blob: 5def3c48870e17f42ac9424a6ee091ac4824dabc [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>
Amritha Nambiarc6345ce2018-06-29 21:26:57 -070081#include <net/busy_poll.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070082
Brian Haleyab32ea52006-09-22 14:15:41 -070083int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Linus Torvalds1da177e2005-04-16 15:20:36 -070084
Linus Torvalds1da177e2005-04-16 15:20:36 -070085#define FLAG_DATA 0x01 /* Incoming frame contained data. */
86#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90#define FLAG_DATA_SACKED 0x20 /* New SACK. */
91#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -070092#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Florian Westphal31770e32017-08-30 19:24:58 +020093#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +000094#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -070095#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -070096#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Neal Cardwelldf92c832017-08-03 09:19:54 -040097#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
Ilpo Järvinencadbd032007-12-31 04:49:21 -080098#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +000099#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -0700100#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
Yuchung Chengeb36be02018-01-17 12:11:00 -0800101#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
103#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
Priyaranjan Jhad09b9e62017-11-03 17:46:55 -0700105#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700109#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Yuchung Chenge662ca42016-02-02 10:33:04 -0800111#define REXMIT_NONE 0 /* no loss recovery to do */
112#define REXMIT_LOST 1 /* retransmit packets marked lost */
113#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
114
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300115#if IS_ENABLED(CONFIG_TLS_DEVICE)
116static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
117
118void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
120{
121 icsk->icsk_clean_acked = cad;
122 static_branch_inc(&clean_acked_data_enabled);
123}
124EXPORT_SYMBOL_GPL(clean_acked_data_enable);
125
126void clean_acked_data_disable(struct inet_connection_sock *icsk)
127{
128 static_branch_dec(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
130}
131EXPORT_SYMBOL_GPL(clean_acked_data_disable);
132#endif
133
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300134static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
135 unsigned int len)
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200136{
137 static bool __once __read_mostly;
138
139 if (!__once) {
140 struct net_device *dev;
141
142 __once = true;
143
144 rcu_read_lock();
145 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300146 if (!dev || len >= dev->mtu)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev ? dev->name : "Unknown driver");
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200149 rcu_read_unlock();
150 }
151}
152
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900153/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900155 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800156static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700158 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900159 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900162 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
166 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800167 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700168 if (len >= icsk->icsk_ack.rcv_mss) {
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200169 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
170 tcp_sk(sk)->advmss);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300171 /* Account for possibly-removed options */
172 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173 MAX_TCP_OPTION_SPACE))
174 tcp_gro_dev_warn(sk, skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 } else {
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
178 *
179 * "len" is invariant segment length, including TCP header.
180 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700181 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000182 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
187 */
188 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700189 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
193 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700194 len -= tcp_sk(sk)->tcp_header_len;
195 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700197 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 return;
199 }
200 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700201 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700203 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 }
205}
206
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700207static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700209 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000210 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800212 if (quickacks == 0)
213 quickacks = 2;
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700214 quickacks = min(quickacks, max_quickacks);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700215 if (quickacks > icsk->icsk_ack.quick)
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700216 icsk->icsk_ack.quick = quickacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217}
218
Yuchung Chenga0496ef22018-07-18 13:56:36 -0700219void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700221 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700222
223 tcp_incr_quickack(sk, max_quickacks);
Wei Wang31954cd2019-01-25 10:53:19 -0800224 inet_csk_exit_pingpong_mode(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700225 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226}
Yuchung Chenga0496ef22018-07-18 13:56:36 -0700227EXPORT_SYMBOL(tcp_enter_quickack_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
229/* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
231 */
232
Jon Maxwell2251ae42015-07-08 10:12:28 +1000233static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700235 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000236 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000237
Jon Maxwell2251ae42015-07-08 10:12:28 +1000238 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
Wei Wang31954cd2019-01-25 10:53:19 -0800239 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240}
241
Florian Westphal735d3832014-09-29 13:08:30 +0200242static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700243{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800244 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700245 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
246}
247
Yuchung Chengfd2123a2018-08-09 09:38:12 -0700248static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700249{
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700250 if (tcp_hdr(skb)->cwr) {
Yuchung Chengfd2123a2018-08-09 09:38:12 -0700251 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700252
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
255 * immediately.
256 */
Yuchung Chengfd2123a2018-08-09 09:38:12 -0700257 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700258 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700259}
260
Florian Westphal735d3832014-09-29 13:08:30 +0200261static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700262{
263 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
264}
265
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700266static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700267{
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700268 struct tcp_sock *tp = tcp_sk(sk);
269
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400270 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000271 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700272 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000273 * and we already seen ECT on a previous segment,
274 * it is probably a retransmit.
275 */
276 if (tp->ecn_flags & TCP_ECN_SEEN)
Eric Dumazet15ecbe92018-06-27 08:47:21 -0700277 tcp_enter_quickack_mode(sk, 2);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000278 break;
279 case INET_ECN_CE:
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700280 if (tcp_ca_needs_ecn(sk))
281 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
Florian Westphal98900922014-09-26 22:37:35 +0200282
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000283 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
284 /* Better not delay acks, sender can have a very low cwnd */
Eric Dumazet15ecbe92018-06-27 08:47:21 -0700285 tcp_enter_quickack_mode(sk, 2);
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000286 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
287 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000288 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200289 break;
290 default:
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700291 if (tcp_ca_needs_ecn(sk))
292 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
Florian Westphal98900922014-09-26 22:37:35 +0200293 tp->ecn_flags |= TCP_ECN_SEEN;
294 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700295 }
296}
297
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700298static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
Florian Westphal735d3832014-09-29 13:08:30 +0200299{
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700300 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
301 __tcp_ecn_check_ce(sk, skb);
Florian Westphal735d3832014-09-29 13:08:30 +0200302}
303
304static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700305{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800306 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700307 tp->ecn_flags &= ~TCP_ECN_OK;
308}
309
Florian Westphal735d3832014-09-29 13:08:30 +0200310static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700311{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700313 tp->ecn_flags &= ~TCP_ECN_OK;
314}
315
Florian Westphal735d3832014-09-29 13:08:30 +0200316static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700317{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800318 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000319 return true;
320 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700321}
322
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323/* Buffer size and advertised window tuning.
324 *
325 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
326 */
327
Eric Dumazet6ae70532013-10-01 10:23:44 -0700328static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700330 const struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400331 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700332 int sndmem, per_mss;
333 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334
Eric Dumazet6ae70532013-10-01 10:23:44 -0700335 /* Worst case is non GSO/TSO : each frame consumes one skb
336 * and skb->head is kmalloced using power of two area of memory
337 */
338 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
339 MAX_TCP_HEADER +
340 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
341
342 per_mss = roundup_pow_of_two(per_mss) +
343 SKB_DATA_ALIGN(sizeof(struct sk_buff));
344
345 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
346 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
347
348 /* Fast Recovery (RFC 5681 3.2) :
349 * Cubic needs 1.7 factor, rounded to 2 to include
Linus Torvaldsa9a08842018-02-11 14:34:03 -0800350 * extra cushion (application might react slowly to EPOLLOUT)
Eric Dumazet6ae70532013-10-01 10:23:44 -0700351 */
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400352 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
353 sndmem *= nr_segs * per_mss;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700354
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000355 if (sk->sk_sndbuf < sndmem)
Eric Dumazet356d1832017-11-07 00:29:28 -0800356 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357}
358
359/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
360 *
361 * All tcp_full_space() is split to two parts: "network" buffer, allocated
362 * forward and advertised in receiver window (tp->rcv_wnd) and
363 * "application buffer", required to isolate scheduling/application
364 * latencies from network.
365 * window_clamp is maximal advertised window. It can be less than
366 * tcp_full_space(), in this case tcp_full_space() - window_clamp
367 * is reserved for "application" buffer. The less window_clamp is
368 * the smoother our behaviour from viewpoint of network, but the lower
369 * throughput and the higher sensitivity of the connection to losses. 8)
370 *
371 * rcv_ssthresh is more strict window_clamp used at "slow start"
372 * phase to predict further behaviour of this connection.
373 * It is used for two goals:
374 * - to enforce header prediction at sender, even when application
375 * requires some significant "application buffer". It is check #1.
376 * - to prevent pruning of receive queue because of misprediction
377 * of receiver window. Check #2.
378 *
379 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800380 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 * in common situations. Otherwise, we have to rely on queue collapsing.
382 */
383
384/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700385static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700387 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 /* Optimize this! */
Eric Dumazet94f08932017-10-26 21:55:09 -0700389 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
Eric Dumazet356d1832017-11-07 00:29:28 -0800390 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
392 while (tp->rcv_ssthresh <= window) {
393 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700394 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
396 truesize >>= 1;
397 window >>= 1;
398 }
399 return 0;
400}
401
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400402static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700404 struct tcp_sock *tp = tcp_sk(sk);
405
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 /* Check #1 */
407 if (tp->rcv_ssthresh < tp->window_clamp &&
408 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700409 !tcp_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 int incr;
411
412 /* Check #2. Increase window, if skb with such overhead
413 * will fit to rcvbuf in future.
414 */
Eric Dumazet94f08932017-10-26 21:55:09 -0700415 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800416 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700418 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
420 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000421 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800422 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
423 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700424 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 }
426 }
427}
428
Yuchung Chenga3375312018-09-27 11:21:19 -0700429/* 3. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 * established state.
431 */
Jerry Chu10467162012-08-31 12:29:11 +0000432void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Eric Dumazet0c126542017-10-26 21:55:08 -0700434 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 struct tcp_sock *tp = tcp_sk(sk);
436 int maxwin;
437
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700439 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Yuchung Cheng041a14d2018-10-01 15:42:32 -0700441 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700442 tcp_mstamp_refresh(tp);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700443 tp->rcvq_space.time = tp->tcp_mstamp;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700444 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
446 maxwin = tcp_full_space(sk);
447
448 if (tp->window_clamp >= maxwin) {
449 tp->window_clamp = maxwin;
450
Eric Dumazet0c126542017-10-26 21:55:08 -0700451 if (tcp_app_win && maxwin > 4 * tp->advmss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 tp->window_clamp = max(maxwin -
Eric Dumazet0c126542017-10-26 21:55:08 -0700453 (maxwin >> tcp_app_win),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 4 * tp->advmss);
455 }
456
457 /* Force reservation of one segment. */
Eric Dumazet0c126542017-10-26 21:55:08 -0700458 if (tcp_app_win &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 tp->window_clamp > 2 * tp->advmss &&
460 tp->window_clamp + tp->advmss > maxwin)
461 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
462
463 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
Eric Dumazetc2203cf2017-05-16 14:00:04 -0700464 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
Yuchung Chenga3375312018-09-27 11:21:19 -0700467/* 4. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700468static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700470 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300471 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet356d1832017-11-07 00:29:28 -0800472 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300474 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475
Eric Dumazet356d1832017-11-07 00:29:28 -0800476 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
John Heffner326f36e2005-11-10 17:11:48 -0800477 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700478 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000479 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800480 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
Eric Dumazet356d1832017-11-07 00:29:28 -0800481 net->ipv4.sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 }
John Heffner326f36e2005-11-10 17:11:48 -0800483 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800484 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
486
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800487/* Initialize RCV_MSS value.
488 * RCV_MSS is an our guess about MSS used by the peer.
489 * We haven't any direct information about the MSS.
490 * It's better to underestimate the RCV_MSS rather than overestimate.
491 * Overestimations make us ACKing less frequently than needed.
492 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493 */
494void tcp_initialize_rcv_mss(struct sock *sk)
495{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400496 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800497 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
498
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800499 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000500 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800501 hint = max(hint, TCP_MIN_MSS);
502
503 inet_csk(sk)->icsk_ack.rcv_mss = hint;
504}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000505EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800506
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507/* Receiver "autotuning" code.
508 *
509 * The algorithm for RTT estimation w/o timestamps is based on
510 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200511 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 *
513 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200514 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 * though this reference is out of date. A new paper
516 * is pending.
517 */
518static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
519{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700520 u32 new_sample = tp->rcv_rtt_est.rtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 long m = sample;
522
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 if (new_sample != 0) {
524 /* If we sample in larger samples in the non-timestamp
525 * case, we could grossly overestimate the RTT especially
526 * with chatty applications or bulk transfer apps which
527 * are stalled on filesystem I/O.
528 *
529 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800530 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800531 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 * long.
533 */
534 if (!win_dep) {
535 m -= (new_sample >> 3);
536 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000537 } else {
538 m <<= 3;
539 if (m < new_sample)
540 new_sample = m;
541 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800543 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 new_sample = m << 3;
545 }
546
Eric Dumazet645f4c62017-04-25 10:15:41 -0700547 tp->rcv_rtt_est.rtt_us = new_sample;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548}
549
550static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
551{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700552 u32 delta_us;
553
Eric Dumazet9a568de2017-05-16 14:00:14 -0700554 if (tp->rcv_rtt_est.time == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 goto new_measure;
556 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
557 return;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700558 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
Wei Wang9ee11bd2017-12-12 16:28:58 -0800559 if (!delta_us)
560 delta_us = 1;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700561 tcp_rcv_rtt_update(tp, delta_us, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
563new_measure:
564 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700565 tp->rcv_rtt_est.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566}
567
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800568static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
569 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700571 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700572
Wei Wang3f6c65d2018-06-19 21:42:50 -0700573 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
574 return;
575 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
576
577 if (TCP_SKB_CB(skb)->end_seq -
578 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
Eric Dumazet9a568de2017-05-16 14:00:14 -0700579 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
Wei Wang9ee11bd2017-12-12 16:28:58 -0800580 u32 delta_us;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700581
Eric Dumazet9efdda42018-11-24 09:12:24 -0800582 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
583 if (!delta)
584 delta = 1;
585 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
586 tcp_rcv_rtt_update(tp, delta_us, 0);
587 }
Eric Dumazet9a568de2017-05-16 14:00:14 -0700588 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589}
590
591/*
592 * This function should be called every time data is copied to user space.
593 * It calculates the appropriate TCP receive buffer space.
594 */
595void tcp_rcv_space_adjust(struct sock *sk)
596{
597 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet607065b2017-12-10 17:55:03 -0800598 u32 copied;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 int time;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900600
Yafang Shao61638492018-04-20 23:18:26 +0800601 trace_tcp_rcv_space_adjust(sk);
602
Eric Dumazet86323852017-12-06 11:08:19 -0800603 tcp_mstamp_refresh(tp);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700604 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700605 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900607
Eric Dumazetb0983d32013-09-20 13:56:58 -0700608 /* Number of bytes copied to user in last RTT */
609 copied = tp->copied_seq - tp->rcvq_space.seq;
610 if (copied <= tp->rcvq_space.space)
611 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612
Eric Dumazetb0983d32013-09-20 13:56:58 -0700613 /* A bit of theory :
614 * copied = bytes received in previous RTT, our base window
615 * To cope with packet losses, we need a 2x factor
616 * To cope with slow start, and sender growing its cwin by 100 %
617 * every RTT, we need a 4x factor, because the ACK we are sending
618 * now is for the next RTT, not the current one :
619 * <prev RTT . ><current RTT .. ><next RTT .... >
620 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Eric Dumazet4540c0c2017-10-27 07:47:22 -0700622 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
Eric Dumazetb0983d32013-09-20 13:56:58 -0700623 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Eric Dumazet607065b2017-12-10 17:55:03 -0800624 int rcvmem, rcvbuf;
Eric Dumazetc3916ad2017-12-10 17:55:04 -0800625 u64 rcvwin, grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626
Eric Dumazetb0983d32013-09-20 13:56:58 -0700627 /* minimal window to cope with packet losses, assuming
628 * steady state. Add some cushion because of small variations.
629 */
Eric Dumazet607065b2017-12-10 17:55:03 -0800630 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
Eric Dumazetc3916ad2017-12-10 17:55:04 -0800632 /* Accommodate for sender rate increase (eg. slow start) */
633 grow = rcvwin * (copied - tp->rcvq_space.space);
634 do_div(grow, tp->rcvq_space.space);
635 rcvwin += (grow << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636
Eric Dumazetb0983d32013-09-20 13:56:58 -0700637 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
Eric Dumazet94f08932017-10-26 21:55:09 -0700638 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
Eric Dumazetb0983d32013-09-20 13:56:58 -0700639 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
Eric Dumazet607065b2017-12-10 17:55:03 -0800641 do_div(rcvwin, tp->advmss);
642 rcvbuf = min_t(u64, rcvwin * rcvmem,
643 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
Eric Dumazetb0983d32013-09-20 13:56:58 -0700644 if (rcvbuf > sk->sk_rcvbuf) {
645 sk->sk_rcvbuf = rcvbuf;
646
647 /* Make the window clamp follow along. */
Eric Dumazet02db5572017-12-10 17:55:02 -0800648 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 }
650 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700651 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900652
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653new_measure:
654 tp->rcvq_space.seq = tp->copied_seq;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700655 tp->rcvq_space.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656}
657
658/* There is something which you must keep in mind when you analyze the
659 * behavior of the tp->ato delayed ack timeout interval. When a
660 * connection starts up, we want to ack as quickly as possible. The
661 * problem is that "good" TCP's do slow start at the beginning of data
662 * transmission. The means that until we send the first few ACK's the
663 * sender will sit on his end and only queue most of his data, because
664 * he can only send snd_cwnd unacked packets at any given time. For
665 * each ACK we send, he increments snd_cwnd and transmits more of his
666 * queue. -DaveM
667 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700668static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700670 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700671 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 u32 now;
673
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700674 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700676 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
678 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900679
Eric Dumazet70eabf02017-05-16 14:00:07 -0700680 now = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700682 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 /* The _first_ data packet received, initialize
684 * delayed ACK engine.
685 */
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700686 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700687 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700689 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800691 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700693 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
694 } else if (m < icsk->icsk_ack.ato) {
695 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
696 if (icsk->icsk_ack.ato > icsk->icsk_rto)
697 icsk->icsk_ack.ato = icsk->icsk_rto;
698 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800699 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 * restart window, so that we send ACKs quickly.
701 */
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700702 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800703 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 }
705 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700706 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700708 tcp_ecn_check_ce(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700711 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712}
713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714/* Called to compute a smoothed rtt estimate. The data fed to this
715 * routine either comes from timestamps, or from segments that were
716 * known _not_ to have been retransmitted [see Karn/Partridge
717 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
718 * piece by Van Jacobson.
719 * NOTE: the next three routines used to be one big routine.
720 * To save cycles in the RFC 1323 implementation it was better to break
721 * it up into three procedures. -- erics
722 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800723static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300725 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800726 long m = mrtt_us; /* RTT */
727 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 /* The following amusing code comes from Jacobson's
730 * article in SIGCOMM '88. Note that rtt and mdev
731 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900732 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 * m stands for "measurement".
734 *
735 * On a 1990 paper the rto value is changed to:
736 * RTO = rtt + 4 * mdev
737 *
738 * Funny. This algorithm seems to be very broken.
739 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800740 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
742 * does not matter how to _calculate_ it. Seems, it was trap
743 * that VJ failed to avoid. 8)
744 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800745 if (srtt != 0) {
746 m -= (srtt >> 3); /* m is now error in rtt est */
747 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 if (m < 0) {
749 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800750 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 /* This is similar to one of Eifel findings.
752 * Eifel blocks mdev updates when rtt decreases.
753 * This solution is a bit different: we use finer gain
754 * for mdev in this case (alpha*beta).
755 * Like Eifel it also prevents growth of rto,
756 * but also it limits too fast rto decreases,
757 * happening in pure Eifel.
758 */
759 if (m > 0)
760 m >>= 3;
761 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800762 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800764 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
765 if (tp->mdev_us > tp->mdev_max_us) {
766 tp->mdev_max_us = tp->mdev_us;
767 if (tp->mdev_max_us > tp->rttvar_us)
768 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 }
770 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800771 if (tp->mdev_max_us < tp->rttvar_us)
772 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800774 tp->mdev_max_us = tcp_rto_min_us(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 }
776 } else {
777 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800778 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800779 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
780 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
781 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 tp->rtt_seq = tp->snd_nxt;
783 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800784 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785}
786
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700787static void tcp_update_pacing_rate(struct sock *sk)
788{
789 const struct tcp_sock *tp = tcp_sk(sk);
790 u64 rate;
791
792 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700793 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
794
795 /* current rate is (cwnd * mss) / srtt
796 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
797 * In Congestion Avoidance phase, set it to 120 % the current rate.
798 *
799 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
800 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
801 * end of slow start and should slow down.
802 */
803 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
Eric Dumazet23a7102a2017-10-27 07:47:31 -0700804 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
Eric Dumazet43e122b2015-08-21 17:38:02 -0700805 else
Eric Dumazetc26e91f2017-10-27 07:47:32 -0700806 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700807
808 rate *= max(tp->snd_cwnd, tp->packets_out);
809
Eric Dumazet740b0f12014-02-26 14:02:48 -0800810 if (likely(tp->srtt_us))
811 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700812
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700813 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
Eric Dumazetba537422013-10-09 17:14:52 -0700814 * without any lock. We want to make sure compiler wont store
815 * intermediate values in this location.
816 */
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700817 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
818 sk->sk_max_pacing_rate));
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700819}
820
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821/* Calculate rto without backoff. This is the second half of Van Jacobson's
822 * routine referred to above.
823 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800824static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700826 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 /* Old crap is replaced with new one. 8)
828 *
829 * More seriously:
830 * 1. If rtt variance happened to be less 50msec, it is hallucination.
831 * It cannot be less due to utterly erratic ACK generation made
832 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
833 * to do with delayed acks, because at cwnd>2 true delack timeout
834 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800835 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000837 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838
839 /* 2. Fixups made earlier cannot be right.
840 * If we do not estimate RTO correctly without them,
841 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800842 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800845 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
846 * guarantees that rto is higher.
847 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000848 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849}
850
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400851__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
853 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
854
Gerrit Renker22b71c82010-08-29 19:23:12 +0000855 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800856 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
858}
859
Ryousei Takano564262c2007-10-25 23:03:52 -0700860/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300861static void tcp_dsack_seen(struct tcp_sock *tp)
862{
Vijay Subramanianab562222011-12-20 13:23:24 +0000863 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Priyaranjan Jha1f255692017-11-03 16:38:48 -0700864 tp->rack.dsack_seen = 1;
Wei Wang7e10b652018-07-31 17:46:23 -0700865 tp->dsack_dups++;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300866}
867
Yuchung Cheng737ff312017-11-08 13:01:27 -0800868/* It's reordering when higher sequence was delivered (i.e. sacked) before
869 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
870 * distance is approximated in full-mss packet distance ("reordering").
871 */
872static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
873 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300875 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -0800876 const u32 mss = tp->mss_cache;
877 u32 fack, metric;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700878
Yuchung Cheng737ff312017-11-08 13:01:27 -0800879 fack = tcp_highest_sack_seq(tp);
880 if (!before(low_seq, fack))
Soheil Hassas Yeganeh6f5b24e2017-05-16 17:39:02 -0400881 return;
882
Yuchung Cheng737ff312017-11-08 13:01:27 -0800883 metric = fack - low_seq;
884 if ((metric > tp->reordering * mss) && mss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000886 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
887 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
888 tp->reordering,
Yuchung Cheng737ff312017-11-08 13:01:27 -0800889 0,
Joe Perches91df42b2012-05-15 14:11:54 +0000890 tp->sacked_out,
891 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892#endif
Yuchung Cheng737ff312017-11-08 13:01:27 -0800893 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
894 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000896
Yuchung Cheng2d2517e2017-04-04 14:15:40 -0700897 /* This exciting event is worth to be remembered. 8) */
Wei Wang7ec65372018-07-31 17:46:24 -0700898 tp->reord_seen++;
Yuchung Cheng737ff312017-11-08 13:01:27 -0800899 NET_INC_STATS(sock_net(sk),
900 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901}
902
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700903/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700904static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
905{
Ian Morris51456b22015-04-03 09:17:26 +0100906 if (!tp->retransmit_skb_hint ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700907 before(TCP_SKB_CB(skb)->seq,
908 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700909 tp->retransmit_skb_hint = skb;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700910}
911
Neal Cardwell0682e692016-09-19 23:39:13 -0400912/* Sum the number of packets on the wire we have marked as lost.
913 * There are two cases we care about here:
914 * a) Packet hasn't been marked lost (nor retransmitted),
915 * and this is the first loss.
916 * b) Packet has been marked both lost and retransmitted,
917 * and this means we think it was lost again.
918 */
919static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
920{
921 __u8 sacked = TCP_SKB_CB(skb)->sacked;
922
923 if (!(sacked & TCPCB_LOST) ||
924 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
925 tp->lost += tcp_skb_pcount(skb);
926}
927
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700928static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
929{
930 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
931 tcp_verify_retransmit_hint(tp, skb);
932
933 tp->lost_out += tcp_skb_pcount(skb);
Neal Cardwell0682e692016-09-19 23:39:13 -0400934 tcp_sum_lost(tp, skb);
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700935 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
936 }
937}
938
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700939void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700940{
941 tcp_verify_retransmit_hint(tp, skb);
942
Neal Cardwell0682e692016-09-19 23:39:13 -0400943 tcp_sum_lost(tp, skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tp->lost_out += tcp_skb_pcount(skb);
946 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
947 }
948}
949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950/* This procedure tags the retransmission queue when SACKs arrive.
