->tcp_header_len
=
5747 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5748 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5749 tcp_store_ts_recent(tp
);
5751 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5754 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5755 tcp_enable_fack(tp
);
5758 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5759 tcp_initialize_rcv_mss(sk
);
5761 /* Remember, tcp_poll() does not lock socket!
5762 * Change state from SYN-SENT only after copied_seq
5763 * is initialized. */
5764 tp
->copied_seq
= tp
->rcv_nxt
;
5768 tcp_finish_connect(sk
, skb
);
5770 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5771 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5773 if (!sock_flag(sk
, SOCK_DEAD
)) {
5774 sk
->sk_state_change(sk
);
5775 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5779 if (sk
->sk_write_pending
||
5780 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5781 icsk
->icsk_ack
.pingpong
) {
5782 /* Save one ACK. Data will be ready after
5783 * several ticks, if write_pending is set.
5785 * It may be deleted, but with this feature tcpdumps
5786 * look so _wonderfully_ clever, that I was not able
5787 * to stand against the temptation 8) --ANK
5789 inet_csk_schedule_ack(sk
);
5790 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
5791 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5792 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5803 /* No ACK in the segment */
5807 * "If the RST bit is set
5809 * Otherwise (no ACK) drop the segment and return."
5812 goto discard_and_undo
;
5816 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5817 tcp_paws_reject(&tp
->rx_opt
, 0))
5818 goto discard_and_undo
;
5821 /* We see SYN without ACK. It is attempt of
5822 * simultaneous connect with crossed SYNs.
5823 * Particularly, it can be connect to self.
5825 tcp_set_state(sk
, TCP_SYN_RECV
);
5827 if (tp
->rx_opt
.saw_tstamp
) {
5828 tp
->rx_opt
.tstamp_ok
= 1;
5829 tcp_store_ts_recent(tp
);
5830 tp
->tcp_header_len
=
5831 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5833 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5836 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5837 tp
->copied_seq
= tp
->rcv_nxt
;
5838 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5840 /* RFC1323: The window in SYN & SYN/ACK segments is
5843 tp
->snd_wnd
= ntohs(th
->window
);
5844 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5845 tp
->max_window
= tp
->snd_wnd
;
5847 tcp_ecn_rcv_syn(tp
, th
);
5850 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5851 tcp_initialize_rcv_mss(sk
);
5853 tcp_send_synack(sk
);
5855 /* Note, we could accept data and URG from this segment.
5856 * There are no obstacles to make this (except that we must
5857 * either change tcp_recvmsg() to prevent it from returning data
5858 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5860 * However, if we ignore data in ACKless segments sometimes,
5861 * we have no reasons to accept it sometimes.
5862 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5863 * is not flawless. So, discard packet for sanity.
5864 * Uncomment this return to process the data.
5871 /* "fifth, if neither of the SYN or RST bits is set then
5872 * drop the segment and return."
5876 tcp_clear_options(&tp
->rx_opt
);
5877 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5881 tcp_clear_options(&tp
->rx_opt
);
5882 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5887 * This function implements the receiving procedure of RFC 793 for
5888 * all states except ESTABLISHED and TIME_WAIT.
5889 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5890 * address independent.
5893 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5895 struct tcp_sock
*tp
= tcp_sk(sk
);
5896 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5897 const struct tcphdr
*th
= tcp_hdr(skb
);
5898 struct request_sock
*req
;
5902 switch (sk
->sk_state
) {
5916 /* It is possible that we process SYN packets from backlog,
5917 * so we need to make sure to disable BH and RCU right there.
5921 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5933 tp
->rx_opt
.saw_tstamp
= 0;
5934 tcp_mstamp_refresh(tp
);
5935 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5939 /* Do step6 onward by hand. */
5940 tcp_urg(sk
, skb
, th
);
5942 tcp_data_snd_check(sk
);
5946 tcp_mstamp_refresh(tp
);
5947 tp
->rx_opt
.saw_tstamp
= 0;
5948 req
= tp
->fastopen_rsk
;
5950 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5951 sk
->sk_state
!= TCP_FIN_WAIT1
);
5953 if (!tcp_check_req(sk
, skb
, req
, true))
5957 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5960 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5963 /* step 5: check the ACK field */
5964 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5965 FLAG_UPDATE_TS_RECENT
|
5966 FLAG_NO_CHALLENGE_ACK
) > 0;
5969 if (sk
->sk_state
== TCP_SYN_RECV
)
5970 return 1; /* send one RST */
5971 tcp_send_challenge_ack(sk
, skb
);
5974 switch (sk
->sk_state
) {
5977 tcp_synack_rtt_meas(sk
, req
);
5979 /* Once we leave TCP_SYN_RECV, we no longer need req
5983 inet_csk(sk
)->icsk_retransmits
= 0;
5984 reqsk_fastopen_remove(sk
, req
, false);
5986 /* Make sure socket is routed, for correct metrics. */
5987 icsk
->icsk_af_ops
->rebuild_header(sk
);
5988 tcp_call_bpf(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
5989 tcp_init_congestion_control(sk
);
5992 tp
->copied_seq
= tp
->rcv_nxt
;
5993 tcp_init_buffer_space(sk
);
5996 tcp_set_state(sk
, TCP_ESTABLISHED
);
5997 sk
->sk_state_change(sk
);
5999 /* Note, that this wakeup is only for marginal crossed SYN case.
