blob: 06e2dbc2b4a212a054fd88e57bb902c55a171b11 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes:
23 * Pedro Roque : Fast Retransmit/Recovery.
24 * Two receive queues.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
28 * Header prediction.
29 * Variable renaming.
30 *
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080043 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 * timestamps.
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
47 * data segments.
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090051 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090053 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * Andi Kleen: Process packets with PSH set in the
55 * fast path.
56 * J Hadi Salim: ECN support
57 * Andrei Gurtov,
58 * Pasi Sarolahti,
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 */
63
Joe Perchesafd465032012-03-12 07:03:32 +000064#define pr_fmt(fmt) "TCP: " fmt
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090067#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070068#include <linux/module.h>
69#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070070#include <linux/kernel.h>
Eric Dumazetad971f62014-10-11 15:17:29 -070071#include <linux/prefetch.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070072#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070073#include <net/tcp.h>
74#include <net/inet_common.h>
75#include <linux/ipsec.h>
76#include <asm/unaligned.h>
Willem de Bruijne1c8a602014-08-04 22:11:50 -040077#include <linux/errqueue.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Brian Haleyab32ea52006-09-22 14:15:41 -070079int sysctl_tcp_timestamps __read_mostly = 1;
80int sysctl_tcp_window_scaling __read_mostly = 1;
81int sysctl_tcp_sack __read_mostly = 1;
Yuchung Cheng94bdc972017-01-12 22:11:42 -080082int sysctl_tcp_fack __read_mostly;
Eric Dumazetdca145f2014-10-27 21:45:24 -070083int sysctl_tcp_max_reordering __read_mostly = 300;
Brian Haleyab32ea52006-09-22 14:15:41 -070084int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000086int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000087EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Eric Dumazet282f23c2012-07-17 10:13:05 +020089/* rfc5961 challenge ack rate limiting */
Eric Dumazet75ff39c2016-07-10 10:04:02 +020090int sysctl_tcp_challenge_ack_limit = 1000;
Eric Dumazet282f23c2012-07-17 10:13:05 +020091
Brian Haleyab32ea52006-09-22 14:15:41 -070092int sysctl_tcp_stdurg __read_mostly;
93int sysctl_tcp_rfc1337 __read_mostly;
94int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070095int sysctl_tcp_frto __read_mostly = 2;
Yuchung Chengf6722582015-10-16 21:57:42 -070096int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
Brian Haleyab32ea52006-09-22 14:15:41 -070097int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000098int sysctl_tcp_early_retrans __read_mostly = 3;
Neal Cardwell032ee422015-02-06 16:04:38 -050099int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101#define FLAG_DATA 0x01 /* Incoming frame contained data. */
102#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106#define FLAG_DATA_SACKED 0x20 /* New SACK. */
107#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -0700108#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +0000110#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700111#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -0700112#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800113#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +0000114#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115
116#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700122#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
Yuchung Chenge662ca42016-02-02 10:33:04 -0800124#define REXMIT_NONE 0 /* no loss recovery to do */
125#define REXMIT_LOST 1 /* retransmit packets marked lost */
126#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
127
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300128static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
129 unsigned int len)
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200130{
131 static bool __once __read_mostly;
132
133 if (!__once) {
134 struct net_device *dev;
135
136 __once = true;
137
138 rcu_read_lock();
139 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300140 if (!dev || len >= dev->mtu)
141 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
142 dev ? dev->name : "Unknown driver");
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200143 rcu_read_unlock();
144 }
145}
146
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900147/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900149 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800150static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700152 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900153 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700154 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900156 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
158 /* skb->len may jitter because of SACKs, even if peer
159 * sends good full-sized frames.
160 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800161 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700162 if (len >= icsk->icsk_ack.rcv_mss) {
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200163 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
164 tcp_sk(sk)->advmss);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300165 /* Account for possibly-removed options */
166 if (unlikely(len > icsk->icsk_ack.rcv_mss +
167 MAX_TCP_OPTION_SPACE))
168 tcp_gro_dev_warn(sk, skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 } else {
170 /* Otherwise, we make more careful check taking into account,
171 * that SACKs block is variable.
172 *
173 * "len" is invariant segment length, including TCP header.
174 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700175 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000176 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 /* If PSH is not set, packet should be
178 * full sized, provided peer TCP is not badly broken.
179 * This observation (if it is correct 8)) allows
180 * to handle super-low mtu links fairly.
181 */
182 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700183 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 /* Subtract also invariant (if peer is RFC compliant),
185 * tcp header plus fixed timestamp option length.
186 * Resulting "len" is MSS free of SACK jitter.
187 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700188 len -= tcp_sk(sk)->tcp_header_len;
189 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700191 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 return;
193 }
194 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700195 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
196 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700197 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 }
199}
200
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700201static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700203 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000204 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800206 if (quickacks == 0)
207 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700208 if (quickacks > icsk->icsk_ack.quick)
209 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210}
211
stephen hemminger1b9f4092010-09-28 19:30:14 +0000212static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700214 struct inet_connection_sock *icsk = inet_csk(sk);
215 tcp_incr_quickack(sk);
216 icsk->icsk_ack.pingpong = 0;
217 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218}
219
220/* Send ACKs quickly, if "quick" count is not exhausted
221 * and the session is not interactive.
222 */
223
Jon Maxwell2251ae42015-07-08 10:12:28 +1000224static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700226 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000227 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000228
Jon Maxwell2251ae42015-07-08 10:12:28 +1000229 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
230 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
Florian Westphal735d3832014-09-29 13:08:30 +0200233static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700234{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800235 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700236 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
237}
238
Florian Westphal735d3832014-09-29 13:08:30 +0200239static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700240{
241 if (tcp_hdr(skb)->cwr)
242 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
243}
244
Florian Westphal735d3832014-09-29 13:08:30 +0200245static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700246{
247 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
248}
249
Florian Westphal735d3832014-09-29 13:08:30 +0200250static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700251{
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400252 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000253 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700254 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000255 * and we already seen ECT on a previous segment,
256 * it is probably a retransmit.
257 */
258 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700259 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000260 break;
261 case INET_ECN_CE:
Florian Westphal98900922014-09-26 22:37:35 +0200262 if (tcp_ca_needs_ecn((struct sock *)tp))
263 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
264
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000265 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
266 /* Better not delay acks, sender can have a very low cwnd */
267 tcp_enter_quickack_mode((struct sock *)tp);
268 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
269 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000270 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200271 break;
272 default:
273 if (tcp_ca_needs_ecn((struct sock *)tp))
274 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
275 tp->ecn_flags |= TCP_ECN_SEEN;
276 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700277 }
278}
279
Florian Westphal735d3832014-09-29 13:08:30 +0200280static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
281{
282 if (tp->ecn_flags & TCP_ECN_OK)
283 __tcp_ecn_check_ce(tp, skb);
284}
285
286static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700287{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800288 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700289 tp->ecn_flags &= ~TCP_ECN_OK;
290}
291
Florian Westphal735d3832014-09-29 13:08:30 +0200292static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700293{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800294 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700295 tp->ecn_flags &= ~TCP_ECN_OK;
296}
297
Florian Westphal735d3832014-09-29 13:08:30 +0200298static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700299{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800300 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000301 return true;
302 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700303}
304
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305/* Buffer size and advertised window tuning.
306 *
307 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
308 */
309
Eric Dumazet6ae70532013-10-01 10:23:44 -0700310static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700312 const struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400313 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700314 int sndmem, per_mss;
315 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
Eric Dumazet6ae70532013-10-01 10:23:44 -0700317 /* Worst case is non GSO/TSO : each frame consumes one skb
318 * and skb->head is kmalloced using power of two area of memory
319 */
320 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
321 MAX_TCP_HEADER +
322 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
323
324 per_mss = roundup_pow_of_two(per_mss) +
325 SKB_DATA_ALIGN(sizeof(struct sk_buff));
326
327 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
328 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
329
330 /* Fast Recovery (RFC 5681 3.2) :
331 * Cubic needs 1.7 factor, rounded to 2 to include
332 * extra cushion (application might react slowly to POLLOUT)
333 */
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400334 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
335 sndmem *= nr_segs * per_mss;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700336
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000337 if (sk->sk_sndbuf < sndmem)
338 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339}
340
341/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
342 *
343 * All tcp_full_space() is split to two parts: "network" buffer, allocated
344 * forward and advertised in receiver window (tp->rcv_wnd) and
345 * "application buffer", required to isolate scheduling/application
346 * latencies from network.
347 * window_clamp is maximal advertised window. It can be less than
348 * tcp_full_space(), in this case tcp_full_space() - window_clamp
349 * is reserved for "application" buffer. The less window_clamp is
350 * the smoother our behaviour from viewpoint of network, but the lower
351 * throughput and the higher sensitivity of the connection to losses. 8)
352 *
353 * rcv_ssthresh is more strict window_clamp used at "slow start"
354 * phase to predict further behaviour of this connection.
355 * It is used for two goals:
356 * - to enforce header prediction at sender, even when application
357 * requires some significant "application buffer". It is check #1.
358 * - to prevent pruning of receive queue because of misprediction
359 * of receiver window. Check #2.
360 *
361 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800362 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 * in common situations. Otherwise, we have to rely on queue collapsing.
364 */
365
366/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700367static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700369 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800371 int truesize = tcp_win_from_space(skb->truesize) >> 1;
372 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374 while (tp->rcv_ssthresh <= window) {
375 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700376 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378 truesize >>= 1;
379 window >>= 1;
380 }
381 return 0;
382}
383
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400384static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700386 struct tcp_sock *tp = tcp_sk(sk);
387
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 /* Check #1 */
389 if (tp->rcv_ssthresh < tp->window_clamp &&
390 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700391 !tcp_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 int incr;
393
394 /* Check #2. Increase window, if skb with such overhead
395 * will fit to rcvbuf in future.
396 */
397 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800398 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700400 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000403 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800404 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
405 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700406 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 }
408 }
409}
410
411/* 3. Tuning rcvbuf, when connection enters established state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412static void tcp_fixup_rcvbuf(struct sock *sk)
413{
Eric Dumazete9266a02011-10-20 16:53:56 -0400414 u32 mss = tcp_sk(sk)->advmss;
Eric Dumazete9266a02011-10-20 16:53:56 -0400415 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Yuchung Cheng85f16522013-06-11 15:35:32 -0700417 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
418 tcp_default_init_rwnd(mss);
Eric Dumazete9266a02011-10-20 16:53:56 -0400419
Eric Dumazetb0983d32013-09-20 13:56:58 -0700420 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
421 * Allow enough cushion so that sender is not limited by our window
422 */
423 if (sysctl_tcp_moderate_rcvbuf)
424 rcvmem <<= 2;
425
Eric Dumazete9266a02011-10-20 16:53:56 -0400426 if (sk->sk_rcvbuf < rcvmem)
427 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800430/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 * established state.
432 */
Jerry Chu10467162012-08-31 12:29:11 +0000433void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
435 struct tcp_sock *tp = tcp_sk(sk);
436 int maxwin;
437
438 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
439 tcp_fixup_rcvbuf(sk);
440 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700441 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442
443 tp->rcvq_space.space = tp->rcv_wnd;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700444 skb_mstamp_get(&tp->tcp_mstamp);
445 tp->rcvq_space.time = tp->tcp_mstamp;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700446 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
448 maxwin = tcp_full_space(sk);
449
450 if (tp->window_clamp >= maxwin) {
451 tp->window_clamp = maxwin;
452
453 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
454 tp->window_clamp = max(maxwin -
455 (maxwin >> sysctl_tcp_app_win),
456 4 * tp->advmss);
457 }
458
459 /* Force reservation of one segment. */
460 if (sysctl_tcp_app_win &&
461 tp->window_clamp > 2 * tp->advmss &&
462 tp->window_clamp + tp->advmss > maxwin)
463 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
464
465 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
466 tp->snd_cwnd_stamp = tcp_time_stamp;
467}
468
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700470static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700472 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300473 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300475 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
John Heffner326f36e2005-11-10 17:11:48 -0800477 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
478 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700479 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000480 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800481 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
482 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 }
John Heffner326f36e2005-11-10 17:11:48 -0800484 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800485 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486}
487
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800488/* Initialize RCV_MSS value.
489 * RCV_MSS is an our guess about MSS used by the peer.
490 * We haven't any direct information about the MSS.
491 * It's better to underestimate the RCV_MSS rather than overestimate.
492 * Overestimations make us ACKing less frequently than needed.
493 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
494 */
495void tcp_initialize_rcv_mss(struct sock *sk)
496{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400497 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800498 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
499
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800500 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000501 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800502 hint = max(hint, TCP_MIN_MSS);
503
504 inet_csk(sk)->icsk_ack.rcv_mss = hint;
505}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000506EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800507
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508/* Receiver "autotuning" code.
509 *
510 * The algorithm for RTT estimation w/o timestamps is based on
511 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200512 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 *
514 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200515 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 * though this reference is out of date. A new paper
517 * is pending.
518 */
519static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
520{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700521 u32 new_sample = tp->rcv_rtt_est.rtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 long m = sample;
523
524 if (m == 0)
525 m = 1;
526
527 if (new_sample != 0) {
528 /* If we sample in larger samples in the non-timestamp
529 * case, we could grossly overestimate the RTT especially
530 * with chatty applications or bulk transfer apps which
531 * are stalled on filesystem I/O.
532 *
533 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800534 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800535 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 * long.
537 */
538 if (!win_dep) {
539 m -= (new_sample >> 3);
540 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000541 } else {
542 m <<= 3;
543 if (m < new_sample)
544 new_sample = m;
545 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800547 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 new_sample = m << 3;
549 }
550
Eric Dumazet645f4c62017-04-25 10:15:41 -0700551 tp->rcv_rtt_est.rtt_us = new_sample;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552}
553
554static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
555{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700556 u32 delta_us;
557
558 if (tp->rcv_rtt_est.time.v64 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 goto new_measure;
560 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
561 return;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700562 delta_us = skb_mstamp_us_delta(&tp->tcp_mstamp, &tp->rcv_rtt_est.time);
563 tcp_rcv_rtt_update(tp, delta_us, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
565new_measure:
566 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700567 tp->rcv_rtt_est.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568}
569
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800570static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
571 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700573 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 if (tp->rx_opt.rcv_tsecr &&
575 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700576 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Eric Dumazet645f4c62017-04-25 10:15:41 -0700577 tcp_rcv_rtt_update(tp,
578 jiffies_to_usecs(tcp_time_stamp -
579 tp->rx_opt.rcv_tsecr),
580 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581}
582
583/*
584 * This function should be called every time data is copied to user space.
585 * It calculates the appropriate TCP receive buffer space.
586 */
587void tcp_rcv_space_adjust(struct sock *sk)
588{
589 struct tcp_sock *tp = tcp_sk(sk);
590 int time;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700591 int copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900592
Eric Dumazet645f4c62017-04-25 10:15:41 -0700593 time = skb_mstamp_us_delta(&tp->tcp_mstamp, &tp->rcvq_space.time);
594 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900596
Eric Dumazetb0983d32013-09-20 13:56:58 -0700597 /* Number of bytes copied to user in last RTT */
598 copied = tp->copied_seq - tp->rcvq_space.seq;
599 if (copied <= tp->rcvq_space.space)
600 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601
Eric Dumazetb0983d32013-09-20 13:56:58 -0700602 /* A bit of theory :
603 * copied = bytes received in previous RTT, our base window
604 * To cope with packet losses, we need a 2x factor
605 * To cope with slow start, and sender growing its cwin by 100 %
606 * every RTT, we need a 4x factor, because the ACK we are sending
607 * now is for the next RTT, not the current one :
608 * <prev RTT . ><current RTT .. ><next RTT .... >
609 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
Eric Dumazetb0983d32013-09-20 13:56:58 -0700611 if (sysctl_tcp_moderate_rcvbuf &&
612 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
613 int rcvwin, rcvmem, rcvbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Eric Dumazetb0983d32013-09-20 13:56:58 -0700615 /* minimal window to cope with packet losses, assuming
616 * steady state. Add some cushion because of small variations.
617 */
618 rcvwin = (copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Eric Dumazetb0983d32013-09-20 13:56:58 -0700620 /* If rate increased by 25%,
621 * assume slow start, rcvwin = 3 * copied
622 * If rate increased by 50%,
623 * assume sender can use 2x growth, rcvwin = 4 * copied
624 */
625 if (copied >=
626 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
627 if (copied >=
628 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
629 rcvwin <<= 1;
630 else
631 rcvwin += (rcvwin >> 1);
632 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633
Eric Dumazetb0983d32013-09-20 13:56:58 -0700634 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
635 while (tcp_win_from_space(rcvmem) < tp->advmss)
636 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
Eric Dumazetb0983d32013-09-20 13:56:58 -0700638 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
639 if (rcvbuf > sk->sk_rcvbuf) {
640 sk->sk_rcvbuf = rcvbuf;
641
642 /* Make the window clamp follow along. */
643 tp->window_clamp = rcvwin;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 }
645 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700646 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900647
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648new_measure:
649 tp->rcvq_space.seq = tp->copied_seq;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700650 tp->rcvq_space.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651}
652
653/* There is something which you must keep in mind when you analyze the
654 * behavior of the tp->ato delayed ack timeout interval. When a
655 * connection starts up, we want to ack as quickly as possible. The
656 * problem is that "good" TCP's do slow start at the beginning of data
657 * transmission. The means that until we send the first few ACK's the
658 * sender will sit on his end and only queue most of his data, because
659 * he can only send snd_cwnd unacked packets at any given time. For
660 * each ACK we send, he increments snd_cwnd and transmits more of his
661 * queue. -DaveM
662 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700663static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700665 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700666 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 u32 now;
668
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700669 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700671 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672
673 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 now = tcp_time_stamp;
676
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700677 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 /* The _first_ data packet received, initialize
679 * delayed ACK engine.
680 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700681 tcp_incr_quickack(sk);
682 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700684 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800686 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700688 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
689 } else if (m < icsk->icsk_ack.ato) {
690 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
691 if (icsk->icsk_ack.ato > icsk->icsk_rto)
692 icsk->icsk_ack.ato = icsk->icsk_rto;
693 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800694 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 * restart window, so that we send ACKs quickly.
696 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700697 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800698 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 }
700 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700701 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702
Florian Westphal735d3832014-09-29 13:08:30 +0200703 tcp_ecn_check_ce(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700706 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707}
708
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709/* Called to compute a smoothed rtt estimate. The data fed to this
710 * routine either comes from timestamps, or from segments that were
711 * known _not_ to have been retransmitted [see Karn/Partridge
712 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
713 * piece by Van Jacobson.
714 * NOTE: the next three routines used to be one big routine.
715 * To save cycles in the RFC 1323 implementation it was better to break
716 * it up into three procedures. -- erics
717 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800718static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300720 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800721 long m = mrtt_us; /* RTT */
722 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 /* The following amusing code comes from Jacobson's
725 * article in SIGCOMM '88. Note that rtt and mdev
726 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900727 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 * m stands for "measurement".
729 *
730 * On a 1990 paper the rto value is changed to:
731 * RTO = rtt + 4 * mdev
732 *
733 * Funny. This algorithm seems to be very broken.
734 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800735 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
737 * does not matter how to _calculate_ it. Seems, it was trap
738 * that VJ failed to avoid. 8)
739 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800740 if (srtt != 0) {
741 m -= (srtt >> 3); /* m is now error in rtt est */
742 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 if (m < 0) {
744 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800745 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 /* This is similar to one of Eifel findings.
747 * Eifel blocks mdev updates when rtt decreases.
748 * This solution is a bit different: we use finer gain
749 * for mdev in this case (alpha*beta).
750 * Like Eifel it also prevents growth of rto,
751 * but also it limits too fast rto decreases,
752 * happening in pure Eifel.
753 */
754 if (m > 0)
755 m >>= 3;
756 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800757 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800759 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
760 if (tp->mdev_us > tp->mdev_max_us) {
761 tp->mdev_max_us = tp->mdev_us;
762 if (tp->mdev_max_us > tp->rttvar_us)
763 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 }
765 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800766 if (tp->mdev_max_us < tp->rttvar_us)
767 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800769 tp->mdev_max_us = tcp_rto_min_us(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 }
771 } else {
772 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800773 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800774 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
775 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
776 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 tp->rtt_seq = tp->snd_nxt;
778 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800779 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780}
781
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700782/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
783 * Note: TCP stack does not yet implement pacing.
784 * FQ packet scheduler can be used to implement cheap but effective
785 * TCP pacing, to smooth the burst on large writes when packets
786 * in flight is significantly lower than cwnd (or rwin)
787 */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700788int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
789int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
790
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700791static void tcp_update_pacing_rate(struct sock *sk)
792{
793 const struct tcp_sock *tp = tcp_sk(sk);
794 u64 rate;
795
796 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700797 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
798
799 /* current rate is (cwnd * mss) / srtt
800 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
801 * In Congestion Avoidance phase, set it to 120 % the current rate.
802 *
803 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
804 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
805 * end of slow start and should slow down.
806 */
807 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
808 rate *= sysctl_tcp_pacing_ss_ratio;
809 else
810 rate *= sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700811
812 rate *= max(tp->snd_cwnd, tp->packets_out);
813
Eric Dumazet740b0f12014-02-26 14:02:48 -0800814 if (likely(tp->srtt_us))
815 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700816
Eric Dumazetba537422013-10-09 17:14:52 -0700817 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
818 * without any lock. We want to make sure compiler wont store
819 * intermediate values in this location.
820 */
821 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
822 sk->sk_max_pacing_rate);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700823}
824
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825/* Calculate rto without backoff. This is the second half of Van Jacobson's
826 * routine referred to above.
827 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800828static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700830 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 /* Old crap is replaced with new one. 8)
832 *
833 * More seriously:
834 * 1. If rtt variance happened to be less 50msec, it is hallucination.
835 * It cannot be less due to utterly erratic ACK generation made
836 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
837 * to do with delayed acks, because at cwnd>2 true delack timeout
838 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800839 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000841 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
843 /* 2. Fixups made earlier cannot be right.
844 * If we do not estimate RTO correctly without them,
845 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800846 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800849 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
850 * guarantees that rto is higher.
