blob: fbc99c939fd85b4a2c41752e7d1ae901351ffe92 [file] [log] [blame]
Willy Tarreaue6b98942007-10-29 01:09:36 +01001/*
2 * AF_INET/AF_INET6 SOCK_STREAM protocol layer (tcp)
3 *
Willy Tarreaud4c33c82013-01-07 21:59:07 +01004 * Copyright 2000-2013 Willy Tarreau <w@1wt.eu>
Willy Tarreaue6b98942007-10-29 01:09:36 +01005 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
11 */
12
13#include <ctype.h>
14#include <errno.h>
15#include <fcntl.h>
16#include <stdio.h>
17#include <stdlib.h>
18#include <string.h>
19#include <time.h>
20
21#include <sys/param.h>
22#include <sys/socket.h>
Willy Tarreaue6b98942007-10-29 01:09:36 +010023#include <sys/types.h>
Willy Tarreaue6b98942007-10-29 01:09:36 +010024
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +040025#include <netinet/tcp.h>
Willy Tarreau2d392c22015-08-24 01:43:45 +020026#include <netinet/in.h>
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +040027
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020028#include <haproxy/api.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020029#include <haproxy/arg.h>
Willy Tarreau7ea393d2020-06-04 18:02:10 +020030#include <haproxy/connection.h>
Willy Tarreau8d366972020-05-27 16:10:29 +020031#include <haproxy/errors.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020032#include <haproxy/fd.h>
Willy Tarreauf268ee82020-06-04 17:05:57 +020033#include <haproxy/global.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020034#include <haproxy/list.h>
Willy Tarreau213e9902020-06-04 14:58:24 +020035#include <haproxy/listener.h>
Willy Tarreauaeed4a82020-06-04 22:01:04 +020036#include <haproxy/log.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020037#include <haproxy/namespace.h>
38#include <haproxy/port_range.h>
39#include <haproxy/proto_tcp.h>
40#include <haproxy/protocol.h>
Willy Tarreaua264d962020-06-04 22:29:18 +020041#include <haproxy/proxy-t.h>
Willy Tarreau18b7df72020-08-28 12:07:22 +020042#include <haproxy/sock.h>
Willy Tarreauf1725582020-08-28 15:30:11 +020043#include <haproxy/sock_inet.h>
Willy Tarreau48fbcae2020-06-03 18:09:46 +020044#include <haproxy/tools.h>
Willy Tarreaue6b98942007-10-29 01:09:36 +010045
Willy Tarreaue6b98942007-10-29 01:09:36 +010046
Emeric Bruncf20bf12010-10-22 16:06:11 +020047static int tcp_bind_listener(struct listener *listener, char *errmsg, int errlen);
Willy Tarreaucb66ea62020-09-25 17:12:32 +020048static int tcp_suspend_receiver(struct receiver *rx);
Willy Tarreau58e6b712020-09-25 19:40:31 +020049static int tcp_resume_receiver(struct receiver *rx);
Willy Tarreau5ddf1ce2020-09-25 19:27:39 +020050static void tcp_enable_listener(struct listener *listener);
51static void tcp_disable_listener(struct listener *listener);
Willy Tarreau9d5be5c2017-09-15 07:55:51 +020052static void tcpv4_add_listener(struct listener *listener, int port);
53static void tcpv6_add_listener(struct listener *listener, int port);
Willy Tarreaue6b98942007-10-29 01:09:36 +010054
55/* Note: must not be declared <const> as its list will be overwritten */
56static struct protocol proto_tcpv4 = {
57 .name = "tcpv4",
Willy Tarreaub0254cb2020-09-04 08:07:11 +020058 .fam = &proto_fam_inet4,
Willy Tarreaua54553f2020-09-16 17:50:45 +020059 .ctrl_type = SOCK_STREAM,
Willy Tarreaue6b98942007-10-29 01:09:36 +010060 .sock_domain = AF_INET,
61 .sock_type = SOCK_STREAM,
62 .sock_prot = IPPROTO_TCP,
Willy Tarreaucb66ea62020-09-25 17:12:32 +020063 .add = tcpv4_add_listener,
64 .listen = tcp_bind_listener,
Willy Tarreau5ddf1ce2020-09-25 19:27:39 +020065 .enable = tcp_enable_listener,
66 .disable = tcp_disable_listener,
Willy Tarreau7b2febd2020-10-09 17:18:29 +020067 .unbind = default_unbind_listener,
Willy Tarreaue03204c2020-10-09 17:02:21 +020068 .suspend = default_suspend_listener,
69 .resume = default_resume_listener,
Willy Tarreau686fa3d2020-09-25 19:09:53 +020070 .rx_enable = sock_enable,
71 .rx_disable = sock_disable,
Willy Tarreauf58b8db2020-10-09 16:32:08 +020072 .rx_unbind = sock_unbind,
Willy Tarreaucb66ea62020-09-25 17:12:32 +020073 .rx_suspend = tcp_suspend_receiver,
Willy Tarreau58e6b712020-09-25 19:40:31 +020074 .rx_resume = tcp_resume_receiver,
Willy Tarreau29185142020-10-13 17:26:00 +020075 .rx_listening = sock_accept_conn,
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020076 .accept = &listener_accept,
Willy Tarreau26d8c592012-05-07 18:12:14 +020077 .connect = tcp_connect_server,
Willy Tarreaud7f331c2020-09-25 17:01:43 +020078 .receivers = LIST_HEAD_INIT(proto_tcpv4.receivers),
79 .nb_receivers = 0,
Willy Tarreaue6b98942007-10-29 01:09:36 +010080};
81
Willy Tarreau0108d902018-11-25 19:14:37 +010082INITCALL1(STG_REGISTER, protocol_register, &proto_tcpv4);
83
Willy Tarreaue6b98942007-10-29 01:09:36 +010084/* Note: must not be declared <const> as its list will be overwritten */
85static struct protocol proto_tcpv6 = {
86 .name = "tcpv6",
Willy Tarreaub0254cb2020-09-04 08:07:11 +020087 .fam = &proto_fam_inet6,
Willy Tarreaua54553f2020-09-16 17:50:45 +020088 .ctrl_type = SOCK_STREAM,
Willy Tarreaue6b98942007-10-29 01:09:36 +010089 .sock_domain = AF_INET6,
90 .sock_type = SOCK_STREAM,
91 .sock_prot = IPPROTO_TCP,
Willy Tarreaucb66ea62020-09-25 17:12:32 +020092 .add = tcpv6_add_listener,
93 .listen = tcp_bind_listener,
Willy Tarreau5ddf1ce2020-09-25 19:27:39 +020094 .enable = tcp_enable_listener,
95 .disable = tcp_disable_listener,
Willy Tarreau7b2febd2020-10-09 17:18:29 +020096 .unbind = default_unbind_listener,
Willy Tarreaue03204c2020-10-09 17:02:21 +020097 .suspend = default_suspend_listener,
98 .resume = default_resume_listener,
Willy Tarreau686fa3d2020-09-25 19:09:53 +020099 .rx_enable = sock_enable,
100 .rx_disable = sock_disable,
Willy Tarreauf58b8db2020-10-09 16:32:08 +0200101 .rx_unbind = sock_unbind,
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200102 .rx_suspend = tcp_suspend_receiver,
Willy Tarreau58e6b712020-09-25 19:40:31 +0200103 .rx_resume = tcp_resume_receiver,
Willy Tarreau29185142020-10-13 17:26:00 +0200104 .rx_listening = sock_accept_conn,
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200105 .accept = &listener_accept,
Willy Tarreau26d8c592012-05-07 18:12:14 +0200106 .connect = tcp_connect_server,
Willy Tarreaud7f331c2020-09-25 17:01:43 +0200107 .receivers = LIST_HEAD_INIT(proto_tcpv6.receivers),
108 .nb_receivers = 0,
Willy Tarreaue6b98942007-10-29 01:09:36 +0100109};
110
Willy Tarreau0108d902018-11-25 19:14:37 +0100111INITCALL1(STG_REGISTER, protocol_register, &proto_tcpv6);
112
David du Colombier6f5ccb12011-03-10 22:26:24 +0100113/* Binds ipv4/ipv6 address <local> to socket <fd>, unless <flags> is set, in which
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100114 * case we try to bind <remote>. <flags> is a 2-bit field consisting of :
115 * - 0 : ignore remote address (may even be a NULL pointer)
116 * - 1 : use provided address
117 * - 2 : use provided port
118 * - 3 : use both
119 *
120 * The function supports multiple foreign binding methods :
121 * - linux_tproxy: we directly bind to the foreign address
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100122 * The second one can be used as a fallback for the first one.
