blob: 82ed026d5beb8e391ecc1339f8b118e4e759bf20 [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 Tarreaudcc048a2020-06-04 19:11:43 +020021#include <haproxy/acl.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020022#include <haproxy/api.h>
Willy Tarreau6be78492020-06-05 00:00:29 +020023#include <haproxy/cfgparse.h>
Willy Tarreau7ea393d2020-06-04 18:02:10 +020024#include <haproxy/connection.h>
Willy Tarreau8d366972020-05-27 16:10:29 +020025#include <haproxy/errors.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020026#include <haproxy/fd.h>
27#include <haproxy/freq_ctr.h>
Willy Tarreauf268ee82020-06-04 17:05:57 +020028#include <haproxy/global.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020029#include <haproxy/list.h>
Willy Tarreau213e9902020-06-04 14:58:24 +020030#include <haproxy/listener.h>
Willy Tarreauaeed4a82020-06-04 22:01:04 +020031#include <haproxy/log.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020032#include <haproxy/proto_sockpair.h>
33#include <haproxy/protocol-t.h>
34#include <haproxy/protocol.h>
35#include <haproxy/sample.h>
Willy Tarreaudfd3de82020-06-04 23:46:14 +020036#include <haproxy/stream.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020037#include <haproxy/task.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020038#include <haproxy/time.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020039#include <haproxy/tools.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020040
Willy Tarreaub648d632007-10-28 22:13:50 +010041
Willy Tarreau26982662012-09-12 23:17:10 +020042/* List head of all known bind keywords */
43static struct bind_kw_list bind_keywords = {
44 .list = LIST_HEAD_INIT(bind_keywords.list)
45};
46
Willy Tarreaua1d97f82019-12-10 11:18:41 +010047/* list of the temporarily limited listeners because of lack of resource */
48static struct mt_list global_listener_queue = MT_LIST_HEAD_INIT(global_listener_queue);
49static struct task *global_listener_queue_task;
50static struct task *manage_global_listener_queue(struct task *t, void *context, unsigned short state);
51
52
Willy Tarreau1efafce2019-01-27 15:37:19 +010053#if defined(USE_THREAD)
54
55struct accept_queue_ring accept_queue_rings[MAX_THREADS] __attribute__((aligned(64))) = { };
56
57/* dequeue and process a pending connection from the local accept queue (single
58 * consumer). Returns the accepted fd or -1 if none was found. The listener is
59 * placed into *li. The address is copied into *addr for no more than *addr_len
60 * bytes, and the address length is returned into *addr_len.
61 */
62int accept_queue_pop_sc(struct accept_queue_ring *ring, struct listener **li, void *addr, int *addr_len)
63{
64 struct accept_queue_entry *e;
65 unsigned int pos, next;
66 struct listener *ptr;
67 int len;
68 int fd;
69
70 pos = ring->head;
71
72 if (pos == ring->tail)
73 return -1;
74
75 next = pos + 1;
76 if (next >= ACCEPT_QUEUE_SIZE)
77 next = 0;
78
79 e = &ring->entry[pos];
80
81 /* wait for the producer to update the listener's pointer */
82 while (1) {
83 ptr = e->listener;
84 __ha_barrier_load();
85 if (ptr)
86 break;
87 pl_cpu_relax();
88 }
89
90 fd = e->fd;
91 len = e->addr_len;
92 if (len > *addr_len)
93 len = *addr_len;
94
95 if (likely(len > 0))
96 memcpy(addr, &e->addr, len);
97
98 /* release the entry */
99 e->listener = NULL;
100
101 __ha_barrier_store();
102 ring->head = next;
103
104 *addr_len = len;
105 *li = ptr;
106
107 return fd;
108}
109
110
111/* tries to push a new accepted connection <fd> into ring <ring> for listener
112 * <li>, from address <addr> whose length is <addr_len>. Returns non-zero if it
113 * succeeds, or zero if the ring is full. Supports multiple producers.
114 */
115int accept_queue_push_mp(struct accept_queue_ring *ring, int fd,
116 struct listener *li, const void *addr, int addr_len)
117{
118 struct accept_queue_entry *e;
119 unsigned int pos, next;
120
121 pos = ring->tail;
122 do {
123 next = pos + 1;
124 if (next >= ACCEPT_QUEUE_SIZE)
125 next = 0;
126 if (next == ring->head)
127 return 0; // ring full
Olivier Houchard64213e92019-03-08 18:52:57 +0100128 } while (unlikely(!_HA_ATOMIC_CAS(&ring->tail, &pos, next)));
Willy Tarreau1efafce2019-01-27 15:37:19 +0100129
130
131 e = &ring->entry[pos];
132
133 if (addr_len > sizeof(e->addr))
134 addr_len = sizeof(e->addr);
135
136 if (addr_len)
137 memcpy(&e->addr, addr, addr_len);
138
139 e->addr_len = addr_len;
140 e->fd = fd;
141
142 __ha_barrier_store();
143 /* now commit the change */
144
145 e->listener = li;
146 return 1;
147}
148
149/* proceed with accepting new connections */
150static struct task *accept_queue_process(struct task *t, void *context, unsigned short state)
151{
152 struct accept_queue_ring *ring = context;
153 struct listener *li;
154 struct sockaddr_storage addr;
Christopher Faulet102854c2019-04-30 12:17:13 +0200155 unsigned int max_accept;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100156 int addr_len;
157 int ret;
158 int fd;
159
Christopher Faulet102854c2019-04-30 12:17:13 +0200160 /* if global.tune.maxaccept is -1, then max_accept is UINT_MAX. It
161 * is not really illimited, but it is probably enough.
162 */
163 max_accept = global.tune.maxaccept ? global.tune.maxaccept : 64;
164 for (; max_accept; max_accept--) {
Willy Tarreau1efafce2019-01-27 15:37:19 +0100165 addr_len = sizeof(addr);
166 fd = accept_queue_pop_sc(ring, &li, &addr, &addr_len);
167 if (fd < 0)
168 break;
169
Olivier Houchard64213e92019-03-08 18:52:57 +0100170 _HA_ATOMIC_ADD(&li->thr_conn[tid], 1);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100171 ret = li->accept(li, fd, &addr);
172 if (ret <= 0) {
173 /* connection was terminated by the application */
174 continue;
175 }
176
177 /* increase the per-process number of cumulated sessions, this
178 * may only be done once l->accept() has accepted the connection.
