blob: 848b01eaeeede33dc455a9135148313d76c6a8a0 [file] [log] [blame]
Willy Tarreaudd815982007-10-16 12:25:14 +02001/*
Willy Tarreaud1d54542012-09-12 22:58:11 +02002 * Listener management functions.
Willy Tarreaudd815982007-10-16 12:25:14 +02003 *
Willy Tarreau0ccb7442013-01-07 22:54:17 +01004 * Copyright 2000-2013 Willy Tarreau <w@1wt.eu>
Willy Tarreaudd815982007-10-16 12:25:14 +02005 *
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
Willy Tarreau44489252014-01-14 17:52:01 +010013#define _GNU_SOURCE
Willy Tarreau6ae1ba62014-05-07 19:01:58 +020014#include <ctype.h>
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020015#include <errno.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020016#include <stdio.h>
17#include <string.h>
Willy Tarreau95ccdde2014-02-01 09:28:36 +010018#include <unistd.h>
19#include <fcntl.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020020
Willy Tarreau1bc4aab2012-10-08 20:11:03 +020021#include <common/accept4.h>
Christopher Fauletf1f0c5f2017-11-22 12:06:43 +010022#include <common/cfgparse.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020023#include <common/config.h>
Willy Tarreaudabf2e22007-10-28 21:59:24 +010024#include <common/errors.h>
Willy Tarreau0108d902018-11-25 19:14:37 +010025#include <common/initcall.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020026#include <common/mini-clist.h>
27#include <common/standard.h>
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020028#include <common/time.h>
29
30#include <types/global.h>
Willy Tarreaud1d54542012-09-12 22:58:11 +020031#include <types/protocol.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020032
Willy Tarreau645513a2010-05-24 20:55:15 +020033#include <proto/acl.h>
Christopher Fauleta717b992018-04-10 14:43:00 +020034#include <proto/connection.h>
Willy Tarreaub648d632007-10-28 22:13:50 +010035#include <proto/fd.h>
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020036#include <proto/freq_ctr.h>
37#include <proto/log.h>
Willy Tarreau7a798e52016-04-14 11:13:20 +020038#include <proto/listener.h>
Willy Tarreau0de59fd2017-09-15 08:10:44 +020039#include <proto/protocol.h>
William Lallemand2fe7dd02018-09-11 16:51:29 +020040#include <proto/proto_sockpair.h>
Willy Tarreau0ccb7442013-01-07 22:54:17 +010041#include <proto/sample.h>
Willy Tarreaufb0afa72015-04-03 14:46:27 +020042#include <proto/stream.h>
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020043#include <proto/task.h>
Willy Tarreaub648d632007-10-28 22:13:50 +010044
Willy Tarreau26982662012-09-12 23:17:10 +020045/* List head of all known bind keywords */
46static struct bind_kw_list bind_keywords = {
47 .list = LIST_HEAD_INIT(bind_keywords.list)
48};
49
Olivier Houchardf73629d2017-04-05 22:33:04 +020050struct xfer_sock_list *xfer_sock_list = NULL;
51
Willy Tarreauf2cb1692019-07-11 10:08:31 +020052/* there is one listener queue per thread so that a thread unblocking the
53 * global queue can wake up listeners bound only to foreing threads by
Willy Tarreau2bd65a72019-09-24 06:55:18 +020054 * moving them to the remote queues and waking up the associated tasklet.
Willy Tarreauf2cb1692019-07-11 10:08:31 +020055 */
56static struct work_list *local_listener_queue;
57
Willy Tarreau1efafce2019-01-27 15:37:19 +010058#if defined(USE_THREAD)
59
60struct accept_queue_ring accept_queue_rings[MAX_THREADS] __attribute__((aligned(64))) = { };
61
62/* dequeue and process a pending connection from the local accept queue (single
63 * consumer). Returns the accepted fd or -1 if none was found. The listener is
64 * placed into *li. The address is copied into *addr for no more than *addr_len
65 * bytes, and the address length is returned into *addr_len.
66 */
67int accept_queue_pop_sc(struct accept_queue_ring *ring, struct listener **li, void *addr, int *addr_len)
68{
69 struct accept_queue_entry *e;
70 unsigned int pos, next;
71 struct listener *ptr;
72 int len;
73 int fd;
74
75 pos = ring->head;
76
77 if (pos == ring->tail)
78 return -1;
79
80 next = pos + 1;
81 if (next >= ACCEPT_QUEUE_SIZE)
82 next = 0;
83
84 e = &ring->entry[pos];
85
86 /* wait for the producer to update the listener's pointer */
87 while (1) {
88 ptr = e->listener;
89 __ha_barrier_load();
90 if (ptr)
91 break;
92 pl_cpu_relax();
93 }
94
95 fd = e->fd;
96 len = e->addr_len;
97 if (len > *addr_len)
98 len = *addr_len;
99
100 if (likely(len > 0))
101 memcpy(addr, &e->addr, len);
102
103 /* release the entry */
104 e->listener = NULL;
105
106 __ha_barrier_store();
107 ring->head = next;
108
109 *addr_len = len;
110 *li = ptr;
111
112 return fd;
113}
114
115
116/* tries to push a new accepted connection <fd> into ring <ring> for listener
117 * <li>, from address <addr> whose length is <addr_len>. Returns non-zero if it
118 * succeeds, or zero if the ring is full. Supports multiple producers.
119 */
120int accept_queue_push_mp(struct accept_queue_ring *ring, int fd,
121 struct listener *li, const void *addr, int addr_len)
122{
123 struct accept_queue_entry *e;
124 unsigned int pos, next;
125
126 pos = ring->tail;
127 do {
128 next = pos + 1;
129 if (next >= ACCEPT_QUEUE_SIZE)
130 next = 0;
131 if (next == ring->head)
132 return 0; // ring full
Olivier Houchard64213e92019-03-08 18:52:57 +0100133 } while (unlikely(!_HA_ATOMIC_CAS(&ring->tail, &pos, next)));
Willy Tarreau1efafce2019-01-27 15:37:19 +0100134
135
136 e = &ring->entry[pos];
137
138 if (addr_len > sizeof(e->addr))
139 addr_len = sizeof(e->addr);
140
141 if (addr_len)
142 memcpy(&e->addr, addr, addr_len);
143
144 e->addr_len = addr_len;
145 e->fd = fd;
146
147 __ha_barrier_store();
148 /* now commit the change */
149
150 e->listener = li;
151 return 1;
152}
153
154/* proceed with accepting new connections */
155static struct task *accept_queue_process(struct task *t, void *context, unsigned short state)
156{
157 struct accept_queue_ring *ring = context;
158 struct listener *li;
159 struct sockaddr_storage addr;
Christopher Faulet102854c2019-04-30 12:17:13 +0200160 unsigned int max_accept;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100161 int addr_len;
162 int ret;
163 int fd;
164
Christopher Faulet102854c2019-04-30 12:17:13 +0200165 /* if global.tune.maxaccept is -1, then max_accept is UINT_MAX. It
166 * is not really illimited, but it is probably enough.
167 */
168 max_accept = global.tune.maxaccept ? global.tune.maxaccept : 64;
169 for (; max_accept; max_accept--) {
Willy Tarreau1efafce2019-01-27 15:37:19 +0100170 addr_len = sizeof(addr);
171 fd = accept_queue_pop_sc(ring, &li, &addr, &addr_len);
172 if (fd < 0)
173 break;
174
Olivier Houchard64213e92019-03-08 18:52:57 +0100175 _HA_ATOMIC_ADD(&li->thr_conn[tid], 1);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100176 ret = li->accept(li, fd, &addr);
177 if (ret <= 0) {
178 /* connection was terminated by the application */
179 continue;
180 }
181
182 /* increase the per-process number of cumulated sessions, this
183 * may only be done once l->accept() has accepted the connection.