851 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000852 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853}
854
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400855__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856{
857 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
858
Gerrit Renker22b71c82010-08-29 19:23:12 +0000859 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800860 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
862}
863
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300864/*
865 * Packet counting of FACK is based on in-order assumptions, therefore TCP
866 * disables it when reordering is detected
867 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700868void tcp_disable_fack(struct tcp_sock *tp)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300869{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800870 /* RFC3517 uses different metric in lost marker => reset on change */
871 if (tcp_is_fack(tp))
872 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000873 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300874}
875
Ryousei Takano564262c2007-10-25 23:03:52 -0700876/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300877static void tcp_dsack_seen(struct tcp_sock *tp)
878{
Vijay Subramanianab562222011-12-20 13:23:24 +0000879 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300880}
881
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300882static void tcp_update_reordering(struct sock *sk, const int metric,
883 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300885 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng2d2517e2017-04-04 14:15:40 -0700886 int mib_idx;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700887
Yuchung Cheng2d2517e2017-04-04 14:15:40 -0700888 if (metric > tp->reordering) {
Eric Dumazetdca145f2014-10-27 21:45:24 -0700889 tp->reordering = min(sysctl_tcp_max_reordering, metric);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000892 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
893 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
894 tp->reordering,
895 tp->fackets_out,
896 tp->sacked_out,
897 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300899 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000901
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700902 tp->rack.reord = 1;
Yuchung Cheng2d2517e2017-04-04 14:15:40 -0700903
904 /* This exciting event is worth to be remembered. 8) */
905 if (ts)
906 mib_idx = LINUX_MIB_TCPTSREORDER;
907 else if (tcp_is_reno(tp))
908 mib_idx = LINUX_MIB_TCPRENOREORDER;
909 else if (tcp_is_fack(tp))
910 mib_idx = LINUX_MIB_TCPFACKREORDER;
911 else
912 mib_idx = LINUX_MIB_TCPSACKREORDER;
913
914 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915}
916
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700917/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700918static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
919{
Ian Morris51456b22015-04-03 09:17:26 +0100920 if (!tp->retransmit_skb_hint ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700921 before(TCP_SKB_CB(skb)->seq,
922 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700923 tp->retransmit_skb_hint = skb;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700924}
925
Neal Cardwell0682e692016-09-19 23:39:13 -0400926/* Sum the number of packets on the wire we have marked as lost.
927 * There are two cases we care about here:
928 * a) Packet hasn't been marked lost (nor retransmitted),
929 * and this is the first loss.
930 * b) Packet has been marked both lost and retransmitted,
931 * and this means we think it was lost again.
932 */
933static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
934{
935 __u8 sacked = TCP_SKB_CB(skb)->sacked;
936
937 if (!(sacked & TCPCB_LOST) ||
938 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
939 tp->lost += tcp_skb_pcount(skb);
940}
941
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700942static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
943{
944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tcp_verify_retransmit_hint(tp, skb);
946
947 tp->lost_out += tcp_skb_pcount(skb);
Neal Cardwell0682e692016-09-19 23:39:13 -0400948 tcp_sum_lost(tp, skb);
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700949 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
950 }
951}
952
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700953void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700954{
955 tcp_verify_retransmit_hint(tp, skb);
956
Neal Cardwell0682e692016-09-19 23:39:13 -0400957 tcp_sum_lost(tp, skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700958 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
959 tp->lost_out += tcp_skb_pcount(skb);
960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
961 }
962}
963
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964/* This procedure tags the retransmission queue when SACKs arrive.
965 *
966 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
967 * Packets in queue with these bits set are counted in variables
968 * sacked_out, retrans_out and lost_out, correspondingly.
969 *
970 * Valid combinations are:
971 * Tag InFlight Description
972 * 0 1 - orig segment is in flight.
973 * S 0 - nothing flies, orig reached receiver.
974 * L 0 - nothing flies, orig lost by net.
975 * R 2 - both orig and retransmit are in flight.
976 * L|R 1 - orig is lost, retransmit is in flight.
977 * S|R 1 - orig reached receiver, retrans is still in flight.
978 * (L|S|R is logically valid, it could occur when L|R is sacked,
979 * but it is equivalent to plain S and code short-curcuits it to S.
980 * L|S is logically invalid, it would mean -1 packet in flight 8))
981 *
982 * These 6 states form finite state machine, controlled by the following events:
983 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
984 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000985 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 * A. Scoreboard estimator decided the packet is lost.
987 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000988 * A''. Its FACK modification, head until snd.fack is lost.
989 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 * segment was retransmitted.
991 * 4. D-SACK added new rule: D-SACK changes any tag to S.
992 *
993 * It is pleasant to note, that state diagram turns out to be commutative,
994 * so that we are allowed not to be bothered by order of our actions,
995 * when multiple events arrive simultaneously. (see the function below).
996 *
997 * Reordering detection.
998 * --------------------
999 * Reordering metric is maximal distance, which a packet can be displaced
1000 * in packet stream. With SACKs we can estimate it:
1001 *
1002 * 1. SACK fills old hole and the corresponding segment was not
1003 * ever retransmitted -> reordering. Alas, we cannot use it
1004 * when segment was retransmitted.
1005 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1006 * for retransmitted and already SACKed segment -> reordering..
1007 * Both of these heuristics are not used in Loss state, when we cannot
1008 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001009 *
1010 * SACK block validation.
1011 * ----------------------
1012 *
1013 * SACK block range validation checks that the received SACK block fits to
1014 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1015 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001016 * it means that the receiver is rather inconsistent with itself reporting
1017 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1018 * perfectly valid, however, in light of RFC2018 which explicitly states
1019 * that "SACK block MUST reflect the newest segment. Even if the newest
1020 * segment is going to be discarded ...", not that it looks very clever
1021 * in case of head skb. Due to potentional receiver driven attacks, we
1022 * choose to avoid immediate execution of a walk in write queue due to
1023 * reneging and defer head skb's loss recovery to standard loss recovery
1024 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001025 *
1026 * Implements also blockage to start_seq wrap-around. Problem lies in the
1027 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1028 * there's no guarantee that it will be before snd_nxt (n). The problem
1029 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1030 * wrap (s_w):
1031 *
1032 * <- outs wnd -> <- wrapzone ->
1033 * u e n u_w e_w s n_w
1034 * | | | | | | |
1035 * |<------------+------+----- TCP seqno space --------------+---------->|
1036 * ...-- <2^31 ->| |<--------...
1037 * ...---- >2^31 ------>| |<--------...
1038 *
1039 * Current code wouldn't be vulnerable but it's better still to discard such
1040 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1041 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1042 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1043 * equal to the ideal case (infinite seqno space without wrap caused issues).
1044 *
1045 * With D-SACK the lower bound is extended to cover sequence space below
1046 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001047 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001048 * for the normal SACK blocks, explained above). But there all simplicity
1049 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1050 * fully below undo_marker they do not affect behavior in anyway and can
1051 * therefore be safely ignored. In rare cases (which are more or less
1052 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1053 * fragmentation and packet reordering past skb's retransmission. To consider
1054 * them correctly, the acceptable range must be extended even more though
1055 * the exact amount is rather hard to quantify. However, tp->max_window can
1056 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001058static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1059 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001060{
1061 /* Too far in future, or reversed (interpretation is ambiguous) */
1062 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001063 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001064
1065 /* Nasty start_seq wrap-around check (see comments above) */
1066 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001067 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001068
Ryousei Takano564262c2007-10-25 23:03:52 -07001069 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001070 * start_seq == snd_una is non-sensical (see comments above)
1071 */
1072 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001073 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001074
1075 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001076 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001077
1078 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001079 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001080 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001081
1082 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001083 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001084
1085 /* Too old */
1086 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001087 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001088
1089 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1090 * start_seq < undo_marker and end_seq >= undo_marker.
1091 */
1092 return !before(start_seq, end_seq - tp->max_window);
1093}
1094
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001095static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096 struct tcp_sack_block_wire *sp, int num_sacks,
1097 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001098{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001099 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001100 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001102 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001103
1104 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001105 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001106 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001107 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001108 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001109 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001111
1112 if (!after(end_seq_0, end_seq_1) &&
1113 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001114 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001115 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001116 NET_INC_STATS(sock_net(sk),
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001117 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001118 }
1119 }
1120
1121 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001122 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001123 !after(end_seq_0, prior_snd_una) &&
1124 after(end_seq_0, tp->undo_marker))
1125 tp->undo_retrans--;
1126
1127 return dup_sack;
1128}
1129
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001130struct tcp_sacktag_state {
Eric Dumazet740b0f12014-02-26 14:02:48 -08001131 int reord;
1132 int fack_count;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001133 /* Timestamps for earliest and latest never-retransmitted segment
1134 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1135 * but congestion control should still get an accurate delay signal.
1136 */
1137 struct skb_mstamp first_sackt;
1138 struct skb_mstamp last_sackt;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001139 struct rate_sample *rate;
Eric Dumazet740b0f12014-02-26 14:02:48 -08001140 int flag;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001141};
1142
Ilpo Järvinend1935942007-10-11 17:34:25 -07001143/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1144 * the incoming SACK may not exactly match but we can find smaller MSS
1145 * aligned portion of it that matches. Therefore we might need to fragment
1146 * which may fail and creates some hassle (caller must handle error case
1147 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001148 *
1149 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001150 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001151static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001152 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001153{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001154 int err;
1155 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001156 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001157 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001158
1159 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1160 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1161
1162 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1163 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001164 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001165 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1166
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001167 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001168 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001169 if (pkt_len < mss)
1170 pkt_len = mss;
1171 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001172 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001173 if (pkt_len < mss)
1174 return -EINVAL;
1175 }
1176
1177 /* Round if necessary so that SACKs cover only full MSSes
1178 * and/or the remaining small portion (if present)
1179 */
1180 if (pkt_len > mss) {
1181 unsigned int new_len = (pkt_len / mss) * mss;
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001182 if (!in_sack && new_len < pkt_len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001183 new_len += mss;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001184 pkt_len = new_len;
1185 }
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001186
1187 if (pkt_len >= skb->len && !in_sack)
1188 return 0;
1189
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001190 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001191 if (err < 0)
1192 return err;
1193 }
1194
1195 return in_sack;
1196}
1197
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001198/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1199static u8 tcp_sacktag_one(struct sock *sk,
1200 struct tcp_sacktag_state *state, u8 sacked,
1201 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001202 int dup_sack, int pcount,
1203 const struct skb_mstamp *xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001204{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001205 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001206 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001207
1208 /* Account D-SACK for retransmitted packet. */
1209 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001210 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001211 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001212 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001213 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001214 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001215 }
1216
1217 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001218 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001219 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001220
1221 if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazetd2329f12017-04-25 10:15:38 -07001222 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001223
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001224 if (sacked & TCPCB_SACKED_RETRANS) {
1225 /* If the segment is not tagged as lost,
1226 * we do not clear RETRANS, believing
1227 * that retransmission is still in flight.
1228 */
1229 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001230 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001231 tp->lost_out -= pcount;
1232 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001233 }
1234 } else {
1235 if (!(sacked & TCPCB_RETRANS)) {
1236 /* New sack for not retransmitted frame,
1237 * which was in hole. It is reordering.
1238 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001239 if (before(start_seq,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001240 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001241 state->reord = min(fack_count,
1242 state->reord);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001243 if (!after(end_seq, tp->high_seq))
1244 state->flag |= FLAG_ORIG_SACK_ACKED;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001245 if (state->first_sackt.v64 == 0)
1246 state->first_sackt = *xmit_time;
1247 state->last_sackt = *xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001248 }
1249
1250 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001251 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001252 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001253 }
1254 }
1255
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001256 sacked |= TCPCB_SACKED_ACKED;
1257 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001258 tp->sacked_out += pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08001259 tp->delivered += pcount; /* Out-of-order packets delivered */
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001260
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001261 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001262
1263 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Ian Morris00db4122015-04-03 09:17:27 +01001264 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001265 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001266 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001267
1268 if (fack_count > tp->fackets_out)
1269 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001270 }
1271
1272 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1273 * frames and clear it. undo_retrans is decreased above, L|R frames
1274 * are accounted above as well.
1275 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001276 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1277 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001278 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001279 }
1280
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001281 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001282}
1283
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001284/* Shift newly-SACKed bytes from this skb to the immediately previous
1285 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1286 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001287static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1288 struct tcp_sacktag_state *state,
1289 unsigned int pcount, int shifted, int mss,
1290 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001291{
1292 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001293 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001294 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1295 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001296
1297 BUG_ON(!pcount);
1298
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001299 /* Adjust counters and hints for the newly sacked sequence
1300 * range but discard the return value since prev is already
1301 * marked. We must tag the range first because the seq
1302 * advancement below implicitly advances
1303 * tcp_highest_sack_seq() when skb is highest_sack.
1304 */
1305 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001306 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001307 &skb->skb_mstamp);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001308 tcp_rate_skb_delivered(sk, skb, state->rate);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001309
1310 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001311 tp->lost_cnt_hint += pcount;
1312
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001313 TCP_SKB_CB(prev)->end_seq += shifted;
1314 TCP_SKB_CB(skb)->seq += shifted;
1315
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001316 tcp_skb_pcount_add(prev, pcount);
1317 BUG_ON(tcp_skb_pcount(skb) < pcount);
1318 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001319
1320 /* When we're adding to gso_segs == 1, gso_size will be zero,
1321 * in theory this shouldn't be necessary but as long as DSACK
1322 * code can come after this skb later on it's better to keep
1323 * setting gso_size to something.
1324 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001325 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1326 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001327
1328 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001329 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001330 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001331
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001332 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1333 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1334
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001335 if (skb->len > 0) {
1336 BUG_ON(!tcp_skb_pcount(skb));
Eric Dumazetc10d9312016-04-29 14:16:47 -07001337 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001338 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001339 }
1340
1341 /* Whole SKB was eaten :-) */
1342
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001343 if (skb == tp->retransmit_skb_hint)
1344 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001345 if (skb == tp->lost_skb_hint) {
1346 tp->lost_skb_hint = prev;
1347 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1348 }
1349
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001350 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001351 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001352 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1353 TCP_SKB_CB(prev)->end_seq++;
1354
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001355 if (skb == tcp_highest_sack(sk))
1356 tcp_advance_highest_sack(sk, skb);
1357
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07001358 tcp_skb_collapse_tstamp(prev, skb);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001359 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1360 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1361
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001362 tcp_unlink_write_queue(skb, sk);
1363 sk_wmem_free_skb(sk, skb);
1364
Eric Dumazetc10d9312016-04-29 14:16:47 -07001365 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001366
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001367 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001368}
1369
1370/* I wish gso_size would have a bit more sane initialization than
1371 * something-or-zero which complicates things
1372 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001373static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001374{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001375 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001376}
1377
1378/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001379static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001380{
1381 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1382}
1383
1384/* Try collapsing SACK blocks spanning across multiple skbs to a single
1385 * skb.
1386 */
1387static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001388 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001389 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001390 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001391{
1392 struct tcp_sock *tp = tcp_sk(sk);
1393 struct sk_buff *prev;
1394 int mss;
1395 int pcount = 0;
1396 int len;
1397 int in_sack;
1398
1399 if (!sk_can_gso(sk))
1400 goto fallback;
1401
1402 /* Normally R but no L won't result in plain S */
1403 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001404 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001405 goto fallback;
1406 if (!skb_can_shift(skb))
1407 goto fallback;
1408 /* This frame is about to be dropped (was ACKed). */
1409 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1410 goto fallback;
1411
1412 /* Can only happen with delayed DSACK + discard craziness */
1413 if (unlikely(skb == tcp_write_queue_head(sk)))
1414 goto fallback;
1415 prev = tcp_write_queue_prev(sk, skb);
1416
1417 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1418 goto fallback;
1419
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001420 if (!tcp_skb_can_collapse_to(prev))
1421 goto fallback;
1422
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001423 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1424 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1425
1426 if (in_sack) {
1427 len = skb->len;
1428 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001429 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001430
1431 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1432 * drop this restriction as unnecessary
1433 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001434 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001435 goto fallback;
1436 } else {
1437 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1438 goto noop;
1439 /* CHECKME: This is non-MSS split case only?, this will
1440 * cause skipped skbs due to advancing loop btw, original
1441 * has that feature too
1442 */
1443 if (tcp_skb_pcount(skb) <= 1)
1444 goto noop;
1445
1446 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1447 if (!in_sack) {
1448 /* TODO: head merge to next could be attempted here
1449 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1450 * though it might not be worth of the additional hassle
1451 *
1452 * ...we can probably just fallback to what was done
1453 * previously. We could try merging non-SACKed ones
1454 * as well but it probably isn't going to buy off
1455 * because later SACKs might again split them, and
1456 * it would make skb timestamp tracking considerably
1457 * harder problem.
1458 */
1459 goto fallback;
1460 }
1461
1462 len = end_seq - TCP_SKB_CB(skb)->seq;
1463 BUG_ON(len < 0);
1464 BUG_ON(len > skb->len);
1465
1466 /* MSS boundaries should be honoured or else pcount will
1467 * severely break even though it makes things bit trickier.
1468 * Optimize common case to avoid most of the divides
1469 */
1470 mss = tcp_skb_mss(skb);
1471
1472 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1473 * drop this restriction as unnecessary
1474 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001475 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001476 goto fallback;
1477
1478 if (len == mss) {
1479 pcount = 1;
1480 } else if (len < mss) {
1481 goto noop;
1482 } else {
1483 pcount = len / mss;
1484 len = pcount * mss;
1485 }
1486 }
1487
Neal Cardwell4648dc92012-03-05 19:35:04 +00001488 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1489 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1490 goto fallback;
1491
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001492 if (!skb_shift(prev, skb, len))
1493 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001494 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001495 goto out;
1496
1497 /* Hole filled allows collapsing with the next as well, this is very
1498 * useful when hole on every nth skb pattern happens
1499 */
1500 if (prev == tcp_write_queue_tail(sk))
1501 goto out;
1502 skb = tcp_write_queue_next(sk, prev);
1503
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001504 if (!skb_can_shift(skb) ||
1505 (skb == tcp_send_head(sk)) ||
1506 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001507 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001508 goto out;
1509
1510 len = skb->len;
1511 if (skb_shift(prev, skb, len)) {
1512 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001513 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001514 }
1515
1516out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001517 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001518 return prev;
1519
1520noop:
1521 return skb;
1522
1523fallback:
Eric Dumazetc10d9312016-04-29 14:16:47 -07001524 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001525 return NULL;
1526}
1527
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001528static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1529 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001530 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001531 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001532 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001533{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001534 struct tcp_sock *tp = tcp_sk(sk);
1535 struct sk_buff *tmp;
1536
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001537 tcp_for_write_queue_from(skb, sk) {
1538 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001539 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001540
1541 if (skb == tcp_send_head(sk))
1542 break;
1543
1544 /* queue is in-order => we can short-circuit the walk early */
1545 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1546 break;
1547
Ian Morris00db4122015-04-03 09:17:27 +01001548 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001549 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1550 in_sack = tcp_match_skb_to_sack(sk, skb,
1551 next_dup->start_seq,
1552 next_dup->end_seq);
1553 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001554 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001555 }
1556
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001557 /* skb reference here is a bit tricky to get right, since
1558 * shifting can eat and free both this skb and the next,
1559 * so not even _safe variant of the loop is enough.
1560 */
1561 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001562 tmp = tcp_shift_skb_data(sk, skb, state,
1563 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001564 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001565 if (tmp != skb) {
1566 skb = tmp;
1567 continue;
1568 }
1569
1570 in_sack = 0;
1571 } else {
1572 in_sack = tcp_match_skb_to_sack(sk, skb,
1573 start_seq,
1574 end_seq);
1575 }
1576 }
1577
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001578 if (unlikely(in_sack < 0))
1579 break;
1580
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001581 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001582 TCP_SKB_CB(skb)->sacked =
1583 tcp_sacktag_one(sk,
1584 state,
1585 TCP_SKB_CB(skb)->sacked,
1586 TCP_SKB_CB(skb)->seq,
1587 TCP_SKB_CB(skb)->end_seq,
1588 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001589 tcp_skb_pcount(skb),
Eric Dumazet740b0f12014-02-26 14:02:48 -08001590 &skb->skb_mstamp);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001591 tcp_rate_skb_delivered(sk, skb, state->rate);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001592
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001593 if (!before(TCP_SKB_CB(skb)->seq,
1594 tcp_highest_sack_seq(tp)))
1595 tcp_advance_highest_sack(sk, skb);
1596 }
1597
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001598 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001599 }
1600 return skb;
1601}
1602
1603/* Avoid all extra work that is being done by sacktag while walking in
1604 * a normal way
1605 */
1606static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001607 struct tcp_sacktag_state *state,
1608 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001609{
1610 tcp_for_write_queue_from(skb, sk) {
1611 if (skb == tcp_send_head(sk))
1612 break;
1613
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001614 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001615 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001616
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001617 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001618 }
1619 return skb;
1620}
1621
1622static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1623 struct sock *sk,
1624 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001625 struct tcp_sacktag_state *state,
1626 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001627{
Ian Morris51456b22015-04-03 09:17:26 +01001628 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001629 return skb;
1630
1631 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001632 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1633 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1634 next_dup->start_seq, next_dup->end_seq,
1635 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001636 }
1637
1638 return skb;
1639}
1640
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001641static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001642{
1643 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1644}
1645
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001647tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001648 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649{
1650 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001651 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1652 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001653 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001654 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001655 struct tcp_sack_block *cache;
1656 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001657 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001658 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001659 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001660 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001661 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001663 state->flag = 0;
1664 state->reord = tp->packets_out;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001665
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001666 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001667 if (WARN_ON(tp->fackets_out))
1668 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001669 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001670 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001672 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001673 num_sacks, prior_snd_una);
Yuchung Chengb9f64822016-09-19 23:39:14 -04001674 if (found_dup_sack) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001675 state->flag |= FLAG_DSACKING_ACK;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001676 tp->delivered++; /* A spurious retransmission is delivered */
1677 }
Baruch Even6f746512007-02-04 23:36:42 -08001678
1679 /* Eliminate too old ACKs, but take into
1680 * account more or less fresh ones, they can
1681 * contain valid SACK info.