123 * This function returns 0 when everything's OK, 1 if it could not bind, to the
124 * local address, 2 if it could not bind to the foreign address.
125 */
David du Colombier6f5ccb12011-03-10 22:26:24 +0100126int tcp_bind_socket(int fd, int flags, struct sockaddr_storage *local, struct sockaddr_storage *remote)
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100127{
David du Colombier6f5ccb12011-03-10 22:26:24 +0100128 struct sockaddr_storage bind_addr;
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100129 int foreign_ok = 0;
130 int ret;
Christopher Faulet1bc04c72017-10-29 20:14:08 +0100131 static THREAD_LOCAL int ip_transp_working = 1;
132 static THREAD_LOCAL int ip6_transp_working = 1;
Pieter Baauwd551fb52013-05-08 22:49:23 +0200133
David du Colombier65c17962012-07-13 14:34:59 +0200134 switch (local->ss_family) {
135 case AF_INET:
136 if (flags && ip_transp_working) {
Pieter Baauwd551fb52013-05-08 22:49:23 +0200137 /* This deserves some explanation. Some platforms will support
138 * multiple combinations of certain methods, so we try the
139 * supported ones until one succeeds.
140 */
Willy Tarreau37bafdc2020-08-28 17:23:40 +0200141 if (sock_inet4_make_foreign(fd))
David du Colombier65c17962012-07-13 14:34:59 +0200142 foreign_ok = 1;
143 else
144 ip_transp_working = 0;
145 }
146 break;
147 case AF_INET6:
148 if (flags && ip6_transp_working) {
Willy Tarreau37bafdc2020-08-28 17:23:40 +0200149 if (sock_inet6_make_foreign(fd))
David du Colombier65c17962012-07-13 14:34:59 +0200150 foreign_ok = 1;
151 else
152 ip6_transp_working = 0;
153 }
154 break;
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100155 }
Pieter Baauwd551fb52013-05-08 22:49:23 +0200156
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100157 if (flags) {
158 memset(&bind_addr, 0, sizeof(bind_addr));
Willy Tarreau96dd0792011-04-19 07:20:57 +0200159 bind_addr.ss_family = remote->ss_family;
David du Colombier6f5ccb12011-03-10 22:26:24 +0100160 switch (remote->ss_family) {
161 case AF_INET:
162 if (flags & 1)
163 ((struct sockaddr_in *)&bind_addr)->sin_addr = ((struct sockaddr_in *)remote)->sin_addr;
164 if (flags & 2)
165 ((struct sockaddr_in *)&bind_addr)->sin_port = ((struct sockaddr_in *)remote)->sin_port;
166 break;
167 case AF_INET6:
168 if (flags & 1)
169 ((struct sockaddr_in6 *)&bind_addr)->sin6_addr = ((struct sockaddr_in6 *)remote)->sin6_addr;
170 if (flags & 2)
171 ((struct sockaddr_in6 *)&bind_addr)->sin6_port = ((struct sockaddr_in6 *)remote)->sin6_port;
172 break;
Willy Tarreau5dc1e982011-12-16 21:25:11 +0100173 default:
174 /* we don't want to try to bind to an unknown address family */
175 foreign_ok = 0;
David du Colombier6f5ccb12011-03-10 22:26:24 +0100176 }
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100177 }
178
Simon Hormande072bd2011-06-24 15:11:37 +0900179 setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &one, sizeof(one));
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100180 if (foreign_ok) {
Willy Tarreau9cf8d3f2014-05-09 22:56:10 +0200181 if (is_inet_addr(&bind_addr)) {
Willy Tarreau5f2877a2012-10-26 19:57:58 +0200182 ret = bind(fd, (struct sockaddr *)&bind_addr, get_addr_len(&bind_addr));
183 if (ret < 0)
184 return 2;
185 }
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100186 }
187 else {
Willy Tarreau9cf8d3f2014-05-09 22:56:10 +0200188 if (is_inet_addr(local)) {
Willy Tarreau5f2877a2012-10-26 19:57:58 +0200189 ret = bind(fd, (struct sockaddr *)local, get_addr_len(local));
190 if (ret < 0)
191 return 1;
192 }
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100193 }
194
195 if (!flags)
196 return 0;
197
Willy Tarreaue8c66af2008-01-13 18:40:14 +0100198 if (!foreign_ok)
199 /* we could not bind to a foreign address */
200 return 2;
201
202 return 0;
203}
Willy Tarreaue6b98942007-10-29 01:09:36 +0100204
Willy Tarreau9650f372009-08-16 14:02:45 +0200205/*
Willy Tarreau14f8e862012-08-30 22:23:13 +0200206 * This function initiates a TCP connection establishment to the target assigned
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200207 * to connection <conn> using (si->{target,dst}). A source address may be
208 * pointed to by conn->src in case of transparent proxying. Normal source
Willy Tarreau14f8e862012-08-30 22:23:13 +0200209 * bind addresses are still determined locally (due to the possible need of a
210 * source port). conn->target may point either to a valid server or to a backend,
Willy Tarreau3fdb3662012-11-12 00:42:33 +0100211 * depending on conn->target. Only OBJ_TYPE_PROXY and OBJ_TYPE_SERVER are
Willy Tarreauf0837b22012-11-24 10:24:27 +0100212 * supported. The <data> parameter is a boolean indicating whether there are data
213 * waiting for being sent or not, in order to adjust data write polling and on
Olivier Houchardfdcb0072019-05-06 18:32:29 +0200214 * some platforms, the ability to avoid an empty initial ACK. The <flags> argument
215 * allows the caller to force using a delayed ACK when establishing the connection
Willy Tarreauf0837b22012-11-24 10:24:27 +0100216 * - 0 = no delayed ACK unless data are advertised and backend has tcp-smart-connect
Olivier Houchardfdcb0072019-05-06 18:32:29 +0200217 * - CONNECT_DELACK_SMART_CONNECT = delayed ACK if backend has tcp-smart-connect, regardless of data
218 * - CONNECT_DELACK_ALWAYS = delayed ACK regardless of backend options
Willy Tarreaub1d67742010-03-29 19:36:59 +0200219 *
Willy Tarreau1ec74bf2013-10-24 21:45:00 +0200220 * Note that a pending send_proxy message accounts for data.
