blob: 734cfc8ff76edf3453921b50620be2986bfcfdb9 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
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.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07009 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 * 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>
20 */
21
22/*
23 * Changes:
24 * Pedro Roque : Fast Retransmit/Recovery.
25 * Two receive queues.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
29 * Header prediction.
30 * Variable renaming.
31 *
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.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080044 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 * timestamps.
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
48 * data segments.
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.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090052 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070053 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090054 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070055 * Andi Kleen: Process packets with PSH set in the
56 * fast path.
57 * J Hadi Salim: ECN support
58 * Andrei Gurtov,
59 * Pasi Sarolahti,
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
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 */
64
Joe Perchesafd465032012-03-12 07:03:32 +000065#define pr_fmt(fmt) "TCP: " fmt
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090068#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#include <linux/module.h>
70#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070071#include <linux/kernel.h>
Eric Dumazetad971f62014-10-11 15:17:29 -070072#include <linux/prefetch.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070073#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#include <net/tcp.h>
75#include <net/inet_common.h>
76#include <linux/ipsec.h>
77#include <asm/unaligned.h>
Willem de Bruijne1c8a602014-08-04 22:11:50 -040078#include <linux/errqueue.h>
Song Liu59415212017-10-23 09:20:25 -070079#include <trace/events/tcp.h>
Ursula Braun60e2a772017-10-25 11:01:45 +020080#include <linux/static_key.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
Brian Haleyab32ea52006-09-22 14:15:41 -070082int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Linus Torvalds1da177e2005-04-16 15:20:36 -070083
Linus Torvalds1da177e2005-04-16 15:20:36 -070084#define FLAG_DATA 0x01 /* Incoming frame contained data. */
85#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89#define FLAG_DATA_SACKED 0x20 /* New SACK. */
90#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -070091#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Florian Westphal31770e32017-08-30 19:24:58 +020092#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +000093#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -070094#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -070095#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Neal Cardwelldf92c832017-08-03 09:19:54 -040096#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
Ilpo Järvinencadbd032007-12-31 04:49:21 -080097#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +000098#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -070099#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
101#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
102#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
Priyaranjan Jhad09b9e62017-11-03 17:46:55 -0700103#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700107#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108
Yuchung Chenge662ca42016-02-02 10:33:04 -0800109#define REXMIT_NONE 0 /* no loss recovery to do */
110#define REXMIT_LOST 1 /* retransmit packets marked lost */
111#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
112
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300113static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
114 unsigned int len)
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200115{
116 static bool __once __read_mostly;
117
118 if (!__once) {
119 struct net_device *dev;
120
121 __once = true;
122
123 rcu_read_lock();
124 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300125 if (!dev || len >= dev->mtu)
126 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
127 dev ? dev->name : "Unknown driver");
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200128 rcu_read_unlock();
129 }
130}
131
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900132/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900134 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800135static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700137 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900138 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700139 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900141 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
143 /* skb->len may jitter because of SACKs, even if peer
144 * sends good full-sized frames.
145 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800146 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700147 if (len >= icsk->icsk_ack.rcv_mss) {
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200148 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
149 tcp_sk(sk)->advmss);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300150 /* Account for possibly-removed options */
151 if (unlikely(len > icsk->icsk_ack.rcv_mss +
152 MAX_TCP_OPTION_SPACE))
153 tcp_gro_dev_warn(sk, skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 } else {
155 /* Otherwise, we make more careful check taking into account,
156 * that SACKs block is variable.
157 *
158 * "len" is invariant segment length, including TCP header.
159 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700160 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000161 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 /* If PSH is not set, packet should be
163 * full sized, provided peer TCP is not badly broken.
164 * This observation (if it is correct 8)) allows
165 * to handle super-low mtu links fairly.
166 */
167 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700168 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 /* Subtract also invariant (if peer is RFC compliant),
170 * tcp header plus fixed timestamp option length.
171 * Resulting "len" is MSS free of SACK jitter.
172 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700173 len -= tcp_sk(sk)->tcp_header_len;
174 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700176 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 return;
178 }
179 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700180 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
181 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700182 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 }
184}
185
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700186static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700188 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000189 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800191 if (quickacks == 0)
192 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700193 if (quickacks > icsk->icsk_ack.quick)
194 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
stephen hemminger1b9f4092010-09-28 19:30:14 +0000197static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700199 struct inet_connection_sock *icsk = inet_csk(sk);
200 tcp_incr_quickack(sk);
201 icsk->icsk_ack.pingpong = 0;
202 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203}
204
205/* Send ACKs quickly, if "quick" count is not exhausted
206 * and the session is not interactive.
207 */
208
Jon Maxwell2251ae42015-07-08 10:12:28 +1000209static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700211 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000212 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000213
Jon Maxwell2251ae42015-07-08 10:12:28 +1000214 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
215 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216}
217
Florian Westphal735d3832014-09-29 13:08:30 +0200218static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700219{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800220 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700221 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
222}
223
Florian Westphal735d3832014-09-29 13:08:30 +0200224static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700225{
226 if (tcp_hdr(skb)->cwr)
227 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
228}
229
Florian Westphal735d3832014-09-29 13:08:30 +0200230static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700231{
232 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
233}
234
Florian Westphal735d3832014-09-29 13:08:30 +0200235static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700236{
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400237 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000238 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700239 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000240 * and we already seen ECT on a previous segment,
241 * it is probably a retransmit.
242 */
243 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700244 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000245 break;
246 case INET_ECN_CE:
Florian Westphal98900922014-09-26 22:37:35 +0200247 if (tcp_ca_needs_ecn((struct sock *)tp))
248 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
249
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000250 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
251 /* Better not delay acks, sender can have a very low cwnd */
252 tcp_enter_quickack_mode((struct sock *)tp);
253 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
254 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000255 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200256 break;
257 default:
258 if (tcp_ca_needs_ecn((struct sock *)tp))
259 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
260 tp->ecn_flags |= TCP_ECN_SEEN;
261 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700262 }
263}
264
Florian Westphal735d3832014-09-29 13:08:30 +0200265static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
266{
267 if (tp->ecn_flags & TCP_ECN_OK)
268 __tcp_ecn_check_ce(tp, skb);
269}
270
271static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700272{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800273 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700274 tp->ecn_flags &= ~TCP_ECN_OK;
275}
276
Florian Westphal735d3832014-09-29 13:08:30 +0200277static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700278{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800279 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700280 tp->ecn_flags &= ~TCP_ECN_OK;
281}
282
Florian Westphal735d3832014-09-29 13:08:30 +0200283static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700284{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800285 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000286 return true;
287 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700288}
289
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290/* Buffer size and advertised window tuning.
291 *
292 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
293 */
294
Eric Dumazet6ae70532013-10-01 10:23:44 -0700295static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700297 const struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400298 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700299 int sndmem, per_mss;
300 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
Eric Dumazet6ae70532013-10-01 10:23:44 -0700302 /* Worst case is non GSO/TSO : each frame consumes one skb
303 * and skb->head is kmalloced using power of two area of memory
304 */
305 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
306 MAX_TCP_HEADER +
307 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
308
309 per_mss = roundup_pow_of_two(per_mss) +
310 SKB_DATA_ALIGN(sizeof(struct sk_buff));
311
312 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
313 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
314
315 /* Fast Recovery (RFC 5681 3.2) :
316 * Cubic needs 1.7 factor, rounded to 2 to include
317 * extra cushion (application might react slowly to POLLOUT)
318 */
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400319 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
320 sndmem *= nr_segs * per_mss;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700321
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000322 if (sk->sk_sndbuf < sndmem)
Eric Dumazet356d1832017-11-07 00:29:28 -0800323 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324}
325
326/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
327 *
328 * All tcp_full_space() is split to two parts: "network" buffer, allocated
329 * forward and advertised in receiver window (tp->rcv_wnd) and
330 * "application buffer", required to isolate scheduling/application
331 * latencies from network.
332 * window_clamp is maximal advertised window. It can be less than
333 * tcp_full_space(), in this case tcp_full_space() - window_clamp
334 * is reserved for "application" buffer. The less window_clamp is
335 * the smoother our behaviour from viewpoint of network, but the lower
336 * throughput and the higher sensitivity of the connection to losses. 8)
337 *
338 * rcv_ssthresh is more strict window_clamp used at "slow start"
339 * phase to predict further behaviour of this connection.
340 * It is used for two goals:
341 * - to enforce header prediction at sender, even when application
342 * requires some significant "application buffer". It is check #1.
343 * - to prevent pruning of receive queue because of misprediction
344 * of receiver window. Check #2.
345 *
346 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800347 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 * in common situations. Otherwise, we have to rely on queue collapsing.
349 */
350
351/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700352static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700354 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 /* Optimize this! */
Eric Dumazet94f08932017-10-26 21:55:09 -0700356 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
Eric Dumazet356d1832017-11-07 00:29:28 -0800357 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
359 while (tp->rcv_ssthresh <= window) {
360 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700361 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
363 truesize >>= 1;
364 window >>= 1;
365 }
366 return 0;
367}
368
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400369static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700371 struct tcp_sock *tp = tcp_sk(sk);
372
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 /* Check #1 */
374 if (tp->rcv_ssthresh < tp->window_clamp &&
375 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700376 !tcp_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 int incr;
378
379 /* Check #2. Increase window, if skb with such overhead
380 * will fit to rcvbuf in future.
381 */
Eric Dumazet94f08932017-10-26 21:55:09 -0700382 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800383 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700385 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
387 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000388 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800389 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
390 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700391 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 }
393 }
394}
395
396/* 3. Tuning rcvbuf, when connection enters established state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397static void tcp_fixup_rcvbuf(struct sock *sk)
398{
Eric Dumazete9266a02011-10-20 16:53:56 -0400399 u32 mss = tcp_sk(sk)->advmss;
Eric Dumazete9266a02011-10-20 16:53:56 -0400400 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Yuchung Cheng85f16522013-06-11 15:35:32 -0700402 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
403 tcp_default_init_rwnd(mss);
Eric Dumazete9266a02011-10-20 16:53:56 -0400404
Eric Dumazetb0983d32013-09-20 13:56:58 -0700405 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
406 * Allow enough cushion so that sender is not limited by our window
407 */
Eric Dumazet4540c0c2017-10-27 07:47:22 -0700408 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
Eric Dumazetb0983d32013-09-20 13:56:58 -0700409 rcvmem <<= 2;
410
Eric Dumazete9266a02011-10-20 16:53:56 -0400411 if (sk->sk_rcvbuf < rcvmem)
Eric Dumazet356d1832017-11-07 00:29:28 -0800412 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413}
414
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800415/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 * established state.
417 */
Jerry Chu10467162012-08-31 12:29:11 +0000418void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419{
Eric Dumazet0c126542017-10-26 21:55:08 -0700420 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 struct tcp_sock *tp = tcp_sk(sk);
422 int maxwin;
423
424 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
425 tcp_fixup_rcvbuf(sk);
426 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700427 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
429 tp->rcvq_space.space = tp->rcv_wnd;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700430 tcp_mstamp_refresh(tp);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700431 tp->rcvq_space.time = tp->tcp_mstamp;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700432 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
434 maxwin = tcp_full_space(sk);
435
436 if (tp->window_clamp >= maxwin) {
437 tp->window_clamp = maxwin;
438
Eric Dumazet0c126542017-10-26 21:55:08 -0700439 if (tcp_app_win && maxwin > 4 * tp->advmss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 tp->window_clamp = max(maxwin -
Eric Dumazet0c126542017-10-26 21:55:08 -0700441 (maxwin >> tcp_app_win),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 4 * tp->advmss);
443 }
444
445 /* Force reservation of one segment. */
Eric Dumazet0c126542017-10-26 21:55:08 -0700446 if (tcp_app_win &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 tp->window_clamp > 2 * tp->advmss &&
448 tp->window_clamp + tp->advmss > maxwin)
449 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
450
451 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
Eric Dumazetc2203cf2017-05-16 14:00:04 -0700452 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453}
454
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700456static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700458 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300459 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet356d1832017-11-07 00:29:28 -0800460 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300462 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Eric Dumazet356d1832017-11-07 00:29:28 -0800464 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
John Heffner326f36e2005-11-10 17:11:48 -0800465 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700466 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000467 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800468 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
Eric Dumazet356d1832017-11-07 00:29:28 -0800469 net->ipv4.sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 }
John Heffner326f36e2005-11-10 17:11:48 -0800471 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800472 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800475/* Initialize RCV_MSS value.
476 * RCV_MSS is an our guess about MSS used by the peer.
477 * We haven't any direct information about the MSS.
478 * It's better to underestimate the RCV_MSS rather than overestimate.
479 * Overestimations make us ACKing less frequently than needed.
480 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
481 */
482void tcp_initialize_rcv_mss(struct sock *sk)
483{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400484 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800485 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
486
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800487 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000488 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800489 hint = max(hint, TCP_MIN_MSS);
490
491 inet_csk(sk)->icsk_ack.rcv_mss = hint;
492}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000493EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800494
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495/* Receiver "autotuning" code.
496 *
497 * The algorithm for RTT estimation w/o timestamps is based on
498 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200499 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 *
501 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200502 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 * though this reference is out of date. A new paper
504 * is pending.
505 */
506static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
507{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700508 u32 new_sample = tp->rcv_rtt_est.rtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 long m = sample;
510
511 if (m == 0)
512 m = 1;
513
514 if (new_sample != 0) {
515 /* If we sample in larger samples in the non-timestamp
516 * case, we could grossly overestimate the RTT especially
517 * with chatty applications or bulk transfer apps which
518 * are stalled on filesystem I/O.
519 *
520 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800521 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800522 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 * long.
524 */
525 if (!win_dep) {
526 m -= (new_sample >> 3);
527 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000528 } else {
529 m <<= 3;
530 if (m < new_sample)
531 new_sample = m;
532 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800534 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 new_sample = m << 3;
536 }
537
Eric Dumazet645f4c62017-04-25 10:15:41 -0700538 tp->rcv_rtt_est.rtt_us = new_sample;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539}
540
541static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
542{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700543 u32 delta_us;
544
Eric Dumazet9a568de2017-05-16 14:00:14 -0700545 if (tp->rcv_rtt_est.time == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 goto new_measure;
547 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
548 return;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700549 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700550 tcp_rcv_rtt_update(tp, delta_us, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551
552new_measure:
553 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700554 tp->rcv_rtt_est.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555}
556
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800557static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
558 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700560 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700561
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 if (tp->rx_opt.rcv_tsecr &&
563 (TCP_SKB_CB(skb)->end_seq -
Eric Dumazet9a568de2017-05-16 14:00:14 -0700564 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
565 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
566 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
567
568 tcp_rcv_rtt_update(tp, delta_us, 0);
569 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
572/*
573 * This function should be called every time data is copied to user space.
574 * It calculates the appropriate TCP receive buffer space.
575 */
576void tcp_rcv_space_adjust(struct sock *sk)
577{
578 struct tcp_sock *tp = tcp_sk(sk);
579 int time;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700580 int copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900581
Eric Dumazet9a568de2017-05-16 14:00:14 -0700582 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700583 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900585
Eric Dumazetb0983d32013-09-20 13:56:58 -0700586 /* Number of bytes copied to user in last RTT */
587 copied = tp->copied_seq - tp->rcvq_space.seq;
588 if (copied <= tp->rcvq_space.space)
589 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590
Eric Dumazetb0983d32013-09-20 13:56:58 -0700591 /* A bit of theory :
592 * copied = bytes received in previous RTT, our base window
593 * To cope with packet losses, we need a 2x factor
594 * To cope with slow start, and sender growing its cwin by 100 %
595 * every RTT, we need a 4x factor, because the ACK we are sending
596 * now is for the next RTT, not the current one :
597 * <prev RTT . ><current RTT .. ><next RTT .... >
598 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
Eric Dumazet4540c0c2017-10-27 07:47:22 -0700600 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
Eric Dumazetb0983d32013-09-20 13:56:58 -0700601 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
602 int rcvwin, rcvmem, rcvbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
Eric Dumazetb0983d32013-09-20 13:56:58 -0700604 /* minimal window to cope with packet losses, assuming
605 * steady state. Add some cushion because of small variations.
606 */
607 rcvwin = (copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Eric Dumazetb0983d32013-09-20 13:56:58 -0700609 /* If rate increased by 25%,
610 * assume slow start, rcvwin = 3 * copied
611 * If rate increased by 50%,
612 * assume sender can use 2x growth, rcvwin = 4 * copied
613 */
614 if (copied >=
615 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
616 if (copied >=
617 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
618 rcvwin <<= 1;
619 else
620 rcvwin += (rcvwin >> 1);
621 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Eric Dumazetb0983d32013-09-20 13:56:58 -0700623 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
Eric Dumazet94f08932017-10-26 21:55:09 -0700624 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
Eric Dumazetb0983d32013-09-20 13:56:58 -0700625 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626
Eric Dumazet356d1832017-11-07 00:29:28 -0800627 rcvbuf = min(rcvwin / tp->advmss * rcvmem,
628 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
Eric Dumazetb0983d32013-09-20 13:56:58 -0700629 if (rcvbuf > sk->sk_rcvbuf) {
630 sk->sk_rcvbuf = rcvbuf;
631
632 /* Make the window clamp follow along. */
633 tp->window_clamp = rcvwin;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 }
635 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700636 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900637
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638new_measure:
639 tp->rcvq_space.seq = tp->copied_seq;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700640 tp->rcvq_space.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641}
642
643/* There is something which you must keep in mind when you analyze the
644 * behavior of the tp->ato delayed ack timeout interval. When a
645 * connection starts up, we want to ack as quickly as possible. The
646 * problem is that "good" TCP's do slow start at the beginning of data
647 * transmission. The means that until we send the first few ACK's the
648 * sender will sit on his end and only queue most of his data, because
649 * he can only send snd_cwnd unacked packets at any given time. For
650 * each ACK we send, he increments snd_cwnd and transmits more of his
651 * queue. -DaveM
652 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700653static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700655 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700656 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 u32 now;
658
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700659 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700661 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900664
Eric Dumazet70eabf02017-05-16 14:00:07 -0700665 now = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700667 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 /* The _first_ data packet received, initialize
669 * delayed ACK engine.
670 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700671 tcp_incr_quickack(sk);
672 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700674 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800676 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700678 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
679 } else if (m < icsk->icsk_ack.ato) {
680 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
681 if (icsk->icsk_ack.ato > icsk->icsk_rto)
682 icsk->icsk_ack.ato = icsk->icsk_rto;
683 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800684 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 * restart window, so that we send ACKs quickly.
686 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700687 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800688 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 }
690 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700691 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Florian Westphal735d3832014-09-29 13:08:30 +0200693 tcp_ecn_check_ce(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
695 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700696 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697}
698
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699/* Called to compute a smoothed rtt estimate. The data fed to this
700 * routine either comes from timestamps, or from segments that were
701 * known _not_ to have been retransmitted [see Karn/Partridge
702 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
703 * piece by Van Jacobson.
704 * NOTE: the next three routines used to be one big routine.
705 * To save cycles in the RFC 1323 implementation it was better to break
706 * it up into three procedures. -- erics
707 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800708static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300710 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800711 long m = mrtt_us; /* RTT */
712 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 /* The following amusing code comes from Jacobson's
715 * article in SIGCOMM '88. Note that rtt and mdev
716 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900717 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 * m stands for "measurement".
719 *
720 * On a 1990 paper the rto value is changed to:
721 * RTO = rtt + 4 * mdev
722 *
723 * Funny. This algorithm seems to be very broken.
724 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800725 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
727 * does not matter how to _calculate_ it. Seems, it was trap
728 * that VJ failed to avoid. 8)
729 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800730 if (srtt != 0) {
731 m -= (srtt >> 3); /* m is now error in rtt est */
732 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 if (m < 0) {
734 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800735 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 /* This is similar to one of Eifel findings.
737 * Eifel blocks mdev updates when rtt decreases.
738 * This solution is a bit different: we use finer gain
739 * for mdev in this case (alpha*beta).
740 * Like Eifel it also prevents growth of rto,
741 * but also it limits too fast rto decreases,
742 * happening in pure Eifel.
743 */
744 if (m > 0)
745 m >>= 3;
746 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800747 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800749 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
750 if (tp->mdev_us > tp->mdev_max_us) {
751 tp->mdev_max_us = tp->mdev_us;
752 if (tp->mdev_max_us > tp->rttvar_us)
753 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 }
755 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800756 if (tp->mdev_max_us < tp->rttvar_us)
757 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800759 tp->mdev_max_us = tcp_rto_min_us(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 }
761 } else {
762 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800763 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800764 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
765 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
766 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 tp->rtt_seq = tp->snd_nxt;
768 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800769 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770}
771
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700772static void tcp_update_pacing_rate(struct sock *sk)
773{
774 const struct tcp_sock *tp = tcp_sk(sk);
775 u64 rate;
776
777 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700778 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
779
780 /* current rate is (cwnd * mss) / srtt
781 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
782 * In Congestion Avoidance phase, set it to 120 % the current rate.
783 *
784 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
785 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
786 * end of slow start and should slow down.
787 */
788 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
Eric Dumazet23a7102a2017-10-27 07:47:31 -0700789 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
Eric Dumazet43e122b2015-08-21 17:38:02 -0700790 else
Eric Dumazetc26e91f2017-10-27 07:47:32 -0700791 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700792
793 rate *= max(tp->snd_cwnd, tp->packets_out);
794
Eric Dumazet740b0f12014-02-26 14:02:48 -0800795 if (likely(tp->srtt_us))
796 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700797
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700798 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
Eric Dumazetba537422013-10-09 17:14:52 -0700799 * without any lock. We want to make sure compiler wont store
800 * intermediate values in this location.
801 */
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700802 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
803 sk->sk_max_pacing_rate));
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700804}
805
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806/* Calculate rto without backoff. This is the second half of Van Jacobson's
807 * routine referred to above.
808 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800809static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700811 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 /* Old crap is replaced with new one. 8)
813 *
814 * More seriously:
815 * 1. If rtt variance happened to be less 50msec, it is hallucination.
816 * It cannot be less due to utterly erratic ACK generation made
817 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
818 * to do with delayed acks, because at cwnd>2 true delack timeout
819 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800820 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000822 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823
824 /* 2. Fixups made earlier cannot be right.
825 * If we do not estimate RTO correctly without them,
826 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800827 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800830 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
831 * guarantees that rto is higher.
832 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000833 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834}
835
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400836__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837{
838 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
839
Gerrit Renker22b71c82010-08-29 19:23:12 +0000840 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800841 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
843}
844
Ryousei Takano564262c2007-10-25 23:03:52 -0700845/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300846static void tcp_dsack_seen(struct tcp_sock *tp)
847{
Vijay Subramanianab562222011-12-20 13:23:24 +0000848 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Priyaranjan Jha1f255692017-11-03 16:38:48 -0700849 tp->rack.dsack_seen = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300850}
851
Yuchung Cheng737ff312017-11-08 13:01:27 -0800852/* It's reordering when higher sequence was delivered (i.e. sacked) before
853 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
854 * distance is approximated in full-mss packet distance ("reordering").