179 */
180 if (!(li->options & LI_O_UNLIMITED)) {
181 HA_ATOMIC_UPDATE_MAX(&global.sps_max,
182 update_freq_ctr(&global.sess_per_sec, 1));
183 if (li->bind_conf && li->bind_conf->is_ssl) {
184 HA_ATOMIC_UPDATE_MAX(&global.ssl_max,
185 update_freq_ctr(&global.ssl_per_sec, 1));
186 }
187 }
188 }
189
190 /* ran out of budget ? Let's come here ASAP */
Christopher Faulet102854c2019-04-30 12:17:13 +0200191 if (!max_accept)
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200192 tasklet_wakeup(ring->tasklet);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100193
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200194 return NULL;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100195}
196
197/* Initializes the accept-queues. Returns 0 on success, otherwise ERR_* flags */
198static int accept_queue_init()
199{
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200200 struct tasklet *t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100201 int i;
202
203 for (i = 0; i < global.nbthread; i++) {
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200204 t = tasklet_new();
Willy Tarreau1efafce2019-01-27 15:37:19 +0100205 if (!t) {
206 ha_alert("Out of memory while initializing accept queue for thread %d\n", i);
207 return ERR_FATAL|ERR_ABORT;
208 }
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200209 t->tid = i;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100210 t->process = accept_queue_process;
211 t->context = &accept_queue_rings[i];
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200212 accept_queue_rings[i].tasklet = t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100213 }
214 return 0;
215}
216
217REGISTER_CONFIG_POSTPARSER("multi-threaded accept queue", accept_queue_init);
218
219#endif // USE_THREAD
220
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200221/* adjust the listener's state and its proxy's listener counters if needed.
222 * It must be called under the listener's lock, but uses atomic ops to change
223 * the proxy's counters so that the proxy lock is not needed.
224 */
Willy Tarreaua37b2442020-09-24 07:23:45 +0200225void listener_set_state(struct listener *l, enum li_state st)
226{
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200227 struct proxy *px = l->bind_conf->frontend;
228
229 if (px) {
230 /* from state */
231 switch (l->state) {
232 case LI_NEW: /* first call */
233 _HA_ATOMIC_ADD(&px->li_all, 1);
234 break;
235 case LI_INIT:
236 case LI_ASSIGNED:
237 break;
238 case LI_PAUSED:
239 _HA_ATOMIC_SUB(&px->li_paused, 1);
240 break;
241 case LI_LISTEN:
242 _HA_ATOMIC_SUB(&px->li_bound, 1);
243 break;
244 case LI_READY:
245 case LI_FULL:
246 case LI_LIMITED:
247 _HA_ATOMIC_SUB(&px->li_ready, 1);
248 break;
249 }
250
251 /* to state */
252 switch (st) {
253 case LI_NEW:
254 case LI_INIT:
255 case LI_ASSIGNED:
256 break;
257 case LI_PAUSED:
258 _HA_ATOMIC_ADD(&px->li_paused, 1);
259 break;
260 case LI_LISTEN:
261 _HA_ATOMIC_ADD(&px->li_bound, 1);
262 break;
263 case LI_READY:
264 case LI_FULL:
265 case LI_LIMITED:
266 _HA_ATOMIC_ADD(&px->li_ready, 1);
267 break;
268 }
269 }
Willy Tarreaua37b2442020-09-24 07:23:45 +0200270 l->state = st;
271}
272
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100273/* This function adds the specified listener's file descriptor to the polling
274 * lists if it is in the LI_LISTEN state. The listener enters LI_READY or
Ilya Shipitsin6fb0f212020-04-02 15:25:26 +0500275 * LI_FULL state depending on its number of connections. In daemon mode, we
Willy Tarreauae302532014-05-07 19:22:24 +0200276 * also support binding only the relevant processes to their respective
277 * listeners. We don't do that in debug mode however.
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100278 */
Willy Tarreau7834a3f2020-09-25 16:40:18 +0200279void enable_listener(struct listener *listener)
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100280{
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100281 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100282 if (listener->state == LI_LISTEN) {
William Lallemand095ba4c2017-06-01 17:38:50 +0200283 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau818a92e2020-09-03 07:50:19 +0200284 !(proc_mask(listener->rx.settings->bind_proc) & pid_bit)) {
Willy Tarreauae302532014-05-07 19:22:24 +0200285 /* we don't want to enable this listener and don't
286 * want any fd event to reach it.
287 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200288 if (!(global.tune.options & GTUNE_SOCKET_TRANSFER))
Christopher Faulet510c0d62018-03-16 10:04:47 +0100289 do_unbind_listener(listener, 1);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200290 else {
Christopher Faulet510c0d62018-03-16 10:04:47 +0100291 do_unbind_listener(listener, 0);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200292 listener_set_state(listener, LI_LISTEN);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200293 }
Willy Tarreauae302532014-05-07 19:22:24 +0200294 }
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100295 else if (!listener->maxconn || listener->nbconn < listener->maxconn) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200296 listener->rx.proto->enable(listener);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200297 listener_set_state(listener, LI_READY);
Willy Tarreauae302532014-05-07 19:22:24 +0200298 }
299 else {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200300 listener_set_state(listener, LI_FULL);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100301 }
302 }
William Lallemande22f11f2018-09-11 10:06:27 +0200303 /* if this listener is supposed to be only in the master, close it in the workers */
304 if ((global.mode & MODE_MWORKER) &&
305 (listener->options & LI_O_MWORKER) &&
306 master == 0) {
307 do_unbind_listener(listener, 1);
308 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100309 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100310}
311
Willy Tarreaube58c382011-07-24 18:28:10 +0200312/* This function tries to temporarily disable a listener, depending on the OS
313 * capabilities. Linux unbinds the listen socket after a SHUT_RD, and ignores
314 * SHUT_WR. Solaris refuses either shutdown(). OpenBSD ignores SHUT_RD but
315 * closes upon SHUT_WR and refuses to rebind. So a common validation path
316 * involves SHUT_WR && listen && SHUT_RD. In case of success, the FD's polling
317 * is disabled. It normally returns non-zero, unless an error is reported.
318 */
319int pause_listener(struct listener *l)
320{
Willy Tarreau58651b42020-09-24 16:03:29 +0200321 struct proxy *px = l->bind_conf->frontend;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200322 int ret = 1;
323
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100324 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200325
Willy Tarreau02e19752020-09-23 17:17:22 +0200326 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
327 !(proc_mask(l->rx.settings->bind_proc) & pid_bit))
328 goto end;
329
Willy Tarreau9b3a9322020-09-24 14:46:34 +0200330 if (l->state <= LI_PAUSED)
331 goto end;
332
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200333 if (l->rx.proto->rx_suspend) {
Willy Tarreau092d8652014-07-07 20:22:12 +0200334 /* Returns < 0 in case of failure, 0 if the listener
335 * was totally stopped, or > 0 if correctly paused.
336 */
Willy Tarreaucb66ea62020-09-25 17:12:32 +0200337 ret = l->rx.proto->rx_suspend(&l->rx);
Willy Tarreaube58c382011-07-24 18:28:10 +0200338
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200339 if (ret < 0) {
340 ret = 0;
341 goto end;
342 }
Willy Tarreaufb76bd52020-09-24 18:07:48 +0200343 ret = 1;
Willy Tarreaub3fb60b2012-10-04 08:56:31 +0200344 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200345
Olivier Houchard859dc802019-08-08 15:47:21 +0200346 MT_LIST_DEL(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200347
Willy Tarreaua37b2442020-09-24 07:23:45 +0200348 listener_set_state(l, LI_PAUSED);
Willy Tarreau58651b42020-09-24 16:03:29 +0200349
350 if (px && !px->li_ready) {
351 ha_warning("Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
352 send_log(px, LOG_WARNING, "Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
353 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200354 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100355 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200356 return ret;
Willy Tarreaube58c382011-07-24 18:28:10 +0200357}
358
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200359/* This function tries to resume a temporarily disabled listener. Paused, full,
360 * limited and disabled listeners are handled, which means that this function
361 * may replace enable_listener(). The resulting state will either be LI_READY
362 * or LI_FULL. 0 is returned in case of failure to resume (eg: dead socket).