184 */
185 if (!(li->options & LI_O_UNLIMITED)) {
186 HA_ATOMIC_UPDATE_MAX(&global.sps_max,
187 update_freq_ctr(&global.sess_per_sec, 1));
188 if (li->bind_conf && li->bind_conf->is_ssl) {
189 HA_ATOMIC_UPDATE_MAX(&global.ssl_max,
190 update_freq_ctr(&global.ssl_per_sec, 1));
191 }
192 }
193 }
194
195 /* ran out of budget ? Let's come here ASAP */
Christopher Faulet102854c2019-04-30 12:17:13 +0200196 if (!max_accept)
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200197 tasklet_wakeup(ring->tasklet);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100198
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200199 return NULL;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100200}
201
202/* Initializes the accept-queues. Returns 0 on success, otherwise ERR_* flags */
203static int accept_queue_init()
204{
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200205 struct tasklet *t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100206 int i;
207
208 for (i = 0; i < global.nbthread; i++) {
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200209 t = tasklet_new();
Willy Tarreau1efafce2019-01-27 15:37:19 +0100210 if (!t) {
211 ha_alert("Out of memory while initializing accept queue for thread %d\n", i);
212 return ERR_FATAL|ERR_ABORT;
213 }
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200214 t->tid = i;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100215 t->process = accept_queue_process;
216 t->context = &accept_queue_rings[i];
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200217 accept_queue_rings[i].tasklet = t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100218 }
219 return 0;
220}
221
222REGISTER_CONFIG_POSTPARSER("multi-threaded accept queue", accept_queue_init);
223
224#endif // USE_THREAD
225
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100226/* This function adds the specified listener's file descriptor to the polling
227 * lists if it is in the LI_LISTEN state. The listener enters LI_READY or
Willy Tarreauae302532014-05-07 19:22:24 +0200228 * LI_FULL state depending on its number of connections. In deamon mode, we
229 * also support binding only the relevant processes to their respective
230 * listeners. We don't do that in debug mode however.
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100231 */
Christopher Fauletf5b8adc2017-06-02 10:00:35 +0200232static void enable_listener(struct listener *listener)
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100233{
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100234 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100235 if (listener->state == LI_LISTEN) {
William Lallemand095ba4c2017-06-01 17:38:50 +0200236 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau6daac192019-02-02 17:39:53 +0100237 !(proc_mask(listener->bind_conf->bind_proc) & pid_bit)) {
Willy Tarreauae302532014-05-07 19:22:24 +0200238 /* we don't want to enable this listener and don't
239 * want any fd event to reach it.
240 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200241 if (!(global.tune.options & GTUNE_SOCKET_TRANSFER))
Christopher Faulet510c0d62018-03-16 10:04:47 +0100242 do_unbind_listener(listener, 1);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200243 else {
Christopher Faulet510c0d62018-03-16 10:04:47 +0100244 do_unbind_listener(listener, 0);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200245 listener->state = LI_LISTEN;
246 }
Willy Tarreauae302532014-05-07 19:22:24 +0200247 }
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100248 else if (!listener->maxconn || listener->nbconn < listener->maxconn) {
Willy Tarreau49b046d2012-08-09 12:11:58 +0200249 fd_want_recv(listener->fd);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100250 listener->state = LI_READY;
Willy Tarreauae302532014-05-07 19:22:24 +0200251 }
252 else {
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100253 listener->state = LI_FULL;
254 }
255 }
William Lallemande22f11f2018-09-11 10:06:27 +0200256 /* if this listener is supposed to be only in the master, close it in the workers */
257 if ((global.mode & MODE_MWORKER) &&
258 (listener->options & LI_O_MWORKER) &&
259 master == 0) {
260 do_unbind_listener(listener, 1);
261 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100262 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100263}
264
265/* This function removes the specified listener's file descriptor from the
266 * polling lists if it is in the LI_READY or in the LI_FULL state. The listener
267 * enters LI_LISTEN.
268 */
Christopher Fauletf5b8adc2017-06-02 10:00:35 +0200269static void disable_listener(struct listener *listener)
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100270{
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100271 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100272 if (listener->state < LI_READY)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200273 goto end;
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100274 if (listener->state == LI_READY)
Willy Tarreau49b046d2012-08-09 12:11:58 +0200275 fd_stop_recv(listener->fd);
Olivier Houchard859dc802019-08-08 15:47:21 +0200276 MT_LIST_DEL(&listener->wait_queue);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100277 listener->state = LI_LISTEN;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200278 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100279 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100280}
281
Willy Tarreaube58c382011-07-24 18:28:10 +0200282/* This function tries to temporarily disable a listener, depending on the OS
283 * capabilities. Linux unbinds the listen socket after a SHUT_RD, and ignores
284 * SHUT_WR. Solaris refuses either shutdown(). OpenBSD ignores SHUT_RD but
285 * closes upon SHUT_WR and refuses to rebind. So a common validation path
286 * involves SHUT_WR && listen && SHUT_RD. In case of success, the FD's polling
287 * is disabled. It normally returns non-zero, unless an error is reported.
288 */
289int pause_listener(struct listener *l)
290{
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200291 int ret = 1;
292
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100293 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200294
Olivier Houchard1fc05162017-04-06 01:05:05 +0200295 if (l->state <= LI_ZOMBIE)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200296 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200297
Willy Tarreau092d8652014-07-07 20:22:12 +0200298 if (l->proto->pause) {
299 /* Returns < 0 in case of failure, 0 if the listener
300 * was totally stopped, or > 0 if correctly paused.
301 */
302 int ret = l->proto->pause(l);
Willy Tarreaube58c382011-07-24 18:28:10 +0200303
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200304 if (ret < 0) {
305 ret = 0;
306 goto end;
307 }
Willy Tarreau092d8652014-07-07 20:22:12 +0200308 else if (ret == 0)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200309 goto end;
Willy Tarreaub3fb60b2012-10-04 08:56:31 +0200310 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200311
Olivier Houchard859dc802019-08-08 15:47:21 +0200312 MT_LIST_DEL(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200313
Willy Tarreau49b046d2012-08-09 12:11:58 +0200314 fd_stop_recv(l->fd);
Willy Tarreaube58c382011-07-24 18:28:10 +0200315 l->state = LI_PAUSED;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200316 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100317 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200318 return ret;
Willy Tarreaube58c382011-07-24 18:28:10 +0200319}
320
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200321/* This function tries to resume a temporarily disabled listener. Paused, full,
322 * limited and disabled listeners are handled, which means that this function
323 * may replace enable_listener(). The resulting state will either be LI_READY
324 * or LI_FULL. 0 is returned in case of failure to resume (eg: dead socket).
Willy Tarreauae302532014-05-07 19:22:24 +0200325 * Listeners bound to a different process are not woken up unless we're in
Willy Tarreauaf2fd582015-04-14 12:07:16 +0200326 * foreground mode, and are ignored. If the listener was only in the assigned
327 * state, it's totally rebound. This can happen if a pause() has completely
328 * stopped it. If the resume fails, 0 is returned and an error might be
329 * displayed.
Willy Tarreaube58c382011-07-24 18:28:10 +0200330 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100331int resume_listener(struct listener *l)
Willy Tarreaube58c382011-07-24 18:28:10 +0200332{
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200333 int ret = 1;
334
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100335 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200336
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200337 /* check that another thread didn't to the job in parallel (e.g. at the
338 * end of listen_accept() while we'd come from dequeue_all_listeners().
339 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200340 if (MT_LIST_ADDED(&l->wait_queue))
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200341 goto end;
342
William Lallemand095ba4c2017-06-01 17:38:50 +0200343 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau6daac192019-02-02 17:39:53 +0100344 !(proc_mask(l->bind_conf->bind_proc) & pid_bit))
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200345 goto end;
Willy Tarreau3569df32017-03-15 12:47:46 +0100346
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200347 if (l->state == LI_ASSIGNED) {
348 char msg[100];
349 int err;
350
351 err = l->proto->bind(l, msg, sizeof(msg));
352 if (err & ERR_ALERT)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100353 ha_alert("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200354 else if (err & ERR_WARN)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100355 ha_warning("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200356
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200357 if (err & (ERR_FATAL | ERR_ABORT)) {
358 ret = 0;
359 goto end;
360 }
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200361 }
362
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200363 if (l->state < LI_PAUSED || l->state == LI_ZOMBIE) {
364 ret = 0;
365 goto end;
366 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200367
Willy Tarreaub3fb60b2012-10-04 08:56:31 +0200368 if (l->proto->sock_prot == IPPROTO_TCP &&
369 l->state == LI_PAUSED &&
Willy Tarreaue2711c72019-02-27 15:39:41 +0100370 listen(l->fd, listener_backlog(l)) != 0) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200371 ret = 0;
372 goto end;
373 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200374
375 if (l->state == LI_READY)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200376 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200377
Olivier Houchard859dc802019-08-08 15:47:21 +0200378 MT_LIST_DEL(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200379
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100380 if (l->maxconn && l->nbconn >= l->maxconn) {
Willy Tarreaube58c382011-07-24 18:28:10 +0200381 l->state = LI_FULL;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200382 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200383 }
384
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200385 if (!(thread_mask(l->bind_conf->bind_thread) & tid_bit)) {
386 /* we're not allowed to touch this listener's FD, let's requeue
387 * the listener into one of its owning thread's queue instead.