1682 */
1683 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1684 return 0;
1685
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001686 if (!tp->packets_out)
1687 goto out;
1688
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001689 used_sacks = 0;
1690 first_sack_index = 0;
1691 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001692 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001693
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001694 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1695 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001696
1697 if (!tcp_is_sackblock_valid(tp, dup_sack,
1698 sp[used_sacks].start_seq,
1699 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001700 int mib_idx;
1701
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001702 if (dup_sack) {
1703 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001704 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001705 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001706 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001707 } else {
1708 /* Don't count olds caused by ACK reordering */
1709 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1710 !after(sp[used_sacks].end_seq, tp->snd_una))
1711 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001712 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001713 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001714
Eric Dumazetc10d9312016-04-29 14:16:47 -07001715 NET_INC_STATS(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001716 if (i == 0)
1717 first_sack_index = -1;
1718 continue;
1719 }
1720
1721 /* Ignore very old stuff early */
1722 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1723 continue;
1724
1725 used_sacks++;
1726 }
1727
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001728 /* order SACK blocks to allow in order walk of the retrans queue */
1729 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001730 for (j = 0; j < i; j++) {
1731 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001732 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001733
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001734 /* Track where the first SACK block goes to */
1735 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001736 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001737 }
1738 }
1739 }
1740
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001741 skb = tcp_write_queue_head(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001742 state->fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001743 i = 0;
1744
1745 if (!tp->sacked_out) {
1746 /* It's already past, so skip checking against it */
1747 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1748 } else {
1749 cache = tp->recv_sack_cache;
1750 /* Skip empty blocks in at head of the cache */
1751 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1752 !cache->end_seq)
1753 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001754 }
1755
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001756 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001757 u32 start_seq = sp[i].start_seq;
1758 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001759 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001760 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001761
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001762 if (found_dup_sack && ((i + 1) == first_sack_index))
1763 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001765 /* Skip too early cached blocks */
1766 while (tcp_sack_cache_ok(tp, cache) &&
1767 !before(start_seq, cache->end_seq))
1768 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001770 /* Can skip some work by looking recv_sack_cache? */
1771 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1772 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001773
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001774 /* Head todo? */
1775 if (before(start_seq, cache->start_seq)) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001776 skb = tcp_sacktag_skip(skb, sk, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001777 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001778 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001779 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001780 start_seq,
1781 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001782 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001783 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001784
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001785 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001786 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001787 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001788
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001789 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001790 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001791 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001792
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001793 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001794 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001795 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001796 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001797 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001798 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001799 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001800 cache++;
1801 goto walk;
1802 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001804 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001805 /* Check overlap against next cached too (past this one already) */
1806 cache++;
1807 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001809
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001810 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1811 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001812 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001813 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001814 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001815 }
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001816 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001817
1818walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001819 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001820 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001821
1822advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001823 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 }
1825
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001826 /* Clear the head of the cache sack blocks so we can skip it next time */
1827 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1828 tp->recv_sack_cache[i].start_seq = 0;
1829 tp->recv_sack_cache[i].end_seq = 0;
1830 }
1831 for (j = 0; j < used_sacks; j++)
1832 tp->recv_sack_cache[i++] = sp[j];
1833
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001834 if ((state->reord < tp->fackets_out) &&
Yuchung Cheng9b441902013-03-20 13:32:58 +00001835 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001836 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001838 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001839out:
1840
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001842 WARN_ON((int)tp->sacked_out < 0);
1843 WARN_ON((int)tp->lost_out < 0);
1844 WARN_ON((int)tp->retrans_out < 0);
1845 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001847 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848}
1849
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001850/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001851 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001852 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001853static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001854{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001855 u32 holes;
1856
1857 holes = max(tp->lost_out, 1U);
1858 holes = min(holes, tp->packets_out);
1859
1860 if ((tp->sacked_out + holes) > tp->packets_out) {
1861 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001862 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001863 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001864 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001865}
1866
1867/* If we receive more dupacks than we expected counting segments
1868 * in assumption of absent reordering, interpret this as reordering.
1869 * The only another reason could be bug in receiver TCP.
1870 */
1871static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1872{
1873 struct tcp_sock *tp = tcp_sk(sk);
1874 if (tcp_limit_reno_sacked(tp))
1875 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001876}
1877
1878/* Emulate SACKs for SACKless connection: account for a new dupack. */
1879
1880static void tcp_add_reno_sack(struct sock *sk)
1881{
1882 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001883 u32 prior_sacked = tp->sacked_out;
1884
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001885 tp->sacked_out++;
1886 tcp_check_reno_reordering(sk, 0);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001887 if (tp->sacked_out > prior_sacked)
1888 tp->delivered++; /* Some out-of-order packet is delivered */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001889 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001890}
1891
1892/* Account for ACK, ACKing some data in Reno Recovery phase. */
1893
1894static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1895{
1896 struct tcp_sock *tp = tcp_sk(sk);
1897
1898 if (acked > 0) {
1899 /* One ACK acked hole. The rest eat duplicate ACKs. */
Yuchung Chengddf1af62016-02-02 10:33:06 -08001900 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001901 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001902 tp->sacked_out = 0;
1903 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001904 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001905 }
1906 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001907 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001908}
1909
1910static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1911{
1912 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001913}
1914
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001915void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001920 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001921 tp->fackets_out = 0;
1922 tp->sacked_out = 0;
1923}
1924
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001925static inline void tcp_init_undo(struct tcp_sock *tp)
1926{
1927 tp->undo_marker = tp->snd_una;
1928 /* Retransmission still in flight may cause DSACKs later. */
1929 tp->undo_retrans = tp->retrans_out ? : -1;
1930}
1931
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001932/* Enter Loss state. If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 * and reset tags completely, otherwise preserve SACKs. If receiver
1934 * dropped its ofo queue, we will know this due to reneging detection.
1935 */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001936void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001938 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov1043e252016-02-03 09:46:52 +02001940 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 struct sk_buff *skb;
Yuchung Chengcc663f42017-04-07 11:42:05 -07001942 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001943 bool is_reneg; /* is receiver reneging on SACKs? */
Neal Cardwell0682e692016-09-19 23:39:13 -04001944 bool mark_lost;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945
1946 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001947 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1948 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001949 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1950 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1951 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1952 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001953 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
1955 tp->snd_cwnd = 1;
1956 tp->snd_cwnd_cnt = 0;
1957 tp->snd_cwnd_stamp = tcp_time_stamp;
1958
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001959 tp->retrans_out = 0;
1960 tp->lost_out = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001961
1962 if (tcp_is_reno(tp))
1963 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001965 skb = tcp_write_queue_head(sk);
1966 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1967 if (is_reneg) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07001968 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001969 tp->sacked_out = 0;
1970 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001971 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07001972 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
David S. Millerfe067e82007-03-07 12:12:44 -08001974 tcp_for_write_queue(skb, sk) {
1975 if (skb == tcp_send_head(sk))
1976 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001977
Neal Cardwell0682e692016-09-19 23:39:13 -04001978 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1979 is_reneg);
1980 if (mark_lost)
1981 tcp_sum_lost(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
Neal Cardwell0682e692016-09-19 23:39:13 -04001983 if (mark_lost) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1985 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1986 tp->lost_out += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 }
1988 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001989 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001991 /* Timeout in disordered state after receiving substantial DUPACKs
1992 * suggests that the degree of reordering is over-estimated.
1993 */
1994 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02001995 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001996 tp->reordering = min_t(unsigned int, tp->reordering,
Nikolay Borisov1043e252016-02-03 09:46:52 +02001997 net->ipv4.sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001998 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02002000 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002001
Yuchung Chengcc663f42017-04-07 11:42:05 -07002002 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2003 * loss recovery is underway except recurring timeout(s) on
2004 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2005 *
2006 * In theory F-RTO can be used repeatedly during loss recovery.
2007 * In practice this interacts badly with broken middle-boxes that
2008 * falsely raise the receive window, which results in repeated
2009 * timeouts and stop-and-go behavior.
Yuchung Chenge33099f2013-03-20 13:33:00 +00002010 */
Yuchung Chengcc663f42017-04-07 11:42:05 -07002011 tp->frto = sysctl_tcp_frto &&
2012 (new_recovery || icsk->icsk_retransmits) &&
2013 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014}
2015
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002016/* If ACK arrived pointing to a remembered SACK, it means that our
2017 * remembered SACKs do not reflect real state of receiver i.e.
2018 * receiver _host_ is heavily congested (or buggy).
2019 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002020 * To avoid big spurious retransmission bursts due to transient SACK
2021 * scoreboard oddities that look like reneging, we give the receiver a
2022 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2023 * restore sanity to the SACK scoreboard. If the apparent reneging
2024 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002025 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002026static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002028 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002029 struct tcp_sock *tp = tcp_sk(sk);
2030 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2031 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002033 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002034 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002035 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002037 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038}
2039
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002040static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002042 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043}
2044
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002045/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2046 * counter when SACK is enabled (without SACK, sacked_out is used for
2047 * that purpose).
2048 *
2049 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2050 * segments up to the highest received SACK block so far and holes in
2051 * between them.
2052 *
2053 * With reordering, holes may still be in flight, so RFC3517 recovery
2054 * uses pure sacked_out (total number of SACKed segments) even though
2055 * it violates the RFC that uses duplicate ACKs, often these are equal
2056 * but when e.g. out-of-window ACKs or packet duplication occurs,
2057 * they differ. Since neither occurs due to loss, TCP should really
2058 * ignore them.
2059 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002060static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002061{
2062 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2063}
2064
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065/* Linux NewReno/SACK/FACK/ECN state machine.
2066 * --------------------------------------
2067 *
2068 * "Open" Normal state, no dubious events, fast path.
2069 * "Disorder" In all the respects it is "Open",
2070 * but requires a bit more attention. It is entered when
2071 * we see some SACKs or dupacks. It is split of "Open"
2072 * mainly to move some processing from fast path to slow one.
2073 * "CWR" CWND was reduced due to some Congestion Notification event.
2074 * It can be ECN, ICMP source quench, local device congestion.
2075 * "Recovery" CWND was reduced, we are fast-retransmitting.
2076 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2077 *
2078 * tcp_fastretrans_alert() is entered:
2079 * - each incoming ACK, if state is not "Open"
2080 * - when arrived ACK is unusual, namely:
2081 * * SACK
2082 * * Duplicate ACK.
2083 * * ECN ECE.
2084 *
2085 * Counting packets in flight is pretty simple.
2086 *
2087 * in_flight = packets_out - left_out + retrans_out
2088 *
2089 * packets_out is SND.NXT-SND.UNA counted in packets.
2090 *
2091 * retrans_out is number of retransmitted segments.
2092 *
2093 * left_out is number of segments left network, but not ACKed yet.
2094 *
2095 * left_out = sacked_out + lost_out
2096 *
2097 * sacked_out: Packets, which arrived to receiver out of order
2098 * and hence not ACKed. With SACKs this number is simply
2099 * amount of SACKed data. Even without SACKs
2100 * it is easy to give pretty reliable estimate of this number,
2101 * counting duplicate ACKs.
2102 *
2103 * lost_out: Packets lost by network. TCP has no explicit
2104 * "loss notification" feedback from network (for now).
2105 * It means that this number can be only _guessed_.
2106 * Actually, it is the heuristics to predict lossage that
2107 * distinguishes different algorithms.
2108 *
2109 * F.e. after RTO, when all the queue is considered as lost,
2110 * lost_out = packets_out and in_flight = retrans_out.
2111 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002112 * Essentially, we have now a few algorithms detecting
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 * lost packets.
2114 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002115 * If the receiver supports SACK:
2116 *
2117 * RFC6675/3517: It is the conventional algorithm. A packet is
2118 * considered lost if the number of higher sequence packets
2119 * SACKed is greater than or equal the DUPACK thoreshold
2120 * (reordering). This is implemented in tcp_mark_head_lost and
2121 * tcp_update_scoreboard.
2122 *
2123 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2124 * (2017-) that checks timing instead of counting DUPACKs.
2125 * Essentially a packet is considered lost if it's not S/ACKed
2126 * after RTT + reordering_window, where both metrics are
2127 * dynamically measured and adjusted. This is implemented in
2128 * tcp_rack_mark_lost.
2129 *
Yuchung Cheng94bdc972017-01-12 22:11:42 -08002130 * FACK (Disabled by default. Subsumbed by RACK):
2131 * It is the simplest heuristics. As soon as we decided
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 * that something is lost, we decide that _all_ not SACKed
2133 * packets until the most forward SACK are lost. I.e.
2134 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2135 * It is absolutely correct estimate, if network does not reorder
2136 * packets. And it loses any connection to reality when reordering
2137 * takes place. We use FACK by default until reordering
2138 * is suspected on the path to this destination.
2139 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002140 * If the receiver does not support SACK:
2141 *
2142 * NewReno (RFC6582): in Recovery we assume that one segment
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 * is lost (classic Reno). While we are in Recovery and
2144 * a partial ACK arrives, we assume that one more packet
2145 * is lost (NewReno). This heuristics are the same in NewReno
2146 * and SACK.
2147 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 * Really tricky (and requiring careful tuning) part of algorithm
2149 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2150 * The first determines the moment _when_ we should reduce CWND and,
2151 * hence, slow down forward transmission. In fact, it determines the moment
2152 * when we decide that hole is caused by loss, rather than by a reorder.
2153 *
2154 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2155 * holes, caused by lost packets.
2156 *
2157 * And the most logically complicated part of algorithm is undo
2158 * heuristics. We detect false retransmits due to both too early
2159 * fast retransmit (reordering) and underestimated RTO, analyzing
2160 * timestamps and D-SACKs. When we detect that some segments were
2161 * retransmitted by mistake and CWND reduction was wrong, we undo
2162 * window reduction and abort recovery phase. This logic is hidden
2163 * inside several functions named tcp_try_undo_<something>.
2164 */
2165
2166/* This function decides, when we should leave Disordered state
2167 * and enter Recovery phase, reducing congestion window.
2168 *
2169 * Main question: may we further continue forward transmission
2170 * with the same cwnd?
2171 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002172static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002174 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175
2176 /* Trick#1: The loss is proven. */
2177 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002178 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179
2180 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002181 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002182 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002184 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185}
2186
Yuchung Cheng974c1232012-01-19 14:42:21 +00002187/* Detect loss in event "A" above by marking head of queue up as lost.
2188 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2189 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2190 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2191 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002192 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002193static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002195 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 struct sk_buff *skb;
Neal Cardwelld88270e2016-01-25 14:01:53 -08002197 int cnt, oldcnt, lost;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002198 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002199 /* Use SACK to deduce losses of new sequences sent during recovery */
2200 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002202 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002203 if (tp->lost_skb_hint) {
2204 skb = tp->lost_skb_hint;
2205 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002206 /* Head already handled? */
2207 if (mark_head && skb != tcp_write_queue_head(sk))
2208 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002209 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002210 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002211 cnt = 0;
2212 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213
David S. Millerfe067e82007-03-07 12:12:44 -08002214 tcp_for_write_queue_from(skb, sk) {
2215 if (skb == tcp_send_head(sk))
2216 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002217 /* TODO: do this better */
2218 /* this is not the most efficient way to do this... */
2219 tp->lost_skb_hint = skb;
2220 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002221
Yuchung Cheng974c1232012-01-19 14:42:21 +00002222 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002223 break;
2224
2225 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002226 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002227 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2228 cnt += tcp_skb_pcount(skb);
2229
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002230 if (cnt > packets) {
Yuchung Chengb3de75592010-09-24 13:22:06 +00002231 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002232 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de75592010-09-24 13:22:06 +00002233 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002234 break;
2235
Eric Dumazetf69ad292015-06-11 09:15:18 -07002236 mss = tcp_skb_mss(skb);
Neal Cardwelld88270e2016-01-25 14:01:53 -08002237 /* If needed, chop off the prefix to mark as lost. */
2238 lost = (packets - oldcnt) * mss;
2239 if (lost < skb->len &&
2240 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002241 break;
2242 cnt = packets;
2243 }
2244
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002245 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002246
2247 if (mark_head)
2248 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002250 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251}
2252
2253/* Account newly detected lost packet(s) */
2254
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002255static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002257 struct tcp_sock *tp = tcp_sk(sk);
2258
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002259 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002260 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002261 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 int lost = tp->fackets_out - tp->reordering;
2263 if (lost <= 0)
2264 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002265 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002267 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002268 if (sacked_upto >= 0)
2269 tcp_mark_head_lost(sk, sacked_upto, 0);
2270 else if (fast_rexmit)
2271 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273}
2274
Yuchung Cheng77c63122015-10-16 21:57:44 -07002275static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2276{
2277 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2278 before(tp->rx_opt.rcv_tsecr, when);
2279}
2280
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002281/* skb is spurious retransmitted if the returned timestamp echo
2282 * reply is prior to the skb transmission time
2283 */
2284static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2285 const struct sk_buff *skb)
2286{
2287 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2288 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2289}
2290
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291/* Nothing was retransmitted or returned timestamp is less
2292 * than timestamp of the first retransmission.
2293 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002294static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295{
2296 return !tp->retrans_stamp ||
Yuchung Cheng77c63122015-10-16 21:57:44 -07002297 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298}
2299
2300/* Undo procedures. */
2301
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002302/* We can clear retrans_stamp when there are no retransmissions in the
2303 * window. It would seem that it is trivially available for us in
2304 * tp->retrans_out, however, that kind of assumptions doesn't consider
2305 * what will happen if errors occur when sending retransmission for the
2306 * second time. ...It could the that such segment has only
2307 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2308 * the head skb is enough except for some reneging corner cases that
2309 * are not worth the effort.
2310 *
2311 * Main reason for all this complexity is the fact that connection dying
2312 * time now depends on the validity of the retrans_stamp, in particular,
2313 * that successive retransmissions of a segment must not advance
2314 * retrans_stamp under any conditions.
2315 */
2316static bool tcp_any_retrans_done(const struct sock *sk)
2317{
2318 const struct tcp_sock *tp = tcp_sk(sk);
2319 struct sk_buff *skb;
2320
2321 if (tp->retrans_out)
2322 return true;
2323
2324 skb = tcp_write_queue_head(sk);
2325 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2326 return true;
2327
2328 return false;
2329}
2330
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002332static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002334 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002336
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002337 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002338 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2339 msg,
2340 &inet->inet_daddr, ntohs(inet->inet_dport),
2341 tp->snd_cwnd, tcp_left_out(tp),
2342 tp->snd_ssthresh, tp->prior_ssthresh,
2343 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002344 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002345#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002346 else if (sk->sk_family == AF_INET6) {
Joe Perches91df42b2012-05-15 14:11:54 +00002347 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2348 msg,
Eric Dumazet019b1c92016-09-22 17:54:00 -07002349 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
Joe Perches91df42b2012-05-15 14:11:54 +00002350 tp->snd_cwnd, tcp_left_out(tp),
2351 tp->snd_ssthresh, tp->prior_ssthresh,
2352 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002353 }
2354#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355}
2356#else
2357#define DBGUNDO(x...) do { } while (0)
2358#endif
2359
Yuchung Cheng7026b912013-05-29 14:20:13 +00002360static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002362 struct tcp_sock *tp = tcp_sk(sk);
2363
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002364 if (unmark_loss) {
2365 struct sk_buff *skb;
2366
2367 tcp_for_write_queue(skb, sk) {
2368 if (skb == tcp_send_head(sk))
2369 break;
2370 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2371 }
2372 tp->lost_out = 0;
2373 tcp_clear_all_retrans_hints(tp);
2374 }
2375
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002377 const struct inet_connection_sock *icsk = inet_csk(sk);
2378
Florian Westphale9799182016-11-21 14:18:38 +01002379 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380
Yuchung Cheng7026b912013-05-29 14:20:13 +00002381 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002383 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002387 tp->undo_marker = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388}
2389
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002390static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002392 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393}
2394
2395/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002396static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002398 struct tcp_sock *tp = tcp_sk(sk);
2399
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002401 int mib_idx;
2402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 /* Happy end! We did not retransmit anything
2404 * or our original transmission succeeded.
2405 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002406 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002407 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002408 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002409 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002411 mib_idx = LINUX_MIB_TCPFULLUNDO;
2412
Eric Dumazetc10d9312016-04-29 14:16:47 -07002413 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002415 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 /* Hold old state until something *above* high_seq
2417 * is ACKed. For Reno it is MUST to prevent false
2418 * fast retransmits (RFC2582). SACK TCP is safe. */
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002419 if (!tcp_any_retrans_done(sk))
2420 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002421 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002423 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002424 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425}
2426
2427/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002428static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002430 struct tcp_sock *tp = tcp_sk(sk);
2431
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002433 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002434 tcp_undo_cwnd_reduction(sk, false);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002435 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002436 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002438 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439}
2440
Yuchung Chenge33099f2013-03-20 13:33:00 +00002441/* Undo during loss recovery after partial ACK or using F-RTO. */
2442static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002444 struct tcp_sock *tp = tcp_sk(sk);
2445
Yuchung Chenge33099f2013-03-20 13:33:00 +00002446 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002447 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002448
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002449 DBGUNDO(sk, "partial loss");
Eric Dumazetc10d9312016-04-29 14:16:47 -07002450 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002451 if (frto_undo)
Eric Dumazetc10d9312016-04-29 14:16:47 -07002452 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002453 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002454 inet_csk(sk)->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002455 if (frto_undo || tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002456 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002457 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002459 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460}
2461
Yuchung Cheng37598242015-07-01 14:11:15 -07002462/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002463 * It computes the number of packets to send (sndcnt) based on packets newly
2464 * delivered:
2465 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2466 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002467 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2468 * But when the retransmits are acked without further losses, PRR
2469 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002470 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002471static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002472{
2473 struct tcp_sock *tp = tcp_sk(sk);
2474
2475 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002476 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002477 tp->snd_cwnd_cnt = 0;
2478 tp->prior_cwnd = tp->snd_cwnd;
2479 tp->prr_delivered = 0;
2480 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002481 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002482 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002483}
2484
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002485void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002486{
2487 struct tcp_sock *tp = tcp_sk(sk);
2488 int sndcnt = 0;
2489 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2490
Yuchung Cheng8b8a3212016-01-06 12:42:38 -08002491 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2492 return;
2493
Yuchung Cheng684bad12012-09-02 17:38:04 +00002494 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002495 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002496 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2497 tp->prior_cwnd - 1;
2498 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yuchung Cheng37598242015-07-01 14:11:15 -07002499 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2500 !(flag & FLAG_LOST_RETRANS)) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002501 sndcnt = min_t(int, delta,
2502 max_t(int, tp->prr_delivered - tp->prr_out,
2503 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002504 } else {
2505 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002506 }
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002507 /* Force a fast retransmit upon entering fast recovery */
2508 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002509 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2510}
2511
Yuchung Cheng684bad12012-09-02 17:38:04 +00002512static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002514 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002515
Yuchung Chengc0402762016-09-19 23:39:21 -04002516 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2517 return;
2518
Yuchung Cheng684bad12012-09-02 17:38:04 +00002519 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2520 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2521 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2522 tp->snd_cwnd = tp->snd_ssthresh;
2523 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002524 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002525 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526}
2527
Yuchung Cheng684bad12012-09-02 17:38:04 +00002528/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002529void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002530{
2531 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002532
2533 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002534 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002535 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002536 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002537 tcp_set_ca_state(sk, TCP_CA_CWR);
2538 }
2539}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002540EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002541
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002542static void tcp_try_keep_open(struct sock *sk)
2543{
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 int state = TCP_CA_Open;
2546
Neal Cardwellf6982042011-11-16 08:58:04 +00002547 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002548 state = TCP_CA_Disorder;
2549
2550 if (inet_csk(sk)->icsk_ca_state != state) {
2551 tcp_set_ca_state(sk, state);
2552 tp->high_seq = tp->snd_nxt;
2553 }
2554}
2555
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002556static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002558 struct tcp_sock *tp = tcp_sk(sk);
2559
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002560 tcp_verify_left_out(tp);
2561
Yuchung Cheng9b441902013-03-20 13:32:58 +00002562 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 tp->retrans_stamp = 0;
2564
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002565 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002566 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002568 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002569 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 }
2571}
2572
John Heffner5d424d52006-03-20 17:53:41 -08002573static void tcp_mtup_probe_failed(struct sock *sk)
2574{
2575 struct inet_connection_sock *icsk = inet_csk(sk);
2576
2577 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2578 icsk->icsk_mtup.probe_size = 0;
Eric Dumazetc10d9312016-04-29 14:16:47 -07002579 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002580}
2581
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002582static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002583{
2584 struct tcp_sock *tp = tcp_sk(sk);
2585 struct inet_connection_sock *icsk = inet_csk(sk);
2586
2587 /* FIXME: breaks with very large cwnd */
2588 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2589 tp->snd_cwnd = tp->snd_cwnd *
2590 tcp_mss_to_mtu(sk, tp->mss_cache) /
2591 icsk->icsk_mtup.probe_size;
2592 tp->snd_cwnd_cnt = 0;
2593 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002594 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002595
2596 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2597 icsk->icsk_mtup.probe_size = 0;
2598 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002600}
2601
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002602/* Do a simple retransmit without using the backoff mechanisms in
2603 * tcp_timer. This is used for path mtu discovery.