Willy Tarreauae302532014-05-07 19:22:24 +0200363 * Listeners bound to a different process are not woken up unless we're in
Willy Tarreauaf2fd582015-04-14 12:07:16 +0200364 * foreground mode, and are ignored. If the listener was only in the assigned
365 * state, it's totally rebound. This can happen if a pause() has completely
366 * stopped it. If the resume fails, 0 is returned and an error might be
367 * displayed.
Willy Tarreaube58c382011-07-24 18:28:10 +0200368 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100369int resume_listener(struct listener *l)
Willy Tarreaube58c382011-07-24 18:28:10 +0200370{
Willy Tarreau58651b42020-09-24 16:03:29 +0200371 struct proxy *px = l->bind_conf->frontend;
372 int was_paused = px && px->li_paused;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200373 int ret = 1;
374
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100375 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200376
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200377 /* check that another thread didn't to the job in parallel (e.g. at the
378 * end of listen_accept() while we'd come from dequeue_all_listeners().
379 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200380 if (MT_LIST_ADDED(&l->wait_queue))
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200381 goto end;
382
William Lallemand095ba4c2017-06-01 17:38:50 +0200383 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau818a92e2020-09-03 07:50:19 +0200384 !(proc_mask(l->rx.settings->bind_proc) & pid_bit))
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200385 goto end;
Willy Tarreau3569df32017-03-15 12:47:46 +0100386
Willy Tarreau5d7f9ce2020-09-24 18:54:11 +0200387 if (l->state == LI_READY)
388 goto end;
389
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200390 if (l->state == LI_ASSIGNED) {
391 char msg[100];
392 int err;
393
Willy Tarreaub3580b12020-09-01 10:26:22 +0200394 err = l->rx.proto->listen(l, msg, sizeof(msg));
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200395 if (err & ERR_ALERT)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100396 ha_alert("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200397 else if (err & ERR_WARN)
Christopher Faulet767a84b2017-11-24 16:50:31 +0100398 ha_warning("Resuming listener: %s\n", msg);
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200399
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200400 if (err & (ERR_FATAL | ERR_ABORT)) {
401 ret = 0;
402 goto end;
403 }
Willy Tarreau1c4b8142014-07-07 21:06:24 +0200404 }
405
Willy Tarreauc6dac6c2020-09-23 17:34:22 +0200406 if (l->state < LI_PAUSED) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200407 ret = 0;
408 goto end;
409 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200410
Willy Tarreau010fe152020-09-25 17:31:05 +0200411 if (l->state == LI_PAUSED && l->rx.proto->rx_resume &&
412 l->rx.proto->rx_resume(&l->rx) <= 0) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200413 ret = 0;
414 goto end;
415 }
Willy Tarreaube58c382011-07-24 18:28:10 +0200416
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100417 if (l->maxconn && l->nbconn >= l->maxconn) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200418 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200419 listener_set_state(l, LI_FULL);
Willy Tarreau58651b42020-09-24 16:03:29 +0200420 goto done;
Willy Tarreaube58c382011-07-24 18:28:10 +0200421 }
422
Willy Tarreau4b51f422020-09-25 20:32:28 +0200423 l->rx.proto->enable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200424 listener_set_state(l, LI_READY);
Willy Tarreau58651b42020-09-24 16:03:29 +0200425
426 done:
427 if (was_paused && !px->li_paused) {
428 ha_warning("Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
429 send_log(px, LOG_WARNING, "Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
430 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200431 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100432 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200433 return ret;
434}
435
Willy Tarreau87b09662015-04-03 00:22:06 +0200436/* Marks a ready listener as full so that the stream code tries to re-enable
Willy Tarreau62793712011-07-24 19:23:38 +0200437 * it upon next close() using resume_listener().
438 */
Christopher Faulet5580ba22017-08-28 15:29:20 +0200439static void listener_full(struct listener *l)
Willy Tarreau62793712011-07-24 19:23:38 +0200440{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100441 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200442 if (l->state >= LI_READY) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200443 MT_LIST_DEL(&l->wait_queue);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100444 if (l->state != LI_FULL) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200445 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200446 listener_set_state(l, LI_FULL);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100447 }
Willy Tarreau62793712011-07-24 19:23:38 +0200448 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100449 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200450}
451
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200452/* Marks a ready listener as limited so that we only try to re-enable it when
453 * resources are free again. It will be queued into the specified queue.
454 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200455static void limit_listener(struct listener *l, struct mt_list *list)
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200456{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100457 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200458 if (l->state == LI_READY) {
Willy Tarreaude4db172020-07-10 08:10:29 +0200459 MT_LIST_TRY_ADDQ(list, &l->wait_queue);
Willy Tarreau4b51f422020-09-25 20:32:28 +0200460 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200461 listener_set_state(l, LI_LIMITED);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200462 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100463 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200464}
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
Willy Tarreaub6607bf2020-09-23 16:24:23 +0200493 * what unbind_listener or unbind_listener_no_close should do. It can also
494 * close a zombie listener's FD when called in early states.
Christopher Faulet510c0d62018-03-16 10:04:47 +0100495 */
496void do_unbind_listener(struct listener *listener, int do_close)
Willy Tarreaub648d632007-10-28 22:13:50 +0100497{
Willy Tarreau4b51f422020-09-25 20:32:28 +0200498 if (listener->state == LI_READY)
499 listener->rx.proto->disable(listener);
Willy Tarreaub648d632007-10-28 22:13:50 +0100500
Olivier Houchard859dc802019-08-08 15:47:21 +0200501 MT_LIST_DEL(&listener->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200502
Willy Tarreaube58c382011-07-24 18:28:10 +0200503 if (listener->state >= LI_PAUSED) {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200504 listener_set_state(listener, LI_ASSIGNED);
Willy Tarreau4b51f422020-09-25 20:32:28 +0200505 listener->rx.proto->rx_disable(&listener->rx);
Willy Tarreaub6607bf2020-09-23 16:24:23 +0200506 }
507
508 if (do_close && listener->rx.fd != -1) {
509 fd_delete(listener->rx.fd);
510 listener->rx.flags &= ~RX_F_BOUND;
511 listener->rx.fd = -1;
Willy Tarreaub648d632007-10-28 22:13:50 +0100512 }
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100513}
514
Olivier Houchard1fc05162017-04-06 01:05:05 +0200515/* This function closes the listening socket for the specified listener,
516 * provided that it's already in a listening state. The listener enters the
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100517 * LI_ASSIGNED state. This function is intended to be used as a generic
518 * function for standard protocols.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200519 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100520void unbind_listener(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200521{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100522 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100523 do_unbind_listener(listener, 1);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100524 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200525}
526
527/* This function pretends the listener is dead, but keeps the FD opened, so
528 * that we can provide it, for conf reloading.
