blob: 2141017cbf7160ec6e22c18af74303d242cc16e1 [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 Tarreau413e9262019-07-11 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
54 * moving them to the remote queues and waking up the associated task.
55 */
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 Tarreau1efafce2019-01-27 15:37:19 +0100197 task_wakeup(t, TASK_WOKEN_IO);
198
199 return t;
200}
201
202/* Initializes the accept-queues. Returns 0 on success, otherwise ERR_* flags */
203static int accept_queue_init()
204{
205 struct task *t;
206 int i;
207
208 for (i = 0; i < global.nbthread; i++) {
209 t = task_new(1UL << i);
210 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 Tarreau0d858442019-04-12 15:25:04 +0200214 t->nice = -1024;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100215 t->process = accept_queue_process;
216 t->context = &accept_queue_rings[i];
217 accept_queue_rings[i].task = t;
218 }
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);
Willy Tarreau01abd022019-02-28 10:27:18 +0100276 LIST_DEL_LOCKED(&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 Tarreauab46c782020-09-23 17:17:22 +0200298 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
299 !(proc_mask(l->bind_conf->bind_proc) & pid_bit))
300 goto end;
301
Willy Tarreau092d8652014-07-07 20:22:12 +0200302 if (l->proto->pause) {
303 /* Returns < 0 in case of failure, 0 if the listener
304 * was totally stopped, or > 0 if correctly paused.
305 */
306 int ret = l->proto->pause(l);
Willy Tarreaube58c382011-07-24 18:28:10 +0200307
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200308 if (ret < 0) {
309 ret = 0;
310 goto end;
311 }
Willy Tarreau092d8652014-07-07 20:22:12 +0200312 else if (ret == 0)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200313 goto end;
Willy Tarreaub3fb60b2012-10-04 08:56:31 +0200314 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200315
Willy Tarreau01abd022019-02-28 10:27:18 +0100316 LIST_DEL_LOCKED(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200317
Willy Tarreau49b046d2012-08-09 12:11:58 +0200318 fd_stop_recv(l->fd);
Willy Tarreaube58c382011-07-24 18:28:10 +0200319 l->state = LI_PAUSED;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200320 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100321 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200322 return ret;
Willy Tarreaube58c382011-07-24 18:28:10 +0200323}
324
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200325/* This function tries to resume a temporarily disabled listener. Paused, full,
326 * limited and disabled listeners are handled, which means that this function
327 * may replace enable_listener(). The resulting state will either be LI_READY
328 * or LI_FULL. 0 is returned in case of failure to resume (eg: dead socket).
Willy Tarreauae302532014-05-07 19:22:24 +0200329 * Listeners bound to a different process are not woken up unless we're in
Willy Tarreauaf2fd582015-04-14 12:07:16 +0200330 * foreground mode, and are ignored. If the listener was only in the assigned
331 * state, it's totally rebound. This can happen if a pause() has completely
332 * stopped it. If the resume fails, 0 is returned and an error might be
333 * displayed.
Willy Tarreaube58c382011-07-24 18:28:10 +0200334 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100335int resume_listener(struct listener *l)
Willy Tarreaube58c382011-07-24 18:28:10 +0200336{
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200337 int ret = 1;
338
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100339 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200340
Willy Tarreau413e9262019-07-11 10:08:31 +0200341 /* check that another thread didn't to the job in parallel (e.g. at the
342 * end of listen_accept() while we'd come from dequeue_all_listeners().
343 */
344 if (LIST_ADDED(&l->wait_queue))
345 goto end;
346
William Lallemand095ba4c2017-06-01 17:38:50 +0200347 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau6daac192019-02-02 17:39:53 +0100348 !(proc_mask(l->bind_conf->bind_proc) & pid_bit))
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200349 goto end;
Willy Tarreau3569df32017-03-15 12:47:46 +0100350
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200351 if (l->state == LI_ASSIGNED) {
352 char msg[100];
353 int err;
354
355 err = l->proto->bind(l, msg, sizeof(msg));
356 if (err & ERR_ALERT)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100357 ha_alert("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200358 else if (err & ERR_WARN)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100359 ha_warning("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200360
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200361 if (err & (ERR_FATAL | ERR_ABORT)) {
362 ret = 0;
363 goto end;
364 }
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200365 }
366
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200367 if (l->state < LI_PAUSED || l->state == LI_ZOMBIE) {
368 ret = 0;
369 goto end;
370 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200371
Willy Tarreaub3fb60b2012-10-04 08:56:31 +0200372 if (l->proto->sock_prot == IPPROTO_TCP &&
373 l->state == LI_PAUSED &&
Willy Tarreaue2711c72019-02-27 15:39:41 +0100374 listen(l->fd, listener_backlog(l)) != 0) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200375 ret = 0;
376 goto end;
377 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200378
379 if (l->state == LI_READY)
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200380 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200381
Willy Tarreau01abd022019-02-28 10:27:18 +0100382 LIST_DEL_LOCKED(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200383
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100384 if (l->maxconn && l->nbconn >= l->maxconn) {
Willy Tarreaube58c382011-07-24 18:28:10 +0200385 l->state = LI_FULL;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200386 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200387 }
388
Willy Tarreau413e9262019-07-11 10:08:31 +0200389 if (!(thread_mask(l->bind_conf->bind_thread) & tid_bit)) {
390 /* we're not allowed to touch this listener's FD, let's requeue
391 * the listener into one of its owning thread's queue instead.
392 */
Willy Tarreau51232032020-02-12 10:01:29 +0100393 int first_thread = my_flsl(thread_mask(l->bind_conf->bind_thread) & all_threads_mask) - 1;
Willy Tarreau413e9262019-07-11 10:08:31 +0200394 work_list_add(&local_listener_queue[first_thread], &l->wait_queue);
395 goto end;
396 }
397
Willy Tarreau49b046d2012-08-09 12:11:58 +0200398 fd_want_recv(l->fd);
Willy Tarreaube58c382011-07-24 18:28:10 +0200399 l->state = LI_READY;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200400 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100401 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200402 return ret;
403}
404
Willy Tarreau87b09662015-04-03 00:22:06 +0200405/* Marks a ready listener as full so that the stream code tries to re-enable
Willy Tarreau62793712011-07-24 19:23:38 +0200406 * it upon next close() using resume_listener().
407 */
Christopher Faulet5580ba22017-08-28 15:29:20 +0200408static void listener_full(struct listener *l)
Willy Tarreau62793712011-07-24 19:23:38 +0200409{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100410 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200411 if (l->state >= LI_READY) {
Willy Tarreau01abd022019-02-28 10:27:18 +0100412 LIST_DEL_LOCKED(&l->wait_queue);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100413 if (l->state != LI_FULL) {
414 fd_stop_recv(l->fd);
415 l->state = LI_FULL;
416 }
Willy Tarreau62793712011-07-24 19:23:38 +0200417 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100418 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200419}
420
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200421/* Marks a ready listener as limited so that we only try to re-enable it when
422 * resources are free again. It will be queued into the specified queue.