388 */
389 int first_thread = my_flsl(thread_mask(l->bind_conf->bind_thread)) - 1;
390 work_list_add(&local_listener_queue[first_thread], &l->wait_queue);
391 goto end;
392 }
393
Willy Tarreau49b046d2012-08-09 12:11:58 +0200394 fd_want_recv(l->fd);
Willy Tarreaube58c382011-07-24 18:28:10 +0200395 l->state = LI_READY;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200396 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100397 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200398 return ret;
399}
400
Willy Tarreau87b09662015-04-03 00:22:06 +0200401/* Marks a ready listener as full so that the stream code tries to re-enable
Willy Tarreau62793712011-07-24 19:23:38 +0200402 * it upon next close() using resume_listener().
403 */
Christopher Faulet5580ba22017-08-28 15:29:20 +0200404static void listener_full(struct listener *l)
Willy Tarreau62793712011-07-24 19:23:38 +0200405{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100406 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200407 if (l->state >= LI_READY) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200408 MT_LIST_DEL(&l->wait_queue);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100409 if (l->state != LI_FULL) {
410 fd_stop_recv(l->fd);
411 l->state = LI_FULL;
412 }
Willy Tarreau62793712011-07-24 19:23:38 +0200413 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100414 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200415}
416
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200417/* Marks a ready listener as limited so that we only try to re-enable it when
418 * resources are free again. It will be queued into the specified queue.
419 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200420static void limit_listener(struct listener *l, struct mt_list *list)
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200421{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100422 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200423 if (l->state == LI_READY) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200424 MT_LIST_ADDQ(list, &l->wait_queue);
Willy Tarreau49b046d2012-08-09 12:11:58 +0200425 fd_stop_recv(l->fd);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200426 l->state = LI_LIMITED;
427 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100428 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200429}
430
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100431/* This function adds all of the protocol's listener's file descriptors to the
432 * polling lists when they are in the LI_LISTEN state. It is intended to be
433 * used as a protocol's generic enable_all() primitive, for use after the
434 * fork(). It puts the listeners into LI_READY or LI_FULL states depending on
435 * their number of connections. It always returns ERR_NONE.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200436 *
437 * Must be called with proto_lock held.
438 *
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100439 */
440int enable_all_listeners(struct protocol *proto)
441{
442 struct listener *listener;
443
444 list_for_each_entry(listener, &proto->listeners, proto_list)
445 enable_listener(listener);
446 return ERR_NONE;
447}
448
449/* This function removes all of the protocol's listener's file descriptors from
450 * the polling lists when they are in the LI_READY or LI_FULL states. It is
451 * intended to be used as a protocol's generic disable_all() primitive. It puts
452 * the listeners into LI_LISTEN, and always returns ERR_NONE.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200453 *
454 * Must be called with proto_lock held.
455 *
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100456 */
457int disable_all_listeners(struct protocol *proto)
458{
459 struct listener *listener;
460
461 list_for_each_entry(listener, &proto->listeners, proto_list)
462 disable_listener(listener);
463 return ERR_NONE;
464}
465
Willy Tarreau241797a2019-12-10 14:10:52 +0100466/* Dequeues all listeners waiting for a resource the global wait queue */
467void dequeue_all_listeners()
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200468{
Willy Tarreau01abd022019-02-28 10:27:18 +0100469 struct listener *listener;
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200470
Willy Tarreau241797a2019-12-10 14:10:52 +0100471 while ((listener = MT_LIST_POP(&global_listener_queue, struct listener *, wait_queue))) {
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200472 /* This cannot fail because the listeners are by definition in
Willy Tarreau01abd022019-02-28 10:27:18 +0100473 * the LI_LIMITED state.
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200474 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100475 resume_listener(listener);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200476 }
477}
478
Willy Tarreau241797a2019-12-10 14:10:52 +0100479/* Dequeues all listeners waiting for a resource in proxy <px>'s queue */
480void dequeue_proxy_listeners(struct proxy *px)
481{
482 struct listener *listener;
483
484 while ((listener = MT_LIST_POP(&px->listener_queue, struct listener *, wait_queue))) {
485 /* This cannot fail because the listeners are by definition in
486 * the LI_LIMITED state.
487 */
488 resume_listener(listener);
489 }
490}
491
Christopher Faulet510c0d62018-03-16 10:04:47 +0100492/* Must be called with the lock held. Depending on <do_close> value, it does
493 * what unbind_listener or unbind_listener_no_close should do.
494 */
495void do_unbind_listener(struct listener *listener, int do_close)
Willy Tarreaub648d632007-10-28 22:13:50 +0100496{
Olivier Houcharda5188562019-03-08 15:35:42 +0100497 if (listener->state == LI_READY && fd_updt)
Willy Tarreau49b046d2012-08-09 12:11:58 +0200498 fd_stop_recv(listener->fd);
Willy Tarreaub648d632007-10-28 22:13:50 +0100499
Olivier Houchard859dc802019-08-08 15:47:21 +0200500 MT_LIST_DEL(&listener->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200501
Willy Tarreaube58c382011-07-24 18:28:10 +0200502 if (listener->state >= LI_PAUSED) {
Olivier Houchard1fc05162017-04-06 01:05:05 +0200503 if (do_close) {
504 fd_delete(listener->fd);
505 listener->fd = -1;
506 }
507 else
508 fd_remove(listener->fd);
Willy Tarreaub648d632007-10-28 22:13:50 +0100509 listener->state = LI_ASSIGNED;
510 }
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100511}
512
Olivier Houchard1fc05162017-04-06 01:05:05 +0200513/* This function closes the listening socket for the specified listener,
514 * provided that it's already in a listening state. The listener enters the
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100515 * LI_ASSIGNED state. This function is intended to be used as a generic
516 * function for standard protocols.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200517 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100518void unbind_listener(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200519{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100520 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100521 do_unbind_listener(listener, 1);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100522 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200523}
524
525/* This function pretends the listener is dead, but keeps the FD opened, so
526 * that we can provide it, for conf reloading.
527 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100528void unbind_listener_no_close(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200529{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100530 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100531 do_unbind_listener(listener, 0);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100532 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200533}
534
Willy Tarreau3acf8c32007-10-28 22:35:41 +0100535/* This function closes all listening sockets bound to the protocol <proto>,
536 * and the listeners end in LI_ASSIGNED state if they were higher. It does not
537 * detach them from the protocol. It always returns ERR_NONE.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200538 *
539 * Must be called with proto_lock held.
540 *
Willy Tarreau3acf8c32007-10-28 22:35:41 +0100541 */
542int unbind_all_listeners(struct protocol *proto)
543{
544 struct listener *listener;
545
546 list_for_each_entry(listener, &proto->listeners, proto_list)
547 unbind_listener(listener);
548 return ERR_NONE;
549}
550
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200551/* creates one or multiple listeners for bind_conf <bc> on sockaddr <ss> on port
552 * range <portl> to <porth>, and possibly attached to fd <fd> (or -1 for auto
553 * allocation). The address family is taken from ss->ss_family. The number of
554 * jobs and listeners is automatically increased by the number of listeners
William Lallemand75ea0a02017-11-15 19:02:58 +0100555 * created. If the <inherited> argument is set to 1, it specifies that the FD
556 * was obtained from a parent process.
557 * It returns non-zero on success, zero on error with the error message
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200558 * set in <err>.
559 */
560int create_listeners(struct bind_conf *bc, const struct sockaddr_storage *ss,
William Lallemand75ea0a02017-11-15 19:02:58 +0100561 int portl, int porth, int fd, int inherited, char **err)
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200562{
563 struct protocol *proto = protocol_by_family(ss->ss_family);
564 struct listener *l;
565 int port;
566
567 if (!proto) {
568 memprintf(err, "unsupported protocol family %d", ss->ss_family);
569 return 0;
570 }
571
572 for (port = portl; port <= porth; port++) {
573 l = calloc(1, sizeof(*l));
574 if (!l) {
575 memprintf(err, "out of memory");
576 return 0;
577 }
578 l->obj_type = OBJ_TYPE_LISTENER;
579 LIST_ADDQ(&bc->frontend->conf.listeners, &l->by_fe);
580 LIST_ADDQ(&bc->listeners, &l->by_bind);
581 l->bind_conf = bc;
582
583 l->fd = fd;
584 memcpy(&l->addr, ss, sizeof(*ss));
Olivier Houchard859dc802019-08-08 15:47:21 +0200585 MT_LIST_INIT(&l->wait_queue);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200586 l->state = LI_INIT;
587
588 proto->add(l, port);
589
William Lallemand75ea0a02017-11-15 19:02:58 +0100590 if (inherited)
591 l->options |= LI_O_INHERITED;
592
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100593 HA_SPIN_INIT(&l->lock);
Olivier Houchard64213e92019-03-08 18:52:57 +0100594 _HA_ATOMIC_ADD(&jobs, 1);
595 _HA_ATOMIC_ADD(&listeners, 1);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200596 }
597 return 1;
598}
599
Willy Tarreau1a64d162007-10-28 22:26:05 +0100600/* Delete a listener from its protocol's list of listeners. The listener's
601 * state is automatically updated from LI_ASSIGNED to LI_INIT. The protocol's
Willy Tarreau2cc5bae2017-09-15 08:18:11 +0200602 * number of listeners is updated, as well as the global number of listeners
603 * and jobs. Note that the listener must have previously been unbound. This
604 * is the generic function to use to remove a listener.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200605 *
606 * Will grab the proto_lock.