2604 * The socket is already locked here.
2605 */
2606void tcp_simple_retransmit(struct sock *sk)
2607{
2608 const struct inet_connection_sock *icsk = inet_csk(sk);
2609 struct tcp_sock *tp = tcp_sk(sk);
2610 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002611 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002612 u32 prior_lost = tp->lost_out;
2613
2614 tcp_for_write_queue(skb, sk) {
2615 if (skb == tcp_send_head(sk))
2616 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002617 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002618 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2619 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2620 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2621 tp->retrans_out -= tcp_skb_pcount(skb);
2622 }
2623 tcp_skb_mark_lost_uncond_verify(tp, skb);
2624 }
2625 }
2626
2627 tcp_clear_retrans_hints_partial(tp);
2628
2629 if (prior_lost == tp->lost_out)
2630 return;
2631
2632 if (tcp_is_reno(tp))
2633 tcp_limit_reno_sacked(tp);
2634
2635 tcp_verify_left_out(tp);
2636
2637 /* Don't muck with the congestion window here.
2638 * Reason is that we do not increase amount of _data_
2639 * in network, but units changed and effective
2640 * cwnd/ssthresh really reduced now.
2641 */
2642 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2643 tp->high_seq = tp->snd_nxt;
2644 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2645 tp->prior_ssthresh = 0;
2646 tp->undo_marker = 0;
2647 tcp_set_ca_state(sk, TCP_CA_Loss);
2648 }
2649 tcp_xmit_retransmit_queue(sk);
2650}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002651EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002652
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002653void tcp_enter_recovery(struct sock *sk, bool ece_ack)
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002654{
2655 struct tcp_sock *tp = tcp_sk(sk);
2656 int mib_idx;
2657
2658 if (tcp_is_reno(tp))
2659 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2660 else
2661 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2662
Eric Dumazetc10d9312016-04-29 14:16:47 -07002663 NET_INC_STATS(sock_net(sk), mib_idx);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002664
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002665 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002666 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002667
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002668 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002669 if (!ece_ack)
2670 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002671 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002672 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002673 tcp_set_ca_state(sk, TCP_CA_Recovery);
2674}
2675
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002676/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2677 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2678 */
Yuchung Chenge662ca42016-02-02 10:33:04 -08002679static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2680 int *rexmit)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002681{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002682 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002683 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002684
Yuchung Chengda34ac72015-05-18 12:31:44 -07002685 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2686 tcp_try_undo_loss(sk, false))
2687 return;
2688
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002689 /* The ACK (s)acks some never-retransmitted data meaning not all
2690 * the data packets before the timeout were lost. Therefore we
2691 * undo the congestion window and state. This is essentially
2692 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2693 * a retransmitted skb is permantly marked, we can apply such an
2694 * operation even if F-RTO was not used.
2695 */
2696 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2697 tcp_try_undo_loss(sk, tp->undo_marker))
2698 return;
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002699
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002700 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002701 if (after(tp->snd_nxt, tp->high_seq)) {
2702 if (flag & FLAG_DATA_SACKED || is_dupack)
2703 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002704 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2705 tp->high_seq = tp->snd_nxt;
Yuchung Chenge662ca42016-02-02 10:33:04 -08002706 /* Step 2.b. Try send new data (but deferred until cwnd
2707 * is updated in tcp_ack()). Otherwise fall back to
2708 * the conventional recovery.
2709 */
2710 if (tcp_send_head(sk) &&
2711 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2712 *rexmit = REXMIT_NEW;
2713 return;
2714 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002715 tp->frto = 0;
2716 }
2717 }
2718
2719 if (recovered) {
2720 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002721 tcp_try_undo_recovery(sk);
2722 return;
2723 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002724 if (tcp_is_reno(tp)) {
2725 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2726 * delivered. Lower inflight to clock out (re)tranmissions.
2727 */
2728 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2729 tcp_add_reno_sack(sk);
2730 else if (flag & FLAG_SND_UNA_ADVANCED)
2731 tcp_reset_reno_sack(tp);
2732 }
Yuchung Chenge662ca42016-02-02 10:33:04 -08002733 *rexmit = REXMIT_LOST;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002734}
2735
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002736/* Undo during fast recovery after partial ACK. */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002737static bool tcp_try_undo_partial(struct sock *sk, const int acked)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002738{
2739 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002740
Yuchung Cheng7026b912013-05-29 14:20:13 +00002741 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002742 /* Plain luck! Hole if filled with delayed
2743 * packet, rather than with a retransmit.
2744 */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002745 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2746
2747 /* We are getting evidence that the reordering degree is higher
2748 * than we realized. If there are no retransmits out then we
2749 * can undo. Otherwise we clock out new packets but do not
2750 * mark more packets lost or retransmit more.
2751 */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002752 if (tp->retrans_out)
Yuchung Cheng7026b912013-05-29 14:20:13 +00002753 return true;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002754
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002755 if (!tcp_any_retrans_done(sk))
2756 tp->retrans_stamp = 0;
2757
Yuchung Cheng7026b912013-05-29 14:20:13 +00002758 DBGUNDO(sk, "partial recovery");
2759 tcp_undo_cwnd_reduction(sk, true);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002760 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002761 tcp_try_keep_open(sk);
2762 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002763 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002764 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002765}
2766
Eric Dumazetefab8f82017-04-25 10:15:35 -07002767static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002768{
2769 struct tcp_sock *tp = tcp_sk(sk);
2770
2771 /* Use RACK to detect loss */
Yuchung Chenga0370b32017-01-12 22:11:36 -08002772 if (sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002773 u32 prior_retrans = tp->retrans_out;
2774
Eric Dumazet128eda82017-04-25 10:15:34 -07002775 tcp_rack_mark_lost(sk);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002776 if (prior_retrans > tp->retrans_out)
2777 *ack_flag |= FLAG_LOST_RETRANS;
2778 }
2779}
2780
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781/* Process an event, which can update packets-in-flight not trivially.
2782 * Main goal of this function is to calculate new estimate for left_out,
2783 * taking into account both packets sitting in receiver's buffer and
2784 * packets lost by network.
2785 *
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002786 * Besides that it updates the congestion state when packet loss or ECN
2787 * is detected. But it does not reduce the cwnd, it is done by the
2788 * congestion control later.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 *
2790 * It does _not_ decide what to send, it is made in function
2791 * tcp_xmit_retransmit_queue().
2792 */
Yuchung Cheng68049732013-05-29 14:20:11 +00002793static void tcp_fastretrans_alert(struct sock *sk, const int acked,
Eric Dumazet1317a9d2017-04-25 10:15:36 -07002794 bool is_dupack, int *ack_flag, int *rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002796 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002798 int fast_rexmit = 0, flag = *ack_flag;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002799 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002800 (tcp_fackets_out(tp) > tp->reordering));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002802 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002804 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 tp->fackets_out = 0;
2806
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002807 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002809 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 tp->prior_ssthresh = 0;
2811
2812 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002813 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 return;
2815
Yuchung Cheng974c1232012-01-19 14:42:21 +00002816 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002817 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818
Yuchung Cheng974c1232012-01-19 14:42:21 +00002819 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002821 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002822 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 tp->retrans_stamp = 0;
2824 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002825 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 case TCP_CA_CWR:
2827 /* CWR is to be held something *above* high_seq
2828 * is ACKed for CWR bit to reach receiver. */
2829 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002830 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002831 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 }
2833 break;
2834
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002836 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002838 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002840 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 break;
2842 }
2843 }
2844
Yuchung Cheng974c1232012-01-19 14:42:21 +00002845 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002846 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002848 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002849 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002850 tcp_add_reno_sack(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002851 } else {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002852 if (tcp_try_undo_partial(sk, acked))
Yuchung Cheng7026b912013-05-29 14:20:13 +00002853 return;
2854 /* Partial ACK arrived. Force fast retransmit. */
2855 do_lost = tcp_is_reno(tp) ||
2856 tcp_fackets_out(tp) > tp->reordering;
2857 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002858 if (tcp_try_undo_dsack(sk)) {
2859 tcp_try_keep_open(sk);
2860 return;
2861 }
Eric Dumazetefab8f82017-04-25 10:15:35 -07002862 tcp_rack_identify_loss(sk, ack_flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 break;
2864 case TCP_CA_Loss:
Yuchung Chenge662ca42016-02-02 10:33:04 -08002865 tcp_process_loss(sk, flag, is_dupack, rexmit);
Eric Dumazetefab8f82017-04-25 10:15:35 -07002866 tcp_rack_identify_loss(sk, ack_flag);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002867 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2868 (*ack_flag & FLAG_LOST_RETRANS)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002870 /* Change state if cwnd is undone or retransmits are lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002872 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002873 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 tcp_reset_reno_sack(tp);
2875 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002876 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 }
2878
Neal Cardwellf6982042011-11-16 08:58:04 +00002879 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002880 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881
Eric Dumazetefab8f82017-04-25 10:15:35 -07002882 tcp_rack_identify_loss(sk, ack_flag);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002883 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002884 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 return;
2886 }
2887
John Heffner5d424d52006-03-20 17:53:41 -08002888 /* MTU probe failure: don't reduce cwnd */
2889 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2890 icsk->icsk_mtup.probe_size &&
John Heffner0e7b13682006-03-20 21:32:58 -08002891 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002892 tcp_mtup_probe_failed(sk);
2893 /* Restores the reduction we did in tcp_mtup_probe() */
2894 tp->snd_cwnd++;
2895 tcp_simple_retransmit(sk);
2896 return;
2897 }
2898
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002900 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002901 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 }
2903
Yuchung Cheng3e59cb02013-05-17 13:45:05 +00002904 if (do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002905 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08002906 *rexmit = REXMIT_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907}
2908
Yuchung Chengf6722582015-10-16 21:57:42 -07002909static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2910{
Neal Cardwell64033892016-09-19 23:39:10 -04002911 struct tcp_sock *tp = tcp_sk(sk);
2912 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
Yuchung Chengf6722582015-10-16 21:57:42 -07002913
Neal Cardwell64033892016-09-19 23:39:10 -04002914 minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2915 rtt_us ? : jiffies_to_usecs(1));
Yuchung Chengf6722582015-10-16 21:57:42 -07002916}
2917
Yuchung Chenged084952013-07-22 16:20:48 -07002918static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
Yuchung Chengf6722582015-10-16 21:57:42 -07002919 long seq_rtt_us, long sack_rtt_us,
2920 long ca_rtt_us)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002921{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002922 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002923
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002924 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2925 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2926 * Karn's algorithm forbids taking RTT if some retransmitted data
2927 * is acked (RFC6298).
2928 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002929 if (seq_rtt_us < 0)
2930 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07002931
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 /* RTTM Rule: A TSecr value received in a segment is used to
2933 * update the averaged RTT measurement only if the segment
2934 * acknowledges some new data, i.e., only if it advances the
2935 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002938 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Yuchung Cheng2909d872013-10-24 08:55:25 -07002939 flag & FLAG_ACKED)
Yuchung Chengf6722582015-10-16 21:57:42 -07002940 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2941 tp->rx_opt.rcv_tsecr);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002942 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002943 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944
Yuchung Chengf6722582015-10-16 21:57:42 -07002945 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2946 * always taken together with ACK, SACK, or TS-opts. Any negative
2947 * values will be skipped with the seq_rtt_us < 0 check above.
2948 */
2949 tcp_update_rtt_min(sk, ca_rtt_us);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002950 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002951 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002953 /* RFC6298: only reset backoff on valid RTT measurement. */
2954 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002955 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956}
2957
Yuchung Cheng375fe022013-07-22 16:20:45 -07002958/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002959void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002960{
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002961 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07002962
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002963 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2964 struct skb_mstamp now;
Yuchung Chengbc15afa2013-10-24 08:44:25 -07002965
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002966 skb_mstamp_get(&now);
2967 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2968 }
2969
Yuchung Chengf6722582015-10-16 21:57:42 -07002970 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002971}
2972
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002973
Eric Dumazet24901552014-05-02 21:18:05 -07002974static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002976 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07002977
2978 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002979 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980}
2981
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982/* Restart timer after forward progress on connection.
2983 * RFC2988 recommends to restart timer to now+rto.
2984 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002985void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002987 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002988 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002989
Jerry Chu168a8f52012-08-31 12:29:13 +00002990 /* If the retrans timer is currently being used by Fast Open
2991 * for SYN-ACK retrans purpose, stay put.
2992 */
2993 if (tp->fastopen_rsk)
2994 return;
2995
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002997 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002999 u32 rto = inet_csk(sk)->icsk_rto;
3000 /* Offset the time elapsed after installing regular RTO */
Yuchung Chengbec41a12017-01-12 22:11:39 -08003001 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003002 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003003 struct sk_buff *skb = tcp_write_queue_head(sk);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003004 const u32 rto_time_stamp =
3005 tcp_skb_timestamp(skb) + rto;
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003006 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3007 /* delta may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003008 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003009 */
3010 if (delta > 0)
3011 rto = delta;
3012 }
3013 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3014 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003016}
3017
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003018/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003019static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020{
3021 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003022 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003024 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025
3026 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003027 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 return 0;
3029 packets_acked -= tcp_skb_pcount(skb);
3030
3031 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003033 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 }
3035
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003036 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037}
3038
Eric Dumazetad971f62014-10-11 15:17:29 -07003039static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3040 u32 prior_snd_una)
3041{
3042 const struct skb_shared_info *shinfo;
3043
3044 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
Soheil Hassas Yeganeh6b084922016-04-02 23:08:08 -04003045 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
Eric Dumazetad971f62014-10-11 15:17:29 -07003046 return;
3047
3048 shinfo = skb_shinfo(skb);
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04003049 if (!before(shinfo->tskey, prior_snd_una) &&
Martin KaFai Lau479f85c2016-04-18 15:39:53 -07003050 before(shinfo->tskey, tcp_sk(sk)->snd_una))
Eric Dumazetad971f62014-10-11 15:17:29 -07003051 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3052}
3053
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003054/* Remove acknowledged frames from the retransmission queue. If our packet
3055 * is before the ack sequence we can discard it as it's confirmed to have
3056 * arrived at the other end.
3057 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003058static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003059 u32 prior_snd_una, int *acked,
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003060 struct tcp_sacktag_state *sack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003062 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Chengb9f64822016-09-19 23:39:14 -04003063 struct skb_mstamp first_ackt, last_ackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003064 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet7e0ca8a2017-04-25 10:15:39 -07003065 struct skb_mstamp *now = &tp->tcp_mstamp;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003066 u32 prior_sacked = tp->sacked_out;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003067 u32 reord = tp->packets_out;
3068 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003069 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003070 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003071 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003072 struct sk_buff *skb;
3073 u32 pkts_acked = 0;
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003074 u32 last_in_flight = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003075 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003076 int flag = 0;
3077
3078 first_ackt.v64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003080 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003081 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003082 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003083 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084
Eric Dumazetad971f62014-10-11 15:17:29 -07003085 tcp_ack_tstamp(sk, skb, prior_snd_una);
Willem de Bruijn712a7222014-08-12 14:53:16 -04003086
Gavin McCullagh2072c222007-12-29 19:11:21 -08003087 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003089 if (tcp_skb_pcount(skb) == 1 ||
3090 !after(tp->snd_una, scb->seq))
3091 break;
3092
Ilpo Järvinen72018832007-12-30 04:37:55 -08003093 acked_pcount = tcp_tso_acked(sk, skb);
3094 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003095 break;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003096 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003097 } else {
Eric Dumazetad971f62014-10-11 15:17:29 -07003098 /* Speedup tcp_unlink_write_queue() and next loop */
3099 prefetchw(skb->next);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003100 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101 }
3102
Eric Dumazetad971f62014-10-11 15:17:29 -07003103 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003104 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003105 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003106 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003107 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003108 last_ackt = skb->skb_mstamp;
Eric Dumazet431a9122014-03-09 17:36:02 -07003109 WARN_ON_ONCE(last_ackt.v64 == 0);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003110 if (!first_ackt.v64)
3111 first_ackt = last_ackt;
3112
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003113 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003114 reord = min(pkts_acked, reord);
3115 if (!after(scb->end_seq, tp->high_seq))
3116 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003117 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003118
Yuchung Chengddf1af62016-02-02 10:33:06 -08003119 if (sacked & TCPCB_SACKED_ACKED) {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003120 tp->sacked_out -= acked_pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003121 } else if (tcp_is_sack(tp)) {
3122 tp->delivered += acked_pcount;
3123 if (!tcp_skb_spurious_retrans(tp, skb))
Yuchung Cheng1d0833d2017-01-12 22:11:34 -08003124 tcp_rack_advance(tp, sacked, scb->end_seq,
Eric Dumazetd2329f12017-04-25 10:15:38 -07003125 &skb->skb_mstamp);
Yuchung Chengddf1af62016-02-02 10:33:06 -08003126 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003127 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003128 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003129
Ilpo Järvinen72018832007-12-30 04:37:55 -08003130 tp->packets_out -= acked_pcount;
3131 pkts_acked += acked_pcount;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003132 tcp_rate_skb_delivered(sk, skb, sack->rate);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003133
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003134 /* Initial outgoing SYN's get put onto the write_queue
3135 * just like anything else we transmit. It is not
3136 * true data, and if we misinform our callers that
3137 * this ACK acks real data, we will erroneously exit
3138 * connection startup slow start one packet too
3139 * quickly. This is severely frowned upon behavior.
3140 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003141 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003142 flag |= FLAG_DATA_ACKED;
3143 } else {
3144 flag |= FLAG_SYN_ACKED;
3145 tp->retrans_stamp = 0;
3146 }
3147
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003148 if (!fully_acked)
3149 break;
3150
David S. Millerfe067e82007-03-07 12:12:44 -08003151 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003152 sk_wmem_free_skb(sk, skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003153 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003154 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003155 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003156 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 }
3158
Francis Yan0f872302016-11-27 23:07:14 -08003159 if (!skb)
3160 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3161
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003162 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3163 tp->snd_up = tp->snd_una;
3164
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003165 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3166 flag |= FLAG_SACK_RENEGING;
3167
Yuchung Cheng9e45a3e2015-10-16 21:57:41 -07003168 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
Yuchung Chengb9f64822016-09-19 23:39:14 -04003169 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3170 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003171 }
3172 if (sack->first_sackt.v64) {
Yuchung Chengb9f64822016-09-19 23:39:14 -04003173 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3174 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003175 }
Yuchung Chengb9f64822016-09-19 23:39:14 -04003176 sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
Yuchung Chengf6722582015-10-16 21:57:42 -07003177 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3178 ca_rtt_us);
Yuchung Chenged084952013-07-22 16:20:48 -07003179
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003180 if (flag & FLAG_ACKED) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003181 tcp_rearm_rto(sk);
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003182 if (unlikely(icsk->icsk_mtup.probe_size &&
3183 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3184 tcp_mtup_probe_success(sk);
3185 }
3186
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003187 if (tcp_is_reno(tp)) {
3188 tcp_remove_reno_sacks(sk, pkts_acked);
3189 } 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 */
3193 if (reord < prior_fackets)
3194 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003195
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003196 delta = tcp_is_fack(tp) ? pkts_acked :
3197 prior_sacked - tp->sacked_out;
3198 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003199 }
3200
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003201 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003202
Eric Dumazet740b0f12014-02-26 14:02:48 -08003203 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
Yuchung Chengb9f64822016-09-19 23:39:14 -04003204 sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003205 /* Do not re-arm RTO if the sack RTT is measured from data sent
3206 * after when the head was last (re)transmitted. Otherwise the
3207 * timeout may continue to extend in loss recovery.
3208 */
3209 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 }
3211
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003212 if (icsk->icsk_ca_ops->pkts_acked) {
3213 struct ack_sample sample = { .pkts_acked = pkts_acked,
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003214 .rtt_us = ca_rtt_us,
3215 .in_flight = last_in_flight };
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003216
3217 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3218 }
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003219
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003221 WARN_ON((int)tp->sacked_out < 0);
3222 WARN_ON((int)tp->lost_out < 0);
3223 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003224 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003225 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003227 pr_debug("Leak l=%u %d\n",
3228 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229 tp->lost_out = 0;
3230 }
3231 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003232 pr_debug("Leak s=%u %d\n",
3233 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 tp->sacked_out = 0;
3235 }
3236 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003237 pr_debug("Leak r=%u %d\n",
3238 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 tp->retrans_out = 0;
3240 }
3241 }
3242#endif
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003243 *acked = pkts_acked;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003244 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245}
3246
3247static void tcp_ack_probe(struct sock *sk)
3248{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003249 const struct tcp_sock *tp = tcp_sk(sk);
3250 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251
3252 /* Was it a usable window open? */
3253
Ilpo Järvinen90840defa2007-12-31 04:48:41 -08003254 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003255 icsk->icsk_backoff = 0;
3256 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 /* Socket must be waked up by subsequent tcp_data_snd_check().
3258 * This function is not for random using!
3259 */
3260 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003261 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003262
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003263 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003264 when, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265 }
3266}
3267
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003268static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003270 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3271 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272}
3273
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003274/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003275static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003277 /* If reordering is high then always grow cwnd whenever data is
3278 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003279 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003280 * new SACK or ECE mark may first advance cwnd here and later reduce
3281 * cwnd in tcp_fastretrans_alert() based on more states.
3282 */
Nikolay Borisov1043e252016-02-03 09:46:52 +02003283 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003284 return flag & FLAG_FORWARD_PROGRESS;
3285
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003286 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287}
3288
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003289/* The "ultimate" congestion control function that aims to replace the rigid
3290 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3291 * It's called toward the end of processing an ACK with precise rate
3292 * information. All transmission or retransmission are delayed afterwards.
3293 */
3294static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
Yuchung Chengc0402762016-09-19 23:39:21 -04003295 int flag, const struct rate_sample *rs)
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003296{
Yuchung Chengc0402762016-09-19 23:39:21 -04003297 const struct inet_connection_sock *icsk = inet_csk(sk);
3298
3299 if (icsk->icsk_ca_ops->cong_control) {
3300 icsk->icsk_ca_ops->cong_control(sk, rs);
3301 return;
3302 }
3303
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003304 if (tcp_in_cwnd_reduction(sk)) {
3305 /* Reduce cwnd if state mandates */
3306 tcp_cwnd_reduction(sk, acked_sacked, flag);
3307 } else if (tcp_may_raise_cwnd(sk, flag)) {
3308 /* Advance cwnd if state allows */
3309 tcp_cong_avoid(sk, ack, acked_sacked);
3310 }
3311 tcp_update_pacing_rate(sk);
3312}
3313
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314/* Check that window update is acceptable.