423 */
Christopher Faulet5580ba22017-08-28 15:29:20 +0200424static void limit_listener(struct listener *l, struct list *list)
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200425{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100426 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200427 if (l->state == LI_READY) {
Willy Tarreau01abd022019-02-28 10:27:18 +0100428 LIST_ADDQ_LOCKED(list, &l->wait_queue);
Willy Tarreau49b046d2012-08-09 12:11:58 +0200429 fd_stop_recv(l->fd);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200430 l->state = LI_LIMITED;
431 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100432 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200433}
434
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100435/* This function adds all of the protocol's listener's file descriptors to the
436 * polling lists when they are in the LI_LISTEN state. It is intended to be
437 * used as a protocol's generic enable_all() primitive, for use after the
438 * fork(). It puts the listeners into LI_READY or LI_FULL states depending on
439 * their number of connections. It always returns ERR_NONE.
Willy Tarreau937604b2019-07-24 16:45:02 +0200440 *
441 * Must be called with proto_lock held.
442 *
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100443 */
444int enable_all_listeners(struct protocol *proto)
445{
446 struct listener *listener;
447
448 list_for_each_entry(listener, &proto->listeners, proto_list)
449 enable_listener(listener);
450 return ERR_NONE;
451}
452
453/* This function removes all of the protocol's listener's file descriptors from
454 * the polling lists when they are in the LI_READY or LI_FULL states. It is
455 * intended to be used as a protocol's generic disable_all() primitive. It puts
456 * the listeners into LI_LISTEN, and always returns ERR_NONE.
Willy Tarreau937604b2019-07-24 16:45:02 +0200457 *
458 * Must be called with proto_lock held.
459 *
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100460 */
461int disable_all_listeners(struct protocol *proto)
462{
463 struct listener *listener;
464
465 list_for_each_entry(listener, &proto->listeners, proto_list)
466 disable_listener(listener);
467 return ERR_NONE;
468}
469
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200470/* Dequeues all of the listeners waiting for a resource in wait queue <queue>. */
471void dequeue_all_listeners(struct list *list)
472{
Willy Tarreau01abd022019-02-28 10:27:18 +0100473 struct listener *listener;
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200474
Willy Tarreau01abd022019-02-28 10:27:18 +0100475 while ((listener = LIST_POP_LOCKED(list, struct listener *, wait_queue))) {
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200476 /* This cannot fail because the listeners are by definition in
Willy Tarreau01abd022019-02-28 10:27:18 +0100477 * the LI_LIMITED state.
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200478 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100479 resume_listener(listener);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200480 }
481}
482
Christopher Faulet510c0d62018-03-16 10:04:47 +0100483/* Must be called with the lock held. Depending on <do_close> value, it does
484 * what unbind_listener or unbind_listener_no_close should do.
485 */
486void do_unbind_listener(struct listener *listener, int do_close)
Willy Tarreaub648d632007-10-28 22:13:50 +0100487{
Olivier Houcharda5188562019-03-08 15:35:42 +0100488 if (listener->state == LI_READY && fd_updt)
Willy Tarreau49b046d2012-08-09 12:11:58 +0200489 fd_stop_recv(listener->fd);
Willy Tarreaub648d632007-10-28 22:13:50 +0100490
Willy Tarreau01abd022019-02-28 10:27:18 +0100491 LIST_DEL_LOCKED(&listener->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200492
Willy Tarreaube58c382011-07-24 18:28:10 +0200493 if (listener->state >= LI_PAUSED) {
Olivier Houchard1fc05162017-04-06 01:05:05 +0200494 if (do_close) {
495 fd_delete(listener->fd);
496 listener->fd = -1;
497 }
498 else
499 fd_remove(listener->fd);
Willy Tarreaub648d632007-10-28 22:13:50 +0100500 listener->state = LI_ASSIGNED;
501 }
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100502}
503
Olivier Houchard1fc05162017-04-06 01:05:05 +0200504/* This function closes the listening socket for the specified listener,
505 * provided that it's already in a listening state. The listener enters the
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100506 * LI_ASSIGNED state. This function is intended to be used as a generic
507 * function for standard protocols.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200508 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100509void unbind_listener(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200510{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100511 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100512 do_unbind_listener(listener, 1);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100513 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200514}
515
516/* This function pretends the listener is dead, but keeps the FD opened, so
517 * that we can provide it, for conf reloading.
518 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100519void unbind_listener_no_close(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200520{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100521 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100522 do_unbind_listener(listener, 0);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100523 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200524}
525
Willy Tarreau3acf8c32007-10-28 22:35:41 +0100526/* This function closes all listening sockets bound to the protocol <proto>,
527 * and the listeners end in LI_ASSIGNED state if they were higher. It does not
528 * detach them from the protocol. It always returns ERR_NONE.
Willy Tarreau937604b2019-07-24 16:45:02 +0200529 *
530 * Must be called with proto_lock held.
531 *
Willy Tarreau3acf8c32007-10-28 22:35:41 +0100532 */
533int unbind_all_listeners(struct protocol *proto)
534{
535 struct listener *listener;
536
537 list_for_each_entry(listener, &proto->listeners, proto_list)
538 unbind_listener(listener);
539 return ERR_NONE;
540}
541
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200542/* creates one or multiple listeners for bind_conf <bc> on sockaddr <ss> on port
543 * range <portl> to <porth>, and possibly attached to fd <fd> (or -1 for auto
544 * allocation). The address family is taken from ss->ss_family. The number of
545 * jobs and listeners is automatically increased by the number of listeners
William Lallemand75ea0a02017-11-15 19:02:58 +0100546 * created. If the <inherited> argument is set to 1, it specifies that the FD
547 * was obtained from a parent process.
548 * It returns non-zero on success, zero on error with the error message
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200549 * set in <err>.
550 */
551int create_listeners(struct bind_conf *bc, const struct sockaddr_storage *ss,
William Lallemand75ea0a02017-11-15 19:02:58 +0100552 int portl, int porth, int fd, int inherited, char **err)
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200553{
554 struct protocol *proto = protocol_by_family(ss->ss_family);
555 struct listener *l;
556 int port;
557
558 if (!proto) {
559 memprintf(err, "unsupported protocol family %d", ss->ss_family);
560 return 0;
561 }
562
563 for (port = portl; port <= porth; port++) {
564 l = calloc(1, sizeof(*l));
565 if (!l) {
566 memprintf(err, "out of memory");
567 return 0;
568 }
569 l->obj_type = OBJ_TYPE_LISTENER;
570 LIST_ADDQ(&bc->frontend->conf.listeners, &l->by_fe);
571 LIST_ADDQ(&bc->listeners, &l->by_bind);
572 l->bind_conf = bc;
573
574 l->fd = fd;
575 memcpy(&l->addr, ss, sizeof(*ss));
Willy Tarreau01abd022019-02-28 10:27:18 +0100576 LIST_INIT(&l->wait_queue);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200577 l->state = LI_INIT;
578
579 proto->add(l, port);
580
William Lallemand75ea0a02017-11-15 19:02:58 +0100581 if (inherited)
582 l->options |= LI_O_INHERITED;
583
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100584 HA_SPIN_INIT(&l->lock);
Olivier Houchard64213e92019-03-08 18:52:57 +0100585 _HA_ATOMIC_ADD(&jobs, 1);
586 _HA_ATOMIC_ADD(&listeners, 1);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200587 }
588 return 1;
589}
590
Willy Tarreau1a64d162007-10-28 22:26:05 +0100591/* Delete a listener from its protocol's list of listeners. The listener's
592 * state is automatically updated from LI_ASSIGNED to LI_INIT. The protocol's
Willy Tarreau2cc5bae2017-09-15 08:18:11 +0200593 * number of listeners is updated, as well as the global number of listeners
594 * and jobs. Note that the listener must have previously been unbound. This
595 * is the generic function to use to remove a listener.