951 *
952 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
953 * Packets in queue with these bits set are counted in variables
954 * sacked_out, retrans_out and lost_out, correspondingly.
955 *
956 * Valid combinations are:
957 * Tag InFlight Description
958 * 0 1 - orig segment is in flight.
959 * S 0 - nothing flies, orig reached receiver.
960 * L 0 - nothing flies, orig lost by net.
961 * R 2 - both orig and retransmit are in flight.
962 * L|R 1 - orig is lost, retransmit is in flight.
963 * S|R 1 - orig reached receiver, retrans is still in flight.
964 * (L|S|R is logically valid, it could occur when L|R is sacked,
965 * but it is equivalent to plain S and code short-curcuits it to S.
966 * L|S is logically invalid, it would mean -1 packet in flight 8))
967 *
968 * These 6 states form finite state machine, controlled by the following events:
969 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
970 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000971 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 * A. Scoreboard estimator decided the packet is lost.
973 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000974 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 * segment was retransmitted.
976 * 4. D-SACK added new rule: D-SACK changes any tag to S.
977 *
978 * It is pleasant to note, that state diagram turns out to be commutative,
979 * so that we are allowed not to be bothered by order of our actions,
980 * when multiple events arrive simultaneously. (see the function below).
981 *
982 * Reordering detection.
983 * --------------------
984 * Reordering metric is maximal distance, which a packet can be displaced
985 * in packet stream. With SACKs we can estimate it:
986 *
987 * 1. SACK fills old hole and the corresponding segment was not
988 * ever retransmitted -> reordering. Alas, we cannot use it
989 * when segment was retransmitted.
990 * 2. The last flaw is solved with D-SACK. D-SACK arrives
991 * for retransmitted and already SACKed segment -> reordering..
992 * Both of these heuristics are not used in Loss state, when we cannot
993 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700994 *
995 * SACK block validation.
996 * ----------------------
997 *
998 * SACK block range validation checks that the received SACK block fits to
999 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1000 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001001 * it means that the receiver is rather inconsistent with itself reporting
1002 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1003 * perfectly valid, however, in light of RFC2018 which explicitly states
1004 * that "SACK block MUST reflect the newest segment. Even if the newest
1005 * segment is going to be discarded ...", not that it looks very clever
1006 * in case of head skb. Due to potentional receiver driven attacks, we
1007 * choose to avoid immediate execution of a walk in write queue due to
1008 * reneging and defer head skb's loss recovery to standard loss recovery
1009 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001010 *
1011 * Implements also blockage to start_seq wrap-around. Problem lies in the
1012 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1013 * there's no guarantee that it will be before snd_nxt (n). The problem
1014 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1015 * wrap (s_w):
1016 *
1017 * <- outs wnd -> <- wrapzone ->
1018 * u e n u_w e_w s n_w
1019 * | | | | | | |
1020 * |<------------+------+----- TCP seqno space --------------+---------->|
1021 * ...-- <2^31 ->| |<--------...
1022 * ...---- >2^31 ------>| |<--------...
1023 *
1024 * Current code wouldn't be vulnerable but it's better still to discard such
1025 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1026 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1027 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1028 * equal to the ideal case (infinite seqno space without wrap caused issues).
1029 *
1030 * With D-SACK the lower bound is extended to cover sequence space below
1031 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001032 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001033 * for the normal SACK blocks, explained above). But there all simplicity
1034 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1035 * fully below undo_marker they do not affect behavior in anyway and can
1036 * therefore be safely ignored. In rare cases (which are more or less
1037 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1038 * fragmentation and packet reordering past skb's retransmission. To consider
1039 * them correctly, the acceptable range must be extended even more though
1040 * the exact amount is rather hard to quantify. However, tp->max_window can
1041 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001043static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1044 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001045{
1046 /* Too far in future, or reversed (interpretation is ambiguous) */
1047 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001048 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001049
1050 /* Nasty start_seq wrap-around check (see comments above) */
1051 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001052 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001053
Ryousei Takano564262c2007-10-25 23:03:52 -07001054 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001055 * start_seq == snd_una is non-sensical (see comments above)
1056 */
1057 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001058 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001059
1060 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001061 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001062
1063 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001064 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001065 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001066
1067 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001068 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001069
1070 /* Too old */
1071 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001072 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001073
1074 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1075 * start_seq < undo_marker and end_seq >= undo_marker.
1076 */
1077 return !before(start_seq, end_seq - tp->max_window);
1078}
1079
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001080static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1081 struct tcp_sack_block_wire *sp, int num_sacks,
1082 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001083{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001084 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001085 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1086 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001087 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001088
1089 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001090 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001091 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001092 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001093 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001094 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1095 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001096
1097 if (!after(end_seq_0, end_seq_1) &&
1098 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001099 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001100 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001101 NET_INC_STATS(sock_net(sk),
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001102 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001103 }
1104 }
1105
1106 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001107 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001108 !after(end_seq_0, prior_snd_una) &&
1109 after(end_seq_0, tp->undo_marker))
1110 tp->undo_retrans--;
1111
1112 return dup_sack;
1113}
1114
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001115struct tcp_sacktag_state {
Yuchung Cheng737ff312017-11-08 13:01:27 -08001116 u32 reord;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001117 /* Timestamps for earliest and latest never-retransmitted segment
1118 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1119 * but congestion control should still get an accurate delay signal.
1120 */
Eric Dumazet9a568de2017-05-16 14:00:14 -07001121 u64 first_sackt;
1122 u64 last_sackt;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001123 struct rate_sample *rate;
Eric Dumazet740b0f12014-02-26 14:02:48 -08001124 int flag;
Eric Dumazet75c119a2017-10-05 22:21:27 -07001125 unsigned int mss_now;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001126};
1127
Ilpo Järvinend1935942007-10-11 17:34:25 -07001128/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1129 * the incoming SACK may not exactly match but we can find smaller MSS
1130 * aligned portion of it that matches. Therefore we might need to fragment
1131 * which may fail and creates some hassle (caller must handle error case
1132 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001133 *
1134 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001135 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001136static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001137 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001138{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001139 int err;
1140 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001141 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001142 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001143
1144 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1145 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1146
1147 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1148 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001149 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1151
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001152 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001153 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001154 if (pkt_len < mss)
1155 pkt_len = mss;
1156 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001157 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001158 if (pkt_len < mss)
1159 return -EINVAL;
1160 }
1161
1162 /* Round if necessary so that SACKs cover only full MSSes
1163 * and/or the remaining small portion (if present)
1164 */
1165 if (pkt_len > mss) {
1166 unsigned int new_len = (pkt_len / mss) * mss;
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001167 if (!in_sack && new_len < pkt_len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001168 new_len += mss;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001169 pkt_len = new_len;
1170 }
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001171
1172 if (pkt_len >= skb->len && !in_sack)
1173 return 0;
1174
Eric Dumazet75c119a2017-10-05 22:21:27 -07001175 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1176 pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001177 if (err < 0)
1178 return err;
1179 }
1180
1181 return in_sack;
1182}
1183
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001184/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1185static u8 tcp_sacktag_one(struct sock *sk,
1186 struct tcp_sacktag_state *state, u8 sacked,
1187 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001188 int dup_sack, int pcount,
Eric Dumazet9a568de2017-05-16 14:00:14 -07001189 u64 xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001190{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001191 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001192
1193 /* Account D-SACK for retransmitted packet. */
1194 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001195 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001196 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001197 tp->undo_retrans--;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001198 if ((sacked & TCPCB_SACKED_ACKED) &&
1199 before(start_seq, state->reord))
1200 state->reord = start_seq;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001201 }
1202
1203 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001204 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001205 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001206
1207 if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazetd2329f12017-04-25 10:15:38 -07001208 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001209
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001210 if (sacked & TCPCB_SACKED_RETRANS) {
1211 /* If the segment is not tagged as lost,
1212 * we do not clear RETRANS, believing
1213 * that retransmission is still in flight.
1214 */
1215 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001216 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001217 tp->lost_out -= pcount;
1218 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001219 }
1220 } else {
1221 if (!(sacked & TCPCB_RETRANS)) {
1222 /* New sack for not retransmitted frame,
1223 * which was in hole. It is reordering.
1224 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001225 if (before(start_seq,
Yuchung Cheng737ff312017-11-08 13:01:27 -08001226 tcp_highest_sack_seq(tp)) &&
1227 before(start_seq, state->reord))
1228 state->reord = start_seq;
1229
Yuchung Chenge33099f2013-03-20 13:33:00 +00001230 if (!after(end_seq, tp->high_seq))
1231 state->flag |= FLAG_ORIG_SACK_ACKED;
Eric Dumazet9a568de2017-05-16 14:00:14 -07001232 if (state->first_sackt == 0)
1233 state->first_sackt = xmit_time;
1234 state->last_sackt = xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001235 }
1236
1237 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001238 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001239 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001240 }
1241 }
1242
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001243 sacked |= TCPCB_SACKED_ACKED;
1244 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001245 tp->sacked_out += pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08001246 tp->delivered += pcount; /* Out-of-order packets delivered */
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001247
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001248 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Yuchung Cheng713bafe2017-11-08 13:01:26 -08001249 if (tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001250 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001251 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001252 }
1253
1254 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255 * frames and clear it. undo_retrans is decreased above, L|R frames
1256 * are accounted above as well.
1257 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001258 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001260 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001261 }
1262
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001263 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001264}
1265
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001266/* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268 */
Eric Dumazetf3319812017-10-05 22:21:26 -07001269static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1270 struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001271 struct tcp_sacktag_state *state,
1272 unsigned int pcount, int shifted, int mss,
1273 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001274{
1275 struct tcp_sock *tp = tcp_sk(sk);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001276 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1277 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001278
1279 BUG_ON(!pcount);
1280
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001281 /* Adjust counters and hints for the newly sacked sequence
1282 * range but discard the return value since prev is already
1283 * marked. We must tag the range first because the seq
1284 * advancement below implicitly advances
1285 * tcp_highest_sack_seq() when skb is highest_sack.
1286 */
1287 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001288 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07001289 tcp_skb_timestamp_us(skb));
Yuchung Chengb9f64822016-09-19 23:39:14 -04001290 tcp_rate_skb_delivered(sk, skb, state->rate);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001291
1292 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001293 tp->lost_cnt_hint += pcount;
1294
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001295 TCP_SKB_CB(prev)->end_seq += shifted;
1296 TCP_SKB_CB(skb)->seq += shifted;
1297
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001298 tcp_skb_pcount_add(prev, pcount);
1299 BUG_ON(tcp_skb_pcount(skb) < pcount);
1300 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001301
1302 /* When we're adding to gso_segs == 1, gso_size will be zero,
1303 * in theory this shouldn't be necessary but as long as DSACK
1304 * code can come after this skb later on it's better to keep
1305 * setting gso_size to something.
1306 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001307 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1308 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001309
1310 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001311 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001312 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001313
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001314 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1315 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1316
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001317 if (skb->len > 0) {
1318 BUG_ON(!tcp_skb_pcount(skb));
Eric Dumazetc10d9312016-04-29 14:16:47 -07001319 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001320 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001321 }
1322
1323 /* Whole SKB was eaten :-) */
1324
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001325 if (skb == tp->retransmit_skb_hint)
1326 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001327 if (skb == tp->lost_skb_hint) {
1328 tp->lost_skb_hint = prev;
1329 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1330 }
1331
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001332 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001333 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001334 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1335 TCP_SKB_CB(prev)->end_seq++;
1336
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001337 if (skb == tcp_highest_sack(sk))
1338 tcp_advance_highest_sack(sk, skb);
1339
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07001340 tcp_skb_collapse_tstamp(prev, skb);
Eric Dumazet9a568de2017-05-16 14:00:14 -07001341 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1342 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001343
Eric Dumazet75c119a2017-10-05 22:21:27 -07001344 tcp_rtx_queue_unlink_and_free(skb, sk);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001345
Eric Dumazetc10d9312016-04-29 14:16:47 -07001346 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001347
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001348 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001349}
1350
1351/* I wish gso_size would have a bit more sane initialization than
1352 * something-or-zero which complicates things
1353 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001354static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001355{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001356 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001357}
1358
1359/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001360static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001361{
1362 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1363}
1364
1365/* Try collapsing SACK blocks spanning across multiple skbs to a single
1366 * skb.
1367 */
1368static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001369 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001370 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001371 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001372{
1373 struct tcp_sock *tp = tcp_sk(sk);
1374 struct sk_buff *prev;
1375 int mss;
1376 int pcount = 0;
1377 int len;
1378 int in_sack;
1379
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001380 /* Normally R but no L won't result in plain S */
1381 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001382 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001383 goto fallback;
1384 if (!skb_can_shift(skb))
1385 goto fallback;
1386 /* This frame is about to be dropped (was ACKed). */
1387 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1388 goto fallback;
1389
1390 /* Can only happen with delayed DSACK + discard craziness */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001391 prev = skb_rb_prev(skb);
1392 if (!prev)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001393 goto fallback;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001394
1395 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1396 goto fallback;
1397
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001398 if (!tcp_skb_can_collapse_to(prev))
1399 goto fallback;
1400
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001401 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1402 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1403
1404 if (in_sack) {
1405 len = skb->len;
1406 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001407 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001408
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1411 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001412 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001413 goto fallback;
1414 } else {
1415 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1416 goto noop;
1417 /* CHECKME: This is non-MSS split case only?, this will
1418 * cause skipped skbs due to advancing loop btw, original
1419 * has that feature too
1420 */
1421 if (tcp_skb_pcount(skb) <= 1)
1422 goto noop;
1423
1424 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1425 if (!in_sack) {
1426 /* TODO: head merge to next could be attempted here
1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1428 * though it might not be worth of the additional hassle
1429 *
1430 * ...we can probably just fallback to what was done
1431 * previously. We could try merging non-SACKed ones
1432 * as well but it probably isn't going to buy off
1433 * because later SACKs might again split them, and
1434 * it would make skb timestamp tracking considerably
1435 * harder problem.
1436 */
1437 goto fallback;
1438 }
1439
1440 len = end_seq - TCP_SKB_CB(skb)->seq;
1441 BUG_ON(len < 0);
1442 BUG_ON(len > skb->len);
1443
1444 /* MSS boundaries should be honoured or else pcount will
1445 * severely break even though it makes things bit trickier.
1446 * Optimize common case to avoid most of the divides
1447 */
1448 mss = tcp_skb_mss(skb);
1449
1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1451 * drop this restriction as unnecessary
1452 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001453 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001454 goto fallback;
1455
1456 if (len == mss) {
1457 pcount = 1;
1458 } else if (len < mss) {
1459 goto noop;
1460 } else {
1461 pcount = len / mss;
1462 len = pcount * mss;
1463 }
1464 }
1465
Neal Cardwell4648dc92012-03-05 19:35:04 +00001466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1467 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1468 goto fallback;
1469
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001470 if (!skb_shift(prev, skb, len))
1471 goto fallback;
Eric Dumazetf3319812017-10-05 22:21:26 -07001472 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001473 goto out;
1474
1475 /* Hole filled allows collapsing with the next as well, this is very
1476 * useful when hole on every nth skb pattern happens
1477 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001478 skb = skb_rb_next(prev);
1479 if (!skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001480 goto out;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001481
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001482 if (!skb_can_shift(skb) ||
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001483 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001484 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001485 goto out;
1486
1487 len = skb->len;
1488 if (skb_shift(prev, skb, len)) {
1489 pcount += tcp_skb_pcount(skb);
Eric Dumazetf3319812017-10-05 22:21:26 -07001490 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1491 len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001492 }
1493
1494out:
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001495 return prev;
1496
1497noop:
1498 return skb;
1499
1500fallback:
Eric Dumazetc10d9312016-04-29 14:16:47 -07001501 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001502 return NULL;
1503}
1504
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001505static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1506 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001507 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001508 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001509 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001510{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001511 struct tcp_sock *tp = tcp_sk(sk);
1512 struct sk_buff *tmp;
1513
Eric Dumazet75c119a2017-10-05 22:21:27 -07001514 skb_rbtree_walk_from(skb) {
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001515 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001516 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001517
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001518 /* queue is in-order => we can short-circuit the walk early */
1519 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1520 break;
1521
Ian Morris00db4122015-04-03 09:17:27 +01001522 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001523 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1524 in_sack = tcp_match_skb_to_sack(sk, skb,
1525 next_dup->start_seq,
1526 next_dup->end_seq);
1527 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001528 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001529 }
1530
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001531 /* skb reference here is a bit tricky to get right, since
1532 * shifting can eat and free both this skb and the next,
1533 * so not even _safe variant of the loop is enough.
1534 */
1535 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001536 tmp = tcp_shift_skb_data(sk, skb, state,
1537 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001538 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001539 if (tmp != skb) {
1540 skb = tmp;
1541 continue;
1542 }
1543
1544 in_sack = 0;
1545 } else {
1546 in_sack = tcp_match_skb_to_sack(sk, skb,
1547 start_seq,
1548 end_seq);
1549 }
1550 }
1551
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001552 if (unlikely(in_sack < 0))
1553 break;
1554
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001555 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001556 TCP_SKB_CB(skb)->sacked =
1557 tcp_sacktag_one(sk,
1558 state,
1559 TCP_SKB_CB(skb)->sacked,
1560 TCP_SKB_CB(skb)->seq,
1561 TCP_SKB_CB(skb)->end_seq,
1562 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001563 tcp_skb_pcount(skb),
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07001564 tcp_skb_timestamp_us(skb));
Yuchung Chengb9f64822016-09-19 23:39:14 -04001565 tcp_rate_skb_delivered(sk, skb, state->rate);
Eric Dumazete2080072017-10-04 12:59:58 -07001566 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1567 list_del_init(&skb->tcp_tsorted_anchor);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001568
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001569 if (!before(TCP_SKB_CB(skb)->seq,
1570 tcp_highest_sack_seq(tp)))
1571 tcp_advance_highest_sack(sk, skb);
1572 }
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001573 }
1574 return skb;
1575}
1576
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001577static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
Eric Dumazet75c119a2017-10-05 22:21:27 -07001578{
1579 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1580 struct sk_buff *skb;
Eric Dumazet75c119a2017-10-05 22:21:27 -07001581
1582 while (*p) {
1583 parent = *p;
1584 skb = rb_to_skb(parent);
1585 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1586 p = &parent->rb_left;
1587 continue;
1588 }
1589 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1590 p = &parent->rb_right;
1591 continue;
1592 }
Eric Dumazet75c119a2017-10-05 22:21:27 -07001593 return skb;
1594 }
1595 return NULL;
1596}
1597
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001598static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001599 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001600{
Eric Dumazet75c119a2017-10-05 22:21:27 -07001601 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1602 return skb;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001603
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001604 return tcp_sacktag_bsearch(sk, skip_to_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001605}
1606
1607static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1608 struct sock *sk,
1609 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001610 struct tcp_sacktag_state *state,
1611 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001612{
Ian Morris51456b22015-04-03 09:17:26 +01001613 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001614 return skb;
1615
1616 if (before(next_dup->start_seq, skip_to_seq)) {
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001617 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001618 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1619 next_dup->start_seq, next_dup->end_seq,
1620 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001621 }
1622
1623 return skb;
1624}
1625
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001626static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001627{
1628 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1629}
1630
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001632tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001633 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634{
1635 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001636 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1637 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001638 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001639 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001640 struct tcp_sack_block *cache;
1641 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001642 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001643 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001644 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001645 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001646 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001648 state->flag = 0;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001649 state->reord = tp->snd_nxt;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001650
Yuchung Cheng737ff312017-11-08 13:01:27 -08001651 if (!tp->sacked_out)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001652 tcp_highest_sack_reset(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001654 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001655 num_sacks, prior_snd_una);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001656 if (found_dup_sack) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001657 state->flag |= FLAG_DSACKING_ACK;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001658 tp->delivered++; /* A spurious retransmission is delivered */
1659 }
Baruch Even6f746512007-02-04 23:36:42 -08001660
1661 /* Eliminate too old ACKs, but take into
1662 * account more or less fresh ones, they can
1663 * contain valid SACK info.
1664 */
1665 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1666 return 0;
1667
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001668 if (!tp->packets_out)
1669 goto out;
1670
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001671 used_sacks = 0;
1672 first_sack_index = 0;
1673 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001674 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001675
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001676 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1677 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001678
1679 if (!tcp_is_sackblock_valid(tp, dup_sack,
1680 sp[used_sacks].start_seq,
1681 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001682 int mib_idx;
1683
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001684 if (dup_sack) {
1685 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001686 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001687 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001688 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001689 } else {
1690 /* Don't count olds caused by ACK reordering */
1691 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1692 !after(sp[used_sacks].end_seq, tp->snd_una))
1693 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001694 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001695 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001696
Eric Dumazetc10d9312016-04-29 14:16:47 -07001697 NET_INC_STATS(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001698 if (i == 0)
1699 first_sack_index = -1;
1700 continue;
1701 }
1702
1703 /* Ignore very old stuff early */
1704 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1705 continue;
1706
1707 used_sacks++;
1708 }
1709
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001710 /* order SACK blocks to allow in order walk of the retrans queue */
1711 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001712 for (j = 0; j < i; j++) {
1713 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001714 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001715
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001716 /* Track where the first SACK block goes to */
1717 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001718 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001719 }
1720 }
1721 }
1722
Eric Dumazet75c119a2017-10-05 22:21:27 -07001723 state->mss_now = tcp_current_mss(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07001724 skb = NULL;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001725 i = 0;
1726
1727 if (!tp->sacked_out) {
1728 /* It's already past, so skip checking against it */
1729 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1730 } else {
1731 cache = tp->recv_sack_cache;
1732 /* Skip empty blocks in at head of the cache */
1733 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1734 !cache->end_seq)
1735 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001736 }
1737
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001738 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001739 u32 start_seq = sp[i].start_seq;
1740 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001741 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001742 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001743
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001744 if (found_dup_sack && ((i + 1) == first_sack_index))
1745 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001747 /* Skip too early cached blocks */
1748 while (tcp_sack_cache_ok(tp, cache) &&
1749 !before(start_seq, cache->end_seq))
1750 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001752 /* Can skip some work by looking recv_sack_cache? */
1753 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1754 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001755
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001756 /* Head todo? */
1757 if (before(start_seq, cache->start_seq)) {
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001758 skb = tcp_sacktag_skip(skb, sk, start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001759 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001760 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001761 start_seq,
1762 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001763 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001764 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001765
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001766 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001767 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001768 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001769
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001770 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001771 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001772 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001773
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001774 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001775 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001776 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001777 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001778 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001779 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001780 cache++;
1781 goto walk;
1782 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001784 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001785 /* Check overlap against next cached too (past this one already) */
1786 cache++;
1787 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001789
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001790 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1791 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001792 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001793 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001794 }
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001795 skb = tcp_sacktag_skip(skb, sk, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001796
1797walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001798 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001799 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001800
1801advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001802 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 }
1804
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001805 /* Clear the head of the cache sack blocks so we can skip it next time */
1806 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1807 tp->recv_sack_cache[i].start_seq = 0;
1808 tp->recv_sack_cache[i].end_seq = 0;
1809 }
1810 for (j = 0; j < used_sacks; j++)
1811 tp->recv_sack_cache[i++] = sp[j];
1812
Yuchung Cheng737ff312017-11-08 13:01:27 -08001813 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1814 tcp_check_sack_reordering(sk, state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001816 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001817out:
1818
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001820 WARN_ON((int)tp->sacked_out < 0);
1821 WARN_ON((int)tp->lost_out < 0);
1822 WARN_ON((int)tp->retrans_out < 0);
1823 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001825 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826}
1827
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001828/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001829 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001830 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001831static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001832{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001833 u32 holes;
1834
1835 holes = max(tp->lost_out, 1U);
1836 holes = min(holes, tp->packets_out);
1837
1838 if ((tp->sacked_out + holes) > tp->packets_out) {
1839 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001840 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001841 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001842 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001843}
1844
1845/* If we receive more dupacks than we expected counting segments
1846 * in assumption of absent reordering, interpret this as reordering.
1847 * The only another reason could be bug in receiver TCP.