607 *
Willy Tarreau1a64d162007-10-28 22:26:05 +0100608 */
609void delete_listener(struct listener *listener)
610{
Willy Tarreau6ee9f8d2019-08-26 10:55:52 +0200611 HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100612 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100613 if (listener->state == LI_ASSIGNED) {
614 listener->state = LI_INIT;
615 LIST_DEL(&listener->proto_list);
616 listener->proto->nb_listeners--;
Olivier Houchard64213e92019-03-08 18:52:57 +0100617 _HA_ATOMIC_SUB(&jobs, 1);
618 _HA_ATOMIC_SUB(&listeners, 1);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100619 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100620 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau6ee9f8d2019-08-26 10:55:52 +0200621 HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
Willy Tarreau1a64d162007-10-28 22:26:05 +0100622}
623
Willy Tarreaue2711c72019-02-27 15:39:41 +0100624/* Returns a suitable value for a listener's backlog. It uses the listener's,
625 * otherwise the frontend's backlog, otherwise the listener's maxconn,
626 * otherwise the frontend's maxconn, otherwise 1024.
627 */
628int listener_backlog(const struct listener *l)
629{
630 if (l->backlog)
631 return l->backlog;
632
633 if (l->bind_conf->frontend->backlog)
634 return l->bind_conf->frontend->backlog;
635
636 if (l->maxconn)
637 return l->maxconn;
638
639 if (l->bind_conf->frontend->maxconn)
640 return l->bind_conf->frontend->maxconn;
641
642 return 1024;
643}
644
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200645/* This function is called on a read event from a listening socket, corresponding
646 * to an accept. It tries to accept as many connections as possible, and for each
647 * calls the listener's accept handler (generally the frontend's accept handler).
648 */
Willy Tarreauafad0e02012-08-09 14:45:22 +0200649void listener_accept(int fd)
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200650{
651 struct listener *l = fdtab[fd].owner;
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100652 struct proxy *p;
Christopher Faulet102854c2019-04-30 12:17:13 +0200653 unsigned int max_accept;
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100654 int next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100655 int next_feconn = 0;
656 int next_actconn = 0;
Willy Tarreaubb660302014-05-07 19:47:02 +0200657 int expire;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200658 int cfd;
659 int ret;
Willy Tarreau818dca52014-01-31 19:40:19 +0100660#ifdef USE_ACCEPT4
661 static int accept4_broken;
662#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200663
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100664 if (!l)
665 return;
666 p = l->bind_conf->frontend;
Christopher Faulet102854c2019-04-30 12:17:13 +0200667
668 /* if l->maxaccept is -1, then max_accept is UINT_MAX. It is not really
669 * illimited, but it is probably enough.
670 */
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100671 max_accept = l->maxaccept ? l->maxaccept : 1;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200672
Willy Tarreau93e7c002013-10-07 18:51:07 +0200673 if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
674 int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
Willy Tarreau93e7c002013-10-07 18:51:07 +0200675
676 if (unlikely(!max)) {
677 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200678 expire = tick_add(now_ms, next_event_delay(&global.sess_per_sec, global.sps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100679 goto limit_global;
Willy Tarreau93e7c002013-10-07 18:51:07 +0200680 }
681
682 if (max_accept > max)
683 max_accept = max;
684 }
685
686 if (!(l->options & LI_O_UNLIMITED) && global.cps_lim) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200687 int max = freq_ctr_remain(&global.conn_per_sec, global.cps_lim, 0);
688
689 if (unlikely(!max)) {
690 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200691 expire = tick_add(now_ms, next_event_delay(&global.conn_per_sec, global.cps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100692 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200693 }
694
695 if (max_accept > max)
696 max_accept = max;
697 }
Willy Tarreaue43d5322013-10-07 20:01:52 +0200698#ifdef USE_OPENSSL
699 if (!(l->options & LI_O_UNLIMITED) && global.ssl_lim && l->bind_conf && l->bind_conf->is_ssl) {
700 int max = freq_ctr_remain(&global.ssl_per_sec, global.ssl_lim, 0);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200701
Willy Tarreaue43d5322013-10-07 20:01:52 +0200702 if (unlikely(!max)) {
703 /* frontend accept rate limit was reached */
Willy Tarreaue43d5322013-10-07 20:01:52 +0200704 expire = tick_add(now_ms, next_event_delay(&global.ssl_per_sec, global.ssl_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100705 goto limit_global;
Willy Tarreaue43d5322013-10-07 20:01:52 +0200706 }
707
708 if (max_accept > max)
709 max_accept = max;
710 }
711#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200712 if (p && p->fe_sps_lim) {
713 int max = freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0);
714
715 if (unlikely(!max)) {
716 /* frontend accept rate limit was reached */
Willy Tarreau0591bf72019-12-10 12:01:21 +0100717 expire = tick_add(now_ms, next_event_delay(&p->fe_sess_per_sec, p->fe_sps_lim, 0));
718 goto limit_proxy;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200719 }
720
721 if (max_accept > max)
722 max_accept = max;
723 }
724
725 /* Note: if we fail to allocate a connection because of configured
726 * limits, we'll schedule a new attempt worst 1 second later in the
727 * worst case. If we fail due to system limits or temporary resource
728 * shortage, we try again 100ms later in the worst case.
729 */
Christopher Faulet102854c2019-04-30 12:17:13 +0200730 for (; max_accept; next_conn = next_feconn = next_actconn = 0, max_accept--) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200731 struct sockaddr_storage addr;
732 socklen_t laddr = sizeof(addr);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200733 unsigned int count;
Willy Tarreau57cb5062019-03-15 17:16:34 +0100734 __decl_hathreads(unsigned long mask);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200735
Willy Tarreau82c97892019-02-27 19:32:32 +0100736 /* pre-increase the number of connections without going too far.
737 * We process the listener, then the proxy, then the process.
738 * We know which ones to unroll based on the next_xxx value.
739 */
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100740 do {
741 count = l->nbconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100742 if (unlikely(l->maxconn && count >= l->maxconn)) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100743 /* the listener was marked full or another
744 * thread is going to do it.
745 */
746 next_conn = 0;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100747 listener_full(l);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100748 goto end;
749 }
750 next_conn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000751 } while (!_HA_ATOMIC_CAS(&l->nbconn, (int *)(&count), next_conn));
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100752
Willy Tarreau82c97892019-02-27 19:32:32 +0100753 if (p) {
754 do {
755 count = p->feconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100756 if (unlikely(count >= p->maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100757 /* the frontend was marked full or another
758 * thread is going to do it.
759 */
760 next_feconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100761 expire = TICK_ETERNITY;
762 goto limit_proxy;
Willy Tarreau82c97892019-02-27 19:32:32 +0100763 }
764 next_feconn = count + 1;
Olivier Houchard64213e92019-03-08 18:52:57 +0100765 } while (!_HA_ATOMIC_CAS(&p->feconn, &count, next_feconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200766 }
767
Willy Tarreau82c97892019-02-27 19:32:32 +0100768 if (!(l->options & LI_O_UNLIMITED)) {
769 do {
770 count = actconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100771 if (unlikely(count >= global.maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100772 /* the process was marked full or another
773 * thread is going to do it.