Willy Tarreau937604b2019-07-24 16:45:02 +0200596 *
597 * Will grab the proto_lock.
598 *
Willy Tarreau1a64d162007-10-28 22:26:05 +0100599 */
600void delete_listener(struct listener *listener)
601{
Willy Tarreaub10c8d72019-08-26 10:55:52 +0200602 HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100603 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100604 if (listener->state == LI_ASSIGNED) {
605 listener->state = LI_INIT;
606 LIST_DEL(&listener->proto_list);
607 listener->proto->nb_listeners--;
Olivier Houchard64213e92019-03-08 18:52:57 +0100608 _HA_ATOMIC_SUB(&jobs, 1);
609 _HA_ATOMIC_SUB(&listeners, 1);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100610 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100611 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaub10c8d72019-08-26 10:55:52 +0200612 HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
Willy Tarreau1a64d162007-10-28 22:26:05 +0100613}
614
Willy Tarreaue2711c72019-02-27 15:39:41 +0100615/* Returns a suitable value for a listener's backlog. It uses the listener's,
616 * otherwise the frontend's backlog, otherwise the listener's maxconn,
617 * otherwise the frontend's maxconn, otherwise 1024.
618 */
619int listener_backlog(const struct listener *l)
620{
621 if (l->backlog)
622 return l->backlog;
623
624 if (l->bind_conf->frontend->backlog)
625 return l->bind_conf->frontend->backlog;
626
627 if (l->maxconn)
628 return l->maxconn;
629
630 if (l->bind_conf->frontend->maxconn)
631 return l->bind_conf->frontend->maxconn;
632
633 return 1024;
634}
635
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200636/* This function is called on a read event from a listening socket, corresponding
637 * to an accept. It tries to accept as many connections as possible, and for each
638 * calls the listener's accept handler (generally the frontend's accept handler).
639 */
Willy Tarreauafad0e02012-08-09 14:45:22 +0200640void listener_accept(int fd)
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200641{
642 struct listener *l = fdtab[fd].owner;
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100643 struct proxy *p;
Christopher Faulet102854c2019-04-30 12:17:13 +0200644 unsigned int max_accept;
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100645 int next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100646 int next_feconn = 0;
647 int next_actconn = 0;
Willy Tarreaubb660302014-05-07 19:47:02 +0200648 int expire;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200649 int cfd;
650 int ret;
Willy Tarreau818dca52014-01-31 19:40:19 +0100651#ifdef USE_ACCEPT4
652 static int accept4_broken;
653#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200654
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100655 if (!l)
656 return;
657 p = l->bind_conf->frontend;
Christopher Faulet102854c2019-04-30 12:17:13 +0200658
659 /* if l->maxaccept is -1, then max_accept is UINT_MAX. It is not really
660 * illimited, but it is probably enough.
661 */
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100662 max_accept = l->maxaccept ? l->maxaccept : 1;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200663
Willy Tarreau93e7c002013-10-07 18:51:07 +0200664 if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
665 int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
Willy Tarreau93e7c002013-10-07 18:51:07 +0200666
667 if (unlikely(!max)) {
668 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200669 expire = tick_add(now_ms, next_event_delay(&global.sess_per_sec, global.sps_lim, 0));
Willy Tarreaubb660302014-05-07 19:47:02 +0200670 goto wait_expire;
Willy Tarreau93e7c002013-10-07 18:51:07 +0200671 }
672
673 if (max_accept > max)
674 max_accept = max;
675 }
676
677 if (!(l->options & LI_O_UNLIMITED) && global.cps_lim) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200678 int max = freq_ctr_remain(&global.conn_per_sec, global.cps_lim, 0);
679
680 if (unlikely(!max)) {
681 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200682 expire = tick_add(now_ms, next_event_delay(&global.conn_per_sec, global.cps_lim, 0));
Willy Tarreaubb660302014-05-07 19:47:02 +0200683 goto wait_expire;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200684 }
685
686 if (max_accept > max)
687 max_accept = max;
688 }
Willy Tarreaue43d5322013-10-07 20:01:52 +0200689#ifdef USE_OPENSSL
690 if (!(l->options & LI_O_UNLIMITED) && global.ssl_lim && l->bind_conf && l->bind_conf->is_ssl) {
691 int max = freq_ctr_remain(&global.ssl_per_sec, global.ssl_lim, 0);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200692
Willy Tarreaue43d5322013-10-07 20:01:52 +0200693 if (unlikely(!max)) {
694 /* frontend accept rate limit was reached */
Willy Tarreaue43d5322013-10-07 20:01:52 +0200695 expire = tick_add(now_ms, next_event_delay(&global.ssl_per_sec, global.ssl_lim, 0));
Willy Tarreaubb660302014-05-07 19:47:02 +0200696 goto wait_expire;
Willy Tarreaue43d5322013-10-07 20:01:52 +0200697 }
698
699 if (max_accept > max)
700 max_accept = max;
701 }
702#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200703 if (p && p->fe_sps_lim) {
704 int max = freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0);
705
706 if (unlikely(!max)) {
707 /* frontend accept rate limit was reached */
708 limit_listener(l, &p->listener_queue);
709 task_schedule(p->task, tick_add(now_ms, next_event_delay(&p->fe_sess_per_sec, p->fe_sps_lim, 0)));
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200710 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200711 }
712
713 if (max_accept > max)
714 max_accept = max;
715 }
716
717 /* Note: if we fail to allocate a connection because of configured
718 * limits, we'll schedule a new attempt worst 1 second later in the
719 * worst case. If we fail due to system limits or temporary resource
720 * shortage, we try again 100ms later in the worst case.
721 */
Christopher Faulet102854c2019-04-30 12:17:13 +0200722 for (; max_accept; next_conn = next_feconn = next_actconn = 0, max_accept--) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200723 struct sockaddr_storage addr;
724 socklen_t laddr = sizeof(addr);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200725 unsigned int count;
Willy Tarreau57cb5062019-03-15 17:16:34 +0100726 __decl_hathreads(unsigned long mask);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200727
Willy Tarreau82c97892019-02-27 19:32:32 +0100728 /* pre-increase the number of connections without going too far.
729 * We process the listener, then the proxy, then the process.
730 * We know which ones to unroll based on the next_xxx value.
731 */
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100732 do {
733 count = l->nbconn;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100734 if (unlikely(l->maxconn && count >= l->maxconn)) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100735 /* the listener was marked full or another
736 * thread is going to do it.
737 */
738 next_conn = 0;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100739 listener_full(l);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100740 goto end;
741 }
742 next_conn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000743 } while (!_HA_ATOMIC_CAS(&l->nbconn, (int *)(&count), next_conn));
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100744
Willy Tarreau82c97892019-02-27 19:32:32 +0100745 if (p) {
746 do {
747 count = p->feconn;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100748 if (unlikely(count >= p->maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100749 /* the frontend was marked full or another
750 * thread is going to do it.