529 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100530void unbind_listener_no_close(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200531{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100532 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100533 do_unbind_listener(listener, 0);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100534 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200535}
536
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200537/* creates one or multiple listeners for bind_conf <bc> on sockaddr <ss> on port
538 * range <portl> to <porth>, and possibly attached to fd <fd> (or -1 for auto
Willy Tarreau9b3178d2020-09-16 17:58:55 +0200539 * allocation). The address family is taken from ss->ss_family, and the protocol
540 * passed in <proto> must be usable on this family. The number of jobs and
541 * listeners is automatically increased by the number of listeners created. It
542 * returns non-zero on success, zero on error with the error message set in <err>.
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200543 */
544int create_listeners(struct bind_conf *bc, const struct sockaddr_storage *ss,
Willy Tarreau9b3178d2020-09-16 17:58:55 +0200545 int portl, int porth, int fd, struct protocol *proto, char **err)
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200546{
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200547 struct listener *l;
548 int port;
549
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200550 for (port = portl; port <= porth; port++) {
551 l = calloc(1, sizeof(*l));
552 if (!l) {
553 memprintf(err, "out of memory");
554 return 0;
555 }
556 l->obj_type = OBJ_TYPE_LISTENER;
557 LIST_ADDQ(&bc->frontend->conf.listeners, &l->by_fe);
558 LIST_ADDQ(&bc->listeners, &l->by_bind);
559 l->bind_conf = bc;
Willy Tarreau0fce6bc2020-09-03 07:46:06 +0200560 l->rx.settings = &bc->settings;
Willy Tarreaueef45422020-09-03 10:05:03 +0200561 l->rx.owner = l;
Willy Tarreau38ba6472020-08-27 08:16:52 +0200562 l->rx.fd = fd;
Willy Tarreau37159062020-08-27 07:48:42 +0200563 memcpy(&l->rx.addr, ss, sizeof(*ss));
Olivier Houchard859dc802019-08-08 15:47:21 +0200564 MT_LIST_INIT(&l->wait_queue);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200565 listener_set_state(l, LI_INIT);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200566
567 proto->add(l, port);
568
Willy Tarreau909c23b2020-09-15 13:50:58 +0200569 if (fd != -1)
Willy Tarreau43046fa2020-09-01 15:41:59 +0200570 l->rx.flags |= RX_F_INHERITED;
William Lallemand75ea0a02017-11-15 19:02:58 +0100571
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100572 HA_SPIN_INIT(&l->lock);
Olivier Houchard64213e92019-03-08 18:52:57 +0100573 _HA_ATOMIC_ADD(&jobs, 1);
574 _HA_ATOMIC_ADD(&listeners, 1);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200575 }
576 return 1;
577}
578
Willy Tarreau1a64d162007-10-28 22:26:05 +0100579/* Delete a listener from its protocol's list of listeners. The listener's
580 * state is automatically updated from LI_ASSIGNED to LI_INIT. The protocol's
Willy Tarreau2cc5bae2017-09-15 08:18:11 +0200581 * number of listeners is updated, as well as the global number of listeners
582 * and jobs. Note that the listener must have previously been unbound. This
583 * is the generic function to use to remove a listener.
Willy Tarreaudaacf362019-07-24 16:45:02 +0200584 *
585 * Will grab the proto_lock.
586 *
Willy Tarreau1a64d162007-10-28 22:26:05 +0100587 */
588void delete_listener(struct listener *listener)
589{
Willy Tarreau6ee9f8d2019-08-26 10:55:52 +0200590 HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100591 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100592 if (listener->state == LI_ASSIGNED) {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200593 listener_set_state(listener, LI_INIT);
Willy Tarreaub7436612020-08-28 19:51:44 +0200594 LIST_DEL(&listener->rx.proto_list);
Willy Tarreaud7f331c2020-09-25 17:01:43 +0200595 listener->rx.proto->nb_receivers--;
Olivier Houchard64213e92019-03-08 18:52:57 +0100596 _HA_ATOMIC_SUB(&jobs, 1);
597 _HA_ATOMIC_SUB(&listeners, 1);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100598 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100599 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau6ee9f8d2019-08-26 10:55:52 +0200600 HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
Willy Tarreau1a64d162007-10-28 22:26:05 +0100601}
602
Willy Tarreaue2711c72019-02-27 15:39:41 +0100603/* Returns a suitable value for a listener's backlog. It uses the listener's,
604 * otherwise the frontend's backlog, otherwise the listener's maxconn,
605 * otherwise the frontend's maxconn, otherwise 1024.
606 */
607int listener_backlog(const struct listener *l)
608{
609 if (l->backlog)
610 return l->backlog;
611
612 if (l->bind_conf->frontend->backlog)
613 return l->bind_conf->frontend->backlog;
614
615 if (l->maxconn)
616 return l->maxconn;
617
618 if (l->bind_conf->frontend->maxconn)
619 return l->bind_conf->frontend->maxconn;
620
621 return 1024;
622}
623
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200624/* This function is called on a read event from a listening socket, corresponding
625 * to an accept. It tries to accept as many connections as possible, and for each
626 * calls the listener's accept handler (generally the frontend's accept handler).
627 */
Willy Tarreauafad0e02012-08-09 14:45:22 +0200628void listener_accept(int fd)
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200629{
630 struct listener *l = fdtab[fd].owner;
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100631 struct proxy *p;
Christopher Faulet102854c2019-04-30 12:17:13 +0200632 unsigned int max_accept;
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100633 int next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100634 int next_feconn = 0;
635 int next_actconn = 0;
Willy Tarreaubb660302014-05-07 19:47:02 +0200636 int expire;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200637 int cfd;
638 int ret;
Willy Tarreau818dca52014-01-31 19:40:19 +0100639#ifdef USE_ACCEPT4
640 static int accept4_broken;
641#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200642
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100643 if (!l)
644 return;
645 p = l->bind_conf->frontend;
Christopher Faulet102854c2019-04-30 12:17:13 +0200646
647 /* if l->maxaccept is -1, then max_accept is UINT_MAX. It is not really
648 * illimited, but it is probably enough.