855 */
856static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
857 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300859 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -0800860 const u32 mss = tp->mss_cache;
861 u32 fack, metric;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700862
Yuchung Cheng737ff312017-11-08 13:01:27 -0800863 fack = tcp_highest_sack_seq(tp);
864 if (!before(low_seq, fack))
Soheil Hassas Yeganeh6f5b24e2017-05-16 17:39:02 -0400865 return;
866
Yuchung Cheng737ff312017-11-08 13:01:27 -0800867 metric = fack - low_seq;
868 if ((metric > tp->reordering * mss) && mss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000870 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
871 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
872 tp->reordering,
Yuchung Cheng737ff312017-11-08 13:01:27 -0800873 0,
Joe Perches91df42b2012-05-15 14:11:54 +0000874 tp->sacked_out,
875 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876#endif
Yuchung Cheng737ff312017-11-08 13:01:27 -0800877 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
878 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000880
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700881 tp->rack.reord = 1;
Yuchung Cheng2d2517e2017-04-04 14:15:40 -0700882 /* This exciting event is worth to be remembered. 8) */
Yuchung Cheng737ff312017-11-08 13:01:27 -0800883 NET_INC_STATS(sock_net(sk),
884 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885}
886
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700887/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700888static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
889{
Ian Morris51456b22015-04-03 09:17:26 +0100890 if (!tp->retransmit_skb_hint ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700891 before(TCP_SKB_CB(skb)->seq,
892 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700893 tp->retransmit_skb_hint = skb;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700894}
895
Neal Cardwell0682e692016-09-19 23:39:13 -0400896/* Sum the number of packets on the wire we have marked as lost.
897 * There are two cases we care about here:
898 * a) Packet hasn't been marked lost (nor retransmitted),
899 * and this is the first loss.
900 * b) Packet has been marked both lost and retransmitted,
901 * and this means we think it was lost again.
902 */
903static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
904{
905 __u8 sacked = TCP_SKB_CB(skb)->sacked;
906
907 if (!(sacked & TCPCB_LOST) ||
908 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
909 tp->lost += tcp_skb_pcount(skb);
910}
911
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700912static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
913{
914 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
915 tcp_verify_retransmit_hint(tp, skb);
916
917 tp->lost_out += tcp_skb_pcount(skb);
Neal Cardwell0682e692016-09-19 23:39:13 -0400918 tcp_sum_lost(tp, skb);
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700919 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
920 }
921}
922
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700923void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700924{
925 tcp_verify_retransmit_hint(tp, skb);
926
Neal Cardwell0682e692016-09-19 23:39:13 -0400927 tcp_sum_lost(tp, skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700928 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
929 tp->lost_out += tcp_skb_pcount(skb);
930 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
931 }
932}
933
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934/* This procedure tags the retransmission queue when SACKs arrive.
935 *
936 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
937 * Packets in queue with these bits set are counted in variables
938 * sacked_out, retrans_out and lost_out, correspondingly.
939 *
940 * Valid combinations are:
941 * Tag InFlight Description
942 * 0 1 - orig segment is in flight.
943 * S 0 - nothing flies, orig reached receiver.
944 * L 0 - nothing flies, orig lost by net.
945 * R 2 - both orig and retransmit are in flight.
946 * L|R 1 - orig is lost, retransmit is in flight.
947 * S|R 1 - orig reached receiver, retrans is still in flight.
948 * (L|S|R is logically valid, it could occur when L|R is sacked,
949 * but it is equivalent to plain S and code short-curcuits it to S.
950 * L|S is logically invalid, it would mean -1 packet in flight 8))
951 *
952 * These 6 states form finite state machine, controlled by the following events:
953 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
954 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000955 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 * A. Scoreboard estimator decided the packet is lost.
957 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000958 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 * segment was retransmitted.
960 * 4. D-SACK added new rule: D-SACK changes any tag to S.
961 *
962 * It is pleasant to note, that state diagram turns out to be commutative,
963 * so that we are allowed not to be bothered by order of our actions,
964 * when multiple events arrive simultaneously. (see the function below).
965 *
966 * Reordering detection.
967 * --------------------
968 * Reordering metric is maximal distance, which a packet can be displaced
969 * in packet stream. With SACKs we can estimate it:
970 *
971 * 1. SACK fills old hole and the corresponding segment was not
972 * ever retransmitted -> reordering. Alas, we cannot use it
973 * when segment was retransmitted.
974 * 2. The last flaw is solved with D-SACK. D-SACK arrives
975 * for retransmitted and already SACKed segment -> reordering..
976 * Both of these heuristics are not used in Loss state, when we cannot
977 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700978 *
979 * SACK block validation.
980 * ----------------------
981 *
982 * SACK block range validation checks that the received SACK block fits to
983 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
984 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -0700985 * it means that the receiver is rather inconsistent with itself reporting
986 * SACK reneging when it should advance SND.UNA. Such SACK block this is
987 * perfectly valid, however, in light of RFC2018 which explicitly states
988 * that "SACK block MUST reflect the newest segment. Even if the newest
989 * segment is going to be discarded ...", not that it looks very clever
990 * in case of head skb. Due to potentional receiver driven attacks, we
991 * choose to avoid immediate execution of a walk in write queue due to
992 * reneging and defer head skb's loss recovery to standard loss recovery
993 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700994 *
995 * Implements also blockage to start_seq wrap-around. Problem lies in the
996 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
997 * there's no guarantee that it will be before snd_nxt (n). The problem
998 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
999 * wrap (s_w):
1000 *
1001 * <- outs wnd -> <- wrapzone ->
1002 * u e n u_w e_w s n_w
1003 * | | | | | | |
1004 * |<------------+------+----- TCP seqno space --------------+---------->|
1005 * ...-- <2^31 ->| |<--------...
1006 * ...---- >2^31 ------>| |<--------...
1007 *
1008 * Current code wouldn't be vulnerable but it's better still to discard such
1009 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1010 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1011 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1012 * equal to the ideal case (infinite seqno space without wrap caused issues).
1013 *
1014 * With D-SACK the lower bound is extended to cover sequence space below
1015 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001016 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001017 * for the normal SACK blocks, explained above). But there all simplicity
1018 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1019 * fully below undo_marker they do not affect behavior in anyway and can
1020 * therefore be safely ignored. In rare cases (which are more or less
1021 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1022 * fragmentation and packet reordering past skb's retransmission. To consider
1023 * them correctly, the acceptable range must be extended even more though
1024 * the exact amount is rather hard to quantify. However, tp->max_window can
1025 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001027static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1028 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001029{
1030 /* Too far in future, or reversed (interpretation is ambiguous) */
1031 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001032 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001033
1034 /* Nasty start_seq wrap-around check (see comments above) */
1035 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001036 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001037
Ryousei Takano564262c2007-10-25 23:03:52 -07001038 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001039 * start_seq == snd_una is non-sensical (see comments above)
1040 */
1041 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001042 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001043
1044 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001045 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001046
1047 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001048 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001049 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001050
1051 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001052 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001053
1054 /* Too old */
1055 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001056 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001057
1058 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1059 * start_seq < undo_marker and end_seq >= undo_marker.
1060 */
1061 return !before(start_seq, end_seq - tp->max_window);
1062}
1063
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001064static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1065 struct tcp_sack_block_wire *sp, int num_sacks,
1066 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001067{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001068 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001069 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1070 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001071 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001072
1073 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001074 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001075 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001076 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001077 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001078 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1079 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001080
1081 if (!after(end_seq_0, end_seq_1) &&
1082 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001083 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001084 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001085 NET_INC_STATS(sock_net(sk),
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001086 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001087 }
1088 }
1089
1090 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001091 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001092 !after(end_seq_0, prior_snd_una) &&
1093 after(end_seq_0, tp->undo_marker))
1094 tp->undo_retrans--;
1095
1096 return dup_sack;
1097}
1098
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001099struct tcp_sacktag_state {
Yuchung Cheng737ff312017-11-08 13:01:27 -08001100 u32 reord;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001101 /* Timestamps for earliest and latest never-retransmitted segment
1102 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1103 * but congestion control should still get an accurate delay signal.
1104 */
Eric Dumazet9a568de2017-05-16 14:00:14 -07001105 u64 first_sackt;
1106 u64 last_sackt;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001107 struct rate_sample *rate;
Eric Dumazet740b0f12014-02-26 14:02:48 -08001108 int flag;
Eric Dumazet75c119a2017-10-05 22:21:27 -07001109 unsigned int mss_now;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001110};
1111
Ilpo Järvinend1935942007-10-11 17:34:25 -07001112/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1113 * the incoming SACK may not exactly match but we can find smaller MSS
1114 * aligned portion of it that matches. Therefore we might need to fragment
1115 * which may fail and creates some hassle (caller must handle error case
1116 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001117 *
1118 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001119 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001120static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001121 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001122{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001123 int err;
1124 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001125 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001126 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001127
1128 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1129 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1130
1131 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1132 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001133 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001134 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1135
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001136 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001137 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001138 if (pkt_len < mss)
1139 pkt_len = mss;
1140 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001141 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001142 if (pkt_len < mss)
1143 return -EINVAL;
1144 }
1145
1146 /* Round if necessary so that SACKs cover only full MSSes
1147 * and/or the remaining small portion (if present)
1148 */
1149 if (pkt_len > mss) {
1150 unsigned int new_len = (pkt_len / mss) * mss;
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001151 if (!in_sack && new_len < pkt_len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001152 new_len += mss;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001153 pkt_len = new_len;
1154 }
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001155
1156 if (pkt_len >= skb->len && !in_sack)
1157 return 0;
1158
Eric Dumazet75c119a2017-10-05 22:21:27 -07001159 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1160 pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001161 if (err < 0)
1162 return err;
1163 }
1164
1165 return in_sack;
1166}
1167
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001168/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1169static u8 tcp_sacktag_one(struct sock *sk,
1170 struct tcp_sacktag_state *state, u8 sacked,
1171 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001172 int dup_sack, int pcount,
Eric Dumazet9a568de2017-05-16 14:00:14 -07001173 u64 xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001174{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001175 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001176
1177 /* Account D-SACK for retransmitted packet. */
1178 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001179 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001180 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001181 tp->undo_retrans--;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001182 if ((sacked & TCPCB_SACKED_ACKED) &&
1183 before(start_seq, state->reord))
1184 state->reord = start_seq;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001185 }
1186
1187 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001188 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001189 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001190
1191 if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazetd2329f12017-04-25 10:15:38 -07001192 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001193
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001194 if (sacked & TCPCB_SACKED_RETRANS) {
1195 /* If the segment is not tagged as lost,
1196 * we do not clear RETRANS, believing
1197 * that retransmission is still in flight.
1198 */
1199 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001200 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001201 tp->lost_out -= pcount;
1202 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001203 }
1204 } else {
1205 if (!(sacked & TCPCB_RETRANS)) {
1206 /* New sack for not retransmitted frame,
1207 * which was in hole. It is reordering.
1208 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001209 if (before(start_seq,
Yuchung Cheng737ff312017-11-08 13:01:27 -08001210 tcp_highest_sack_seq(tp)) &&
1211 before(start_seq, state->reord))
1212 state->reord = start_seq;
1213
Yuchung Chenge33099f2013-03-20 13:33:00 +00001214 if (!after(end_seq, tp->high_seq))
1215 state->flag |= FLAG_ORIG_SACK_ACKED;
Eric Dumazet9a568de2017-05-16 14:00:14 -07001216 if (state->first_sackt == 0)
1217 state->first_sackt = xmit_time;
1218 state->last_sackt = xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001219 }
1220
1221 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001222 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001223 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001224 }
1225 }
1226
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001227 sacked |= TCPCB_SACKED_ACKED;
1228 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001229 tp->sacked_out += pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08001230 tp->delivered += pcount; /* Out-of-order packets delivered */
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001231
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001232 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Yuchung Cheng713bafe2017-11-08 13:01:26 -08001233 if (tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001234 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001235 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001236 }
1237
1238 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1239 * frames and clear it. undo_retrans is decreased above, L|R frames
1240 * are accounted above as well.
1241 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001242 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1243 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001244 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001245 }
1246
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001247 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001248}
1249
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001250/* Shift newly-SACKed bytes from this skb to the immediately previous
1251 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1252 */
Eric Dumazetf3319812017-10-05 22:21:26 -07001253static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1254 struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001255 struct tcp_sacktag_state *state,
1256 unsigned int pcount, int shifted, int mss,
1257 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001258{
1259 struct tcp_sock *tp = tcp_sk(sk);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001260 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1261 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001262
1263 BUG_ON(!pcount);
1264
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001265 /* Adjust counters and hints for the newly sacked sequence
1266 * range but discard the return value since prev is already
1267 * marked. We must tag the range first because the seq
1268 * advancement below implicitly advances
1269 * tcp_highest_sack_seq() when skb is highest_sack.
1270 */
1271 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001272 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet9a568de2017-05-16 14:00:14 -07001273 skb->skb_mstamp);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001274 tcp_rate_skb_delivered(sk, skb, state->rate);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001275
1276 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001277 tp->lost_cnt_hint += pcount;
1278
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001279 TCP_SKB_CB(prev)->end_seq += shifted;
1280 TCP_SKB_CB(skb)->seq += shifted;
1281
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001282 tcp_skb_pcount_add(prev, pcount);
1283 BUG_ON(tcp_skb_pcount(skb) < pcount);
1284 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001285
1286 /* When we're adding to gso_segs == 1, gso_size will be zero,
1287 * in theory this shouldn't be necessary but as long as DSACK
1288 * code can come after this skb later on it's better to keep
1289 * setting gso_size to something.
1290 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001291 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1292 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001293
1294 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001295 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001296 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001297
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001298 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1299 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1300
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001301 if (skb->len > 0) {
1302 BUG_ON(!tcp_skb_pcount(skb));
Eric Dumazetc10d9312016-04-29 14:16:47 -07001303 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001304 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001305 }
1306
1307 /* Whole SKB was eaten :-) */
1308
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001309 if (skb == tp->retransmit_skb_hint)
1310 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001311 if (skb == tp->lost_skb_hint) {
1312 tp->lost_skb_hint = prev;
1313 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1314 }
1315
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001316 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001317 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001318 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1319 TCP_SKB_CB(prev)->end_seq++;
1320
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001321 if (skb == tcp_highest_sack(sk))
1322 tcp_advance_highest_sack(sk, skb);
1323
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07001324 tcp_skb_collapse_tstamp(prev, skb);
Eric Dumazet9a568de2017-05-16 14:00:14 -07001325 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1326 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001327
Eric Dumazet75c119a2017-10-05 22:21:27 -07001328 tcp_rtx_queue_unlink_and_free(skb, sk);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001329
Eric Dumazetc10d9312016-04-29 14:16:47 -07001330 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001331
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001332 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001333}
1334
1335/* I wish gso_size would have a bit more sane initialization than
1336 * something-or-zero which complicates things
1337 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001338static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001339{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001340 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001341}
1342
1343/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001344static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001345{
1346 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1347}
1348
1349/* Try collapsing SACK blocks spanning across multiple skbs to a single
1350 * skb.
1351 */
1352static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001353 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001354 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001355 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001356{
1357 struct tcp_sock *tp = tcp_sk(sk);
1358 struct sk_buff *prev;
1359 int mss;
1360 int pcount = 0;
1361 int len;
1362 int in_sack;
1363
1364 if (!sk_can_gso(sk))
1365 goto fallback;
1366
1367 /* Normally R but no L won't result in plain S */
1368 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001369 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001370 goto fallback;
1371 if (!skb_can_shift(skb))
1372 goto fallback;
1373 /* This frame is about to be dropped (was ACKed). */
1374 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1375 goto fallback;
1376
1377 /* Can only happen with delayed DSACK + discard craziness */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001378 prev = skb_rb_prev(skb);
1379 if (!prev)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001380 goto fallback;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001381
1382 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1383 goto fallback;
1384
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001385 if (!tcp_skb_can_collapse_to(prev))
1386 goto fallback;
1387
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001388 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1389 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1390
1391 if (in_sack) {
1392 len = skb->len;
1393 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001394 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001395
1396 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1397 * drop this restriction as unnecessary
1398 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001399 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001400 goto fallback;
1401 } else {
1402 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1403 goto noop;
1404 /* CHECKME: This is non-MSS split case only?, this will
1405 * cause skipped skbs due to advancing loop btw, original
1406 * has that feature too
1407 */
1408 if (tcp_skb_pcount(skb) <= 1)
1409 goto noop;
1410
1411 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1412 if (!in_sack) {
1413 /* TODO: head merge to next could be attempted here
1414 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1415 * though it might not be worth of the additional hassle
1416 *
1417 * ...we can probably just fallback to what was done
1418 * previously. We could try merging non-SACKed ones
1419 * as well but it probably isn't going to buy off
1420 * because later SACKs might again split them, and
1421 * it would make skb timestamp tracking considerably
1422 * harder problem.
1423 */
1424 goto fallback;
1425 }
1426
1427 len = end_seq - TCP_SKB_CB(skb)->seq;
1428 BUG_ON(len < 0);
1429 BUG_ON(len > skb->len);
1430
1431 /* MSS boundaries should be honoured or else pcount will
1432 * severely break even though it makes things bit trickier.
1433 * Optimize common case to avoid most of the divides
1434 */
1435 mss = tcp_skb_mss(skb);
1436
1437 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1438 * drop this restriction as unnecessary
1439 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001440 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001441 goto fallback;
1442
1443 if (len == mss) {
1444 pcount = 1;
1445 } else if (len < mss) {
1446 goto noop;
1447 } else {
1448 pcount = len / mss;
1449 len = pcount * mss;
1450 }
1451 }
1452
Neal Cardwell4648dc92012-03-05 19:35:04 +00001453 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1454 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1455 goto fallback;
1456
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001457 if (!skb_shift(prev, skb, len))
1458 goto fallback;
Eric Dumazetf3319812017-10-05 22:21:26 -07001459 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001460 goto out;
1461
1462 /* Hole filled allows collapsing with the next as well, this is very
1463 * useful when hole on every nth skb pattern happens
1464 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001465 skb = skb_rb_next(prev);
1466 if (!skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001467 goto out;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001468
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001469 if (!skb_can_shift(skb) ||
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001470 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001471 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001472 goto out;
1473
1474 len = skb->len;
1475 if (skb_shift(prev, skb, len)) {
1476 pcount += tcp_skb_pcount(skb);
Eric Dumazetf3319812017-10-05 22:21:26 -07001477 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1478 len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001479 }
1480
1481out:
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001482 return prev;
1483
1484noop:
1485 return skb;
1486
1487fallback:
Eric Dumazetc10d9312016-04-29 14:16:47 -07001488 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001489 return NULL;
1490}
1491
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001492static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1493 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001494 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001495 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001496 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001497{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001498 struct tcp_sock *tp = tcp_sk(sk);
1499 struct sk_buff *tmp;
1500
Eric Dumazet75c119a2017-10-05 22:21:27 -07001501 skb_rbtree_walk_from(skb) {
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001502 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001503 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001504
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001505 /* queue is in-order => we can short-circuit the walk early */
1506 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1507 break;
1508
Ian Morris00db4122015-04-03 09:17:27 +01001509 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001510 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1511 in_sack = tcp_match_skb_to_sack(sk, skb,
1512 next_dup->start_seq,
1513 next_dup->end_seq);
1514 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001515 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001516 }
1517
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001518 /* skb reference here is a bit tricky to get right, since
1519 * shifting can eat and free both this skb and the next,
1520 * so not even _safe variant of the loop is enough.
1521 */
1522 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001523 tmp = tcp_shift_skb_data(sk, skb, state,
1524 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001525 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001526 if (tmp != skb) {
1527 skb = tmp;
1528 continue;
1529 }
1530
1531 in_sack = 0;
1532 } else {
1533 in_sack = tcp_match_skb_to_sack(sk, skb,
1534 start_seq,
1535 end_seq);
1536 }
1537 }
1538
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001539 if (unlikely(in_sack < 0))
1540 break;
1541
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001542 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001543 TCP_SKB_CB(skb)->sacked =
1544 tcp_sacktag_one(sk,
1545 state,
1546 TCP_SKB_CB(skb)->sacked,
1547 TCP_SKB_CB(skb)->seq,
1548 TCP_SKB_CB(skb)->end_seq,
1549 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001550 tcp_skb_pcount(skb),
Eric Dumazet9a568de2017-05-16 14:00:14 -07001551 skb->skb_mstamp);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001552 tcp_rate_skb_delivered(sk, skb, state->rate);
Eric Dumazete2080072017-10-04 12:59:58 -07001553 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1554 list_del_init(&skb->tcp_tsorted_anchor);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001555
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001556 if (!before(TCP_SKB_CB(skb)->seq,
1557 tcp_highest_sack_seq(tp)))
1558 tcp_advance_highest_sack(sk, skb);
1559 }
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001560 }
1561 return skb;
1562}
1563
Eric Dumazet75c119a2017-10-05 22:21:27 -07001564static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1565 struct tcp_sacktag_state *state,
1566 u32 seq)
1567{
1568 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1569 struct sk_buff *skb;
Eric Dumazet75c119a2017-10-05 22:21:27 -07001570
1571 while (*p) {
1572 parent = *p;
1573 skb = rb_to_skb(parent);
1574 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1575 p = &parent->rb_left;
1576 continue;
1577 }
1578 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1579 p = &parent->rb_right;
1580 continue;
1581 }
Eric Dumazet75c119a2017-10-05 22:21:27 -07001582 return skb;
1583 }
1584 return NULL;
1585}
1586
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001587static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001588 struct tcp_sacktag_state *state,
1589 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001590{
Eric Dumazet75c119a2017-10-05 22:21:27 -07001591 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1592 return skb;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001593
Eric Dumazet75c119a2017-10-05 22:21:27 -07001594 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001595}
1596
1597static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1598 struct sock *sk,
1599 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001600 struct tcp_sacktag_state *state,
1601 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001602{
Ian Morris51456b22015-04-03 09:17:26 +01001603 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001604 return skb;
1605
1606 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001607 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1608 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1609 next_dup->start_seq, next_dup->end_seq,
1610 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001611 }
1612
1613 return skb;
1614}
1615
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001616static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001617{
1618 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1619}
1620
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001622tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001623 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624{
1625 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001626 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1627 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001628 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001629 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001630 struct tcp_sack_block *cache;
1631 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001632 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001633 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001634 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001635 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001636 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001638 state->flag = 0;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001639 state->reord = tp->snd_nxt;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001640
Yuchung Cheng737ff312017-11-08 13:01:27 -08001641 if (!tp->sacked_out)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001642 tcp_highest_sack_reset(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001644 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001645 num_sacks, prior_snd_una);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001646 if (found_dup_sack) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001647 state->flag |= FLAG_DSACKING_ACK;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001648 tp->delivered++; /* A spurious retransmission is delivered */
1649 }
Baruch Even6f746512007-02-04 23:36:42 -08001650
1651 /* Eliminate too old ACKs, but take into
1652 * account more or less fresh ones, they can
1653 * contain valid SACK info.