221 *
Willy Tarreau9650f372009-08-16 14:02:45 +0200222 * It can return one of :
Willy Tarreaue7dff022015-04-03 01:14:29 +0200223 * - SF_ERR_NONE if everything's OK
224 * - SF_ERR_SRVTO if there are no more servers
225 * - SF_ERR_SRVCL if the connection was refused by the server
226 * - SF_ERR_PRXCOND if the connection has been limited by the proxy (maxconn)
227 * - SF_ERR_RESOURCE if a system resource is lacking (eg: fd limits, ports, ...)
228 * - SF_ERR_INTERNAL for any other purely internal errors
Tim Düsterhus4896c442016-11-29 02:15:19 +0100229 * Additionally, in the case of SF_ERR_RESOURCE, an emergency log will be emitted.
Willy Tarreau6b0a8502012-11-23 08:51:32 +0100230 *
Willy Tarreaue7dff022015-04-03 01:14:29 +0200231 * The connection's fd is inserted only when SF_ERR_NONE is returned, otherwise
Willy Tarreau6b0a8502012-11-23 08:51:32 +0100232 * it's invalid and the caller has nothing to do.
Willy Tarreau9650f372009-08-16 14:02:45 +0200233 */
Willy Tarreauf1536862011-03-03 18:27:32 +0100234
Olivier Houchardfdcb0072019-05-06 18:32:29 +0200235int tcp_connect_server(struct connection *conn, int flags)
Willy Tarreau9650f372009-08-16 14:02:45 +0200236{
237 int fd;
Willy Tarreauac825402011-03-04 22:04:29 +0100238 struct server *srv;
239 struct proxy *be;
Willy Tarreaua4380b42012-12-08 22:49:11 +0100240 struct conn_src *src;
Willy Tarreau034c88c2017-01-23 23:36:45 +0100241 int use_fastopen = 0;
Alexander Liu2a54bb72019-05-22 19:44:48 +0800242 struct sockaddr_storage *addr;
Willy Tarreauac825402011-03-04 22:04:29 +0100243
Olivier Houchard637b6952018-11-23 14:23:07 +0100244 conn->flags |= CO_FL_WAIT_L4_CONN; /* connection in progress */
Willy Tarreau9ce70132014-01-24 16:08:19 +0100245
Willy Tarreau3fdb3662012-11-12 00:42:33 +0100246 switch (obj_type(conn->target)) {
247 case OBJ_TYPE_PROXY:
248 be = objt_proxy(conn->target);
Willy Tarreauac825402011-03-04 22:04:29 +0100249 srv = NULL;
250 break;
Willy Tarreau3fdb3662012-11-12 00:42:33 +0100251 case OBJ_TYPE_SERVER:
252 srv = objt_server(conn->target);
Willy Tarreauac825402011-03-04 22:04:29 +0100253 be = srv->proxy;
Willy Tarreau034c88c2017-01-23 23:36:45 +0100254 /* Make sure we check that we have data before activating
255 * TFO, or we could trigger a kernel issue whereby after
256 * a successful connect() == 0, any subsequent connect()
257 * will return EINPROGRESS instead of EISCONN.
258 */
259 use_fastopen = (srv->flags & SRV_F_FASTOPEN) &&
260 ((flags & (CONNECT_CAN_USE_TFO | CONNECT_HAS_DATA)) ==
261 (CONNECT_CAN_USE_TFO | CONNECT_HAS_DATA));
Willy Tarreauac825402011-03-04 22:04:29 +0100262 break;
263 default:
Willy Tarreau9ce70132014-01-24 16:08:19 +0100264 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200265 return SF_ERR_INTERNAL;
Willy Tarreauac825402011-03-04 22:04:29 +0100266 }
Willy Tarreau9650f372009-08-16 14:02:45 +0200267
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200268 if (!conn->dst) {
269 conn->flags |= CO_FL_ERROR;
270 return SF_ERR_INTERNAL;
271 }
272
Willy Tarreau18b7df72020-08-28 12:07:22 +0200273 fd = conn->handle.fd = sock_create_server_socket(conn);
KOVACS Krisztianb3e54fe2014-11-17 15:11:45 +0100274
275 if (fd == -1) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200276 qfprintf(stderr, "Cannot get a server socket.\n");
277
Willy Tarreau9ce70132014-01-24 16:08:19 +0100278 if (errno == ENFILE) {
279 conn->err_code = CO_ER_SYS_FDLIM;
Willy Tarreau9650f372009-08-16 14:02:45 +0200280 send_log(be, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100281 "Proxy %s reached system FD limit (maxsock=%d). Please check system tunables.\n",
282 be->id, global.maxsock);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100283 }
284 else if (errno == EMFILE) {
285 conn->err_code = CO_ER_PROC_FDLIM;
Willy Tarreau9650f372009-08-16 14:02:45 +0200286 send_log(be, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100287 "Proxy %s reached process FD limit (maxsock=%d). Please check 'ulimit-n' and restart.\n",
288 be->id, global.maxsock);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100289 }
290 else if (errno == ENOBUFS || errno == ENOMEM) {
291 conn->err_code = CO_ER_SYS_MEMLIM;
Willy Tarreau9650f372009-08-16 14:02:45 +0200292 send_log(be, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100293 "Proxy %s reached system memory limit (maxsock=%d). Please check system tunables.\n",
294 be->id, global.maxsock);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100295 }
296 else if (errno == EAFNOSUPPORT || errno == EPROTONOSUPPORT) {
297 conn->err_code = CO_ER_NOPROTO;
298 }
299 else
300 conn->err_code = CO_ER_SOCK_ERR;
301
Willy Tarreau9650f372009-08-16 14:02:45 +0200302 /* this is a resource error */
Willy Tarreau9ce70132014-01-24 16:08:19 +0100303 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200304 return SF_ERR_RESOURCE;
Willy Tarreau9650f372009-08-16 14:02:45 +0200305 }
306
307 if (fd >= global.maxsock) {
308 /* do not log anything there, it's a normal condition when this option
309 * is used to serialize connections to a server !