751 */
752 next_feconn = 0;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100753 limit_listener(l, &p->listener_queue);
Willy Tarreau82c97892019-02-27 19:32:32 +0100754 goto end;
755 }
756 next_feconn = count + 1;
Olivier Houchard64213e92019-03-08 18:52:57 +0100757 } while (!_HA_ATOMIC_CAS(&p->feconn, &count, next_feconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200758 }
759
Willy Tarreau82c97892019-02-27 19:32:32 +0100760 if (!(l->options & LI_O_UNLIMITED)) {
761 do {
762 count = actconn;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100763 if (unlikely(count >= global.maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100764 /* the process was marked full or another
765 * thread is going to do it.
766 */
767 next_actconn = 0;
Willy Tarreaueb1d4862019-11-15 10:20:07 +0100768 limit_listener(l, &global_listener_queue);
769 task_schedule(global_listener_queue_task, tick_add(now_ms, 1000)); /* try again in 1 second */
Willy Tarreau82c97892019-02-27 19:32:32 +0100770 goto end;
771 }
772 next_actconn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000773 } while (!_HA_ATOMIC_CAS(&actconn, (int *)(&count), next_actconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200774 }
775
William Lallemand2fe7dd02018-09-11 16:51:29 +0200776 /* with sockpair@ we don't want to do an accept */
777 if (unlikely(l->addr.ss_family == AF_CUST_SOCKPAIR)) {
778 if ((cfd = recv_fd_uxst(fd)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100779 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau888d5672019-01-27 18:34:12 +0100780 /* just like with UNIX sockets, only the family is filled */
781 addr.ss_family = AF_UNIX;
782 laddr = sizeof(addr.ss_family);
William Lallemand2fe7dd02018-09-11 16:51:29 +0200783 } else
784
Willy Tarreau1bc4aab2012-10-08 20:11:03 +0200785#ifdef USE_ACCEPT4
Willy Tarreau818dca52014-01-31 19:40:19 +0100786 /* only call accept4() if it's known to be safe, otherwise
787 * fallback to the legacy accept() + fcntl().
788 */
789 if (unlikely(accept4_broken ||
William Lallemandd9138002018-11-27 12:02:39 +0100790 ((cfd = accept4(fd, (struct sockaddr *)&addr, &laddr, SOCK_NONBLOCK)) == -1 &&
Willy Tarreau818dca52014-01-31 19:40:19 +0100791 (errno == ENOSYS || errno == EINVAL || errno == EBADF) &&
792 (accept4_broken = 1))))
793#endif
Willy Tarreau6b3b0d42012-10-22 19:32:55 +0200794 if ((cfd = accept(fd, (struct sockaddr *)&addr, &laddr)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100795 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau818dca52014-01-31 19:40:19 +0100796
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200797 if (unlikely(cfd == -1)) {
798 switch (errno) {
799 case EAGAIN:
Willy Tarreaubbee29c2019-12-10 08:42:21 +0100800 if (fdtab[fd].ev & (FD_POLL_HUP|FD_POLL_ERR)) {
Willy Tarreaubb660302014-05-07 19:47:02 +0200801 /* the listening socket might have been disabled in a shared
802 * process and we're a collateral victim. We'll just pause for
803 * a while in case it comes back. In the mean time, we need to
804 * clear this sticky flag.
805 */
Willy Tarreaubbee29c2019-12-10 08:42:21 +0100806 _HA_ATOMIC_AND(&fdtab[fd].ev, ~(FD_POLL_HUP|FD_POLL_ERR));
Willy Tarreaubb660302014-05-07 19:47:02 +0200807 goto transient_error;
808 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200809 goto end; /* nothing more to accept */
Willy Tarreaubb660302014-05-07 19:47:02 +0200810 case EINVAL:
811 /* might be trying to accept on a shut fd (eg: soft stop) */
812 goto transient_error;
Willy Tarreaua593ec52014-01-20 21:21:30 +0100813 case EINTR:
814 case ECONNABORTED:
Olivier Houchard64213e92019-03-08 18:52:57 +0100815 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100816 if (p)
Olivier Houchard64213e92019-03-08 18:52:57 +0100817 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100818 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +0100819 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreaua593ec52014-01-20 21:21:30 +0100820 continue;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200821 case ENFILE:
822 if (p)
823 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100824 "Proxy %s reached system FD limit (maxsock=%d). Please check system tunables.\n",
825 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200826 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200827 case EMFILE:
828 if (p)
829 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100830 "Proxy %s reached process FD limit (maxsock=%d). Please check 'ulimit-n' and restart.\n",
831 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200832 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200833 case ENOBUFS:
834 case ENOMEM:
835 if (p)
836 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100837 "Proxy %s reached system memory limit (maxsock=%d). Please check system tunables.\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 default:
Willy Tarreaua593ec52014-01-20 21:21:30 +0100841 /* unexpected result, let's give up and let other tasks run */
Willy Tarreau07e13222019-12-10 09:30:05 +0100842 max_accept = 0;
843 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200844 }
845 }
846
William Lallemandd9138002018-11-27 12:02:39 +0100847 /* we don't want to leak the FD upon reload if it's in the master */
848 if (unlikely(master == 1))
849 fcntl(cfd, F_SETFD, FD_CLOEXEC);
850
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100851 /* The connection was accepted, it must be counted as such */
852 if (l->counters)
853 HA_ATOMIC_UPDATE_MAX(&l->counters->conn_max, next_conn);
854
Willy Tarreau82c97892019-02-27 19:32:32 +0100855 if (p)
856 HA_ATOMIC_UPDATE_MAX(&p->fe_counters.conn_max, next_feconn);
857
858 proxy_inc_fe_conn_ctr(l, p);
859
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100860 if (!(l->options & LI_O_UNLIMITED)) {
861 count = update_freq_ctr(&global.conn_per_sec, 1);
862 HA_ATOMIC_UPDATE_MAX(&global.cps_max, count);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100863 }
864
Willy Tarreau64a9c052019-04-12 15:27:17 +0200865 _HA_ATOMIC_ADD(&activity[tid].accepted, 1);
866
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200867 if (unlikely(cfd >= global.maxsock)) {
868 send_log(p, LOG_EMERG,
869 "Proxy %s reached the configured maximum connection limit. Please check the global 'maxconn' value.\n",
870 p->id);
871 close(cfd);
872 limit_listener(l, &global_listener_queue);
873 task_schedule(global_listener_queue_task, tick_add(now_ms, 1000)); /* try again in 1 second */
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200874 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200875 }
876
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100877 /* past this point, l->accept() will automatically decrement
Willy Tarreau82c97892019-02-27 19:32:32 +0100878 * l->nbconn, feconn and actconn once done. Setting next_*conn=0
879 * allows the error path not to rollback on nbconn. It's more
880 * convenient than duplicating all exit labels.