3315 * The function assumes that snd_una<=ack<=snd_next.
3316 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003317static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003318 const u32 ack, const u32 ack_seq,
3319 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003321 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003323 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324}
3325
Eric Dumazet0df48c22015-04-28 15:28:17 -07003326/* If we update tp->snd_una, also update tp->bytes_acked */
3327static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3328{
3329 u32 delta = ack - tp->snd_una;
3330
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003331 sock_owned_by_me((struct sock *)tp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003332 tp->bytes_acked += delta;
3333 tp->snd_una = ack;
3334}
3335
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003336/* If we update tp->rcv_nxt, also update tp->bytes_received */
3337static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3338{
3339 u32 delta = seq - tp->rcv_nxt;
3340
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003341 sock_owned_by_me((struct sock *)tp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003342 tp->bytes_received += delta;
3343 tp->rcv_nxt = seq;
3344}
3345
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346/* Update our send window.
3347 *
3348 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3349 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3350 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003351static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003352 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003354 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003356 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003358 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 nwin <<= tp->rx_opt.snd_wscale;
3360
3361 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3362 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003363 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364
3365 if (tp->snd_wnd != nwin) {
3366 tp->snd_wnd = nwin;
3367
3368 /* Note, it is the only place, where
3369 * fast path is recovered for sending TCP.
3370 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003371 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003372 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
Eric Dumazet6f021c62015-08-21 12:30:00 -07003374 if (tcp_send_head(sk))
3375 tcp_slow_start_after_idle_check(sk);
3376
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 if (nwin > tp->max_window) {
3378 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003379 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 }
3381 }
3382 }
3383
Eric Dumazet0df48c22015-04-28 15:28:17 -07003384 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385
3386 return flag;
3387}
3388
Jason Baron083ae302016-07-14 11:38:40 -04003389static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3390 u32 *last_oow_ack_time)
3391{
3392 if (*last_oow_ack_time) {
3393 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3394
3395 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3396 NET_INC_STATS(net, mib_idx);
3397 return true; /* rate-limited: don't send yet! */
3398 }
3399 }
3400
3401 *last_oow_ack_time = tcp_time_stamp;
3402
3403 return false; /* not rate-limited: go ahead, send dupack now! */
3404}
3405
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003406/* Return true if we're currently rate-limiting out-of-window ACKs and
3407 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3408 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3409 * attacks that send repeated SYNs or ACKs for the same connection. To
3410 * do this, we do not send a duplicate SYNACK or ACK if the remote
3411 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3412 */
3413bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3414 int mib_idx, u32 *last_oow_ack_time)
3415{
3416 /* Data packets without SYNs are not likely part of an ACK loop. */
3417 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3418 !tcp_hdr(skb)->syn)
Jason Baron083ae302016-07-14 11:38:40 -04003419 return false;
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003420
Jason Baron083ae302016-07-14 11:38:40 -04003421 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003422}
3423
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003424/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003425static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003426{
3427 /* unprotected vars, we dont care of overwrites */
3428 static u32 challenge_timestamp;
3429 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003430 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003431 u32 count, now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003432
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003433 /* First check our per-socket dupack rate limit. */
Jason Baron083ae302016-07-14 11:38:40 -04003434 if (__tcp_oow_rate_limited(sock_net(sk),
3435 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3436 &tp->last_oow_ack_time))
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003437 return;
3438
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003439 /* Then check host-wide RFC 5961 rate limit. */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003440 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003441 if (now != challenge_timestamp) {
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003442 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3443
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003444 challenge_timestamp = now;
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003445 WRITE_ONCE(challenge_count, half +
3446 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003447 }
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003448 count = READ_ONCE(challenge_count);
3449 if (count > 0) {
3450 WRITE_ONCE(challenge_count, count - 1);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003451 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003452 tcp_send_ack(sk);
3453 }
3454}
3455
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003456static void tcp_store_ts_recent(struct tcp_sock *tp)
3457{
3458 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3459 tp->rx_opt.ts_recent_stamp = get_seconds();
3460}
3461
3462static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3463{
3464 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3465 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3466 * extra check below makes sure this can only happen
3467 * for pure ACK frames. -DaveM
3468 *
3469 * Not only, also it occurs for expired timestamps.
3470 */
3471
3472 if (tcp_paws_check(&tp->rx_opt, 0))
3473 tcp_store_ts_recent(tp);
3474 }
3475}
3476
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003477/* This routine deals with acks during a TLP episode.
Sébastien Barré08abdff2015-01-12 10:30:40 +01003478 * We mark the end of a TLP episode on receiving TLP dupack or when
3479 * ack is after tlp_high_seq.
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003480 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3481 */
3482static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3483{
3484 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003485
Sébastien Barré08abdff2015-01-12 10:30:40 +01003486 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003487 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003488
Sébastien Barré08abdff2015-01-12 10:30:40 +01003489 if (flag & FLAG_DSACKING_ACK) {
3490 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003491 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003492 } else if (after(ack, tp->tlp_high_seq)) {
3493 /* ACK advances: there was a loss, so reduce cwnd. Reset
3494 * tlp_high_seq in tcp_init_cwnd_reduction()
3495 */
3496 tcp_init_cwnd_reduction(sk);
3497 tcp_set_ca_state(sk, TCP_CA_CWR);
3498 tcp_end_cwnd_reduction(sk);
3499 tcp_try_keep_open(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003500 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003501 LINUX_MIB_TCPLOSSPROBERECOVERY);
Sébastien Barré08abdff2015-01-12 10:30:40 +01003502 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3503 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3504 /* Pure dupack: original and TLP probe arrived; no loss */
3505 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003506 }
3507}
3508
Florian Westphal7354c8c2014-09-26 22:37:34 +02003509static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3510{
3511 const struct inet_connection_sock *icsk = inet_csk(sk);
3512
3513 if (icsk->icsk_ca_ops->in_ack_event)
3514 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3515}
3516
Yuchung Chenge662ca42016-02-02 10:33:04 -08003517/* Congestion control has updated the cwnd already. So if we're in
3518 * loss recovery then now we do any new sends (for FRTO) or
3519 * retransmits (for CA_Loss or CA_recovery) that make sense.
3520 */
3521static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3522{
3523 struct tcp_sock *tp = tcp_sk(sk);
3524
3525 if (rexmit == REXMIT_NONE)
3526 return;
3527
3528 if (unlikely(rexmit == 2)) {
3529 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3530 TCP_NAGLE_OFF);
3531 if (after(tp->snd_nxt, tp->high_seq))
3532 return;
3533 tp->frto = 0;
3534 }
3535 tcp_xmit_retransmit_queue(sk);
3536}
3537
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003539static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003541 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003543 struct tcp_sacktag_state sack_state;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003544 struct rate_sample rs = { .prior_delivered = 0 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 u32 prior_snd_una = tp->snd_una;
3546 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3547 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003548 bool is_dupack = false;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003549 u32 prior_fackets;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003550 int prior_packets = tp->packets_out;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003551 u32 delivered = tp->delivered;
3552 u32 lost = tp->lost;
Yuchung Cheng68049732013-05-29 14:20:11 +00003553 int acked = 0; /* Number of packets newly acked */
Yuchung Chenge662ca42016-02-02 10:33:04 -08003554 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003555
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003556 sack_state.first_sackt.v64 = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003557 sack_state.rate = &rs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558
Eric Dumazetad971f62014-10-11 15:17:29 -07003559 /* We very likely will need to access write queue head. */
3560 prefetchw(sk->sk_write_queue.next);
3561
John Dykstra96e0bf42009-03-22 21:49:57 -07003562 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 * then we can probably ignore it.
3564 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003565 if (before(ack, prior_snd_una)) {
3566 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3567 if (before(ack, prior_snd_una - tp->max_window)) {
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003568 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003569 return -1;
3570 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003572 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573
John Dykstra96e0bf42009-03-22 21:49:57 -07003574 /* If the ack includes data we haven't sent yet, discard
3575 * this segment (RFC793 Section 3.9).
3576 */
3577 if (after(ack, tp->snd_nxt))
3578 goto invalid_ack;
3579
Yuchung Chengbec41a12017-01-12 22:11:39 -08003580 if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003581 tcp_rearm_rto(sk);
3582
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003583 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003584 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003585 icsk->icsk_retransmits = 0;
3586 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003587
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003588 prior_fackets = tp->fackets_out;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003589 rs.prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003590
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003591 /* ts_recent update must be made after we are sure that the packet
3592 * is in window.
3593 */
3594 if (flag & FLAG_UPDATE_TS_RECENT)
3595 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3596
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003597 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 /* Window is constant, pure forward advance.
3599 * No more checks are required.
3600 * Note, we use the fact that SND.UNA>=SND.WL2.
3601 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003602 tcp_update_wl(tp, ack_seq);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003603 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 flag |= FLAG_WIN_UPDATE;
3605
Florian Westphal98900922014-09-26 22:37:35 +02003606 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003607
Eric Dumazetc10d9312016-04-29 14:16:47 -07003608 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 } else {
Florian Westphal98900922014-09-26 22:37:35 +02003610 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3611
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3613 flag |= FLAG_DATA;
3614 else
Eric Dumazetc10d9312016-04-29 14:16:47 -07003615 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003617 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618
3619 if (TCP_SKB_CB(skb)->sacked)
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003620 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003621 &sack_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Florian Westphal735d3832014-09-29 13:08:30 +02003623 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 flag |= FLAG_ECE;
Florian Westphal98900922014-09-26 22:37:35 +02003625 ack_ev_flags |= CA_ACK_ECE;
3626 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627
Florian Westphal98900922014-09-26 22:37:35 +02003628 if (flag & FLAG_WIN_UPDATE)
3629 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3630
3631 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 }
3633
3634 /* We passed data and got it acked, remove any soft error
3635 * log. Something worked...
3636 */
3637 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003638 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 tp->rcv_tstamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640 if (!prior_packets)
3641 goto no_queue;
3642
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003644 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003645 &sack_state);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003646
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003647 if (tcp_ack_is_dubious(sk, flag)) {
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003648 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Eric Dumazet1317a9d2017-04-25 10:15:36 -07003649 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003651 if (tp->tlp_high_seq)
3652 tcp_process_tlp_ack(sk, ack, flag);
3653
Julian Anastasovc3a2e832017-02-06 23:14:14 +02003654 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3655 sk_dst_confirm(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003656
3657 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3658 tcp_schedule_loss_probe(sk);
Yuchung Chengb9f64822016-09-19 23:39:14 -04003659 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3660 lost = tp->lost - lost; /* freshly marked lost */
Eric Dumazet88d5c652017-04-25 10:15:37 -07003661 tcp_rate_gen(sk, delivered, lost, sack_state.rate);
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003662 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003663 tcp_xmit_recovery(sk, rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 return 1;
3665
3666no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003667 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3668 if (flag & FLAG_DSACKING_ACK)
Eric Dumazet1317a9d2017-04-25 10:15:36 -07003669 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 /* If this ack opens up a zero window, clear backoff. It was
3671 * being used to time the probes, and is probably far higher than
3672 * it needs to be for normal retransmission.
3673 */
David S. Millerfe067e82007-03-07 12:12:44 -08003674 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003676
3677 if (tp->tlp_high_seq)
3678 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 return 1;
3680
John Dykstra96e0bf42009-03-22 21:49:57 -07003681invalid_ack:
3682 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3683 return -1;
3684
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003686 /* If data was SACKed, tag it and see if we should send more data.
3687 * If data was DSACKed, see if we can undo a cwnd reduction.
3688 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003689 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003690 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003691 &sack_state);
Eric Dumazet1317a9d2017-04-25 10:15:36 -07003692 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003693 tcp_xmit_recovery(sk, rexmit);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695
John Dykstra96e0bf42009-03-22 21:49:57 -07003696 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 return 0;
3698}
3699
Daniel Lee7f9b8382015-04-06 14:37:26 -07003700static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3701 bool syn, struct tcp_fastopen_cookie *foc,
3702 bool exp_opt)
3703{
3704 /* Valid only in SYN or SYN-ACK with an even length. */
3705 if (!foc || !syn || len < 0 || (len & 1))
3706 return;
3707
3708 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3709 len <= TCP_FASTOPEN_COOKIE_MAX)
3710 memcpy(foc->val, cookie, len);
3711 else if (len != 0)
3712 len = -1;
3713 foc->len = len;
3714 foc->exp = exp_opt;
3715}
3716
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3718 * But, this can also be called on packets in the established flow when
3719 * the fast version below fails.
3720 */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003721void tcp_parse_options(const struct sk_buff *skb,
3722 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003723 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003725 const unsigned char *ptr;
3726 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003727 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003729 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 opt_rx->saw_tstamp = 0;
3731
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003732 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003733 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 int opsize;
3735
3736 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003737 case TCPOPT_EOL:
3738 return;
3739 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3740 length--;
3741 continue;
3742 default:
3743 opsize = *ptr++;
3744 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003746 if (opsize > length)
3747 return; /* don't parse partial options */
3748 switch (opcode) {
3749 case TCPOPT_MSS:
3750 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003751 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003752 if (in_mss) {
3753 if (opt_rx->user_mss &&
3754 opt_rx->user_mss < in_mss)
3755 in_mss = opt_rx->user_mss;
3756 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003758 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003759 break;
3760 case TCPOPT_WINDOW:
3761 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003762 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003763 __u8 snd_wscale = *(__u8 *)ptr;
3764 opt_rx->wscale_ok = 1;
Gao Feng589c49c2017-04-04 21:09:48 +08003765 if (snd_wscale > TCP_MAX_WSCALE) {
3766 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
Joe Perchese87cc472012-05-13 21:56:26 +00003767 __func__,
Gao Feng589c49c2017-04-04 21:09:48 +08003768 snd_wscale,
3769 TCP_MAX_WSCALE);
3770 snd_wscale = TCP_MAX_WSCALE;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003771 }
3772 opt_rx->snd_wscale = snd_wscale;
3773 }
3774 break;
3775 case TCPOPT_TIMESTAMP:
3776 if ((opsize == TCPOLEN_TIMESTAMP) &&
3777 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003778 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003779 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003780 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3781 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003782 }
3783 break;
3784 case TCPOPT_SACK_PERM:
3785 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003786 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003787 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003788 tcp_sack_reset(opt_rx);
3789 }
3790 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003791
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003792 case TCPOPT_SACK:
3793 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3794 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3795 opt_rx->sack_ok) {
3796 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3797 }
3798 break;
3799#ifdef CONFIG_TCP_MD5SIG
3800 case TCPOPT_MD5SIG:
3801 /*
3802 * The MD5 Hash has already been
3803 * checked (see tcp_v{4,6}_do_rcv()).
3804 */
3805 break;
3806#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07003807 case TCPOPT_FASTOPEN:
3808 tcp_parse_fastopen_option(
3809 opsize - TCPOLEN_FASTOPEN_BASE,
3810 ptr, th->syn, foc, false);
3811 break;
3812
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003813 case TCPOPT_EXP:
3814 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07003815 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003816 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07003817 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3818 get_unaligned_be16(ptr) ==
3819 TCPOPT_FASTOPEN_MAGIC)
3820 tcp_parse_fastopen_option(opsize -
3821 TCPOLEN_EXP_FASTOPEN_BASE,
3822 ptr + 2, th->syn, foc, true);
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003823 break;
3824
3825 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003826 ptr += opsize-2;
3827 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003828 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 }
3830}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003831EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003833static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003834{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003835 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003836
3837 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3838 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3839 tp->rx_opt.saw_tstamp = 1;
3840 ++ptr;
3841 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3842 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003843 if (*ptr)
3844 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3845 else
3846 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003847 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003848 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003849 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003850}
3851
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852/* Fast parse options. This hopes to only see timestamps.
3853 * If it is wrong it falls back on tcp_parse_options().
3854 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003855static bool tcp_fast_parse_options(const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003856 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003858 /* In the spirit of fast parsing, compare doff directly to constant
3859 * values. Because equality is used, short doff can be ignored here.
3860 */
3861 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003863 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003865 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003866 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003867 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003868 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003869
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003870 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003871 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003872 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3873
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003874 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875}
3876
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003877#ifdef CONFIG_TCP_MD5SIG
3878/*
3879 * Parse MD5 Signature option
3880 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003881const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003882{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003883 int length = (th->doff << 2) - sizeof(*th);
3884 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003885
3886 /* If the TCP option is too short, we can short cut */
3887 if (length < TCPOLEN_MD5SIG)
3888 return NULL;
3889
3890 while (length > 0) {
3891 int opcode = *ptr++;
3892 int opsize;
3893
Weilong Chena22318e2013-12-23 14:37:26 +08003894 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003895 case TCPOPT_EOL:
3896 return NULL;
3897 case TCPOPT_NOP:
3898 length--;
3899 continue;
3900 default:
3901 opsize = *ptr++;
3902 if (opsize < 2 || opsize > length)
3903 return NULL;
3904 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003905 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003906 }
3907 ptr += opsize - 2;
3908 length -= opsize;
3909 }
3910 return NULL;
3911}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003912EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003913#endif
3914
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3916 *
3917 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3918 * it can pass through stack. So, the following predicate verifies that
3919 * this segment is not used for anything but congestion avoidance or
3920 * fast retransmit. Moreover, we even are able to eliminate most of such
3921 * second order effects, if we apply some small "replay" window (~RTO)
3922 * to timestamp space.
3923 *
3924 * All these measures still do not guarantee that we reject wrapped ACKs
3925 * on networks with high bandwidth, when sequence space is recycled fastly,
3926 * but it guarantees that such events will be very rare and do not affect
3927 * connection seriously. This doesn't look nice, but alas, PAWS is really
3928 * buggy extension.
3929 *
3930 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3931 * states that events when retransmit arrives after original data are rare.
3932 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3933 * the biggest problem on large power networks even with minor reordering.
3934 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3935 * up to bandwidth of 18Gigabit/sec. 8) ]
3936 */
3937
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003938static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003940 const struct tcp_sock *tp = tcp_sk(sk);
3941 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 u32 seq = TCP_SKB_CB(skb)->seq;
3943 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3944
3945 return (/* 1. Pure ACK with correct sequence number. */
3946 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3947
3948 /* 2. ... and duplicate ACK. */
3949 ack == tp->snd_una &&
3950
3951 /* 3. ... and does not update window. */
3952 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3953
3954 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003955 (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 -07003956}
3957
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003958static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003959 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003961 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003962
3963 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3964 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965}
3966
3967/* Check segment sequence number for validity.
3968 *
3969 * Segment controls are considered valid, if the segment
3970 * fits to the window after truncation to the window. Acceptability
3971 * of data (and SYN, FIN, of course) is checked separately.
3972 * See tcp_data_queue(), for example.
3973 *
3974 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3975 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3976 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3977 * (borrowed from freebsd)
3978 */
3979
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003980static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981{
3982 return !before(end_seq, tp->rcv_wup) &&
3983 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3984}
3985
3986/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00003987void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988{
3989 /* We want the right error as BSD sees it (and indeed as we do). */
3990 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003991 case TCP_SYN_SENT:
3992 sk->sk_err = ECONNREFUSED;
3993 break;
3994 case TCP_CLOSE_WAIT:
3995 sk->sk_err = EPIPE;
3996 break;
3997 case TCP_CLOSE:
3998 return;
3999 default:
4000 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004002 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4003 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004
Eric Dumazet3d476262017-04-18 09:45:51 -07004005 tcp_done(sk);
4006
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007 if (!sock_flag(sk, SOCK_DEAD))
4008 sk->sk_error_report(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009}
4010
4011/*
4012 * Process the FIN bit. This now behaves as it is supposed to work
4013 * and the FIN takes effect when it is validly part of sequence
4014 * space. Not before when we get holes.
4015 *
4016 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4017 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4018 * TIME-WAIT)
4019 *
4020 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4021 * close and we go into CLOSING (and later onto TIME-WAIT)
4022 *
4023 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4024 */
Eric Dumazete3e17b72016-02-06 11:16:28 -08004025void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026{
4027 struct tcp_sock *tp = tcp_sk(sk);
4028
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004029 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030
4031 sk->sk_shutdown |= RCV_SHUTDOWN;
4032 sock_set_flag(sk, SOCK_DONE);
4033
4034 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004035 case TCP_SYN_RECV:
4036 case TCP_ESTABLISHED:
4037 /* Move to CLOSE_WAIT */
4038 tcp_set_state(sk, TCP_CLOSE_WAIT);
Jon Maxwell2251ae42015-07-08 10:12:28 +10004039 inet_csk(sk)->icsk_ack.pingpong = 1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004040 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004042 case TCP_CLOSE_WAIT:
4043 case TCP_CLOSING:
4044 /* Received a retransmission of the FIN, do
4045 * nothing.
4046 */
4047 break;
4048 case TCP_LAST_ACK:
4049 /* RFC793: Remain in the LAST-ACK state. */
4050 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004052 case TCP_FIN_WAIT1:
4053 /* This case occurs when a simultaneous close
4054 * happens, we must ack the received FIN and
4055 * enter the CLOSING state.
4056 */
4057 tcp_send_ack(sk);
4058 tcp_set_state(sk, TCP_CLOSING);
4059 break;
4060 case TCP_FIN_WAIT2:
4061 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4062 tcp_send_ack(sk);
4063 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4064 break;
4065 default:
4066 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4067 * cases we should never reach this piece of code.
4068 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004069 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004070 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004071 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004072 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
4074 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4075 * Probably, we should reset in this case. For now drop them.
4076 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004077 skb_rbtree_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004078 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004080 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081
4082 if (!sock_flag(sk, SOCK_DEAD)) {
4083 sk->sk_state_change(sk);
4084
4085 /* Do not send POLL_HUP for half duplex close. */
4086 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4087 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004088 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004090 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091 }
4092}
4093
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004094static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004095 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096{
4097 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4098 if (before(seq, sp->start_seq))
4099 sp->start_seq = seq;
4100 if (after(end_seq, sp->end_seq))
4101 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004102 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004104 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105}
4106
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004107static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004109 struct tcp_sock *tp = tcp_sk(sk);
4110
David S. Millerbb5b7c12009-12-15 20:56:42 -08004111 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004112 int mib_idx;
4113
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004115 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004117 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4118
Eric Dumazetc10d9312016-04-29 14:16:47 -07004119 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
4121 tp->rx_opt.dsack = 1;
4122 tp->duplicate_sack[0].start_seq = seq;
4123 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124 }
4125}
4126
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004127static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004129 struct tcp_sock *tp = tcp_sk(sk);
4130
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004132 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133 else
4134 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4135}
4136
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004137static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138{
4139 struct tcp_sock *tp = tcp_sk(sk);
4140
4141 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4142 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004143 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004144 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
David S. Millerbb5b7c12009-12-15 20:56:42 -08004146 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4148
4149 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4150 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004151 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 }
4153 }
4154
4155 tcp_send_ack(sk);
4156}
4157
4158/* These routines update the SACK block as out-of-order packets arrive or
4159 * in-order packets close up the sequence space.