649 */
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100650 max_accept = l->maxaccept ? l->maxaccept : 1;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200651
Willy Tarreau93e7c002013-10-07 18:51:07 +0200652 if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
653 int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
Willy Tarreau93e7c002013-10-07 18:51:07 +0200654
655 if (unlikely(!max)) {
656 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200657 expire = tick_add(now_ms, next_event_delay(&global.sess_per_sec, global.sps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100658 goto limit_global;
Willy Tarreau93e7c002013-10-07 18:51:07 +0200659 }
660
661 if (max_accept > max)
662 max_accept = max;
663 }
664
665 if (!(l->options & LI_O_UNLIMITED) && global.cps_lim) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200666 int max = freq_ctr_remain(&global.conn_per_sec, global.cps_lim, 0);
667
668 if (unlikely(!max)) {
669 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200670 expire = tick_add(now_ms, next_event_delay(&global.conn_per_sec, global.cps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100671 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200672 }
673
674 if (max_accept > max)
675 max_accept = max;
676 }
Willy Tarreaue43d5322013-10-07 20:01:52 +0200677#ifdef USE_OPENSSL
678 if (!(l->options & LI_O_UNLIMITED) && global.ssl_lim && l->bind_conf && l->bind_conf->is_ssl) {
679 int max = freq_ctr_remain(&global.ssl_per_sec, global.ssl_lim, 0);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200680
Willy Tarreaue43d5322013-10-07 20:01:52 +0200681 if (unlikely(!max)) {
682 /* frontend accept rate limit was reached */
Willy Tarreaue43d5322013-10-07 20:01:52 +0200683 expire = tick_add(now_ms, next_event_delay(&global.ssl_per_sec, global.ssl_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100684 goto limit_global;
Willy Tarreaue43d5322013-10-07 20:01:52 +0200685 }
686
687 if (max_accept > max)
688 max_accept = max;
689 }
690#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200691 if (p && p->fe_sps_lim) {
692 int max = freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0);
693
694 if (unlikely(!max)) {
695 /* frontend accept rate limit was reached */
Willy Tarreau0591bf72019-12-10 12:01:21 +0100696 expire = tick_add(now_ms, next_event_delay(&p->fe_sess_per_sec, p->fe_sps_lim, 0));
697 goto limit_proxy;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200698 }
699
700 if (max_accept > max)
701 max_accept = max;
702 }
703
704 /* Note: if we fail to allocate a connection because of configured
705 * limits, we'll schedule a new attempt worst 1 second later in the
706 * worst case. If we fail due to system limits or temporary resource
707 * shortage, we try again 100ms later in the worst case.
708 */
Christopher Faulet102854c2019-04-30 12:17:13 +0200709 for (; max_accept; next_conn = next_feconn = next_actconn = 0, max_accept--) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200710 struct sockaddr_storage addr;
711 socklen_t laddr = sizeof(addr);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200712 unsigned int count;
Willy Tarreauaf613e82020-06-05 08:40:51 +0200713 __decl_thread(unsigned long mask);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200714
Willy Tarreau82c97892019-02-27 19:32:32 +0100715 /* pre-increase the number of connections without going too far.
716 * We process the listener, then the proxy, then the process.
717 * We know which ones to unroll based on the next_xxx value.
718 */
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100719 do {
720 count = l->nbconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100721 if (unlikely(l->maxconn && count >= l->maxconn)) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100722 /* the listener was marked full or another
723 * thread is going to do it.
724 */
725 next_conn = 0;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100726 listener_full(l);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100727 goto end;
728 }
729 next_conn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000730 } while (!_HA_ATOMIC_CAS(&l->nbconn, (int *)(&count), next_conn));
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100731
Willy Tarreau82c97892019-02-27 19:32:32 +0100732 if (p) {
733 do {
734 count = p->feconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100735 if (unlikely(count >= p->maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100736 /* the frontend was marked full or another
737 * thread is going to do it.
738 */
739 next_feconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100740 expire = TICK_ETERNITY;
741 goto limit_proxy;
Willy Tarreau82c97892019-02-27 19:32:32 +0100742 }
743 next_feconn = count + 1;
Olivier Houchard64213e92019-03-08 18:52:57 +0100744 } while (!_HA_ATOMIC_CAS(&p->feconn, &count, next_feconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200745 }
746
Willy Tarreau82c97892019-02-27 19:32:32 +0100747 if (!(l->options & LI_O_UNLIMITED)) {
748 do {
749 count = actconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100750 if (unlikely(count >= global.maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100751 /* the process was marked full or another
752 * thread is going to do it.
753 */
754 next_actconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100755 expire = tick_add(now_ms, 1000); /* try again in 1 second */
756 goto limit_global;
Willy Tarreau82c97892019-02-27 19:32:32 +0100757 }
758 next_actconn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000759 } while (!_HA_ATOMIC_CAS(&actconn, (int *)(&count), next_actconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200760 }
761
William Lallemand2fe7dd02018-09-11 16:51:29 +0200762 /* with sockpair@ we don't want to do an accept */
Willy Tarreau37159062020-08-27 07:48:42 +0200763 if (unlikely(l->rx.addr.ss_family == AF_CUST_SOCKPAIR)) {
William Lallemand2fe7dd02018-09-11 16:51:29 +0200764 if ((cfd = recv_fd_uxst(fd)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100765 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau888d5672019-01-27 18:34:12 +0100766 /* just like with UNIX sockets, only the family is filled */
767 addr.ss_family = AF_UNIX;
768 laddr = sizeof(addr.ss_family);
William Lallemand2fe7dd02018-09-11 16:51:29 +0200769 } else
770
Willy Tarreau1bc4aab2012-10-08 20:11:03 +0200771#ifdef USE_ACCEPT4
Willy Tarreau818dca52014-01-31 19:40:19 +0100772 /* only call accept4() if it's known to be safe, otherwise
773 * fallback to the legacy accept() + fcntl().
774 */
775 if (unlikely(accept4_broken ||
William Lallemandd9138002018-11-27 12:02:39 +0100776 ((cfd = accept4(fd, (struct sockaddr *)&addr, &laddr, SOCK_NONBLOCK)) == -1 &&
Willy Tarreau818dca52014-01-31 19:40:19 +0100777 (errno == ENOSYS || errno == EINVAL || errno == EBADF) &&
778 (accept4_broken = 1))))
779#endif
Willy Tarreau6b3b0d42012-10-22 19:32:55 +0200780 if ((cfd = accept(fd, (struct sockaddr *)&addr, &laddr)) != -1)
William Lallemandd9138002018-11-27 12:02:39 +0100781 fcntl(cfd, F_SETFL, O_NONBLOCK);
Willy Tarreau818dca52014-01-31 19:40:19 +0100782
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200783 if (unlikely(cfd == -1)) {
784 switch (errno) {
785 case EAGAIN:
Willy Tarreau20aeb1c2019-12-10 08:42:21 +0100786 if (fdtab[fd].ev & (FD_POLL_HUP|FD_POLL_ERR)) {
Willy Tarreaubb660302014-05-07 19:47:02 +0200787 /* the listening socket might have been disabled in a shared
788 * process and we're a collateral victim. We'll just pause for
789 * a while in case it comes back. In the mean time, we need to
790 * clear this sticky flag.