1654 */
1655 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1656 return 0;
1657
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001658 if (!tp->packets_out)
1659 goto out;
1660
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001661 used_sacks = 0;
1662 first_sack_index = 0;
1663 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001664 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001665
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001666 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1667 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001668
1669 if (!tcp_is_sackblock_valid(tp, dup_sack,
1670 sp[used_sacks].start_seq,
1671 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001672 int mib_idx;
1673
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001674 if (dup_sack) {
1675 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001676 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001677 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001678 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001679 } else {
1680 /* Don't count olds caused by ACK reordering */
1681 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1682 !after(sp[used_sacks].end_seq, tp->snd_una))
1683 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001684 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001685 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001686
Eric Dumazetc10d9312016-04-29 14:16:47 -07001687 NET_INC_STATS(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001688 if (i == 0)
1689 first_sack_index = -1;
1690 continue;
1691 }
1692
1693 /* Ignore very old stuff early */
1694 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1695 continue;
1696
1697 used_sacks++;
1698 }
1699
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001700 /* order SACK blocks to allow in order walk of the retrans queue */
1701 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001702 for (j = 0; j < i; j++) {
1703 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001704 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001705
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001706 /* Track where the first SACK block goes to */
1707 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001708 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001709 }
1710 }
1711 }
1712
Eric Dumazet75c119a2017-10-05 22:21:27 -07001713 state->mss_now = tcp_current_mss(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07001714 skb = NULL;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001715 i = 0;
1716
1717 if (!tp->sacked_out) {
1718 /* It's already past, so skip checking against it */
1719 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1720 } else {
1721 cache = tp->recv_sack_cache;
1722 /* Skip empty blocks in at head of the cache */
1723 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1724 !cache->end_seq)
1725 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001726 }
1727
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001728 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001729 u32 start_seq = sp[i].start_seq;
1730 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001731 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001732 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001733
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001734 if (found_dup_sack && ((i + 1) == first_sack_index))
1735 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001737 /* Skip too early cached blocks */
1738 while (tcp_sack_cache_ok(tp, cache) &&
1739 !before(start_seq, cache->end_seq))
1740 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001742 /* Can skip some work by looking recv_sack_cache? */
1743 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1744 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001745
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001746 /* Head todo? */
1747 if (before(start_seq, cache->start_seq)) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001748 skb = tcp_sacktag_skip(skb, sk, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001749 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001750 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001751 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001752 start_seq,
1753 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001754 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001755 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001756
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001757 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001758 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001759 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001760
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001761 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001762 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001763 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001764
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001765 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001766 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001767 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001768 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001769 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001770 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001771 cache++;
1772 goto walk;
1773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001775 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001776 /* Check overlap against next cached too (past this one already) */
1777 cache++;
1778 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001780
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001781 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1782 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001783 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001784 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001785 }
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001786 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001787
1788walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001789 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001790 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001791
1792advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001793 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 }
1795
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001796 /* Clear the head of the cache sack blocks so we can skip it next time */
1797 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1798 tp->recv_sack_cache[i].start_seq = 0;
1799 tp->recv_sack_cache[i].end_seq = 0;
1800 }
1801 for (j = 0; j < used_sacks; j++)
1802 tp->recv_sack_cache[i++] = sp[j];
1803
Yuchung Cheng737ff312017-11-08 13:01:27 -08001804 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1805 tcp_check_sack_reordering(sk, state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001807 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001808out:
1809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001811 WARN_ON((int)tp->sacked_out < 0);
1812 WARN_ON((int)tp->lost_out < 0);
1813 WARN_ON((int)tp->retrans_out < 0);
1814 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001816 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817}
1818
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001819/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001820 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001821 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001822static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001823{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001824 u32 holes;
1825
1826 holes = max(tp->lost_out, 1U);
1827 holes = min(holes, tp->packets_out);
1828
1829 if ((tp->sacked_out + holes) > tp->packets_out) {
1830 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001831 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001832 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001833 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001834}
1835
1836/* If we receive more dupacks than we expected counting segments
1837 * in assumption of absent reordering, interpret this as reordering.
1838 * The only another reason could be bug in receiver TCP.
1839 */
1840static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1841{
1842 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -08001843
1844 if (!tcp_limit_reno_sacked(tp))
1845 return;
1846
1847 tp->reordering = min_t(u32, tp->packets_out + addend,
1848 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1849 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001850}
1851
1852/* Emulate SACKs for SACKless connection: account for a new dupack. */
1853
1854static void tcp_add_reno_sack(struct sock *sk)
1855{
1856 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001857 u32 prior_sacked = tp->sacked_out;
1858
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001859 tp->sacked_out++;
1860 tcp_check_reno_reordering(sk, 0);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001861 if (tp->sacked_out > prior_sacked)
1862 tp->delivered++; /* Some out-of-order packet is delivered */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001863 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001864}
1865
1866/* Account for ACK, ACKing some data in Reno Recovery phase. */
1867
1868static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1869{
1870 struct tcp_sock *tp = tcp_sk(sk);
1871
1872 if (acked > 0) {
1873 /* One ACK acked hole. The rest eat duplicate ACKs. */
Yuchung Chengddf1af62016-02-02 10:33:06 -08001874 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001875 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001876 tp->sacked_out = 0;
1877 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001878 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001879 }
1880 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001881 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001882}
1883
1884static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1885{
1886 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001887}
1888
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001889void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001894 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001895 tp->sacked_out = 0;
1896}
1897
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001898static inline void tcp_init_undo(struct tcp_sock *tp)
1899{
1900 tp->undo_marker = tp->snd_una;
1901 /* Retransmission still in flight may cause DSACKs later. */
1902 tp->undo_retrans = tp->retrans_out ? : -1;
1903}
1904
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001905/* Enter Loss state. If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 * and reset tags completely, otherwise preserve SACKs. If receiver
1907 * dropped its ofo queue, we will know this due to reneging detection.
1908 */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001909void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001911 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov1043e252016-02-03 09:46:52 +02001913 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 struct sk_buff *skb;
Yuchung Chengcc663f42017-04-07 11:42:05 -07001915 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001916 bool is_reneg; /* is receiver reneging on SACKs? */
Neal Cardwell0682e692016-09-19 23:39:13 -04001917 bool mark_lost;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
1919 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001920 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1921 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001922 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1923 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Yuchung Cheng4faf7832017-08-03 20:38:51 -07001924 tp->prior_cwnd = tp->snd_cwnd;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001925 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1926 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001927 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 }
1929 tp->snd_cwnd = 1;
1930 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07001931 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001933 tp->retrans_out = 0;
1934 tp->lost_out = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001935
1936 if (tcp_is_reno(tp))
1937 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
Eric Dumazet75c119a2017-10-05 22:21:27 -07001939 skb = tcp_rtx_queue_head(sk);
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001940 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1941 if (is_reneg) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07001942 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001943 tp->sacked_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001944 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07001945 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946
Eric Dumazet75c119a2017-10-05 22:21:27 -07001947 skb_rbtree_walk_from(skb) {
Neal Cardwell0682e692016-09-19 23:39:13 -04001948 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1949 is_reneg);
1950 if (mark_lost)
1951 tcp_sum_lost(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
Neal Cardwell0682e692016-09-19 23:39:13 -04001953 if (mark_lost) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1955 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1956 tp->lost_out += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 }
1958 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001959 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001961 /* Timeout in disordered state after receiving substantial DUPACKs
1962 * suggests that the degree of reordering is over-estimated.
1963 */
1964 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02001965 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001966 tp->reordering = min_t(unsigned int, tp->reordering,
Nikolay Borisov1043e252016-02-03 09:46:52 +02001967 net->ipv4.sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001968 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02001970 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001971
Yuchung Chengcc663f42017-04-07 11:42:05 -07001972 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1973 * loss recovery is underway except recurring timeout(s) on
1974 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1975 *
1976 * In theory F-RTO can be used repeatedly during loss recovery.
1977 * In practice this interacts badly with broken middle-boxes that
1978 * falsely raise the receive window, which results in repeated
1979 * timeouts and stop-and-go behavior.
Yuchung Chenge33099f2013-03-20 13:33:00 +00001980 */
Eric Dumazetaf9b69a2017-10-26 21:55:10 -07001981 tp->frto = net->ipv4.sysctl_tcp_frto &&
Yuchung Chengcc663f42017-04-07 11:42:05 -07001982 (new_recovery || icsk->icsk_retransmits) &&
1983 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984}
1985
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001986/* If ACK arrived pointing to a remembered SACK, it means that our
1987 * remembered SACKs do not reflect real state of receiver i.e.
1988 * receiver _host_ is heavily congested (or buggy).
1989 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001990 * To avoid big spurious retransmission bursts due to transient SACK
1991 * scoreboard oddities that look like reneging, we give the receiver a
1992 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1993 * restore sanity to the SACK scoreboard. If the apparent reneging
1994 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001995 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001996static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001998 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001999 struct tcp_sock *tp = tcp_sk(sk);
2000 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2001 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002004 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002005 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002007 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008}
2009
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002010/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2011 * counter when SACK is enabled (without SACK, sacked_out is used for
2012 * that purpose).
2013 *
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002014 * With reordering, holes may still be in flight, so RFC3517 recovery
2015 * uses pure sacked_out (total number of SACKed segments) even though
2016 * it violates the RFC that uses duplicate ACKs, often these are equal
2017 * but when e.g. out-of-window ACKs or packet duplication occurs,
2018 * they differ. Since neither occurs due to loss, TCP should really
2019 * ignore them.
2020 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002021static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002022{
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002023 return tp->sacked_out + 1;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002024}
2025
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002026/* Linux NewReno/SACK/ECN state machine.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 * --------------------------------------
2028 *
2029 * "Open" Normal state, no dubious events, fast path.
2030 * "Disorder" In all the respects it is "Open",
2031 * but requires a bit more attention. It is entered when
2032 * we see some SACKs or dupacks. It is split of "Open"
2033 * mainly to move some processing from fast path to slow one.
2034 * "CWR" CWND was reduced due to some Congestion Notification event.
2035 * It can be ECN, ICMP source quench, local device congestion.
2036 * "Recovery" CWND was reduced, we are fast-retransmitting.
2037 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2038 *
2039 * tcp_fastretrans_alert() is entered:
2040 * - each incoming ACK, if state is not "Open"
2041 * - when arrived ACK is unusual, namely:
2042 * * SACK
2043 * * Duplicate ACK.
2044 * * ECN ECE.
2045 *
2046 * Counting packets in flight is pretty simple.
2047 *
2048 * in_flight = packets_out - left_out + retrans_out
2049 *
2050 * packets_out is SND.NXT-SND.UNA counted in packets.
2051 *
2052 * retrans_out is number of retransmitted segments.
2053 *
2054 * left_out is number of segments left network, but not ACKed yet.
2055 *
2056 * left_out = sacked_out + lost_out
2057 *
2058 * sacked_out: Packets, which arrived to receiver out of order
2059 * and hence not ACKed. With SACKs this number is simply
2060 * amount of SACKed data. Even without SACKs
2061 * it is easy to give pretty reliable estimate of this number,
2062 * counting duplicate ACKs.
2063 *
2064 * lost_out: Packets lost by network. TCP has no explicit
2065 * "loss notification" feedback from network (for now).
2066 * It means that this number can be only _guessed_.
2067 * Actually, it is the heuristics to predict lossage that
2068 * distinguishes different algorithms.
2069 *
2070 * F.e. after RTO, when all the queue is considered as lost,
2071 * lost_out = packets_out and in_flight = retrans_out.
2072 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002073 * Essentially, we have now a few algorithms detecting
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 * lost packets.
2075 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002076 * If the receiver supports SACK:
2077 *
2078 * RFC6675/3517: It is the conventional algorithm. A packet is
2079 * considered lost if the number of higher sequence packets
2080 * SACKed is greater than or equal the DUPACK thoreshold
2081 * (reordering). This is implemented in tcp_mark_head_lost and
2082 * tcp_update_scoreboard.
2083 *
2084 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2085 * (2017-) that checks timing instead of counting DUPACKs.
2086 * Essentially a packet is considered lost if it's not S/ACKed
2087 * after RTT + reordering_window, where both metrics are
2088 * dynamically measured and adjusted. This is implemented in
2089 * tcp_rack_mark_lost.
2090 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002091 * If the receiver does not support SACK:
2092 *
2093 * NewReno (RFC6582): in Recovery we assume that one segment
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 * is lost (classic Reno). While we are in Recovery and
2095 * a partial ACK arrives, we assume that one more packet
2096 * is lost (NewReno). This heuristics are the same in NewReno
2097 * and SACK.
2098 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 * Really tricky (and requiring careful tuning) part of algorithm
2100 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2101 * The first determines the moment _when_ we should reduce CWND and,
2102 * hence, slow down forward transmission. In fact, it determines the moment
2103 * when we decide that hole is caused by loss, rather than by a reorder.
2104 *
2105 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2106 * holes, caused by lost packets.
2107 *
2108 * And the most logically complicated part of algorithm is undo
2109 * heuristics. We detect false retransmits due to both too early
2110 * fast retransmit (reordering) and underestimated RTO, analyzing
2111 * timestamps and D-SACKs. When we detect that some segments were
2112 * retransmitted by mistake and CWND reduction was wrong, we undo
2113 * window reduction and abort recovery phase. This logic is hidden
2114 * inside several functions named tcp_try_undo_<something>.
2115 */
2116
2117/* This function decides, when we should leave Disordered state
2118 * and enter Recovery phase, reducing congestion window.
2119 *
2120 * Main question: may we further continue forward transmission
2121 * with the same cwnd?
2122 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002123static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002125 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126
2127 /* Trick#1: The loss is proven. */
2128 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002129 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
2131 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002132 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002133 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002135 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136}
2137
Yuchung Cheng974c1232012-01-19 14:42:21 +00002138/* Detect loss in event "A" above by marking head of queue up as lost.
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002139 * For non-SACK(Reno) senders, the first "packets" number of segments
Yuchung Cheng974c1232012-01-19 14:42:21 +00002140 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2141 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2142 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002143 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002144static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002146 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 struct sk_buff *skb;
Neal Cardwelld88270e2016-01-25 14:01:53 -08002148 int cnt, oldcnt, lost;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002149 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002150 /* Use SACK to deduce losses of new sequences sent during recovery */
2151 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002153 WARN_ON(packets > tp->packets_out);
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002154 skb = tp->lost_skb_hint;
2155 if (skb) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002156 /* Head already handled? */
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002157 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002158 return;
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002159 cnt = tp->lost_cnt_hint;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002160 } else {
Eric Dumazet75c119a2017-10-05 22:21:27 -07002161 skb = tcp_rtx_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002162 cnt = 0;
2163 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164
Eric Dumazet75c119a2017-10-05 22:21:27 -07002165 skb_rbtree_walk_from(skb) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002166 /* TODO: do this better */
2167 /* this is not the most efficient way to do this... */
2168 tp->lost_skb_hint = skb;
2169 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002170
Yuchung Cheng974c1232012-01-19 14:42:21 +00002171 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002172 break;
2173
2174 oldcnt = cnt;
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002175 if (tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002176 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2177 cnt += tcp_skb_pcount(skb);
2178
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002179 if (cnt > packets) {
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002180 if (tcp_is_sack(tp) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002181 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de75592010-09-24 13:22:06 +00002182 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002183 break;
2184
Eric Dumazetf69ad292015-06-11 09:15:18 -07002185 mss = tcp_skb_mss(skb);
Neal Cardwelld88270e2016-01-25 14:01:53 -08002186 /* If needed, chop off the prefix to mark as lost. */
2187 lost = (packets - oldcnt) * mss;
2188 if (lost < skb->len &&
Eric Dumazet75c119a2017-10-05 22:21:27 -07002189 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2190 lost, mss, GFP_ATOMIC) < 0)
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002191 break;
2192 cnt = packets;
2193 }
2194
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002195 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002196
2197 if (mark_head)
2198 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002200 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201}
2202
2203/* Account newly detected lost packet(s) */
2204
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002205static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002207 struct tcp_sock *tp = tcp_sk(sk);
2208
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002209 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002210 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002212 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002213 if (sacked_upto >= 0)
2214 tcp_mark_head_lost(sk, sacked_upto, 0);
2215 else if (fast_rexmit)
2216 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218}
2219
Yuchung Cheng77c63122015-10-16 21:57:44 -07002220static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2221{
2222 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2223 before(tp->rx_opt.rcv_tsecr, when);
2224}
2225
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002226/* skb is spurious retransmitted if the returned timestamp echo
2227 * reply is prior to the skb transmission time
2228 */
2229static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2230 const struct sk_buff *skb)
2231{
2232 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2233 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2234}
2235
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236/* Nothing was retransmitted or returned timestamp is less
2237 * than timestamp of the first retransmission.
2238 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002239static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240{
2241 return !tp->retrans_stamp ||
Yuchung Cheng77c63122015-10-16 21:57:44 -07002242 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243}
2244
2245/* Undo procedures. */
2246
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002247/* We can clear retrans_stamp when there are no retransmissions in the
2248 * window. It would seem that it is trivially available for us in
2249 * tp->retrans_out, however, that kind of assumptions doesn't consider
2250 * what will happen if errors occur when sending retransmission for the
2251 * second time. ...It could the that such segment has only
2252 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2253 * the head skb is enough except for some reneging corner cases that
2254 * are not worth the effort.
2255 *
2256 * Main reason for all this complexity is the fact that connection dying
2257 * time now depends on the validity of the retrans_stamp, in particular,
2258 * that successive retransmissions of a segment must not advance
2259 * retrans_stamp under any conditions.
2260 */
2261static bool tcp_any_retrans_done(const struct sock *sk)
2262{
2263 const struct tcp_sock *tp = tcp_sk(sk);
2264 struct sk_buff *skb;
2265
2266 if (tp->retrans_out)
2267 return true;
2268
Eric Dumazet75c119a2017-10-05 22:21:27 -07002269 skb = tcp_rtx_queue_head(sk);
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002270 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2271 return true;
2272
2273 return false;
2274}
2275
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002276static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277{
Joe Perches39347882017-09-10 19:02:25 -07002278#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002279 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002281
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002282 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002283 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2284 msg,
2285 &inet->inet_daddr, ntohs(inet->inet_dport),
2286 tp->snd_cwnd, tcp_left_out(tp),
2287 tp->snd_ssthresh, tp->prior_ssthresh,
2288 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002289 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002290#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002291 else if (sk->sk_family == AF_INET6) {
Joe Perches91df42b2012-05-15 14:11:54 +00002292 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2293 msg,
Eric Dumazet019b1c92016-09-22 17:54:00 -07002294 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
Joe Perches91df42b2012-05-15 14:11:54 +00002295 tp->snd_cwnd, tcp_left_out(tp),
2296 tp->snd_ssthresh, tp->prior_ssthresh,
2297 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002298 }
2299#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300#endif
Joe Perches39347882017-09-10 19:02:25 -07002301}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302
Yuchung Cheng7026b912013-05-29 14:20:13 +00002303static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002305 struct tcp_sock *tp = tcp_sk(sk);
2306
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002307 if (unmark_loss) {
2308 struct sk_buff *skb;
2309
Eric Dumazet75c119a2017-10-05 22:21:27 -07002310 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002311 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2312 }
2313 tp->lost_out = 0;
2314 tcp_clear_all_retrans_hints(tp);
2315 }
2316
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002318 const struct inet_connection_sock *icsk = inet_csk(sk);
2319
Florian Westphale9799182016-11-21 14:18:38 +01002320 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321
Yuchung Cheng7026b912013-05-29 14:20:13 +00002322 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002324 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 }
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002327 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002328 tp->undo_marker = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329}
2330
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002331static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002333 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334}
2335
2336/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002337static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002339 struct tcp_sock *tp = tcp_sk(sk);
2340
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002342 int mib_idx;
2343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 /* Happy end! We did not retransmit anything
2345 * or our original transmission succeeded.
2346 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002347 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002348 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002349 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002350 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002352 mib_idx = LINUX_MIB_TCPFULLUNDO;
2353
Eric Dumazetc10d9312016-04-29 14:16:47 -07002354 NET_INC_STATS(sock_net(sk), mib_idx);
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002355 } else if (tp->rack.reo_wnd_persist) {
2356 tp->rack.reo_wnd_persist--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002358 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 /* Hold old state until something *above* high_seq
2360 * is ACKed. For Reno it is MUST to prevent false
2361 * fast retransmits (RFC2582). SACK TCP is safe. */
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002362 if (!tcp_any_retrans_done(sk))
2363 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002364 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002366 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002367 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368}
2369
2370/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002371static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002373 struct tcp_sock *tp = tcp_sk(sk);
2374
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 if (tp->undo_marker && !tp->undo_retrans) {
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002376 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2377 tp->rack.reo_wnd_persist + 1);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002378 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002379 tcp_undo_cwnd_reduction(sk, false);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002380 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002381 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002383 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384}
2385
Yuchung Chenge33099f2013-03-20 13:33:00 +00002386/* Undo during loss recovery after partial ACK or using F-RTO. */
2387static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002389 struct tcp_sock *tp = tcp_sk(sk);
2390
Yuchung Chenge33099f2013-03-20 13:33:00 +00002391 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002392 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002393
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002394 DBGUNDO(sk, "partial loss");
Eric Dumazetc10d9312016-04-29 14:16:47 -07002395 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002396 if (frto_undo)
Eric Dumazetc10d9312016-04-29 14:16:47 -07002397 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002398 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002399 inet_csk(sk)->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002400 if (frto_undo || tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002401 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002402 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002404 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405}
2406
Yuchung Cheng37598242015-07-01 14:11:15 -07002407/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002408 * It computes the number of packets to send (sndcnt) based on packets newly
2409 * delivered:
2410 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2411 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002412 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2413 * But when the retransmits are acked without further losses, PRR
2414 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002415 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002416static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002417{
2418 struct tcp_sock *tp = tcp_sk(sk);
2419
2420 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002421 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002422 tp->snd_cwnd_cnt = 0;
2423 tp->prior_cwnd = tp->snd_cwnd;
2424 tp->prr_delivered = 0;
2425 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002426 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002427 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002428}
2429
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002430void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002431{
2432 struct tcp_sock *tp = tcp_sk(sk);
2433 int sndcnt = 0;
2434 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2435
Yuchung Cheng8b8a3212016-01-06 12:42:38 -08002436 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2437 return;
2438
Yuchung Cheng684bad12012-09-02 17:38:04 +00002439 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002440 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002441 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2442 tp->prior_cwnd - 1;
2443 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yuchung Cheng37598242015-07-01 14:11:15 -07002444 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2445 !(flag & FLAG_LOST_RETRANS)) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002446 sndcnt = min_t(int, delta,
2447 max_t(int, tp->prr_delivered - tp->prr_out,
2448 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002449 } else {
2450 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002451 }
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002452 /* Force a fast retransmit upon entering fast recovery */
2453 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002454 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2455}
2456
Yuchung Cheng684bad12012-09-02 17:38:04 +00002457static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002459 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002460
Yuchung Chengc0402762016-09-19 23:39:21 -04002461 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2462 return;
2463
Yuchung Cheng684bad12012-09-02 17:38:04 +00002464 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
Yuchung Chenged254972017-08-01 13:22:32 -07002465 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2466 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002467 tp->snd_cwnd = tp->snd_ssthresh;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002468 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002469 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002470 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471}
2472
Yuchung Cheng684bad12012-09-02 17:38:04 +00002473/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002474void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002475{
2476 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002477
2478 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002479 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002480 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002481 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002482 tcp_set_ca_state(sk, TCP_CA_CWR);
2483 }
2484}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002485EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002486
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002487static void tcp_try_keep_open(struct sock *sk)
2488{
2489 struct tcp_sock *tp = tcp_sk(sk);
2490 int state = TCP_CA_Open;
2491
Neal Cardwellf6982042011-11-16 08:58:04 +00002492 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002493 state = TCP_CA_Disorder;
2494
2495 if (inet_csk(sk)->icsk_ca_state != state) {
2496 tcp_set_ca_state(sk, state);
2497 tp->high_seq = tp->snd_nxt;
2498 }
2499}
2500
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002501static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002503 struct tcp_sock *tp = tcp_sk(sk);
2504
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002505 tcp_verify_left_out(tp);
2506
Yuchung Cheng9b441902013-03-20 13:32:58 +00002507 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 tp->retrans_stamp = 0;
2509
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002510 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002511 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002513 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002514 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 }
2516}
2517
John Heffner5d424d52006-03-20 17:53:41 -08002518static void tcp_mtup_probe_failed(struct sock *sk)
2519{
2520 struct inet_connection_sock *icsk = inet_csk(sk);
2521
2522 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2523 icsk->icsk_mtup.probe_size = 0;
Eric Dumazetc10d9312016-04-29 14:16:47 -07002524 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002525}
2526
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002527static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002528{
2529 struct tcp_sock *tp = tcp_sk(sk);
2530 struct inet_connection_sock *icsk = inet_csk(sk);
2531
2532 /* FIXME: breaks with very large cwnd */
2533 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2534 tp->snd_cwnd = tp->snd_cwnd *
2535 tcp_mss_to_mtu(sk, tp->mss_cache) /
2536 icsk->icsk_mtup.probe_size;
2537 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002538 tp->snd_cwnd_stamp = tcp_jiffies32;
John Heffner9c6d5e52010-10-06 21:18:02 -07002539 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002540
2541 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2542 icsk->icsk_mtup.probe_size = 0;
2543 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002544 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002545}
2546
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002547/* Do a simple retransmit without using the backoff mechanisms in
2548 * tcp_timer. This is used for path mtu discovery.