310 */
Christopher Faulet767a84b2017-11-24 16:50:31 +0100311 ha_alert("socket(): not enough free sockets. Raise -n argument. Giving up.\n");
Willy Tarreau9650f372009-08-16 14:02:45 +0200312 close(fd);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100313 conn->err_code = CO_ER_CONF_FDLIM;
314 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200315 return SF_ERR_PRXCOND; /* it is a configuration limit */
Willy Tarreau9650f372009-08-16 14:02:45 +0200316 }
317
318 if ((fcntl(fd, F_SETFL, O_NONBLOCK)==-1) ||
Simon Hormande072bd2011-06-24 15:11:37 +0900319 (setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &one, sizeof(one)) == -1)) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200320 qfprintf(stderr,"Cannot set client socket to non blocking mode.\n");
321 close(fd);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100322 conn->err_code = CO_ER_SOCK_ERR;
323 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200324 return SF_ERR_INTERNAL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200325 }
326
William Lallemandc03eb012018-11-27 12:02:37 +0100327 if (master == 1 && (fcntl(fd, F_SETFD, FD_CLOEXEC) == -1)) {
328 ha_alert("Cannot set CLOEXEC on client socket.\n");
329 close(fd);
330 conn->err_code = CO_ER_SOCK_ERR;
331 conn->flags |= CO_FL_ERROR;
332 return SF_ERR_INTERNAL;
333 }
334
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900335 if (be->options & PR_O_TCP_SRV_KA) {
Simon Hormande072bd2011-06-24 15:11:37 +0900336 setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, &one, sizeof(one));
Willy Tarreau9650f372009-08-16 14:02:45 +0200337
Willy Tarreau52543212020-07-09 05:58:51 +0200338#ifdef TCP_KEEPCNT
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900339 if (be->srvtcpka_cnt)
340 setsockopt(fd, IPPROTO_TCP, TCP_KEEPCNT, &be->srvtcpka_cnt, sizeof(be->srvtcpka_cnt));
Willy Tarreau52543212020-07-09 05:58:51 +0200341#endif
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900342
Willy Tarreau52543212020-07-09 05:58:51 +0200343#ifdef TCP_KEEPIDLE
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900344 if (be->srvtcpka_idle)
345 setsockopt(fd, IPPROTO_TCP, TCP_KEEPIDLE, &be->srvtcpka_idle, sizeof(be->srvtcpka_idle));
Willy Tarreau52543212020-07-09 05:58:51 +0200346#endif
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900347
Willy Tarreau52543212020-07-09 05:58:51 +0200348#ifdef TCP_KEEPINTVL
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900349 if (be->srvtcpka_intvl)
350 setsockopt(fd, IPPROTO_TCP, TCP_KEEPINTVL, &be->srvtcpka_intvl, sizeof(be->srvtcpka_intvl));
Willy Tarreau52543212020-07-09 05:58:51 +0200351#endif
MIZUTA Takeshib24bc0d2020-07-09 11:13:20 +0900352 }
353
Willy Tarreau9650f372009-08-16 14:02:45 +0200354 /* allow specific binding :
355 * - server-specific at first
356 * - proxy-specific next
357 */
Willy Tarreaua4380b42012-12-08 22:49:11 +0100358 if (srv && srv->conn_src.opts & CO_SRC_BIND)
359 src = &srv->conn_src;
360 else if (be->conn_src.opts & CO_SRC_BIND)
361 src = &be->conn_src;
362 else
363 src = NULL;
364
365 if (src) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200366 int ret, flags = 0;
367
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200368 if (conn->src && is_inet_addr(conn->src)) {
Willy Tarreaua4380b42012-12-08 22:49:11 +0100369 switch (src->opts & CO_SRC_TPROXY_MASK) {
Willy Tarreauef9a3602012-12-08 22:29:20 +0100370 case CO_SRC_TPROXY_CLI:
Christopher Faulet21ddc742020-07-01 15:26:14 +0200371 conn_set_private(conn);
Willy Tarreau387ebf82015-08-04 19:24:13 +0200372 /* fall through */
373 case CO_SRC_TPROXY_ADDR:
Willy Tarreau5f2877a2012-10-26 19:57:58 +0200374 flags = 3;
375 break;
Willy Tarreauef9a3602012-12-08 22:29:20 +0100376 case CO_SRC_TPROXY_CIP:
377 case CO_SRC_TPROXY_DYN:
Christopher Faulet21ddc742020-07-01 15:26:14 +0200378 conn_set_private(conn);
Willy Tarreau5f2877a2012-10-26 19:57:58 +0200379 flags = 1;
380 break;
381 }
Willy Tarreau9650f372009-08-16 14:02:45 +0200382 }
Willy Tarreaub1d67742010-03-29 19:36:59 +0200383
Willy Tarreau9650f372009-08-16 14:02:45 +0200384#ifdef SO_BINDTODEVICE
385 /* Note: this might fail if not CAP_NET_RAW */
Willy Tarreaua4380b42012-12-08 22:49:11 +0100386 if (src->iface_name)
387 setsockopt(fd, SOL_SOCKET, SO_BINDTODEVICE, src->iface_name, src->iface_len + 1);
Willy Tarreau9650f372009-08-16 14:02:45 +0200388#endif
389
Willy Tarreaua4380b42012-12-08 22:49:11 +0100390 if (src->sport_range) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200391 int attempts = 10; /* should be more than enough to find a spare port */
Willy Tarreaua4380b42012-12-08 22:49:11 +0100392 struct sockaddr_storage sa;
Willy Tarreau9650f372009-08-16 14:02:45 +0200393
394 ret = 1;
Vincent Bernat6e615892016-05-18 16:17:44 +0200395 memcpy(&sa, &src->source_addr, sizeof(sa));
Willy Tarreau9650f372009-08-16 14:02:45 +0200396
397 do {
398 /* note: in case of retry, we may have to release a previously
399 * allocated port, hence this loop's construct.