791 */
Willy Tarreau20aeb1c2019-12-10 08:42:21 +0100792 _HA_ATOMIC_AND(&fdtab[fd].ev, ~(FD_POLL_HUP|FD_POLL_ERR));
Willy Tarreaubb660302014-05-07 19:47:02 +0200793 goto transient_error;
794 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200795 goto end; /* nothing more to accept */
Willy Tarreaubb660302014-05-07 19:47:02 +0200796 case EINVAL:
797 /* might be trying to accept on a shut fd (eg: soft stop) */
798 goto transient_error;
Willy Tarreaua593ec52014-01-20 21:21:30 +0100799 case EINTR:
800 case ECONNABORTED:
Olivier Houchard64213e92019-03-08 18:52:57 +0100801 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100802 if (p)
Olivier Houchard64213e92019-03-08 18:52:57 +0100803 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100804 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +0100805 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreaua593ec52014-01-20 21:21:30 +0100806 continue;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200807 case ENFILE:
808 if (p)
809 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100810 "Proxy %s reached system FD limit (maxsock=%d). Please check system tunables.\n",
811 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200812 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200813 case EMFILE:
814 if (p)
815 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100816 "Proxy %s reached process FD limit (maxsock=%d). Please check 'ulimit-n' and restart.\n",
817 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200818 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200819 case ENOBUFS:
820 case ENOMEM:
821 if (p)
822 send_log(p, LOG_EMERG,
Willy Tarreauc5532ac2018-01-29 15:06:04 +0100823 "Proxy %s reached system memory limit (maxsock=%d). Please check system tunables.\n",
824 p->id, global.maxsock);
Willy Tarreaubb660302014-05-07 19:47:02 +0200825 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200826 default:
Willy Tarreaua593ec52014-01-20 21:21:30 +0100827 /* unexpected result, let's give up and let other tasks run */
Willy Tarreau92079932019-12-10 09:30:05 +0100828 max_accept = 0;
829 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200830 }
831 }
832
William Lallemandd9138002018-11-27 12:02:39 +0100833 /* we don't want to leak the FD upon reload if it's in the master */
834 if (unlikely(master == 1))
835 fcntl(cfd, F_SETFD, FD_CLOEXEC);
836
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100837 /* The connection was accepted, it must be counted as such */
838 if (l->counters)
839 HA_ATOMIC_UPDATE_MAX(&l->counters->conn_max, next_conn);
840
Willy Tarreau82c97892019-02-27 19:32:32 +0100841 if (p)
842 HA_ATOMIC_UPDATE_MAX(&p->fe_counters.conn_max, next_feconn);
843
844 proxy_inc_fe_conn_ctr(l, p);
845
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100846 if (!(l->options & LI_O_UNLIMITED)) {
847 count = update_freq_ctr(&global.conn_per_sec, 1);
848 HA_ATOMIC_UPDATE_MAX(&global.cps_max, count);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100849 }
850
Willy Tarreau64a9c052019-04-12 15:27:17 +0200851 _HA_ATOMIC_ADD(&activity[tid].accepted, 1);
852
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200853 if (unlikely(cfd >= global.maxsock)) {
854 send_log(p, LOG_EMERG,
855 "Proxy %s reached the configured maximum connection limit. Please check the global 'maxconn' value.\n",
856 p->id);
857 close(cfd);
Willy Tarreau0591bf72019-12-10 12:01:21 +0100858 expire = tick_add(now_ms, 1000); /* try again in 1 second */
859 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200860 }
861
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100862 /* past this point, l->accept() will automatically decrement
Willy Tarreau82c97892019-02-27 19:32:32 +0100863 * l->nbconn, feconn and actconn once done. Setting next_*conn=0
864 * allows the error path not to rollback on nbconn. It's more
865 * convenient than duplicating all exit labels.
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100866 */
867 next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100868 next_feconn = 0;
869 next_actconn = 0;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200870
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100871#if defined(USE_THREAD)
Willy Tarreau818a92e2020-09-03 07:50:19 +0200872 mask = thread_mask(l->rx.settings->bind_thread) & all_threads_mask;
Willy Tarreaua7da5e82020-03-12 17:33:29 +0100873 if (atleast2(mask) && (global.tune.options & GTUNE_LISTENER_MQ) && !stopping) {
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100874 struct accept_queue_ring *ring;
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100875 unsigned int t, t0, t1, t2;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100876
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100877 /* The principle is that we have two running indexes,
878 * each visiting in turn all threads bound to this
879 * listener. The connection will be assigned to the one
880 * with the least connections, and the other one will
881 * be updated. This provides a good fairness on short
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100882 * connections (round robin) and on long ones (conn
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100883 * count), without ever missing any idle thread.
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100884 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100885
886 /* keep a copy for the final update. thr_idx is composite
887 * and made of (t2<<16) + t1.
888 */
Willy Tarreau0cf33172019-03-06 15:26:33 +0100889 t0 = l->thr_idx;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100890 do {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100891 unsigned long m1, m2;
892 int q1, q2;
893
894 t2 = t1 = t0;
895 t2 >>= 16;
896 t1 &= 0xFFFF;
897
898 /* t1 walks low to high bits ;
899 * t2 walks high to low.
900 */
901 m1 = mask >> t1;
902 m2 = mask & (t2 ? nbits(t2 + 1) : ~0UL);
903
Willy Tarreau85d04242019-04-16 18:09:13 +0200904 if (unlikely(!(m1 & 1))) {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100905 m1 &= ~1UL;
906 if (!m1) {
907 m1 = mask;
908 t1 = 0;
909 }
910 t1 += my_ffsl(m1) - 1;
911 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100912
Willy Tarreau85d04242019-04-16 18:09:13 +0200913 if (unlikely(!(m2 & (1UL << t2)) || t1 == t2)) {
914 /* highest bit not set */
915 if (!m2)
916 m2 = mask;
917
918 t2 = my_flsl(m2) - 1;
919 }
920
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100921 /* now we have two distinct thread IDs belonging to the mask */
922 q1 = accept_queue_rings[t1].tail - accept_queue_rings[t1].head + ACCEPT_QUEUE_SIZE;
923 if (q1 >= ACCEPT_QUEUE_SIZE)
924 q1 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100925
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100926 q2 = accept_queue_rings[t2].tail - accept_queue_rings[t2].head + ACCEPT_QUEUE_SIZE;
927 if (q2 >= ACCEPT_QUEUE_SIZE)
928 q2 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100929
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100930 /* we have 3 possibilities now :
931 * q1 < q2 : t1 is less loaded than t2, so we pick it
932 * and update t2 (since t1 might still be
933 * lower than another thread)
934 * q1 > q2 : t2 is less loaded than t1, so we pick it
935 * and update t1 (since t2 might still be
936 * lower than another thread)
937 * q1 = q2 : both are equally loaded, thus we pick t1
938 * and update t1 as it will become more loaded
939 * than t2.
940 */
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100941
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100942 q1 += l->thr_conn[t1];
943 q2 += l->thr_conn[t2];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100944
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100945 if (q1 - q2 < 0) {
946 t = t1;
947 t2 = t2 ? t2 - 1 : LONGBITS - 1;
948 }
949 else if (q1 - q2 > 0) {
950 t = t2;
951 t1++;
952 if (t1 >= LONGBITS)
953 t1 = 0;
954 }
955 else {
956 t = t1;
957 t1++;
958 if (t1 >= LONGBITS)
959 t1 = 0;
960 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100961
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100962 /* new value for thr_idx */
963 t1 += (t2 << 16);
Olivier Houchard64213e92019-03-08 18:52:57 +0100964 } while (unlikely(!_HA_ATOMIC_CAS(&l->thr_idx, &t0, t1)));
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100965
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100966 /* We successfully selected the best thread "t" for this
967 * connection. We use deferred accepts even if it's the
968 * local thread because tests show that it's the best
969 * performing model, likely due to better cache locality
970 * when processing this loop.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100971 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100972 ring = &accept_queue_rings[t];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100973 if (accept_queue_push_mp(ring, cfd, l, &addr, laddr)) {
Olivier Houchard64213e92019-03-08 18:52:57 +0100974 _HA_ATOMIC_ADD(&activity[t].accq_pushed, 1);
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200975 tasklet_wakeup(ring->tasklet);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100976 continue;
977 }
978 /* If the ring is full we do a synchronous accept on
979 * the local thread here.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100980 */
Olivier Houchard64213e92019-03-08 18:52:57 +0100981 _HA_ATOMIC_ADD(&activity[t].accq_full, 1);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100982 }
983#endif // USE_THREAD
984
Olivier Houchard64213e92019-03-08 18:52:57 +0100985 _HA_ATOMIC_ADD(&l->thr_conn[tid], 1);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200986 ret = l->accept(l, cfd, &addr);
987 if (unlikely(ret <= 0)) {
Willy Tarreau87b09662015-04-03 00:22:06 +0200988 /* The connection was closed by stream_accept(). Either
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200989 * we just have to ignore it (ret == 0) or it's a critical
990 * error due to a resource shortage, and we must stop the
991 * listener (ret < 0).