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100881 */
882 next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100883 next_feconn = 0;
884 next_actconn = 0;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200885
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100886#if defined(USE_THREAD)
Willy Tarreau897e2c52019-03-13 15:03:53 +0100887 mask = thread_mask(l->bind_conf->bind_thread) & all_threads_mask;
Willy Tarreauc72d2642020-03-12 17:33:29 +0100888 if (atleast2(mask) && (global.tune.options & GTUNE_LISTENER_MQ) && !stopping) {
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100889 struct accept_queue_ring *ring;
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100890 unsigned int t, t0, t1, t2;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100891
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100892 /* The principle is that we have two running indexes,
893 * each visiting in turn all threads bound to this
894 * listener. The connection will be assigned to the one
895 * with the least connections, and the other one will
896 * be updated. This provides a good fairness on short
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100897 * connections (round robin) and on long ones (conn
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100898 * count), without ever missing any idle thread.
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100899 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100900
901 /* keep a copy for the final update. thr_idx is composite
902 * and made of (t2<<16) + t1.
903 */
Willy Tarreau0cf33172019-03-06 15:26:33 +0100904 t0 = l->thr_idx;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100905 do {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100906 unsigned long m1, m2;
907 int q1, q2;
908
909 t2 = t1 = t0;
910 t2 >>= 16;
911 t1 &= 0xFFFF;
912
913 /* t1 walks low to high bits ;
914 * t2 walks high to low.
915 */
916 m1 = mask >> t1;
917 m2 = mask & (t2 ? nbits(t2 + 1) : ~0UL);
918
Willy Tarreau85d04242019-04-16 18:09:13 +0200919 if (unlikely(!(m1 & 1))) {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100920 m1 &= ~1UL;
921 if (!m1) {
922 m1 = mask;
923 t1 = 0;
924 }
925 t1 += my_ffsl(m1) - 1;
926 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100927
Willy Tarreau85d04242019-04-16 18:09:13 +0200928 if (unlikely(!(m2 & (1UL << t2)) || t1 == t2)) {
929 /* highest bit not set */
930 if (!m2)
931 m2 = mask;
932
933 t2 = my_flsl(m2) - 1;
934 }
935
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100936 /* now we have two distinct thread IDs belonging to the mask */
937 q1 = accept_queue_rings[t1].tail - accept_queue_rings[t1].head + ACCEPT_QUEUE_SIZE;
938 if (q1 >= ACCEPT_QUEUE_SIZE)
939 q1 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100940
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100941 q2 = accept_queue_rings[t2].tail - accept_queue_rings[t2].head + ACCEPT_QUEUE_SIZE;
942 if (q2 >= ACCEPT_QUEUE_SIZE)
943 q2 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100944
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100945 /* we have 3 possibilities now :
946 * q1 < q2 : t1 is less loaded than t2, so we pick it
947 * and update t2 (since t1 might still be
948 * lower than another thread)
949 * q1 > q2 : t2 is less loaded than t1, so we pick it
950 * and update t1 (since t2 might still be
951 * lower than another thread)
952 * q1 = q2 : both are equally loaded, thus we pick t1
953 * and update t1 as it will become more loaded
954 * than t2.
955 */
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100956
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100957 q1 += l->thr_conn[t1];
958 q2 += l->thr_conn[t2];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100959
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100960 if (q1 - q2 < 0) {
961 t = t1;
962 t2 = t2 ? t2 - 1 : LONGBITS - 1;
963 }
964 else if (q1 - q2 > 0) {
965 t = t2;
966 t1++;
967 if (t1 >= LONGBITS)
968 t1 = 0;
969 }
970 else {
971 t = t1;
972 t1++;
973 if (t1 >= LONGBITS)
974 t1 = 0;
975 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100976
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100977 /* new value for thr_idx */
978 t1 += (t2 << 16);
Olivier Houchard64213e92019-03-08 18:52:57 +0100979 } while (unlikely(!_HA_ATOMIC_CAS(&l->thr_idx, &t0, t1)));
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100980
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100981 /* We successfully selected the best thread "t" for this
982 * connection. We use deferred accepts even if it's the
983 * local thread because tests show that it's the best
984 * performing model, likely due to better cache locality
985 * when processing this loop.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100986 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100987 ring = &accept_queue_rings[t];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100988 if (accept_queue_push_mp(ring, cfd, l, &addr, laddr)) {
Olivier Houchard64213e92019-03-08 18:52:57 +0100989 _HA_ATOMIC_ADD(&activity[t].accq_pushed, 1);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100990 task_wakeup(ring->task, TASK_WOKEN_IO);
991 continue;
992 }
993 /* If the ring is full we do a synchronous accept on
994 * the local thread here.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100995 */
Olivier Houchard64213e92019-03-08 18:52:57 +0100996 _HA_ATOMIC_ADD(&activity[t].accq_full, 1);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100997 }
998#endif // USE_THREAD
999
Olivier Houchard64213e92019-03-08 18:52:57 +01001000 _HA_ATOMIC_ADD(&l->thr_conn[tid], 1);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001001 ret = l->accept(l, cfd, &addr);
1002 if (unlikely(ret <= 0)) {
Willy Tarreau87b09662015-04-03 00:22:06 +02001003 /* The connection was closed by stream_accept(). Either
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001004 * we just have to ignore it (ret == 0) or it's a critical
1005 * error due to a resource shortage, and we must stop the
1006 * listener (ret < 0).
1007 */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001008 if (ret == 0) /* successful termination */
1009 continue;
1010
Willy Tarreaubb660302014-05-07 19:47:02 +02001011 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001012 }
1013
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001014 /* increase the per-process number of cumulated sessions, this
1015 * may only be done once l->accept() has accepted the connection.
1016 */
Willy Tarreau93e7c002013-10-07 18:51:07 +02001017 if (!(l->options & LI_O_UNLIMITED)) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001018 count = update_freq_ctr(&global.sess_per_sec, 1);
1019 HA_ATOMIC_UPDATE_MAX(&global.sps_max, count);
Willy Tarreau93e7c002013-10-07 18:51:07 +02001020 }
Willy Tarreaue43d5322013-10-07 20:01:52 +02001021#ifdef USE_OPENSSL
1022 if (!(l->options & LI_O_UNLIMITED) && l->bind_conf && l->bind_conf->is_ssl) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001023 count = update_freq_ctr(&global.ssl_per_sec, 1);
1024 HA_ATOMIC_UPDATE_MAX(&global.ssl_max, count);
Willy Tarreaue43d5322013-10-07 20:01:52 +02001025 }
1026#endif
Willy Tarreau93e7c002013-10-07 18:51:07 +02001027
Willy Tarreau9fbaa3c2020-05-01 09:51:11 +02001028 ti->flags &= ~TI_FL_STUCK; // this thread is still running
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001029 } /* end of for (max_accept--) */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001030
Willy Tarreauaece46a2012-07-06 12:25:58 +02001031 /* we've exhausted max_accept, so there is no need to poll again */
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001032 goto end;
Willy Tarreaubb660302014-05-07 19:47:02 +02001033
1034 transient_error:
Willy Tarreau07e13222019-12-10 09:30:05 +01001035 /* pause the listener for up to 100 ms */
Willy Tarreaubb660302014-05-07 19:47:02 +02001036 expire = tick_add(now_ms, 100);
1037
1038 wait_expire:
Willy Tarreau07e13222019-12-10 09:30:05 +01001039 /* switch the listener to LI_LIMITED and wait until up to <expire> in the global queue */
Willy Tarreaubb660302014-05-07 19:47:02 +02001040 limit_listener(l, &global_listener_queue);
1041 task_schedule(global_listener_queue_task, tick_first(expire, global_listener_queue_task->expire));
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001042 end:
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001043 if (next_conn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001044 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau741b4d62019-02-25 15:02:04 +01001045
Willy Tarreau82c97892019-02-27 19:32:32 +01001046 if (p && next_feconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001047 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001048
1049 if (next_actconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001050 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001051
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001052 if ((l->state == LI_FULL && (!l->maxconn || l->nbconn < l->maxconn)) ||
Willy Tarreau5de8d1f2019-12-11 15:06:30 +01001053 (l->state == LI_LIMITED &&
1054 ((!p || p->feconn < p->maxconn) && (actconn < global.maxconn) &&
1055 (!tick_isset(global_listener_queue_task->expire) ||
1056 tick_is_expired(global_listener_queue_task->expire, now_ms))))) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001057 /* at least one thread has to this when quitting */
1058 resume_listener(l);
1059
1060 /* Dequeues all of the listeners waiting for a resource */
1061 if (!LIST_ISEMPTY(&global_listener_queue))
1062 dequeue_all_listeners(&global_listener_queue);
1063
Christopher Fauletbe2c1062019-09-10 10:01:26 +02001064 if (p && !LIST_ISEMPTY(&p->listener_queue) &&
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001065 (!p->fe_sps_lim || freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0) > 0))
1066 dequeue_all_listeners(&p->listener_queue);
1067 }
Willy Tarreau79116102019-12-05 07:40:32 +01001068
Willy Tarreau07e13222019-12-10 09:30:05 +01001069 /* Now it's getting tricky. The listener was supposed to be in LI_READY
1070 * state but in the mean time we might have changed it to LI_FULL or
1071 * LI_LIMITED, and another thread might also have turned it to
1072 * LI_PAUSED, LI_LISTEN or even LI_INI when stopping a proxy. We must
1073 * be certain to keep the FD enabled when in the READY state but we
1074 * must also stop it for other states that we might have switched to
1075 * while others re-enabled polling.