1848 */
1849static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1850{
1851 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -08001852
1853 if (!tcp_limit_reno_sacked(tp))
1854 return;
1855
1856 tp->reordering = min_t(u32, tp->packets_out + addend,
1857 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
Wei Wang7ec65372018-07-31 17:46:24 -07001858 tp->reord_seen++;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001859 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001860}
1861
1862/* Emulate SACKs for SACKless connection: account for a new dupack. */
1863
Eric Dumazet19119f22018-11-27 14:42:01 -08001864static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001865{
Eric Dumazet19119f22018-11-27 14:42:01 -08001866 if (num_dupack) {
1867 struct tcp_sock *tp = tcp_sk(sk);
1868 u32 prior_sacked = tp->sacked_out;
1869 s32 delivered;
Yuchung Chengddf1af62016-02-02 10:33:06 -08001870
Eric Dumazet19119f22018-11-27 14:42:01 -08001871 tp->sacked_out += num_dupack;
1872 tcp_check_reno_reordering(sk, 0);
1873 delivered = tp->sacked_out - prior_sacked;
1874 if (delivered > 0)
1875 tp->delivered += delivered;
1876 tcp_verify_left_out(tp);
1877 }
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001878}
1879
1880/* Account for ACK, ACKing some data in Reno Recovery phase. */
1881
1882static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1883{
1884 struct tcp_sock *tp = tcp_sk(sk);
1885
1886 if (acked > 0) {
1887 /* One ACK acked hole. The rest eat duplicate ACKs. */
Yuchung Chengddf1af62016-02-02 10:33:06 -08001888 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001889 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001890 tp->sacked_out = 0;
1891 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001892 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001893 }
1894 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001895 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001896}
1897
1898static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1899{
1900 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001901}
1902
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001903void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001908 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001909 tp->sacked_out = 0;
1910}
1911
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001912static inline void tcp_init_undo(struct tcp_sock *tp)
1913{
1914 tp->undo_marker = tp->snd_una;
1915 /* Retransmission still in flight may cause DSACKs later. */
1916 tp->undo_retrans = tp->retrans_out ? : -1;
1917}
1918
Yuchung Chengb8fef652018-05-16 16:40:16 -07001919static bool tcp_is_rack(const struct sock *sk)
1920{
1921 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1922}
1923
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001924/* If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 * and reset tags completely, otherwise preserve SACKs. If receiver
1926 * dropped its ofo queue, we will know this due to reneging detection.
1927 */
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001928static void tcp_timeout_mark_lost(struct sock *sk)
1929{
1930 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07001931 struct sk_buff *skb, *head;
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001932 bool is_reneg; /* is receiver reneging on SACKs? */
1933
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07001934 head = tcp_rtx_queue_head(sk);
1935 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001936 if (is_reneg) {
1937 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1938 tp->sacked_out = 0;
1939 /* Mark SACK reneging until we recover from this loss event. */
1940 tp->is_sack_reneg = 1;
1941 } else if (tcp_is_reno(tp)) {
1942 tcp_reset_reno_sack(tp);
1943 }
1944
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07001945 skb = head;
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001946 skb_rbtree_walk_from(skb) {
1947 if (is_reneg)
1948 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07001949 else if (tcp_is_rack(sk) && skb != head &&
1950 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1951 continue; /* Don't mark recently sent ones lost yet */
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07001952 tcp_mark_skb_lost(sk, skb);
1953 }
1954 tcp_verify_left_out(tp);
1955 tcp_clear_all_retrans_hints(tp);
1956}
1957
1958/* Enter Loss state. */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001959void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001961 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov1043e252016-02-03 09:46:52 +02001963 struct net *net = sock_net(sk);
Yuchung Chengcc663f42017-04-07 11:42:05 -07001964 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965
Yuchung Chengc77d62f2018-05-16 16:40:15 -07001966 tcp_timeout_mark_lost(sk);
1967
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001969 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1970 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001971 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1972 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Yuchung Cheng4faf7832017-08-03 20:38:51 -07001973 tp->prior_cwnd = tp->snd_cwnd;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001974 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1975 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001976 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 }
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07001978 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07001980 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001982 /* Timeout in disordered state after receiving substantial DUPACKs
1983 * suggests that the degree of reordering is over-estimated.
1984 */
1985 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02001986 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001987 tp->reordering = min_t(unsigned int, tp->reordering,
Nikolay Borisov1043e252016-02-03 09:46:52 +02001988 net->ipv4.sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001989 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02001991 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001992
Yuchung Chengcc663f42017-04-07 11:42:05 -07001993 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1994 * loss recovery is underway except recurring timeout(s) on
1995 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
Yuchung Chenge33099f2013-03-20 13:33:00 +00001996 */
Eric Dumazetaf9b69a2017-10-26 21:55:10 -07001997 tp->frto = net->ipv4.sysctl_tcp_frto &&
Yuchung Chengcc663f42017-04-07 11:42:05 -07001998 (new_recovery || icsk->icsk_retransmits) &&
1999 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000}
2001
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002002/* If ACK arrived pointing to a remembered SACK, it means that our
2003 * remembered SACKs do not reflect real state of receiver i.e.
2004 * receiver _host_ is heavily congested (or buggy).
2005 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002006 * To avoid big spurious retransmission bursts due to transient SACK
2007 * scoreboard oddities that look like reneging, we give the receiver a
2008 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2009 * restore sanity to the SACK scoreboard. If the apparent reneging
2010 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002011 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002012static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002014 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002015 struct tcp_sock *tp = tcp_sk(sk);
2016 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2017 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002019 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002020 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002021 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002023 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024}
2025
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002026/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2027 * counter when SACK is enabled (without SACK, sacked_out is used for
2028 * that purpose).
2029 *
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002030 * With reordering, holes may still be in flight, so RFC3517 recovery
2031 * uses pure sacked_out (total number of SACKed segments) even though
2032 * it violates the RFC that uses duplicate ACKs, often these are equal
2033 * but when e.g. out-of-window ACKs or packet duplication occurs,
2034 * they differ. Since neither occurs due to loss, TCP should really
2035 * ignore them.
2036 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002037static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002038{
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002039 return tp->sacked_out + 1;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002040}
2041
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002042/* Linux NewReno/SACK/ECN state machine.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 * --------------------------------------
2044 *
2045 * "Open" Normal state, no dubious events, fast path.
2046 * "Disorder" In all the respects it is "Open",
2047 * but requires a bit more attention. It is entered when
2048 * we see some SACKs or dupacks. It is split of "Open"
2049 * mainly to move some processing from fast path to slow one.
2050 * "CWR" CWND was reduced due to some Congestion Notification event.
2051 * It can be ECN, ICMP source quench, local device congestion.
2052 * "Recovery" CWND was reduced, we are fast-retransmitting.
2053 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2054 *
2055 * tcp_fastretrans_alert() is entered:
2056 * - each incoming ACK, if state is not "Open"
2057 * - when arrived ACK is unusual, namely:
2058 * * SACK
2059 * * Duplicate ACK.
2060 * * ECN ECE.
2061 *
2062 * Counting packets in flight is pretty simple.
2063 *
2064 * in_flight = packets_out - left_out + retrans_out
2065 *
2066 * packets_out is SND.NXT-SND.UNA counted in packets.
2067 *
2068 * retrans_out is number of retransmitted segments.
2069 *
2070 * left_out is number of segments left network, but not ACKed yet.
2071 *
2072 * left_out = sacked_out + lost_out
2073 *
2074 * sacked_out: Packets, which arrived to receiver out of order
2075 * and hence not ACKed. With SACKs this number is simply
2076 * amount of SACKed data. Even without SACKs
2077 * it is easy to give pretty reliable estimate of this number,
2078 * counting duplicate ACKs.
2079 *
2080 * lost_out: Packets lost by network. TCP has no explicit
2081 * "loss notification" feedback from network (for now).
2082 * It means that this number can be only _guessed_.
2083 * Actually, it is the heuristics to predict lossage that
2084 * distinguishes different algorithms.
2085 *
2086 * F.e. after RTO, when all the queue is considered as lost,
2087 * lost_out = packets_out and in_flight = retrans_out.
2088 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002089 * Essentially, we have now a few algorithms detecting
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 * lost packets.
2091 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002092 * If the receiver supports SACK:
2093 *
2094 * RFC6675/3517: It is the conventional algorithm. A packet is
2095 * considered lost if the number of higher sequence packets
2096 * SACKed is greater than or equal the DUPACK thoreshold
2097 * (reordering). This is implemented in tcp_mark_head_lost and
2098 * tcp_update_scoreboard.
2099 *
2100 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2101 * (2017-) that checks timing instead of counting DUPACKs.
2102 * Essentially a packet is considered lost if it's not S/ACKed
2103 * after RTT + reordering_window, where both metrics are
2104 * dynamically measured and adjusted. This is implemented in
2105 * tcp_rack_mark_lost.
2106 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002107 * If the receiver does not support SACK:
2108 *
2109 * NewReno (RFC6582): in Recovery we assume that one segment
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 * is lost (classic Reno). While we are in Recovery and
2111 * a partial ACK arrives, we assume that one more packet
2112 * is lost (NewReno). This heuristics are the same in NewReno
2113 * and SACK.
2114 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 * Really tricky (and requiring careful tuning) part of algorithm
2116 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2117 * The first determines the moment _when_ we should reduce CWND and,
2118 * hence, slow down forward transmission. In fact, it determines the moment
2119 * when we decide that hole is caused by loss, rather than by a reorder.
2120 *
2121 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2122 * holes, caused by lost packets.
2123 *
2124 * And the most logically complicated part of algorithm is undo
2125 * heuristics. We detect false retransmits due to both too early
2126 * fast retransmit (reordering) and underestimated RTO, analyzing
2127 * timestamps and D-SACKs. When we detect that some segments were
2128 * retransmitted by mistake and CWND reduction was wrong, we undo
2129 * window reduction and abort recovery phase. This logic is hidden
2130 * inside several functions named tcp_try_undo_<something>.
2131 */
2132
2133/* This function decides, when we should leave Disordered state
2134 * and enter Recovery phase, reducing congestion window.
2135 *
2136 * Main question: may we further continue forward transmission
2137 * with the same cwnd?
2138 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002139static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002141 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142
2143 /* Trick#1: The loss is proven. */
2144 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002145 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146
2147 /* Not-A-Trick#2 : Classic rule... */
Yuchung Chengb38a51f2018-05-16 16:40:11 -07002148 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002149 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002151 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152}
2153
Yuchung Cheng974c1232012-01-19 14:42:21 +00002154/* Detect loss in event "A" above by marking head of queue up as lost.
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002155 * For non-SACK(Reno) senders, the first "packets" number of segments
Yuchung Cheng974c1232012-01-19 14:42:21 +00002156 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2157 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2158 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002159 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002160static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002162 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 struct sk_buff *skb;
Neal Cardwelld88270e2016-01-25 14:01:53 -08002164 int cnt, oldcnt, lost;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002165 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002166 /* Use SACK to deduce losses of new sequences sent during recovery */
2167 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002169 WARN_ON(packets > tp->packets_out);
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002170 skb = tp->lost_skb_hint;
2171 if (skb) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002172 /* Head already handled? */
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002173 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002174 return;
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002175 cnt = tp->lost_cnt_hint;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002176 } else {
Eric Dumazet75c119a2017-10-05 22:21:27 -07002177 skb = tcp_rtx_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002178 cnt = 0;
2179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
Eric Dumazet75c119a2017-10-05 22:21:27 -07002181 skb_rbtree_walk_from(skb) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002182 /* TODO: do this better */
2183 /* this is not the most efficient way to do this... */
2184 tp->lost_skb_hint = skb;
2185 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002186
Yuchung Cheng974c1232012-01-19 14:42:21 +00002187 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002188 break;
2189
2190 oldcnt = cnt;
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002191 if (tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002192 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2193 cnt += tcp_skb_pcount(skb);
2194
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002195 if (cnt > packets) {
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002196 if (tcp_is_sack(tp) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002197 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de75592010-09-24 13:22:06 +00002198 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002199 break;
2200
Eric Dumazetf69ad292015-06-11 09:15:18 -07002201 mss = tcp_skb_mss(skb);
Neal Cardwelld88270e2016-01-25 14:01:53 -08002202 /* If needed, chop off the prefix to mark as lost. */
2203 lost = (packets - oldcnt) * mss;
2204 if (lost < skb->len &&
Eric Dumazet75c119a2017-10-05 22:21:27 -07002205 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2206 lost, mss, GFP_ATOMIC) < 0)
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002207 break;
2208 cnt = packets;
2209 }
2210
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002211 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002212
2213 if (mark_head)
2214 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002216 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217}
2218
2219/* Account newly detected lost packet(s) */
2220
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002221static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002223 struct tcp_sock *tp = tcp_sk(sk);
2224
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002225 if (tcp_is_sack(tp)) {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002226 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002227 if (sacked_upto >= 0)
2228 tcp_mark_head_lost(sk, sacked_upto, 0);
2229 else if (fast_rexmit)
2230 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232}
2233
Yuchung Cheng77c63122015-10-16 21:57:44 -07002234static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2235{
2236 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2237 before(tp->rx_opt.rcv_tsecr, when);
2238}
2239
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002240/* skb is spurious retransmitted if the returned timestamp echo
2241 * reply is prior to the skb transmission time
2242 */
2243static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2244 const struct sk_buff *skb)
2245{
2246 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2247 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2248}
2249
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250/* Nothing was retransmitted or returned timestamp is less
2251 * than timestamp of the first retransmission.
2252 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002253static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254{
2255 return !tp->retrans_stamp ||
Yuchung Cheng77c63122015-10-16 21:57:44 -07002256 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257}
2258
2259/* Undo procedures. */
2260
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002261/* We can clear retrans_stamp when there are no retransmissions in the
2262 * window. It would seem that it is trivially available for us in
2263 * tp->retrans_out, however, that kind of assumptions doesn't consider
2264 * what will happen if errors occur when sending retransmission for the
2265 * second time. ...It could the that such segment has only
2266 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2267 * the head skb is enough except for some reneging corner cases that
2268 * are not worth the effort.
2269 *
2270 * Main reason for all this complexity is the fact that connection dying
2271 * time now depends on the validity of the retrans_stamp, in particular,
2272 * that successive retransmissions of a segment must not advance
2273 * retrans_stamp under any conditions.
2274 */
2275static bool tcp_any_retrans_done(const struct sock *sk)
2276{
2277 const struct tcp_sock *tp = tcp_sk(sk);
2278 struct sk_buff *skb;
2279
2280 if (tp->retrans_out)
2281 return true;
2282
Eric Dumazet75c119a2017-10-05 22:21:27 -07002283 skb = tcp_rtx_queue_head(sk);
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002284 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2285 return true;
2286
2287 return false;
2288}
2289
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002290static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291{
Joe Perches39347882017-09-10 19:02:25 -07002292#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002293 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002295
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002296 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002297 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2298 msg,
2299 &inet->inet_daddr, ntohs(inet->inet_dport),
2300 tp->snd_cwnd, tcp_left_out(tp),
2301 tp->snd_ssthresh, tp->prior_ssthresh,
2302 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002303 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002304#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002305 else if (sk->sk_family == AF_INET6) {
Joe Perches91df42b2012-05-15 14:11:54 +00002306 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2307 msg,
Eric Dumazet019b1c92016-09-22 17:54:00 -07002308 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
Joe Perches91df42b2012-05-15 14:11:54 +00002309 tp->snd_cwnd, tcp_left_out(tp),
2310 tp->snd_ssthresh, tp->prior_ssthresh,
2311 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002312 }
2313#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314#endif
Joe Perches39347882017-09-10 19:02:25 -07002315}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316
Yuchung Cheng7026b912013-05-29 14:20:13 +00002317static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002319 struct tcp_sock *tp = tcp_sk(sk);
2320
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002321 if (unmark_loss) {
2322 struct sk_buff *skb;
2323
Eric Dumazet75c119a2017-10-05 22:21:27 -07002324 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002325 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2326 }
2327 tp->lost_out = 0;
2328 tcp_clear_all_retrans_hints(tp);
2329 }
2330
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002332 const struct inet_connection_sock *icsk = inet_csk(sk);
2333
Florian Westphale9799182016-11-21 14:18:38 +01002334 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
Yuchung Cheng7026b912013-05-29 14:20:13 +00002336 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002338 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 }
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002341 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002342 tp->undo_marker = 0;
Yuchung Chengcd1fc852017-12-07 11:33:31 -08002343 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344}
2345
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002346static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002348 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349}
2350
2351/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002352static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002354 struct tcp_sock *tp = tcp_sk(sk);
2355
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002357 int mib_idx;
2358
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 /* Happy end! We did not retransmit anything
2360 * or our original transmission succeeded.
2361 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002362 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002363 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002364 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002365 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002367 mib_idx = LINUX_MIB_TCPFULLUNDO;
2368
Eric Dumazetc10d9312016-04-29 14:16:47 -07002369 NET_INC_STATS(sock_net(sk), mib_idx);
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002370 } else if (tp->rack.reo_wnd_persist) {
2371 tp->rack.reo_wnd_persist--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002373 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 /* Hold old state until something *above* high_seq
2375 * is ACKed. For Reno it is MUST to prevent false
2376 * fast retransmits (RFC2582). SACK TCP is safe. */
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002377 if (!tcp_any_retrans_done(sk))
2378 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002379 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002381 tcp_set_ca_state(sk, TCP_CA_Open);
Yousuk Seungd4761752017-12-07 13:41:34 -08002382 tp->is_sack_reneg = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002383 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384}
2385
2386/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002387static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002389 struct tcp_sock *tp = tcp_sk(sk);
2390
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 if (tp->undo_marker && !tp->undo_retrans) {
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002392 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2393 tp->rack.reo_wnd_persist + 1);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002394 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002395 tcp_undo_cwnd_reduction(sk, false);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002396 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002397 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002399 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400}
2401
Yuchung Chenge33099f2013-03-20 13:33:00 +00002402/* Undo during loss recovery after partial ACK or using F-RTO. */
2403static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002405 struct tcp_sock *tp = tcp_sk(sk);
2406
Yuchung Chenge33099f2013-03-20 13:33:00 +00002407 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002408 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002409
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002410 DBGUNDO(sk, "partial loss");
Eric Dumazetc10d9312016-04-29 14:16:47 -07002411 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002412 if (frto_undo)
Eric Dumazetc10d9312016-04-29 14:16:47 -07002413 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002414 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002415 inet_csk(sk)->icsk_retransmits = 0;
Yousuk Seungd4761752017-12-07 13:41:34 -08002416 if (frto_undo || tcp_is_sack(tp)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002417 tcp_set_ca_state(sk, TCP_CA_Open);
Yousuk Seungd4761752017-12-07 13:41:34 -08002418 tp->is_sack_reneg = 0;
2419 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002420 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002422 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423}
2424
Yuchung Cheng37598242015-07-01 14:11:15 -07002425/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002426 * It computes the number of packets to send (sndcnt) based on packets newly
2427 * delivered:
2428 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2429 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002430 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2431 * But when the retransmits are acked without further losses, PRR
2432 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002433 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002434static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002435{
2436 struct tcp_sock *tp = tcp_sk(sk);
2437
2438 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002439 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002440 tp->snd_cwnd_cnt = 0;
2441 tp->prior_cwnd = tp->snd_cwnd;
2442 tp->prr_delivered = 0;
2443 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002444 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002445 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002446}
2447
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002448void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002449{
2450 struct tcp_sock *tp = tcp_sk(sk);
2451 int sndcnt = 0;
2452 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2453
Yuchung Cheng8b8a3212016-01-06 12:42:38 -08002454 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2455 return;
2456
Yuchung Cheng684bad12012-09-02 17:38:04 +00002457 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002458 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002459 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2460 tp->prior_cwnd - 1;
2461 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yafang Shao5e13a0d2018-11-11 20:10:10 +08002462 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2463 FLAG_RETRANS_DATA_ACKED) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002464 sndcnt = min_t(int, delta,
2465 max_t(int, tp->prr_delivered - tp->prr_out,
2466 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002467 } else {
2468 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002469 }
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002470 /* Force a fast retransmit upon entering fast recovery */
2471 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002472 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2473}
2474
Yuchung Cheng684bad12012-09-02 17:38:04 +00002475static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002477 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002478
Yuchung Chengc0402762016-09-19 23:39:21 -04002479 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2480 return;
2481
Yuchung Cheng684bad12012-09-02 17:38:04 +00002482 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
Yuchung Chenged254972017-08-01 13:22:32 -07002483 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2484 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002485 tp->snd_cwnd = tp->snd_ssthresh;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002486 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002487 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002488 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489}
2490
Yuchung Cheng684bad12012-09-02 17:38:04 +00002491/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002492void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002493{
2494 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002495
2496 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002497 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002498 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002499 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002500 tcp_set_ca_state(sk, TCP_CA_CWR);
2501 }
2502}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002503EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002504
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002505static void tcp_try_keep_open(struct sock *sk)
2506{
2507 struct tcp_sock *tp = tcp_sk(sk);
2508 int state = TCP_CA_Open;
2509
Neal Cardwellf6982042011-11-16 08:58:04 +00002510 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002511 state = TCP_CA_Disorder;
2512
2513 if (inet_csk(sk)->icsk_ca_state != state) {
2514 tcp_set_ca_state(sk, state);
2515 tp->high_seq = tp->snd_nxt;
2516 }
2517}
2518
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002519static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002521 struct tcp_sock *tp = tcp_sk(sk);
2522
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002523 tcp_verify_left_out(tp);
2524
Yuchung Cheng9b441902013-03-20 13:32:58 +00002525 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 tp->retrans_stamp = 0;
2527
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002528 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002529 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002531 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002532 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 }
2534}
2535
John Heffner5d424d52006-03-20 17:53:41 -08002536static void tcp_mtup_probe_failed(struct sock *sk)
2537{
2538 struct inet_connection_sock *icsk = inet_csk(sk);
2539
2540 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2541 icsk->icsk_mtup.probe_size = 0;
Eric Dumazetc10d9312016-04-29 14:16:47 -07002542 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002543}
2544
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002545static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002546{
2547 struct tcp_sock *tp = tcp_sk(sk);
2548 struct inet_connection_sock *icsk = inet_csk(sk);
2549
2550 /* FIXME: breaks with very large cwnd */
2551 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2552 tp->snd_cwnd = tp->snd_cwnd *
2553 tcp_mss_to_mtu(sk, tp->mss_cache) /
2554 icsk->icsk_mtup.probe_size;
2555 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002556 tp->snd_cwnd_stamp = tcp_jiffies32;
John Heffner9c6d5e52010-10-06 21:18:02 -07002557 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002558
2559 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2560 icsk->icsk_mtup.probe_size = 0;
2561 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002563}
2564
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002565/* Do a simple retransmit without using the backoff mechanisms in
2566 * tcp_timer. This is used for path mtu discovery.
2567 * The socket is already locked here.
2568 */
2569void tcp_simple_retransmit(struct sock *sk)
2570{
2571 const struct inet_connection_sock *icsk = inet_csk(sk);
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002574 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002575
Eric Dumazet75c119a2017-10-05 22:21:27 -07002576 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002577 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002578 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2579 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2580 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2581 tp->retrans_out -= tcp_skb_pcount(skb);
2582 }
2583 tcp_skb_mark_lost_uncond_verify(tp, skb);
2584 }
2585 }
2586
2587 tcp_clear_retrans_hints_partial(tp);
2588
Yuchung Cheng0eb96bf2017-11-07 15:33:43 -08002589 if (!tp->lost_out)
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002590 return;
2591
2592 if (tcp_is_reno(tp))
2593 tcp_limit_reno_sacked(tp);
2594
2595 tcp_verify_left_out(tp);
2596
2597 /* Don't muck with the congestion window here.
2598 * Reason is that we do not increase amount of _data_
2599 * in network, but units changed and effective
2600 * cwnd/ssthresh really reduced now.
2601 */
2602 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2603 tp->high_seq = tp->snd_nxt;
2604 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2605 tp->prior_ssthresh = 0;
2606 tp->undo_marker = 0;
2607 tcp_set_ca_state(sk, TCP_CA_Loss);
2608 }
2609 tcp_xmit_retransmit_queue(sk);
2610}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002611EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002612
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002613void tcp_enter_recovery(struct sock *sk, bool ece_ack)
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002614{
2615 struct tcp_sock *tp = tcp_sk(sk);
2616 int mib_idx;
2617
2618 if (tcp_is_reno(tp))
2619 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2620 else
2621 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2622
Eric Dumazetc10d9312016-04-29 14:16:47 -07002623 NET_INC_STATS(sock_net(sk), mib_idx);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002624
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002625 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002626 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002627
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002628 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002629 if (!ece_ack)
2630 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002631 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002632 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002633 tcp_set_ca_state(sk, TCP_CA_Recovery);
2634}
2635
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002636/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2637 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2638 */
Eric Dumazet19119f22018-11-27 14:42:01 -08002639static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
Yuchung Chenge662ca42016-02-02 10:33:04 -08002640 int *rexmit)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002641{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002642 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002643 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002644
Yuchung Chengda34ac72015-05-18 12:31:44 -07002645 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2646 tcp_try_undo_loss(sk, false))
2647 return;
2648
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002649 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Chengfc68e172018-02-27 14:15:02 -08002650 /* Step 3.b. A timeout is spurious if not all data are
2651 * lost, i.e., never-retransmitted data are (s)acked.
2652 */
2653 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2654 tcp_try_undo_loss(sk, true))
2655 return;
2656
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002657 if (after(tp->snd_nxt, tp->high_seq)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08002658 if (flag & FLAG_DATA_SACKED || num_dupack)
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002659 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002660 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2661 tp->high_seq = tp->snd_nxt;
Yuchung Chenge662ca42016-02-02 10:33:04 -08002662 /* Step 2.b. Try send new data (but deferred until cwnd
2663 * is updated in tcp_ack()). Otherwise fall back to
2664 * the conventional recovery.