400 */
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200401 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
402 fdinfo[fd].port_range = NULL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200403
404 if (!attempts)
405 break;
406 attempts--;
407
Willy Tarreaua4380b42012-12-08 22:49:11 +0100408 fdinfo[fd].local_port = port_range_alloc_port(src->sport_range);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100409 if (!fdinfo[fd].local_port) {
410 conn->err_code = CO_ER_PORT_RANGE;
Willy Tarreau9650f372009-08-16 14:02:45 +0200411 break;
Willy Tarreau9ce70132014-01-24 16:08:19 +0100412 }
Willy Tarreau9650f372009-08-16 14:02:45 +0200413
Willy Tarreaua4380b42012-12-08 22:49:11 +0100414 fdinfo[fd].port_range = src->sport_range;
415 set_host_port(&sa, fdinfo[fd].local_port);
Willy Tarreau9650f372009-08-16 14:02:45 +0200416
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200417 ret = tcp_bind_socket(fd, flags, &sa, conn->src);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100418 if (ret != 0)
419 conn->err_code = CO_ER_CANT_BIND;
Willy Tarreau9650f372009-08-16 14:02:45 +0200420 } while (ret != 0); /* binding NOK */
421 }
422 else {
Lukas Tribus7d56c6d2016-09-13 09:51:15 +0000423#ifdef IP_BIND_ADDRESS_NO_PORT
Christopher Faulet1bc04c72017-10-29 20:14:08 +0100424 static THREAD_LOCAL int bind_address_no_port = 1;
Lukas Tribus7d56c6d2016-09-13 09:51:15 +0000425 setsockopt(fd, SOL_IP, IP_BIND_ADDRESS_NO_PORT, (const void *) &bind_address_no_port, sizeof(int));
426#endif
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200427 ret = tcp_bind_socket(fd, flags, &src->source_addr, conn->src);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100428 if (ret != 0)
429 conn->err_code = CO_ER_CANT_BIND;
Willy Tarreau9650f372009-08-16 14:02:45 +0200430 }
431
Willy Tarreaua4380b42012-12-08 22:49:11 +0100432 if (unlikely(ret != 0)) {
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200433 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
434 fdinfo[fd].port_range = NULL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200435 close(fd);
436
Willy Tarreau9650f372009-08-16 14:02:45 +0200437 if (ret == 1) {
Christopher Faulet767a84b2017-11-24 16:50:31 +0100438 ha_alert("Cannot bind to source address before connect() for backend %s. Aborting.\n",
439 be->id);
Willy Tarreau9650f372009-08-16 14:02:45 +0200440 send_log(be, LOG_EMERG,
Willy Tarreaua4380b42012-12-08 22:49:11 +0100441 "Cannot bind to source address before connect() for backend %s.\n",
Willy Tarreau9650f372009-08-16 14:02:45 +0200442 be->id);
443 } else {
Christopher Faulet767a84b2017-11-24 16:50:31 +0100444 ha_alert("Cannot bind to tproxy source address before connect() for backend %s. Aborting.\n",
445 be->id);
Willy Tarreau9650f372009-08-16 14:02:45 +0200446 send_log(be, LOG_EMERG,
Willy Tarreaua4380b42012-12-08 22:49:11 +0100447 "Cannot bind to tproxy source address before connect() for backend %s.\n",
Willy Tarreau9650f372009-08-16 14:02:45 +0200448 be->id);
449 }
Willy Tarreau9ce70132014-01-24 16:08:19 +0100450 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200451 return SF_ERR_RESOURCE;
Willy Tarreau9650f372009-08-16 14:02:45 +0200452 }
453 }
454
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +0400455#if defined(TCP_QUICKACK)
Willy Tarreau9650f372009-08-16 14:02:45 +0200456 /* disabling tcp quick ack now allows the first request to leave the
457 * machine with the first ACK. We only do this if there are pending
Willy Tarreauf0837b22012-11-24 10:24:27 +0100458 * data in the buffer.
Willy Tarreau9650f372009-08-16 14:02:45 +0200459 */
Olivier Houchardfdcb0072019-05-06 18:32:29 +0200460 if (flags & (CONNECT_DELACK_ALWAYS) ||
461 ((flags & CONNECT_DELACK_SMART_CONNECT ||
462 (flags & CONNECT_HAS_DATA) || conn->send_proxy_ofs) &&
463 (be->options2 & PR_O2_SMARTCON)))
Simon Hormande072bd2011-06-24 15:11:37 +0900464 setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, &zero, sizeof(zero));
Willy Tarreau9650f372009-08-16 14:02:45 +0200465#endif
466
Willy Tarreau163d4622015-10-13 16:16:41 +0200467#ifdef TCP_USER_TIMEOUT
468 /* there is not much more we can do here when it fails, it's still minor */
469 if (srv && srv->tcp_ut)
470 setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &srv->tcp_ut, sizeof(srv->tcp_ut));
471#endif
Willy Tarreau034c88c2017-01-23 23:36:45 +0100472
473 if (use_fastopen) {
474#if defined(TCP_FASTOPEN_CONNECT)
475 setsockopt(fd, IPPROTO_TCP, TCP_FASTOPEN_CONNECT, &one, sizeof(one));
476#endif
477 }
Willy Tarreaue803de22010-01-21 17:43:04 +0100478 if (global.tune.server_sndbuf)
479 setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &global.tune.server_sndbuf, sizeof(global.tune.server_sndbuf));
480
481 if (global.tune.server_rcvbuf)
482 setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &global.tune.server_rcvbuf, sizeof(global.tune.server_rcvbuf));
483
Willy Tarreau7bbc4a52019-07-17 15:41:35 +0200484 addr = (conn->flags & CO_FL_SOCKS4) ? &srv->socks4_addr : conn->dst;
Alexander Liu2a54bb72019-05-22 19:44:48 +0800485 if (connect(fd, (const struct sockaddr *)addr, get_addr_len(addr)) == -1) {
Willy Tarreau819efbf2017-01-25 14:12:22 +0100486 if (errno == EINPROGRESS || errno == EALREADY) {
487 /* common case, let's wait for connect status */
488 conn->flags |= CO_FL_WAIT_L4_CONN;
489 }
490 else if (errno == EISCONN) {
491 /* should normally not happen but if so, indicates that it's OK */
492 conn->flags &= ~CO_FL_WAIT_L4_CONN;
493 }
494 else if (errno == EAGAIN || errno == EADDRINUSE || errno == EADDRNOTAVAIL) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200495 char *msg;
Willy Tarreau9ce70132014-01-24 16:08:19 +0100496 if (errno == EAGAIN || errno == EADDRNOTAVAIL) {
Willy Tarreau9650f372009-08-16 14:02:45 +0200497 msg = "no free ports";
Willy Tarreau9ce70132014-01-24 16:08:19 +0100498 conn->err_code = CO_ER_FREE_PORTS;
499 }
500 else {
Willy Tarreau9650f372009-08-16 14:02:45 +0200501 msg = "local address already in use";
Willy Tarreau9ce70132014-01-24 16:08:19 +0100502 conn->err_code = CO_ER_ADDR_INUSE;