992 */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200993 if (ret == 0) /* successful termination */
994 continue;
995
Willy Tarreaubb660302014-05-07 19:47:02 +0200996 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200997 }
998
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100999 /* increase the per-process number of cumulated sessions, this
1000 * may only be done once l->accept() has accepted the connection.
1001 */
Willy Tarreau93e7c002013-10-07 18:51:07 +02001002 if (!(l->options & LI_O_UNLIMITED)) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001003 count = update_freq_ctr(&global.sess_per_sec, 1);
1004 HA_ATOMIC_UPDATE_MAX(&global.sps_max, count);
Willy Tarreau93e7c002013-10-07 18:51:07 +02001005 }
Willy Tarreaue43d5322013-10-07 20:01:52 +02001006#ifdef USE_OPENSSL
1007 if (!(l->options & LI_O_UNLIMITED) && l->bind_conf && l->bind_conf->is_ssl) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001008 count = update_freq_ctr(&global.ssl_per_sec, 1);
1009 HA_ATOMIC_UPDATE_MAX(&global.ssl_max, count);
Willy Tarreaue43d5322013-10-07 20:01:52 +02001010 }
1011#endif
Willy Tarreau93e7c002013-10-07 18:51:07 +02001012
Willy Tarreau8d2c98b2020-05-01 09:51:11 +02001013 ti->flags &= ~TI_FL_STUCK; // this thread is still running
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001014 } /* end of for (max_accept--) */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001015
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001016 end:
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001017 if (next_conn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001018 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau741b4d62019-02-25 15:02:04 +01001019
Willy Tarreau82c97892019-02-27 19:32:32 +01001020 if (p && next_feconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001021 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001022
1023 if (next_actconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001024 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001025
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001026 if ((l->state == LI_FULL && (!l->maxconn || l->nbconn < l->maxconn)) ||
Willy Tarreaucdcba112019-12-11 15:06:30 +01001027 (l->state == LI_LIMITED &&
1028 ((!p || p->feconn < p->maxconn) && (actconn < global.maxconn) &&
1029 (!tick_isset(global_listener_queue_task->expire) ||
1030 tick_is_expired(global_listener_queue_task->expire, now_ms))))) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001031 /* at least one thread has to this when quitting */
1032 resume_listener(l);
1033
1034 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001035 dequeue_all_listeners();
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001036
Olivier Houchard859dc802019-08-08 15:47:21 +02001037 if (p && !MT_LIST_ISEMPTY(&p->listener_queue) &&
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001038 (!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 +01001039 dequeue_proxy_listeners(p);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001040 }
Willy Tarreau4c044e22019-12-05 07:40:32 +01001041
Willy Tarreau92079932019-12-10 09:30:05 +01001042 /* Now it's getting tricky. The listener was supposed to be in LI_READY
1043 * state but in the mean time we might have changed it to LI_FULL or
1044 * LI_LIMITED, and another thread might also have turned it to
1045 * LI_PAUSED, LI_LISTEN or even LI_INI when stopping a proxy. We must
1046 * be certain to keep the FD enabled when in the READY state but we
1047 * must also stop it for other states that we might have switched to
1048 * while others re-enabled polling.
1049 */
1050 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
1051 if (l->state == LI_READY) {
1052 if (max_accept > 0)
1053 fd_cant_recv(fd);
1054 else
1055 fd_done_recv(fd);
1056 } else if (l->state > LI_ASSIGNED) {
Willy Tarreau38ba6472020-08-27 08:16:52 +02001057 fd_stop_recv(l->rx.fd);
Willy Tarreau92079932019-12-10 09:30:05 +01001058 }
1059 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau0591bf72019-12-10 12:01:21 +01001060 return;
1061
1062 transient_error:
1063 /* pause the listener for up to 100 ms */
1064 expire = tick_add(now_ms, 100);
1065
1066 limit_global:
1067 /* (re-)queue the listener to the global queue and set it to expire no
1068 * later than <expire> ahead. The listener turns to LI_LIMITED.
1069 */
1070 limit_listener(l, &global_listener_queue);
1071 task_schedule(global_listener_queue_task, expire);
1072 goto end;
1073
1074 limit_proxy:
1075 /* (re-)queue the listener to the proxy's queue and set it to expire no
1076 * later than <expire> ahead. The listener turns to LI_LIMITED.
1077 */
1078 limit_listener(l, &p->listener_queue);
Willy Tarreaueeea8082020-01-08 19:15:07 +01001079 if (p->task && tick_isset(expire))
1080 task_schedule(p->task, expire);
Willy Tarreau0591bf72019-12-10 12:01:21 +01001081 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001082}
1083
Willy Tarreau05f50472017-09-15 09:19:58 +02001084/* Notify the listener that a connection initiated from it was released. This
1085 * is used to keep the connection count consistent and to possibly re-open
1086 * listening when it was limited.
1087 */
1088void listener_release(struct listener *l)
1089{
1090 struct proxy *fe = l->bind_conf->frontend;
1091
1092 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +01001093 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001094 if (fe)
Olivier Houchard64213e92019-03-08 18:52:57 +01001095 _HA_ATOMIC_SUB(&fe->feconn, 1);
1096 _HA_ATOMIC_SUB(&l->nbconn, 1);
1097 _HA_ATOMIC_SUB(&l->thr_conn[tid], 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001098
1099 if (l->state == LI_FULL || l->state == LI_LIMITED)
Willy Tarreau05f50472017-09-15 09:19:58 +02001100 resume_listener(l);
1101
1102 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001103 dequeue_all_listeners();
Willy Tarreau05f50472017-09-15 09:19:58 +02001104
Olivier Houchard859dc802019-08-08 15:47:21 +02001105 if (!MT_LIST_ISEMPTY(&fe->listener_queue) &&
Willy Tarreau05f50472017-09-15 09:19:58 +02001106 (!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 +01001107 dequeue_proxy_listeners(fe);
Willy Tarreau05f50472017-09-15 09:19:58 +02001108}
1109
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001110/* Initializes the listener queues. Returns 0 on success, otherwise ERR_* flags */
1111static int listener_queue_init()
1112{
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001113 global_listener_queue_task = task_new(MAX_THREADS_MASK);
1114 if (!global_listener_queue_task) {
1115 ha_alert("Out of memory when initializing global listener queue\n");
1116 return ERR_FATAL|ERR_ABORT;
1117 }
1118 /* very simple initialization, users will queue the task if needed */
1119 global_listener_queue_task->context = NULL; /* not even a context! */
1120 global_listener_queue_task->process = manage_global_listener_queue;
1121
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001122 return 0;
1123}
1124
1125static void listener_queue_deinit()
1126{
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001127 task_destroy(global_listener_queue_task);
1128 global_listener_queue_task = NULL;
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001129}
1130
1131REGISTER_CONFIG_POSTPARSER("multi-threaded listener queue", listener_queue_init);
1132REGISTER_POST_DEINIT(listener_queue_deinit);
1133
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001134
1135/* This is the global management task for listeners. It enables listeners waiting
1136 * for global resources when there are enough free resource, or at least once in
1137 * a while. It is designed to be called as a task.