4160 */
4161static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4162{
4163 int this_sack;
4164 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004165 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166
4167 /* See if the recent change to the first SACK eats into
4168 * or hits the sequence space of other SACK blocks, if so coalesce.
4169 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004170 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4172 int i;
4173
4174 /* Zap SWALK, by moving every further SACK up by one slot.
4175 * Decrease num_sacks.
4176 */
4177 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004178 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4179 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 continue;
4181 }
4182 this_sack++, swalk++;
4183 }
4184}
4185
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4187{
4188 struct tcp_sock *tp = tcp_sk(sk);
4189 struct tcp_sack_block *sp = &tp->selective_acks[0];
4190 int cur_sacks = tp->rx_opt.num_sacks;
4191 int this_sack;
4192
4193 if (!cur_sacks)
4194 goto new_sack;
4195
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004196 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197 if (tcp_sack_extend(sp, seq, end_seq)) {
4198 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004199 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004200 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 if (cur_sacks > 1)
4202 tcp_sack_maybe_coalesce(tp);
4203 return;
4204 }
4205 }
4206
4207 /* Could not find an adjacent existing SACK, build a new one,
4208 * put it at the front, and shift everyone else down. We
4209 * always know there is at least one SACK present already here.
4210 *
4211 * If the sack array is full, forget about the last one.
4212 */
Adam Langley4389dde2008-07-19 00:07:02 -07004213 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 this_sack--;
4215 tp->rx_opt.num_sacks--;
4216 sp--;
4217 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004218 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004219 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220
4221new_sack:
4222 /* Build the new head SACK, and we're done. */
4223 sp->start_seq = seq;
4224 sp->end_seq = end_seq;
4225 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226}
4227
4228/* RCV.NXT advances, some SACKs should be eaten. */
4229
4230static void tcp_sack_remove(struct tcp_sock *tp)
4231{
4232 struct tcp_sack_block *sp = &tp->selective_acks[0];
4233 int num_sacks = tp->rx_opt.num_sacks;
4234 int this_sack;
4235
4236 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004237 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 return;
4240 }
4241
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004242 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243 /* Check if the start of the sack is covered by RCV.NXT. */
4244 if (!before(tp->rcv_nxt, sp->start_seq)) {
4245 int i;
4246
4247 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004248 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249
4250 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004251 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 tp->selective_acks[i-1] = tp->selective_acks[i];
4253 num_sacks--;
4254 continue;
4255 }
4256 this_sack++;
4257 sp++;
4258 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004259 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260}
4261
Eric Dumazet1402d362012-04-23 07:11:42 +00004262/**
4263 * tcp_try_coalesce - try to merge skb to prior one
4264 * @sk: socket
4265 * @to: prior buffer
4266 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004267 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004268 *
4269 * Before queueing skb @from after @to, try to merge them
4270 * to reduce overall memory use and queue lengths, if cost is small.
4271 * Packets in ofo or receive queues can stay a long time.
4272 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004273 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004274 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004275static bool tcp_try_coalesce(struct sock *sk,
4276 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004277 struct sk_buff *from,
4278 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004279{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004280 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004281
Eric Dumazet329033f2012-04-27 00:38:33 +00004282 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004283
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004284 /* Its possible this segment overlaps with prior segment in queue */
4285 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4286 return false;
4287
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004288 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004289 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004290
Alexander Duyck34a802a2012-05-02 21:19:09 +00004291 atomic_add(delta, &sk->sk_rmem_alloc);
4292 sk_mem_charge(sk, delta);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004293 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004294 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4295 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004296 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004297 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004298}
4299
Eric Dumazet532182c2016-04-01 08:52:19 -07004300static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4301{
4302 sk_drops_add(sk, skb);
4303 __kfree_skb(skb);
4304}
4305
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306/* This one checks to see if we can put data from the
4307 * out_of_order queue into the receive_queue.
4308 */
4309static void tcp_ofo_queue(struct sock *sk)
4310{
4311 struct tcp_sock *tp = tcp_sk(sk);
4312 __u32 dsack_high = tp->rcv_nxt;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004313 bool fin, fragstolen, eaten;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004314 struct sk_buff *skb, *tail;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004315 struct rb_node *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004317 p = rb_first(&tp->out_of_order_queue);
4318 while (p) {
4319 skb = rb_entry(p, struct sk_buff, rbnode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4321 break;
4322
4323 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4324 __u32 dsack = dsack_high;
4325 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4326 dsack_high = TCP_SKB_CB(skb)->end_seq;
4327 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4328 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004329 p = rb_next(p);
4330 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004332 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333 SOCK_DEBUG(sk, "ofo packet was already received\n");
Eric Dumazet532182c2016-04-01 08:52:19 -07004334 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335 continue;
4336 }
4337 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4338 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4339 TCP_SKB_CB(skb)->end_seq);
4340
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004341 tail = skb_peek_tail(&sk->sk_receive_queue);
4342 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004343 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004344 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004345 if (!eaten)
4346 __skb_queue_tail(&sk->sk_receive_queue, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004347 else
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004348 kfree_skb_partial(skb, fragstolen);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004349
4350 if (unlikely(fin)) {
4351 tcp_fin(sk);
4352 /* tcp_fin() purges tp->out_of_order_queue,
4353 * so we must end this loop right now.
4354 */
4355 break;
4356 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357 }
4358}
4359
4360static bool tcp_prune_ofo_queue(struct sock *sk);
4361static int tcp_prune_queue(struct sock *sk);
4362
4363static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4364 unsigned int size)
4365{
4366 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4367 !sk_rmem_schedule(sk, skb, size)) {
4368
4369 if (tcp_prune_queue(sk) < 0)
4370 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004371
Eric Dumazet36a65032016-08-17 14:17:09 -07004372 while (!sk_rmem_schedule(sk, skb, size)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373 if (!tcp_prune_ofo_queue(sk))
4374 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375 }
4376 }
4377 return 0;
4378}
4379
Eric Dumazete86b29192012-03-18 11:06:44 +00004380static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4381{
4382 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004383 struct rb_node **p, *q, *parent;
Eric Dumazete86b29192012-03-18 11:06:44 +00004384 struct sk_buff *skb1;
4385 u32 seq, end_seq;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004386 bool fragstolen;
Eric Dumazete86b29192012-03-18 11:06:44 +00004387
Florian Westphal735d3832014-09-29 13:08:30 +02004388 tcp_ecn_check_ce(tp, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004389
Mel Gormanc76562b2012-07-31 16:44:41 -07004390 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004391 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazet532182c2016-04-01 08:52:19 -07004392 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004393 return;
4394 }
4395
4396 /* Disable header prediction. */
4397 tp->pred_flags = 0;
4398 inet_csk_schedule_ack(sk);
4399
Eric Dumazetc10d9312016-04-29 14:16:47 -07004400 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004401 seq = TCP_SKB_CB(skb)->seq;
4402 end_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004403 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004404 tp->rcv_nxt, seq, end_seq);
Eric Dumazete86b29192012-03-18 11:06:44 +00004405
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004406 p = &tp->out_of_order_queue.rb_node;
4407 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004408 /* Initial out of order segment, build 1 SACK. */
4409 if (tcp_is_sack(tp)) {
4410 tp->rx_opt.num_sacks = 1;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004411 tp->selective_acks[0].start_seq = seq;
4412 tp->selective_acks[0].end_seq = end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004413 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004414 rb_link_node(&skb->rbnode, NULL, p);
4415 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4416 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004417 goto end;
4418 }
4419
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004420 /* In the typical case, we are adding an skb to the end of the list.
4421 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4422 */
4423 if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4424coalesce_done:
4425 tcp_grow_window(sk, skb);
4426 kfree_skb_partial(skb, fragstolen);
4427 skb = NULL;
4428 goto add_sack;
Eric Dumazete86b29192012-03-18 11:06:44 +00004429 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004430 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4431 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4432 parent = &tp->ooo_last_skb->rbnode;
4433 p = &parent->rb_right;
4434 goto insert;
Eric Dumazete86b29192012-03-18 11:06:44 +00004435 }
4436
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004437 /* Find place to insert this segment. Handle overlaps on the way. */
4438 parent = NULL;
4439 while (*p) {
4440 parent = *p;
4441 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4442 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4443 p = &parent->rb_left;
4444 continue;
Eric Dumazete86b29192012-03-18 11:06:44 +00004445 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004446 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4447 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4448 /* All the bits are present. Drop. */
4449 NET_INC_STATS(sock_net(sk),
4450 LINUX_MIB_TCPOFOMERGE);
4451 __kfree_skb(skb);
4452 skb = NULL;
4453 tcp_dsack_set(sk, seq, end_seq);
4454 goto add_sack;
4455 }
4456 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4457 /* Partial overlap. */
4458 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4459 } else {
4460 /* skb's seq == skb1's seq and skb covers skb1.
4461 * Replace skb1 with skb.
4462 */
4463 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4464 &tp->out_of_order_queue);
4465 tcp_dsack_extend(sk,
4466 TCP_SKB_CB(skb1)->seq,
4467 TCP_SKB_CB(skb1)->end_seq);
4468 NET_INC_STATS(sock_net(sk),
4469 LINUX_MIB_TCPOFOMERGE);
4470 __kfree_skb(skb1);
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004471 goto merge_right;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004472 }
4473 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4474 goto coalesce_done;
4475 }
4476 p = &parent->rb_right;
Eric Dumazete86b29192012-03-18 11:06:44 +00004477 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004478insert:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004479 /* Insert segment into RB tree. */
4480 rb_link_node(&skb->rbnode, parent, p);
4481 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004482
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004483merge_right:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004484 /* Remove other segments covered by skb. */
4485 while ((q = rb_next(&skb->rbnode)) != NULL) {
4486 skb1 = rb_entry(q, struct sk_buff, rbnode);
Eric Dumazete86b29192012-03-18 11:06:44 +00004487
4488 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4489 break;
4490 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4491 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4492 end_seq);
4493 break;
4494 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004495 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004496 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4497 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004498 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004499 tcp_drop(sk, skb1);
Eric Dumazete86b29192012-03-18 11:06:44 +00004500 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004501 /* If there is no skb after us, we are the last_skb ! */
4502 if (!q)
4503 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004504
4505add_sack:
4506 if (tcp_is_sack(tp))
4507 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4508end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004509 if (skb) {
4510 tcp_grow_window(sk, skb);
Eric Dumazet60b1af32017-01-24 14:57:36 -08004511 skb_condense(skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004512 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004513 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004514}
4515
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004516static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004517 bool *fragstolen)
4518{
4519 int eaten;
4520 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4521
4522 __skb_pull(skb, hdrlen);
4523 eaten = (tail &&
4524 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004525 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004526 if (!eaten) {
4527 __skb_queue_tail(&sk->sk_receive_queue, skb);
4528 skb_set_owner_r(skb, sk);
4529 }
4530 return eaten;
4531}
Eric Dumazete86b29192012-03-18 11:06:44 +00004532
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004533int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4534{
Eric Dumazetcb934712014-09-17 03:14:42 -07004535 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004536 int err = -ENOMEM;
4537 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004538 bool fragstolen;
4539
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004540 if (size == 0)
4541 return 0;
4542
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004543 if (size > PAGE_SIZE) {
4544 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4545
4546 data_len = npages << PAGE_SHIFT;
4547 size = data_len + (size & ~PAGE_MASK);
4548 }
4549 skb = alloc_skb_with_frags(size - data_len, data_len,
4550 PAGE_ALLOC_COSTLY_ORDER,
4551 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004552 if (!skb)
4553 goto err;
4554
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004555 skb_put(skb, size - data_len);
4556 skb->data_len = data_len;
4557 skb->len = size;
4558
Eric Dumazetcb934712014-09-17 03:14:42 -07004559 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
Mel Gormanc76562b2012-07-31 16:44:41 -07004560 goto err_free;
4561
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004562 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4563 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004564 goto err_free;
4565
4566 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4567 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4568 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4569
Eric Dumazetcb934712014-09-17 03:14:42 -07004570 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004571 WARN_ON_ONCE(fragstolen); /* should not happen */
4572 __kfree_skb(skb);
4573 }
4574 return size;
4575
4576err_free:
4577 kfree_skb(skb);
4578err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004579 return err;
4580
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004581}
4582
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4584{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004585 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet329033f2012-04-27 00:38:33 +00004586 bool fragstolen = false;
Eric Dumazet532182c2016-04-01 08:52:19 -07004587 int eaten = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004588
Eric Dumazet532182c2016-04-01 08:52:19 -07004589 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4590 __kfree_skb(skb);
4591 return;
4592 }
Eric Dumazetf84af322010-04-28 15:31:51 -07004593 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004594 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595
Florian Westphal735d3832014-09-29 13:08:30 +02004596 tcp_ecn_accept_cwr(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004597
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004598 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599
4600 /* Queue data for delivery to the user.
4601 * Packets in sequence go to the receive queue.
4602 * Out of sequence packets to the out_of_order_queue.
4603 */
4604 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4605 if (tcp_receive_window(tp) == 0)
4606 goto out_of_window;
4607
4608 /* Ok. In sequence. In window. */
4609 if (tp->ucopy.task == current &&
4610 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4611 sock_owned_by_user(sk) && !tp->urg_data) {
4612 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004613 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614
4615 __set_current_state(TASK_RUNNING);
4616
Al Virof4362a22014-11-24 13:26:06 -05004617 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004618 tp->ucopy.len -= chunk;
4619 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004620 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004621 tcp_rcv_space_adjust(sk);
4622 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004623 }
4624
4625 if (eaten <= 0) {
4626queue_and_out:
Eric Dumazet76dfa602015-05-15 12:39:29 -07004627 if (eaten < 0) {
4628 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4629 sk_forced_mem_schedule(sk, skb->truesize);
4630 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4631 goto drop;
4632 }
Eric Dumazetb081f852012-05-02 09:58:29 +00004633 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004634 }
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004635 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004636 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004637 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004638 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004639 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004640
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004641 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004642 tcp_ofo_queue(sk);
4643
4644 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4645 * gap in queue is filled.
4646 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004647 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004648 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004649 }
4650
4651 if (tp->rx_opt.num_sacks)
4652 tcp_sack_remove(tp);
4653
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004654 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004655
Eric Dumazet923dd342012-05-02 07:55:58 +00004656 if (eaten > 0)
4657 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004658 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04004659 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004660 return;
4661 }
4662
4663 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4664 /* A retransmit, 2nd most common case. Force an immediate ack. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004665 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004666 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004667
4668out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004669 tcp_enter_quickack_mode(sk);
4670 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671drop:
Eric Dumazet532182c2016-04-01 08:52:19 -07004672 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004673 return;
4674 }
4675
4676 /* Out of window. F.e. zero window probe. */
4677 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4678 goto out_of_window;
4679
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004680 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004681
4682 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4683 /* Partial packet, seq < rcv_next < end_seq */
4684 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4685 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4686 TCP_SKB_CB(skb)->end_seq);
4687
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004688 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004689
Linus Torvalds1da177e2005-04-16 15:20:36 -07004690 /* If window is closed, drop tail of packet. But after
4691 * remembering D-SACK for its head made in previous line.
4692 */
4693 if (!tcp_receive_window(tp))
4694 goto out_of_window;
4695 goto queue_and_out;
4696 }
4697
Eric Dumazete86b29192012-03-18 11:06:44 +00004698 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004699}
4700
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004701static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004702{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004703 if (list)
4704 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
David S. Miller91521942009-05-28 21:35:47 -07004705
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004706 return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4707}
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004708
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004709static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4710 struct sk_buff_head *list,
4711 struct rb_root *root)
4712{
4713 struct sk_buff *next = tcp_skb_next(skb, list);
4714
4715 if (list)
4716 __skb_unlink(skb, list);
4717 else
4718 rb_erase(&skb->rbnode, root);
4719
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004720 __kfree_skb(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004721 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004722
4723 return next;
4724}
4725
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004726/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4727static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4728{
4729 struct rb_node **p = &root->rb_node;
4730 struct rb_node *parent = NULL;
4731 struct sk_buff *skb1;
4732
4733 while (*p) {
4734 parent = *p;
4735 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4736 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4737 p = &parent->rb_left;
4738 else
4739 p = &parent->rb_right;
4740 }
4741 rb_link_node(&skb->rbnode, parent, p);
4742 rb_insert_color(&skb->rbnode, root);
4743}
4744
Linus Torvalds1da177e2005-04-16 15:20:36 -07004745/* Collapse contiguous sequence of skbs head..tail with
4746 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004747 *
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004748 * If tail is NULL, this means until the end of the queue.
David S. Miller91521942009-05-28 21:35:47 -07004749 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004750 * Segments with FIN/SYN are not collapsed (only because this
4751 * simplifies code)
4752 */
4753static void
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004754tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4755 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004757 struct sk_buff *skb = head, *n;
4758 struct sk_buff_head tmp;
David S. Miller91521942009-05-28 21:35:47 -07004759 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004760
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004761 /* First, check that queue is collapsible and find
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004762 * the point where collapsing can be useful.
4763 */
David S. Miller91521942009-05-28 21:35:47 -07004764restart:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004765 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4766 n = tcp_skb_next(skb, list);
4767
Linus Torvalds1da177e2005-04-16 15:20:36 -07004768 /* No new bits? It is possible on ofo queue. */
4769 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004770 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004771 if (!skb)
4772 break;
4773 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004774 }
4775
4776 /* The first skb to collapse is:
4777 * - not SYN/FIN and
4778 * - bloated or contains data before "start" or
4779 * overlaps to the next one.
4780 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004781 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004782 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004783 before(TCP_SKB_CB(skb)->seq, start))) {
4784 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004785 break;
David S. Miller91521942009-05-28 21:35:47 -07004786 }
4787
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004788 if (n && n != tail &&
4789 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4790 end_of_skbs = false;
4791 break;
David S. Miller91521942009-05-28 21:35:47 -07004792 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004793
4794 /* Decided to skip this, advance start seq. */
4795 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004796 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004797 if (end_of_skbs ||
4798 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004799 return;
4800
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004801 __skb_queue_head_init(&tmp);
4802
Linus Torvalds1da177e2005-04-16 15:20:36 -07004803 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004804 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004805 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004806
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004807 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004808 if (!nskb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004809 break;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004810
Linus Torvalds1da177e2005-04-16 15:20:36 -07004811 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4812 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004813 if (list)
4814 __skb_queue_before(list, skb, nskb);
4815 else
4816 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004817 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004818
4819 /* Copy data, releasing collapsed skbs. */
4820 while (copy > 0) {
4821 int offset = start - TCP_SKB_CB(skb)->seq;
4822 int size = TCP_SKB_CB(skb)->end_seq - start;
4823
Kris Katterjohn09a62662006-01-08 22:24:28 -08004824 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825 if (size > 0) {
4826 size = min(copy, size);
4827 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4828 BUG();
4829 TCP_SKB_CB(nskb)->end_seq += size;
4830 copy -= size;
4831 start += size;
4832 }
4833 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004834 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07004835 if (!skb ||
4836 skb == tail ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004837 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004838 goto end;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004839 }
4840 }
4841 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004842end:
4843 skb_queue_walk_safe(&tmp, skb, n)
4844 tcp_rbtree_insert(root, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004845}
4846
4847/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4848 * and tcp_collapse() them until all the queue is collapsed.
4849 */
4850static void tcp_collapse_ofo_queue(struct sock *sk)
4851{
4852 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004853 struct sk_buff *skb, *head;
4854 struct rb_node *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004855 u32 start, end;
4856
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004857 p = rb_first(&tp->out_of_order_queue);
4858 skb = rb_entry_safe(p, struct sk_buff, rbnode);
4859new_range:
4860 if (!skb) {
4861 p = rb_last(&tp->out_of_order_queue);
4862 /* Note: This is possible p is NULL here. We do not
4863 * use rb_entry_safe(), as ooo_last_skb is valid only
4864 * if rbtree is not empty.
4865 */
4866 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004867 return;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004868 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004869 start = TCP_SKB_CB(skb)->seq;
4870 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004871
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004872 for (head = skb;;) {
4873 skb = tcp_skb_next(skb, NULL);
David S. Miller91521942009-05-28 21:35:47 -07004874
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004875 /* Range is terminated when we see a gap or when
4876 * we are at the queue end.
4877 */
David S. Miller91521942009-05-28 21:35:47 -07004878 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004879 after(TCP_SKB_CB(skb)->seq, end) ||
4880 before(TCP_SKB_CB(skb)->end_seq, start)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004881 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
David S. Miller8728b832005-08-09 19:25:21 -07004882 head, skb, start, end);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004883 goto new_range;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004885
4886 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4887 start = TCP_SKB_CB(skb)->seq;
4888 if (after(TCP_SKB_CB(skb)->end_seq, end))
4889 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004890 }
4891}
4892
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004893/*
Eric Dumazet36a65032016-08-17 14:17:09 -07004894 * Clean the out-of-order queue to make room.
4895 * We drop high sequences packets to :
4896 * 1) Let a chance for holes to be filled.
4897 * 2) not add too big latencies if thousands of packets sit there.
4898 * (But if application shrinks SO_RCVBUF, we could still end up
4899 * freeing whole queue here)
4900 *
4901 * Return true if queue has shrunk.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004902 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004903static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004904{
4905 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004906 struct rb_node *node, *prev;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004907
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004908 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Eric Dumazet36a65032016-08-17 14:17:09 -07004909 return false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004910
Eric Dumazet36a65032016-08-17 14:17:09 -07004911 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004912 node = &tp->ooo_last_skb->rbnode;
4913 do {
4914 prev = rb_prev(node);
4915 rb_erase(node, &tp->out_of_order_queue);
4916 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004917 sk_mem_reclaim(sk);
Eric Dumazet36a65032016-08-17 14:17:09 -07004918 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4919 !tcp_under_memory_pressure(sk))
4920 break;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004921 node = prev;
4922 } while (node);
4923 tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
Eric Dumazet36a65032016-08-17 14:17:09 -07004924
4925 /* Reset SACK state. A conforming SACK implementation will
4926 * do the same at a timeout based retransmit. When a connection
4927 * is in a sad state like this, we care only about integrity
4928 * of the connection not performance.
4929 */
4930 if (tp->rx_opt.sack_ok)
4931 tcp_sack_reset(&tp->rx_opt);
4932 return true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004933}
4934
Linus Torvalds1da177e2005-04-16 15:20:36 -07004935/* Reduce allocated memory if we can, trying to get
4936 * the socket within its memory limits again.
4937 *
4938 * Return less than zero if we should start dropping frames
4939 * until the socket owning process reads some of the data
4940 * to stabilize the situation.