2549 * The socket is already locked here.
2550 */
2551void tcp_simple_retransmit(struct sock *sk)
2552{
2553 const struct inet_connection_sock *icsk = inet_csk(sk);
2554 struct tcp_sock *tp = tcp_sk(sk);
2555 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002556 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002557
Eric Dumazet75c119a2017-10-05 22:21:27 -07002558 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002559 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002560 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2561 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2562 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2563 tp->retrans_out -= tcp_skb_pcount(skb);
2564 }
2565 tcp_skb_mark_lost_uncond_verify(tp, skb);
2566 }
2567 }
2568
2569 tcp_clear_retrans_hints_partial(tp);
2570
Yuchung Cheng0eb96bf2017-11-07 15:33:43 -08002571 if (!tp->lost_out)
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002572 return;
2573
2574 if (tcp_is_reno(tp))
2575 tcp_limit_reno_sacked(tp);
2576
2577 tcp_verify_left_out(tp);
2578
2579 /* Don't muck with the congestion window here.
2580 * Reason is that we do not increase amount of _data_
2581 * in network, but units changed and effective
2582 * cwnd/ssthresh really reduced now.
2583 */
2584 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2585 tp->high_seq = tp->snd_nxt;
2586 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2587 tp->prior_ssthresh = 0;
2588 tp->undo_marker = 0;
2589 tcp_set_ca_state(sk, TCP_CA_Loss);
2590 }
2591 tcp_xmit_retransmit_queue(sk);
2592}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002593EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002594
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002595void tcp_enter_recovery(struct sock *sk, bool ece_ack)
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002596{
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 int mib_idx;
2599
2600 if (tcp_is_reno(tp))
2601 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2602 else
2603 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2604
Eric Dumazetc10d9312016-04-29 14:16:47 -07002605 NET_INC_STATS(sock_net(sk), mib_idx);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002606
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002607 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002608 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002609
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002610 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002611 if (!ece_ack)
2612 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002613 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002614 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002615 tcp_set_ca_state(sk, TCP_CA_Recovery);
2616}
2617
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002618/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2619 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2620 */
Yuchung Chenge662ca42016-02-02 10:33:04 -08002621static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2622 int *rexmit)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002623{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002624 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002625 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002626
Yuchung Chengda34ac72015-05-18 12:31:44 -07002627 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2628 tcp_try_undo_loss(sk, false))
2629 return;
2630
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002631 /* The ACK (s)acks some never-retransmitted data meaning not all
2632 * the data packets before the timeout were lost. Therefore we
2633 * undo the congestion window and state. This is essentially
2634 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2635 * a retransmitted skb is permantly marked, we can apply such an
2636 * operation even if F-RTO was not used.
2637 */
2638 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2639 tcp_try_undo_loss(sk, tp->undo_marker))
2640 return;
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002641
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002642 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002643 if (after(tp->snd_nxt, tp->high_seq)) {
2644 if (flag & FLAG_DATA_SACKED || is_dupack)
2645 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002646 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2647 tp->high_seq = tp->snd_nxt;
Yuchung Chenge662ca42016-02-02 10:33:04 -08002648 /* Step 2.b. Try send new data (but deferred until cwnd
2649 * is updated in tcp_ack()). Otherwise fall back to
2650 * the conventional recovery.
2651 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07002652 if (!tcp_write_queue_empty(sk) &&
Yuchung Chenge662ca42016-02-02 10:33:04 -08002653 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2654 *rexmit = REXMIT_NEW;
2655 return;
2656 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002657 tp->frto = 0;
2658 }
2659 }
2660
2661 if (recovered) {
2662 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002663 tcp_try_undo_recovery(sk);
2664 return;
2665 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002666 if (tcp_is_reno(tp)) {
2667 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2668 * delivered. Lower inflight to clock out (re)tranmissions.
2669 */
2670 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2671 tcp_add_reno_sack(sk);
2672 else if (flag & FLAG_SND_UNA_ADVANCED)
2673 tcp_reset_reno_sack(tp);
2674 }
Yuchung Chenge662ca42016-02-02 10:33:04 -08002675 *rexmit = REXMIT_LOST;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002676}
2677
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002678/* Undo during fast recovery after partial ACK. */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002679static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002680{
2681 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002682
Yuchung Cheng7026b912013-05-29 14:20:13 +00002683 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002684 /* Plain luck! Hole if filled with delayed
Yuchung Cheng737ff312017-11-08 13:01:27 -08002685 * packet, rather than with a retransmit. Check reordering.
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002686 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002687 tcp_check_sack_reordering(sk, prior_snd_una, 1);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002688
2689 /* We are getting evidence that the reordering degree is higher
2690 * than we realized. If there are no retransmits out then we
2691 * can undo. Otherwise we clock out new packets but do not
2692 * mark more packets lost or retransmit more.
2693 */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002694 if (tp->retrans_out)
Yuchung Cheng7026b912013-05-29 14:20:13 +00002695 return true;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002696
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002697 if (!tcp_any_retrans_done(sk))
2698 tp->retrans_stamp = 0;
2699
Yuchung Cheng7026b912013-05-29 14:20:13 +00002700 DBGUNDO(sk, "partial recovery");
2701 tcp_undo_cwnd_reduction(sk, true);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002702 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002703 tcp_try_keep_open(sk);
2704 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002705 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002706 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002707}
2708
Eric Dumazetefab8f82017-04-25 10:15:35 -07002709static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002710{
2711 struct tcp_sock *tp = tcp_sk(sk);
2712
2713 /* Use RACK to detect loss */
Eric Dumazete20223f2017-10-26 21:54:57 -07002714 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002715 u32 prior_retrans = tp->retrans_out;
2716
Eric Dumazet128eda82017-04-25 10:15:34 -07002717 tcp_rack_mark_lost(sk);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002718 if (prior_retrans > tp->retrans_out)
2719 *ack_flag |= FLAG_LOST_RETRANS;
2720 }
2721}
2722
Yuchung Cheng737ff312017-11-08 13:01:27 -08002723static bool tcp_force_fast_retransmit(struct sock *sk)
2724{
2725 struct tcp_sock *tp = tcp_sk(sk);
2726
2727 return after(tcp_highest_sack_seq(tp),
2728 tp->snd_una + tp->reordering * tp->mss_cache);
2729}
2730
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731/* Process an event, which can update packets-in-flight not trivially.
2732 * Main goal of this function is to calculate new estimate for left_out,
2733 * taking into account both packets sitting in receiver's buffer and
2734 * packets lost by network.
2735 *
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002736 * Besides that it updates the congestion state when packet loss or ECN
2737 * is detected. But it does not reduce the cwnd, it is done by the
2738 * congestion control later.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 *
2740 * It does _not_ decide what to send, it is made in function
2741 * tcp_xmit_retransmit_queue().
2742 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002743static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
Eric Dumazet1317a9d2017-04-25 10:15:36 -07002744 bool is_dupack, int *ack_flag, int *rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002746 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002748 int fast_rexmit = 0, flag = *ack_flag;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002749 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Yuchung Cheng737ff312017-11-08 13:01:27 -08002750 tcp_force_fast_retransmit(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751
Eric Dumazet8ba6dda2017-10-05 22:21:25 -07002752 if (!tp->packets_out && tp->sacked_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 tp->sacked_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002755 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002757 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 tp->prior_ssthresh = 0;
2759
2760 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002761 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 return;
2763
Yuchung Cheng974c1232012-01-19 14:42:21 +00002764 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002765 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
Yuchung Cheng974c1232012-01-19 14:42:21 +00002767 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002769 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002770 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 tp->retrans_stamp = 0;
2772 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002773 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 case TCP_CA_CWR:
2775 /* CWR is to be held something *above* high_seq
2776 * is ACKed for CWR bit to reach receiver. */
2777 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002778 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002779 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 }
2781 break;
2782
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002784 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002786 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002788 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 break;
2790 }
2791 }
2792
Yuchung Cheng974c1232012-01-19 14:42:21 +00002793 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002794 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002796 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002797 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002798 tcp_add_reno_sack(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002799 } else {
Yuchung Cheng737ff312017-11-08 13:01:27 -08002800 if (tcp_try_undo_partial(sk, prior_snd_una))
Yuchung Cheng7026b912013-05-29 14:20:13 +00002801 return;
2802 /* Partial ACK arrived. Force fast retransmit. */
2803 do_lost = tcp_is_reno(tp) ||
Yuchung Cheng737ff312017-11-08 13:01:27 -08002804 tcp_force_fast_retransmit(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002805 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002806 if (tcp_try_undo_dsack(sk)) {
2807 tcp_try_keep_open(sk);
2808 return;
2809 }
Eric Dumazetefab8f82017-04-25 10:15:35 -07002810 tcp_rack_identify_loss(sk, ack_flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 break;
2812 case TCP_CA_Loss:
Yuchung Chenge662ca42016-02-02 10:33:04 -08002813 tcp_process_loss(sk, flag, is_dupack, rexmit);
Eric Dumazetefab8f82017-04-25 10:15:35 -07002814 tcp_rack_identify_loss(sk, ack_flag);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002815 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2816 (*ack_flag & FLAG_LOST_RETRANS)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002818 /* Change state if cwnd is undone or retransmits are lost */
Gustavo A. R. Silvafcfd6df2017-10-16 15:48:55 -05002819 /* fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002821 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002822 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 tcp_reset_reno_sack(tp);
2824 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002825 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 }
2827
Neal Cardwellf6982042011-11-16 08:58:04 +00002828 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002829 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
Eric Dumazetefab8f82017-04-25 10:15:35 -07002831 tcp_rack_identify_loss(sk, ack_flag);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002832 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002833 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 return;
2835 }
2836
John Heffner5d424d52006-03-20 17:53:41 -08002837 /* MTU probe failure: don't reduce cwnd */
2838 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2839 icsk->icsk_mtup.probe_size &&
John Heffner0e7b13682006-03-20 21:32:58 -08002840 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002841 tcp_mtup_probe_failed(sk);
2842 /* Restores the reduction we did in tcp_mtup_probe() */
2843 tp->snd_cwnd++;
2844 tcp_simple_retransmit(sk);
2845 return;
2846 }
2847
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002849 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002850 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 }
2852
Yuchung Cheng3e59cb02013-05-17 13:45:05 +00002853 if (do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002854 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08002855 *rexmit = REXMIT_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856}
2857
Yuchung Chengf6722582015-10-16 21:57:42 -07002858static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2859{
Eric Dumazetbd239702017-10-27 07:47:28 -07002860 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
Neal Cardwell64033892016-09-19 23:39:10 -04002861 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengf6722582015-10-16 21:57:42 -07002862
Eric Dumazetac9517f2017-05-16 14:00:13 -07002863 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
Neal Cardwell64033892016-09-19 23:39:10 -04002864 rtt_us ? : jiffies_to_usecs(1));
Yuchung Chengf6722582015-10-16 21:57:42 -07002865}
2866
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002867static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2868 long seq_rtt_us, long sack_rtt_us,
2869 long ca_rtt_us, struct rate_sample *rs)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002870{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002871 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002872
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002873 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2874 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2875 * Karn's algorithm forbids taking RTT if some retransmitted data
2876 * is acked (RFC6298).
2877 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002878 if (seq_rtt_us < 0)
2879 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07002880
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 /* RTTM Rule: A TSecr value received in a segment is used to
2882 * update the averaged RTT measurement only if the segment
2883 * acknowledges some new data, i.e., only if it advances the
2884 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002887 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Eric Dumazet9a568de2017-05-16 14:00:14 -07002888 flag & FLAG_ACKED) {
2889 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2890 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2891
2892 seq_rtt_us = ca_rtt_us = delta_us;
2893 }
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002894 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002895 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002896 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897
Yuchung Chengf6722582015-10-16 21:57:42 -07002898 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2899 * always taken together with ACK, SACK, or TS-opts. Any negative
2900 * values will be skipped with the seq_rtt_us < 0 check above.
2901 */
2902 tcp_update_rtt_min(sk, ca_rtt_us);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002903 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002904 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002906 /* RFC6298: only reset backoff on valid RTT measurement. */
2907 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002908 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909}
2910
Yuchung Cheng375fe022013-07-22 16:20:45 -07002911/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002912void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002913{
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002914 struct rate_sample rs;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002915 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07002916
Eric Dumazet9a568de2017-05-16 14:00:14 -07002917 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2918 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002919
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002920 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002921}
2922
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002923
Eric Dumazet24901552014-05-02 21:18:05 -07002924static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002926 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07002927
2928 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002929 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930}
2931
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932/* Restart timer after forward progress on connection.
2933 * RFC2988 recommends to restart timer to now+rto.
2934 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002935void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002937 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002938 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002939
Jerry Chu168a8f52012-08-31 12:29:13 +00002940 /* If the retrans timer is currently being used by Fast Open
2941 * for SYN-ACK retrans purpose, stay put.
2942 */
2943 if (tp->fastopen_rsk)
2944 return;
2945
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002947 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002949 u32 rto = inet_csk(sk)->icsk_rto;
2950 /* Offset the time elapsed after installing regular RTO */
Yuchung Chengbec41a12017-01-12 22:11:39 -08002951 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002952 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Neal Cardwelle1a10ef2017-08-03 09:19:52 -04002953 s64 delta_us = tcp_rto_delta_us(sk);
Eric Dumazetb17b8a22017-05-18 09:15:58 -07002954 /* delta_us may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002955 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002956 */
Neal Cardwellcdbeb632017-08-16 17:53:36 -04002957 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002958 }
2959 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2960 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002962}
2963
Neal Cardwelldf92c832017-08-03 09:19:54 -04002964/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2965static void tcp_set_xmit_timer(struct sock *sk)
2966{
Neal Cardwelled66dfa2017-11-17 21:06:14 -05002967 if (!tcp_schedule_loss_probe(sk, true))
Neal Cardwelldf92c832017-08-03 09:19:54 -04002968 tcp_rearm_rto(sk);
2969}
2970
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002971/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002972static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973{
2974 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002975 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002977 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978
2979 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002980 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 return 0;
2982 packets_acked -= tcp_skb_pcount(skb);
2983
2984 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002986 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 }
2988
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002989 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990}
2991
Eric Dumazetad971f62014-10-11 15:17:29 -07002992static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
2993 u32 prior_snd_una)
2994{
2995 const struct skb_shared_info *shinfo;
2996
2997 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
Soheil Hassas Yeganeh6b084922016-04-02 23:08:08 -04002998 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
Eric Dumazetad971f62014-10-11 15:17:29 -07002999 return;
3000
3001 shinfo = skb_shinfo(skb);
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04003002 if (!before(shinfo->tskey, prior_snd_una) &&
Eric Dumazete2080072017-10-04 12:59:58 -07003003 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3004 tcp_skb_tsorted_save(skb) {
3005 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3006 } tcp_skb_tsorted_restore(skb);
3007 }
Eric Dumazetad971f62014-10-11 15:17:29 -07003008}
3009
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003010/* Remove acknowledged frames from the retransmission queue. If our packet
3011 * is before the ack sequence we can discard it as it's confirmed to have
3012 * arrived at the other end.
3013 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003014static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3015 u32 prior_snd_una,
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003016 struct tcp_sacktag_state *sack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003018 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -07003019 u64 first_ackt, last_ackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003020 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003021 u32 prior_sacked = tp->sacked_out;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003022 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003023 struct sk_buff *skb, *next;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003024 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003025 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003026 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003027 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003028 u32 pkts_acked = 0;
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003029 u32 last_in_flight = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003030 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003031 int flag = 0;
3032
Eric Dumazet9a568de2017-05-16 14:00:14 -07003033 first_ackt = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034
Eric Dumazet75c119a2017-10-05 22:21:27 -07003035 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003036 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Yuchung Cheng737ff312017-11-08 13:01:27 -08003037 const u32 start_seq = scb->seq;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003038 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003039 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040
Eric Dumazetad971f62014-10-11 15:17:29 -07003041 tcp_ack_tstamp(sk, skb, prior_snd_una);
Willem de Bruijn712a7222014-08-12 14:53:16 -04003042
Gavin McCullagh2072c222007-12-29 19:11:21 -08003043 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003045 if (tcp_skb_pcount(skb) == 1 ||
3046 !after(tp->snd_una, scb->seq))
3047 break;
3048
Ilpo Järvinen72018832007-12-30 04:37:55 -08003049 acked_pcount = tcp_tso_acked(sk, skb);
3050 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003051 break;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003052 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003053 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003054 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 }
3056
Eric Dumazetad971f62014-10-11 15:17:29 -07003057 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003058 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003059 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003060 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003061 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003062 last_ackt = skb->skb_mstamp;
Eric Dumazet9a568de2017-05-16 14:00:14 -07003063 WARN_ON_ONCE(last_ackt == 0);
3064 if (!first_ackt)
Eric Dumazet740b0f12014-02-26 14:02:48 -08003065 first_ackt = last_ackt;
3066
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003067 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003068 if (before(start_seq, reord))
3069 reord = start_seq;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003070 if (!after(scb->end_seq, tp->high_seq))
3071 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003072 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003073
Yuchung Chengddf1af62016-02-02 10:33:06 -08003074 if (sacked & TCPCB_SACKED_ACKED) {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003075 tp->sacked_out -= acked_pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003076 } else if (tcp_is_sack(tp)) {
3077 tp->delivered += acked_pcount;
3078 if (!tcp_skb_spurious_retrans(tp, skb))
Yuchung Cheng1d0833d2017-01-12 22:11:34 -08003079 tcp_rack_advance(tp, sacked, scb->end_seq,
Eric Dumazet9a568de2017-05-16 14:00:14 -07003080 skb->skb_mstamp);
Yuchung Chengddf1af62016-02-02 10:33:06 -08003081 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003082 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003083 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003084
Ilpo Järvinen72018832007-12-30 04:37:55 -08003085 tp->packets_out -= acked_pcount;
3086 pkts_acked += acked_pcount;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003087 tcp_rate_skb_delivered(sk, skb, sack->rate);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003088
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003089 /* Initial outgoing SYN's get put onto the write_queue
3090 * just like anything else we transmit. It is not
3091 * true data, and if we misinform our callers that
3092 * this ACK acks real data, we will erroneously exit
3093 * connection startup slow start one packet too
3094 * quickly. This is severely frowned upon behavior.
3095 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003096 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003097 flag |= FLAG_DATA_ACKED;
3098 } else {
3099 flag |= FLAG_SYN_ACKED;
3100 tp->retrans_stamp = 0;
3101 }
3102
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003103 if (!fully_acked)
3104 break;
3105
Eric Dumazet75c119a2017-10-05 22:21:27 -07003106 next = skb_rb_next(skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003107 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003108 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003109 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003110 tp->lost_skb_hint = NULL;
Eric Dumazet75c119a2017-10-05 22:21:27 -07003111 tcp_rtx_queue_unlink_and_free(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 }
3113
Francis Yan0f872302016-11-27 23:07:14 -08003114 if (!skb)
3115 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3116
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003117 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3118 tp->snd_up = tp->snd_una;
3119
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003120 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3121 flag |= FLAG_SACK_RENEGING;
3122
Eric Dumazet9a568de2017-05-16 14:00:14 -07003123 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3124 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3125 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003126 }
Eric Dumazet9a568de2017-05-16 14:00:14 -07003127 if (sack->first_sackt) {
3128 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3129 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003130 }
Yuchung Chengf6722582015-10-16 21:57:42 -07003131 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003132 ca_rtt_us, sack->rate);
Yuchung Chenged084952013-07-22 16:20:48 -07003133
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003134 if (flag & FLAG_ACKED) {
Neal Cardwelldf92c832017-08-03 09:19:54 -04003135 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003136 if (unlikely(icsk->icsk_mtup.probe_size &&
3137 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3138 tcp_mtup_probe_success(sk);
3139 }
3140
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003141 if (tcp_is_reno(tp)) {
3142 tcp_remove_reno_sacks(sk, pkts_acked);
3143 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003144 int delta;
3145
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003146 /* Non-retransmitted hole got filled? That's reordering */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003147 if (before(reord, prior_fack))
3148 tcp_check_sack_reordering(sk, reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003149
Yuchung Cheng713bafe2017-11-08 13:01:26 -08003150 delta = prior_sacked - tp->sacked_out;
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003151 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003152 }
Eric Dumazet740b0f12014-02-26 14:02:48 -08003153 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
Eric Dumazet9a568de2017-05-16 14:00:14 -07003154 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003155 /* Do not re-arm RTO if the sack RTT is measured from data sent
3156 * after when the head was last (re)transmitted. Otherwise the
3157 * timeout may continue to extend in loss recovery.
3158 */
Neal Cardwelldf92c832017-08-03 09:19:54 -04003159 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 }
3161
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003162 if (icsk->icsk_ca_ops->pkts_acked) {
3163 struct ack_sample sample = { .pkts_acked = pkts_acked,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003164 .rtt_us = sack->rate->rtt_us,
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003165 .in_flight = last_in_flight };
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003166
3167 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3168 }
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003169
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003171 WARN_ON((int)tp->sacked_out < 0);
3172 WARN_ON((int)tp->lost_out < 0);
3173 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003174 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003175 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003177 pr_debug("Leak l=%u %d\n",
3178 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 tp->lost_out = 0;
3180 }
3181 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003182 pr_debug("Leak s=%u %d\n",
3183 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 tp->sacked_out = 0;
3185 }
3186 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003187 pr_debug("Leak r=%u %d\n",
3188 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 tp->retrans_out = 0;
3190 }
3191 }
3192#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003193 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194}
3195
3196static void tcp_ack_probe(struct sock *sk)
3197{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003198 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07003199 struct sk_buff *head = tcp_send_head(sk);
3200 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
3202 /* Was it a usable window open? */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003203 if (!head)
3204 return;
3205 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003206 icsk->icsk_backoff = 0;
3207 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 /* Socket must be waked up by subsequent tcp_data_snd_check().
3209 * This function is not for random using!
3210 */
3211 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003212 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003213
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003214 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003215 when, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 }
3217}
3218
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003219static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003221 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3222 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223}
3224
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003225/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003226static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003228 /* If reordering is high then always grow cwnd whenever data is
3229 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003230 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003231 * new SACK or ECE mark may first advance cwnd here and later reduce
3232 * cwnd in tcp_fastretrans_alert() based on more states.
3233 */
Nikolay Borisov1043e252016-02-03 09:46:52 +02003234 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003235 return flag & FLAG_FORWARD_PROGRESS;
3236
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003237 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238}
3239
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003240/* The "ultimate" congestion control function that aims to replace the rigid
3241 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3242 * It's called toward the end of processing an ACK with precise rate
3243 * information. All transmission or retransmission are delayed afterwards.