503 }
Willy Tarreau9650f372009-08-16 14:02:45 +0200504
Willy Tarreaub1719512012-12-08 23:03:28 +0100505 qfprintf(stderr,"Connect() failed for backend %s: %s.\n", be->id, msg);
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200506 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
507 fdinfo[fd].port_range = NULL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200508 close(fd);
Willy Tarreaub1719512012-12-08 23:03:28 +0100509 send_log(be, LOG_ERR, "Connect() failed for backend %s: %s.\n", be->id, msg);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100510 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200511 return SF_ERR_RESOURCE;
Willy Tarreau9650f372009-08-16 14:02:45 +0200512 } else if (errno == ETIMEDOUT) {
513 //qfprintf(stderr,"Connect(): ETIMEDOUT");
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200514 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
515 fdinfo[fd].port_range = NULL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200516 close(fd);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100517 conn->err_code = CO_ER_SOCK_ERR;
518 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200519 return SF_ERR_SRVTO;
Willy Tarreau9650f372009-08-16 14:02:45 +0200520 } else {
521 // (errno == ECONNREFUSED || errno == ENETUNREACH || errno == EACCES || errno == EPERM)
522 //qfprintf(stderr,"Connect(): %d", errno);
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200523 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
524 fdinfo[fd].port_range = NULL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200525 close(fd);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100526 conn->err_code = CO_ER_SOCK_ERR;
527 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200528 return SF_ERR_SRVCL;
Willy Tarreau9650f372009-08-16 14:02:45 +0200529 }
530 }
Willy Tarreau819efbf2017-01-25 14:12:22 +0100531 else {
532 /* connect() == 0, this is great! */
533 conn->flags &= ~CO_FL_WAIT_L4_CONN;
534 }
Willy Tarreau9650f372009-08-16 14:02:45 +0200535
Willy Tarreaufc8f1f02012-12-08 18:53:44 +0100536 conn->flags |= CO_FL_ADDR_TO_SET;
Willy Tarreaueeda90e2012-05-11 19:53:32 +0200537
Willy Tarreauf79c8172013-10-21 16:30:56 +0200538 conn_ctrl_init(conn); /* registers the FD */
Willy Tarreauad38ace2013-12-15 14:19:38 +0100539 fdtab[fd].linger_risk = 1; /* close hard if needed */
Willy Tarreau15678ef2012-08-31 13:54:11 +0200540
Willy Tarreau4c69cff2020-03-04 16:38:00 +0100541 if (conn->flags & CO_FL_WAIT_L4_CONN) {
542 fd_want_send(fd);
543 fd_cant_send(fd);
Willy Tarreau8dbd1a22020-07-31 08:59:09 +0200544 fd_cant_recv(fd);
Willy Tarreau4c69cff2020-03-04 16:38:00 +0100545 }
Willy Tarreauccf3f6d2019-09-05 17:05:05 +0200546
Willy Tarreauf7bc57c2012-10-03 00:19:48 +0200547 if (conn_xprt_init(conn) < 0) {
Willy Tarreau3f2770b2017-10-05 18:01:29 +0200548 conn_full_close(conn);
Willy Tarreau9ce70132014-01-24 16:08:19 +0100549 conn->flags |= CO_FL_ERROR;
Willy Tarreaue7dff022015-04-03 01:14:29 +0200550 return SF_ERR_RESOURCE;
Willy Tarreau184636e2012-09-06 14:04:41 +0200551 }
Willy Tarreau15678ef2012-08-31 13:54:11 +0200552
Willy Tarreaue7dff022015-04-03 01:14:29 +0200553 return SF_ERR_NONE; /* connection is OK */
Willy Tarreau9650f372009-08-16 14:02:45 +0200554}
555
Willy Tarreaue6b98942007-10-29 01:09:36 +0100556/* This function tries to bind a TCPv4/v6 listener. It may return a warning or
Willy Tarreau8ab505b2013-01-24 01:41:38 +0100557 * an error message in <errmsg> if the message is at most <errlen> bytes long
558 * (including '\0'). Note that <errmsg> may be NULL if <errlen> is also zero.
559 * The return value is composed from ERR_ABORT, ERR_WARN,
Willy Tarreaue6b98942007-10-29 01:09:36 +0100560 * ERR_ALERT, ERR_RETRYABLE and ERR_FATAL. ERR_NONE indicates that everything
561 * was alright and that no message was returned. ERR_RETRYABLE means that an
562 * error occurred but that it may vanish after a retry (eg: port in use), and
Aman Guptad94991d2012-04-06 17:39:26 -0700563 * ERR_FATAL indicates a non-fixable error. ERR_WARN and ERR_ALERT do not alter
Willy Tarreaue6b98942007-10-29 01:09:36 +0100564 * the meaning of the error, but just indicate that a message is present which
565 * should be displayed with the respective level. Last, ERR_ABORT indicates
566 * that it's pointless to try to start other listeners. No error message is
567 * returned if errlen is NULL.
568 */
569int tcp_bind_listener(struct listener *listener, char *errmsg, int errlen)
570{
Willy Tarreaue6b98942007-10-29 01:09:36 +0100571 int fd, err;
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200572 int ready;
Willy Tarreau40aa0702013-03-10 23:51:38 +0100573 socklen_t ready_len;
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200574 char *msg = NULL;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100575
Willy Tarreauad33acf2020-09-02 18:40:02 +0200576 err = ERR_NONE;
577
Willy Tarreaue6b98942007-10-29 01:09:36 +0100578 /* ensure we never return garbage */
Willy Tarreau8ab505b2013-01-24 01:41:38 +0100579 if (errlen)
Willy Tarreaue6b98942007-10-29 01:09:36 +0100580 *errmsg = 0;
581
582 if (listener->state != LI_ASSIGNED)
583 return ERR_NONE; /* already bound */
584
Willy Tarreauad33acf2020-09-02 18:40:02 +0200585 if (!(listener->rx.flags & RX_F_BOUND)) {
586 msg = "receiving socket not bound";
587 goto tcp_return;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100588 }
589
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200590 fd = listener->rx.fd;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100591
592 if (listener->options & LI_O_NOLINGER)
Simon Hormande072bd2011-06-24 15:11:37 +0900593 setsockopt(fd, SOL_SOCKET, SO_LINGER, &nolinger, sizeof(struct linger));
Olivier Houchard153659f2017-04-05 22:39:56 +0200594 else {
595 struct linger tmplinger;
596 socklen_t len = sizeof(tmplinger);
597 if (getsockopt(fd, SOL_SOCKET, SO_LINGER, &tmplinger, &len) == 0 &&
598 (tmplinger.l_onoff == 1 || tmplinger.l_linger == 0)) {
599 tmplinger.l_onoff = 0;