774 */
775 next_actconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100776 expire = tick_add(now_ms, 1000); /* try again in 1 second */
777 goto limit_global;
Willy Tarreau82c97892019-02-27 19:32:32 +0100778 }
779 next_actconn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000780 } while (!_HA_ATOMIC_CAS(&actconn, (int *)(&count), next_actconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200781 }
782
William Lallemand2fe7dd02018-09-11 16:51:29 +0200783 /* with sockpair@ we don't want to do an accept */
784 if (unlikely(l->addr.ss_family == AF_CUST_SOCKPAIR)) {
785 if ((cfd = recv_fd_uxst(fd)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100786 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau888d5672019-01-27 18:34:12 +0100787 /* just like with UNIX sockets, only the family is filled */
788 addr.ss_family = AF_UNIX;
789 laddr = sizeof(addr.ss_family);
William Lallemand2fe7dd02018-09-11 16:51:29 +0200790 } else
791
Willy Tarreau1bc4aab2012-10-08 20:11:03 +0200792#ifdef USE_ACCEPT4
Willy Tarreau818dca52014-01-31 19:40:19 +0100793 /* only call accept4() if it's known to be safe, otherwise
794 * fallback to the legacy accept() + fcntl().
795 */
796 if (unlikely(accept4_broken ||
William Lallemandd9138002018-11-27 12:02:39 +0100797 ((cfd = accept4(fd, (struct sockaddr *)&addr, &laddr, SOCK_NONBLOCK)) == -1 &&
Willy Tarreau818dca52014-01-31 19:40:19 +0100798 (errno == ENOSYS || errno == EINVAL || errno == EBADF) &&
799 (accept4_broken = 1))))
800#endif
Willy Tarreau6b3b0d42012-10-22 19:32:55 +0200801 if ((cfd = accept(fd, (struct sockaddr *)&addr, &laddr)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100802 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau818dca52014-01-31 19:40:19 +0100803
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200804 if (unlikely(cfd == -1)) {
805 switch (errno) {
806 case EAGAIN:
Willy Tarreau20aeb1c2019-12-10 08:42:21 +0100807 if (fdtab[fd].ev & (FD_POLL_HUP|FD_POLL_ERR)) {
Willy Tarreaubb660302014-05-07 19:47:02 +0200808 /* the listening socket might have been disabled in a shared
809 * process and we're a collateral victim. We'll just pause for
810 * a while in case it comes back. In the mean time, we need to
811 * clear this sticky flag.
812 */
Willy Tarreau20aeb1c2019-12-10 08:42:21 +0100813 _HA_ATOMIC_AND(&fdtab[fd].ev, ~(FD_POLL_HUP|FD_POLL_ERR));
Willy Tarreaubb660302014-05-07 19:47:02 +0200814 goto transient_error;
815 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200816 goto end; /* nothing more to accept */
Willy Tarreaubb660302014-05-07 19:47:02 +0200817 case EINVAL:
818 /* might be trying to accept on a shut fd (eg: soft stop) */
819 goto transient_error;
Willy Tarreaua593ec52014-01-20 21:21:30 +0100820 case EINTR:
821 case ECONNABORTED:
Olivier Houchard64213e92019-03-08 18:52:57 +0100822 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100823 if (p)
Olivier Houchard64213e92019-03-08 18:52:57 +0100824 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100825 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +0100826 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreaua593ec52014-01-20 21:21:30 +0100827 continue;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200828 case ENFILE:
829 if (p)
830 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100831 "Proxy %s reached system FD limit (maxsock=%d). Please check system tunables.\n",
832 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200833 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200834 case EMFILE:
835 if (p)
836 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100837 "Proxy %s reached process FD limit (maxsock=%d). Please check 'ulimit-n' and restart.\n",
838 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200839 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200840 case ENOBUFS:
841 case ENOMEM:
842 if (p)
843 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100844 "Proxy %s reached system memory limit (maxsock=%d). Please check system tunables.\n",
845 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200846 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200847 default:
Willy Tarreaua593ec52014-01-20 21:21:30 +0100848 /* unexpected result, let's give up and let other tasks run */
Willy Tarreau92079932019-12-10 09:30:05 +0100849 max_accept = 0;
850 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200851 }
852 }
853
William Lallemandd9138002018-11-27 12:02:39 +0100854 /* we don't want to leak the FD upon reload if it's in the master */
855 if (unlikely(master == 1))
856 fcntl(cfd, F_SETFD, FD_CLOEXEC);
857
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100858 /* The connection was accepted, it must be counted as such */
859 if (l->counters)
860 HA_ATOMIC_UPDATE_MAX(&l->counters->conn_max, next_conn);
861
Willy Tarreau82c97892019-02-27 19:32:32 +0100862 if (p)
863 HA_ATOMIC_UPDATE_MAX(&p->fe_counters.conn_max, next_feconn);
864
865 proxy_inc_fe_conn_ctr(l, p);
866
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100867 if (!(l->options & LI_O_UNLIMITED)) {
868 count = update_freq_ctr(&global.conn_per_sec, 1);
869 HA_ATOMIC_UPDATE_MAX(&global.cps_max, count);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100870 }
871
Willy Tarreau64a9c052019-04-12 15:27:17 +0200872 _HA_ATOMIC_ADD(&activity[tid].accepted, 1);
873
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200874 if (unlikely(cfd >= global.maxsock)) {
875 send_log(p, LOG_EMERG,
876 "Proxy %s reached the configured maximum connection limit. Please check the global 'maxconn' value.\n",
877 p->id);
878 close(cfd);
Willy Tarreau0591bf72019-12-10 12:01:21 +0100879 expire = tick_add(now_ms, 1000); /* try again in 1 second */
880 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200881 }
882
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100883 /* past this point, l->accept() will automatically decrement
Willy Tarreau82c97892019-02-27 19:32:32 +0100884 * l->nbconn, feconn and actconn once done. Setting next_*conn=0
885 * allows the error path not to rollback on nbconn. It's more
886 * convenient than duplicating all exit labels.
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100887 */
888 next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100889 next_feconn = 0;
890 next_actconn = 0;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200891
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100892#if defined(USE_THREAD)
Willy Tarreau897e2c52019-03-13 15:03:53 +0100893 mask = thread_mask(l->bind_conf->bind_thread) & all_threads_mask;
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100894 if (atleast2(mask) && (global.tune.options & GTUNE_LISTENER_MQ)) {
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100895 struct accept_queue_ring *ring;
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100896 unsigned int t, t0, t1, t2;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100897
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100898 /* The principle is that we have two running indexes,
899 * each visiting in turn all threads bound to this
900 * listener. The connection will be assigned to the one
901 * with the least connections, and the other one will
902 * be updated. This provides a good fairness on short
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100903 * connections (round robin) and on long ones (conn
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100904 * count), without ever missing any idle thread.
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100905 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100906
907 /* keep a copy for the final update. thr_idx is composite
908 * and made of (t2<<16) + t1.
909 */
Willy Tarreau0cf33172019-03-06 15:26:33 +0100910 t0 = l->thr_idx;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100911 do {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100912 unsigned long m1, m2;
913 int q1, q2;
914
915 t2 = t1 = t0;
916 t2 >>= 16;
917 t1 &= 0xFFFF;
918
919 /* t1 walks low to high bits ;
920 * t2 walks high to low.
921 */
922 m1 = mask >> t1;
923 m2 = mask & (t2 ? nbits(t2 + 1) : ~0UL);
924
Willy Tarreau85d04242019-04-16 18:09:13 +0200925 if (unlikely(!(m1 & 1))) {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100926 m1 &= ~1UL;
927 if (!m1) {
928 m1 = mask;
929 t1 = 0;
930 }
931 t1 += my_ffsl(m1) - 1;
932 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100933
Willy Tarreau85d04242019-04-16 18:09:13 +0200934 if (unlikely(!(m2 & (1UL << t2)) || t1 == t2)) {
935 /* highest bit not set */
936 if (!m2)
937 m2 = mask;
938
939 t2 = my_flsl(m2) - 1;
940 }
941
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100942 /* now we have two distinct thread IDs belonging to the mask */
943 q1 = accept_queue_rings[t1].tail - accept_queue_rings[t1].head + ACCEPT_QUEUE_SIZE;
944 if (q1 >= ACCEPT_QUEUE_SIZE)
945 q1 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100946
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100947 q2 = accept_queue_rings[t2].tail - accept_queue_rings[t2].head + ACCEPT_QUEUE_SIZE;
948 if (q2 >= ACCEPT_QUEUE_SIZE)
949 q2 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100950
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100951 /* we have 3 possibilities now :
952 * q1 < q2 : t1 is less loaded than t2, so we pick it
953 * and update t2 (since t1 might still be
954 * lower than another thread)
955 * q1 > q2 : t2 is less loaded than t1, so we pick it
956 * and update t1 (since t2 might still be
957 * lower than another thread)
958 * q1 = q2 : both are equally loaded, thus we pick t1
959 * and update t1 as it will become more loaded
960 * than t2.