6000 * Passively open sockets are not waked up, because
6001 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6004 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
6006 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6007 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
6008 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6010 if (tp
->rx_opt
.tstamp_ok
)
6011 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6014 /* Re-arm the timer because data may have been sent out.
6015 * This is similar to the regular data transmission case
6016 * when new data has just been ack'ed.
6018 * (TFO) - we could try to be more aggressive and
6019 * retransmitting any data sooner based on when they
6024 tcp_init_metrics(sk
);
6026 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
6027 tcp_update_pacing_rate(sk
);
6029 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6030 tp
->lsndtime
= tcp_jiffies32
;
6032 tcp_initialize_rcv_mss(sk
);
6033 tcp_fast_path_on(tp
);
6036 case TCP_FIN_WAIT1
: {
6039 /* If we enter the TCP_FIN_WAIT1 state and we are a
6040 * Fast Open socket and this is the first acceptable
6041 * ACK we have received, this would have acknowledged
6042 * our SYNACK so stop the SYNACK timer.
6045 /* We no longer need the request sock. */
6046 reqsk_fastopen_remove(sk
, req
, false);
6049 if (tp
->snd_una
!= tp
->write_seq
)
6052 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6053 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6057 if (!sock_flag(sk
, SOCK_DEAD
)) {
6058 /* Wake up lingering close() */
6059 sk
->sk_state_change(sk
);
6063 if (tp
->linger2
< 0) {
6065 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6068 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6069 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6070 /* Receive out of order FIN after close() */
6071 if (tp
->syn_fastopen
&& th
->fin
)
6072 tcp_fastopen_active_disable(sk
);
6074 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6078 tmo
= tcp_fin_time(sk
);
6079 if (tmo
> TCP_TIMEWAIT_LEN
) {
6080 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6081 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6082 /* Bad case. We could lose such FIN otherwise.
6083 * It is not a big problem, but it looks confusing
6084 * and not so rare event. We still can lose it now,
6085 * if it spins in bh_lock_sock(), but it is really
6088 inet_csk_reset_keepalive_timer(sk
, tmo
);
6090 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6097 if (tp
->snd_una
== tp
->write_seq
) {
6098 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6104 if (tp
->snd_una
== tp
->write_seq
) {
6105 tcp_update_metrics(sk
);
6112 /* step 6: check the URG bit */
6113 tcp_urg(sk
, skb
, th
);
6115 /* step 7: process the segment text */
6116 switch (sk
->sk_state
) {
6117 case TCP_CLOSE_WAIT
:
6120 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6124 /* RFC 793 says to queue data in these states,
6125 * RFC 1122 says we MUST send a reset.
6126 * BSD 4.4 also does reset.
6128 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6129 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6130 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6131 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6137 case TCP_ESTABLISHED
:
6138 tcp_data_queue(sk
, skb
);
6143 /* tcp_data could move socket to TIME-WAIT */
6144 if (sk
->sk_state
!= TCP_CLOSE
) {
6145 tcp_data_snd_check(sk
);
6146 tcp_ack_snd_check(sk
);
6155 EXPORT_SYMBOL(tcp_rcv_state_process
);
6157 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6159 struct inet_request_sock
*ireq
= inet_rsk(req
);
6161 if (family
== AF_INET
)
6162 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6163 &ireq
->ir_rmt_addr
, port
);
6164 #if IS_ENABLED(CONFIG_IPV6)
6165 else if (family
== AF_INET6
)
6166 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6167 &ireq
->ir_v6_rmt_addr
, port
);
6171 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6173 * If we receive a SYN packet with these bits set, it means a
6174 * network is playing bad games with TOS bits. In order to
6175 * avoid possible false congestion notifications, we disable
6176 * TCP ECN negotiation.
6178 * Exception: tcp_ca wants ECN. This is required for DCTCP
6179 * congestion control: Linux DCTCP asserts ECT on all packets,
6180 * including SYN, which is most optimal solution; however,
6181 * others, such as FreeBSD do not.