2665 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07002666 if (!tcp_write_queue_empty(sk) &&
Yuchung Chenge662ca42016-02-02 10:33:04 -08002667 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2668 *rexmit = REXMIT_NEW;
2669 return;
2670 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002671 tp->frto = 0;
2672 }
2673 }
2674
2675 if (recovered) {
2676 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002677 tcp_try_undo_recovery(sk);
2678 return;
2679 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002680 if (tcp_is_reno(tp)) {
2681 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2682 * delivered. Lower inflight to clock out (re)tranmissions.
2683 */
Eric Dumazet19119f22018-11-27 14:42:01 -08002684 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2685 tcp_add_reno_sack(sk, num_dupack);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002686 else if (flag & FLAG_SND_UNA_ADVANCED)
2687 tcp_reset_reno_sack(tp);
2688 }
Yuchung Chenge662ca42016-02-02 10:33:04 -08002689 *rexmit = REXMIT_LOST;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002690}
2691
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002692/* Undo during fast recovery after partial ACK. */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002693static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002694{
2695 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002696
Yuchung Cheng7026b912013-05-29 14:20:13 +00002697 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002698 /* Plain luck! Hole if filled with delayed
Yuchung Cheng737ff312017-11-08 13:01:27 -08002699 * packet, rather than with a retransmit. Check reordering.
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002700 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002701 tcp_check_sack_reordering(sk, prior_snd_una, 1);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002702
2703 /* We are getting evidence that the reordering degree is higher
2704 * than we realized. If there are no retransmits out then we
2705 * can undo. Otherwise we clock out new packets but do not
2706 * mark more packets lost or retransmit more.
2707 */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002708 if (tp->retrans_out)
Yuchung Cheng7026b912013-05-29 14:20:13 +00002709 return true;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002710
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002711 if (!tcp_any_retrans_done(sk))
2712 tp->retrans_stamp = 0;
2713
Yuchung Cheng7026b912013-05-29 14:20:13 +00002714 DBGUNDO(sk, "partial recovery");
2715 tcp_undo_cwnd_reduction(sk, true);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002716 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002717 tcp_try_keep_open(sk);
2718 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002719 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002720 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002721}
2722
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002723static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002724{
2725 struct tcp_sock *tp = tcp_sk(sk);
2726
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002727 if (tcp_rtx_queue_empty(sk))
2728 return;
2729
2730 if (unlikely(tcp_is_reno(tp))) {
2731 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2732 } else if (tcp_is_rack(sk)) {
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002733 u32 prior_retrans = tp->retrans_out;
2734
Eric Dumazet128eda82017-04-25 10:15:34 -07002735 tcp_rack_mark_lost(sk);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002736 if (prior_retrans > tp->retrans_out)
2737 *ack_flag |= FLAG_LOST_RETRANS;
2738 }
2739}
2740
Yuchung Cheng737ff312017-11-08 13:01:27 -08002741static bool tcp_force_fast_retransmit(struct sock *sk)
2742{
2743 struct tcp_sock *tp = tcp_sk(sk);
2744
2745 return after(tcp_highest_sack_seq(tp),
2746 tp->snd_una + tp->reordering * tp->mss_cache);
2747}
2748
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749/* Process an event, which can update packets-in-flight not trivially.
2750 * Main goal of this function is to calculate new estimate for left_out,
2751 * taking into account both packets sitting in receiver's buffer and
2752 * packets lost by network.
2753 *
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002754 * Besides that it updates the congestion state when packet loss or ECN
2755 * is detected. But it does not reduce the cwnd, it is done by the
2756 * congestion control later.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 *
2758 * It does _not_ decide what to send, it is made in function
2759 * tcp_xmit_retransmit_queue().
2760 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002761static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
Eric Dumazet19119f22018-11-27 14:42:01 -08002762 int num_dupack, int *ack_flag, int *rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002764 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002766 int fast_rexmit = 0, flag = *ack_flag;
Eric Dumazet19119f22018-11-27 14:42:01 -08002767 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2768 tcp_force_fast_retransmit(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769
Eric Dumazet8ba6dda2017-10-05 22:21:25 -07002770 if (!tp->packets_out && tp->sacked_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 tp->sacked_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002773 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002775 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 tp->prior_ssthresh = 0;
2777
2778 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002779 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 return;
2781
Yuchung Cheng974c1232012-01-19 14:42:21 +00002782 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002783 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
Yuchung Cheng974c1232012-01-19 14:42:21 +00002785 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002787 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002788 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 tp->retrans_stamp = 0;
2790 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002791 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 case TCP_CA_CWR:
2793 /* CWR is to be held something *above* high_seq
2794 * is ACKed for CWR bit to reach receiver. */
2795 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002796 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002797 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 }
2799 break;
2800
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002802 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002804 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002806 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 break;
2808 }
2809 }
2810
Yuchung Cheng974c1232012-01-19 14:42:21 +00002811 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002812 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002814 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08002815 if (tcp_is_reno(tp))
2816 tcp_add_reno_sack(sk, num_dupack);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002817 } else {
Yuchung Cheng737ff312017-11-08 13:01:27 -08002818 if (tcp_try_undo_partial(sk, prior_snd_una))
Yuchung Cheng7026b912013-05-29 14:20:13 +00002819 return;
2820 /* Partial ACK arrived. Force fast retransmit. */
2821 do_lost = tcp_is_reno(tp) ||
Yuchung Cheng737ff312017-11-08 13:01:27 -08002822 tcp_force_fast_retransmit(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002823 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002824 if (tcp_try_undo_dsack(sk)) {
2825 tcp_try_keep_open(sk);
2826 return;
2827 }
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002828 tcp_identify_packet_loss(sk, ack_flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 break;
2830 case TCP_CA_Loss:
Eric Dumazet19119f22018-11-27 14:42:01 -08002831 tcp_process_loss(sk, flag, num_dupack, rexmit);
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002832 tcp_identify_packet_loss(sk, ack_flag);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002833 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2834 (*ack_flag & FLAG_LOST_RETRANS)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002836 /* Change state if cwnd is undone or retransmits are lost */
Gustavo A. R. Silvafcfd6df2017-10-16 15:48:55 -05002837 /* fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002839 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002840 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 tcp_reset_reno_sack(tp);
Eric Dumazet19119f22018-11-27 14:42:01 -08002842 tcp_add_reno_sack(sk, num_dupack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 }
2844
Neal Cardwellf6982042011-11-16 08:58:04 +00002845 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002846 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002848 tcp_identify_packet_loss(sk, ack_flag);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002849 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002850 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 return;
2852 }
2853
John Heffner5d424d52006-03-20 17:53:41 -08002854 /* MTU probe failure: don't reduce cwnd */
2855 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2856 icsk->icsk_mtup.probe_size &&
John Heffner0e7b13682006-03-20 21:32:58 -08002857 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002858 tcp_mtup_probe_failed(sk);
2859 /* Restores the reduction we did in tcp_mtup_probe() */
2860 tp->snd_cwnd++;
2861 tcp_simple_retransmit(sk);
2862 return;
2863 }
2864
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002866 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002867 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 }
2869
Yuchung Chengb38a51f2018-05-16 16:40:11 -07002870 if (!tcp_is_rack(sk) && do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002871 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08002872 *rexmit = REXMIT_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873}
2874
Yuchung Chengeb36be02018-01-17 12:11:00 -08002875static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
Yuchung Chengf6722582015-10-16 21:57:42 -07002876{
Eric Dumazetbd239702017-10-27 07:47:28 -07002877 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
Neal Cardwell64033892016-09-19 23:39:10 -04002878 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengf6722582015-10-16 21:57:42 -07002879
Yuchung Chengeb36be02018-01-17 12:11:00 -08002880 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2881 /* If the remote keeps returning delayed ACKs, eventually
2882 * the min filter would pick it up and overestimate the
2883 * prop. delay when it expires. Skip suspected delayed ACKs.
2884 */
2885 return;
2886 }
Eric Dumazetac9517f2017-05-16 14:00:13 -07002887 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
Neal Cardwell64033892016-09-19 23:39:10 -04002888 rtt_us ? : jiffies_to_usecs(1));
Yuchung Chengf6722582015-10-16 21:57:42 -07002889}
2890
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002891static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2892 long seq_rtt_us, long sack_rtt_us,
2893 long ca_rtt_us, struct rate_sample *rs)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002894{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002895 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002896
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002897 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2898 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2899 * Karn's algorithm forbids taking RTT if some retransmitted data
2900 * is acked (RFC6298).
2901 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002902 if (seq_rtt_us < 0)
2903 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07002904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 /* RTTM Rule: A TSecr value received in a segment is used to
2906 * update the averaged RTT measurement only if the segment
2907 * acknowledges some new data, i.e., only if it advances the
2908 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002911 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Eric Dumazet9a568de2017-05-16 14:00:14 -07002912 flag & FLAG_ACKED) {
2913 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
Eric Dumazet9a568de2017-05-16 14:00:14 -07002914
Eric Dumazet9efdda42018-11-24 09:12:24 -08002915 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2916 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2917 ca_rtt_us = seq_rtt_us;
2918 }
Eric Dumazet9a568de2017-05-16 14:00:14 -07002919 }
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002920 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002921 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002922 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923
Yuchung Chengf6722582015-10-16 21:57:42 -07002924 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2925 * always taken together with ACK, SACK, or TS-opts. Any negative
2926 * values will be skipped with the seq_rtt_us < 0 check above.
2927 */
Yuchung Chengeb36be02018-01-17 12:11:00 -08002928 tcp_update_rtt_min(sk, ca_rtt_us, flag);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002929 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002930 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002932 /* RFC6298: only reset backoff on valid RTT measurement. */
2933 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002934 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935}
2936
Yuchung Cheng375fe022013-07-22 16:20:45 -07002937/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002938void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002939{
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002940 struct rate_sample rs;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002941 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07002942
Eric Dumazet9a568de2017-05-16 14:00:14 -07002943 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2944 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002945
Yuchung Cheng775e68a2017-05-31 11:30:53 -07002946 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002947}
2948
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002949
Eric Dumazet24901552014-05-02 21:18:05 -07002950static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002952 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07002953
2954 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002955 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956}
2957
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958/* Restart timer after forward progress on connection.
2959 * RFC2988 recommends to restart timer to now+rto.
2960 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002961void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002963 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002964 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002965
Jerry Chu168a8f52012-08-31 12:29:13 +00002966 /* If the retrans timer is currently being used by Fast Open
2967 * for SYN-ACK retrans purpose, stay put.
2968 */
2969 if (tp->fastopen_rsk)
2970 return;
2971
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002973 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002975 u32 rto = inet_csk(sk)->icsk_rto;
2976 /* Offset the time elapsed after installing regular RTO */
Yuchung Chengbec41a12017-01-12 22:11:39 -08002977 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002978 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Neal Cardwelle1a10ef2017-08-03 09:19:52 -04002979 s64 delta_us = tcp_rto_delta_us(sk);
Eric Dumazetb17b8a22017-05-18 09:15:58 -07002980 /* delta_us may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002981 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002982 */
Neal Cardwellcdbeb632017-08-16 17:53:36 -04002983 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002984 }
Eric Dumazet3f80e082018-10-23 11:54:16 -07002985 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2986 TCP_RTO_MAX, tcp_rtx_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002988}
2989
Neal Cardwelldf92c832017-08-03 09:19:54 -04002990/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2991static void tcp_set_xmit_timer(struct sock *sk)
2992{
Neal Cardwelled66dfa2017-11-17 21:06:14 -05002993 if (!tcp_schedule_loss_probe(sk, true))
Neal Cardwelldf92c832017-08-03 09:19:54 -04002994 tcp_rearm_rto(sk);
2995}
2996
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002997/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002998static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999{
3000 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003001 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003003 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
3005 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003006 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 return 0;
3008 packets_acked -= tcp_skb_pcount(skb);
3009
3010 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003012 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 }
3014
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003015 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016}
3017
Eric Dumazetad971f62014-10-11 15:17:29 -07003018static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3019 u32 prior_snd_una)
3020{
3021 const struct skb_shared_info *shinfo;
3022
3023 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
Soheil Hassas Yeganeh6b084922016-04-02 23:08:08 -04003024 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
Eric Dumazetad971f62014-10-11 15:17:29 -07003025 return;
3026
3027 shinfo = skb_shinfo(skb);
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04003028 if (!before(shinfo->tskey, prior_snd_una) &&
Eric Dumazete2080072017-10-04 12:59:58 -07003029 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3030 tcp_skb_tsorted_save(skb) {
3031 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3032 } tcp_skb_tsorted_restore(skb);
3033 }
Eric Dumazetad971f62014-10-11 15:17:29 -07003034}
3035
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003036/* Remove acknowledged frames from the retransmission queue. If our packet
3037 * is before the ack sequence we can discard it as it's confirmed to have
3038 * arrived at the other end.
3039 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003040static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3041 u32 prior_snd_una,
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003042 struct tcp_sacktag_state *sack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003044 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -07003045 u64 first_ackt, last_ackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003046 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003047 u32 prior_sacked = tp->sacked_out;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003048 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003049 struct sk_buff *skb, *next;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003050 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003051 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003052 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003053 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003054 u32 pkts_acked = 0;
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003055 u32 last_in_flight = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003056 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003057 int flag = 0;
3058
Eric Dumazet9a568de2017-05-16 14:00:14 -07003059 first_ackt = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060
Eric Dumazet75c119a2017-10-05 22:21:27 -07003061 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003062 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Yuchung Cheng737ff312017-11-08 13:01:27 -08003063 const u32 start_seq = scb->seq;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003064 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003065 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066
Eric Dumazetad971f62014-10-11 15:17:29 -07003067 tcp_ack_tstamp(sk, skb, prior_snd_una);
Willem de Bruijn712a7222014-08-12 14:53:16 -04003068
Gavin McCullagh2072c222007-12-29 19:11:21 -08003069 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003071 if (tcp_skb_pcount(skb) == 1 ||
3072 !after(tp->snd_una, scb->seq))
3073 break;
3074
Ilpo Järvinen72018832007-12-30 04:37:55 -08003075 acked_pcount = tcp_tso_acked(sk, skb);
3076 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003077 break;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003078 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003079 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003080 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 }
3082
Eric Dumazetad971f62014-10-11 15:17:29 -07003083 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003084 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003085 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003086 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003087 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003088 last_ackt = tcp_skb_timestamp_us(skb);
Eric Dumazet9a568de2017-05-16 14:00:14 -07003089 WARN_ON_ONCE(last_ackt == 0);
3090 if (!first_ackt)
Eric Dumazet740b0f12014-02-26 14:02:48 -08003091 first_ackt = last_ackt;
3092
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003093 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003094 if (before(start_seq, reord))
3095 reord = start_seq;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003096 if (!after(scb->end_seq, tp->high_seq))
3097 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003098 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003099
Yuchung Chengddf1af62016-02-02 10:33:06 -08003100 if (sacked & TCPCB_SACKED_ACKED) {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003101 tp->sacked_out -= acked_pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003102 } else if (tcp_is_sack(tp)) {
3103 tp->delivered += acked_pcount;
3104 if (!tcp_skb_spurious_retrans(tp, skb))
Yuchung Cheng1d0833d2017-01-12 22:11:34 -08003105 tcp_rack_advance(tp, sacked, scb->end_seq,
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003106 tcp_skb_timestamp_us(skb));
Yuchung Chengddf1af62016-02-02 10:33:06 -08003107 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003108 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003109 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003110
Ilpo Järvinen72018832007-12-30 04:37:55 -08003111 tp->packets_out -= acked_pcount;
3112 pkts_acked += acked_pcount;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003113 tcp_rate_skb_delivered(sk, skb, sack->rate);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003114
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003115 /* Initial outgoing SYN's get put onto the write_queue
3116 * just like anything else we transmit. It is not
3117 * true data, and if we misinform our callers that
3118 * this ACK acks real data, we will erroneously exit
3119 * connection startup slow start one packet too
3120 * quickly. This is severely frowned upon behavior.
3121 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003122 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003123 flag |= FLAG_DATA_ACKED;
3124 } else {
3125 flag |= FLAG_SYN_ACKED;
3126 tp->retrans_stamp = 0;
3127 }
3128
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003129 if (!fully_acked)
3130 break;
3131
Eric Dumazet75c119a2017-10-05 22:21:27 -07003132 next = skb_rb_next(skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003133 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003134 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003135 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003136 tp->lost_skb_hint = NULL;
Eric Dumazet75c119a2017-10-05 22:21:27 -07003137 tcp_rtx_queue_unlink_and_free(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 }
3139
Francis Yan0f872302016-11-27 23:07:14 -08003140 if (!skb)
3141 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3142
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003143 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3144 tp->snd_up = tp->snd_una;
3145
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003146 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3147 flag |= FLAG_SACK_RENEGING;
3148
Eric Dumazet9a568de2017-05-16 14:00:14 -07003149 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3150 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3151 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
Yuchung Chengeb36be02018-01-17 12:11:00 -08003152
3153 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3154 last_in_flight && !prior_sacked && fully_acked &&
3155 sack->rate->prior_delivered + 1 == tp->delivered &&
3156 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3157 /* Conservatively mark a delayed ACK. It's typically
3158 * from a lone runt packet over the round trip to
3159 * a receiver w/o out-of-order or CE events.
3160 */
3161 flag |= FLAG_ACK_MAYBE_DELAYED;
3162 }
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003163 }
Eric Dumazet9a568de2017-05-16 14:00:14 -07003164 if (sack->first_sackt) {
3165 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3166 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003167 }
Yuchung Chengf6722582015-10-16 21:57:42 -07003168 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003169 ca_rtt_us, sack->rate);
Yuchung Chenged084952013-07-22 16:20:48 -07003170
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003171 if (flag & FLAG_ACKED) {
Neal Cardwelldf92c832017-08-03 09:19:54 -04003172 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003173 if (unlikely(icsk->icsk_mtup.probe_size &&
3174 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3175 tcp_mtup_probe_success(sk);
3176 }
3177
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003178 if (tcp_is_reno(tp)) {
3179 tcp_remove_reno_sacks(sk, pkts_acked);
Ilpo Järvinen1236f222018-06-29 13:07:53 +03003180
3181 /* If any of the cumulatively ACKed segments was
3182 * retransmitted, non-SACK case cannot confirm that
3183 * progress was due to original transmission due to
3184 * lack of TCPCB_SACKED_ACKED bits even if some of
3185 * the packets may have been never retransmitted.
3186 */
3187 if (flag & FLAG_RETRANS_DATA_ACKED)
3188 flag &= ~FLAG_ORIG_SACK_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003189 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003190 int delta;
3191
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003192 /* Non-retransmitted hole got filled? That's reordering */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003193 if (before(reord, prior_fack))
3194 tcp_check_sack_reordering(sk, reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003195
Yuchung Cheng713bafe2017-11-08 13:01:26 -08003196 delta = prior_sacked - tp->sacked_out;
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003197 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003198 }
Eric Dumazet740b0f12014-02-26 14:02:48 -08003199 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003200 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3201 tcp_skb_timestamp_us(skb))) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003202 /* Do not re-arm RTO if the sack RTT is measured from data sent
3203 * after when the head was last (re)transmitted. Otherwise the
3204 * timeout may continue to extend in loss recovery.
3205 */
Neal Cardwelldf92c832017-08-03 09:19:54 -04003206 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 }
3208
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003209 if (icsk->icsk_ca_ops->pkts_acked) {
3210 struct ack_sample sample = { .pkts_acked = pkts_acked,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003211 .rtt_us = sack->rate->rtt_us,
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003212 .in_flight = last_in_flight };
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003213
3214 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3215 }
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003216
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003218 WARN_ON((int)tp->sacked_out < 0);
3219 WARN_ON((int)tp->lost_out < 0);
3220 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003221 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003222 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003224 pr_debug("Leak l=%u %d\n",
3225 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 tp->lost_out = 0;
3227 }
3228 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003229 pr_debug("Leak s=%u %d\n",
3230 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231 tp->sacked_out = 0;
3232 }
3233 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003234 pr_debug("Leak r=%u %d\n",
3235 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 tp->retrans_out = 0;
3237 }
3238 }
3239#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003240 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241}
3242
3243static void tcp_ack_probe(struct sock *sk)
3244{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003245 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07003246 struct sk_buff *head = tcp_send_head(sk);
3247 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248
3249 /* Was it a usable window open? */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003250 if (!head)
3251 return;
3252 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003253 icsk->icsk_backoff = 0;
3254 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 /* Socket must be waked up by subsequent tcp_data_snd_check().
3256 * This function is not for random using!
3257 */
3258 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003259 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003260
Eric Dumazet3f80e082018-10-23 11:54:16 -07003261 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3262 when, TCP_RTO_MAX, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263 }
3264}
3265
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003266static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003268 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3269 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270}
3271
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003272/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003273static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003275 /* If reordering is high then always grow cwnd whenever data is
3276 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003277 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003278 * new SACK or ECE mark may first advance cwnd here and later reduce
3279 * cwnd in tcp_fastretrans_alert() based on more states.
3280 */
Nikolay Borisov1043e252016-02-03 09:46:52 +02003281 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003282 return flag & FLAG_FORWARD_PROGRESS;
3283
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003284 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285}
3286
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003287/* The "ultimate" congestion control function that aims to replace the rigid
3288 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3289 * It's called toward the end of processing an ACK with precise rate
3290 * information. All transmission or retransmission are delayed afterwards.
3291 */
3292static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
Yuchung Chengc0402762016-09-19 23:39:21 -04003293 int flag, const struct rate_sample *rs)
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003294{
Yuchung Chengc0402762016-09-19 23:39:21 -04003295 const struct inet_connection_sock *icsk = inet_csk(sk);
3296
3297 if (icsk->icsk_ca_ops->cong_control) {
3298 icsk->icsk_ca_ops->cong_control(sk, rs);
3299 return;
3300 }
3301
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003302 if (tcp_in_cwnd_reduction(sk)) {
3303 /* Reduce cwnd if state mandates */
3304 tcp_cwnd_reduction(sk, acked_sacked, flag);
3305 } else if (tcp_may_raise_cwnd(sk, flag)) {
3306 /* Advance cwnd if state allows */
3307 tcp_cong_avoid(sk, ack, acked_sacked);
3308 }
3309 tcp_update_pacing_rate(sk);
3310}
3311
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312/* Check that window update is acceptable.
3313 * The function assumes that snd_una<=ack<=snd_next.
3314 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003315static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003316 const u32 ack, const u32 ack_seq,
3317 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003319 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003321 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322}
3323
Eric Dumazet0df48c22015-04-28 15:28:17 -07003324/* If we update tp->snd_una, also update tp->bytes_acked */
3325static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3326{
3327 u32 delta = ack - tp->snd_una;
3328
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003329 sock_owned_by_me((struct sock *)tp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003330 tp->bytes_acked += delta;
3331 tp->snd_una = ack;
3332}
3333
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003334/* If we update tp->rcv_nxt, also update tp->bytes_received */
3335static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3336{
3337 u32 delta = seq - tp->rcv_nxt;
3338
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003339 sock_owned_by_me((struct sock *)tp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003340 tp->bytes_received += delta;
3341 tp->rcv_nxt = seq;
3342}
3343
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344/* Update our send window.
3345 *
3346 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3347 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3348 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003349static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003350 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003352 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003354 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003356 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 nwin <<= tp->rx_opt.snd_wscale;
3358
3359 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3360 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003361 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
3363 if (tp->snd_wnd != nwin) {
3364 tp->snd_wnd = nwin;
3365
Florian Westphal31770e32017-08-30 19:24:58 +02003366 /* Note, it is the only place, where
3367 * fast path is recovered for sending TCP.