961 */
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100962
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100963 q1 += l->thr_conn[t1];
964 q2 += l->thr_conn[t2];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100965
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100966 if (q1 - q2 < 0) {
967 t = t1;
968 t2 = t2 ? t2 - 1 : LONGBITS - 1;
969 }
970 else if (q1 - q2 > 0) {
971 t = t2;
972 t1++;
973 if (t1 >= LONGBITS)
974 t1 = 0;
975 }
976 else {
977 t = t1;
978 t1++;
979 if (t1 >= LONGBITS)
980 t1 = 0;
981 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100982
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100983 /* new value for thr_idx */
984 t1 += (t2 << 16);
Olivier Houchard64213e92019-03-08 18:52:57 +0100985 } while (unlikely(!_HA_ATOMIC_CAS(&l->thr_idx, &t0, t1)));
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100986
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100987 /* We successfully selected the best thread "t" for this
988 * connection. We use deferred accepts even if it's the
989 * local thread because tests show that it's the best
990 * performing model, likely due to better cache locality
991 * when processing this loop.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100992 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100993 ring = &accept_queue_rings[t];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100994 if (accept_queue_push_mp(ring, cfd, l, &addr, laddr)) {
Olivier Houchard64213e92019-03-08 18:52:57 +0100995 _HA_ATOMIC_ADD(&activity[t].accq_pushed, 1);
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200996 tasklet_wakeup(ring->tasklet);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100997 continue;
998 }
999 /* If the ring is full we do a synchronous accept on
1000 * the local thread here.
Willy Tarreaue0e9c482019-01-27 15:37:19 +01001001 */
Olivier Houchard64213e92019-03-08 18:52:57 +01001002 _HA_ATOMIC_ADD(&activity[t].accq_full, 1);
Willy Tarreaue0e9c482019-01-27 15:37:19 +01001003 }
1004#endif // USE_THREAD
1005
Olivier Houchard64213e92019-03-08 18:52:57 +01001006 _HA_ATOMIC_ADD(&l->thr_conn[tid], 1);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001007 ret = l->accept(l, cfd, &addr);
1008 if (unlikely(ret <= 0)) {
Willy Tarreau87b09662015-04-03 00:22:06 +02001009 /* The connection was closed by stream_accept(). Either
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001010 * we just have to ignore it (ret == 0) or it's a critical
1011 * error due to a resource shortage, and we must stop the
1012 * listener (ret < 0).
1013 */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001014 if (ret == 0) /* successful termination */
1015 continue;
1016
Willy Tarreaubb660302014-05-07 19:47:02 +02001017 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001018 }
1019
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001020 /* increase the per-process number of cumulated sessions, this
1021 * may only be done once l->accept() has accepted the connection.
1022 */
Willy Tarreau93e7c002013-10-07 18:51:07 +02001023 if (!(l->options & LI_O_UNLIMITED)) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001024 count = update_freq_ctr(&global.sess_per_sec, 1);
1025 HA_ATOMIC_UPDATE_MAX(&global.sps_max, count);
Willy Tarreau93e7c002013-10-07 18:51:07 +02001026 }
Willy Tarreaue43d5322013-10-07 20:01:52 +02001027#ifdef USE_OPENSSL
1028 if (!(l->options & LI_O_UNLIMITED) && l->bind_conf && l->bind_conf->is_ssl) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001029 count = update_freq_ctr(&global.ssl_per_sec, 1);
1030 HA_ATOMIC_UPDATE_MAX(&global.ssl_max, count);
Willy Tarreaue43d5322013-10-07 20:01:52 +02001031 }
1032#endif
Willy Tarreau93e7c002013-10-07 18:51:07 +02001033
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001034 } /* end of for (max_accept--) */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001035
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001036 end:
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001037 if (next_conn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001038 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau741b4d62019-02-25 15:02:04 +01001039
Willy Tarreau82c97892019-02-27 19:32:32 +01001040 if (p && next_feconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001041 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001042
1043 if (next_actconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001044 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001045
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001046 if ((l->state == LI_FULL && (!l->maxconn || l->nbconn < l->maxconn)) ||
Willy Tarreau82c97892019-02-27 19:32:32 +01001047 (l->state == LI_LIMITED && ((!p || p->feconn < p->maxconn) && (actconn < global.maxconn)))) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001048 /* at least one thread has to this when quitting */
1049 resume_listener(l);
1050
1051 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001052 dequeue_all_listeners();
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001053
Olivier Houchard859dc802019-08-08 15:47:21 +02001054 if (p && !MT_LIST_ISEMPTY(&p->listener_queue) &&
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001055 (!p->fe_sps_lim || freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0) > 0))
Willy Tarreau241797a2019-12-10 14:10:52 +01001056 dequeue_proxy_listeners(p);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001057 }
Willy Tarreau4c044e22019-12-05 07:40:32 +01001058
Willy Tarreau92079932019-12-10 09:30:05 +01001059 /* Now it's getting tricky. The listener was supposed to be in LI_READY
1060 * state but in the mean time we might have changed it to LI_FULL or
1061 * LI_LIMITED, and another thread might also have turned it to
1062 * LI_PAUSED, LI_LISTEN or even LI_INI when stopping a proxy. We must
1063 * be certain to keep the FD enabled when in the READY state but we
1064 * must also stop it for other states that we might have switched to
1065 * while others re-enabled polling.
1066 */
1067 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
1068 if (l->state == LI_READY) {
1069 if (max_accept > 0)
1070 fd_cant_recv(fd);
1071 else
1072 fd_done_recv(fd);
1073 } else if (l->state > LI_ASSIGNED) {
Willy Tarreau4c044e22019-12-05 07:40:32 +01001074 fd_stop_recv(l->fd);
Willy Tarreau92079932019-12-10 09:30:05 +01001075 }
1076 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau0591bf72019-12-10 12:01:21 +01001077 return;
1078
1079 transient_error:
1080 /* pause the listener for up to 100 ms */
1081 expire = tick_add(now_ms, 100);
1082
1083 limit_global:
1084 /* (re-)queue the listener to the global queue and set it to expire no
1085 * later than <expire> ahead. The listener turns to LI_LIMITED.
1086 */
1087 limit_listener(l, &global_listener_queue);
1088 task_schedule(global_listener_queue_task, expire);
1089 goto end;
1090
1091 limit_proxy:
1092 /* (re-)queue the listener to the proxy's queue and set it to expire no
1093 * later than <expire> ahead. The listener turns to LI_LIMITED.
1094 */
1095 limit_listener(l, &p->listener_queue);
1096 task_schedule(p->task, expire);
1097 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001098}
1099
Willy Tarreau05f50472017-09-15 09:19:58 +02001100/* Notify the listener that a connection initiated from it was released. This
1101 * is used to keep the connection count consistent and to possibly re-open
1102 * listening when it was limited.
1103 */
1104void listener_release(struct listener *l)
1105{
1106 struct proxy *fe = l->bind_conf->frontend;
1107
1108 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +01001109 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001110 if (fe)
Olivier Houchard64213e92019-03-08 18:52:57 +01001111 _HA_ATOMIC_SUB(&fe->feconn, 1);
1112 _HA_ATOMIC_SUB(&l->nbconn, 1);
1113 _HA_ATOMIC_SUB(&l->thr_conn[tid], 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001114
1115 if (l->state == LI_FULL || l->state == LI_LIMITED)
Willy Tarreau05f50472017-09-15 09:19:58 +02001116 resume_listener(l);
1117
1118 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001119 dequeue_all_listeners();
Willy Tarreau05f50472017-09-15 09:19:58 +02001120
Olivier Houchard859dc802019-08-08 15:47:21 +02001121 if (!MT_LIST_ISEMPTY(&fe->listener_queue) &&
Willy Tarreau05f50472017-09-15 09:19:58 +02001122 (!fe->fe_sps_lim || freq_ctr_remain(&fe->fe_sess_per_sec, fe->fe_sps_lim, 0) > 0))
Willy Tarreau241797a2019-12-10 14:10:52 +01001123 dequeue_proxy_listeners(fe);
Willy Tarreau05f50472017-09-15 09:19:58 +02001124}
1125
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001126/* resume listeners waiting in the local listener queue. They are still in LI_LIMITED state */
1127static struct task *listener_queue_process(struct task *t, void *context, unsigned short state)
1128{
1129 struct work_list *wl = context;
1130 struct listener *l;
1131
Olivier Houchard859dc802019-08-08 15:47:21 +02001132 while ((l = MT_LIST_POP(&wl->head, struct listener *, wait_queue))) {
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001133 /* The listeners are still in the LI_LIMITED state */
1134 resume_listener(l);
1135 }
1136 return t;
1137}
1138
1139/* Initializes the listener queues. Returns 0 on success, otherwise ERR_* flags */
1140static int listener_queue_init()
1141{
1142 local_listener_queue = work_list_create(global.nbthread, listener_queue_process, NULL);
1143 if (!local_listener_queue) {
1144 ha_alert("Out of memory while initializing listener queues.\n");
1145 return ERR_FATAL|ERR_ABORT;
1146 }
1147 return 0;
1148}
1149
1150static void listener_queue_deinit()
1151{
1152 work_list_destroy(local_listener_queue, global.nbthread);
1153}
1154
1155REGISTER_CONFIG_POSTPARSER("multi-threaded listener queue", listener_queue_init);
1156REGISTER_POST_DEINIT(listener_queue_deinit);
1157
Willy Tarreau26982662012-09-12 23:17:10 +02001158/*
1159 * Registers the bind keyword list <kwl> as a list of valid keywords for next
1160 * parsing sessions.