1076 */
1077 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
1078 if (l->state == LI_READY) {
1079 if (max_accept > 0)
1080 fd_cant_recv(fd);
1081 else
1082 fd_done_recv(fd);
1083 } else if (l->state > LI_ASSIGNED) {
Willy Tarreau79116102019-12-05 07:40:32 +01001084 fd_stop_recv(l->fd);
Willy Tarreau07e13222019-12-10 09:30:05 +01001085 }
1086 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001087}
1088
Willy Tarreau05f50472017-09-15 09:19:58 +02001089/* Notify the listener that a connection initiated from it was released. This
1090 * is used to keep the connection count consistent and to possibly re-open
1091 * listening when it was limited.
1092 */
1093void listener_release(struct listener *l)
1094{
1095 struct proxy *fe = l->bind_conf->frontend;
1096
1097 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +01001098 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001099 if (fe)
Olivier Houchard64213e92019-03-08 18:52:57 +01001100 _HA_ATOMIC_SUB(&fe->feconn, 1);
1101 _HA_ATOMIC_SUB(&l->nbconn, 1);
1102 _HA_ATOMIC_SUB(&l->thr_conn[tid], 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001103
1104 if (l->state == LI_FULL || l->state == LI_LIMITED)
Willy Tarreau05f50472017-09-15 09:19:58 +02001105 resume_listener(l);
1106
1107 /* Dequeues all of the listeners waiting for a resource */
1108 if (!LIST_ISEMPTY(&global_listener_queue))
1109 dequeue_all_listeners(&global_listener_queue);
1110
1111 if (!LIST_ISEMPTY(&fe->listener_queue) &&
1112 (!fe->fe_sps_lim || freq_ctr_remain(&fe->fe_sess_per_sec, fe->fe_sps_lim, 0) > 0))
1113 dequeue_all_listeners(&fe->listener_queue);
1114}
1115
Willy Tarreau413e9262019-07-11 10:08:31 +02001116/* resume listeners waiting in the local listener queue. They are still in LI_LIMITED state */
1117static struct task *listener_queue_process(struct task *t, void *context, unsigned short state)
1118{
1119 struct work_list *wl = context;
1120 struct listener *l;
1121
1122 while ((l = LIST_POP_LOCKED(&wl->head, struct listener *, wait_queue))) {
1123 /* The listeners are still in the LI_LIMITED state */
1124 resume_listener(l);
1125 }
1126 return t;
1127}
1128
1129/* Initializes the listener queues. Returns 0 on success, otherwise ERR_* flags */
1130static int listener_queue_init()
1131{
1132 local_listener_queue = work_list_create(global.nbthread, listener_queue_process, NULL);
1133 if (!local_listener_queue) {
1134 ha_alert("Out of memory while initializing listener queues.\n");
1135 return ERR_FATAL|ERR_ABORT;
1136 }
1137 return 0;
1138}
1139
1140static void listener_queue_deinit()
1141{
1142 work_list_destroy(local_listener_queue, global.nbthread);
1143}
1144
1145REGISTER_CONFIG_POSTPARSER("multi-threaded listener queue", listener_queue_init);
1146REGISTER_POST_DEINIT(listener_queue_deinit);
1147
Willy Tarreau26982662012-09-12 23:17:10 +02001148/*
1149 * Registers the bind keyword list <kwl> as a list of valid keywords for next
1150 * parsing sessions.
1151 */
1152void bind_register_keywords(struct bind_kw_list *kwl)
1153{
1154 LIST_ADDQ(&bind_keywords.list, &kwl->list);
1155}
1156
1157/* Return a pointer to the bind keyword <kw>, or NULL if not found. If the
1158 * keyword is found with a NULL ->parse() function, then an attempt is made to
1159 * find one with a valid ->parse() function. This way it is possible to declare
1160 * platform-dependant, known keywords as NULL, then only declare them as valid
1161 * if some options are met. Note that if the requested keyword contains an
1162 * opening parenthesis, everything from this point is ignored.
1163 */
1164struct bind_kw *bind_find_kw(const char *kw)
1165{
1166 int index;
1167 const char *kwend;
1168 struct bind_kw_list *kwl;
1169 struct bind_kw *ret = NULL;
1170
1171 kwend = strchr(kw, '(');
1172 if (!kwend)
1173 kwend = kw + strlen(kw);
1174
1175 list_for_each_entry(kwl, &bind_keywords.list, list) {
1176 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1177 if ((strncmp(kwl->kw[index].kw, kw, kwend - kw) == 0) &&
1178 kwl->kw[index].kw[kwend-kw] == 0) {
1179 if (kwl->kw[index].parse)
1180 return &kwl->kw[index]; /* found it !*/
1181 else
1182 ret = &kwl->kw[index]; /* may be OK */
1183 }
1184 }
1185 }
1186 return ret;
1187}
1188
Willy Tarreau8638f482012-09-18 18:01:17 +02001189/* Dumps all registered "bind" keywords to the <out> string pointer. The
1190 * unsupported keywords are only dumped if their supported form was not
1191 * found.