3368 */
3369 tp->pred_flags = 0;
3370 tcp_fast_path_check(sk);
3371
Eric Dumazet75c119a2017-10-05 22:21:27 -07003372 if (!tcp_write_queue_empty(sk))
Eric Dumazet6f021c62015-08-21 12:30:00 -07003373 tcp_slow_start_after_idle_check(sk);
3374
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 if (nwin > tp->max_window) {
3376 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003377 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 }
3379 }
3380 }
3381
Eric Dumazet0df48c22015-04-28 15:28:17 -07003382 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383
3384 return flag;
3385}
3386
Jason Baron083ae302016-07-14 11:38:40 -04003387static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3388 u32 *last_oow_ack_time)
3389{
3390 if (*last_oow_ack_time) {
Eric Dumazet594208a2017-05-16 14:00:10 -07003391 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
Jason Baron083ae302016-07-14 11:38:40 -04003392
Eric Dumazet4170ba62017-10-27 07:47:30 -07003393 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
Jason Baron083ae302016-07-14 11:38:40 -04003394 NET_INC_STATS(net, mib_idx);
3395 return true; /* rate-limited: don't send yet! */
3396 }
3397 }
3398
Eric Dumazet594208a2017-05-16 14:00:10 -07003399 *last_oow_ack_time = tcp_jiffies32;
Jason Baron083ae302016-07-14 11:38:40 -04003400
3401 return false; /* not rate-limited: go ahead, send dupack now! */
3402}
3403
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003404/* Return true if we're currently rate-limiting out-of-window ACKs and
3405 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3406 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3407 * attacks that send repeated SYNs or ACKs for the same connection. To
3408 * do this, we do not send a duplicate SYNACK or ACK if the remote
3409 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3410 */
3411bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3412 int mib_idx, u32 *last_oow_ack_time)
3413{
3414 /* Data packets without SYNs are not likely part of an ACK loop. */
3415 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3416 !tcp_hdr(skb)->syn)
Jason Baron083ae302016-07-14 11:38:40 -04003417 return false;
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003418
Jason Baron083ae302016-07-14 11:38:40 -04003419 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003420}
3421
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003422/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003423static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003424{
3425 /* unprotected vars, we dont care of overwrites */
3426 static u32 challenge_timestamp;
3427 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003428 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetb530b682017-10-27 07:47:26 -07003429 struct net *net = sock_net(sk);
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003430 u32 count, now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003431
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003432 /* First check our per-socket dupack rate limit. */
Eric Dumazetb530b682017-10-27 07:47:26 -07003433 if (__tcp_oow_rate_limited(net,
Jason Baron083ae302016-07-14 11:38:40 -04003434 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3435 &tp->last_oow_ack_time))
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003436 return;
3437
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003438 /* Then check host-wide RFC 5961 rate limit. */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003439 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003440 if (now != challenge_timestamp) {
Eric Dumazetb530b682017-10-27 07:47:26 -07003441 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3442 u32 half = (ack_limit + 1) >> 1;
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003443
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003444 challenge_timestamp = now;
Eric Dumazetb530b682017-10-27 07:47:26 -07003445 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003446 }
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003447 count = READ_ONCE(challenge_count);
3448 if (count > 0) {
3449 WRITE_ONCE(challenge_count, count - 1);
Eric Dumazetb530b682017-10-27 07:47:26 -07003450 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003451 tcp_send_ack(sk);
3452 }
3453}
3454
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003455static void tcp_store_ts_recent(struct tcp_sock *tp)
3456{
3457 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
Arnd Bergmanncca9bab2018-07-11 12:16:12 +02003458 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003459}
3460
3461static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3462{
3463 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3464 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3465 * extra check below makes sure this can only happen
3466 * for pure ACK frames. -DaveM
3467 *
3468 * Not only, also it occurs for expired timestamps.
3469 */
3470
3471 if (tcp_paws_check(&tp->rx_opt, 0))
3472 tcp_store_ts_recent(tp);
3473 }
3474}
3475
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003476/* This routine deals with acks during a TLP episode.
Sébastien Barré08abdff2015-01-12 10:30:40 +01003477 * We mark the end of a TLP episode on receiving TLP dupack or when
3478 * ack is after tlp_high_seq.
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003479 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3480 */
3481static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3482{
3483 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003484
Sébastien Barré08abdff2015-01-12 10:30:40 +01003485 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003486 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003487
Sébastien Barré08abdff2015-01-12 10:30:40 +01003488 if (flag & FLAG_DSACKING_ACK) {
3489 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003490 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003491 } else if (after(ack, tp->tlp_high_seq)) {
3492 /* ACK advances: there was a loss, so reduce cwnd. Reset
3493 * tlp_high_seq in tcp_init_cwnd_reduction()
3494 */
3495 tcp_init_cwnd_reduction(sk);
3496 tcp_set_ca_state(sk, TCP_CA_CWR);
3497 tcp_end_cwnd_reduction(sk);
3498 tcp_try_keep_open(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003499 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003500 LINUX_MIB_TCPLOSSPROBERECOVERY);
Sébastien Barré08abdff2015-01-12 10:30:40 +01003501 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3502 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3503 /* Pure dupack: original and TLP probe arrived; no loss */
3504 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003505 }
3506}
3507
Florian Westphal7354c8c2014-09-26 22:37:34 +02003508static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3509{
3510 const struct inet_connection_sock *icsk = inet_csk(sk);
3511
3512 if (icsk->icsk_ca_ops->in_ack_event)
3513 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3514}
3515
Yuchung Chenge662ca42016-02-02 10:33:04 -08003516/* Congestion control has updated the cwnd already. So if we're in
3517 * loss recovery then now we do any new sends (for FRTO) or
3518 * retransmits (for CA_Loss or CA_recovery) that make sense.
3519 */
3520static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3521{
3522 struct tcp_sock *tp = tcp_sk(sk);
3523
3524 if (rexmit == REXMIT_NONE)
3525 return;
3526
3527 if (unlikely(rexmit == 2)) {
3528 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3529 TCP_NAGLE_OFF);
3530 if (after(tp->snd_nxt, tp->high_seq))
3531 return;
3532 tp->frto = 0;
3533 }
3534 tcp_xmit_retransmit_queue(sk);
3535}
3536
Yuchung Chenga77fa012018-04-17 23:18:47 -07003537/* Returns the number of packets newly acked or sacked by the current ACK */
3538static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3539{
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003540 const struct net *net = sock_net(sk);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003541 struct tcp_sock *tp = tcp_sk(sk);
3542 u32 delivered;
3543
3544 delivered = tp->delivered - prior_delivered;
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003545 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3546 if (flag & FLAG_ECE) {
Yuchung Chenge21db6f2018-04-17 23:18:48 -07003547 tp->delivered_ce += delivered;
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003548 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3549 }
Yuchung Chenga77fa012018-04-17 23:18:47 -07003550 return delivered;
3551}
3552
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003554static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003556 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003558 struct tcp_sacktag_state sack_state;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003559 struct rate_sample rs = { .prior_delivered = 0 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 u32 prior_snd_una = tp->snd_una;
Yousuk Seungd4761752017-12-07 13:41:34 -08003561 bool is_sack_reneg = tp->is_sack_reneg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3563 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Eric Dumazet19119f22018-11-27 14:42:01 -08003564 int num_dupack = 0;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003565 int prior_packets = tp->packets_out;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003566 u32 delivered = tp->delivered;
3567 u32 lost = tp->lost;
Yuchung Chenge662ca42016-02-02 10:33:04 -08003568 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003569 u32 prior_fack;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003570
Eric Dumazet9a568de2017-05-16 14:00:14 -07003571 sack_state.first_sackt = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003572 sack_state.rate = &rs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573
Eric Dumazet75c119a2017-10-05 22:21:27 -07003574 /* We very likely will need to access rtx queue. */
3575 prefetch(sk->tcp_rtx_queue.rb_node);
Eric Dumazetad971f62014-10-11 15:17:29 -07003576
John Dykstra96e0bf42009-03-22 21:49:57 -07003577 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 * then we can probably ignore it.
3579 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003580 if (before(ack, prior_snd_una)) {
3581 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3582 if (before(ack, prior_snd_una - tp->max_window)) {
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07003583 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3584 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003585 return -1;
3586 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003588 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
John Dykstra96e0bf42009-03-22 21:49:57 -07003590 /* If the ack includes data we haven't sent yet, discard
3591 * this segment (RFC793 Section 3.9).
3592 */
3593 if (after(ack, tp->snd_nxt))
Yafang Shao9946b342019-02-25 18:33:48 +08003594 return -1;
John Dykstra96e0bf42009-03-22 21:49:57 -07003595
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003596 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003597 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003598 icsk->icsk_retransmits = 0;
Ilya Lesokhin6dac1522018-04-30 10:16:10 +03003599
3600#if IS_ENABLED(CONFIG_TLS_DEVICE)
3601 if (static_branch_unlikely(&clean_acked_data_enabled))
3602 if (icsk->icsk_clean_acked)
3603 icsk->icsk_clean_acked(sk, ack);
3604#endif
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003605 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003606
Eric Dumazet50895b92017-11-14 21:02:19 -08003607 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003608 rs.prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003609
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003610 /* ts_recent update must be made after we are sure that the packet
3611 * is in window.
3612 */
3613 if (flag & FLAG_UPDATE_TS_RECENT)
3614 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3615
Yafang Shao5e13a0d2018-11-11 20:10:10 +08003616 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3617 FLAG_SND_UNA_ADVANCED) {
Florian Westphal31770e32017-08-30 19:24:58 +02003618 /* Window is constant, pure forward advance.
3619 * No more checks are required.
3620 * Note, we use the fact that SND.UNA>=SND.WL2.
3621 */
3622 tcp_update_wl(tp, ack_seq);
3623 tcp_snd_una_update(tp, ack);
3624 flag |= FLAG_WIN_UPDATE;
3625
3626 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3627
3628 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3629 } else {
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003630 u32 ack_ev_flags = CA_ACK_SLOWPATH;
Florian Westphal98900922014-09-26 22:37:35 +02003631
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003632 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3633 flag |= FLAG_DATA;
3634 else
3635 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003637 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003639 if (TCP_SKB_CB(skb)->sacked)
3640 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3641 &sack_state);
3642
3643 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3644 flag |= FLAG_ECE;
3645 ack_ev_flags |= CA_ACK_ECE;
3646 }
3647
3648 if (flag & FLAG_WIN_UPDATE)
3649 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3650
3651 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 }
3653
3654 /* We passed data and got it acked, remove any soft error
3655 * log. Something worked...
3656 */
3657 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003658 icsk->icsk_probes_out = 0;
Eric Dumazet70eabf02017-05-16 14:00:07 -07003659 tp->rcv_tstamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 if (!prior_packets)
3661 goto no_queue;
3662
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003664 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003665
Priyaranjan Jha1f255692017-11-03 16:38:48 -07003666 tcp_rack_update_reo_wnd(sk, &rs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667
Neal Cardwelldf92c832017-08-03 09:19:54 -04003668 if (tp->tlp_high_seq)
3669 tcp_process_tlp_ack(sk, ack, flag);
3670 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3671 if (flag & FLAG_SET_XMIT_TIMER)
3672 tcp_set_xmit_timer(sk);
3673
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003674 if (tcp_ack_is_dubious(sk, flag)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08003675 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3676 num_dupack = 1;
3677 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3678 if (!(flag & FLAG_DATA))
3679 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3680 }
3681 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003682 &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003684
Julian Anastasovc3a2e832017-02-06 23:14:14 +02003685 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3686 sk_dst_confirm(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003687
Yuchung Chenga77fa012018-04-17 23:18:47 -07003688 delivered = tcp_newly_delivered(sk, delivered, flag);
Yuchung Chengb9f64822016-09-19 23:39:14 -04003689 lost = tp->lost - lost; /* freshly marked lost */
Yuchung Chenge4286602018-01-17 12:11:01 -08003690 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
Yousuk Seungd4761752017-12-07 13:41:34 -08003691 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003692 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003693 tcp_xmit_recovery(sk, rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 return 1;
3695
3696no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003697 /* If data was DSACKed, see if we can undo a cwnd reduction. */
Yuchung Chenga77fa012018-04-17 23:18:47 -07003698 if (flag & FLAG_DSACKING_ACK) {
Eric Dumazet19119f22018-11-27 14:42:01 -08003699 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003700 &rexmit);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003701 tcp_newly_delivered(sk, delivered, flag);
3702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 /* If this ack opens up a zero window, clear backoff. It was
3704 * being used to time the probes, and is probably far higher than
3705 * it needs to be for normal retransmission.
3706 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003707 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003708
3709 if (tp->tlp_high_seq)
3710 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711 return 1;
3712
3713old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003714 /* If data was SACKed, tag it and see if we should send more data.
3715 * If data was DSACKed, see if we can undo a cwnd reduction.
3716 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003717 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003718 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003719 &sack_state);
Eric Dumazet19119f22018-11-27 14:42:01 -08003720 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003721 &rexmit);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003722 tcp_newly_delivered(sk, delivered, flag);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003723 tcp_xmit_recovery(sk, rexmit);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003724 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 return 0;
3727}
3728
Daniel Lee7f9b8382015-04-06 14:37:26 -07003729static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3730 bool syn, struct tcp_fastopen_cookie *foc,
3731 bool exp_opt)
3732{
3733 /* Valid only in SYN or SYN-ACK with an even length. */
3734 if (!foc || !syn || len < 0 || (len & 1))
3735 return;
3736
3737 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3738 len <= TCP_FASTOPEN_COOKIE_MAX)
3739 memcpy(foc->val, cookie, len);
3740 else if (len != 0)
3741 len = -1;
3742 foc->len = len;
3743 foc->exp = exp_opt;
3744}
3745
Ursula Braun60e2a772017-10-25 11:01:45 +02003746static void smc_parse_options(const struct tcphdr *th,
3747 struct tcp_options_received *opt_rx,
3748 const unsigned char *ptr,
3749 int opsize)
3750{
3751#if IS_ENABLED(CONFIG_SMC)
3752 if (static_branch_unlikely(&tcp_have_smc)) {
3753 if (th->syn && !(opsize & 1) &&
3754 opsize >= TCPOLEN_EXP_SMC_BASE &&
3755 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3756 opt_rx->smc_ok = 1;
3757 }
3758#endif
3759}
3760
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3762 * But, this can also be called on packets in the established flow when
3763 * the fast version below fails.
3764 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07003765void tcp_parse_options(const struct net *net,
3766 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003767 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003768 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003770 const unsigned char *ptr;
3771 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003772 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003774 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 opt_rx->saw_tstamp = 0;
3776
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003777 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003778 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779 int opsize;
3780
3781 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003782 case TCPOPT_EOL:
3783 return;
3784 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3785 length--;
3786 continue;
3787 default:
3788 opsize = *ptr++;
3789 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003791 if (opsize > length)
3792 return; /* don't parse partial options */
3793 switch (opcode) {
3794 case TCPOPT_MSS:
3795 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003796 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003797 if (in_mss) {
3798 if (opt_rx->user_mss &&
3799 opt_rx->user_mss < in_mss)
3800 in_mss = opt_rx->user_mss;
3801 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003803 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003804 break;
3805 case TCPOPT_WINDOW:
3806 if (opsize == TCPOLEN_WINDOW && th->syn &&
Eric Dumazet9bb37ef2017-06-07 10:34:38 -07003807 !estab && net->ipv4.sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003808 __u8 snd_wscale = *(__u8 *)ptr;
3809 opt_rx->wscale_ok = 1;
Gao Feng589c49c2017-04-04 21:09:48 +08003810 if (snd_wscale > TCP_MAX_WSCALE) {
3811 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
Joe Perchese87cc472012-05-13 21:56:26 +00003812 __func__,
Gao Feng589c49c2017-04-04 21:09:48 +08003813 snd_wscale,
3814 TCP_MAX_WSCALE);
3815 snd_wscale = TCP_MAX_WSCALE;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003816 }
3817 opt_rx->snd_wscale = snd_wscale;
3818 }
3819 break;
3820 case TCPOPT_TIMESTAMP:
3821 if ((opsize == TCPOLEN_TIMESTAMP) &&
3822 ((estab && opt_rx->tstamp_ok) ||
Eric Dumazet5d2ed052017-06-07 10:34:39 -07003823 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003824 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003825 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3826 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003827 }
3828 break;
3829 case TCPOPT_SACK_PERM:
3830 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
Eric Dumazetf9301032017-06-07 10:34:37 -07003831 !estab && net->ipv4.sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003832 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003833 tcp_sack_reset(opt_rx);
3834 }
3835 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003836
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003837 case TCPOPT_SACK:
3838 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3839 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3840 opt_rx->sack_ok) {
3841 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3842 }
3843 break;
3844#ifdef CONFIG_TCP_MD5SIG
3845 case TCPOPT_MD5SIG:
3846 /*
3847 * The MD5 Hash has already been
3848 * checked (see tcp_v{4,6}_do_rcv()).
3849 */
3850 break;
3851#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07003852 case TCPOPT_FASTOPEN:
3853 tcp_parse_fastopen_option(
3854 opsize - TCPOLEN_FASTOPEN_BASE,
3855 ptr, th->syn, foc, false);
3856 break;
3857
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003858 case TCPOPT_EXP:
3859 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07003860 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003861 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07003862 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3863 get_unaligned_be16(ptr) ==
3864 TCPOPT_FASTOPEN_MAGIC)
3865 tcp_parse_fastopen_option(opsize -
3866 TCPOLEN_EXP_FASTOPEN_BASE,
3867 ptr + 2, th->syn, foc, true);
Ursula Braun60e2a772017-10-25 11:01:45 +02003868 else
3869 smc_parse_options(th, opt_rx, ptr,
3870 opsize);
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003871 break;
3872
3873 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003874 ptr += opsize-2;
3875 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 }
3878}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003879EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003881static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003882{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003883 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003884
3885 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3886 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3887 tp->rx_opt.saw_tstamp = 1;
3888 ++ptr;
3889 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3890 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003891 if (*ptr)
3892 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3893 else
3894 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003895 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003896 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003897 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003898}
3899
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900/* Fast parse options. This hopes to only see timestamps.
3901 * If it is wrong it falls back on tcp_parse_options().
3902 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07003903static bool tcp_fast_parse_options(const struct net *net,
3904 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003905 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003907 /* In the spirit of fast parsing, compare doff directly to constant
3908 * values. Because equality is used, short doff can be ignored here.
3909 */
3910 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003912 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003914 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003915 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003916 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003918
Eric Dumazeteed29f12017-06-07 10:34:36 -07003919 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003920 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003921 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3922
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003923 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924}
3925
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003926#ifdef CONFIG_TCP_MD5SIG
3927/*
3928 * Parse MD5 Signature option
3929 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003930const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003931{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003932 int length = (th->doff << 2) - sizeof(*th);
3933 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003934
Jann Horn7e5a2062018-04-20 15:57:30 +02003935 /* If not enough data remaining, we can short cut */
3936 while (length >= TCPOLEN_MD5SIG) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003937 int opcode = *ptr++;
3938 int opsize;
3939
Weilong Chena22318e2013-12-23 14:37:26 +08003940 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003941 case TCPOPT_EOL:
3942 return NULL;
3943 case TCPOPT_NOP:
3944 length--;
3945 continue;
3946 default:
3947 opsize = *ptr++;
3948 if (opsize < 2 || opsize > length)
3949 return NULL;
3950 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003951 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003952 }
3953 ptr += opsize - 2;
3954 length -= opsize;
3955 }
3956 return NULL;
3957}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003958EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003959#endif
3960
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3962 *
3963 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3964 * it can pass through stack. So, the following predicate verifies that
3965 * this segment is not used for anything but congestion avoidance or
3966 * fast retransmit. Moreover, we even are able to eliminate most of such
3967 * second order effects, if we apply some small "replay" window (~RTO)
3968 * to timestamp space.
3969 *
3970 * All these measures still do not guarantee that we reject wrapped ACKs
3971 * on networks with high bandwidth, when sequence space is recycled fastly,
3972 * but it guarantees that such events will be very rare and do not affect
3973 * connection seriously. This doesn't look nice, but alas, PAWS is really
3974 * buggy extension.
3975 *
3976 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3977 * states that events when retransmit arrives after original data are rare.
3978 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3979 * the biggest problem on large power networks even with minor reordering.
3980 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3981 * up to bandwidth of 18Gigabit/sec. 8) ]
3982 */
3983
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003984static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003986 const struct tcp_sock *tp = tcp_sk(sk);
3987 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 u32 seq = TCP_SKB_CB(skb)->seq;
3989 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3990
3991 return (/* 1. Pure ACK with correct sequence number. */
3992 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3993
3994 /* 2. ... and duplicate ACK. */
3995 ack == tp->snd_una &&
3996
3997 /* 3. ... and does not update window. */
3998 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3999
4000 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004001 (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 -07004002}
4003
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004004static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004005 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004007 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00004008
4009 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4010 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011}
4012
4013/* Check segment sequence number for validity.
4014 *
4015 * Segment controls are considered valid, if the segment
4016 * fits to the window after truncation to the window. Acceptability
4017 * of data (and SYN, FIN, of course) is checked separately.
4018 * See tcp_data_queue(), for example.
4019 *
4020 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4021 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4022 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4023 * (borrowed from freebsd)
4024 */
4025
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004026static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027{
4028 return !before(end_seq, tp->rcv_wup) &&
4029 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4030}
4031
4032/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00004033void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034{
Song Liu59415212017-10-23 09:20:25 -07004035 trace_tcp_receive_reset(sk);
4036
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 /* We want the right error as BSD sees it (and indeed as we do). */
4038 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004039 case TCP_SYN_SENT:
4040 sk->sk_err = ECONNREFUSED;
4041 break;
4042 case TCP_CLOSE_WAIT:
4043 sk->sk_err = EPIPE;
4044 break;
4045 case TCP_CLOSE:
4046 return;
4047 default:
4048 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004050 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4051 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Soheil Hassas Yeganeha27fd7a2018-02-27 18:32:18 -05004053 tcp_write_queue_purge(sk);
Eric Dumazet3d476262017-04-18 09:45:51 -07004054 tcp_done(sk);
4055
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 if (!sock_flag(sk, SOCK_DEAD))
4057 sk->sk_error_report(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058}
4059
4060/*
4061 * Process the FIN bit. This now behaves as it is supposed to work
4062 * and the FIN takes effect when it is validly part of sequence
4063 * space. Not before when we get holes.
4064 *
4065 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4066 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4067 * TIME-WAIT)
4068 *
4069 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4070 * close and we go into CLOSING (and later onto TIME-WAIT)
4071 *
4072 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4073 */
Eric Dumazete3e17b72016-02-06 11:16:28 -08004074void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075{
4076 struct tcp_sock *tp = tcp_sk(sk);
4077
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004078 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079
4080 sk->sk_shutdown |= RCV_SHUTDOWN;
4081 sock_set_flag(sk, SOCK_DONE);
4082
4083 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004084 case TCP_SYN_RECV:
4085 case TCP_ESTABLISHED:
4086 /* Move to CLOSE_WAIT */
4087 tcp_set_state(sk, TCP_CLOSE_WAIT);
Wei Wang31954cd2019-01-25 10:53:19 -08004088 inet_csk_enter_pingpong_mode(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004089 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004091 case TCP_CLOSE_WAIT:
4092 case TCP_CLOSING:
4093 /* Received a retransmission of the FIN, do
4094 * nothing.
4095 */
4096 break;
4097 case TCP_LAST_ACK:
4098 /* RFC793: Remain in the LAST-ACK state. */
4099 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004101 case TCP_FIN_WAIT1:
4102 /* This case occurs when a simultaneous close
4103 * happens, we must ack the received FIN and
4104 * enter the CLOSING state.
4105 */
4106 tcp_send_ack(sk);
4107 tcp_set_state(sk, TCP_CLOSING);
4108 break;
4109 case TCP_FIN_WAIT2:
4110 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4111 tcp_send_ack(sk);
4112 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4113 break;
4114 default:
4115 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4116 * cases we should never reach this piece of code.
4117 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004118 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004119 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004120 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122
4123 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4124 * Probably, we should reset in this case. For now drop them.
4125 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004126 skb_rbtree_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004127 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004129 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130
4131 if (!sock_flag(sk, SOCK_DEAD)) {
4132 sk->sk_state_change(sk);
4133
4134 /* Do not send POLL_HUP for half duplex close. */
4135 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4136 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004137 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004139 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 }
4141}
4142
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004143static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004144 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145{
4146 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4147 if (before(seq, sp->start_seq))
4148 sp->start_seq = seq;
4149 if (after(end_seq, sp->end_seq))
4150 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004151 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004153 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154}
4155
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004156static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004158 struct tcp_sock *tp = tcp_sk(sk);
4159
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004160 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004161 int mib_idx;
4162
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004164 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004166 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4167
Eric Dumazetc10d9312016-04-29 14:16:47 -07004168 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169
4170 tp->rx_opt.dsack = 1;
4171 tp->duplicate_sack[0].start_seq = seq;
4172 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 }
4174}
4175
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004176static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004178 struct tcp_sock *tp = tcp_sk(sk);
4179
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004181 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182 else
4183 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4184}
4185
Yuchung Cheng77881742018-08-29 14:53:56 -07004186static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4187{
4188 /* When the ACK path fails or drops most ACKs, the sender would
4189 * timeout and spuriously retransmit the same segment repeatedly.
4190 * The receiver remembers and reflects via DSACKs. Leverage the
4191 * DSACK state and change the txhash to re-route speculatively.
4192 */
4193 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4194 sk_rethink_txhash(sk);
4195}
4196
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004197static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198{
4199 struct tcp_sock *tp = tcp_sk(sk);
4200
4201 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4202 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004203 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07004204 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004206 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4208
Yuchung Cheng77881742018-08-29 14:53:56 -07004209 tcp_rcv_spurious_retrans(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4211 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004212 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 }
4214 }
4215
4216 tcp_send_ack(sk);
4217}
4218
4219/* These routines update the SACK block as out-of-order packets arrive or
4220 * in-order packets close up the sequence space.
4221 */
4222static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4223{
4224 int this_sack;
4225 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004226 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227
4228 /* See if the recent change to the first SACK eats into
4229 * or hits the sequence space of other SACK blocks, if so coalesce.
4230 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004231 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4233 int i;
4234
4235 /* Zap SWALK, by moving every further SACK up by one slot.
4236 * Decrease num_sacks.
4237 */
4238 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004239 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4240 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241 continue;
4242 }
4243 this_sack++, swalk++;
4244 }
4245}
4246
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4248{
4249 struct tcp_sock *tp = tcp_sk(sk);
4250 struct tcp_sack_block *sp = &tp->selective_acks[0];
4251 int cur_sacks = tp->rx_opt.num_sacks;
4252 int this_sack;
4253
4254 if (!cur_sacks)
4255 goto new_sack;
4256
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004257 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258 if (tcp_sack_extend(sp, seq, end_seq)) {
4259 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004260 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004261 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262 if (cur_sacks > 1)
4263 tcp_sack_maybe_coalesce(tp);
4264 return;
4265 }
4266 }
4267
4268 /* Could not find an adjacent existing SACK, build a new one,
4269 * put it at the front, and shift everyone else down. We
4270 * always know there is at least one SACK present already here.