1161 */
1162void bind_register_keywords(struct bind_kw_list *kwl)
1163{
1164 LIST_ADDQ(&bind_keywords.list, &kwl->list);
1165}
1166
1167/* Return a pointer to the bind keyword <kw>, or NULL if not found. If the
1168 * keyword is found with a NULL ->parse() function, then an attempt is made to
1169 * find one with a valid ->parse() function. This way it is possible to declare
1170 * platform-dependant, known keywords as NULL, then only declare them as valid
1171 * if some options are met. Note that if the requested keyword contains an
1172 * opening parenthesis, everything from this point is ignored.
1173 */
1174struct bind_kw *bind_find_kw(const char *kw)
1175{
1176 int index;
1177 const char *kwend;
1178 struct bind_kw_list *kwl;
1179 struct bind_kw *ret = NULL;
1180
1181 kwend = strchr(kw, '(');
1182 if (!kwend)
1183 kwend = kw + strlen(kw);
1184
1185 list_for_each_entry(kwl, &bind_keywords.list, list) {
1186 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1187 if ((strncmp(kwl->kw[index].kw, kw, kwend - kw) == 0) &&
1188 kwl->kw[index].kw[kwend-kw] == 0) {
1189 if (kwl->kw[index].parse)
1190 return &kwl->kw[index]; /* found it !*/
1191 else
1192 ret = &kwl->kw[index]; /* may be OK */
1193 }
1194 }
1195 }
1196 return ret;
1197}
1198
Willy Tarreau8638f482012-09-18 18:01:17 +02001199/* Dumps all registered "bind" keywords to the <out> string pointer. The
1200 * unsupported keywords are only dumped if their supported form was not
1201 * found.
1202 */
1203void bind_dump_kws(char **out)
1204{
1205 struct bind_kw_list *kwl;
1206 int index;
1207
1208 *out = NULL;
1209 list_for_each_entry(kwl, &bind_keywords.list, list) {
1210 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1211 if (kwl->kw[index].parse ||
1212 bind_find_kw(kwl->kw[index].kw) == &kwl->kw[index]) {
Willy Tarreau51fb7652012-09-18 18:24:39 +02001213 memprintf(out, "%s[%4s] %s%s%s\n", *out ? *out : "",
1214 kwl->scope,
Willy Tarreau8638f482012-09-18 18:01:17 +02001215 kwl->kw[index].kw,
Willy Tarreau51fb7652012-09-18 18:24:39 +02001216 kwl->kw[index].skip ? " <arg>" : "",
1217 kwl->kw[index].parse ? "" : " (not supported)");
Willy Tarreau8638f482012-09-18 18:01:17 +02001218 }
1219 }
1220 }
1221}
1222
Willy Tarreau645513a2010-05-24 20:55:15 +02001223/************************************************************************/
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001224/* All supported sample and ACL keywords must be declared here. */
Willy Tarreau645513a2010-05-24 20:55:15 +02001225/************************************************************************/
1226
Willy Tarreaua5e37562011-12-16 17:06:15 +01001227/* set temp integer to the number of connexions to the same listening socket */
Willy Tarreau645513a2010-05-24 20:55:15 +02001228static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001229smp_fetch_dconn(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau645513a2010-05-24 20:55:15 +02001230{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001231 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001232 smp->data.u.sint = smp->sess->listener->nbconn;
Willy Tarreau645513a2010-05-24 20:55:15 +02001233 return 1;
1234}
1235
Willy Tarreaua5e37562011-12-16 17:06:15 +01001236/* set temp integer to the id of the socket (listener) */
Willy Tarreau645513a2010-05-24 20:55:15 +02001237static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001238smp_fetch_so_id(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau37406352012-04-23 16:16:37 +02001239{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001240 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001241 smp->data.u.sint = smp->sess->listener->luid;
Willy Tarreau645513a2010-05-24 20:55:15 +02001242 return 1;
1243}
1244
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001245/* parse the "accept-proxy" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001246static int bind_parse_accept_proxy(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001247{
1248 struct listener *l;
1249
Willy Tarreau4348fad2012-09-20 16:48:07 +02001250 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001251 l->options |= LI_O_ACC_PROXY;
1252
1253 return 0;
1254}
1255
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001256/* parse the "accept-netscaler-cip" bind keyword */
1257static int bind_parse_accept_netscaler_cip(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1258{
1259 struct listener *l;
1260 uint32_t val;
1261
1262 if (!*args[cur_arg + 1]) {
1263 memprintf(err, "'%s' : missing value", args[cur_arg]);
1264 return ERR_ALERT | ERR_FATAL;
1265 }
1266
1267 val = atol(args[cur_arg + 1]);
1268 if (val <= 0) {
Willy Tarreaue2711c72019-02-27 15:39:41 +01001269 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001270 return ERR_ALERT | ERR_FATAL;
1271 }
1272
1273 list_for_each_entry(l, &conf->listeners, by_bind) {
1274 l->options |= LI_O_ACC_CIP;
1275 conf->ns_cip_magic = val;
1276 }
1277
1278 return 0;
1279}
1280
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001281/* parse the "backlog" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001282static int bind_parse_backlog(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001283{
1284 struct listener *l;
1285 int val;
1286
1287 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001288 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001289 return ERR_ALERT | ERR_FATAL;
1290 }
1291
1292 val = atol(args[cur_arg + 1]);
Willy Tarreaue2711c72019-02-27 15:39:41 +01001293 if (val < 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001294 memprintf(err, "'%s' : invalid value %d, must be > 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001295 return ERR_ALERT | ERR_FATAL;
1296 }
1297
Willy Tarreau4348fad2012-09-20 16:48:07 +02001298 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001299 l->backlog = val;
1300
1301 return 0;
1302}
1303
1304/* parse the "id" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001305static int bind_parse_id(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001306{
1307 struct eb32_node *node;
Willy Tarreau4348fad2012-09-20 16:48:07 +02001308 struct listener *l, *new;
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001309 char *error;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001310
Willy Tarreau4348fad2012-09-20 16:48:07 +02001311 if (conf->listeners.n != conf->listeners.p) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001312 memprintf(err, "'%s' can only be used with a single socket", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001313 return ERR_ALERT | ERR_FATAL;
1314 }
1315
1316 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001317 memprintf(err, "'%s' : expects an integer argument", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001318 return ERR_ALERT | ERR_FATAL;
1319 }
1320
Willy Tarreau4348fad2012-09-20 16:48:07 +02001321 new = LIST_NEXT(&conf->listeners, struct listener *, by_bind);
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001322 new->luid = strtol(args[cur_arg + 1], &error, 10);
1323 if (*error != '\0') {
1324 memprintf(err, "'%s' : expects an integer argument, found '%s'", args[cur_arg], args[cur_arg + 1]);
1325 return ERR_ALERT | ERR_FATAL;
1326 }
Willy Tarreau4348fad2012-09-20 16:48:07 +02001327 new->conf.id.key = new->luid;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001328
Willy Tarreau4348fad2012-09-20 16:48:07 +02001329 if (new->luid <= 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001330 memprintf(err, "'%s' : custom id has to be > 0", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001331 return ERR_ALERT | ERR_FATAL;
1332 }
1333
Willy Tarreau4348fad2012-09-20 16:48:07 +02001334 node = eb32_lookup(&px->conf.used_listener_id, new->luid);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001335 if (node) {
1336 l = container_of(node, struct listener, conf.id);
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001337 memprintf(err, "'%s' : custom id %d already used at %s:%d ('bind %s')",
1338 args[cur_arg], l->luid, l->bind_conf->file, l->bind_conf->line,
1339 l->bind_conf->arg);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001340 return ERR_ALERT | ERR_FATAL;
1341 }
1342
Willy Tarreau4348fad2012-09-20 16:48:07 +02001343 eb32_insert(&px->conf.used_listener_id, &new->conf.id);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001344 return 0;
1345}
1346