3244 */
3245static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
Yuchung Chengc0402762016-09-19 23:39:21 -04003246 int flag, const struct rate_sample *rs)
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003247{
Yuchung Chengc0402762016-09-19 23:39:21 -04003248 const struct inet_connection_sock *icsk = inet_csk(sk);
3249
3250 if (icsk->icsk_ca_ops->cong_control) {
3251 icsk->icsk_ca_ops->cong_control(sk, rs);
3252 return;
3253 }
3254
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003255 if (tcp_in_cwnd_reduction(sk)) {
3256 /* Reduce cwnd if state mandates */
3257 tcp_cwnd_reduction(sk, acked_sacked, flag);
3258 } else if (tcp_may_raise_cwnd(sk, flag)) {
3259 /* Advance cwnd if state allows */
3260 tcp_cong_avoid(sk, ack, acked_sacked);
3261 }
3262 tcp_update_pacing_rate(sk);
3263}
3264
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265/* Check that window update is acceptable.
3266 * The function assumes that snd_una<=ack<=snd_next.
3267 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003268static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003269 const u32 ack, const u32 ack_seq,
3270 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003272 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003274 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275}
3276
Eric Dumazet0df48c22015-04-28 15:28:17 -07003277/* If we update tp->snd_una, also update tp->bytes_acked */
3278static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3279{
3280 u32 delta = ack - tp->snd_una;
3281
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003282 sock_owned_by_me((struct sock *)tp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003283 tp->bytes_acked += delta;
3284 tp->snd_una = ack;
3285}
3286
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003287/* If we update tp->rcv_nxt, also update tp->bytes_received */
3288static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3289{
3290 u32 delta = seq - tp->rcv_nxt;
3291
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003292 sock_owned_by_me((struct sock *)tp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003293 tp->bytes_received += delta;
3294 tp->rcv_nxt = seq;
3295}
3296
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297/* Update our send window.
3298 *
3299 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3300 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3301 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003302static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003303 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003305 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003307 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003309 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 nwin <<= tp->rx_opt.snd_wscale;
3311
3312 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3313 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003314 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315
3316 if (tp->snd_wnd != nwin) {
3317 tp->snd_wnd = nwin;
3318
Florian Westphal31770e32017-08-30 19:24:58 +02003319 /* Note, it is the only place, where
3320 * fast path is recovered for sending TCP.
3321 */
3322 tp->pred_flags = 0;
3323 tcp_fast_path_check(sk);
3324
Eric Dumazet75c119a2017-10-05 22:21:27 -07003325 if (!tcp_write_queue_empty(sk))
Eric Dumazet6f021c62015-08-21 12:30:00 -07003326 tcp_slow_start_after_idle_check(sk);
3327
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 if (nwin > tp->max_window) {
3329 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003330 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 }
3332 }
3333 }
3334
Eric Dumazet0df48c22015-04-28 15:28:17 -07003335 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336
3337 return flag;
3338}
3339
Jason Baron083ae302016-07-14 11:38:40 -04003340static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3341 u32 *last_oow_ack_time)
3342{
3343 if (*last_oow_ack_time) {
Eric Dumazet594208a2017-05-16 14:00:10 -07003344 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
Jason Baron083ae302016-07-14 11:38:40 -04003345
Eric Dumazet4170ba62017-10-27 07:47:30 -07003346 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
Jason Baron083ae302016-07-14 11:38:40 -04003347 NET_INC_STATS(net, mib_idx);
3348 return true; /* rate-limited: don't send yet! */
3349 }
3350 }
3351
Eric Dumazet594208a2017-05-16 14:00:10 -07003352 *last_oow_ack_time = tcp_jiffies32;
Jason Baron083ae302016-07-14 11:38:40 -04003353
3354 return false; /* not rate-limited: go ahead, send dupack now! */
3355}
3356
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003357/* Return true if we're currently rate-limiting out-of-window ACKs and
3358 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3359 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3360 * attacks that send repeated SYNs or ACKs for the same connection. To
3361 * do this, we do not send a duplicate SYNACK or ACK if the remote
3362 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3363 */
3364bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3365 int mib_idx, u32 *last_oow_ack_time)
3366{
3367 /* Data packets without SYNs are not likely part of an ACK loop. */
3368 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3369 !tcp_hdr(skb)->syn)
Jason Baron083ae302016-07-14 11:38:40 -04003370 return false;
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003371
Jason Baron083ae302016-07-14 11:38:40 -04003372 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003373}
3374
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003375/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003376static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003377{
3378 /* unprotected vars, we dont care of overwrites */
3379 static u32 challenge_timestamp;
3380 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003381 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetb530b682017-10-27 07:47:26 -07003382 struct net *net = sock_net(sk);
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003383 u32 count, now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003384
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003385 /* First check our per-socket dupack rate limit. */
Eric Dumazetb530b682017-10-27 07:47:26 -07003386 if (__tcp_oow_rate_limited(net,
Jason Baron083ae302016-07-14 11:38:40 -04003387 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3388 &tp->last_oow_ack_time))
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003389 return;
3390
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003391 /* Then check host-wide RFC 5961 rate limit. */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003392 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003393 if (now != challenge_timestamp) {
Eric Dumazetb530b682017-10-27 07:47:26 -07003394 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3395 u32 half = (ack_limit + 1) >> 1;
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003396
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003397 challenge_timestamp = now;
Eric Dumazetb530b682017-10-27 07:47:26 -07003398 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003399 }
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003400 count = READ_ONCE(challenge_count);
3401 if (count > 0) {
3402 WRITE_ONCE(challenge_count, count - 1);
Eric Dumazetb530b682017-10-27 07:47:26 -07003403 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003404 tcp_send_ack(sk);
3405 }
3406}
3407
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003408static void tcp_store_ts_recent(struct tcp_sock *tp)
3409{
3410 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3411 tp->rx_opt.ts_recent_stamp = get_seconds();
3412}
3413
3414static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3415{
3416 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3417 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3418 * extra check below makes sure this can only happen
3419 * for pure ACK frames. -DaveM
3420 *
3421 * Not only, also it occurs for expired timestamps.
3422 */
3423
3424 if (tcp_paws_check(&tp->rx_opt, 0))
3425 tcp_store_ts_recent(tp);
3426 }
3427}
3428
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003429/* This routine deals with acks during a TLP episode.
Sébastien Barré08abdff2015-01-12 10:30:40 +01003430 * We mark the end of a TLP episode on receiving TLP dupack or when
3431 * ack is after tlp_high_seq.
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003432 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3433 */
3434static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3435{
3436 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003437
Sébastien Barré08abdff2015-01-12 10:30:40 +01003438 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003439 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003440
Sébastien Barré08abdff2015-01-12 10:30:40 +01003441 if (flag & FLAG_DSACKING_ACK) {
3442 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003443 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003444 } else if (after(ack, tp->tlp_high_seq)) {
3445 /* ACK advances: there was a loss, so reduce cwnd. Reset
3446 * tlp_high_seq in tcp_init_cwnd_reduction()
3447 */
3448 tcp_init_cwnd_reduction(sk);
3449 tcp_set_ca_state(sk, TCP_CA_CWR);
3450 tcp_end_cwnd_reduction(sk);
3451 tcp_try_keep_open(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003452 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003453 LINUX_MIB_TCPLOSSPROBERECOVERY);
Sébastien Barré08abdff2015-01-12 10:30:40 +01003454 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3455 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3456 /* Pure dupack: original and TLP probe arrived; no loss */
3457 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003458 }
3459}
3460
Florian Westphal7354c8c2014-09-26 22:37:34 +02003461static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3462{
3463 const struct inet_connection_sock *icsk = inet_csk(sk);
3464
3465 if (icsk->icsk_ca_ops->in_ack_event)
3466 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3467}
3468
Yuchung Chenge662ca42016-02-02 10:33:04 -08003469/* Congestion control has updated the cwnd already. So if we're in
3470 * loss recovery then now we do any new sends (for FRTO) or
3471 * retransmits (for CA_Loss or CA_recovery) that make sense.
3472 */
3473static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3474{
3475 struct tcp_sock *tp = tcp_sk(sk);
3476
3477 if (rexmit == REXMIT_NONE)
3478 return;
3479
3480 if (unlikely(rexmit == 2)) {
3481 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3482 TCP_NAGLE_OFF);
3483 if (after(tp->snd_nxt, tp->high_seq))
3484 return;
3485 tp->frto = 0;
3486 }
3487 tcp_xmit_retransmit_queue(sk);
3488}
3489
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003491static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003493 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003495 struct tcp_sacktag_state sack_state;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003496 struct rate_sample rs = { .prior_delivered = 0 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 u32 prior_snd_una = tp->snd_una;
3498 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3499 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003500 bool is_dupack = false;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003501 int prior_packets = tp->packets_out;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003502 u32 delivered = tp->delivered;
3503 u32 lost = tp->lost;
Yuchung Chenge662ca42016-02-02 10:33:04 -08003504 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003505 u32 prior_fack;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003506
Eric Dumazet9a568de2017-05-16 14:00:14 -07003507 sack_state.first_sackt = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003508 sack_state.rate = &rs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509
Eric Dumazet75c119a2017-10-05 22:21:27 -07003510 /* We very likely will need to access rtx queue. */
3511 prefetch(sk->tcp_rtx_queue.rb_node);
Eric Dumazetad971f62014-10-11 15:17:29 -07003512
John Dykstra96e0bf42009-03-22 21:49:57 -07003513 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 * then we can probably ignore it.
3515 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003516 if (before(ack, prior_snd_una)) {
3517 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3518 if (before(ack, prior_snd_una - tp->max_window)) {
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07003519 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3520 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003521 return -1;
3522 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525
John Dykstra96e0bf42009-03-22 21:49:57 -07003526 /* If the ack includes data we haven't sent yet, discard
3527 * this segment (RFC793 Section 3.9).
3528 */
3529 if (after(ack, tp->snd_nxt))
3530 goto invalid_ack;
3531
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003532 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003533 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003534 icsk->icsk_retransmits = 0;
3535 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003536
Eric Dumazet50895b92017-11-14 21:02:19 -08003537 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003538 rs.prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003539
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003540 /* ts_recent update must be made after we are sure that the packet
3541 * is in window.
3542 */
3543 if (flag & FLAG_UPDATE_TS_RECENT)
3544 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3545
Florian Westphal31770e32017-08-30 19:24:58 +02003546 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3547 /* Window is constant, pure forward advance.
3548 * No more checks are required.
3549 * Note, we use the fact that SND.UNA>=SND.WL2.
3550 */
3551 tcp_update_wl(tp, ack_seq);
3552 tcp_snd_una_update(tp, ack);
3553 flag |= FLAG_WIN_UPDATE;
3554
3555 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3556
3557 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3558 } else {
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003559 u32 ack_ev_flags = CA_ACK_SLOWPATH;
Florian Westphal98900922014-09-26 22:37:35 +02003560
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003561 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3562 flag |= FLAG_DATA;
3563 else
3564 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003566 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003568 if (TCP_SKB_CB(skb)->sacked)
3569 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3570 &sack_state);
3571
3572 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3573 flag |= FLAG_ECE;
3574 ack_ev_flags |= CA_ACK_ECE;
3575 }
3576
3577 if (flag & FLAG_WIN_UPDATE)
3578 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3579
3580 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 }
3582
3583 /* We passed data and got it acked, remove any soft error
3584 * log. Something worked...
3585 */
3586 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003587 icsk->icsk_probes_out = 0;
Eric Dumazet70eabf02017-05-16 14:00:07 -07003588 tp->rcv_tstamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 if (!prior_packets)
3590 goto no_queue;
3591
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003593 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003594
Priyaranjan Jha1f255692017-11-03 16:38:48 -07003595 tcp_rack_update_reo_wnd(sk, &rs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596
Neal Cardwelldf92c832017-08-03 09:19:54 -04003597 if (tp->tlp_high_seq)
3598 tcp_process_tlp_ack(sk, ack, flag);
3599 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3600 if (flag & FLAG_SET_XMIT_TIMER)
3601 tcp_set_xmit_timer(sk);
3602
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003603 if (tcp_ack_is_dubious(sk, flag)) {
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003604 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Yuchung Cheng737ff312017-11-08 13:01:27 -08003605 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3606 &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003608
Julian Anastasovc3a2e832017-02-06 23:14:14 +02003609 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3610 sk_dst_confirm(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003611
Yuchung Chengb9f64822016-09-19 23:39:14 -04003612 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3613 lost = tp->lost - lost; /* freshly marked lost */
Eric Dumazet88d5c652017-04-25 10:15:37 -07003614 tcp_rate_gen(sk, delivered, lost, sack_state.rate);
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003615 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003616 tcp_xmit_recovery(sk, rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 return 1;
3618
3619no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003620 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3621 if (flag & FLAG_DSACKING_ACK)
Yuchung Cheng737ff312017-11-08 13:01:27 -08003622 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3623 &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 /* If this ack opens up a zero window, clear backoff. It was
3625 * being used to time the probes, and is probably far higher than
3626 * it needs to be for normal retransmission.
3627 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003628 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003629
3630 if (tp->tlp_high_seq)
3631 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 return 1;
3633
John Dykstra96e0bf42009-03-22 21:49:57 -07003634invalid_ack:
3635 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3636 return -1;
3637
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003639 /* If data was SACKed, tag it and see if we should send more data.
3640 * If data was DSACKed, see if we can undo a cwnd reduction.
3641 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003642 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003643 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003644 &sack_state);
Yuchung Cheng737ff312017-11-08 13:01:27 -08003645 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3646 &rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003647 tcp_xmit_recovery(sk, rexmit);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649
John Dykstra96e0bf42009-03-22 21:49:57 -07003650 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 return 0;
3652}
3653
Daniel Lee7f9b8382015-04-06 14:37:26 -07003654static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3655 bool syn, struct tcp_fastopen_cookie *foc,
3656 bool exp_opt)
3657{
3658 /* Valid only in SYN or SYN-ACK with an even length. */
3659 if (!foc || !syn || len < 0 || (len & 1))
3660 return;
3661
3662 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3663 len <= TCP_FASTOPEN_COOKIE_MAX)
3664 memcpy(foc->val, cookie, len);
3665 else if (len != 0)
3666 len = -1;
3667 foc->len = len;
3668 foc->exp = exp_opt;
3669}
3670
Ursula Braun60e2a772017-10-25 11:01:45 +02003671static void smc_parse_options(const struct tcphdr *th,
3672 struct tcp_options_received *opt_rx,
3673 const unsigned char *ptr,
3674 int opsize)
3675{
3676#if IS_ENABLED(CONFIG_SMC)
3677 if (static_branch_unlikely(&tcp_have_smc)) {
3678 if (th->syn && !(opsize & 1) &&
3679 opsize >= TCPOLEN_EXP_SMC_BASE &&
3680 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3681 opt_rx->smc_ok = 1;
3682 }
3683#endif
3684}
3685
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3687 * But, this can also be called on packets in the established flow when
3688 * the fast version below fails.
3689 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07003690void tcp_parse_options(const struct net *net,
3691 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003692 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003693 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003695 const unsigned char *ptr;
3696 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003697 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003699 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 opt_rx->saw_tstamp = 0;
3701
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003702 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003703 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704 int opsize;
3705
3706 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003707 case TCPOPT_EOL:
3708 return;
3709 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3710 length--;
3711 continue;
3712 default:
3713 opsize = *ptr++;
3714 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003716 if (opsize > length)
3717 return; /* don't parse partial options */
3718 switch (opcode) {
3719 case TCPOPT_MSS:
3720 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003721 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003722 if (in_mss) {
3723 if (opt_rx->user_mss &&
3724 opt_rx->user_mss < in_mss)
3725 in_mss = opt_rx->user_mss;
3726 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003728 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003729 break;
3730 case TCPOPT_WINDOW:
3731 if (opsize == TCPOLEN_WINDOW && th->syn &&
Eric Dumazet9bb37ef2017-06-07 10:34:38 -07003732 !estab && net->ipv4.sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003733 __u8 snd_wscale = *(__u8 *)ptr;
3734 opt_rx->wscale_ok = 1;
Gao Feng589c49c2017-04-04 21:09:48 +08003735 if (snd_wscale > TCP_MAX_WSCALE) {
3736 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
Joe Perchese87cc472012-05-13 21:56:26 +00003737 __func__,
Gao Feng589c49c2017-04-04 21:09:48 +08003738 snd_wscale,
3739 TCP_MAX_WSCALE);
3740 snd_wscale = TCP_MAX_WSCALE;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003741 }
3742 opt_rx->snd_wscale = snd_wscale;
3743 }
3744 break;
3745 case TCPOPT_TIMESTAMP:
3746 if ((opsize == TCPOLEN_TIMESTAMP) &&
3747 ((estab && opt_rx->tstamp_ok) ||
Eric Dumazet5d2ed052017-06-07 10:34:39 -07003748 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003749 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003750 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3751 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003752 }
3753 break;
3754 case TCPOPT_SACK_PERM:
3755 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
Eric Dumazetf9301032017-06-07 10:34:37 -07003756 !estab && net->ipv4.sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003757 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003758 tcp_sack_reset(opt_rx);
3759 }
3760 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003761
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003762 case TCPOPT_SACK:
3763 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3764 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3765 opt_rx->sack_ok) {
3766 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3767 }
3768 break;
3769#ifdef CONFIG_TCP_MD5SIG
3770 case TCPOPT_MD5SIG:
3771 /*
3772 * The MD5 Hash has already been
3773 * checked (see tcp_v{4,6}_do_rcv()).
3774 */
3775 break;
3776#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07003777 case TCPOPT_FASTOPEN:
3778 tcp_parse_fastopen_option(
3779 opsize - TCPOLEN_FASTOPEN_BASE,
3780 ptr, th->syn, foc, false);
3781 break;
3782
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003783 case TCPOPT_EXP:
3784 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07003785 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003786 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07003787 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3788 get_unaligned_be16(ptr) ==
3789 TCPOPT_FASTOPEN_MAGIC)
3790 tcp_parse_fastopen_option(opsize -
3791 TCPOLEN_EXP_FASTOPEN_BASE,
3792 ptr + 2, th->syn, foc, true);
Ursula Braun60e2a772017-10-25 11:01:45 +02003793 else
3794 smc_parse_options(th, opt_rx, ptr,
3795 opsize);
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003796 break;
3797
3798 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003799 ptr += opsize-2;
3800 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003801 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 }
3803}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003804EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003806static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003807{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003808 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003809
3810 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3811 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3812 tp->rx_opt.saw_tstamp = 1;
3813 ++ptr;
3814 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3815 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003816 if (*ptr)
3817 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3818 else
3819 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003820 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003821 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003822 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003823}
3824
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825/* Fast parse options. This hopes to only see timestamps.
3826 * If it is wrong it falls back on tcp_parse_options().
3827 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07003828static bool tcp_fast_parse_options(const struct net *net,
3829 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003830 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003832 /* In the spirit of fast parsing, compare doff directly to constant
3833 * values. Because equality is used, short doff can be ignored here.
3834 */
3835 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003837 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003839 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003840 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003841 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003843
Eric Dumazeteed29f12017-06-07 10:34:36 -07003844 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003845 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003846 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3847
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003848 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849}
3850
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003851#ifdef CONFIG_TCP_MD5SIG
3852/*
3853 * Parse MD5 Signature option
3854 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003855const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003856{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003857 int length = (th->doff << 2) - sizeof(*th);
3858 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003859
3860 /* If the TCP option is too short, we can short cut */
3861 if (length < TCPOLEN_MD5SIG)
3862 return NULL;
3863
3864 while (length > 0) {
3865 int opcode = *ptr++;
3866 int opsize;
3867
Weilong Chena22318e2013-12-23 14:37:26 +08003868 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003869 case TCPOPT_EOL:
3870 return NULL;
3871 case TCPOPT_NOP:
3872 length--;
3873 continue;
3874 default:
3875 opsize = *ptr++;
3876 if (opsize < 2 || opsize > length)
3877 return NULL;
3878 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003879 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003880 }
3881 ptr += opsize - 2;
3882 length -= opsize;
3883 }
3884 return NULL;
3885}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003886EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003887#endif
3888
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3890 *
3891 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3892 * it can pass through stack. So, the following predicate verifies that
3893 * this segment is not used for anything but congestion avoidance or
3894 * fast retransmit. Moreover, we even are able to eliminate most of such
3895 * second order effects, if we apply some small "replay" window (~RTO)
3896 * to timestamp space.
3897 *
3898 * All these measures still do not guarantee that we reject wrapped ACKs
3899 * on networks with high bandwidth, when sequence space is recycled fastly,
3900 * but it guarantees that such events will be very rare and do not affect
3901 * connection seriously. This doesn't look nice, but alas, PAWS is really
3902 * buggy extension.
3903 *
3904 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3905 * states that events when retransmit arrives after original data are rare.
3906 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3907 * the biggest problem on large power networks even with minor reordering.
3908 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3909 * up to bandwidth of 18Gigabit/sec. 8) ]
3910 */
3911
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003912static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003914 const struct tcp_sock *tp = tcp_sk(sk);
3915 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 u32 seq = TCP_SKB_CB(skb)->seq;
3917 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3918
3919 return (/* 1. Pure ACK with correct sequence number. */
3920 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3921
3922 /* 2. ... and duplicate ACK. */
3923 ack == tp->snd_una &&
3924
3925 /* 3. ... and does not update window. */
3926 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3927
3928 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003929 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930}
3931
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003932static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003933 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003935 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003936
3937 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3938 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939}
3940
3941/* Check segment sequence number for validity.
3942 *
3943 * Segment controls are considered valid, if the segment
3944 * fits to the window after truncation to the window. Acceptability
3945 * of data (and SYN, FIN, of course) is checked separately.
3946 * See tcp_data_queue(), for example.
3947 *
3948 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3949 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3950 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3951 * (borrowed from freebsd)
3952 */
3953
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003954static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955{
3956 return !before(end_seq, tp->rcv_wup) &&
3957 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3958}
3959
3960/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00003961void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962{
Song Liu59415212017-10-23 09:20:25 -07003963 trace_tcp_receive_reset(sk);
3964
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 /* We want the right error as BSD sees it (and indeed as we do). */
3966 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003967 case TCP_SYN_SENT:
3968 sk->sk_err = ECONNREFUSED;
3969 break;
3970 case TCP_CLOSE_WAIT:
3971 sk->sk_err = EPIPE;
3972 break;
3973 case TCP_CLOSE:
3974 return;
3975 default:
3976 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 }
Tom Marshalla4d25802010-09-20 15:42:05 -07003978 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3979 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980
Eric Dumazet3d476262017-04-18 09:45:51 -07003981 tcp_done(sk);
3982
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 if (!sock_flag(sk, SOCK_DEAD))
3984 sk->sk_error_report(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985}
3986
3987/*
3988 * Process the FIN bit. This now behaves as it is supposed to work
3989 * and the FIN takes effect when it is validly part of sequence
3990 * space. Not before when we get holes.
3991 *
3992 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3993 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3994 * TIME-WAIT)
3995 *
3996 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3997 * close and we go into CLOSING (and later onto TIME-WAIT)
3998 *
3999 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4000 */
Eric Dumazete3e17b72016-02-06 11:16:28 -08004001void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002{
4003 struct tcp_sock *tp = tcp_sk(sk);
4004
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004005 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
4007 sk->sk_shutdown |= RCV_SHUTDOWN;
4008 sock_set_flag(sk, SOCK_DONE);
4009
4010 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004011 case TCP_SYN_RECV:
4012 case TCP_ESTABLISHED:
4013 /* Move to CLOSE_WAIT */
4014 tcp_set_state(sk, TCP_CLOSE_WAIT);
Jon Maxwell2251ae42015-07-08 10:12:28 +10004015 inet_csk(sk)->icsk_ack.pingpong = 1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004016 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004018 case TCP_CLOSE_WAIT:
4019 case TCP_CLOSING:
4020 /* Received a retransmission of the FIN, do
4021 * nothing.