4271 *
4272 * If the sack array is full, forget about the last one.
4273 */
Adam Langley4389dde2008-07-19 00:07:02 -07004274 if (this_sack >= TCP_NUM_SACKS) {
Eric Dumazet86de5922018-11-20 05:53:59 -08004275 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
Eric Dumazet5d9f4262018-05-17 14:47:26 -07004276 tcp_send_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 this_sack--;
4278 tp->rx_opt.num_sacks--;
4279 sp--;
4280 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004281 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004282 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283
4284new_sack:
4285 /* Build the new head SACK, and we're done. */
4286 sp->start_seq = seq;
4287 sp->end_seq = end_seq;
4288 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289}
4290
4291/* RCV.NXT advances, some SACKs should be eaten. */
4292
4293static void tcp_sack_remove(struct tcp_sock *tp)
4294{
4295 struct tcp_sack_block *sp = &tp->selective_acks[0];
4296 int num_sacks = tp->rx_opt.num_sacks;
4297 int this_sack;
4298
4299 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004300 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302 return;
4303 }
4304
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004305 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306 /* Check if the start of the sack is covered by RCV.NXT. */
4307 if (!before(tp->rcv_nxt, sp->start_seq)) {
4308 int i;
4309
4310 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004311 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312
4313 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004314 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315 tp->selective_acks[i-1] = tp->selective_acks[i];
4316 num_sacks--;
4317 continue;
4318 }
4319 this_sack++;
4320 sp++;
4321 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004322 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004323}
4324
Eric Dumazet1402d362012-04-23 07:11:42 +00004325/**
4326 * tcp_try_coalesce - try to merge skb to prior one
4327 * @sk: socket
Mike Maloney98aaa912017-08-22 17:08:48 -04004328 * @dest: destination queue
Eric Dumazet1402d362012-04-23 07:11:42 +00004329 * @to: prior buffer
4330 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004331 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004332 *
4333 * Before queueing skb @from after @to, try to merge them
4334 * to reduce overall memory use and queue lengths, if cost is small.
4335 * Packets in ofo or receive queues can stay a long time.
4336 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004337 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004338 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004339static bool tcp_try_coalesce(struct sock *sk,
4340 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004341 struct sk_buff *from,
4342 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004343{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004344 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004345
Eric Dumazet329033f2012-04-27 00:38:33 +00004346 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004347
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004348 /* Its possible this segment overlaps with prior segment in queue */
4349 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4350 return false;
4351
Boris Pismenny41ed9c02018-07-13 14:33:38 +03004352#ifdef CONFIG_TLS_DEVICE
4353 if (from->decrypted != to->decrypted)
4354 return false;
4355#endif
4356
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004357 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004358 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004359
Alexander Duyck34a802a2012-05-02 21:19:09 +00004360 atomic_add(delta, &sk->sk_rmem_alloc);
4361 sk_mem_charge(sk, delta);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004362 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004363 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4364 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004365 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Mike Maloney98aaa912017-08-22 17:08:48 -04004366
4367 if (TCP_SKB_CB(from)->has_rxtstamp) {
4368 TCP_SKB_CB(to)->has_rxtstamp = true;
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004369 to->tstamp = from->tstamp;
Stephen Malloncadf9df2018-11-20 19:15:02 +11004370 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
Mike Maloney98aaa912017-08-22 17:08:48 -04004371 }
4372
Alexander Duyck34a802a2012-05-02 21:19:09 +00004373 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004374}
4375
Eric Dumazet58152ec2018-07-23 09:28:21 -07004376static bool tcp_ooo_try_coalesce(struct sock *sk,
4377 struct sk_buff *to,
4378 struct sk_buff *from,
4379 bool *fragstolen)
4380{
4381 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4382
4383 /* In case tcp_drop() is called later, update to->gso_segs */
4384 if (res) {
4385 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4386 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4387
4388 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4389 }
4390 return res;
4391}
4392
Eric Dumazet532182c2016-04-01 08:52:19 -07004393static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4394{
4395 sk_drops_add(sk, skb);
4396 __kfree_skb(skb);
4397}
4398
Linus Torvalds1da177e2005-04-16 15:20:36 -07004399/* This one checks to see if we can put data from the
4400 * out_of_order queue into the receive_queue.
4401 */
4402static void tcp_ofo_queue(struct sock *sk)
4403{
4404 struct tcp_sock *tp = tcp_sk(sk);
4405 __u32 dsack_high = tp->rcv_nxt;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004406 bool fin, fragstolen, eaten;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004407 struct sk_buff *skb, *tail;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004408 struct rb_node *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004409
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004410 p = rb_first(&tp->out_of_order_queue);
4411 while (p) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004412 skb = rb_to_skb(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004413 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4414 break;
4415
4416 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4417 __u32 dsack = dsack_high;
4418 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4419 dsack_high = TCP_SKB_CB(skb)->end_seq;
4420 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4421 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004422 p = rb_next(p);
4423 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004424
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004425 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
Eric Dumazet532182c2016-04-01 08:52:19 -07004426 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427 continue;
4428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004429
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004430 tail = skb_peek_tail(&sk->sk_receive_queue);
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004431 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004432 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004433 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004434 if (!eaten)
4435 __skb_queue_tail(&sk->sk_receive_queue, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004436 else
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004437 kfree_skb_partial(skb, fragstolen);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004438
4439 if (unlikely(fin)) {
4440 tcp_fin(sk);
4441 /* tcp_fin() purges tp->out_of_order_queue,
4442 * so we must end this loop right now.
4443 */
4444 break;
4445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004446 }
4447}
4448
4449static bool tcp_prune_ofo_queue(struct sock *sk);
4450static int tcp_prune_queue(struct sock *sk);
4451
4452static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4453 unsigned int size)
4454{
4455 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4456 !sk_rmem_schedule(sk, skb, size)) {
4457
4458 if (tcp_prune_queue(sk) < 0)
4459 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004460
Eric Dumazet36a65032016-08-17 14:17:09 -07004461 while (!sk_rmem_schedule(sk, skb, size)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462 if (!tcp_prune_ofo_queue(sk))
4463 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004464 }
4465 }
4466 return 0;
4467}
4468
Eric Dumazete86b29192012-03-18 11:06:44 +00004469static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4470{
4471 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004472 struct rb_node **p, *parent;
Eric Dumazete86b29192012-03-18 11:06:44 +00004473 struct sk_buff *skb1;
4474 u32 seq, end_seq;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004475 bool fragstolen;
Eric Dumazete86b29192012-03-18 11:06:44 +00004476
Yousuk Seungf4c9f852018-06-04 15:29:51 -07004477 tcp_ecn_check_ce(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004478
Mel Gormanc76562b2012-07-31 16:44:41 -07004479 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004480 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazet532182c2016-04-01 08:52:19 -07004481 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004482 return;
4483 }
4484
Florian Westphal31770e32017-08-30 19:24:58 +02004485 /* Disable header prediction. */
4486 tp->pred_flags = 0;
Eric Dumazete86b29192012-03-18 11:06:44 +00004487 inet_csk_schedule_ack(sk);
4488
Eric Dumazetc10d9312016-04-29 14:16:47 -07004489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004490 seq = TCP_SKB_CB(skb)->seq;
4491 end_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004492
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004493 p = &tp->out_of_order_queue.rb_node;
4494 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004495 /* Initial out of order segment, build 1 SACK. */
4496 if (tcp_is_sack(tp)) {
4497 tp->rx_opt.num_sacks = 1;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004498 tp->selective_acks[0].start_seq = seq;
4499 tp->selective_acks[0].end_seq = end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004500 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004501 rb_link_node(&skb->rbnode, NULL, p);
4502 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4503 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004504 goto end;
4505 }
4506
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004507 /* In the typical case, we are adding an skb to the end of the list.
4508 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4509 */
Eric Dumazet58152ec2018-07-23 09:28:21 -07004510 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4511 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004512coalesce_done:
4513 tcp_grow_window(sk, skb);
4514 kfree_skb_partial(skb, fragstolen);
4515 skb = NULL;
4516 goto add_sack;
Eric Dumazete86b29192012-03-18 11:06:44 +00004517 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004518 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4519 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4520 parent = &tp->ooo_last_skb->rbnode;
4521 p = &parent->rb_right;
4522 goto insert;
Eric Dumazete86b29192012-03-18 11:06:44 +00004523 }
4524
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004525 /* Find place to insert this segment. Handle overlaps on the way. */
4526 parent = NULL;
4527 while (*p) {
4528 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004529 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004530 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4531 p = &parent->rb_left;
4532 continue;
Eric Dumazete86b29192012-03-18 11:06:44 +00004533 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004534 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4535 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4536 /* All the bits are present. Drop. */
4537 NET_INC_STATS(sock_net(sk),
4538 LINUX_MIB_TCPOFOMERGE);
Eric Dumazet8541b212018-07-23 09:28:20 -07004539 tcp_drop(sk, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004540 skb = NULL;
4541 tcp_dsack_set(sk, seq, end_seq);
4542 goto add_sack;
4543 }
4544 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4545 /* Partial overlap. */
4546 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4547 } else {
4548 /* skb's seq == skb1's seq and skb covers skb1.
4549 * Replace skb1 with skb.
4550 */
4551 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4552 &tp->out_of_order_queue);
4553 tcp_dsack_extend(sk,
4554 TCP_SKB_CB(skb1)->seq,
4555 TCP_SKB_CB(skb1)->end_seq);
4556 NET_INC_STATS(sock_net(sk),
4557 LINUX_MIB_TCPOFOMERGE);
Eric Dumazet8541b212018-07-23 09:28:20 -07004558 tcp_drop(sk, skb1);
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004559 goto merge_right;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004560 }
Eric Dumazet58152ec2018-07-23 09:28:21 -07004561 } else if (tcp_ooo_try_coalesce(sk, skb1,
4562 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004563 goto coalesce_done;
4564 }
4565 p = &parent->rb_right;
Eric Dumazete86b29192012-03-18 11:06:44 +00004566 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004567insert:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004568 /* Insert segment into RB tree. */
4569 rb_link_node(&skb->rbnode, parent, p);
4570 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004571
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004572merge_right:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004573 /* Remove other segments covered by skb. */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004574 while ((skb1 = skb_rb_next(skb)) != NULL) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004575 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4576 break;
4577 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4578 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4579 end_seq);
4580 break;
4581 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004582 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004583 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4584 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004585 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004586 tcp_drop(sk, skb1);
Eric Dumazete86b29192012-03-18 11:06:44 +00004587 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004588 /* If there is no skb after us, we are the last_skb ! */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004589 if (!skb1)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004590 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004591
4592add_sack:
4593 if (tcp_is_sack(tp))
4594 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4595end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004596 if (skb) {
4597 tcp_grow_window(sk, skb);
Eric Dumazet60b1af32017-01-24 14:57:36 -08004598 skb_condense(skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004599 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004600 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004601}
4602
Eric Dumazete7395f1f2018-11-26 14:49:12 -08004603static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4604 bool *fragstolen)
Eric Dumazetb081f852012-05-02 09:58:29 +00004605{
4606 int eaten;
4607 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4608
Eric Dumazetb081f852012-05-02 09:58:29 +00004609 eaten = (tail &&
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004610 tcp_try_coalesce(sk, tail,
Mike Maloney98aaa912017-08-22 17:08:48 -04004611 skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004612 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004613 if (!eaten) {
4614 __skb_queue_tail(&sk->sk_receive_queue, skb);
4615 skb_set_owner_r(skb, sk);
4616 }
4617 return eaten;
4618}
Eric Dumazete86b29192012-03-18 11:06:44 +00004619
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004620int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4621{
Eric Dumazetcb934712014-09-17 03:14:42 -07004622 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004623 int err = -ENOMEM;
4624 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004625 bool fragstolen;
4626
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004627 if (size == 0)
4628 return 0;
4629
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004630 if (size > PAGE_SIZE) {
4631 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4632
4633 data_len = npages << PAGE_SHIFT;
4634 size = data_len + (size & ~PAGE_MASK);
4635 }
4636 skb = alloc_skb_with_frags(size - data_len, data_len,
4637 PAGE_ALLOC_COSTLY_ORDER,
4638 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004639 if (!skb)
4640 goto err;
4641
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004642 skb_put(skb, size - data_len);
4643 skb->data_len = data_len;
4644 skb->len = size;
4645
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004646 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4647 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
Mel Gormanc76562b2012-07-31 16:44:41 -07004648 goto err_free;
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004649 }
Mel Gormanc76562b2012-07-31 16:44:41 -07004650
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004651 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4652 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004653 goto err_free;
4654
4655 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4656 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4657 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4658
Eric Dumazete7395f1f2018-11-26 14:49:12 -08004659 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004660 WARN_ON_ONCE(fragstolen); /* should not happen */
4661 __kfree_skb(skb);
4662 }
4663 return size;
4664
4665err_free:
4666 kfree_skb(skb);
4667err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004668 return err;
4669
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004670}
4671
Eric Dumazet03f45c82018-04-16 10:33:37 -07004672void tcp_data_ready(struct sock *sk)
4673{
4674 const struct tcp_sock *tp = tcp_sk(sk);
4675 int avail = tp->rcv_nxt - tp->copied_seq;
4676
4677 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4678 return;
4679
4680 sk->sk_data_ready(sk);
4681}
4682
Linus Torvalds1da177e2005-04-16 15:20:36 -07004683static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4684{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004685 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004686 bool fragstolen;
4687 int eaten;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004688
Eric Dumazet532182c2016-04-01 08:52:19 -07004689 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4690 __kfree_skb(skb);
4691 return;
4692 }
Eric Dumazetf84af322010-04-28 15:31:51 -07004693 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004694 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004695
Yuchung Chengfd2123a2018-08-09 09:38:12 -07004696 tcp_ecn_accept_cwr(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004697
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004698 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004699
4700 /* Queue data for delivery to the user.
4701 * Packets in sequence go to the receive queue.
4702 * Out of sequence packets to the out_of_order_queue.
4703 */
4704 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
Yafang Shaofb223502018-06-24 10:02:54 -04004705 if (tcp_receive_window(tp) == 0) {
4706 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004707 goto out_of_window;
Yafang Shaofb223502018-06-24 10:02:54 -04004708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004709
4710 /* Ok. In sequence. In window. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004711queue_and_out:
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004712 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4713 sk_forced_mem_schedule(sk, skb->truesize);
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004714 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4715 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004716 goto drop;
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004717 }
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004718
Eric Dumazete7395f1f2018-11-26 14:49:12 -08004719 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004720 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004721 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004722 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004723 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004725 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004726 tcp_ofo_queue(sk);
4727
Yuchung Cheng15bdd562018-08-09 09:38:11 -07004728 /* RFC5681. 4.2. SHOULD send immediate ACK, when
Linus Torvalds1da177e2005-04-16 15:20:36 -07004729 * gap in queue is filled.
4730 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004731 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Yuchung Cheng15bdd562018-08-09 09:38:11 -07004732 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733 }
4734
4735 if (tp->rx_opt.num_sacks)
4736 tcp_sack_remove(tp);
4737
Florian Westphal31770e32017-08-30 19:24:58 +02004738 tcp_fast_path_check(sk);
4739
Eric Dumazet923dd342012-05-02 07:55:58 +00004740 if (eaten > 0)
4741 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004742 if (!sock_flag(sk, SOCK_DEAD))
Eric Dumazet03f45c82018-04-16 10:33:37 -07004743 tcp_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004744 return;
4745 }
4746
4747 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
Yuchung Cheng77881742018-08-29 14:53:56 -07004748 tcp_rcv_spurious_retrans(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749 /* A retransmit, 2nd most common case. Force an immediate ack. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004750 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004751 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752
4753out_of_window:
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07004754 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004755 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756drop:
Eric Dumazet532182c2016-04-01 08:52:19 -07004757 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758 return;
4759 }
4760
4761 /* Out of window. F.e. zero window probe. */
4762 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4763 goto out_of_window;
4764
Linus Torvalds1da177e2005-04-16 15:20:36 -07004765 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4766 /* Partial packet, seq < rcv_next < end_seq */
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004767 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004768
Linus Torvalds1da177e2005-04-16 15:20:36 -07004769 /* If window is closed, drop tail of packet. But after
4770 * remembering D-SACK for its head made in previous line.
4771 */
Yafang Shaofb223502018-06-24 10:02:54 -04004772 if (!tcp_receive_window(tp)) {
4773 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004774 goto out_of_window;
Yafang Shaofb223502018-06-24 10:02:54 -04004775 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004776 goto queue_and_out;
4777 }
4778
Eric Dumazete86b29192012-03-18 11:06:44 +00004779 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780}
4781
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004782static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004783{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004784 if (list)
4785 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
David S. Miller91521942009-05-28 21:35:47 -07004786
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004787 return skb_rb_next(skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004788}
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004789
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004790static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4791 struct sk_buff_head *list,
4792 struct rb_root *root)
4793{
4794 struct sk_buff *next = tcp_skb_next(skb, list);
4795
4796 if (list)
4797 __skb_unlink(skb, list);
4798 else
4799 rb_erase(&skb->rbnode, root);
4800
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004801 __kfree_skb(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004802 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004803
4804 return next;
4805}
4806
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004807/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07004808void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004809{
4810 struct rb_node **p = &root->rb_node;
4811 struct rb_node *parent = NULL;
4812 struct sk_buff *skb1;
4813
4814 while (*p) {
4815 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004816 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004817 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4818 p = &parent->rb_left;
4819 else
4820 p = &parent->rb_right;
4821 }
4822 rb_link_node(&skb->rbnode, parent, p);
4823 rb_insert_color(&skb->rbnode, root);
4824}
4825
Linus Torvalds1da177e2005-04-16 15:20:36 -07004826/* Collapse contiguous sequence of skbs head..tail with
4827 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004828 *
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004829 * If tail is NULL, this means until the end of the queue.
David S. Miller91521942009-05-28 21:35:47 -07004830 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004831 * Segments with FIN/SYN are not collapsed (only because this
4832 * simplifies code)
4833 */
4834static void
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004835tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4836 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004837{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004838 struct sk_buff *skb = head, *n;
4839 struct sk_buff_head tmp;
David S. Miller91521942009-05-28 21:35:47 -07004840 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004841
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004842 /* First, check that queue is collapsible and find
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004843 * the point where collapsing can be useful.
4844 */
David S. Miller91521942009-05-28 21:35:47 -07004845restart:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004846 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4847 n = tcp_skb_next(skb, list);
4848
Linus Torvalds1da177e2005-04-16 15:20:36 -07004849 /* No new bits? It is possible on ofo queue. */
4850 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004851 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004852 if (!skb)
4853 break;
4854 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004855 }
4856
4857 /* The first skb to collapse is:
4858 * - not SYN/FIN and
4859 * - bloated or contains data before "start" or
4860 * overlaps to the next one.
4861 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004862 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Eric Dumazet94f08932017-10-26 21:55:09 -07004863 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004864 before(TCP_SKB_CB(skb)->seq, start))) {
4865 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004866 break;
David S. Miller91521942009-05-28 21:35:47 -07004867 }
4868
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004869 if (n && n != tail &&
4870 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4871 end_of_skbs = false;
4872 break;
David S. Miller91521942009-05-28 21:35:47 -07004873 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004874
4875 /* Decided to skip this, advance start seq. */
4876 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004877 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004878 if (end_of_skbs ||
4879 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004880 return;
4881
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004882 __skb_queue_head_init(&tmp);
4883
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004885 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004886 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004887
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004888 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004889 if (!nskb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004890 break;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004891
Linus Torvalds1da177e2005-04-16 15:20:36 -07004892 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
Boris Pismenny41ed9c02018-07-13 14:33:38 +03004893#ifdef CONFIG_TLS_DEVICE
4894 nskb->decrypted = skb->decrypted;
4895#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004896 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004897 if (list)
4898 __skb_queue_before(list, skb, nskb);
4899 else
4900 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004901 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004902
4903 /* Copy data, releasing collapsed skbs. */
4904 while (copy > 0) {
4905 int offset = start - TCP_SKB_CB(skb)->seq;
4906 int size = TCP_SKB_CB(skb)->end_seq - start;
4907
Kris Katterjohn09a62662006-01-08 22:24:28 -08004908 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004909 if (size > 0) {
4910 size = min(copy, size);
4911 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4912 BUG();
4913 TCP_SKB_CB(nskb)->end_seq += size;
4914 copy -= size;
4915 start += size;
4916 }
4917 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004918 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004919 if (!skb ||
4920 skb == tail ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004921 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004922 goto end;
Boris Pismenny41ed9c02018-07-13 14:33:38 +03004923#ifdef CONFIG_TLS_DEVICE
4924 if (skb->decrypted != nskb->decrypted)
4925 goto end;
4926#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004927 }
4928 }
4929 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004930end:
4931 skb_queue_walk_safe(&tmp, skb, n)
4932 tcp_rbtree_insert(root, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004933}
4934
4935/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4936 * and tcp_collapse() them until all the queue is collapsed.
4937 */
4938static void tcp_collapse_ofo_queue(struct sock *sk)
4939{
4940 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet3d4bf932018-07-23 09:28:19 -07004941 u32 range_truesize, sum_tiny = 0;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004942 struct sk_buff *skb, *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004943 u32 start, end;
4944
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004945 skb = skb_rb_first(&tp->out_of_order_queue);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004946new_range:
4947 if (!skb) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004948 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949 return;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004950 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004951 start = TCP_SKB_CB(skb)->seq;
4952 end = TCP_SKB_CB(skb)->end_seq;
Eric Dumazet3d4bf932018-07-23 09:28:19 -07004953 range_truesize = skb->truesize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004954
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004955 for (head = skb;;) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004956 skb = skb_rb_next(skb);
David S. Miller91521942009-05-28 21:35:47 -07004957
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004958 /* Range is terminated when we see a gap or when
4959 * we are at the queue end.
4960 */
David S. Miller91521942009-05-28 21:35:47 -07004961 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962 after(TCP_SKB_CB(skb)->seq, end) ||
4963 before(TCP_SKB_CB(skb)->end_seq, start)) {
Eric Dumazet3d4bf932018-07-23 09:28:19 -07004964 /* Do not attempt collapsing tiny skbs */
4965 if (range_truesize != head->truesize ||
4966 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4967 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4968 head, skb, start, end);
4969 } else {
4970 sum_tiny += range_truesize;
4971 if (sum_tiny > sk->sk_rcvbuf >> 3)
4972 return;
4973 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004974 goto new_range;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004975 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004976
Eric Dumazet3d4bf932018-07-23 09:28:19 -07004977 range_truesize += skb->truesize;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004978 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4979 start = TCP_SKB_CB(skb)->seq;
4980 if (after(TCP_SKB_CB(skb)->end_seq, end))
4981 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004982 }
4983}
4984
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004985/*
Eric Dumazet36a65032016-08-17 14:17:09 -07004986 * Clean the out-of-order queue to make room.
4987 * We drop high sequences packets to :
4988 * 1) Let a chance for holes to be filled.
4989 * 2) not add too big latencies if thousands of packets sit there.
4990 * (But if application shrinks SO_RCVBUF, we could still end up
4991 * freeing whole queue here)
Eric Dumazet72cd43b2018-07-23 09:28:17 -07004992 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
Eric Dumazet36a65032016-08-17 14:17:09 -07004993 *
4994 * Return true if queue has shrunk.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004995 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004996static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004997{
4998 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004999 struct rb_node *node, *prev;
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005000 int goal;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005001
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005002 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Eric Dumazet36a65032016-08-17 14:17:09 -07005003 return false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005004
Eric Dumazet36a65032016-08-17 14:17:09 -07005005 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005006 goal = sk->sk_rcvbuf >> 3;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005007 node = &tp->ooo_last_skb->rbnode;
5008 do {
5009 prev = rb_prev(node);
5010 rb_erase(node, &tp->out_of_order_queue);
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005011 goal -= rb_to_skb(node)->truesize;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005012 tcp_drop(sk, rb_to_skb(node));
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005013 if (!prev || goal <= 0) {
5014 sk_mem_reclaim(sk);
5015 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5016 !tcp_under_memory_pressure(sk))
5017 break;
5018 goal = sk->sk_rcvbuf >> 3;
5019 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005020 node = prev;
5021 } while (node);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005022 tp->ooo_last_skb = rb_to_skb(prev);
Eric Dumazet36a65032016-08-17 14:17:09 -07005023
5024 /* Reset SACK state. A conforming SACK implementation will
5025 * do the same at a timeout based retransmit. When a connection
5026 * is in a sad state like this, we care only about integrity
5027 * of the connection not performance.
5028 */
5029 if (tp->rx_opt.sack_ok)
5030 tcp_sack_reset(&tp->rx_opt);
5031 return true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005032}
5033
Linus Torvalds1da177e2005-04-16 15:20:36 -07005034/* Reduce allocated memory if we can, trying to get
5035 * the socket within its memory limits again.
5036 *
5037 * Return less than zero if we should start dropping frames
5038 * until the socket owning process reads some of the data
5039 * to stabilize the situation.