4941 */
4942static int tcp_prune_queue(struct sock *sk)
4943{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004944 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004945
4946 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4947
Eric Dumazetc10d9312016-04-29 14:16:47 -07004948 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949
4950 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004951 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004952 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004953 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4954
4955 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004956 if (!skb_queue_empty(&sk->sk_receive_queue))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004957 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
David S. Miller91521942009-05-28 21:35:47 -07004958 skb_peek(&sk->sk_receive_queue),
4959 NULL,
4960 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004961 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962
4963 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4964 return 0;
4965
4966 /* Collapsing did not help, destructive actions follow.
4967 * This must not ever occur. */
4968
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004969 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004970
4971 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4972 return 0;
4973
4974 /* If we are really being abused, tell the caller to silently
4975 * drop receive data on the floor. It will get retransmitted
4976 * and hopefully then we'll have sufficient space.
4977 */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004978 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004979
4980 /* Massive buffer overcommit. */
4981 tp->pred_flags = 0;
4982 return -1;
4983}
4984
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004985static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004986{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004987 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004988
David S. Miller0d9901d2005-07-05 15:21:10 -07004989 /* If the user specified a specific send buffer setting, do
4990 * not modify it.
4991 */
4992 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004993 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004994
4995 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004996 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004997 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004998
4999 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00005000 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005001 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005002
5003 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05005004 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005005 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005006
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005007 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07005008}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005009
5010/* When incoming ACK allowed to free some skb from write_queue,
5011 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5012 * on the exit from tcp input handler.
5013 *
5014 * PROBLEM: sndbuf expansion does not work well with largesend.
5015 */
5016static void tcp_new_space(struct sock *sk)
5017{
5018 struct tcp_sock *tp = tcp_sk(sk);
5019
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005020 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07005021 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005022 tp->snd_cwnd_stamp = tcp_time_stamp;
5023 }
5024
5025 sk->sk_write_space(sk);
5026}
5027
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005028static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005029{
5030 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5031 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
jbaron@akamai.com3c715122015-04-20 20:05:07 +00005032 /* pairs with tcp_poll() */
Jason Baron56d80622017-01-24 21:49:41 -05005033 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005034 if (sk->sk_socket &&
Francis Yanb0f71bd2016-11-27 23:07:16 -08005035 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005036 tcp_new_space(sk);
Francis Yanb0f71bd2016-11-27 23:07:16 -08005037 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5038 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5039 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005040 }
5041}
5042
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005043static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005044{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005045 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005046 tcp_check_space(sk);
5047}
5048
5049/*
5050 * Check if sending an ack is needed.
5051 */
5052static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5053{
5054 struct tcp_sock *tp = tcp_sk(sk);
5055
5056 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005057 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005058 /* ... and right edge of window advances far enough.
5059 * (tcp_recvmsg() will send ACK otherwise). Or...
5060 */
Joe Perches9d4fb272009-11-23 10:41:23 -08005061 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005062 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005063 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005064 /* We have out of order data. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005065 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066 /* Then ack it now */
5067 tcp_send_ack(sk);
5068 } else {
5069 /* Else, send delayed ack. */
5070 tcp_send_delayed_ack(sk);
5071 }
5072}
5073
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005074static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005075{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005076 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077 /* We sent a data segment already. */
5078 return;
5079 }
5080 __tcp_ack_snd_check(sk, 1);
5081}
5082
5083/*
5084 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005085 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005086 * moved inline now as tcp_urg is only called from one
5087 * place. We handle URGent data wrong. We have to - as
5088 * BSD still doesn't use the correction from RFC961.
5089 * For 1003.1g we should support a new option TCP_STDURG to permit
5090 * either form (or just set the sysctl tcp_stdurg).
5091 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005092
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005093static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094{
5095 struct tcp_sock *tp = tcp_sk(sk);
5096 u32 ptr = ntohs(th->urg_ptr);
5097
5098 if (ptr && !sysctl_tcp_stdurg)
5099 ptr--;
5100 ptr += ntohl(th->seq);
5101
5102 /* Ignore urgent data that we've already seen and read. */
5103 if (after(tp->copied_seq, ptr))
5104 return;
5105
5106 /* Do not replay urg ptr.
5107 *
5108 * NOTE: interesting situation not covered by specs.
5109 * Misbehaving sender may send urg ptr, pointing to segment,
5110 * which we already have in ofo queue. We are not able to fetch
5111 * such data and will stay in TCP_URG_NOTYET until will be eaten
5112 * by recvmsg(). Seems, we are not obliged to handle such wicked
5113 * situations. But it is worth to think about possibility of some
5114 * DoSes using some hypothetical application level deadlock.
5115 */
5116 if (before(ptr, tp->rcv_nxt))
5117 return;
5118
5119 /* Do we already have a newer (or duplicate) urgent pointer? */
5120 if (tp->urg_data && !after(ptr, tp->urg_seq))
5121 return;
5122
5123 /* Tell the world about our new urgent pointer. */
5124 sk_send_sigurg(sk);
5125
5126 /* We may be adding urgent data when the last byte read was
5127 * urgent. To do this requires some care. We cannot just ignore
5128 * tp->copied_seq since we would read the last urgent byte again
5129 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005130 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131 *
5132 * NOTE. Double Dutch. Rendering to plain English: author of comment
5133 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5134 * and expect that both A and B disappear from stream. This is _wrong_.
5135 * Though this happens in BSD with high probability, this is occasional.
5136 * Any application relying on this is buggy. Note also, that fix "works"
5137 * only in this artificial test. Insert some normal data between A and B and we will
5138 * decline of BSD again. Verdict: it is better to remove to trap
5139 * buggy users.
5140 */
5141 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005142 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005143 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5144 tp->copied_seq++;
5145 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005146 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147 __kfree_skb(skb);
5148 }
5149 }
5150
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005151 tp->urg_data = TCP_URG_NOTYET;
5152 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005153
5154 /* Disable header prediction. */
5155 tp->pred_flags = 0;
5156}
5157
5158/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005159static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005160{
5161 struct tcp_sock *tp = tcp_sk(sk);
5162
5163 /* Check if we get a new urgent pointer - normally not. */
5164 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005165 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005166
5167 /* Do we wait for any urgent data? - normally not... */
5168 if (tp->urg_data == TCP_URG_NOTYET) {
5169 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5170 th->syn;
5171
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005172 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005173 if (ptr < skb->len) {
5174 u8 tmp;
5175 if (skb_copy_bits(skb, ptr, &tmp, 1))
5176 BUG();
5177 tp->urg_data = TCP_URG_VALID | tmp;
5178 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005179 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005180 }
5181 }
5182}
5183
5184static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5185{
5186 struct tcp_sock *tp = tcp_sk(sk);
5187 int chunk = skb->len - hlen;
5188 int err;
5189
Herbert Xu60476372007-04-09 11:59:39 -07005190 if (skb_csum_unnecessary(skb))
Al Virof4362a22014-11-24 13:26:06 -05005191 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005192 else
Al Virof4362a22014-11-24 13:26:06 -05005193 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194
5195 if (!err) {
5196 tp->ucopy.len -= chunk;
5197 tp->copied_seq += chunk;
5198 tcp_rcv_space_adjust(sk);
5199 }
5200
Linus Torvalds1da177e2005-04-16 15:20:36 -07005201 return err;
5202}
5203
Jason Baron0e40f4c2017-01-17 13:37:19 -05005204/* Accept RST for rcv_nxt - 1 after a FIN.
5205 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5206 * FIN is sent followed by a RST packet. The RST is sent with the same
5207 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5208 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5209 * ACKs on the closed socket. In addition middleboxes can drop either the
5210 * challenge ACK or a subsequent RST.
5211 */
5212static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5213{
5214 struct tcp_sock *tp = tcp_sk(sk);
5215
5216 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5217 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5218 TCPF_CLOSING));
5219}
5220
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005221/* Does PAWS and seqno based validation of an incoming segment, flags will
5222 * play significant role here.
5223 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005224static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5225 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005226{
5227 struct tcp_sock *tp = tcp_sk(sk);
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005228 bool rst_seq_match = false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005229
5230 /* RFC1323: H1. Apply PAWS check first. */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005231 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005232 tcp_paws_discard(sk, skb)) {
5233 if (!th->rst) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005234 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005235 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5236 LINUX_MIB_TCPACKSKIPPEDPAWS,
5237 &tp->last_oow_ack_time))
5238 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005239 goto discard;
5240 }
5241 /* Reset is accepted even if it did not pass PAWS. */
5242 }
5243
5244 /* Step 1: check sequence number */
5245 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5246 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5247 * (RST) segments are validated by checking their SEQ-fields."
5248 * And page 69: "If an incoming segment is not acceptable,
5249 * an acknowledgment should be sent in reply (unless the RST
5250 * bit is set, if so drop the segment and return)".
5251 */
Eric Dumazete3715892012-07-17 12:29:30 +00005252 if (!th->rst) {
5253 if (th->syn)
5254 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005255 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5256 LINUX_MIB_TCPACKSKIPPEDSEQ,
5257 &tp->last_oow_ack_time))
5258 tcp_send_dupack(sk, skb);
Jason Baron0e40f4c2017-01-17 13:37:19 -05005259 } else if (tcp_reset_check(sk, skb)) {
5260 tcp_reset(sk);
Eric Dumazete3715892012-07-17 12:29:30 +00005261 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005262 goto discard;
5263 }
5264
5265 /* Step 2: check RST bit */
5266 if (th->rst) {
Jason Baron0e40f4c2017-01-17 13:37:19 -05005267 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5268 * FIN and SACK too if available):
5269 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5270 * the right-most SACK block,
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005271 * then
Eric Dumazet282f23c2012-07-17 10:13:05 +02005272 * RESET the connection
5273 * else
5274 * Send a challenge ACK
5275 */
Jason Baron0e40f4c2017-01-17 13:37:19 -05005276 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5277 tcp_reset_check(sk, skb)) {
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005278 rst_seq_match = true;
5279 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5280 struct tcp_sack_block *sp = &tp->selective_acks[0];
5281 int max_sack = sp[0].end_seq;
5282 int this_sack;
5283
5284 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5285 ++this_sack) {
5286 max_sack = after(sp[this_sack].end_seq,
5287 max_sack) ?
5288 sp[this_sack].end_seq : max_sack;
5289 }
5290
5291 if (TCP_SKB_CB(skb)->seq == max_sack)
5292 rst_seq_match = true;
5293 }
5294
5295 if (rst_seq_match)
Eric Dumazet282f23c2012-07-17 10:13:05 +02005296 tcp_reset(sk);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005297 else {
5298 /* Disable TFO if RST is out-of-order
5299 * and no data has been received
5300 * for current active TFO socket
5301 */
5302 if (tp->syn_fastopen && !tp->data_segs_in &&
5303 sk->sk_state == TCP_ESTABLISHED)
Wei Wang46c2fa32017-04-20 14:45:47 -07005304 tcp_fastopen_active_disable(sk);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005305 tcp_send_challenge_ack(sk, skb);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005306 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005307 goto discard;
5308 }
5309
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005310 /* step 3: check security and precedence [ignored] */
5311
Eric Dumazet0c246042012-07-17 01:41:30 +00005312 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005313 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005314 */
5315 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005316syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005317 if (syn_inerr)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005318 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5319 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005320 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005321 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005322 }
5323
Eric Dumazet0c246042012-07-17 01:41:30 +00005324 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005325
5326discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005327 tcp_drop(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005328 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005329}
5330
Linus Torvalds1da177e2005-04-16 15:20:36 -07005331/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005332 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005334 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005335 * disabled when:
5336 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005337 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338 * - Out of order segments arrived.
5339 * - Urgent data is expected.
5340 * - There is no buffer space left
5341 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005342 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005343 * - Data is sent in both directions. Fast path only supports pure senders
5344 * or pure receivers (this means either the sequence number or the ack
5345 * value must stay constant)
5346 * - Unexpected TCP option.
5347 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005348 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005349 * receive procedure patterned after RFC793 to handle all cases.
5350 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005351 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005352 * tcp_data_queue when everything is OK.
5353 */
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005354void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5355 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356{
5357 struct tcp_sock *tp = tcp_sk(sk);
5358
Eric Dumazet69e996c2017-04-25 10:15:32 -07005359 skb_mstamp_get(&tp->tcp_mstamp);
Ian Morris51456b22015-04-03 09:17:26 +01005360 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005361 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362 /*
5363 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005364 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005365 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005366 *
5367 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005368 * on a device interrupt, to call tcp_recv function
5369 * on the receive process context and checksum and copy
5370 * the buffer to user space. smart...
5371 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005372 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005373 * extra cost of the net_bh soft interrupt processing...
5374 * We do checksum and copy also but from device to kernel.
5375 */
5376
5377 tp->rx_opt.saw_tstamp = 0;
5378
5379 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005380 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005381 * 'S' will always be tp->tcp_header_len >> 2
5382 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005383 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005384 * space for instance)
5385 * PSH flag is ignored.
5386 */
5387
5388 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005389 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5390 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005391 int tcp_header_len = tp->tcp_header_len;
5392
5393 /* Timestamp header prediction: tcp_header_len
5394 * is automatically equal to th->doff*4 due to pred_flags
5395 * match.
5396 */
5397
5398 /* Check timestamp */
5399 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005400 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005401 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005402 goto slow_path;
5403
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404 /* If PAWS failed, check it more carefully in slow path */
5405 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5406 goto slow_path;
5407
5408 /* DO NOT update ts_recent here, if checksum fails
5409 * and timestamp was corrupted part, it will result
5410 * in a hung connection since we will drop all
5411 * future packets due to the PAWS test.
5412 */
5413 }
5414
5415 if (len <= tcp_header_len) {
5416 /* Bulk data transfer: sender */
5417 if (len == tcp_header_len) {
5418 /* Predicted packet is in window by definition.
5419 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5420 * Hence, check seq<=rcv_wup reduces to:
5421 */
5422 if (tcp_header_len ==
5423 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5424 tp->rcv_nxt == tp->rcv_wup)
5425 tcp_store_ts_recent(tp);
5426
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427 /* We know that such packets are checksummed
5428 * on entry.
5429 */
5430 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005431 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005432 tcp_data_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005433 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005434 } else { /* Header too small */
Eric Dumazetc10d9312016-04-29 14:16:47 -07005435 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436 goto discard;
5437 }
5438 } else {
5439 int eaten = 0;
Eric Dumazetb081f852012-05-02 09:58:29 +00005440 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005441
Dan Williamsd27f9bc12013-12-30 11:37:15 -08005442 if (tp->ucopy.task == current &&
5443 tp->copied_seq == tp->rcv_nxt &&
5444 len - tcp_header_len <= tp->ucopy.len &&
5445 sock_owned_by_user(sk)) {
5446 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005447
Dan Williamsd27f9bc12013-12-30 11:37:15 -08005448 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005449 /* Predicted packet is in window by definition.
5450 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5451 * Hence, check seq<=rcv_wup reduces to:
5452 */
5453 if (tcp_header_len ==
5454 (sizeof(struct tcphdr) +
5455 TCPOLEN_TSTAMP_ALIGNED) &&
5456 tp->rcv_nxt == tp->rcv_wup)
5457 tcp_store_ts_recent(tp);
5458
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005459 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005460
5461 __skb_pull(skb, tcp_header_len);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07005462 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005463 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005464 LINUX_MIB_TCPHPHITSTOUSER);
Dan Williamsd27f9bc12013-12-30 11:37:15 -08005465 eaten = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005466 }
5467 }
5468 if (!eaten) {
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005469 if (tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005470 goto csum_error;
5471
Neal Cardwellaab2b4b2013-03-04 06:23:05 +00005472 if ((int)skb->truesize > sk->sk_forward_alloc)
5473 goto step5;
5474
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475 /* Predicted packet is in window by definition.
5476 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5477 * Hence, check seq<=rcv_wup reduces to:
5478 */
5479 if (tcp_header_len ==
5480 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5481 tp->rcv_nxt == tp->rcv_wup)
5482 tcp_store_ts_recent(tp);
5483
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005484 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005485
Eric Dumazetc10d9312016-04-29 14:16:47 -07005486 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005487
5488 /* Bulk data transfer: receiver */
Eric Dumazetb081f852012-05-02 09:58:29 +00005489 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5490 &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005491 }
5492
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005493 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005494
5495 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5496 /* Well, only one small jumplet in fast path... */
5497 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005498 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005499 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500 goto no_ack;
5501 }
5502
Dan Williamsd27f9bc12013-12-30 11:37:15 -08005503 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005504no_ack:
5505 if (eaten)
Eric Dumazetb081f852012-05-02 09:58:29 +00005506 kfree_skb_partial(skb, fragstolen);
David S. Miller676d2362014-04-11 16:15:36 -04005507 sk->sk_data_ready(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005508 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005509 }
5510 }
5511
5512slow_path:
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005513 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005514 goto csum_error;
5515
Calvin Owens0c228e82014-11-20 15:09:53 -08005516 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005517 goto discard;
5518
Linus Torvalds1da177e2005-04-16 15:20:36 -07005519 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005520 * Standard slow path.
5521 */
5522
Eric Dumazet0c246042012-07-17 01:41:30 +00005523 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005524 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525
5526step5:
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00005527 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005528 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005529
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005530 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005531
5532 /* Process urgent data. */
5533 tcp_urg(sk, skb, th);
5534
5535 /* step 7: process the segment text */
5536 tcp_data_queue(sk, skb);
5537
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005538 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005539 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005540 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005541
5542csum_error:
Eric Dumazetc10d9312016-04-29 14:16:47 -07005543 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5544 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005545
5546discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005547 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005548}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005549EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005550
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005551void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5552{
5553 struct tcp_sock *tp = tcp_sk(sk);
5554 struct inet_connection_sock *icsk = inet_csk(sk);
5555
5556 tcp_set_state(sk, TCP_ESTABLISHED);
Eric Dumazet15bb7742017-03-22 08:10:21 -07005557 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005558
Ian Morris00db4122015-04-03 09:17:27 +01005559 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005560 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005561 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005562 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005563
5564 /* Make sure socket is routed, for correct metrics. */
5565 icsk->icsk_af_ops->rebuild_header(sk);
5566
5567 tcp_init_metrics(sk);
5568
5569 tcp_init_congestion_control(sk);
5570
5571 /* Prevent spurious tcp_cwnd_restart() on first data
5572 * packet.
5573 */
5574 tp->lsndtime = tcp_time_stamp;
5575
5576 tcp_init_buffer_space(sk);
5577
5578 if (sock_flag(sk, SOCK_KEEPOPEN))
5579 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5580
5581 if (!tp->rx_opt.snd_wscale)
5582 __tcp_fast_path_on(tp, tp->snd_wnd);
5583 else
5584 tp->pred_flags = 0;
5585
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005586}
5587
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005588static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5589 struct tcp_fastopen_cookie *cookie)
5590{
5591 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005592 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07005593 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5594 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005595
5596 if (mss == tp->rx_opt.user_mss) {
5597 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005598
5599 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5600 tcp_clear_options(&opt);
5601 opt.user_mss = opt.mss_clamp = 0;
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005602 tcp_parse_options(synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005603 mss = opt.mss_clamp;
5604 }
5605
Daniel Lee2646c832015-04-06 14:37:27 -07005606 if (!tp->syn_fastopen) {
5607 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005608 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07005609 } else if (tp->total_retrans) {
5610 /* SYN timed out and the SYN-ACK neither has a cookie nor
5611 * acknowledges data. Presumably the remote received only
5612 * the retransmitted (regular) SYNs: either the original
5613 * SYN-data or the corresponding SYN-ACK was dropped.
5614 */
5615 syn_drop = (cookie->len < 0 && data);
5616 } else if (cookie->len < 0 && !tp->syn_data) {
5617 /* We requested a cookie but didn't get it. If we did not use
5618 * the (old) exp opt format then try so next time (try_exp=1).
5619 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5620 */
5621 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5622 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005623
Daniel Lee2646c832015-04-06 14:37:27 -07005624 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005625
5626 if (data) { /* Retransmit unacked data in SYN */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005627 tcp_for_write_queue_from(data, sk) {
5628 if (data == tcp_send_head(sk) ||
Eric Dumazet10d3be52016-04-21 10:55:23 -07005629 __tcp_retransmit_skb(sk, data, 1))
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005630 break;
5631 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005632 tcp_rearm_rto(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005633 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005634 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005635 return true;
5636 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005637 tp->syn_data_acked = tp->syn_data;
Yuchung Chengf19c29e2014-03-03 12:31:36 -08005638 if (tp->syn_data_acked)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005639 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005640 LINUX_MIB_TCPFASTOPENACTIVE);
Eric Dumazet61d2bca2016-02-01 21:03:07 -08005641
5642 tcp_fastopen_add_skb(sk, synack);
5643
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005644 return false;
5645}
5646
Linus Torvalds1da177e2005-04-16 15:20:36 -07005647static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07005648 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005649{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005650 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005651 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005652 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005653 int saved_clamp = tp->rx_opt.mss_clamp;
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005654 bool fastopen_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005655
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005656 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005657 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005658 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005659
5660 if (th->ack) {
5661 /* rfc793:
5662 * "If the state is SYN-SENT then
5663 * first check the ACK bit
5664 * If the ACK bit is set
5665 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5666 * a reset (unless the RST bit is set, if so drop
5667 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005668 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005669 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5670 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005671 goto reset_and_undo;
5672
5673 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5674 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5675 tcp_time_stamp)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005676 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005677 LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005678 goto reset_and_undo;
5679 }
5680
5681 /* Now ACK is acceptable.
5682 *
5683 * "If the RST bit is set
5684 * If the ACK was acceptable then signal the user "error:
5685 * connection reset", drop the segment, enter CLOSED state,
5686 * delete TCB, and return."
5687 */
5688
5689 if (th->rst) {
5690 tcp_reset(sk);
5691 goto discard;
5692 }
5693
5694 /* rfc793:
5695 * "fifth, if neither of the SYN or RST bits is set then
5696 * drop the segment and return."
5697 *
5698 * See note below!
5699 * --ANK(990513)
5700 */
5701 if (!th->syn)
5702 goto discard_and_undo;
5703
5704 /* rfc793:
5705 * "If the SYN bit is on ...
5706 * are acceptable then ...
5707 * (our SYN has been ACKed), change the connection
5708 * state to ESTABLISHED..."
5709 */
5710
Florian Westphal735d3832014-09-29 13:08:30 +02005711 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005713 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005714 tcp_ack(sk, skb, FLAG_SLOWPATH);
5715
5716 /* Ok.. it's good. Set up sequence numbers and
5717 * move to established.
5718 */
5719 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5720 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5721
5722 /* RFC1323: The window in SYN & SYN/ACK segments is
5723 * never scaled.