1347/* parse the "maxconn" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001348static int bind_parse_maxconn(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001349{
1350 struct listener *l;
1351 int val;
1352
1353 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001354 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001355 return ERR_ALERT | ERR_FATAL;
1356 }
1357
1358 val = atol(args[cur_arg + 1]);
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001359 if (val < 0) {
1360 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001361 return ERR_ALERT | ERR_FATAL;
1362 }
1363
Willy Tarreau4348fad2012-09-20 16:48:07 +02001364 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001365 l->maxconn = val;
1366
1367 return 0;
1368}
1369
1370/* parse the "name" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001371static int bind_parse_name(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001372{
1373 struct listener *l;
1374
1375 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001376 memprintf(err, "'%s' : missing name", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001377 return ERR_ALERT | ERR_FATAL;
1378 }
1379
Willy Tarreau4348fad2012-09-20 16:48:07 +02001380 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001381 l->name = strdup(args[cur_arg + 1]);
1382
1383 return 0;
1384}
1385
1386/* parse the "nice" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001387static int bind_parse_nice(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001388{
1389 struct listener *l;
1390 int val;
1391
1392 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001393 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001394 return ERR_ALERT | ERR_FATAL;
1395 }
1396
1397 val = atol(args[cur_arg + 1]);
1398 if (val < -1024 || val > 1024) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001399 memprintf(err, "'%s' : invalid value %d, allowed range is -1024..1024", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001400 return ERR_ALERT | ERR_FATAL;
1401 }
1402
Willy Tarreau4348fad2012-09-20 16:48:07 +02001403 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001404 l->nice = val;
1405
1406 return 0;
1407}
1408
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001409/* parse the "process" bind keyword */
1410static int bind_parse_process(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1411{
Christopher Fauletc644fa92017-11-23 22:44:11 +01001412 char *slash;
1413 unsigned long proc = 0, thread = 0;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001414
Christopher Fauletc644fa92017-11-23 22:44:11 +01001415 if ((slash = strchr(args[cur_arg + 1], '/')) != NULL)
1416 *slash = 0;
1417
Willy Tarreauff9c9142019-02-07 10:39:36 +01001418 if (parse_process_number(args[cur_arg + 1], &proc, MAX_PROCS, NULL, err)) {
Christopher Fauletf1f0c5f2017-11-22 12:06:43 +01001419 memprintf(err, "'%s' : %s", args[cur_arg], *err);
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001420 return ERR_ALERT | ERR_FATAL;
1421 }
1422
Christopher Fauletc644fa92017-11-23 22:44:11 +01001423 if (slash) {
Willy Tarreauc9a82e42019-01-26 13:25:14 +01001424 if (parse_process_number(slash+1, &thread, MAX_THREADS, NULL, err)) {
Christopher Fauletc644fa92017-11-23 22:44:11 +01001425 memprintf(err, "'%s' : %s", args[cur_arg], *err);
1426 return ERR_ALERT | ERR_FATAL;
1427 }
1428 *slash = '/';
1429 }
1430
1431 conf->bind_proc |= proc;
Willy Tarreaua36b3242019-02-02 13:14:34 +01001432 conf->bind_thread |= thread;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001433 return 0;
1434}
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001435
Christopher Fauleta717b992018-04-10 14:43:00 +02001436/* parse the "proto" bind keyword */
1437static int bind_parse_proto(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1438{
1439 struct ist proto;
1440
1441 if (!*args[cur_arg + 1]) {
1442 memprintf(err, "'%s' : missing value", args[cur_arg]);
1443 return ERR_ALERT | ERR_FATAL;
1444 }
1445
1446 proto = ist2(args[cur_arg + 1], strlen(args[cur_arg + 1]));
1447 conf->mux_proto = get_mux_proto(proto);
1448 if (!conf->mux_proto) {
1449 memprintf(err, "'%s' : unknown MUX protocol '%s'", args[cur_arg], args[cur_arg+1]);
1450 return ERR_ALERT | ERR_FATAL;
1451 }
Christopher Fauleta717b992018-04-10 14:43:00 +02001452 return 0;
1453}
1454
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001455/* config parser for global "tune.listener.multi-queue", accepts "on" or "off" */
1456static int cfg_parse_tune_listener_mq(char **args, int section_type, struct proxy *curpx,
1457 struct proxy *defpx, const char *file, int line,
1458 char **err)
1459{
1460 if (too_many_args(1, args, err, NULL))
1461 return -1;
1462
1463 if (strcmp(args[1], "on") == 0)
1464 global.tune.options |= GTUNE_LISTENER_MQ;
1465 else if (strcmp(args[1], "off") == 0)
1466 global.tune.options &= ~GTUNE_LISTENER_MQ;
1467 else {
1468 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
1469 return -1;
1470 }
1471 return 0;
1472}
1473
Willy Tarreau61612d42012-04-19 18:42:05 +02001474/* Note: must not be declared <const> as its list will be overwritten.
1475 * Please take care of keeping this list alphabetically sorted.
1476 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001477static struct sample_fetch_kw_list smp_kws = {ILH, {
Thierry FOURNIER07ee64e2015-07-06 23:43:03 +02001478 { "dst_conn", smp_fetch_dconn, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
1479 { "so_id", smp_fetch_so_id, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001480 { /* END */ },
1481}};
1482
Willy Tarreau0108d902018-11-25 19:14:37 +01001483INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
1484
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001485/* Note: must not be declared <const> as its list will be overwritten.
1486 * Please take care of keeping this list alphabetically sorted.
1487 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001488static struct acl_kw_list acl_kws = {ILH, {
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001489 { /* END */ },
Willy Tarreau645513a2010-05-24 20:55:15 +02001490}};
1491
Willy Tarreau0108d902018-11-25 19:14:37 +01001492INITCALL1(STG_REGISTER, acl_register_keywords, &acl_kws);
1493
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001494/* Note: must not be declared <const> as its list will be overwritten.
1495 * Please take care of keeping this list alphabetically sorted, doing so helps
1496 * all code contributors.
1497 * Optional keywords are also declared with a NULL ->parse() function so that
1498 * the config parser can report an appropriate error when a known keyword was
1499 * not enabled.
1500 */
Willy Tarreau51fb7652012-09-18 18:24:39 +02001501static struct bind_kw_list bind_kws = { "ALL", { }, {
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001502 { "accept-netscaler-cip", bind_parse_accept_netscaler_cip, 1 }, /* enable NetScaler Client IP insertion protocol */
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001503 { "accept-proxy", bind_parse_accept_proxy, 0 }, /* enable PROXY protocol */
1504 { "backlog", bind_parse_backlog, 1 }, /* set backlog of listening socket */
1505 { "id", bind_parse_id, 1 }, /* set id of listening socket */
1506 { "maxconn", bind_parse_maxconn, 1 }, /* set maxconn of listening socket */
1507 { "name", bind_parse_name, 1 }, /* set name of listening socket */
1508 { "nice", bind_parse_nice, 1 }, /* set nice of listening socket */
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001509 { "process", bind_parse_process, 1 }, /* set list of allowed process for this socket */
Christopher Fauleta717b992018-04-10 14:43:00 +02001510 { "proto", bind_parse_proto, 1 }, /* set the proto to use for all incoming connections */
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001511 { /* END */ },
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001512}};
1513
Willy Tarreau0108d902018-11-25 19:14:37 +01001514INITCALL1(STG_REGISTER, bind_register_keywords, &bind_kws);
1515
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001516/* config keyword parsers */
1517static struct cfg_kw_list cfg_kws = {ILH, {
1518 { CFG_GLOBAL, "tune.listener.multi-queue", cfg_parse_tune_listener_mq },
1519 { 0, NULL, NULL }
1520}};
1521
1522INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
1523
Willy Tarreau645513a2010-05-24 20:55:15 +02001524/*
1525 * Local variables:
1526 * c-indent-level: 8
1527 * c-basic-offset: 8
1528 * End:
1529 */