1192 */
1193void bind_dump_kws(char **out)
1194{
1195 struct bind_kw_list *kwl;
1196 int index;
1197
1198 *out = NULL;
1199 list_for_each_entry(kwl, &bind_keywords.list, list) {
1200 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1201 if (kwl->kw[index].parse ||
1202 bind_find_kw(kwl->kw[index].kw) == &kwl->kw[index]) {
Willy Tarreau51fb7652012-09-18 18:24:39 +02001203 memprintf(out, "%s[%4s] %s%s%s\n", *out ? *out : "",
1204 kwl->scope,
Willy Tarreau8638f482012-09-18 18:01:17 +02001205 kwl->kw[index].kw,
Willy Tarreau51fb7652012-09-18 18:24:39 +02001206 kwl->kw[index].skip ? " <arg>" : "",
1207 kwl->kw[index].parse ? "" : " (not supported)");
Willy Tarreau8638f482012-09-18 18:01:17 +02001208 }
1209 }
1210 }
1211}
1212
Willy Tarreau645513a2010-05-24 20:55:15 +02001213/************************************************************************/
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001214/* All supported sample and ACL keywords must be declared here. */
Willy Tarreau645513a2010-05-24 20:55:15 +02001215/************************************************************************/
1216
Willy Tarreaua5e37562011-12-16 17:06:15 +01001217/* set temp integer to the number of connexions to the same listening socket */
Willy Tarreau645513a2010-05-24 20:55:15 +02001218static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001219smp_fetch_dconn(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau645513a2010-05-24 20:55:15 +02001220{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001221 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001222 smp->data.u.sint = smp->sess->listener->nbconn;
Willy Tarreau645513a2010-05-24 20:55:15 +02001223 return 1;
1224}
1225
Willy Tarreaua5e37562011-12-16 17:06:15 +01001226/* set temp integer to the id of the socket (listener) */
Willy Tarreau645513a2010-05-24 20:55:15 +02001227static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001228smp_fetch_so_id(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau37406352012-04-23 16:16:37 +02001229{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001230 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001231 smp->data.u.sint = smp->sess->listener->luid;
Willy Tarreau645513a2010-05-24 20:55:15 +02001232 return 1;
1233}
Jerome Magnin28b90332020-03-27 22:08:40 +01001234static int
1235smp_fetch_so_name(const struct arg *args, struct sample *smp, const char *kw, void *private)
1236{
1237 smp->data.u.str.area = smp->sess->listener->name;
1238 if (!smp->data.u.str.area)
1239 return 0;
1240
1241 smp->data.type = SMP_T_STR;
1242 smp->flags = SMP_F_CONST;
1243 smp->data.u.str.data = strlen(smp->data.u.str.area);
1244 return 1;
1245}
Willy Tarreau645513a2010-05-24 20:55:15 +02001246
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001247/* parse the "accept-proxy" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001248static 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 +02001249{
1250 struct listener *l;
1251
Willy Tarreau4348fad2012-09-20 16:48:07 +02001252 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001253 l->options |= LI_O_ACC_PROXY;
1254
1255 return 0;
1256}
1257
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001258/* parse the "accept-netscaler-cip" bind keyword */
1259static int bind_parse_accept_netscaler_cip(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1260{
1261 struct listener *l;
1262 uint32_t val;
1263
1264 if (!*args[cur_arg + 1]) {
1265 memprintf(err, "'%s' : missing value", args[cur_arg]);
1266 return ERR_ALERT | ERR_FATAL;
1267 }
1268
1269 val = atol(args[cur_arg + 1]);
1270 if (val <= 0) {
Willy Tarreaue2711c72019-02-27 15:39:41 +01001271 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001272 return ERR_ALERT | ERR_FATAL;
1273 }
1274
1275 list_for_each_entry(l, &conf->listeners, by_bind) {
1276 l->options |= LI_O_ACC_CIP;
1277 conf->ns_cip_magic = val;
1278 }
1279
1280 return 0;
1281}
1282
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001283/* parse the "backlog" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001284static 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 +02001285{
1286 struct listener *l;
1287 int val;
1288
1289 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001290 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001291 return ERR_ALERT | ERR_FATAL;
1292 }
1293
1294 val = atol(args[cur_arg + 1]);
Willy Tarreaue2711c72019-02-27 15:39:41 +01001295 if (val < 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001296 memprintf(err, "'%s' : invalid value %d, must be > 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001297 return ERR_ALERT | ERR_FATAL;
1298 }
1299
Willy Tarreau4348fad2012-09-20 16:48:07 +02001300 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001301 l->backlog = val;
1302
1303 return 0;
1304}
1305
1306/* parse the "id" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001307static 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 +02001308{
1309 struct eb32_node *node;
Willy Tarreau4348fad2012-09-20 16:48:07 +02001310 struct listener *l, *new;
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001311 char *error;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001312
Willy Tarreau4348fad2012-09-20 16:48:07 +02001313 if (conf->listeners.n != conf->listeners.p) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001314 memprintf(err, "'%s' can only be used with a single socket", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001315 return ERR_ALERT | ERR_FATAL;
1316 }
1317
1318 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001319 memprintf(err, "'%s' : expects an integer argument", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001320 return ERR_ALERT | ERR_FATAL;
1321 }
1322
Willy Tarreau4348fad2012-09-20 16:48:07 +02001323 new = LIST_NEXT(&conf->listeners, struct listener *, by_bind);
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001324 new->luid = strtol(args[cur_arg + 1], &error, 10);
1325 if (*error != '\0') {
1326 memprintf(err, "'%s' : expects an integer argument, found '%s'", args[cur_arg], args[cur_arg + 1]);
1327 return ERR_ALERT | ERR_FATAL;
1328 }
Willy Tarreau4348fad2012-09-20 16:48:07 +02001329 new->conf.id.key = new->luid;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001330
Willy Tarreau4348fad2012-09-20 16:48:07 +02001331 if (new->luid <= 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001332 memprintf(err, "'%s' : custom id has to be > 0", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001333 return ERR_ALERT | ERR_FATAL;
1334 }
1335
Willy Tarreau4348fad2012-09-20 16:48:07 +02001336 node = eb32_lookup(&px->conf.used_listener_id, new->luid);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001337 if (node) {
1338 l = container_of(node, struct listener, conf.id);
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001339 memprintf(err, "'%s' : custom id %d already used at %s:%d ('bind %s')",
1340 args[cur_arg], l->luid, l->bind_conf->file, l->bind_conf->line,
1341 l->bind_conf->arg);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001342 return ERR_ALERT | ERR_FATAL;
1343 }
1344
Willy Tarreau4348fad2012-09-20 16:48:07 +02001345 eb32_insert(&px->conf.used_listener_id, &new->conf.id);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001346 return 0;
1347}
1348
1349/* parse the "maxconn" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001350static 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 +02001351{
1352 struct listener *l;
1353 int val;
1354
1355 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001356 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001357 return ERR_ALERT | ERR_FATAL;
1358 }
1359
1360 val = atol(args[cur_arg + 1]);
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001361 if (val < 0) {
1362 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001363 return ERR_ALERT | ERR_FATAL;
1364 }
1365