4022 */
4023 break;
4024 case TCP_LAST_ACK:
4025 /* RFC793: Remain in the LAST-ACK state. */
4026 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004028 case TCP_FIN_WAIT1:
4029 /* This case occurs when a simultaneous close
4030 * happens, we must ack the received FIN and
4031 * enter the CLOSING state.
4032 */
4033 tcp_send_ack(sk);
4034 tcp_set_state(sk, TCP_CLOSING);
4035 break;
4036 case TCP_FIN_WAIT2:
4037 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4038 tcp_send_ack(sk);
4039 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4040 break;
4041 default:
4042 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4043 * cases we should never reach this piece of code.
4044 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004045 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004046 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004047 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004048 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049
4050 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4051 * Probably, we should reset in this case. For now drop them.
4052 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004053 skb_rbtree_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004054 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004056 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057
4058 if (!sock_flag(sk, SOCK_DEAD)) {
4059 sk->sk_state_change(sk);
4060
4061 /* Do not send POLL_HUP for half duplex close. */
4062 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4063 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004064 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004066 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 }
4068}
4069
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004070static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004071 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072{
4073 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4074 if (before(seq, sp->start_seq))
4075 sp->start_seq = seq;
4076 if (after(end_seq, sp->end_seq))
4077 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004078 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004080 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081}
4082
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004083static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004085 struct tcp_sock *tp = tcp_sk(sk);
4086
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004087 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004088 int mib_idx;
4089
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004091 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004093 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4094
Eric Dumazetc10d9312016-04-29 14:16:47 -07004095 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096
4097 tp->rx_opt.dsack = 1;
4098 tp->duplicate_sack[0].start_seq = seq;
4099 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 }
4101}
4102
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004103static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004105 struct tcp_sock *tp = tcp_sk(sk);
4106
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004108 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109 else
4110 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4111}
4112
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004113static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114{
4115 struct tcp_sock *tp = tcp_sk(sk);
4116
4117 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4118 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004119 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004120 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004122 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4124
4125 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4126 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004127 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 }
4129 }
4130
4131 tcp_send_ack(sk);
4132}
4133
4134/* These routines update the SACK block as out-of-order packets arrive or
4135 * in-order packets close up the sequence space.
4136 */
4137static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4138{
4139 int this_sack;
4140 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004141 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142
4143 /* See if the recent change to the first SACK eats into
4144 * or hits the sequence space of other SACK blocks, if so coalesce.
4145 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004146 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4148 int i;
4149
4150 /* Zap SWALK, by moving every further SACK up by one slot.
4151 * Decrease num_sacks.
4152 */
4153 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004154 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4155 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 continue;
4157 }
4158 this_sack++, swalk++;
4159 }
4160}
4161
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4163{
4164 struct tcp_sock *tp = tcp_sk(sk);
4165 struct tcp_sack_block *sp = &tp->selective_acks[0];
4166 int cur_sacks = tp->rx_opt.num_sacks;
4167 int this_sack;
4168
4169 if (!cur_sacks)
4170 goto new_sack;
4171
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004172 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 if (tcp_sack_extend(sp, seq, end_seq)) {
4174 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004175 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004176 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177 if (cur_sacks > 1)
4178 tcp_sack_maybe_coalesce(tp);
4179 return;
4180 }
4181 }
4182
4183 /* Could not find an adjacent existing SACK, build a new one,
4184 * put it at the front, and shift everyone else down. We
4185 * always know there is at least one SACK present already here.
4186 *
4187 * If the sack array is full, forget about the last one.
4188 */
Adam Langley4389dde2008-07-19 00:07:02 -07004189 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 this_sack--;
4191 tp->rx_opt.num_sacks--;
4192 sp--;
4193 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004194 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004195 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
4197new_sack:
4198 /* Build the new head SACK, and we're done. */
4199 sp->start_seq = seq;
4200 sp->end_seq = end_seq;
4201 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202}
4203
4204/* RCV.NXT advances, some SACKs should be eaten. */
4205
4206static void tcp_sack_remove(struct tcp_sock *tp)
4207{
4208 struct tcp_sack_block *sp = &tp->selective_acks[0];
4209 int num_sacks = tp->rx_opt.num_sacks;
4210 int this_sack;
4211
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004213 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 return;
4216 }
4217
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004218 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219 /* Check if the start of the sack is covered by RCV.NXT. */
4220 if (!before(tp->rcv_nxt, sp->start_seq)) {
4221 int i;
4222
4223 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004224 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225
4226 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004227 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004228 tp->selective_acks[i-1] = tp->selective_acks[i];
4229 num_sacks--;
4230 continue;
4231 }
4232 this_sack++;
4233 sp++;
4234 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004235 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236}
4237
Eric Dumazet1402d362012-04-23 07:11:42 +00004238/**
4239 * tcp_try_coalesce - try to merge skb to prior one
4240 * @sk: socket
Mike Maloney98aaa912017-08-22 17:08:48 -04004241 * @dest: destination queue
Eric Dumazet1402d362012-04-23 07:11:42 +00004242 * @to: prior buffer
4243 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004244 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004245 *
4246 * Before queueing skb @from after @to, try to merge them
4247 * to reduce overall memory use and queue lengths, if cost is small.
4248 * Packets in ofo or receive queues can stay a long time.
4249 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004250 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004251 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004252static bool tcp_try_coalesce(struct sock *sk,
4253 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004254 struct sk_buff *from,
4255 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004256{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004257 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004258
Eric Dumazet329033f2012-04-27 00:38:33 +00004259 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004260
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004261 /* Its possible this segment overlaps with prior segment in queue */
4262 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4263 return false;
4264
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004265 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004266 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004267
Alexander Duyck34a802a2012-05-02 21:19:09 +00004268 atomic_add(delta, &sk->sk_rmem_alloc);
4269 sk_mem_charge(sk, delta);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004270 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004271 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4272 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004273 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Mike Maloney98aaa912017-08-22 17:08:48 -04004274
4275 if (TCP_SKB_CB(from)->has_rxtstamp) {
4276 TCP_SKB_CB(to)->has_rxtstamp = true;
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004277 to->tstamp = from->tstamp;
Mike Maloney98aaa912017-08-22 17:08:48 -04004278 }
4279
Alexander Duyck34a802a2012-05-02 21:19:09 +00004280 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004281}
4282
Eric Dumazet532182c2016-04-01 08:52:19 -07004283static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4284{
4285 sk_drops_add(sk, skb);
4286 __kfree_skb(skb);
4287}
4288
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289/* This one checks to see if we can put data from the
4290 * out_of_order queue into the receive_queue.
4291 */
4292static void tcp_ofo_queue(struct sock *sk)
4293{
4294 struct tcp_sock *tp = tcp_sk(sk);
4295 __u32 dsack_high = tp->rcv_nxt;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004296 bool fin, fragstolen, eaten;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004297 struct sk_buff *skb, *tail;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004298 struct rb_node *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004299
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004300 p = rb_first(&tp->out_of_order_queue);
4301 while (p) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004302 skb = rb_to_skb(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4304 break;
4305
4306 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4307 __u32 dsack = dsack_high;
4308 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4309 dsack_high = TCP_SKB_CB(skb)->end_seq;
4310 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4311 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004312 p = rb_next(p);
4313 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004315 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 SOCK_DEBUG(sk, "ofo packet was already received\n");
Eric Dumazet532182c2016-04-01 08:52:19 -07004317 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004318 continue;
4319 }
4320 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4321 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4322 TCP_SKB_CB(skb)->end_seq);
4323
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004324 tail = skb_peek_tail(&sk->sk_receive_queue);
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004325 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004326 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004327 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004328 if (!eaten)
4329 __skb_queue_tail(&sk->sk_receive_queue, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004330 else
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004331 kfree_skb_partial(skb, fragstolen);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004332
4333 if (unlikely(fin)) {
4334 tcp_fin(sk);
4335 /* tcp_fin() purges tp->out_of_order_queue,
4336 * so we must end this loop right now.
4337 */
4338 break;
4339 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340 }
4341}
4342
4343static bool tcp_prune_ofo_queue(struct sock *sk);
4344static int tcp_prune_queue(struct sock *sk);
4345
4346static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4347 unsigned int size)
4348{
4349 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4350 !sk_rmem_schedule(sk, skb, size)) {
4351
4352 if (tcp_prune_queue(sk) < 0)
4353 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004354
Eric Dumazet36a65032016-08-17 14:17:09 -07004355 while (!sk_rmem_schedule(sk, skb, size)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 if (!tcp_prune_ofo_queue(sk))
4357 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004358 }
4359 }
4360 return 0;
4361}
4362
Eric Dumazete86b29192012-03-18 11:06:44 +00004363static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4364{
4365 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004366 struct rb_node **p, *parent;
Eric Dumazete86b29192012-03-18 11:06:44 +00004367 struct sk_buff *skb1;
4368 u32 seq, end_seq;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004369 bool fragstolen;
Eric Dumazete86b29192012-03-18 11:06:44 +00004370
Florian Westphal735d3832014-09-29 13:08:30 +02004371 tcp_ecn_check_ce(tp, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004372
Mel Gormanc76562b2012-07-31 16:44:41 -07004373 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004374 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazet532182c2016-04-01 08:52:19 -07004375 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004376 return;
4377 }
4378
Florian Westphal31770e32017-08-30 19:24:58 +02004379 /* Disable header prediction. */
4380 tp->pred_flags = 0;
Eric Dumazete86b29192012-03-18 11:06:44 +00004381 inet_csk_schedule_ack(sk);
4382
Eric Dumazetc10d9312016-04-29 14:16:47 -07004383 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004384 seq = TCP_SKB_CB(skb)->seq;
4385 end_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004386 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004387 tp->rcv_nxt, seq, end_seq);
Eric Dumazete86b29192012-03-18 11:06:44 +00004388
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004389 p = &tp->out_of_order_queue.rb_node;
4390 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004391 /* Initial out of order segment, build 1 SACK. */
4392 if (tcp_is_sack(tp)) {
4393 tp->rx_opt.num_sacks = 1;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004394 tp->selective_acks[0].start_seq = seq;
4395 tp->selective_acks[0].end_seq = end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004396 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004397 rb_link_node(&skb->rbnode, NULL, p);
4398 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4399 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004400 goto end;
4401 }
4402
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004403 /* In the typical case, we are adding an skb to the end of the list.
4404 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4405 */
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004406 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
Mike Maloney98aaa912017-08-22 17:08:48 -04004407 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004408coalesce_done:
4409 tcp_grow_window(sk, skb);
4410 kfree_skb_partial(skb, fragstolen);
4411 skb = NULL;
4412 goto add_sack;
Eric Dumazete86b29192012-03-18 11:06:44 +00004413 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004414 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4415 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4416 parent = &tp->ooo_last_skb->rbnode;
4417 p = &parent->rb_right;
4418 goto insert;
Eric Dumazete86b29192012-03-18 11:06:44 +00004419 }
4420
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004421 /* Find place to insert this segment. Handle overlaps on the way. */
4422 parent = NULL;
4423 while (*p) {
4424 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004425 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004426 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4427 p = &parent->rb_left;
4428 continue;
Eric Dumazete86b29192012-03-18 11:06:44 +00004429 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004430 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4431 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4432 /* All the bits are present. Drop. */
4433 NET_INC_STATS(sock_net(sk),
4434 LINUX_MIB_TCPOFOMERGE);
4435 __kfree_skb(skb);
4436 skb = NULL;
4437 tcp_dsack_set(sk, seq, end_seq);
4438 goto add_sack;
4439 }
4440 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4441 /* Partial overlap. */
4442 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4443 } else {
4444 /* skb's seq == skb1's seq and skb covers skb1.
4445 * Replace skb1 with skb.
4446 */
4447 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4448 &tp->out_of_order_queue);
4449 tcp_dsack_extend(sk,
4450 TCP_SKB_CB(skb1)->seq,
4451 TCP_SKB_CB(skb1)->end_seq);
4452 NET_INC_STATS(sock_net(sk),
4453 LINUX_MIB_TCPOFOMERGE);
4454 __kfree_skb(skb1);
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004455 goto merge_right;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004456 }
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004457 } else if (tcp_try_coalesce(sk, skb1,
Mike Maloney98aaa912017-08-22 17:08:48 -04004458 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004459 goto coalesce_done;
4460 }
4461 p = &parent->rb_right;
Eric Dumazete86b29192012-03-18 11:06:44 +00004462 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004463insert:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004464 /* Insert segment into RB tree. */
4465 rb_link_node(&skb->rbnode, parent, p);
4466 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004467
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004468merge_right:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004469 /* Remove other segments covered by skb. */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004470 while ((skb1 = skb_rb_next(skb)) != NULL) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004471 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4472 break;
4473 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4474 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4475 end_seq);
4476 break;
4477 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004478 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004479 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4480 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004481 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004482 tcp_drop(sk, skb1);
Eric Dumazete86b29192012-03-18 11:06:44 +00004483 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004484 /* If there is no skb after us, we are the last_skb ! */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004485 if (!skb1)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004486 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004487
4488add_sack:
4489 if (tcp_is_sack(tp))
4490 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4491end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004492 if (skb) {
4493 tcp_grow_window(sk, skb);
Eric Dumazet60b1af32017-01-24 14:57:36 -08004494 skb_condense(skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004495 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004496 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004497}
4498
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004499static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004500 bool *fragstolen)
4501{
4502 int eaten;
4503 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4504
4505 __skb_pull(skb, hdrlen);
4506 eaten = (tail &&
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004507 tcp_try_coalesce(sk, tail,
Mike Maloney98aaa912017-08-22 17:08:48 -04004508 skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004509 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004510 if (!eaten) {
4511 __skb_queue_tail(&sk->sk_receive_queue, skb);
4512 skb_set_owner_r(skb, sk);
4513 }
4514 return eaten;
4515}
Eric Dumazete86b29192012-03-18 11:06:44 +00004516
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004517int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4518{
Eric Dumazetcb934712014-09-17 03:14:42 -07004519 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004520 int err = -ENOMEM;
4521 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004522 bool fragstolen;
4523
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004524 if (size == 0)
4525 return 0;
4526
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004527 if (size > PAGE_SIZE) {
4528 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4529
4530 data_len = npages << PAGE_SHIFT;
4531 size = data_len + (size & ~PAGE_MASK);
4532 }
4533 skb = alloc_skb_with_frags(size - data_len, data_len,
4534 PAGE_ALLOC_COSTLY_ORDER,
4535 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004536 if (!skb)
4537 goto err;
4538
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004539 skb_put(skb, size - data_len);
4540 skb->data_len = data_len;
4541 skb->len = size;
4542
Eric Dumazetcb934712014-09-17 03:14:42 -07004543 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
Mel Gormanc76562b2012-07-31 16:44:41 -07004544 goto err_free;
4545
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004546 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4547 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004548 goto err_free;
4549
4550 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4551 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4552 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4553
Eric Dumazetcb934712014-09-17 03:14:42 -07004554 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004555 WARN_ON_ONCE(fragstolen); /* should not happen */
4556 __kfree_skb(skb);
4557 }
4558 return size;
4559
4560err_free:
4561 kfree_skb(skb);
4562err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004563 return err;
4564
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004565}
4566
Linus Torvalds1da177e2005-04-16 15:20:36 -07004567static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4568{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004569 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004570 bool fragstolen;
4571 int eaten;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004572
Eric Dumazet532182c2016-04-01 08:52:19 -07004573 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4574 __kfree_skb(skb);
4575 return;
4576 }
Eric Dumazetf84af322010-04-28 15:31:51 -07004577 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004578 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004579
Florian Westphal735d3832014-09-29 13:08:30 +02004580 tcp_ecn_accept_cwr(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004581
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004582 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583
4584 /* Queue data for delivery to the user.
4585 * Packets in sequence go to the receive queue.
4586 * Out of sequence packets to the out_of_order_queue.
4587 */
4588 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4589 if (tcp_receive_window(tp) == 0)
4590 goto out_of_window;
4591
4592 /* Ok. In sequence. In window. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004593queue_and_out:
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004594 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4595 sk_forced_mem_schedule(sk, skb->truesize);
4596 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4597 goto drop;
4598
Florian Westphalc13ee2a2017-07-30 03:57:19 +02004599 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004600 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004601 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004602 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004603 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004604 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004605
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004606 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004607 tcp_ofo_queue(sk);
4608
4609 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4610 * gap in queue is filled.
4611 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004612 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004613 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614 }
4615
4616 if (tp->rx_opt.num_sacks)
4617 tcp_sack_remove(tp);
4618
Florian Westphal31770e32017-08-30 19:24:58 +02004619 tcp_fast_path_check(sk);
4620
Eric Dumazet923dd342012-05-02 07:55:58 +00004621 if (eaten > 0)
4622 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004623 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04004624 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625 return;
4626 }
4627
4628 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4629 /* A retransmit, 2nd most common case. Force an immediate ack. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004630 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004631 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004632
4633out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004634 tcp_enter_quickack_mode(sk);
4635 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004636drop:
Eric Dumazet532182c2016-04-01 08:52:19 -07004637 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004638 return;
4639 }
4640
4641 /* Out of window. F.e. zero window probe. */
4642 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4643 goto out_of_window;
4644
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004645 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004646
4647 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4648 /* Partial packet, seq < rcv_next < end_seq */
4649 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4650 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4651 TCP_SKB_CB(skb)->end_seq);
4652
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004653 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004654
Linus Torvalds1da177e2005-04-16 15:20:36 -07004655 /* If window is closed, drop tail of packet. But after
4656 * remembering D-SACK for its head made in previous line.
4657 */
4658 if (!tcp_receive_window(tp))
4659 goto out_of_window;
4660 goto queue_and_out;
4661 }
4662
Eric Dumazete86b29192012-03-18 11:06:44 +00004663 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004664}
4665
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004666static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004667{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004668 if (list)
4669 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
David S. Miller91521942009-05-28 21:35:47 -07004670
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004671 return skb_rb_next(skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004672}
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004673
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004674static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4675 struct sk_buff_head *list,
4676 struct rb_root *root)
4677{
4678 struct sk_buff *next = tcp_skb_next(skb, list);
4679
4680 if (list)
4681 __skb_unlink(skb, list);
4682 else
4683 rb_erase(&skb->rbnode, root);
4684
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004685 __kfree_skb(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004686 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004687
4688 return next;
4689}
4690
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004691/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07004692void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004693{
4694 struct rb_node **p = &root->rb_node;
4695 struct rb_node *parent = NULL;
4696 struct sk_buff *skb1;
4697
4698 while (*p) {
4699 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004700 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004701 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4702 p = &parent->rb_left;
4703 else
4704 p = &parent->rb_right;
4705 }
4706 rb_link_node(&skb->rbnode, parent, p);
4707 rb_insert_color(&skb->rbnode, root);
4708}
4709
Linus Torvalds1da177e2005-04-16 15:20:36 -07004710/* Collapse contiguous sequence of skbs head..tail with
4711 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004712 *
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004713 * If tail is NULL, this means until the end of the queue.
David S. Miller91521942009-05-28 21:35:47 -07004714 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004715 * Segments with FIN/SYN are not collapsed (only because this
4716 * simplifies code)
4717 */
4718static void
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004719tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4720 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004721{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004722 struct sk_buff *skb = head, *n;
4723 struct sk_buff_head tmp;
David S. Miller91521942009-05-28 21:35:47 -07004724 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004725
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004726 /* First, check that queue is collapsible and find
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004727 * the point where collapsing can be useful.
4728 */
David S. Miller91521942009-05-28 21:35:47 -07004729restart:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004730 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4731 n = tcp_skb_next(skb, list);
4732
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733 /* No new bits? It is possible on ofo queue. */
4734 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004735 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004736 if (!skb)
4737 break;
4738 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739 }
4740
4741 /* The first skb to collapse is:
4742 * - not SYN/FIN and
4743 * - bloated or contains data before "start" or
4744 * overlaps to the next one.
4745 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004746 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Eric Dumazet94f08932017-10-26 21:55:09 -07004747 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004748 before(TCP_SKB_CB(skb)->seq, start))) {
4749 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004750 break;
David S. Miller91521942009-05-28 21:35:47 -07004751 }
4752
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004753 if (n && n != tail &&
4754 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4755 end_of_skbs = false;
4756 break;
David S. Miller91521942009-05-28 21:35:47 -07004757 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758
4759 /* Decided to skip this, advance start seq. */
4760 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004762 if (end_of_skbs ||
4763 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004764 return;
4765
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004766 __skb_queue_head_init(&tmp);
4767
Linus Torvalds1da177e2005-04-16 15:20:36 -07004768 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004769 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004771
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004772 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004773 if (!nskb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004774 break;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004775
Linus Torvalds1da177e2005-04-16 15:20:36 -07004776 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4777 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004778 if (list)
4779 __skb_queue_before(list, skb, nskb);
4780 else
4781 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004782 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004783
4784 /* Copy data, releasing collapsed skbs. */
4785 while (copy > 0) {
4786 int offset = start - TCP_SKB_CB(skb)->seq;
4787 int size = TCP_SKB_CB(skb)->end_seq - start;
4788
Kris Katterjohn09a62662006-01-08 22:24:28 -08004789 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 if (size > 0) {
4791 size = min(copy, size);
4792 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4793 BUG();
4794 TCP_SKB_CB(nskb)->end_seq += size;
4795 copy -= size;
4796 start += size;
4797 }
4798 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004799 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004800 if (!skb ||
4801 skb == tail ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004802 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004803 goto end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004804 }
4805 }
4806 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004807end:
4808 skb_queue_walk_safe(&tmp, skb, n)
4809 tcp_rbtree_insert(root, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004810}
4811
4812/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4813 * and tcp_collapse() them until all the queue is collapsed.
4814 */
4815static void tcp_collapse_ofo_queue(struct sock *sk)
4816{
4817 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004818 struct sk_buff *skb, *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004819 u32 start, end;
4820
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004821 skb = skb_rb_first(&tp->out_of_order_queue);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004822new_range:
4823 if (!skb) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004824 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825 return;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004826 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004827 start = TCP_SKB_CB(skb)->seq;
4828 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004829
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004830 for (head = skb;;) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004831 skb = skb_rb_next(skb);
David S. Miller91521942009-05-28 21:35:47 -07004832
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004833 /* Range is terminated when we see a gap or when
4834 * we are at the queue end.
4835 */
David S. Miller91521942009-05-28 21:35:47 -07004836 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004837 after(TCP_SKB_CB(skb)->seq, end) ||
4838 before(TCP_SKB_CB(skb)->end_seq, start)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004839 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
David S. Miller8728b832005-08-09 19:25:21 -07004840 head, skb, start, end);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004841 goto new_range;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004842 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004843
4844 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4845 start = TCP_SKB_CB(skb)->seq;
4846 if (after(TCP_SKB_CB(skb)->end_seq, end))
4847 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004848 }
4849}
4850
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004851/*
Eric Dumazet36a65032016-08-17 14:17:09 -07004852 * Clean the out-of-order queue to make room.
4853 * We drop high sequences packets to :
4854 * 1) Let a chance for holes to be filled.
4855 * 2) not add too big latencies if thousands of packets sit there.
4856 * (But if application shrinks SO_RCVBUF, we could still end up
4857 * freeing whole queue here)
4858 *
4859 * Return true if queue has shrunk.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004860 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004861static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004862{
4863 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004864 struct rb_node *node, *prev;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004865
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004866 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Eric Dumazet36a65032016-08-17 14:17:09 -07004867 return false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004868
Eric Dumazet36a65032016-08-17 14:17:09 -07004869 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004870 node = &tp->ooo_last_skb->rbnode;
4871 do {
4872 prev = rb_prev(node);
4873 rb_erase(node, &tp->out_of_order_queue);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004874 tcp_drop(sk, rb_to_skb(node));
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004875 sk_mem_reclaim(sk);
Eric Dumazet36a65032016-08-17 14:17:09 -07004876 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4877 !tcp_under_memory_pressure(sk))
4878 break;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004879 node = prev;
4880 } while (node);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004881 tp->ooo_last_skb = rb_to_skb(prev);
Eric Dumazet36a65032016-08-17 14:17:09 -07004882
4883 /* Reset SACK state. A conforming SACK implementation will
4884 * do the same at a timeout based retransmit. When a connection
4885 * is in a sad state like this, we care only about integrity
4886 * of the connection not performance.