5040 */
5041static int tcp_prune_queue(struct sock *sk)
5042{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005043 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005044
Eric Dumazetc10d9312016-04-29 14:16:47 -07005045 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005046
5047 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005048 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07005049 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005050 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5051
Eric Dumazetf4a33132018-07-23 09:28:18 -07005052 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5053 return 0;
5054
Linus Torvalds1da177e2005-04-16 15:20:36 -07005055 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07005056 if (!skb_queue_empty(&sk->sk_receive_queue))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005057 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
David S. Miller91521942009-05-28 21:35:47 -07005058 skb_peek(&sk->sk_receive_queue),
5059 NULL,
5060 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005061 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005062
5063 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5064 return 0;
5065
5066 /* Collapsing did not help, destructive actions follow.
5067 * This must not ever occur. */
5068
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005069 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070
5071 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5072 return 0;
5073
5074 /* If we are really being abused, tell the caller to silently
5075 * drop receive data on the floor. It will get retransmitted
5076 * and hopefully then we'll have sufficient space.
5077 */
Eric Dumazetc10d9312016-04-29 14:16:47 -07005078 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005079
5080 /* Massive buffer overcommit. */
Florian Westphal31770e32017-08-30 19:24:58 +02005081 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005082 return -1;
5083}
5084
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005085static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07005086{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005087 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005088
David S. Miller0d9901d2005-07-05 15:21:10 -07005089 /* If the user specified a specific send buffer setting, do
5090 * not modify it.
5091 */
5092 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005093 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005094
5095 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07005096 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005097 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005098
5099 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00005100 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005101 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005102
5103 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05005104 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005105 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005106
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005107 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07005108}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005109
5110/* When incoming ACK allowed to free some skb from write_queue,
5111 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5112 * on the exit from tcp input handler.
5113 *
5114 * PROBLEM: sndbuf expansion does not work well with largesend.
5115 */
5116static void tcp_new_space(struct sock *sk)
5117{
5118 struct tcp_sock *tp = tcp_sk(sk);
5119
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005120 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07005121 tcp_sndbuf_expand(sk);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07005122 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005123 }
5124
5125 sk->sk_write_space(sk);
5126}
5127
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005128static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129{
5130 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5131 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
jbaron@akamai.com3c715122015-04-20 20:05:07 +00005132 /* pairs with tcp_poll() */
Jason Baron56d80622017-01-24 21:49:41 -05005133 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134 if (sk->sk_socket &&
Francis Yanb0f71bd2016-11-27 23:07:16 -08005135 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005136 tcp_new_space(sk);
Francis Yanb0f71bd2016-11-27 23:07:16 -08005137 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5138 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5139 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005140 }
5141}
5142
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005143static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005144{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005145 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005146 tcp_check_space(sk);
5147}
5148
5149/*
5150 * Check if sending an ack is needed.
5151 */
5152static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5153{
5154 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005155 unsigned long rtt, delay;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005156
5157 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005158 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005159 /* ... and right edge of window advances far enough.
Eric Dumazet796f82e2018-04-16 10:33:36 -07005160 * (tcp_recvmsg() will send ACK otherwise).
5161 * If application uses SO_RCVLOWAT, we want send ack now if
5162 * we have not received enough bytes to satisfy the condition.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163 */
Eric Dumazet796f82e2018-04-16 10:33:36 -07005164 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5165 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005166 /* We ACK each frame or... */
Yuchung Cheng466466d2018-08-09 09:38:09 -07005167 tcp_in_quickack_mode(sk) ||
5168 /* Protocol state mandates a one-time immediate ACK */
5169 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005170send_now:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005171 tcp_send_ack(sk);
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005172 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005173 }
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005174
5175 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5176 tcp_send_delayed_ack(sk);
5177 return;
5178 }
5179
Eric Dumazet9c21d2f2018-05-17 14:47:29 -07005180 if (!tcp_is_sack(tp) ||
5181 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005182 goto send_now;
Eric Dumazet86de5922018-11-20 05:53:59 -08005183
5184 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5185 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5186 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5187 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5188 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5189 tp->compressed_ack = 0;
5190 }
5191
5192 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5193 goto send_now;
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005194
5195 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5196 return;
5197
Eric Dumazet6d82aa22018-05-17 14:47:28 -07005198 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005199
5200 rtt = tp->rcv_rtt_est.rtt_us;
5201 if (tp->srtt_us && tp->srtt_us < rtt)
5202 rtt = tp->srtt_us;
5203
Eric Dumazet6d82aa22018-05-17 14:47:28 -07005204 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005205 rtt * (NSEC_PER_USEC >> 3)/20);
5206 sock_hold(sk);
5207 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5208 HRTIMER_MODE_REL_PINNED_SOFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005209}
5210
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005211static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005212{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005213 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005214 /* We sent a data segment already. */
5215 return;
5216 }
5217 __tcp_ack_snd_check(sk, 1);
5218}
5219
5220/*
5221 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005222 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005223 * moved inline now as tcp_urg is only called from one
5224 * place. We handle URGent data wrong. We have to - as
5225 * BSD still doesn't use the correction from RFC961.
5226 * For 1003.1g we should support a new option TCP_STDURG to permit
5227 * either form (or just set the sysctl tcp_stdurg).
5228 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005229
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005230static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005231{
5232 struct tcp_sock *tp = tcp_sk(sk);
5233 u32 ptr = ntohs(th->urg_ptr);
5234
Eric Dumazet3f4c7c62017-10-26 21:55:01 -07005235 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005236 ptr--;
5237 ptr += ntohl(th->seq);
5238
5239 /* Ignore urgent data that we've already seen and read. */
5240 if (after(tp->copied_seq, ptr))
5241 return;
5242
5243 /* Do not replay urg ptr.
5244 *
5245 * NOTE: interesting situation not covered by specs.
5246 * Misbehaving sender may send urg ptr, pointing to segment,
5247 * which we already have in ofo queue. We are not able to fetch
5248 * such data and will stay in TCP_URG_NOTYET until will be eaten
5249 * by recvmsg(). Seems, we are not obliged to handle such wicked
5250 * situations. But it is worth to think about possibility of some
5251 * DoSes using some hypothetical application level deadlock.
5252 */
5253 if (before(ptr, tp->rcv_nxt))
5254 return;
5255
5256 /* Do we already have a newer (or duplicate) urgent pointer? */
5257 if (tp->urg_data && !after(ptr, tp->urg_seq))
5258 return;
5259
5260 /* Tell the world about our new urgent pointer. */
5261 sk_send_sigurg(sk);
5262
5263 /* We may be adding urgent data when the last byte read was
5264 * urgent. To do this requires some care. We cannot just ignore
5265 * tp->copied_seq since we would read the last urgent byte again
5266 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005267 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005268 *
5269 * NOTE. Double Dutch. Rendering to plain English: author of comment
5270 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5271 * and expect that both A and B disappear from stream. This is _wrong_.
5272 * Though this happens in BSD with high probability, this is occasional.
5273 * Any application relying on this is buggy. Note also, that fix "works"
5274 * only in this artificial test. Insert some normal data between A and B and we will
5275 * decline of BSD again. Verdict: it is better to remove to trap
5276 * buggy users.
5277 */
5278 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005279 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005280 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5281 tp->copied_seq++;
5282 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005283 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005284 __kfree_skb(skb);
5285 }
5286 }
5287
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005288 tp->urg_data = TCP_URG_NOTYET;
5289 tp->urg_seq = ptr;
Florian Westphal31770e32017-08-30 19:24:58 +02005290
5291 /* Disable header prediction. */
5292 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293}
5294
5295/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005296static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297{
5298 struct tcp_sock *tp = tcp_sk(sk);
5299
5300 /* Check if we get a new urgent pointer - normally not. */
5301 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005302 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005303
5304 /* Do we wait for any urgent data? - normally not... */
5305 if (tp->urg_data == TCP_URG_NOTYET) {
5306 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5307 th->syn;
5308
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005309 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005310 if (ptr < skb->len) {
5311 u8 tmp;
5312 if (skb_copy_bits(skb, ptr, &tmp, 1))
5313 BUG();
5314 tp->urg_data = TCP_URG_VALID | tmp;
5315 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005316 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005317 }
5318 }
5319}
5320
Jason Baron0e40f4c2017-01-17 13:37:19 -05005321/* Accept RST for rcv_nxt - 1 after a FIN.
5322 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5323 * FIN is sent followed by a RST packet. The RST is sent with the same
5324 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5325 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5326 * ACKs on the closed socket. In addition middleboxes can drop either the
5327 * challenge ACK or a subsequent RST.
5328 */
5329static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5330{
5331 struct tcp_sock *tp = tcp_sk(sk);
5332
5333 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5334 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5335 TCPF_CLOSING));
5336}
5337
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005338/* Does PAWS and seqno based validation of an incoming segment, flags will
5339 * play significant role here.
5340 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005341static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5342 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005343{
5344 struct tcp_sock *tp = tcp_sk(sk);
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005345 bool rst_seq_match = false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005346
5347 /* RFC1323: H1. Apply PAWS check first. */
Eric Dumazeteed29f12017-06-07 10:34:36 -07005348 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5349 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005350 tcp_paws_discard(sk, skb)) {
5351 if (!th->rst) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005352 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005353 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5354 LINUX_MIB_TCPACKSKIPPEDPAWS,
5355 &tp->last_oow_ack_time))
5356 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005357 goto discard;
5358 }
5359 /* Reset is accepted even if it did not pass PAWS. */
5360 }
5361
5362 /* Step 1: check sequence number */
5363 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5364 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5365 * (RST) segments are validated by checking their SEQ-fields."
5366 * And page 69: "If an incoming segment is not acceptable,
5367 * an acknowledgment should be sent in reply (unless the RST
5368 * bit is set, if so drop the segment and return)".
5369 */
Eric Dumazete3715892012-07-17 12:29:30 +00005370 if (!th->rst) {
5371 if (th->syn)
5372 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005373 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5374 LINUX_MIB_TCPACKSKIPPEDSEQ,
5375 &tp->last_oow_ack_time))
5376 tcp_send_dupack(sk, skb);
Jason Baron0e40f4c2017-01-17 13:37:19 -05005377 } else if (tcp_reset_check(sk, skb)) {
5378 tcp_reset(sk);
Eric Dumazete3715892012-07-17 12:29:30 +00005379 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005380 goto discard;
5381 }
5382
5383 /* Step 2: check RST bit */
5384 if (th->rst) {
Jason Baron0e40f4c2017-01-17 13:37:19 -05005385 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5386 * FIN and SACK too if available):
5387 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5388 * the right-most SACK block,
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005389 * then
Eric Dumazet282f23c2012-07-17 10:13:05 +02005390 * RESET the connection
5391 * else
5392 * Send a challenge ACK
5393 */
Jason Baron0e40f4c2017-01-17 13:37:19 -05005394 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5395 tcp_reset_check(sk, skb)) {
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005396 rst_seq_match = true;
5397 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5398 struct tcp_sack_block *sp = &tp->selective_acks[0];
5399 int max_sack = sp[0].end_seq;
5400 int this_sack;
5401
5402 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5403 ++this_sack) {
5404 max_sack = after(sp[this_sack].end_seq,
5405 max_sack) ?
5406 sp[this_sack].end_seq : max_sack;
5407 }
5408
5409 if (TCP_SKB_CB(skb)->seq == max_sack)
5410 rst_seq_match = true;
5411 }
5412
5413 if (rst_seq_match)
Eric Dumazet282f23c2012-07-17 10:13:05 +02005414 tcp_reset(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005415 else {
5416 /* Disable TFO if RST is out-of-order
5417 * and no data has been received
5418 * for current active TFO socket
5419 */
5420 if (tp->syn_fastopen && !tp->data_segs_in &&
5421 sk->sk_state == TCP_ESTABLISHED)
Wei Wang46c2fa32017-04-20 14:45:47 -07005422 tcp_fastopen_active_disable(sk);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005423 tcp_send_challenge_ack(sk, skb);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005424 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005425 goto discard;
5426 }
5427
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005428 /* step 3: check security and precedence [ignored] */
5429
Eric Dumazet0c246042012-07-17 01:41:30 +00005430 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005431 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005432 */
5433 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005434syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005435 if (syn_inerr)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005436 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5437 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005438 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005439 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005440 }
5441
Eric Dumazet0c246042012-07-17 01:41:30 +00005442 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005443
5444discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005445 tcp_drop(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005446 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005447}
5448
Linus Torvalds1da177e2005-04-16 15:20:36 -07005449/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005450 * TCP receive function for the ESTABLISHED state.
Florian Westphal31770e32017-08-30 19:24:58 +02005451 *
5452 * It is split into a fast path and a slow path. The fast path is
5453 * disabled when:
5454 * - A zero window was announced from us - zero window probing
5455 * is only handled properly in the slow path.
5456 * - Out of order segments arrived.
5457 * - Urgent data is expected.
5458 * - There is no buffer space left
5459 * - Unexpected TCP flags/window values/header lengths are received
5460 * (detected by checking the TCP header against pred_flags)
5461 * - Data is sent in both directions. Fast path only supports pure senders
5462 * or pure receivers (this means either the sequence number or the ack
5463 * value must stay constant)
5464 * - Unexpected TCP option.
5465 *
5466 * When these conditions are not satisfied it drops into a standard
5467 * receive procedure patterned after RFC793 to handle all cases.
5468 * The first three cases are guaranteed by proper pred_flags setting,
5469 * the rest is checked inline. Fast processing is turned on in
5470 * tcp_data_queue when everything is OK.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005471 */
Yafang Shao3d97d882018-05-29 23:27:31 +08005472void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005473{
Yafang Shao3d97d882018-05-29 23:27:31 +08005474 const struct tcphdr *th = (const struct tcphdr *)skb->data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475 struct tcp_sock *tp = tcp_sk(sk);
Yafang Shao3d97d882018-05-29 23:27:31 +08005476 unsigned int len = skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005477
Masami Hiramatsuc3fde1b2017-12-29 11:45:51 +09005478 /* TCP congestion window tracking */
5479 trace_tcp_probe(sk, skb);
5480
Eric Dumazet9a568de2017-05-16 14:00:14 -07005481 tcp_mstamp_refresh(tp);
Ian Morris51456b22015-04-03 09:17:26 +01005482 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005483 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Florian Westphal31770e32017-08-30 19:24:58 +02005484 /*
5485 * Header prediction.
5486 * The code loosely follows the one in the famous
5487 * "30 instruction TCP receive" Van Jacobson mail.
5488 *
5489 * Van's trick is to deposit buffers into socket queue
5490 * on a device interrupt, to call tcp_recv function
5491 * on the receive process context and checksum and copy
5492 * the buffer to user space. smart...
5493 *
5494 * Our current scheme is not silly either but we take the
5495 * extra cost of the net_bh soft interrupt processing...
5496 * We do checksum and copy also but from device to kernel.
5497 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005498
5499 tp->rx_opt.saw_tstamp = 0;
5500
Florian Westphal31770e32017-08-30 19:24:58 +02005501 /* pred_flags is 0xS?10 << 16 + snd_wnd
5502 * if header_prediction is to be made
5503 * 'S' will always be tp->tcp_header_len >> 2
5504 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5505 * turn it off (when there are holes in the receive
5506 * space for instance)
5507 * PSH flag is ignored.
5508 */
5509
5510 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5511 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5512 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5513 int tcp_header_len = tp->tcp_header_len;
5514
5515 /* Timestamp header prediction: tcp_header_len
5516 * is automatically equal to th->doff*4 due to pred_flags
5517 * match.
5518 */
5519
5520 /* Check timestamp */
5521 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5522 /* No? Slow path! */
5523 if (!tcp_parse_aligned_timestamp(tp, th))
5524 goto slow_path;
5525
5526 /* If PAWS failed, check it more carefully in slow path */
5527 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5528 goto slow_path;
5529
5530 /* DO NOT update ts_recent here, if checksum fails
5531 * and timestamp was corrupted part, it will result
5532 * in a hung connection since we will drop all
5533 * future packets due to the PAWS test.
5534 */
5535 }
5536
5537 if (len <= tcp_header_len) {
5538 /* Bulk data transfer: sender */
5539 if (len == tcp_header_len) {
5540 /* Predicted packet is in window by definition.
5541 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5542 * Hence, check seq<=rcv_wup reduces to:
5543 */
5544 if (tcp_header_len ==
5545 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5546 tp->rcv_nxt == tp->rcv_wup)
5547 tcp_store_ts_recent(tp);
5548
5549 /* We know that such packets are checksummed
5550 * on entry.
5551 */
5552 tcp_ack(sk, skb, 0);
5553 __kfree_skb(skb);
5554 tcp_data_snd_check(sk);
Wei Wang3f6c65d2018-06-19 21:42:50 -07005555 /* When receiving pure ack in fast path, update
5556 * last ts ecr directly instead of calling
5557 * tcp_rcv_rtt_measure_ts()
5558 */
5559 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
Florian Westphal31770e32017-08-30 19:24:58 +02005560 return;
5561 } else { /* Header too small */
5562 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5563 goto discard;
5564 }
5565 } else {
5566 int eaten = 0;
5567 bool fragstolen = false;
5568
5569 if (tcp_checksum_complete(skb))
5570 goto csum_error;
5571
5572 if ((int)skb->truesize > sk->sk_forward_alloc)
5573 goto step5;
5574
5575 /* Predicted packet is in window by definition.
5576 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5577 * Hence, check seq<=rcv_wup reduces to:
5578 */
5579 if (tcp_header_len ==
5580 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5581 tp->rcv_nxt == tp->rcv_wup)
5582 tcp_store_ts_recent(tp);
5583
5584 tcp_rcv_rtt_measure_ts(sk, skb);
5585
5586 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5587
5588 /* Bulk data transfer: receiver */
Eric Dumazete7395f1f2018-11-26 14:49:12 -08005589 __skb_pull(skb, tcp_header_len);
5590 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
Florian Westphal31770e32017-08-30 19:24:58 +02005591
5592 tcp_event_data_recv(sk, skb);
5593
5594 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5595 /* Well, only one small jumplet in fast path... */
5596 tcp_ack(sk, skb, FLAG_DATA);
5597 tcp_data_snd_check(sk);
5598 if (!inet_csk_ack_scheduled(sk))
5599 goto no_ack;
5600 }
5601
5602 __tcp_ack_snd_check(sk, 0);
5603no_ack:
5604 if (eaten)
5605 kfree_skb_partial(skb, fragstolen);
Eric Dumazet03f45c82018-04-16 10:33:37 -07005606 tcp_data_ready(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02005607 return;
5608 }
5609 }
5610
5611slow_path:
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005612 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005613 goto csum_error;
5614
Calvin Owens0c228e82014-11-20 15:09:53 -08005615 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005616 goto discard;
5617
Florian Westphal31770e32017-08-30 19:24:58 +02005618 /*
5619 * Standard slow path.
5620 */
5621
Eric Dumazet0c246042012-07-17 01:41:30 +00005622 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005623 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005624
Florian Westphal31770e32017-08-30 19:24:58 +02005625step5:
5626 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005627 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005628
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005629 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630
5631 /* Process urgent data. */
5632 tcp_urg(sk, skb, th);
5633
5634 /* step 7: process the segment text */
5635 tcp_data_queue(sk, skb);
5636
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005637 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005638 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005639 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005640
5641csum_error:
Eric Dumazetc10d9312016-04-29 14:16:47 -07005642 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5643 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005644
5645discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005646 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005647}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005648EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005649
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005650void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5651{
5652 struct tcp_sock *tp = tcp_sk(sk);
5653 struct inet_connection_sock *icsk = inet_csk(sk);
5654
5655 tcp_set_state(sk, TCP_ESTABLISHED);
Eric Dumazet70eabf02017-05-16 14:00:07 -07005656 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005657
Ian Morris00db4122015-04-03 09:17:27 +01005658 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005659 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005660 security_inet_conn_established(sk, skb);
Amritha Nambiarc6345ce2018-06-29 21:26:57 -07005661 sk_mark_napi_id(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005662 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005663
Wei Wang27204aa2017-10-04 10:03:44 -07005664 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005665
5666 /* Prevent spurious tcp_cwnd_restart() on first data
5667 * packet.
5668 */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07005669 tp->lsndtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005670
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005671 if (sock_flag(sk, SOCK_KEEPOPEN))
5672 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Florian Westphal31770e32017-08-30 19:24:58 +02005673
5674 if (!tp->rx_opt.snd_wscale)
5675 __tcp_fast_path_on(tp, tp->snd_wnd);
5676 else
5677 tp->pred_flags = 0;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005678}
5679
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005680static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5681 struct tcp_fastopen_cookie *cookie)
5682{
5683 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07005684 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07005685 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5686 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005687
5688 if (mss == tp->rx_opt.user_mss) {
5689 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005690
5691 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5692 tcp_clear_options(&opt);
5693 opt.user_mss = opt.mss_clamp = 0;
Eric Dumazeteed29f12017-06-07 10:34:36 -07005694 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005695 mss = opt.mss_clamp;
5696 }
5697
Daniel Lee2646c832015-04-06 14:37:27 -07005698 if (!tp->syn_fastopen) {
5699 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005700 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07005701 } else if (tp->total_retrans) {
5702 /* SYN timed out and the SYN-ACK neither has a cookie nor
5703 * acknowledges data. Presumably the remote received only
5704 * the retransmitted (regular) SYNs: either the original
5705 * SYN-data or the corresponding SYN-ACK was dropped.
5706 */
5707 syn_drop = (cookie->len < 0 && data);
5708 } else if (cookie->len < 0 && !tp->syn_data) {
5709 /* We requested a cookie but didn't get it. If we did not use
5710 * the (old) exp opt format then try so next time (try_exp=1).
5711 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5712 */
5713 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5714 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005715
Daniel Lee2646c832015-04-06 14:37:27 -07005716 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005717
5718 if (data) { /* Retransmit unacked data in SYN */
Eric Dumazet75c119a2017-10-05 22:21:27 -07005719 skb_rbtree_walk_from(data) {
5720 if (__tcp_retransmit_skb(sk, data, 1))
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005721 break;
5722 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005723 tcp_rearm_rto(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005724 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005725 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005726 return true;
5727 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005728 tp->syn_data_acked = tp->syn_data;
Yuchung Chengbef57672018-04-17 23:18:46 -07005729 if (tp->syn_data_acked) {
5730 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5731 /* SYN-data is counted as two separate packets in tcp_ack() */
5732 if (tp->delivered > 1)
5733 --tp->delivered;
5734 }
Eric Dumazet61d2bca2016-02-01 21:03:07 -08005735
5736 tcp_fastopen_add_skb(sk, synack);
5737
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005738 return false;
5739}
5740
Ursula Braun60e2a772017-10-25 11:01:45 +02005741static void smc_check_reset_syn(struct tcp_sock *tp)
5742{
5743#if IS_ENABLED(CONFIG_SMC)
5744 if (static_branch_unlikely(&tcp_have_smc)) {
5745 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5746 tp->syn_smc = 0;
5747 }
5748#endif
5749}
5750
Linus Torvalds1da177e2005-04-16 15:20:36 -07005751static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07005752 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005753{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005754 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005755 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005756 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005757 int saved_clamp = tp->rx_opt.mss_clamp;
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005758 bool fastopen_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005759
Eric Dumazeteed29f12017-06-07 10:34:36 -07005760 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005761 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005762 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005763
5764 if (th->ack) {
5765 /* rfc793:
5766 * "If the state is SYN-SENT then
5767 * first check the ACK bit
5768 * If the ACK bit is set
5769 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5770 * a reset (unless the RST bit is set, if so drop
5771 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005772 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005773 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5774 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005775 goto reset_and_undo;
5776
5777 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5778 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
Eric Dumazet9a568de2017-05-16 14:00:14 -07005779 tcp_time_stamp(tp))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005780 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005781 LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005782 goto reset_and_undo;
5783 }
5784
5785 /* Now ACK is acceptable.
5786 *
5787 * "If the RST bit is set
5788 * If the ACK was acceptable then signal the user "error:
5789 * connection reset", drop the segment, enter CLOSED state,
5790 * delete TCB, and return."
5791 */
5792
5793 if (th->rst) {
5794 tcp_reset(sk);
5795 goto discard;
5796 }
5797
5798 /* rfc793:
5799 * "fifth, if neither of the SYN or RST bits is set then
5800 * drop the segment and return."
5801 *
5802 * See note below!
5803 * --ANK(990513)
5804 */
5805 if (!th->syn)
5806 goto discard_and_undo;
5807
5808 /* rfc793:
5809 * "If the SYN bit is on ...
5810 * are acceptable then ...
5811 * (our SYN has been ACKed), change the connection
5812 * state to ESTABLISHED..."
5813 */
5814
Florian Westphal735d3832014-09-29 13:08:30 +02005815 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005816
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005817 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Florian Westphal31770e32017-08-30 19:24:58 +02005818 tcp_ack(sk, skb, FLAG_SLOWPATH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005819
5820 /* Ok.. it's good. Set up sequence numbers and
5821 * move to established.
5822 */
5823 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5824 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5825
5826 /* RFC1323: The window in SYN & SYN/ACK segments is
5827 * never scaled.