Willy Tarreau4348fad2012-09-20 16:48:07 +02001366 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001367 l->maxconn = val;
1368
1369 return 0;
1370}
1371
1372/* parse the "name" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001373static 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 +02001374{
1375 struct listener *l;
1376
1377 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001378 memprintf(err, "'%s' : missing name", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001379 return ERR_ALERT | ERR_FATAL;
1380 }
1381
Willy Tarreau4348fad2012-09-20 16:48:07 +02001382 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001383 l->name = strdup(args[cur_arg + 1]);
1384
1385 return 0;
1386}
1387
1388/* parse the "nice" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001389static 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 +02001390{
1391 struct listener *l;
1392 int val;
1393
1394 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001395 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001396 return ERR_ALERT | ERR_FATAL;
1397 }
1398
1399 val = atol(args[cur_arg + 1]);
1400 if (val < -1024 || val > 1024) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001401 memprintf(err, "'%s' : invalid value %d, allowed range is -1024..1024", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001402 return ERR_ALERT | ERR_FATAL;
1403 }
1404
Willy Tarreau4348fad2012-09-20 16:48:07 +02001405 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001406 l->nice = val;
1407
1408 return 0;
1409}
1410
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001411/* parse the "process" bind keyword */
1412static int bind_parse_process(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1413{
Christopher Fauletc644fa92017-11-23 22:44:11 +01001414 char *slash;
1415 unsigned long proc = 0, thread = 0;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001416
Christopher Fauletc644fa92017-11-23 22:44:11 +01001417 if ((slash = strchr(args[cur_arg + 1], '/')) != NULL)
1418 *slash = 0;
1419
Willy Tarreauff9c9142019-02-07 10:39:36 +01001420 if (parse_process_number(args[cur_arg + 1], &proc, MAX_PROCS, NULL, err)) {
Christopher Fauletf1f0c5f2017-11-22 12:06:43 +01001421 memprintf(err, "'%s' : %s", args[cur_arg], *err);
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001422 return ERR_ALERT | ERR_FATAL;
1423 }
1424
Christopher Fauletc644fa92017-11-23 22:44:11 +01001425 if (slash) {
Willy Tarreauc9a82e42019-01-26 13:25:14 +01001426 if (parse_process_number(slash+1, &thread, MAX_THREADS, NULL, err)) {
Christopher Fauletc644fa92017-11-23 22:44:11 +01001427 memprintf(err, "'%s' : %s", args[cur_arg], *err);
1428 return ERR_ALERT | ERR_FATAL;
1429 }
1430 *slash = '/';
1431 }
1432
1433 conf->bind_proc |= proc;
Willy Tarreaua36b3242019-02-02 13:14:34 +01001434 conf->bind_thread |= thread;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001435 return 0;
1436}
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001437
Christopher Fauleta717b992018-04-10 14:43:00 +02001438/* parse the "proto" bind keyword */
1439static int bind_parse_proto(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1440{
1441 struct ist proto;
1442
1443 if (!*args[cur_arg + 1]) {
1444 memprintf(err, "'%s' : missing value", args[cur_arg]);
1445 return ERR_ALERT | ERR_FATAL;
1446 }
1447
1448 proto = ist2(args[cur_arg + 1], strlen(args[cur_arg + 1]));
1449 conf->mux_proto = get_mux_proto(proto);
1450 if (!conf->mux_proto) {
1451 memprintf(err, "'%s' : unknown MUX protocol '%s'", args[cur_arg], args[cur_arg+1]);
1452 return ERR_ALERT | ERR_FATAL;
1453 }
Christopher Fauleta717b992018-04-10 14:43:00 +02001454 return 0;
1455}
1456
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001457/* config parser for global "tune.listener.multi-queue", accepts "on" or "off" */
1458static int cfg_parse_tune_listener_mq(char **args, int section_type, struct proxy *curpx,
1459 struct proxy *defpx, const char *file, int line,
1460 char **err)
1461{
1462 if (too_many_args(1, args, err, NULL))
1463 return -1;
1464
1465 if (strcmp(args[1], "on") == 0)
1466 global.tune.options |= GTUNE_LISTENER_MQ;
1467 else if (strcmp(args[1], "off") == 0)
1468 global.tune.options &= ~GTUNE_LISTENER_MQ;
1469 else {
1470 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
1471 return -1;
1472 }
1473 return 0;
1474}
1475
Willy Tarreau61612d42012-04-19 18:42:05 +02001476/* Note: must not be declared <const> as its list will be overwritten.
1477 * Please take care of keeping this list alphabetically sorted.
1478 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001479static struct sample_fetch_kw_list smp_kws = {ILH, {
Thierry FOURNIER07ee64e2015-07-06 23:43:03 +02001480 { "dst_conn", smp_fetch_dconn, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
1481 { "so_id", smp_fetch_so_id, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
Jerome Magnin28b90332020-03-27 22:08:40 +01001482 { "so_name", smp_fetch_so_name, 0, NULL, SMP_T_STR, SMP_USE_FTEND, },
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001483 { /* END */ },
1484}};
1485
Willy Tarreau0108d902018-11-25 19:14:37 +01001486INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
1487
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001488/* Note: must not be declared <const> as its list will be overwritten.
1489 * Please take care of keeping this list alphabetically sorted.
1490 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001491static struct acl_kw_list acl_kws = {ILH, {
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001492 { /* END */ },
Willy Tarreau645513a2010-05-24 20:55:15 +02001493}};
1494
Willy Tarreau0108d902018-11-25 19:14:37 +01001495INITCALL1(STG_REGISTER, acl_register_keywords, &acl_kws);
1496
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001497/* Note: must not be declared <const> as its list will be overwritten.
1498 * Please take care of keeping this list alphabetically sorted, doing so helps
1499 * all code contributors.
1500 * Optional keywords are also declared with a NULL ->parse() function so that
1501 * the config parser can report an appropriate error when a known keyword was
1502 * not enabled.
1503 */
Willy Tarreau51fb7652012-09-18 18:24:39 +02001504static struct bind_kw_list bind_kws = { "ALL", { }, {
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001505 { "accept-netscaler-cip", bind_parse_accept_netscaler_cip, 1 }, /* enable NetScaler Client IP insertion protocol */
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001506 { "accept-proxy", bind_parse_accept_proxy, 0 }, /* enable PROXY protocol */
1507 { "backlog", bind_parse_backlog, 1 }, /* set backlog of listening socket */
1508 { "id", bind_parse_id, 1 }, /* set id of listening socket */
1509 { "maxconn", bind_parse_maxconn, 1 }, /* set maxconn of listening socket */
1510 { "name", bind_parse_name, 1 }, /* set name of listening socket */
1511 { "nice", bind_parse_nice, 1 }, /* set nice of listening socket */
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001512 { "process", bind_parse_process, 1 }, /* set list of allowed process for this socket */
Christopher Fauleta717b992018-04-10 14:43:00 +02001513 { "proto", bind_parse_proto, 1 }, /* set the proto to use for all incoming connections */
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001514 { /* END */ },
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001515}};
1516
Willy Tarreau0108d902018-11-25 19:14:37 +01001517INITCALL1(STG_REGISTER, bind_register_keywords, &bind_kws);
1518
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001519/* config keyword parsers */
1520static struct cfg_kw_list cfg_kws = {ILH, {
1521 { CFG_GLOBAL, "tune.listener.multi-queue", cfg_parse_tune_listener_mq },
1522 { 0, NULL, NULL }
1523}};
1524
1525INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
1526
Willy Tarreau645513a2010-05-24 20:55:15 +02001527/*
1528 * Local variables:
1529 * c-indent-level: 8
1530 * c-basic-offset: 8
1531 * End:
1532 */