4887 */
4888 if (tp->rx_opt.sack_ok)
4889 tcp_sack_reset(&tp->rx_opt);
4890 return true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004891}
4892
Linus Torvalds1da177e2005-04-16 15:20:36 -07004893/* Reduce allocated memory if we can, trying to get
4894 * the socket within its memory limits again.
4895 *
4896 * Return less than zero if we should start dropping frames
4897 * until the socket owning process reads some of the data
4898 * to stabilize the situation.
4899 */
4900static int tcp_prune_queue(struct sock *sk)
4901{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004902 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004903
4904 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4905
Eric Dumazetc10d9312016-04-29 14:16:47 -07004906 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004907
4908 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004909 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004910 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004911 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4912
4913 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004914 if (!skb_queue_empty(&sk->sk_receive_queue))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004915 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
David S. Miller91521942009-05-28 21:35:47 -07004916 skb_peek(&sk->sk_receive_queue),
4917 NULL,
4918 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004919 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004920
4921 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4922 return 0;
4923
4924 /* Collapsing did not help, destructive actions follow.
4925 * This must not ever occur. */
4926
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004927 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004928
4929 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4930 return 0;
4931
4932 /* If we are really being abused, tell the caller to silently
4933 * drop receive data on the floor. It will get retransmitted
4934 * and hopefully then we'll have sufficient space.
4935 */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004936 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004937
4938 /* Massive buffer overcommit. */
Florian Westphal31770e32017-08-30 19:24:58 +02004939 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004940 return -1;
4941}
4942
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004943static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004944{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004945 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004946
David S. Miller0d9901d2005-07-05 15:21:10 -07004947 /* If the user specified a specific send buffer setting, do
4948 * not modify it.
4949 */
4950 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004951 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004952
4953 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004954 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004955 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004956
4957 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004958 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004959 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004960
4961 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05004962 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004963 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004964
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004965 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07004966}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967
4968/* When incoming ACK allowed to free some skb from write_queue,
4969 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4970 * on the exit from tcp input handler.
4971 *
4972 * PROBLEM: sndbuf expansion does not work well with largesend.
4973 */
4974static void tcp_new_space(struct sock *sk)
4975{
4976 struct tcp_sock *tp = tcp_sk(sk);
4977
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004978 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07004979 tcp_sndbuf_expand(sk);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07004980 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004981 }
4982
4983 sk->sk_write_space(sk);
4984}
4985
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004986static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004987{
4988 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4989 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
jbaron@akamai.com3c715122015-04-20 20:05:07 +00004990 /* pairs with tcp_poll() */
Jason Baron56d80622017-01-24 21:49:41 -05004991 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004992 if (sk->sk_socket &&
Francis Yanb0f71bd2016-11-27 23:07:16 -08004993 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004994 tcp_new_space(sk);
Francis Yanb0f71bd2016-11-27 23:07:16 -08004995 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4996 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
4997 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004998 }
4999}
5000
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005001static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005002{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005003 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005004 tcp_check_space(sk);
5005}
5006
5007/*
5008 * Check if sending an ack is needed.
5009 */
5010static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5011{
5012 struct tcp_sock *tp = tcp_sk(sk);
5013
5014 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005015 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005016 /* ... and right edge of window advances far enough.
5017 * (tcp_recvmsg() will send ACK otherwise). Or...
5018 */
Joe Perches9d4fb272009-11-23 10:41:23 -08005019 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005020 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005021 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005022 /* We have out of order data. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005023 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005024 /* Then ack it now */
5025 tcp_send_ack(sk);
5026 } else {
5027 /* Else, send delayed ack. */
5028 tcp_send_delayed_ack(sk);
5029 }
5030}
5031
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005032static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005033{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005034 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035 /* We sent a data segment already. */
5036 return;
5037 }
5038 __tcp_ack_snd_check(sk, 1);
5039}
5040
5041/*
5042 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005043 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005044 * moved inline now as tcp_urg is only called from one
5045 * place. We handle URGent data wrong. We have to - as
5046 * BSD still doesn't use the correction from RFC961.
5047 * For 1003.1g we should support a new option TCP_STDURG to permit
5048 * either form (or just set the sysctl tcp_stdurg).
5049 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005050
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005051static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005052{
5053 struct tcp_sock *tp = tcp_sk(sk);
5054 u32 ptr = ntohs(th->urg_ptr);
5055
Eric Dumazet3f4c7c62017-10-26 21:55:01 -07005056 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057 ptr--;
5058 ptr += ntohl(th->seq);
5059
5060 /* Ignore urgent data that we've already seen and read. */
5061 if (after(tp->copied_seq, ptr))
5062 return;
5063
5064 /* Do not replay urg ptr.
5065 *
5066 * NOTE: interesting situation not covered by specs.
5067 * Misbehaving sender may send urg ptr, pointing to segment,
5068 * which we already have in ofo queue. We are not able to fetch
5069 * such data and will stay in TCP_URG_NOTYET until will be eaten
5070 * by recvmsg(). Seems, we are not obliged to handle such wicked
5071 * situations. But it is worth to think about possibility of some
5072 * DoSes using some hypothetical application level deadlock.
5073 */
5074 if (before(ptr, tp->rcv_nxt))
5075 return;
5076
5077 /* Do we already have a newer (or duplicate) urgent pointer? */
5078 if (tp->urg_data && !after(ptr, tp->urg_seq))
5079 return;
5080
5081 /* Tell the world about our new urgent pointer. */
5082 sk_send_sigurg(sk);
5083
5084 /* We may be adding urgent data when the last byte read was
5085 * urgent. To do this requires some care. We cannot just ignore
5086 * tp->copied_seq since we would read the last urgent byte again
5087 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005088 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005089 *
5090 * NOTE. Double Dutch. Rendering to plain English: author of comment
5091 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5092 * and expect that both A and B disappear from stream. This is _wrong_.
5093 * Though this happens in BSD with high probability, this is occasional.
5094 * Any application relying on this is buggy. Note also, that fix "works"
5095 * only in this artificial test. Insert some normal data between A and B and we will
5096 * decline of BSD again. Verdict: it is better to remove to trap
5097 * buggy users.
5098 */
5099 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005100 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005101 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5102 tp->copied_seq++;
5103 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005104 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005105 __kfree_skb(skb);
5106 }
5107 }
5108
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005109 tp->urg_data = TCP_URG_NOTYET;
5110 tp->urg_seq = ptr;
Florian Westphal31770e32017-08-30 19:24:58 +02005111
5112 /* Disable header prediction. */
5113 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005114}
5115
5116/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005117static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005118{
5119 struct tcp_sock *tp = tcp_sk(sk);
5120
5121 /* Check if we get a new urgent pointer - normally not. */
5122 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005123 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124
5125 /* Do we wait for any urgent data? - normally not... */
5126 if (tp->urg_data == TCP_URG_NOTYET) {
5127 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5128 th->syn;
5129
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005130 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131 if (ptr < skb->len) {
5132 u8 tmp;
5133 if (skb_copy_bits(skb, ptr, &tmp, 1))
5134 BUG();
5135 tp->urg_data = TCP_URG_VALID | tmp;
5136 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005137 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005138 }
5139 }
5140}
5141
Jason Baron0e40f4c2017-01-17 13:37:19 -05005142/* Accept RST for rcv_nxt - 1 after a FIN.
5143 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5144 * FIN is sent followed by a RST packet. The RST is sent with the same
5145 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5146 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5147 * ACKs on the closed socket. In addition middleboxes can drop either the
5148 * challenge ACK or a subsequent RST.
5149 */
5150static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5151{
5152 struct tcp_sock *tp = tcp_sk(sk);
5153
5154 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5155 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5156 TCPF_CLOSING));
5157}
5158
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005159/* Does PAWS and seqno based validation of an incoming segment, flags will
5160 * play significant role here.
5161 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005162static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5163 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005164{
5165 struct tcp_sock *tp = tcp_sk(sk);
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005166 bool rst_seq_match = false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005167
5168 /* RFC1323: H1. Apply PAWS check first. */
Eric Dumazeteed29f12017-06-07 10:34:36 -07005169 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5170 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005171 tcp_paws_discard(sk, skb)) {
5172 if (!th->rst) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005173 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005174 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5175 LINUX_MIB_TCPACKSKIPPEDPAWS,
5176 &tp->last_oow_ack_time))
5177 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005178 goto discard;
5179 }
5180 /* Reset is accepted even if it did not pass PAWS. */
5181 }
5182
5183 /* Step 1: check sequence number */
5184 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5185 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5186 * (RST) segments are validated by checking their SEQ-fields."
5187 * And page 69: "If an incoming segment is not acceptable,
5188 * an acknowledgment should be sent in reply (unless the RST
5189 * bit is set, if so drop the segment and return)".
5190 */
Eric Dumazete3715892012-07-17 12:29:30 +00005191 if (!th->rst) {
5192 if (th->syn)
5193 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005194 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5195 LINUX_MIB_TCPACKSKIPPEDSEQ,
5196 &tp->last_oow_ack_time))
5197 tcp_send_dupack(sk, skb);
Jason Baron0e40f4c2017-01-17 13:37:19 -05005198 } else if (tcp_reset_check(sk, skb)) {
5199 tcp_reset(sk);
Eric Dumazete3715892012-07-17 12:29:30 +00005200 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005201 goto discard;
5202 }
5203
5204 /* Step 2: check RST bit */
5205 if (th->rst) {
Jason Baron0e40f4c2017-01-17 13:37:19 -05005206 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5207 * FIN and SACK too if available):
5208 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5209 * the right-most SACK block,
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005210 * then
Eric Dumazet282f23c2012-07-17 10:13:05 +02005211 * RESET the connection
5212 * else
5213 * Send a challenge ACK
5214 */
Jason Baron0e40f4c2017-01-17 13:37:19 -05005215 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5216 tcp_reset_check(sk, skb)) {
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005217 rst_seq_match = true;
5218 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5219 struct tcp_sack_block *sp = &tp->selective_acks[0];
5220 int max_sack = sp[0].end_seq;
5221 int this_sack;
5222
5223 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5224 ++this_sack) {
5225 max_sack = after(sp[this_sack].end_seq,
5226 max_sack) ?
5227 sp[this_sack].end_seq : max_sack;
5228 }
5229
5230 if (TCP_SKB_CB(skb)->seq == max_sack)
5231 rst_seq_match = true;
5232 }
5233
5234 if (rst_seq_match)
Eric Dumazet282f23c2012-07-17 10:13:05 +02005235 tcp_reset(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005236 else {
5237 /* Disable TFO if RST is out-of-order
5238 * and no data has been received
5239 * for current active TFO socket
5240 */
5241 if (tp->syn_fastopen && !tp->data_segs_in &&
5242 sk->sk_state == TCP_ESTABLISHED)
Wei Wang46c2fa32017-04-20 14:45:47 -07005243 tcp_fastopen_active_disable(sk);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005244 tcp_send_challenge_ack(sk, skb);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005245 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005246 goto discard;
5247 }
5248
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005249 /* step 3: check security and precedence [ignored] */
5250
Eric Dumazet0c246042012-07-17 01:41:30 +00005251 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005252 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005253 */
5254 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005255syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005256 if (syn_inerr)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005257 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5258 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005259 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005260 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005261 }
5262
Eric Dumazet0c246042012-07-17 01:41:30 +00005263 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005264
5265discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005266 tcp_drop(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005267 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005268}
5269
Linus Torvalds1da177e2005-04-16 15:20:36 -07005270/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005271 * TCP receive function for the ESTABLISHED state.
Florian Westphal31770e32017-08-30 19:24:58 +02005272 *
5273 * It is split into a fast path and a slow path. The fast path is
5274 * disabled when:
5275 * - A zero window was announced from us - zero window probing
5276 * is only handled properly in the slow path.
5277 * - Out of order segments arrived.
5278 * - Urgent data is expected.
5279 * - There is no buffer space left
5280 * - Unexpected TCP flags/window values/header lengths are received
5281 * (detected by checking the TCP header against pred_flags)
5282 * - Data is sent in both directions. Fast path only supports pure senders
5283 * or pure receivers (this means either the sequence number or the ack
5284 * value must stay constant)
5285 * - Unexpected TCP option.
5286 *
5287 * When these conditions are not satisfied it drops into a standard
5288 * receive procedure patterned after RFC793 to handle all cases.
5289 * The first three cases are guaranteed by proper pred_flags setting,
5290 * the rest is checked inline. Fast processing is turned on in
5291 * tcp_data_queue when everything is OK.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005292 */
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005293void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
Matvejchikov Ilyae42e24c2017-07-24 16:02:12 +04005294 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005295{
Matvejchikov Ilyae42e24c2017-07-24 16:02:12 +04005296 unsigned int len = skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297 struct tcp_sock *tp = tcp_sk(sk);
5298
Eric Dumazet9a568de2017-05-16 14:00:14 -07005299 tcp_mstamp_refresh(tp);
Ian Morris51456b22015-04-03 09:17:26 +01005300 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005301 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Florian Westphal31770e32017-08-30 19:24:58 +02005302 /*
5303 * Header prediction.
5304 * The code loosely follows the one in the famous
5305 * "30 instruction TCP receive" Van Jacobson mail.
5306 *
5307 * Van's trick is to deposit buffers into socket queue
5308 * on a device interrupt, to call tcp_recv function
5309 * on the receive process context and checksum and copy
5310 * the buffer to user space. smart...
5311 *
5312 * Our current scheme is not silly either but we take the
5313 * extra cost of the net_bh soft interrupt processing...
5314 * We do checksum and copy also but from device to kernel.
5315 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005316
5317 tp->rx_opt.saw_tstamp = 0;
5318
Florian Westphal31770e32017-08-30 19:24:58 +02005319 /* pred_flags is 0xS?10 << 16 + snd_wnd
5320 * if header_prediction is to be made
5321 * 'S' will always be tp->tcp_header_len >> 2
5322 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5323 * turn it off (when there are holes in the receive
5324 * space for instance)
5325 * PSH flag is ignored.
5326 */
5327
5328 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5329 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5330 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5331 int tcp_header_len = tp->tcp_header_len;
5332
5333 /* Timestamp header prediction: tcp_header_len
5334 * is automatically equal to th->doff*4 due to pred_flags
5335 * match.
5336 */
5337
5338 /* Check timestamp */
5339 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5340 /* No? Slow path! */
5341 if (!tcp_parse_aligned_timestamp(tp, th))
5342 goto slow_path;
5343
5344 /* If PAWS failed, check it more carefully in slow path */
5345 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5346 goto slow_path;
5347
5348 /* DO NOT update ts_recent here, if checksum fails
5349 * and timestamp was corrupted part, it will result
5350 * in a hung connection since we will drop all
5351 * future packets due to the PAWS test.
5352 */
5353 }
5354
5355 if (len <= tcp_header_len) {
5356 /* Bulk data transfer: sender */
5357 if (len == tcp_header_len) {
5358 /* Predicted packet is in window by definition.
5359 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5360 * Hence, check seq<=rcv_wup reduces to:
5361 */
5362 if (tcp_header_len ==
5363 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5364 tp->rcv_nxt == tp->rcv_wup)
5365 tcp_store_ts_recent(tp);
5366
5367 /* We know that such packets are checksummed
5368 * on entry.
5369 */
5370 tcp_ack(sk, skb, 0);
5371 __kfree_skb(skb);
5372 tcp_data_snd_check(sk);
5373 return;
5374 } else { /* Header too small */
5375 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5376 goto discard;
5377 }
5378 } else {
5379 int eaten = 0;
5380 bool fragstolen = false;
5381
5382 if (tcp_checksum_complete(skb))
5383 goto csum_error;
5384
5385 if ((int)skb->truesize > sk->sk_forward_alloc)
5386 goto step5;
5387
5388 /* Predicted packet is in window by definition.
5389 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5390 * Hence, check seq<=rcv_wup reduces to:
5391 */
5392 if (tcp_header_len ==
5393 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5394 tp->rcv_nxt == tp->rcv_wup)
5395 tcp_store_ts_recent(tp);
5396
5397 tcp_rcv_rtt_measure_ts(sk, skb);
5398
5399 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5400
5401 /* Bulk data transfer: receiver */
5402 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5403 &fragstolen);
5404
5405 tcp_event_data_recv(sk, skb);
5406
5407 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5408 /* Well, only one small jumplet in fast path... */
5409 tcp_ack(sk, skb, FLAG_DATA);
5410 tcp_data_snd_check(sk);
5411 if (!inet_csk_ack_scheduled(sk))
5412 goto no_ack;
5413 }
5414
5415 __tcp_ack_snd_check(sk, 0);
5416no_ack:
5417 if (eaten)
5418 kfree_skb_partial(skb, fragstolen);
5419 sk->sk_data_ready(sk);
5420 return;
5421 }
5422 }
5423
5424slow_path:
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005425 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005426 goto csum_error;
5427
Calvin Owens0c228e82014-11-20 15:09:53 -08005428 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005429 goto discard;
5430
Florian Westphal31770e32017-08-30 19:24:58 +02005431 /*
5432 * Standard slow path.
5433 */
5434
Eric Dumazet0c246042012-07-17 01:41:30 +00005435 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005436 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005437
Florian Westphal31770e32017-08-30 19:24:58 +02005438step5:
5439 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005440 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005441
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005442 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005443
5444 /* Process urgent data. */
5445 tcp_urg(sk, skb, th);
5446
5447 /* step 7: process the segment text */
5448 tcp_data_queue(sk, skb);
5449
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005450 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005451 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005452 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005453
5454csum_error:
Eric Dumazetc10d9312016-04-29 14:16:47 -07005455 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5456 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005457
5458discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005459 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005460}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005461EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005463void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5464{
5465 struct tcp_sock *tp = tcp_sk(sk);
5466 struct inet_connection_sock *icsk = inet_csk(sk);
5467
5468 tcp_set_state(sk, TCP_ESTABLISHED);
Eric Dumazet70eabf02017-05-16 14:00:07 -07005469 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005470
Ian Morris00db4122015-04-03 09:17:27 +01005471 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005472 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005473 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005474 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005475
Wei Wang27204aa2017-10-04 10:03:44 -07005476 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005477
5478 /* Prevent spurious tcp_cwnd_restart() on first data
5479 * packet.
5480 */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07005481 tp->lsndtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005482
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005483 if (sock_flag(sk, SOCK_KEEPOPEN))
5484 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Florian Westphal31770e32017-08-30 19:24:58 +02005485
5486 if (!tp->rx_opt.snd_wscale)
5487 __tcp_fast_path_on(tp, tp->snd_wnd);
5488 else
5489 tp->pred_flags = 0;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005490}
5491
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005492static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5493 struct tcp_fastopen_cookie *cookie)
5494{
5495 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07005496 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07005497 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5498 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005499
5500 if (mss == tp->rx_opt.user_mss) {
5501 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005502
5503 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5504 tcp_clear_options(&opt);
5505 opt.user_mss = opt.mss_clamp = 0;
Eric Dumazeteed29f12017-06-07 10:34:36 -07005506 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005507 mss = opt.mss_clamp;
5508 }
5509
Daniel Lee2646c832015-04-06 14:37:27 -07005510 if (!tp->syn_fastopen) {
5511 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005512 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07005513 } else if (tp->total_retrans) {
5514 /* SYN timed out and the SYN-ACK neither has a cookie nor
5515 * acknowledges data. Presumably the remote received only
5516 * the retransmitted (regular) SYNs: either the original
5517 * SYN-data or the corresponding SYN-ACK was dropped.
5518 */
5519 syn_drop = (cookie->len < 0 && data);
5520 } else if (cookie->len < 0 && !tp->syn_data) {
5521 /* We requested a cookie but didn't get it. If we did not use
5522 * the (old) exp opt format then try so next time (try_exp=1).
5523 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5524 */
5525 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5526 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005527
Daniel Lee2646c832015-04-06 14:37:27 -07005528 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005529
5530 if (data) { /* Retransmit unacked data in SYN */
Eric Dumazet75c119a2017-10-05 22:21:27 -07005531 skb_rbtree_walk_from(data) {
5532 if (__tcp_retransmit_skb(sk, data, 1))
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005533 break;
5534 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005535 tcp_rearm_rto(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005536 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005537 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005538 return true;
5539 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005540 tp->syn_data_acked = tp->syn_data;
Yuchung Chengf19c29e2014-03-03 12:31:36 -08005541 if (tp->syn_data_acked)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005542 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005543 LINUX_MIB_TCPFASTOPENACTIVE);
Eric Dumazet61d2bca2016-02-01 21:03:07 -08005544
5545 tcp_fastopen_add_skb(sk, synack);
5546
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005547 return false;
5548}
5549
Ursula Braun60e2a772017-10-25 11:01:45 +02005550static void smc_check_reset_syn(struct tcp_sock *tp)
5551{
5552#if IS_ENABLED(CONFIG_SMC)
5553 if (static_branch_unlikely(&tcp_have_smc)) {
5554 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5555 tp->syn_smc = 0;
5556 }
5557#endif
5558}
5559
Linus Torvalds1da177e2005-04-16 15:20:36 -07005560static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07005561 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005562{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005563 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005564 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005565 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005566 int saved_clamp = tp->rx_opt.mss_clamp;
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005567 bool fastopen_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005568
Eric Dumazeteed29f12017-06-07 10:34:36 -07005569 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005570 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005571 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005572
5573 if (th->ack) {
5574 /* rfc793:
5575 * "If the state is SYN-SENT then
5576 * first check the ACK bit
5577 * If the ACK bit is set
5578 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5579 * a reset (unless the RST bit is set, if so drop
5580 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005581 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005582 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5583 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584 goto reset_and_undo;
5585
5586 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5587 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
Eric Dumazet9a568de2017-05-16 14:00:14 -07005588 tcp_time_stamp(tp))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005589 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005590 LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005591 goto reset_and_undo;
5592 }
5593
5594 /* Now ACK is acceptable.
5595 *
5596 * "If the RST bit is set
5597 * If the ACK was acceptable then signal the user "error:
5598 * connection reset", drop the segment, enter CLOSED state,
5599 * delete TCB, and return."
5600 */
5601
5602 if (th->rst) {
5603 tcp_reset(sk);
5604 goto discard;
5605 }
5606
5607 /* rfc793:
5608 * "fifth, if neither of the SYN or RST bits is set then
5609 * drop the segment and return."
5610 *
5611 * See note below!
5612 * --ANK(990513)
5613 */
5614 if (!th->syn)
5615 goto discard_and_undo;
5616
5617 /* rfc793:
5618 * "If the SYN bit is on ...
5619 * are acceptable then ...
5620 * (our SYN has been ACKed), change the connection
5621 * state to ESTABLISHED..."
5622 */
5623
Florian Westphal735d3832014-09-29 13:08:30 +02005624 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005625
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005626 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Florian Westphal31770e32017-08-30 19:24:58 +02005627 tcp_ack(sk, skb, FLAG_SLOWPATH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005628
5629 /* Ok.. it's good. Set up sequence numbers and
5630 * move to established.
5631 */
5632 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5633 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5634
5635 /* RFC1323: The window in SYN & SYN/ACK segments is
5636 * never scaled.