5828 */
5829 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005830
5831 if (!tp->rx_opt.wscale_ok) {
5832 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5833 tp->window_clamp = min(tp->window_clamp, 65535U);
5834 }
5835
5836 if (tp->rx_opt.saw_tstamp) {
5837 tp->rx_opt.tstamp_ok = 1;
5838 tp->tcp_header_len =
5839 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5840 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5841 tcp_store_ts_recent(tp);
5842 } else {
5843 tp->tcp_header_len = sizeof(struct tcphdr);
5844 }
5845
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005846 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005847 tcp_initialize_rcv_mss(sk);
5848
5849 /* Remember, tcp_poll() does not lock socket!
5850 * Change state from SYN-SENT only after copied_seq
5851 * is initialized. */
5852 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005853
Ursula Braun60e2a772017-10-25 11:01:45 +02005854 smc_check_reset_syn(tp);
5855
Ralf Baechlee16aa202006-12-07 00:11:33 -08005856 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005857
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005858 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005859
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005860 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5861 tcp_rcv_fastopen_synack(sk, skb, &foc);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005862
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005863 if (!sock_flag(sk, SOCK_DEAD)) {
5864 sk->sk_state_change(sk);
5865 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5866 }
5867 if (fastopen_fail)
5868 return -1;
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005869 if (sk->sk_write_pending ||
5870 icsk->icsk_accept_queue.rskq_defer_accept ||
Wei Wang31954cd2019-01-25 10:53:19 -08005871 inet_csk_in_pingpong_mode(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005872 /* Save one ACK. Data will be ready after
5873 * several ticks, if write_pending is set.
5874 *
5875 * It may be deleted, but with this feature tcpdumps
5876 * look so _wonderfully_ clever, that I was not able
5877 * to stand against the temptation 8) --ANK
5878 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005879 inet_csk_schedule_ack(sk);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07005880 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005881 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5882 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005883
5884discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005885 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005886 return 0;
5887 } else {
5888 tcp_send_ack(sk);
5889 }
5890 return -1;
5891 }
5892
5893 /* No ACK in the segment */
5894
5895 if (th->rst) {
5896 /* rfc793:
5897 * "If the RST bit is set
5898 *
5899 * Otherwise (no ACK) drop the segment and return."
5900 */
5901
5902 goto discard_and_undo;
5903 }
5904
5905 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005906 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005907 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005908 goto discard_and_undo;
5909
5910 if (th->syn) {
5911 /* We see SYN without ACK. It is attempt of
5912 * simultaneous connect with crossed SYNs.
5913 * Particularly, it can be connect to self.
5914 */
5915 tcp_set_state(sk, TCP_SYN_RECV);
5916
5917 if (tp->rx_opt.saw_tstamp) {
5918 tp->rx_opt.tstamp_ok = 1;
5919 tcp_store_ts_recent(tp);
5920 tp->tcp_header_len =
5921 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5922 } else {
5923 tp->tcp_header_len = sizeof(struct tcphdr);
5924 }
5925
5926 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet142a2e72015-11-26 08:18:14 -08005927 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005928 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5929
5930 /* RFC1323: The window in SYN & SYN/ACK segments is
5931 * never scaled.
5932 */
5933 tp->snd_wnd = ntohs(th->window);
5934 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5935 tp->max_window = tp->snd_wnd;
5936
Florian Westphal735d3832014-09-29 13:08:30 +02005937 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005938
John Heffner5d424d52006-03-20 17:53:41 -08005939 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005940 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005941 tcp_initialize_rcv_mss(sk);
5942
Linus Torvalds1da177e2005-04-16 15:20:36 -07005943 tcp_send_synack(sk);
5944#if 0
5945 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005946 * There are no obstacles to make this (except that we must
5947 * either change tcp_recvmsg() to prevent it from returning data
5948 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005949 *
5950 * However, if we ignore data in ACKless segments sometimes,
5951 * we have no reasons to accept it sometimes.
5952 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5953 * is not flawless. So, discard packet for sanity.
5954 * Uncomment this return to process the data.
5955 */
5956 return -1;
5957#else
5958 goto discard;
5959#endif
5960 }
5961 /* "fifth, if neither of the SYN or RST bits is set then
5962 * drop the segment and return."
5963 */
5964
5965discard_and_undo:
5966 tcp_clear_options(&tp->rx_opt);
5967 tp->rx_opt.mss_clamp = saved_clamp;
5968 goto discard;
5969
5970reset_and_undo:
5971 tcp_clear_options(&tp->rx_opt);
5972 tp->rx_opt.mss_clamp = saved_clamp;
5973 return 1;
5974}
5975
Linus Torvalds1da177e2005-04-16 15:20:36 -07005976/*
5977 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005978 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005979 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5980 * address independent.
5981 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005982
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005983int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005984{
5985 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005986 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005987 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00005988 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005989 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005990 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005991
Linus Torvalds1da177e2005-04-16 15:20:36 -07005992 switch (sk->sk_state) {
5993 case TCP_CLOSE:
5994 goto discard;
5995
5996 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005997 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005998 return 1;
5999
Stephen Hemminger2de979b2007-03-08 20:45:19 -08006000 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006001 goto discard;
6002
Stephen Hemminger2de979b2007-03-08 20:45:19 -08006003 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00006004 if (th->fin)
6005 goto discard;
Eric Dumazet449809a2017-03-01 08:39:49 -08006006 /* It is possible that we process SYN packets from backlog,
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006007 * so we need to make sure to disable BH and RCU right there.
Eric Dumazet449809a2017-03-01 08:39:49 -08006008 */
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006009 rcu_read_lock();
Eric Dumazet449809a2017-03-01 08:39:49 -08006010 local_bh_disable();
6011 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6012 local_bh_enable();
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006013 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006014
Eric Dumazet449809a2017-03-01 08:39:49 -08006015 if (!acceptable)
6016 return 1;
Eric Dumazet0aea76d2016-04-21 22:13:01 -07006017 consume_skb(skb);
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08006018 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006019 }
6020 goto discard;
6021
6022 case TCP_SYN_SENT:
Eric Dumazet8804b272016-04-13 22:05:40 -07006023 tp->rx_opt.saw_tstamp = 0;
Eric Dumazet9a568de2017-05-16 14:00:14 -07006024 tcp_mstamp_refresh(tp);
Eric Dumazetbda07a62015-09-29 07:42:40 -07006025 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006026 if (queued >= 0)
6027 return queued;
6028
6029 /* Do step6 onward by hand. */
6030 tcp_urg(sk, skb, th);
6031 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006032 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006033 return 0;
6034 }
6035
Eric Dumazet9a568de2017-05-16 14:00:14 -07006036 tcp_mstamp_refresh(tp);
Eric Dumazet8804b272016-04-13 22:05:40 -07006037 tp->rx_opt.saw_tstamp = 0;
Jerry Chu168a8f52012-08-31 12:29:13 +00006038 req = tp->fastopen_rsk;
Ian Morris00db4122015-04-03 09:17:27 +01006039 if (req) {
Eric Dumazete0f97592018-02-13 06:14:12 -08006040 bool req_stolen;
6041
Jerry Chu37561f62012-10-22 11:26:36 +00006042 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00006043 sk->sk_state != TCP_FIN_WAIT1);
6044
Eric Dumazete0f97592018-02-13 06:14:12 -08006045 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
Jerry Chu168a8f52012-08-31 12:29:13 +00006046 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00006047 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006048
Calvin Owens0c228e82014-11-20 15:09:53 -08006049 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006050 goto discard;
6051
Neal Cardwelle69bebd2012-09-22 04:18:57 +00006052 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00006053 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006054
6055 /* step 5: check the ACK field */
Florian Westphal31770e32017-08-30 19:24:58 +02006056 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6057 FLAG_UPDATE_TS_RECENT |
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07006058 FLAG_NO_CHALLENGE_ACK) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006059
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07006060 if (!acceptable) {
6061 if (sk->sk_state == TCP_SYN_RECV)
6062 return 1; /* send one RST */
6063 tcp_send_challenge_ack(sk, skb);
6064 goto discard;
6065 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006066 switch (sk->sk_state) {
6067 case TCP_SYN_RECV:
Yuchung Chengbef57672018-04-17 23:18:46 -07006068 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07006069 if (!tp->srtt_us)
6070 tcp_synack_rtt_meas(sk, req);
6071
Joe Perches61eb90032013-05-24 18:36:13 +00006072 /* Once we leave TCP_SYN_RECV, we no longer need req
6073 * so release it.
6074 */
6075 if (req) {
Yuchung Cheng7e32b442016-09-21 16:16:15 -07006076 inet_csk(sk)->icsk_retransmits = 0;
Joe Perches61eb90032013-05-24 18:36:13 +00006077 reqsk_fastopen_remove(sk, req, false);
Wei Wang6d050812017-10-04 10:04:04 -07006078 /* Re-arm the timer because data may have been sent out.
6079 * This is similar to the regular data transmission case
6080 * when new data has just been ack'ed.
6081 *
6082 * (TFO) - we could try to be more aggressive and
6083 * retransmitting any data sooner based on when they
6084 * are sent out.
6085 */
6086 tcp_rearm_rto(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00006087 } else {
Wei Wang27204aa2017-10-04 10:03:44 -07006088 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
Joe Perches61eb90032013-05-24 18:36:13 +00006089 tp->copied_seq = tp->rcv_nxt;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006090 }
Joe Perches61eb90032013-05-24 18:36:13 +00006091 smp_mb();
6092 tcp_set_state(sk, TCP_ESTABLISHED);
6093 sk->sk_state_change(sk);
6094
6095 /* Note, that this wakeup is only for marginal crossed SYN case.
6096 * Passively open sockets are not waked up, because
6097 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6098 */
6099 if (sk->sk_socket)
6100 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6101
6102 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6103 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6104 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6105
6106 if (tp->rx_opt.tstamp_ok)
6107 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6108
Yuchung Chengc0402762016-09-19 23:39:21 -04006109 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6110 tcp_update_pacing_rate(sk);
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04006111
Joe Perches61eb90032013-05-24 18:36:13 +00006112 /* Prevent spurious tcp_cwnd_restart() on first data packet */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07006113 tp->lsndtime = tcp_jiffies32;
Joe Perches61eb90032013-05-24 18:36:13 +00006114
6115 tcp_initialize_rcv_mss(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02006116 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006117 break;
6118
Joe Perchesc48b22d2013-05-24 18:06:58 +00006119 case TCP_FIN_WAIT1: {
Joe Perchesc48b22d2013-05-24 18:06:58 +00006120 int tmo;
6121
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006122 /* If we enter the TCP_FIN_WAIT1 state and we are a
6123 * Fast Open socket and this is the first acceptable
6124 * ACK we have received, this would have acknowledged
6125 * our SYNACK so stop the SYNACK timer.
6126 */
Ian Morris00db4122015-04-03 09:17:27 +01006127 if (req) {
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006128 /* We no longer need the request sock. */
6129 reqsk_fastopen_remove(sk, req, false);
6130 tcp_rearm_rto(sk);
6131 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00006132 if (tp->snd_una != tp->write_seq)
6133 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006134
Joe Perchesc48b22d2013-05-24 18:06:58 +00006135 tcp_set_state(sk, TCP_FIN_WAIT2);
6136 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006137
Julian Anastasovc3a2e832017-02-06 23:14:14 +02006138 sk_dst_confirm(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006139
Joe Perchesc48b22d2013-05-24 18:06:58 +00006140 if (!sock_flag(sk, SOCK_DEAD)) {
6141 /* Wake up lingering close() */
6142 sk->sk_state_change(sk);
6143 break;
6144 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006145
Wei Wangcf1ef3f2017-04-20 14:45:46 -07006146 if (tp->linger2 < 0) {
6147 tcp_done(sk);
6148 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6149 return 1;
6150 }
6151 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6152 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6153 /* Receive out of order FIN after close() */
6154 if (tp->syn_fastopen && th->fin)
Wei Wang46c2fa32017-04-20 14:45:47 -07006155 tcp_fastopen_active_disable(sk);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006156 tcp_done(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07006157 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006158 return 1;
6159 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006160
Joe Perchesc48b22d2013-05-24 18:06:58 +00006161 tmo = tcp_fin_time(sk);
6162 if (tmo > TCP_TIMEWAIT_LEN) {
6163 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6164 } else if (th->fin || sock_owned_by_user(sk)) {
6165 /* Bad case. We could lose such FIN otherwise.
6166 * It is not a big problem, but it looks confusing
6167 * and not so rare event. We still can lose it now,
6168 * if it spins in bh_lock_sock(), but it is really
6169 * marginal case.
6170 */
6171 inet_csk_reset_keepalive_timer(sk, tmo);
6172 } else {
6173 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6174 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006175 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006176 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00006177 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006178
6179 case TCP_CLOSING:
6180 if (tp->snd_una == tp->write_seq) {
6181 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6182 goto discard;
6183 }
6184 break;
6185
6186 case TCP_LAST_ACK:
6187 if (tp->snd_una == tp->write_seq) {
6188 tcp_update_metrics(sk);
6189 tcp_done(sk);
6190 goto discard;
6191 }
6192 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006193 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006194
6195 /* step 6: check the URG bit */
6196 tcp_urg(sk, skb, th);
6197
6198 /* step 7: process the segment text */
6199 switch (sk->sk_state) {
6200 case TCP_CLOSE_WAIT:
6201 case TCP_CLOSING:
6202 case TCP_LAST_ACK:
6203 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6204 break;
Gustavo A. R. Silvafcfd6df2017-10-16 15:48:55 -05006205 /* fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006206 case TCP_FIN_WAIT1:
6207 case TCP_FIN_WAIT2:
6208 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006209 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006210 * BSD 4.4 also does reset.
6211 */
6212 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6213 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6214 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006215 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006216 tcp_reset(sk);
6217 return 1;
6218 }
6219 }
6220 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006221 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006222 tcp_data_queue(sk, skb);
6223 queued = 1;
6224 break;
6225 }
6226
6227 /* tcp_data could move socket to TIME-WAIT */
6228 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006229 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006230 tcp_ack_snd_check(sk);
6231 }
6232
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006233 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006234discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006235 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006236 }
6237 return 0;
6238}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006239EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006240
6241static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6242{
6243 struct inet_request_sock *ireq = inet_rsk(req);
6244
6245 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006246 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6247 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006248#if IS_ENABLED(CONFIG_IPV6)
6249 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006250 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6251 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006252#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006253}
6254
Florian Westphald82bd122014-09-29 13:08:29 +02006255/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6256 *
6257 * If we receive a SYN packet with these bits set, it means a
6258 * network is playing bad games with TOS bits. In order to
6259 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006260 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006261 *
6262 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006263 * congestion control: Linux DCTCP asserts ECT on all packets,
6264 * including SYN, which is most optimal solution; however,
6265 * others, such as FreeBSD do not.
Florian Westphald82bd122014-09-29 13:08:29 +02006266 */
6267static void tcp_ecn_create_request(struct request_sock *req,
6268 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006269 const struct sock *listen_sk,
6270 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006271{
6272 const struct tcphdr *th = tcp_hdr(skb);
6273 const struct net *net = sock_net(listen_sk);
6274 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006275 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006276 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006277
6278 if (!th_ecn)
6279 return;
6280
6281 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006282 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6283 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006284
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006285 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
Lawrence Brakmo91b5b212017-06-30 20:02:49 -07006286 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6287 tcp_bpf_ca_needs_ecn((struct sock *)req))
Florian Westphald82bd122014-09-29 13:08:29 +02006288 inet_rsk(req)->ecn_ok = 1;
6289}
6290
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006291static void tcp_openreq_init(struct request_sock *req,
6292 const struct tcp_options_received *rx_opt,
6293 struct sk_buff *skb, const struct sock *sk)
6294{
6295 struct inet_request_sock *ireq = inet_rsk(req);
6296
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006297 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006298 req->cookie_ts = 0;
6299 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6300 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet9a568de2017-05-16 14:00:14 -07006301 tcp_rsk(req)->snt_synack = tcp_clock_us();
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006302 tcp_rsk(req)->last_oow_ack_time = 0;
6303 req->mss = rx_opt->mss_clamp;
6304 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6305 ireq->tstamp_ok = rx_opt->tstamp_ok;
6306 ireq->sack_ok = rx_opt->sack_ok;
6307 ireq->snd_wscale = rx_opt->snd_wscale;
6308 ireq->wscale_ok = rx_opt->wscale_ok;
6309 ireq->acked = 0;
6310 ireq->ecn_ok = 0;
6311 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6312 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6313 ireq->ir_mark = inet_request_mark(sk, skb);
Ursula Braun60e2a772017-10-25 11:01:45 +02006314#if IS_ENABLED(CONFIG_SMC)
6315 ireq->smc_ok = rx_opt->smc_ok;
6316#endif
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006317}
6318
Eric Dumazete49bb332015-03-17 18:32:27 -07006319struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006320 struct sock *sk_listener,
6321 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006322{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006323 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6324 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006325
6326 if (req) {
6327 struct inet_request_sock *ireq = inet_rsk(req);
6328
Eric Dumazetc92e8c02017-10-20 09:04:13 -07006329 ireq->ireq_opt = NULL;
Huw Davies56ac42b2016-06-27 15:05:28 -04006330#if IS_ENABLED(CONFIG_IPV6)
6331 ireq->pktopts = NULL;
6332#endif
Yafang Shaoa06ac0d2018-04-24 23:07:45 +08006333 atomic64_set(&ireq->ir_cookie, 0);
Eric Dumazete49bb332015-03-17 18:32:27 -07006334 ireq->ireq_state = TCP_NEW_SYN_RECV;
6335 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006336 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006337 }
6338
6339 return req;
6340}
6341EXPORT_SYMBOL(inet_reqsk_alloc);
6342
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006343/*
6344 * Return true if a syncookie should be sent
6345 */
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006346static bool tcp_syn_flood_action(const struct sock *sk,
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006347 const struct sk_buff *skb,
6348 const char *proto)
6349{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006350 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006351 const char *msg = "Dropping request";
6352 bool want_cookie = false;
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006353 struct net *net = sock_net(sk);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006354
6355#ifdef CONFIG_SYN_COOKIES
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006356 if (net->ipv4.sysctl_tcp_syncookies) {
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006357 msg = "Sending cookies";
6358 want_cookie = true;
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006359 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006360 } else
6361#endif
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006362 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006363
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006364 if (!queue->synflood_warned &&
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006365 net->ipv4.sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006366 xchg(&queue->synflood_warned, 1) == 0)
Willem de Bruijn0297c1c2018-09-09 19:12:12 -04006367 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6368 proto, ntohs(tcp_hdr(skb)->dest), msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006369
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006370 return want_cookie;
6371}
6372
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006373static void tcp_reqsk_record_syn(const struct sock *sk,
6374 struct request_sock *req,
6375 const struct sk_buff *skb)
6376{
6377 if (tcp_sk(sk)->save_syn) {
6378 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6379 u32 *copy;
6380
6381 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6382 if (copy) {
6383 copy[0] = len;
6384 memcpy(&copy[1], skb_network_header(skb), len);
6385 req->saved_syn = copy;
6386 }
6387 }
6388}
6389
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006390int tcp_conn_request(struct request_sock_ops *rsk_ops,
6391 const struct tcp_request_sock_ops *af_ops,
6392 struct sock *sk, struct sk_buff *skb)
6393{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006394 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006395 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6396 struct tcp_options_received tmp_opt;
6397 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006398 struct net *net = sock_net(sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006399 struct sock *fastopen_sk = NULL;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006400 struct request_sock *req;
6401 bool want_cookie = false;
Eric Dumazeteaa72dc2017-08-29 15:16:01 -07006402 struct dst_entry *dst;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006403 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006404
6405 /* TW buckets are converted to open requests without
6406 * limitations, they conserve resources and peer is
6407 * evidently real one.
6408 */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006409 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006410 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6411 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6412 if (!want_cookie)
6413 goto drop;
6414 }
6415
Eric Dumazet5ea8ea22016-10-26 09:27:57 -07006416 if (sk_acceptq_is_full(sk)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006417 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006418 goto drop;
6419 }
6420
Eric Dumazeta1a53442015-10-04 21:08:11 -07006421 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006422 if (!req)
6423 goto drop;
6424
6425 tcp_rsk(req)->af_specific = af_ops;
Florian Westphal95a22ca2016-12-01 11:32:06 +01006426 tcp_rsk(req)->ts_off = 0;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006427
6428 tcp_clear_options(&tmp_opt);
6429 tmp_opt.mss_clamp = af_ops->mss_clamp;
6430 tmp_opt.user_mss = tp->rx_opt.user_mss;
Eric Dumazeteed29f12017-06-07 10:34:36 -07006431 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6432 want_cookie ? NULL : &foc);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006433
6434 if (want_cookie && !tmp_opt.saw_tstamp)
6435 tcp_clear_options(&tmp_opt);
6436
Hans Wippelbc58a1b2018-03-23 11:05:45 +01006437 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6438 tmp_opt.smc_ok = 0;
6439
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006440 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6441 tcp_openreq_init(req, &tmp_opt, skb, sk);
KOVACS Krisztian7a682572016-09-23 11:27:42 +02006442 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006443
Eric Dumazet16f86162015-03-13 15:51:10 -07006444 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
David Ahern6dd9a142015-12-16 13:20:44 -08006445 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
Eric Dumazet16f86162015-03-13 15:51:10 -07006446
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006447 af_ops->init_req(req, sk, skb);
6448
6449 if (security_inet_conn_request(sk, skb, req))
6450 goto drop_and_free;
6451
Eric Dumazet84b114b2017-05-05 06:56:54 -07006452 if (tmp_opt.tstamp_ok)
Eric Dumazet5d2ed052017-06-07 10:34:39 -07006453 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
Florian Westphal95a22ca2016-12-01 11:32:06 +01006454
Tonghao Zhang49c71582017-08-21 23:33:48 -07006455 dst = af_ops->route_req(sk, &fl, req);
6456 if (!dst)
6457 goto drop_and_free;
6458
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006459 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006460 /* Kill the following clause, if you dislike this way. */
Soheil Hassas Yeganeh4396e462017-03-15 16:30:46 -04006461 if (!net->ipv4.sysctl_tcp_syncookies &&
6462 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6463 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6464 !tcp_peer_is_proven(req, dst)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006465 /* Without syncookies last quarter of
6466 * backlog is filled with destinations,
6467 * proven to be alive.
6468 * It means that we continue to communicate
6469 * to destinations, already remembered
6470 * to the moment of synflood.
6471 */
6472 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6473 rsk_ops->family);
6474 goto drop_and_release;
6475 }
6476
Eric Dumazet84b114b2017-05-05 06:56:54 -07006477 isn = af_ops->init_seq(skb);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006478 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006479
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006480 tcp_ecn_create_request(req, skb, sk, dst);
6481
6482 if (want_cookie) {
6483 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6484 req->cookie_ts = tmp_opt.tstamp_ok;
6485 if (!tmp_opt.tstamp_ok)
6486 inet_rsk(req)->ecn_ok = 0;
6487 }
6488
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006489 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006490 tcp_rsk(req)->txhash = net_tx_rndhash();
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006491 tcp_openreq_init_rwin(req, sk, dst);
Amritha Nambiarc6345ce2018-06-29 21:26:57 -07006492 sk_rx_queue_set(req_to_sk(req), skb);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006493 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006494 tcp_reqsk_record_syn(sk, req, skb);
Christoph Paasch71c02372017-10-23 13:22:23 -07006495 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006496 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006497 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006498 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Eric Dumazetb3d05142016-04-13 22:05:39 -07006499 &foc, TCP_SYNACK_FASTOPEN);
Eric Dumazet7656d842015-10-04 21:08:07 -07006500 /* Add the child socket directly into the accept queue */
Guillaume Nault9d3e1362019-03-08 22:09:47 +01006501 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6502 reqsk_fastopen_remove(fastopen_sk, req, false);
6503 bh_unlock_sock(fastopen_sk);
6504 sock_put(fastopen_sk);
6505 reqsk_put(req);
6506 goto drop;
6507 }
Eric Dumazet7656d842015-10-04 21:08:07 -07006508 sk->sk_data_ready(sk);
6509 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006510 sock_put(fastopen_sk);
6511 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006512 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006513 if (!want_cookie)
Lawrence Brakmo8550f322017-06-30 20:02:42 -07006514 inet_csk_reqsk_queue_hash_add(sk, req,
6515 tcp_timeout_init((struct sock *)req));
Eric Dumazetb3d05142016-04-13 22:05:39 -07006516 af_ops->send_synack(sk, dst, &fl, req, &foc,
6517 !want_cookie ? TCP_SYNACK_NORMAL :
6518 TCP_SYNACK_COOKIE);
Eric Dumazet9caad862016-04-01 08:52:20 -07006519 if (want_cookie) {
6520 reqsk_free(req);
6521 return 0;
6522 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006523 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006524 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006525 return 0;
6526
6527drop_and_release:
6528 dst_release(dst);
6529drop_and_free:
6530 reqsk_free(req);
6531drop:
Eric Dumazet9caad862016-04-01 08:52:20 -07006532 tcp_listendrop(sk);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006533 return 0;
6534}
6535EXPORT_SYMBOL(tcp_conn_request);