6183 static void tcp_ecn_create_request(struct request_sock
*req
,
6184 const struct sk_buff
*skb
,
6185 const struct sock
*listen_sk
,
6186 const struct dst_entry
*dst
)
6188 const struct tcphdr
*th
= tcp_hdr(skb
);
6189 const struct net
*net
= sock_net(listen_sk
);
6190 bool th_ecn
= th
->ece
&& th
->cwr
;
6197 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6198 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6199 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6201 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6202 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6203 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6204 inet_rsk(req
)->ecn_ok
= 1;
6207 static void tcp_openreq_init(struct request_sock
*req
,
6208 const struct tcp_options_received
*rx_opt
,
6209 struct sk_buff
*skb
, const struct sock
*sk
)
6211 struct inet_request_sock
*ireq
= inet_rsk(req
);
6213 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6215 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6216 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6217 tcp_rsk(req
)->snt_synack
= tcp_clock_us();
6218 tcp_rsk(req
)->last_oow_ack_time
= 0;
6219 req
->mss
= rx_opt
->mss_clamp
;
6220 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6221 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6222 ireq
->sack_ok
= rx_opt
->sack_ok
;
6223 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6224 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6227 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6228 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6229 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6232 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6233 struct sock
*sk_listener
,
6234 bool attach_listener
)
6236 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6240 struct inet_request_sock
*ireq
= inet_rsk(req
);
6242 ireq
->ireq_opt
= NULL
;
6243 #if IS_ENABLED(CONFIG_IPV6)
6244 ireq
->pktopts
= NULL
;
6246 atomic64_set(&ireq
->ir_cookie
, 0);
6247 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6248 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6249 ireq
->ireq_family
= sk_listener
->sk_family
;
6254 EXPORT_SYMBOL(inet_reqsk_alloc
);
6257 * Return true if a syncookie should be sent
6259 static bool tcp_syn_flood_action(const struct sock
*sk
,
6260 const struct sk_buff
*skb
,
6263 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6264 const char *msg
= "Dropping request";
6265 bool want_cookie
= false;
6266 struct net
*net
= sock_net(sk
);
6268 #ifdef CONFIG_SYN_COOKIES
6269 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6270 msg
= "Sending cookies";
6272 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6275 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6277 if (!queue
->synflood_warned
&&
6278 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6279 xchg(&queue
->synflood_warned
, 1) == 0)
6280 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6281 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6286 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6287 struct request_sock
*req
,
6288 const struct sk_buff
*skb
)
6290 if (tcp_sk(sk
)->save_syn
) {
6291 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6294 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6297 memcpy(©
[1], skb_network_header(skb
), len
);
6298 req
->saved_syn
= copy
;
6303 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6304 const struct tcp_request_sock_ops
*af_ops
,
6305 struct sock
*sk
, struct sk_buff
*skb
)
6307 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6308 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6309 struct tcp_options_received tmp_opt
;
6310 struct tcp_sock
*tp
= tcp_sk(sk
);
6311 struct net
*net
= sock_net(sk
);
6312 struct sock
*fastopen_sk
= NULL
;
6313 struct request_sock
*req
;
6314 bool want_cookie
= false;
6315 struct dst_entry
*dst
;
6318 /* TW buckets are converted to open requests without
6319 * limitations, they conserve resources and peer is
6320 * evidently real one.
6322 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6323 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6324 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6329 if (sk_acceptq_is_full(sk
)) {
6330 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6334 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6338 tcp_rsk(req
)->af_specific
= af_ops
;
6339 tcp_rsk(req
)->ts_off
= 0;
6341 tcp_clear_options(&tmp_opt
);
6342 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6343 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6344 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6345 want_cookie
? NULL
: &foc
);
6347 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6348 tcp_clear_options(&tmp_opt
);
6350 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6351 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6352 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6354 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6355 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6357 af_ops
->init_req(req
, sk
, skb
);
6359 if (security_inet_conn_request(sk
, skb
, req
))
6362 if (tmp_opt
.tstamp_ok
)
6363 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6365 dst
= af_ops
->route_req(sk
, &fl
, req
);
6369 if (!want_cookie
&& !isn
) {
6370 /* Kill the following clause, if you dislike this way. */
6371 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6372 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6373 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6374 !tcp_peer_is_proven(req
, dst
)) {
6375 /* Without syncookies last quarter of
6376 * backlog is filled with destinations,
6377 * proven to be alive.
6378 * It means that we continue to communicate
6379 * to destinations, already remembered
6380 * to the moment of synflood.
6382 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6384 goto drop_and_release
;
6387 isn
= af_ops
->init_seq(skb
);
6390 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6393 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6394 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6395 if (!tmp_opt
.tstamp_ok
)
6396 inet_rsk(req
)->ecn_ok
= 0;
6399 tcp_rsk(req
)->snt_isn
= isn
;
6400 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6401 tcp_openreq_init_rwin(req
, sk
, dst
);
6403 tcp_reqsk_record_syn(sk
, req
, skb
);
6404 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
);
6407 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6408 &foc
, TCP_SYNACK_FASTOPEN
);
6409 /* Add the child socket directly into the accept queue */
6410 if (!inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
)) {
6411 reqsk_fastopen_remove(fastopen_sk
, req
, false);
6412 bh_unlock_sock(fastopen_sk
);
6413 sock_put(fastopen_sk
);
6417 sk
->sk_data_ready(sk
);
6418 bh_unlock_sock(fastopen_sk
);
6419 sock_put(fastopen_sk
);
6421 tcp_rsk(req
)->tfo_listener
= false;
6423 inet_csk_reqsk_queue_hash_add(sk
, req
,
6424 tcp_timeout_init((struct sock
*)req
));
6425 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6426 !want_cookie
? TCP_SYNACK_NORMAL
:
6444 EXPORT_SYMBOL(tcp_conn_request
);