5637 */
5638 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005639
5640 if (!tp->rx_opt.wscale_ok) {
5641 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5642 tp->window_clamp = min(tp->window_clamp, 65535U);
5643 }
5644
5645 if (tp->rx_opt.saw_tstamp) {
5646 tp->rx_opt.tstamp_ok = 1;
5647 tp->tcp_header_len =
5648 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5649 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5650 tcp_store_ts_recent(tp);
5651 } else {
5652 tp->tcp_header_len = sizeof(struct tcphdr);
5653 }
5654
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005655 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005656 tcp_initialize_rcv_mss(sk);
5657
5658 /* Remember, tcp_poll() does not lock socket!
5659 * Change state from SYN-SENT only after copied_seq
5660 * is initialized. */
5661 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005662
Ursula Braun60e2a772017-10-25 11:01:45 +02005663 smc_check_reset_syn(tp);
5664
Ralf Baechlee16aa202006-12-07 00:11:33 -08005665 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005666
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005667 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005668
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005669 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5670 tcp_rcv_fastopen_synack(sk, skb, &foc);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005671
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005672 if (!sock_flag(sk, SOCK_DEAD)) {
5673 sk->sk_state_change(sk);
5674 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5675 }
5676 if (fastopen_fail)
5677 return -1;
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005678 if (sk->sk_write_pending ||
5679 icsk->icsk_accept_queue.rskq_defer_accept ||
5680 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005681 /* Save one ACK. Data will be ready after
5682 * several ticks, if write_pending is set.
5683 *
5684 * It may be deleted, but with this feature tcpdumps
5685 * look so _wonderfully_ clever, that I was not able
5686 * to stand against the temptation 8) --ANK
5687 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005688 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005689 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005690 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5691 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005692
5693discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005694 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005695 return 0;
5696 } else {
5697 tcp_send_ack(sk);
5698 }
5699 return -1;
5700 }
5701
5702 /* No ACK in the segment */
5703
5704 if (th->rst) {
5705 /* rfc793:
5706 * "If the RST bit is set
5707 *
5708 * Otherwise (no ACK) drop the segment and return."
5709 */
5710
5711 goto discard_and_undo;
5712 }
5713
5714 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005715 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005716 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005717 goto discard_and_undo;
5718
5719 if (th->syn) {
5720 /* We see SYN without ACK. It is attempt of
5721 * simultaneous connect with crossed SYNs.
5722 * Particularly, it can be connect to self.
5723 */
5724 tcp_set_state(sk, TCP_SYN_RECV);
5725
5726 if (tp->rx_opt.saw_tstamp) {
5727 tp->rx_opt.tstamp_ok = 1;
5728 tcp_store_ts_recent(tp);
5729 tp->tcp_header_len =
5730 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5731 } else {
5732 tp->tcp_header_len = sizeof(struct tcphdr);
5733 }
5734
5735 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet142a2e72015-11-26 08:18:14 -08005736 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005737 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5738
5739 /* RFC1323: The window in SYN & SYN/ACK segments is
5740 * never scaled.
5741 */
5742 tp->snd_wnd = ntohs(th->window);
5743 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5744 tp->max_window = tp->snd_wnd;
5745
Florian Westphal735d3832014-09-29 13:08:30 +02005746 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005747
John Heffner5d424d52006-03-20 17:53:41 -08005748 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005749 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005750 tcp_initialize_rcv_mss(sk);
5751
Linus Torvalds1da177e2005-04-16 15:20:36 -07005752 tcp_send_synack(sk);
5753#if 0
5754 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005755 * There are no obstacles to make this (except that we must
5756 * either change tcp_recvmsg() to prevent it from returning data
5757 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005758 *
5759 * However, if we ignore data in ACKless segments sometimes,
5760 * we have no reasons to accept it sometimes.
5761 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5762 * is not flawless. So, discard packet for sanity.
5763 * Uncomment this return to process the data.
5764 */
5765 return -1;
5766#else
5767 goto discard;
5768#endif
5769 }
5770 /* "fifth, if neither of the SYN or RST bits is set then
5771 * drop the segment and return."
5772 */
5773
5774discard_and_undo:
5775 tcp_clear_options(&tp->rx_opt);
5776 tp->rx_opt.mss_clamp = saved_clamp;
5777 goto discard;
5778
5779reset_and_undo:
5780 tcp_clear_options(&tp->rx_opt);
5781 tp->rx_opt.mss_clamp = saved_clamp;
5782 return 1;
5783}
5784
Linus Torvalds1da177e2005-04-16 15:20:36 -07005785/*
5786 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005787 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005788 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5789 * address independent.
5790 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005791
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005792int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005793{
5794 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005795 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005796 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00005797 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005798 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005799 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005800
Linus Torvalds1da177e2005-04-16 15:20:36 -07005801 switch (sk->sk_state) {
5802 case TCP_CLOSE:
5803 goto discard;
5804
5805 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005806 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005807 return 1;
5808
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005809 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005810 goto discard;
5811
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005812 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005813 if (th->fin)
5814 goto discard;
Eric Dumazet449809a2017-03-01 08:39:49 -08005815 /* It is possible that we process SYN packets from backlog,
5816 * so we need to make sure to disable BH right there.
5817 */
5818 local_bh_disable();
5819 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5820 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005821
Eric Dumazet449809a2017-03-01 08:39:49 -08005822 if (!acceptable)
5823 return 1;
Eric Dumazet0aea76d2016-04-21 22:13:01 -07005824 consume_skb(skb);
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005825 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005826 }
5827 goto discard;
5828
5829 case TCP_SYN_SENT:
Eric Dumazet8804b272016-04-13 22:05:40 -07005830 tp->rx_opt.saw_tstamp = 0;
Eric Dumazet9a568de2017-05-16 14:00:14 -07005831 tcp_mstamp_refresh(tp);
Eric Dumazetbda07a62015-09-29 07:42:40 -07005832 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833 if (queued >= 0)
5834 return queued;
5835
5836 /* Do step6 onward by hand. */
5837 tcp_urg(sk, skb, th);
5838 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005839 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005840 return 0;
5841 }
5842
Eric Dumazet9a568de2017-05-16 14:00:14 -07005843 tcp_mstamp_refresh(tp);
Eric Dumazet8804b272016-04-13 22:05:40 -07005844 tp->rx_opt.saw_tstamp = 0;
Jerry Chu168a8f52012-08-31 12:29:13 +00005845 req = tp->fastopen_rsk;
Ian Morris00db4122015-04-03 09:17:27 +01005846 if (req) {
Jerry Chu37561f62012-10-22 11:26:36 +00005847 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005848 sk->sk_state != TCP_FIN_WAIT1);
5849
Ian Morris51456b22015-04-03 09:17:26 +01005850 if (!tcp_check_req(sk, skb, req, true))
Jerry Chu168a8f52012-08-31 12:29:13 +00005851 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005852 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005853
Calvin Owens0c228e82014-11-20 15:09:53 -08005854 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005855 goto discard;
5856
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005857 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005858 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005859
5860 /* step 5: check the ACK field */
Florian Westphal31770e32017-08-30 19:24:58 +02005861 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5862 FLAG_UPDATE_TS_RECENT |
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07005863 FLAG_NO_CHALLENGE_ACK) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005864
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07005865 if (!acceptable) {
5866 if (sk->sk_state == TCP_SYN_RECV)
5867 return 1; /* send one RST */
5868 tcp_send_challenge_ack(sk, skb);
5869 goto discard;
5870 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005871 switch (sk->sk_state) {
5872 case TCP_SYN_RECV:
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07005873 if (!tp->srtt_us)
5874 tcp_synack_rtt_meas(sk, req);
5875
Joe Perches61eb90032013-05-24 18:36:13 +00005876 /* Once we leave TCP_SYN_RECV, we no longer need req
5877 * so release it.
5878 */
5879 if (req) {
Yuchung Cheng7e32b442016-09-21 16:16:15 -07005880 inet_csk(sk)->icsk_retransmits = 0;
Joe Perches61eb90032013-05-24 18:36:13 +00005881 reqsk_fastopen_remove(sk, req, false);
Wei Wang6d050812017-10-04 10:04:04 -07005882 /* Re-arm the timer because data may have been sent out.
5883 * This is similar to the regular data transmission case
5884 * when new data has just been ack'ed.
5885 *
5886 * (TFO) - we could try to be more aggressive and
5887 * retransmitting any data sooner based on when they
5888 * are sent out.
5889 */
5890 tcp_rearm_rto(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00005891 } else {
Wei Wang27204aa2017-10-04 10:03:44 -07005892 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
Joe Perches61eb90032013-05-24 18:36:13 +00005893 tp->copied_seq = tp->rcv_nxt;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005894 }
Joe Perches61eb90032013-05-24 18:36:13 +00005895 smp_mb();
5896 tcp_set_state(sk, TCP_ESTABLISHED);
5897 sk->sk_state_change(sk);
5898
5899 /* Note, that this wakeup is only for marginal crossed SYN case.
5900 * Passively open sockets are not waked up, because
5901 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5902 */
5903 if (sk->sk_socket)
5904 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5905
5906 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5907 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5908 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5909
5910 if (tp->rx_opt.tstamp_ok)
5911 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5912
Yuchung Chengc0402762016-09-19 23:39:21 -04005913 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
5914 tcp_update_pacing_rate(sk);
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04005915
Joe Perches61eb90032013-05-24 18:36:13 +00005916 /* Prevent spurious tcp_cwnd_restart() on first data packet */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07005917 tp->lsndtime = tcp_jiffies32;
Joe Perches61eb90032013-05-24 18:36:13 +00005918
5919 tcp_initialize_rcv_mss(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02005920 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005921 break;
5922
Joe Perchesc48b22d2013-05-24 18:06:58 +00005923 case TCP_FIN_WAIT1: {
Joe Perchesc48b22d2013-05-24 18:06:58 +00005924 int tmo;
5925
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005926 /* If we enter the TCP_FIN_WAIT1 state and we are a
5927 * Fast Open socket and this is the first acceptable
5928 * ACK we have received, this would have acknowledged
5929 * our SYNACK so stop the SYNACK timer.
5930 */
Ian Morris00db4122015-04-03 09:17:27 +01005931 if (req) {
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005932 /* We no longer need the request sock. */
5933 reqsk_fastopen_remove(sk, req, false);
5934 tcp_rearm_rto(sk);
5935 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00005936 if (tp->snd_una != tp->write_seq)
5937 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005938
Joe Perchesc48b22d2013-05-24 18:06:58 +00005939 tcp_set_state(sk, TCP_FIN_WAIT2);
5940 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005941
Julian Anastasovc3a2e832017-02-06 23:14:14 +02005942 sk_dst_confirm(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005943
Joe Perchesc48b22d2013-05-24 18:06:58 +00005944 if (!sock_flag(sk, SOCK_DEAD)) {
5945 /* Wake up lingering close() */
5946 sk->sk_state_change(sk);
5947 break;
5948 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005949
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005950 if (tp->linger2 < 0) {
5951 tcp_done(sk);
5952 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5953 return 1;
5954 }
5955 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5956 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5957 /* Receive out of order FIN after close() */
5958 if (tp->syn_fastopen && th->fin)
Wei Wang46c2fa32017-04-20 14:45:47 -07005959 tcp_fastopen_active_disable(sk);
Joe Perchesc48b22d2013-05-24 18:06:58 +00005960 tcp_done(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005961 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Joe Perchesc48b22d2013-05-24 18:06:58 +00005962 return 1;
5963 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005964
Joe Perchesc48b22d2013-05-24 18:06:58 +00005965 tmo = tcp_fin_time(sk);
5966 if (tmo > TCP_TIMEWAIT_LEN) {
5967 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5968 } else if (th->fin || sock_owned_by_user(sk)) {
5969 /* Bad case. We could lose such FIN otherwise.
5970 * It is not a big problem, but it looks confusing
5971 * and not so rare event. We still can lose it now,
5972 * if it spins in bh_lock_sock(), but it is really
5973 * marginal case.
5974 */
5975 inet_csk_reset_keepalive_timer(sk, tmo);
5976 } else {
5977 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5978 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005979 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005980 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00005981 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005982
5983 case TCP_CLOSING:
5984 if (tp->snd_una == tp->write_seq) {
5985 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5986 goto discard;
5987 }
5988 break;
5989
5990 case TCP_LAST_ACK:
5991 if (tp->snd_una == tp->write_seq) {
5992 tcp_update_metrics(sk);
5993 tcp_done(sk);
5994 goto discard;
5995 }
5996 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005997 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005998
5999 /* step 6: check the URG bit */
6000 tcp_urg(sk, skb, th);
6001
6002 /* step 7: process the segment text */
6003 switch (sk->sk_state) {
6004 case TCP_CLOSE_WAIT:
6005 case TCP_CLOSING:
6006 case TCP_LAST_ACK:
6007 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6008 break;
Gustavo A. R. Silvafcfd6df2017-10-16 15:48:55 -05006009 /* fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006010 case TCP_FIN_WAIT1:
6011 case TCP_FIN_WAIT2:
6012 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006013 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006014 * BSD 4.4 also does reset.
6015 */
6016 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6017 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6018 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006019 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006020 tcp_reset(sk);
6021 return 1;
6022 }
6023 }
6024 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006025 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006026 tcp_data_queue(sk, skb);
6027 queued = 1;
6028 break;
6029 }
6030
6031 /* tcp_data could move socket to TIME-WAIT */
6032 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006033 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006034 tcp_ack_snd_check(sk);
6035 }
6036
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006037 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006038discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006039 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006040 }
6041 return 0;
6042}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006043EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006044
6045static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6046{
6047 struct inet_request_sock *ireq = inet_rsk(req);
6048
6049 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006050 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6051 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006052#if IS_ENABLED(CONFIG_IPV6)
6053 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006054 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6055 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006056#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006057}
6058
Florian Westphald82bd122014-09-29 13:08:29 +02006059/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6060 *
6061 * If we receive a SYN packet with these bits set, it means a
6062 * network is playing bad games with TOS bits. In order to
6063 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006064 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006065 *
6066 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006067 * congestion control: Linux DCTCP asserts ECT on all packets,
6068 * including SYN, which is most optimal solution; however,
6069 * others, such as FreeBSD do not.
Florian Westphald82bd122014-09-29 13:08:29 +02006070 */
6071static void tcp_ecn_create_request(struct request_sock *req,
6072 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006073 const struct sock *listen_sk,
6074 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006075{
6076 const struct tcphdr *th = tcp_hdr(skb);
6077 const struct net *net = sock_net(listen_sk);
6078 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006079 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006080 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006081
6082 if (!th_ecn)
6083 return;
6084
6085 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006086 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6087 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006088
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006089 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
Lawrence Brakmo91b5b212017-06-30 20:02:49 -07006090 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6091 tcp_bpf_ca_needs_ecn((struct sock *)req))
Florian Westphald82bd122014-09-29 13:08:29 +02006092 inet_rsk(req)->ecn_ok = 1;
6093}
6094
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006095static void tcp_openreq_init(struct request_sock *req,
6096 const struct tcp_options_received *rx_opt,
6097 struct sk_buff *skb, const struct sock *sk)
6098{
6099 struct inet_request_sock *ireq = inet_rsk(req);
6100
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006101 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006102 req->cookie_ts = 0;
6103 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6104 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet9a568de2017-05-16 14:00:14 -07006105 tcp_rsk(req)->snt_synack = tcp_clock_us();
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006106 tcp_rsk(req)->last_oow_ack_time = 0;
6107 req->mss = rx_opt->mss_clamp;
6108 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6109 ireq->tstamp_ok = rx_opt->tstamp_ok;
6110 ireq->sack_ok = rx_opt->sack_ok;
6111 ireq->snd_wscale = rx_opt->snd_wscale;
6112 ireq->wscale_ok = rx_opt->wscale_ok;
6113 ireq->acked = 0;
6114 ireq->ecn_ok = 0;
6115 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6116 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6117 ireq->ir_mark = inet_request_mark(sk, skb);
Ursula Braun60e2a772017-10-25 11:01:45 +02006118#if IS_ENABLED(CONFIG_SMC)
6119 ireq->smc_ok = rx_opt->smc_ok;
6120#endif
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006121}
6122
Eric Dumazete49bb332015-03-17 18:32:27 -07006123struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006124 struct sock *sk_listener,
6125 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006126{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006127 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6128 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006129
6130 if (req) {
6131 struct inet_request_sock *ireq = inet_rsk(req);
6132
Eric Dumazetc92e8c02017-10-20 09:04:13 -07006133 ireq->ireq_opt = NULL;
Huw Davies56ac42b2016-06-27 15:05:28 -04006134#if IS_ENABLED(CONFIG_IPV6)
6135 ireq->pktopts = NULL;
6136#endif
Eric Dumazete49bb332015-03-17 18:32:27 -07006137 atomic64_set(&ireq->ir_cookie, 0);
6138 ireq->ireq_state = TCP_NEW_SYN_RECV;
6139 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006140 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006141 }
6142
6143 return req;
6144}
6145EXPORT_SYMBOL(inet_reqsk_alloc);
6146
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006147/*
6148 * Return true if a syncookie should be sent
6149 */
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006150static bool tcp_syn_flood_action(const struct sock *sk,
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006151 const struct sk_buff *skb,
6152 const char *proto)
6153{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006154 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006155 const char *msg = "Dropping request";
6156 bool want_cookie = false;
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006157 struct net *net = sock_net(sk);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006158
6159#ifdef CONFIG_SYN_COOKIES
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006160 if (net->ipv4.sysctl_tcp_syncookies) {
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006161 msg = "Sending cookies";
6162 want_cookie = true;
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006163 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006164 } else
6165#endif
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006166 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006167
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006168 if (!queue->synflood_warned &&
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006169 net->ipv4.sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006170 xchg(&queue->synflood_warned, 1) == 0)
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006171 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6172 proto, ntohs(tcp_hdr(skb)->dest), msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006173
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006174 return want_cookie;
6175}
6176
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006177static void tcp_reqsk_record_syn(const struct sock *sk,
6178 struct request_sock *req,
6179 const struct sk_buff *skb)
6180{
6181 if (tcp_sk(sk)->save_syn) {
6182 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6183 u32 *copy;
6184
6185 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6186 if (copy) {
6187 copy[0] = len;
6188 memcpy(&copy[1], skb_network_header(skb), len);
6189 req->saved_syn = copy;
6190 }
6191 }
6192}
6193
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006194int tcp_conn_request(struct request_sock_ops *rsk_ops,
6195 const struct tcp_request_sock_ops *af_ops,
6196 struct sock *sk, struct sk_buff *skb)
6197{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006198 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006199 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6200 struct tcp_options_received tmp_opt;
6201 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006202 struct net *net = sock_net(sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006203 struct sock *fastopen_sk = NULL;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006204 struct request_sock *req;
6205 bool want_cookie = false;
Eric Dumazeteaa72dc2017-08-29 15:16:01 -07006206 struct dst_entry *dst;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006207 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006208
6209 /* TW buckets are converted to open requests without
6210 * limitations, they conserve resources and peer is
6211 * evidently real one.
6212 */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006213 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006214 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6215 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6216 if (!want_cookie)
6217 goto drop;
6218 }
6219
Eric Dumazet5ea8ea22016-10-26 09:27:57 -07006220 if (sk_acceptq_is_full(sk)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006221 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006222 goto drop;
6223 }
6224
Eric Dumazeta1a53442015-10-04 21:08:11 -07006225 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006226 if (!req)
6227 goto drop;
6228
6229 tcp_rsk(req)->af_specific = af_ops;
Florian Westphal95a22ca2016-12-01 11:32:06 +01006230 tcp_rsk(req)->ts_off = 0;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006231
6232 tcp_clear_options(&tmp_opt);
6233 tmp_opt.mss_clamp = af_ops->mss_clamp;
6234 tmp_opt.user_mss = tp->rx_opt.user_mss;
Eric Dumazeteed29f12017-06-07 10:34:36 -07006235 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6236 want_cookie ? NULL : &foc);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006237
6238 if (want_cookie && !tmp_opt.saw_tstamp)
6239 tcp_clear_options(&tmp_opt);
6240
6241 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6242 tcp_openreq_init(req, &tmp_opt, skb, sk);
KOVACS Krisztian7a682572016-09-23 11:27:42 +02006243 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006244
Eric Dumazet16f86162015-03-13 15:51:10 -07006245 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
David Ahern6dd9a142015-12-16 13:20:44 -08006246 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
Eric Dumazet16f86162015-03-13 15:51:10 -07006247
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006248 af_ops->init_req(req, sk, skb);
6249
6250 if (security_inet_conn_request(sk, skb, req))
6251 goto drop_and_free;
6252
Eric Dumazet84b114b2017-05-05 06:56:54 -07006253 if (tmp_opt.tstamp_ok)
Eric Dumazet5d2ed052017-06-07 10:34:39 -07006254 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
Florian Westphal95a22ca2016-12-01 11:32:06 +01006255
Tonghao Zhang49c71582017-08-21 23:33:48 -07006256 dst = af_ops->route_req(sk, &fl, req);
6257 if (!dst)
6258 goto drop_and_free;
6259
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006260 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006261 /* Kill the following clause, if you dislike this way. */
Soheil Hassas Yeganeh4396e462017-03-15 16:30:46 -04006262 if (!net->ipv4.sysctl_tcp_syncookies &&
6263 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6264 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6265 !tcp_peer_is_proven(req, dst)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006266 /* Without syncookies last quarter of
6267 * backlog is filled with destinations,
6268 * proven to be alive.
6269 * It means that we continue to communicate
6270 * to destinations, already remembered
6271 * to the moment of synflood.
6272 */
6273 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6274 rsk_ops->family);
6275 goto drop_and_release;
6276 }
6277
Eric Dumazet84b114b2017-05-05 06:56:54 -07006278 isn = af_ops->init_seq(skb);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006279 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006280
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006281 tcp_ecn_create_request(req, skb, sk, dst);
6282
6283 if (want_cookie) {
6284 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6285 req->cookie_ts = tmp_opt.tstamp_ok;
6286 if (!tmp_opt.tstamp_ok)
6287 inet_rsk(req)->ecn_ok = 0;
6288 }
6289
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006290 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006291 tcp_rsk(req)->txhash = net_tx_rndhash();
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006292 tcp_openreq_init_rwin(req, sk, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006293 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006294 tcp_reqsk_record_syn(sk, req, skb);
Christoph Paasch71c02372017-10-23 13:22:23 -07006295 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006296 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006297 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006298 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Eric Dumazetb3d05142016-04-13 22:05:39 -07006299 &foc, TCP_SYNACK_FASTOPEN);
Eric Dumazet7656d842015-10-04 21:08:07 -07006300 /* Add the child socket directly into the accept queue */
6301 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6302 sk->sk_data_ready(sk);
6303 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006304 sock_put(fastopen_sk);
6305 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006306 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006307 if (!want_cookie)
Lawrence Brakmo8550f322017-06-30 20:02:42 -07006308 inet_csk_reqsk_queue_hash_add(sk, req,
6309 tcp_timeout_init((struct sock *)req));
Eric Dumazetb3d05142016-04-13 22:05:39 -07006310 af_ops->send_synack(sk, dst, &fl, req, &foc,
6311 !want_cookie ? TCP_SYNACK_NORMAL :
6312 TCP_SYNACK_COOKIE);
Eric Dumazet9caad862016-04-01 08:52:20 -07006313 if (want_cookie) {
6314 reqsk_free(req);
6315 return 0;
6316 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006317 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006318 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006319 return 0;
6320
6321drop_and_release:
6322 dst_release(dst);
6323drop_and_free:
6324 reqsk_free(req);
6325drop:
Eric Dumazet9caad862016-04-01 08:52:20 -07006326 tcp_listendrop(sk);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006327 return 0;
6328}
6329EXPORT_SYMBOL(tcp_conn_request);