5724 */
5725 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005726
5727 if (!tp->rx_opt.wscale_ok) {
5728 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5729 tp->window_clamp = min(tp->window_clamp, 65535U);
5730 }
5731
5732 if (tp->rx_opt.saw_tstamp) {
5733 tp->rx_opt.tstamp_ok = 1;
5734 tp->tcp_header_len =
5735 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5736 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5737 tcp_store_ts_recent(tp);
5738 } else {
5739 tp->tcp_header_len = sizeof(struct tcphdr);
5740 }
5741
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005742 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5743 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005744
John Heffner5d424d52006-03-20 17:53:41 -08005745 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005746 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005747 tcp_initialize_rcv_mss(sk);
5748
5749 /* Remember, tcp_poll() does not lock socket!
5750 * Change state from SYN-SENT only after copied_seq
5751 * is initialized. */
5752 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005753
Ralf Baechlee16aa202006-12-07 00:11:33 -08005754 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005755
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005756 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005757
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005758 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5759 tcp_rcv_fastopen_synack(sk, skb, &foc);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005760
Eric Dumazet0f9fa832017-04-18 09:45:52 -07005761 if (!sock_flag(sk, SOCK_DEAD)) {
5762 sk->sk_state_change(sk);
5763 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5764 }
5765 if (fastopen_fail)
5766 return -1;
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005767 if (sk->sk_write_pending ||
5768 icsk->icsk_accept_queue.rskq_defer_accept ||
5769 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005770 /* Save one ACK. Data will be ready after
5771 * several ticks, if write_pending is set.
5772 *
5773 * It may be deleted, but with this feature tcpdumps
5774 * look so _wonderfully_ clever, that I was not able
5775 * to stand against the temptation 8) --ANK
5776 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005777 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005778 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005779 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5780 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005781
5782discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005783 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005784 return 0;
5785 } else {
5786 tcp_send_ack(sk);
5787 }
5788 return -1;
5789 }
5790
5791 /* No ACK in the segment */
5792
5793 if (th->rst) {
5794 /* rfc793:
5795 * "If the RST bit is set
5796 *
5797 * Otherwise (no ACK) drop the segment and return."
5798 */
5799
5800 goto discard_and_undo;
5801 }
5802
5803 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005804 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005805 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806 goto discard_and_undo;
5807
5808 if (th->syn) {
5809 /* We see SYN without ACK. It is attempt of
5810 * simultaneous connect with crossed SYNs.
5811 * Particularly, it can be connect to self.
5812 */
5813 tcp_set_state(sk, TCP_SYN_RECV);
5814
5815 if (tp->rx_opt.saw_tstamp) {
5816 tp->rx_opt.tstamp_ok = 1;
5817 tcp_store_ts_recent(tp);
5818 tp->tcp_header_len =
5819 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5820 } else {
5821 tp->tcp_header_len = sizeof(struct tcphdr);
5822 }
5823
5824 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet142a2e72015-11-26 08:18:14 -08005825 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005826 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5827
5828 /* RFC1323: The window in SYN & SYN/ACK segments is
5829 * never scaled.
5830 */
5831 tp->snd_wnd = ntohs(th->window);
5832 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5833 tp->max_window = tp->snd_wnd;
5834
Florian Westphal735d3832014-09-29 13:08:30 +02005835 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836
John Heffner5d424d52006-03-20 17:53:41 -08005837 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005838 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005839 tcp_initialize_rcv_mss(sk);
5840
Linus Torvalds1da177e2005-04-16 15:20:36 -07005841 tcp_send_synack(sk);
5842#if 0
5843 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005844 * There are no obstacles to make this (except that we must
5845 * either change tcp_recvmsg() to prevent it from returning data
5846 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005847 *
5848 * However, if we ignore data in ACKless segments sometimes,
5849 * we have no reasons to accept it sometimes.
5850 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5851 * is not flawless. So, discard packet for sanity.
5852 * Uncomment this return to process the data.
5853 */
5854 return -1;
5855#else
5856 goto discard;
5857#endif
5858 }
5859 /* "fifth, if neither of the SYN or RST bits is set then
5860 * drop the segment and return."
5861 */
5862
5863discard_and_undo:
5864 tcp_clear_options(&tp->rx_opt);
5865 tp->rx_opt.mss_clamp = saved_clamp;
5866 goto discard;
5867
5868reset_and_undo:
5869 tcp_clear_options(&tp->rx_opt);
5870 tp->rx_opt.mss_clamp = saved_clamp;
5871 return 1;
5872}
5873
Linus Torvalds1da177e2005-04-16 15:20:36 -07005874/*
5875 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005876 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005877 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5878 * address independent.
5879 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005880
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005881int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005882{
5883 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005884 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005885 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00005886 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005887 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005888 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005889
Linus Torvalds1da177e2005-04-16 15:20:36 -07005890 switch (sk->sk_state) {
5891 case TCP_CLOSE:
5892 goto discard;
5893
5894 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005895 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005896 return 1;
5897
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005898 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005899 goto discard;
5900
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005901 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005902 if (th->fin)
5903 goto discard;
Eric Dumazet449809a2017-03-01 08:39:49 -08005904 /* It is possible that we process SYN packets from backlog,
5905 * so we need to make sure to disable BH right there.
5906 */
5907 local_bh_disable();
5908 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5909 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005910
Eric Dumazet449809a2017-03-01 08:39:49 -08005911 if (!acceptable)
5912 return 1;
Eric Dumazet0aea76d2016-04-21 22:13:01 -07005913 consume_skb(skb);
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005914 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005915 }
5916 goto discard;
5917
5918 case TCP_SYN_SENT:
Eric Dumazet8804b272016-04-13 22:05:40 -07005919 tp->rx_opt.saw_tstamp = 0;
Eric Dumazet69e996c2017-04-25 10:15:32 -07005920 skb_mstamp_get(&tp->tcp_mstamp);
Eric Dumazetbda07a62015-09-29 07:42:40 -07005921 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005922 if (queued >= 0)
5923 return queued;
5924
5925 /* Do step6 onward by hand. */
5926 tcp_urg(sk, skb, th);
5927 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005928 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005929 return 0;
5930 }
5931
Eric Dumazet69e996c2017-04-25 10:15:32 -07005932 skb_mstamp_get(&tp->tcp_mstamp);
Eric Dumazet8804b272016-04-13 22:05:40 -07005933 tp->rx_opt.saw_tstamp = 0;
Jerry Chu168a8f52012-08-31 12:29:13 +00005934 req = tp->fastopen_rsk;
Ian Morris00db4122015-04-03 09:17:27 +01005935 if (req) {
Jerry Chu37561f62012-10-22 11:26:36 +00005936 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005937 sk->sk_state != TCP_FIN_WAIT1);
5938
Ian Morris51456b22015-04-03 09:17:26 +01005939 if (!tcp_check_req(sk, skb, req, true))
Jerry Chu168a8f52012-08-31 12:29:13 +00005940 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005941 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005942
Calvin Owens0c228e82014-11-20 15:09:53 -08005943 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005944 goto discard;
5945
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005946 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005947 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005948
5949 /* step 5: check the ACK field */
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005950 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5951 FLAG_UPDATE_TS_RECENT) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005952
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005953 switch (sk->sk_state) {
5954 case TCP_SYN_RECV:
Joe Perches61eb90032013-05-24 18:36:13 +00005955 if (!acceptable)
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005956 return 1;
Joe Perches61eb90032013-05-24 18:36:13 +00005957
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07005958 if (!tp->srtt_us)
5959 tcp_synack_rtt_meas(sk, req);
5960
Joe Perches61eb90032013-05-24 18:36:13 +00005961 /* Once we leave TCP_SYN_RECV, we no longer need req
5962 * so release it.
5963 */
5964 if (req) {
Yuchung Cheng7e32b442016-09-21 16:16:15 -07005965 inet_csk(sk)->icsk_retransmits = 0;
Joe Perches61eb90032013-05-24 18:36:13 +00005966 reqsk_fastopen_remove(sk, req, false);
5967 } else {
5968 /* Make sure socket is routed, for correct metrics. */
5969 icsk->icsk_af_ops->rebuild_header(sk);
5970 tcp_init_congestion_control(sk);
5971
5972 tcp_mtup_init(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00005973 tp->copied_seq = tp->rcv_nxt;
Eric Dumazetb0983d32013-09-20 13:56:58 -07005974 tcp_init_buffer_space(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005975 }
Joe Perches61eb90032013-05-24 18:36:13 +00005976 smp_mb();
5977 tcp_set_state(sk, TCP_ESTABLISHED);
5978 sk->sk_state_change(sk);
5979
5980 /* Note, that this wakeup is only for marginal crossed SYN case.
5981 * Passively open sockets are not waked up, because
5982 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5983 */
5984 if (sk->sk_socket)
5985 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5986
5987 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5988 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5989 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5990
5991 if (tp->rx_opt.tstamp_ok)
5992 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5993
5994 if (req) {
5995 /* Re-arm the timer because data may have been sent out.
5996 * This is similar to the regular data transmission case
5997 * when new data has just been ack'ed.
5998 *
5999 * (TFO) - we could try to be more aggressive and
6000 * retransmitting any data sooner based on when they
6001 * are sent out.
6002 */
6003 tcp_rearm_rto(sk);
6004 } else
6005 tcp_init_metrics(sk);
6006
Yuchung Chengc0402762016-09-19 23:39:21 -04006007 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6008 tcp_update_pacing_rate(sk);
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04006009
Joe Perches61eb90032013-05-24 18:36:13 +00006010 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6011 tp->lsndtime = tcp_time_stamp;
6012
6013 tcp_initialize_rcv_mss(sk);
6014 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006015 break;
6016
Joe Perchesc48b22d2013-05-24 18:06:58 +00006017 case TCP_FIN_WAIT1: {
Joe Perchesc48b22d2013-05-24 18:06:58 +00006018 int tmo;
6019
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006020 /* If we enter the TCP_FIN_WAIT1 state and we are a
6021 * Fast Open socket and this is the first acceptable
6022 * ACK we have received, this would have acknowledged
6023 * our SYNACK so stop the SYNACK timer.
6024 */
Ian Morris00db4122015-04-03 09:17:27 +01006025 if (req) {
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006026 /* Return RST if ack_seq is invalid.
6027 * Note that RFC793 only says to generate a
6028 * DUPACK for it but for TCP Fast Open it seems
6029 * better to treat this case like TCP_SYN_RECV
6030 * above.
6031 */
6032 if (!acceptable)
6033 return 1;
6034 /* We no longer need the request sock. */
6035 reqsk_fastopen_remove(sk, req, false);
6036 tcp_rearm_rto(sk);
6037 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00006038 if (tp->snd_una != tp->write_seq)
6039 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006040
Joe Perchesc48b22d2013-05-24 18:06:58 +00006041 tcp_set_state(sk, TCP_FIN_WAIT2);
6042 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006043
Julian Anastasovc3a2e832017-02-06 23:14:14 +02006044 sk_dst_confirm(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006045
Joe Perchesc48b22d2013-05-24 18:06:58 +00006046 if (!sock_flag(sk, SOCK_DEAD)) {
6047 /* Wake up lingering close() */
6048 sk->sk_state_change(sk);
6049 break;
6050 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006051
Wei Wangcf1ef3f2017-04-20 14:45:46 -07006052 if (tp->linger2 < 0) {
6053 tcp_done(sk);
6054 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6055 return 1;
6056 }
6057 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6058 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6059 /* Receive out of order FIN after close() */
6060 if (tp->syn_fastopen && th->fin)
Wei Wang46c2fa32017-04-20 14:45:47 -07006061 tcp_fastopen_active_disable(sk);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006062 tcp_done(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07006063 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006064 return 1;
6065 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006066
Joe Perchesc48b22d2013-05-24 18:06:58 +00006067 tmo = tcp_fin_time(sk);
6068 if (tmo > TCP_TIMEWAIT_LEN) {
6069 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6070 } else if (th->fin || sock_owned_by_user(sk)) {
6071 /* Bad case. We could lose such FIN otherwise.
6072 * It is not a big problem, but it looks confusing
6073 * and not so rare event. We still can lose it now,
6074 * if it spins in bh_lock_sock(), but it is really
6075 * marginal case.
6076 */
6077 inet_csk_reset_keepalive_timer(sk, tmo);
6078 } else {
6079 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6080 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006081 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006082 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00006083 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006084
6085 case TCP_CLOSING:
6086 if (tp->snd_una == tp->write_seq) {
6087 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6088 goto discard;
6089 }
6090 break;
6091
6092 case TCP_LAST_ACK:
6093 if (tp->snd_una == tp->write_seq) {
6094 tcp_update_metrics(sk);
6095 tcp_done(sk);
6096 goto discard;
6097 }
6098 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006099 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006100
6101 /* step 6: check the URG bit */
6102 tcp_urg(sk, skb, th);
6103
6104 /* step 7: process the segment text */
6105 switch (sk->sk_state) {
6106 case TCP_CLOSE_WAIT:
6107 case TCP_CLOSING:
6108 case TCP_LAST_ACK:
6109 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6110 break;
6111 case TCP_FIN_WAIT1:
6112 case TCP_FIN_WAIT2:
6113 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006114 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006115 * BSD 4.4 also does reset.
6116 */
6117 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6118 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6119 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006120 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006121 tcp_reset(sk);
6122 return 1;
6123 }
6124 }
6125 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006126 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006127 tcp_data_queue(sk, skb);
6128 queued = 1;
6129 break;
6130 }
6131
6132 /* tcp_data could move socket to TIME-WAIT */
6133 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006134 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006135 tcp_ack_snd_check(sk);
6136 }
6137
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006138 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006139discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006140 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006141 }
6142 return 0;
6143}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006144EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006145
6146static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6147{
6148 struct inet_request_sock *ireq = inet_rsk(req);
6149
6150 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006151 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6152 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006153#if IS_ENABLED(CONFIG_IPV6)
6154 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006155 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6156 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006157#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006158}
6159
Florian Westphald82bd122014-09-29 13:08:29 +02006160/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6161 *
6162 * If we receive a SYN packet with these bits set, it means a
6163 * network is playing bad games with TOS bits. In order to
6164 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006165 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006166 *
6167 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006168 * congestion control: Linux DCTCP asserts ECT on all packets,
6169 * including SYN, which is most optimal solution; however,
6170 * others, such as FreeBSD do not.
Florian Westphald82bd122014-09-29 13:08:29 +02006171 */
6172static void tcp_ecn_create_request(struct request_sock *req,
6173 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006174 const struct sock *listen_sk,
6175 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006176{
6177 const struct tcphdr *th = tcp_hdr(skb);
6178 const struct net *net = sock_net(listen_sk);
6179 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006180 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006181 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006182
6183 if (!th_ecn)
6184 return;
6185
6186 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006187 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6188 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006189
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006190 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6191 (ecn_ok_dst & DST_FEATURE_ECN_CA))
Florian Westphald82bd122014-09-29 13:08:29 +02006192 inet_rsk(req)->ecn_ok = 1;
6193}
6194
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006195static void tcp_openreq_init(struct request_sock *req,
6196 const struct tcp_options_received *rx_opt,
6197 struct sk_buff *skb, const struct sock *sk)
6198{
6199 struct inet_request_sock *ireq = inet_rsk(req);
6200
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006201 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006202 req->cookie_ts = 0;
6203 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6204 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07006205 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006206 tcp_rsk(req)->last_oow_ack_time = 0;
6207 req->mss = rx_opt->mss_clamp;
6208 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6209 ireq->tstamp_ok = rx_opt->tstamp_ok;
6210 ireq->sack_ok = rx_opt->sack_ok;
6211 ireq->snd_wscale = rx_opt->snd_wscale;
6212 ireq->wscale_ok = rx_opt->wscale_ok;
6213 ireq->acked = 0;
6214 ireq->ecn_ok = 0;
6215 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6216 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6217 ireq->ir_mark = inet_request_mark(sk, skb);
6218}
6219
Eric Dumazete49bb332015-03-17 18:32:27 -07006220struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006221 struct sock *sk_listener,
6222 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006223{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006224 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6225 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006226
6227 if (req) {
6228 struct inet_request_sock *ireq = inet_rsk(req);
6229
6230 kmemcheck_annotate_bitfield(ireq, flags);
6231 ireq->opt = NULL;
Huw Davies56ac42b2016-06-27 15:05:28 -04006232#if IS_ENABLED(CONFIG_IPV6)
6233 ireq->pktopts = NULL;
6234#endif
Eric Dumazete49bb332015-03-17 18:32:27 -07006235 atomic64_set(&ireq->ir_cookie, 0);
6236 ireq->ireq_state = TCP_NEW_SYN_RECV;
6237 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006238 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006239 }
6240
6241 return req;
6242}
6243EXPORT_SYMBOL(inet_reqsk_alloc);
6244
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006245/*
6246 * Return true if a syncookie should be sent
6247 */
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006248static bool tcp_syn_flood_action(const struct sock *sk,
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006249 const struct sk_buff *skb,
6250 const char *proto)
6251{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006252 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006253 const char *msg = "Dropping request";
6254 bool want_cookie = false;
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006255 struct net *net = sock_net(sk);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006256
6257#ifdef CONFIG_SYN_COOKIES
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006258 if (net->ipv4.sysctl_tcp_syncookies) {
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006259 msg = "Sending cookies";
6260 want_cookie = true;
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006261 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006262 } else
6263#endif
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006264 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006265
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006266 if (!queue->synflood_warned &&
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006267 net->ipv4.sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006268 xchg(&queue->synflood_warned, 1) == 0)
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006269 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6270 proto, ntohs(tcp_hdr(skb)->dest), msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006271
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006272 return want_cookie;
6273}
6274
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006275static void tcp_reqsk_record_syn(const struct sock *sk,
6276 struct request_sock *req,
6277 const struct sk_buff *skb)
6278{
6279 if (tcp_sk(sk)->save_syn) {
6280 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6281 u32 *copy;
6282
6283 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6284 if (copy) {
6285 copy[0] = len;
6286 memcpy(&copy[1], skb_network_header(skb), len);
6287 req->saved_syn = copy;
6288 }
6289 }
6290}
6291
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006292int tcp_conn_request(struct request_sock_ops *rsk_ops,
6293 const struct tcp_request_sock_ops *af_ops,
6294 struct sock *sk, struct sk_buff *skb)
6295{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006296 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006297 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6298 struct tcp_options_received tmp_opt;
6299 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006300 struct net *net = sock_net(sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006301 struct sock *fastopen_sk = NULL;
6302 struct dst_entry *dst = NULL;
6303 struct request_sock *req;
6304 bool want_cookie = false;
6305 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006306
6307 /* TW buckets are converted to open requests without
6308 * limitations, they conserve resources and peer is
6309 * evidently real one.
6310 */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006311 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006312 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6313 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6314 if (!want_cookie)
6315 goto drop;
6316 }
6317
Eric Dumazet5ea8ea22016-10-26 09:27:57 -07006318 if (sk_acceptq_is_full(sk)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006319 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006320 goto drop;
6321 }
6322
Eric Dumazeta1a53442015-10-04 21:08:11 -07006323 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006324 if (!req)
6325 goto drop;
6326
6327 tcp_rsk(req)->af_specific = af_ops;
Florian Westphal95a22ca2016-12-01 11:32:06 +01006328 tcp_rsk(req)->ts_off = 0;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006329
6330 tcp_clear_options(&tmp_opt);
6331 tmp_opt.mss_clamp = af_ops->mss_clamp;
6332 tmp_opt.user_mss = tp->rx_opt.user_mss;
6333 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6334
6335 if (want_cookie && !tmp_opt.saw_tstamp)
6336 tcp_clear_options(&tmp_opt);
6337
6338 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6339 tcp_openreq_init(req, &tmp_opt, skb, sk);
KOVACS Krisztian7a682572016-09-23 11:27:42 +02006340 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006341
Eric Dumazet16f86162015-03-13 15:51:10 -07006342 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
David Ahern6dd9a142015-12-16 13:20:44 -08006343 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
Eric Dumazet16f86162015-03-13 15:51:10 -07006344
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006345 af_ops->init_req(req, sk, skb);
6346
6347 if (security_inet_conn_request(sk, skb, req))
6348 goto drop_and_free;
6349
Eric Dumazet84b114b2017-05-05 06:56:54 -07006350 if (tmp_opt.tstamp_ok)
6351 tcp_rsk(req)->ts_off = af_ops->init_ts_off(skb);
Florian Westphal95a22ca2016-12-01 11:32:06 +01006352
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006353 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006354 /* Kill the following clause, if you dislike this way. */
Soheil Hassas Yeganeh4396e462017-03-15 16:30:46 -04006355 if (!net->ipv4.sysctl_tcp_syncookies &&
6356 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6357 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6358 !tcp_peer_is_proven(req, dst)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006359 /* Without syncookies last quarter of
6360 * backlog is filled with destinations,
6361 * proven to be alive.
6362 * It means that we continue to communicate
6363 * to destinations, already remembered
6364 * to the moment of synflood.
6365 */
6366 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6367 rsk_ops->family);
6368 goto drop_and_release;
6369 }
6370
Eric Dumazet84b114b2017-05-05 06:56:54 -07006371 isn = af_ops->init_seq(skb);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006372 }
6373 if (!dst) {
Soheil Hassas Yeganeh4396e462017-03-15 16:30:46 -04006374 dst = af_ops->route_req(sk, &fl, req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006375 if (!dst)
6376 goto drop_and_free;
6377 }
6378
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006379 tcp_ecn_create_request(req, skb, sk, dst);
6380
6381 if (want_cookie) {
6382 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6383 req->cookie_ts = tmp_opt.tstamp_ok;
6384 if (!tmp_opt.tstamp_ok)
6385 inet_rsk(req)->ecn_ok = 0;
6386 }
6387
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006388 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006389 tcp_rsk(req)->txhash = net_tx_rndhash();
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006390 tcp_openreq_init_rwin(req, sk, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006391 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006392 tcp_reqsk_record_syn(sk, req, skb);
Eric Dumazet7656d842015-10-04 21:08:07 -07006393 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006394 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006395 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006396 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Eric Dumazetb3d05142016-04-13 22:05:39 -07006397 &foc, TCP_SYNACK_FASTOPEN);
Eric Dumazet7656d842015-10-04 21:08:07 -07006398 /* Add the child socket directly into the accept queue */
6399 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6400 sk->sk_data_ready(sk);
6401 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006402 sock_put(fastopen_sk);
6403 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006404 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006405 if (!want_cookie)
6406 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
Eric Dumazetb3d05142016-04-13 22:05:39 -07006407 af_ops->send_synack(sk, dst, &fl, req, &foc,
6408 !want_cookie ? TCP_SYNACK_NORMAL :
6409 TCP_SYNACK_COOKIE);
Eric Dumazet9caad862016-04-01 08:52:20 -07006410 if (want_cookie) {
6411 reqsk_free(req);
6412 return 0;
6413 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006414 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006415 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006416 return 0;
6417
6418drop_and_release:
6419 dst_release(dst);
6420drop_and_free:
6421 reqsk_free(req);
6422drop:
Eric Dumazet9caad862016-04-01 08:52:20 -07006423 tcp_listendrop(sk);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006424 return 0;
6425}
6426EXPORT_SYMBOL(tcp_conn_request);