600 tmplinger.l_linger = 0;
601 setsockopt(fd, SOL_SOCKET, SO_LINGER, &tmplinger,
602 sizeof(tmplinger));
603 }
604 }
Willy Tarreauedcf6682008-11-30 23:15:34 +0100605
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +0400606#if defined(TCP_MAXSEG)
Willy Tarreau48a7e722010-12-24 15:26:39 +0100607 if (listener->maxseg > 0) {
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +0400608 if (setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG,
Willy Tarreaube1b9182009-06-14 18:48:19 +0200609 &listener->maxseg, sizeof(listener->maxseg)) == -1) {
610 msg = "cannot set MSS";
611 err |= ERR_WARN;
612 }
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200613 } else {
614 /* we may want to try to restore the default MSS if the socket was inherited */
Olivier Houchard153659f2017-04-05 22:39:56 +0200615 int tmpmaxseg = -1;
616 int defaultmss;
617 socklen_t len = sizeof(tmpmaxseg);
618
Willy Tarreau37159062020-08-27 07:48:42 +0200619 if (listener->rx.addr.ss_family == AF_INET)
Willy Tarreaue5bdc512020-08-28 18:03:10 +0200620 defaultmss = sock_inet_tcp_maxseg_default;
Olivier Houchard153659f2017-04-05 22:39:56 +0200621 else
Willy Tarreaue5bdc512020-08-28 18:03:10 +0200622 defaultmss = sock_inet6_tcp_maxseg_default;
Olivier Houchard153659f2017-04-05 22:39:56 +0200623
624 getsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &tmpmaxseg, &len);
William Dauchy97a7bda2020-02-12 15:53:04 +0100625 if (defaultmss > 0 &&
626 tmpmaxseg != defaultmss &&
627 setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &defaultmss, sizeof(defaultmss)) == -1) {
Olivier Houchard153659f2017-04-05 22:39:56 +0200628 msg = "cannot set MSS";
629 err |= ERR_WARN;
630 }
Willy Tarreaube1b9182009-06-14 18:48:19 +0200631 }
632#endif
Willy Tarreau2af207a2015-02-04 00:45:58 +0100633#if defined(TCP_USER_TIMEOUT)
634 if (listener->tcp_ut) {
635 if (setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT,
636 &listener->tcp_ut, sizeof(listener->tcp_ut)) == -1) {
637 msg = "cannot set TCP User Timeout";
638 err |= ERR_WARN;
639 }
Olivier Houchard153659f2017-04-05 22:39:56 +0200640 } else
641 setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &zero,
642 sizeof(zero));
Willy Tarreau2af207a2015-02-04 00:45:58 +0100643#endif
Willy Tarreaucb6cd432009-10-13 07:34:14 +0200644#if defined(TCP_DEFER_ACCEPT)
645 if (listener->options & LI_O_DEF_ACCEPT) {
646 /* defer accept by up to one second */
647 int accept_delay = 1;
648 if (setsockopt(fd, IPPROTO_TCP, TCP_DEFER_ACCEPT, &accept_delay, sizeof(accept_delay)) == -1) {
649 msg = "cannot enable DEFER_ACCEPT";
650 err |= ERR_WARN;
651 }
Olivier Houchard153659f2017-04-05 22:39:56 +0200652 } else
653 setsockopt(fd, IPPROTO_TCP, TCP_DEFER_ACCEPT, &zero,
654 sizeof(zero));
Willy Tarreaucb6cd432009-10-13 07:34:14 +0200655#endif
Willy Tarreau1c862c52012-10-05 16:21:00 +0200656#if defined(TCP_FASTOPEN)
657 if (listener->options & LI_O_TCP_FO) {
658 /* TFO needs a queue length, let's use the configured backlog */
Willy Tarreaue2711c72019-02-27 15:39:41 +0100659 int qlen = listener_backlog(listener);
Willy Tarreau1c862c52012-10-05 16:21:00 +0200660 if (setsockopt(fd, IPPROTO_TCP, TCP_FASTOPEN, &qlen, sizeof(qlen)) == -1) {
661 msg = "cannot enable TCP_FASTOPEN";
662 err |= ERR_WARN;
663 }
Olivier Houchard153659f2017-04-05 22:39:56 +0200664 } else {
665 socklen_t len;
666 int qlen;
667 len = sizeof(qlen);
668 /* Only disable fast open if it was enabled, we don't want
669 * the kernel to create a fast open queue if there's none.
670 */
671 if (getsockopt(fd, IPPROTO_TCP, TCP_FASTOPEN, &qlen, &len) == 0 &&
672 qlen != 0) {
673 if (setsockopt(fd, IPPROTO_TCP, TCP_FASTOPEN, &zero,
674 sizeof(zero)) == -1) {
675 msg = "cannot disable TCP_FASTOPEN";
676 err |= ERR_WARN;
677 }
678 }
Willy Tarreau1c862c52012-10-05 16:21:00 +0200679 }
680#endif
Willy Tarreau40aa0702013-03-10 23:51:38 +0100681 ready = 0;
682 ready_len = sizeof(ready);
683 if (getsockopt(fd, SOL_SOCKET, SO_ACCEPTCONN, &ready, &ready_len) == -1)
684 ready = 0;
685
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200686 if (!ready && /* only listen if not already done by external process */
Willy Tarreaue2711c72019-02-27 15:39:41 +0100687 listen(fd, listener_backlog(listener)) == -1) {
Willy Tarreaue6b98942007-10-29 01:09:36 +0100688 err |= ERR_RETRYABLE | ERR_ALERT;
689 msg = "cannot listen to socket";
690 goto tcp_close_return;
691 }
Willy Tarreauedcf6682008-11-30 23:15:34 +0100692
Dmitry Sivachenkocaf58982009-08-24 15:11:06 +0400693#if defined(TCP_QUICKACK)
Willy Tarreau9ea05a72009-06-14 12:07:01 +0200694 if (listener->options & LI_O_NOQUICKACK)
Simon Hormande072bd2011-06-24 15:11:37 +0900695 setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, &zero, sizeof(zero));
Olivier Houchard153659f2017-04-05 22:39:56 +0200696 else
697 setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, &one, sizeof(one));
Willy Tarreau9ea05a72009-06-14 12:07:01 +0200698#endif
699
Willy Tarreaue6b98942007-10-29 01:09:36 +0100700 /* the socket is ready */
Willy Tarreaua37b2442020-09-24 07:23:45 +0200701 listener_set_state(listener, LI_LISTEN);
Christopher Fauleta1000032020-10-07 11:14:47 +0200702 goto tcp_return;
Willy Tarreaubb1caff2020-08-19 10:00:57 +0200703
Willy Tarreauaf9a7f52020-09-01 16:12:50 +0200704 tcp_close_return:
705 close(fd);
Willy Tarreauad33acf2020-09-02 18:40:02 +0200706 tcp_return:
Cyril Bonté43ba1b32010-11-01 19:26:01 +0100707 if (msg && errlen) {
708 char pn[INET6_ADDRSTRLEN];
709
Willy Tarreau37159062020-08-27 07:48:42 +0200710 addr_to_str(&listener->rx.addr, pn, sizeof(pn));
711 snprintf(errmsg, errlen, "%s [%s:%d]", msg, pn, get_host_port(&listener->rx.addr));
Cyril Bonté43ba1b32010-11-01 19:26:01 +0100712 }
Willy Tarreaue6b98942007-10-29 01:09:36 +0100713 return err;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100714}
715
Willy Tarreau32282382017-09-15 07:44:44 +0200716/* Add <listener> to the list of tcpv4 listeners, on port <port>. The
717 * listener's state is automatically updated from LI_INIT to LI_ASSIGNED.
718 * The number of listeners for the protocol is updated.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200719 *
720 * Must be called with proto_lock held.
721 *
Willy Tarreaue6b98942007-10-29 01:09:36 +0100722 */
Willy Tarreau9d5be5c2017-09-15 07:55:51 +0200723static void tcpv4_add_listener(struct listener *listener, int port)
Willy Tarreaue6b98942007-10-29 01:09:36 +0100724{
725 if (listener->state != LI_INIT)
726 return;
Willy Tarreaua37b2442020-09-24 07:23:45 +0200727 listener_set_state(listener, LI_ASSIGNED);
Willy Tarreaub7436612020-08-28 19:51:44 +0200728 listener->rx.proto = &proto_tcpv4;
Willy Tarreau37159062020-08-27 07:48:42 +0200729 ((struct sockaddr_in *)(&listener->rx.addr))->sin_port = htons(port);
Willy Tarreaud7f331c2020-09-25 17:01:43 +0200730 LIST_ADDQ(&proto_tcpv4.receivers, &listener->rx.proto_list);
731 proto_tcpv4.nb_receivers++;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100732}
733
Willy Tarreau32282382017-09-15 07:44:44 +0200734/* Add <listener> to the list of tcpv6 listeners, on port <port>. The
735 * listener's state is automatically updated from LI_INIT to LI_ASSIGNED.
736 * The number of listeners for the protocol is updated.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200737 *
738 * Must be called with proto_lock held.
739 *
Willy Tarreaue6b98942007-10-29 01:09:36 +0100740 */
Willy Tarreau9d5be5c2017-09-15 07:55:51 +0200741static void tcpv6_add_listener(struct listener *listener, int port)
Willy Tarreaue6b98942007-10-29 01:09:36 +0100742{
743 if (listener->state != LI_INIT)
744 return;
Willy Tarreaua37b2442020-09-24 07:23:45 +0200745 listener_set_state(listener, LI_ASSIGNED);
Willy Tarreaub7436612020-08-28 19:51:44 +0200746 listener->rx.proto = &proto_tcpv6;
Willy Tarreau37159062020-08-27 07:48:42 +0200747 ((struct sockaddr_in *)(&listener->rx.addr))->sin_port = htons(port);
Willy Tarreaud7f331c2020-09-25 17:01:43 +0200748 LIST_ADDQ(&proto_tcpv6.receivers, &listener->rx.proto_list);
749 proto_tcpv6.nb_receivers++;
Willy Tarreaue6b98942007-10-29 01:09:36 +0100750}
751
Willy Tarreau5ddf1ce2020-09-25 19:27:39 +0200752/* Enable receipt of incoming connections for listener <l>. The receiver must
753 * still be valid. Does nothing in early boot (needs fd_updt).
754 */
755static void tcp_enable_listener(struct listener *l)
756{
757 if (fd_updt)
758 fd_want_recv(l->rx.fd);
759}
760
761/* Disable receipt of incoming connections for listener <l>. The receiver must
762 * still be valid. Does nothing in early boot (needs fd_updt).
763 */
764static void tcp_disable_listener(struct listener *l)
765{
766 if (fd_updt)
767 fd_stop_recv(l->rx.fd);
768}
769
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200770/* Suspend a receiver. Returns < 0 in case of failure, 0 if the receiver
771 * was totally stopped, or > 0 if correctly suspended.
Willy Tarreau092d8652014-07-07 20:22:12 +0200772 */
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200773static int tcp_suspend_receiver(struct receiver *rx)
Willy Tarreau092d8652014-07-07 20:22:12 +0200774{
Willy Tarreau91c614d2020-10-08 16:51:09 +0200775 socklen_t opt_val, opt_len;
Willy Tarreau8b6fc3d2020-10-13 16:34:19 +0200776 struct sockaddr sa;
Willy Tarreau092d8652014-07-07 20:22:12 +0200777
Willy Tarreau8b6fc3d2020-10-13 16:34:19 +0200778 sa.sa_family = AF_UNSPEC;
779 if (connect(rx->fd, &sa, sizeof(sa)) < 0)
780 goto check_already_done;
Willy Tarreau91c614d2020-10-08 16:51:09 +0200781
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200782 fd_stop_recv(rx->fd);
Willy Tarreau092d8652014-07-07 20:22:12 +0200783 return 1;
Willy Tarreau91c614d2020-10-08 16:51:09 +0200784
785 check_already_done:
786 /* in case one of the shutdown() above fails, it might be because we're
787 * dealing with a socket that is shared with other processes doing the
788 * same. Let's check if it's still accepting connections.
789 */
790 opt_val = 0;
791 opt_len = sizeof(opt_val);
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200792 if (getsockopt(rx->fd, SOL_SOCKET, SO_ACCEPTCONN, &opt_val, &opt_len) == -1) {
793 fd_stop_recv(rx->fd);
Willy Tarreau91c614d2020-10-08 16:51:09 +0200794 return 0; /* the socket is really unrecoverable */
Willy Tarreaue53608b2020-09-24 18:20:37 +0200795 }
Willy Tarreau91c614d2020-10-08 16:51:09 +0200796
Willy Tarreaue53608b2020-09-24 18:20:37 +0200797 if (!opt_val) {
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200798 fd_stop_recv(rx->fd);
Willy Tarreau91c614d2020-10-08 16:51:09 +0200799 return 1; /* already paused by another process */
Willy Tarreaue53608b2020-09-24 18:20:37 +0200800 }
Willy Tarreau91c614d2020-10-08 16:51:09 +0200801
802 /* something looks fishy here */
803 return -1;
Willy Tarreau092d8652014-07-07 20:22:12 +0200804}
805
Willy Tarreau58e6b712020-09-25 19:40:31 +0200806/* Resume a receiver. Returns < 0 in case of failure, 0 if the receiver
807 * was totally stopped, or > 0 if correctly suspended.
808 */
809static int tcp_resume_receiver(struct receiver *rx)
810{
811 struct listener *l = LIST_ELEM(rx, struct listener *, rx);
812
813 if (rx->fd < 0)
814 return 0;
815
816 if (listen(rx->fd, listener_backlog(l)) == 0) {
817 fd_want_recv(l->rx.fd);
818 return 1;
819 }
820 return -1;
821}
822
Willy Tarreaue6b98942007-10-29 01:09:36 +0100823
824/*
825 * Local variables:
826 * c-indent-level: 8
827 * c-basic-offset: 8
828 * End:
829 */