1138 */
1139static struct task *manage_global_listener_queue(struct task *t, void *context, unsigned short state)
1140{
1141 /* If there are still too many concurrent connections, let's wait for
1142 * some of them to go away. We don't need to re-arm the timer because
1143 * each of them will scan the queue anyway.
1144 */
1145 if (unlikely(actconn >= global.maxconn))
1146 goto out;
1147
1148 /* We should periodically try to enable listeners waiting for a global
1149 * resource here, because it is possible, though very unlikely, that
1150 * they have been blocked by a temporary lack of global resource such
1151 * as a file descriptor or memory and that the temporary condition has
1152 * disappeared.
1153 */
1154 dequeue_all_listeners();
1155
1156 out:
1157 t->expire = TICK_ETERNITY;
1158 task_queue(t);
1159 return t;
1160}
1161
Willy Tarreau26982662012-09-12 23:17:10 +02001162/*
1163 * Registers the bind keyword list <kwl> as a list of valid keywords for next
1164 * parsing sessions.
1165 */
1166void bind_register_keywords(struct bind_kw_list *kwl)
1167{
1168 LIST_ADDQ(&bind_keywords.list, &kwl->list);
1169}
1170
1171/* Return a pointer to the bind keyword <kw>, or NULL if not found. If the
1172 * keyword is found with a NULL ->parse() function, then an attempt is made to
1173 * find one with a valid ->parse() function. This way it is possible to declare
1174 * platform-dependant, known keywords as NULL, then only declare them as valid
1175 * if some options are met. Note that if the requested keyword contains an
1176 * opening parenthesis, everything from this point is ignored.
1177 */
1178struct bind_kw *bind_find_kw(const char *kw)
1179{
1180 int index;
1181 const char *kwend;
1182 struct bind_kw_list *kwl;
1183 struct bind_kw *ret = NULL;
1184
1185 kwend = strchr(kw, '(');
1186 if (!kwend)
1187 kwend = kw + strlen(kw);
1188
1189 list_for_each_entry(kwl, &bind_keywords.list, list) {
1190 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1191 if ((strncmp(kwl->kw[index].kw, kw, kwend - kw) == 0) &&
1192 kwl->kw[index].kw[kwend-kw] == 0) {
1193 if (kwl->kw[index].parse)
1194 return &kwl->kw[index]; /* found it !*/
1195 else
1196 ret = &kwl->kw[index]; /* may be OK */
1197 }
1198 }
1199 }
1200 return ret;
1201}
1202
Willy Tarreau8638f482012-09-18 18:01:17 +02001203/* Dumps all registered "bind" keywords to the <out> string pointer. The
1204 * unsupported keywords are only dumped if their supported form was not
1205 * found.
1206 */
1207void bind_dump_kws(char **out)
1208{
1209 struct bind_kw_list *kwl;
1210 int index;
1211
Christopher Faulet784063e2020-05-18 12:14:18 +02001212 if (!out)
1213 return;
1214
Willy Tarreau8638f482012-09-18 18:01:17 +02001215 *out = NULL;
1216 list_for_each_entry(kwl, &bind_keywords.list, list) {
1217 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1218 if (kwl->kw[index].parse ||
1219 bind_find_kw(kwl->kw[index].kw) == &kwl->kw[index]) {
Willy Tarreau51fb7652012-09-18 18:24:39 +02001220 memprintf(out, "%s[%4s] %s%s%s\n", *out ? *out : "",
1221 kwl->scope,
Willy Tarreau8638f482012-09-18 18:01:17 +02001222 kwl->kw[index].kw,
Willy Tarreau51fb7652012-09-18 18:24:39 +02001223 kwl->kw[index].skip ? " <arg>" : "",
1224 kwl->kw[index].parse ? "" : " (not supported)");
Willy Tarreau8638f482012-09-18 18:01:17 +02001225 }
1226 }
1227 }
1228}
1229
Willy Tarreau645513a2010-05-24 20:55:15 +02001230/************************************************************************/
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001231/* All supported sample and ACL keywords must be declared here. */
Willy Tarreau645513a2010-05-24 20:55:15 +02001232/************************************************************************/
1233
Willy Tarreaua5e37562011-12-16 17:06:15 +01001234/* set temp integer to the number of connexions to the same listening socket */
Willy Tarreau645513a2010-05-24 20:55:15 +02001235static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001236smp_fetch_dconn(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau645513a2010-05-24 20:55:15 +02001237{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001238 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001239 smp->data.u.sint = smp->sess->listener->nbconn;
Willy Tarreau645513a2010-05-24 20:55:15 +02001240 return 1;
1241}
1242
Willy Tarreaua5e37562011-12-16 17:06:15 +01001243/* set temp integer to the id of the socket (listener) */
Willy Tarreau645513a2010-05-24 20:55:15 +02001244static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001245smp_fetch_so_id(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau37406352012-04-23 16:16:37 +02001246{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001247 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001248 smp->data.u.sint = smp->sess->listener->luid;
Willy Tarreau645513a2010-05-24 20:55:15 +02001249 return 1;
1250}
Jerome Magnineb421b22020-03-27 22:08:40 +01001251static int
1252smp_fetch_so_name(const struct arg *args, struct sample *smp, const char *kw, void *private)
1253{
1254 smp->data.u.str.area = smp->sess->listener->name;
1255 if (!smp->data.u.str.area)
1256 return 0;
1257
1258 smp->data.type = SMP_T_STR;
1259 smp->flags = SMP_F_CONST;
1260 smp->data.u.str.data = strlen(smp->data.u.str.area);
1261 return 1;
1262}
Willy Tarreau645513a2010-05-24 20:55:15 +02001263
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001264/* parse the "accept-proxy" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001265static 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 +02001266{
1267 struct listener *l;
1268
Willy Tarreau4348fad2012-09-20 16:48:07 +02001269 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001270 l->options |= LI_O_ACC_PROXY;
1271
1272 return 0;
1273}
1274
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001275/* parse the "accept-netscaler-cip" bind keyword */
1276static int bind_parse_accept_netscaler_cip(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1277{
1278 struct listener *l;
1279 uint32_t val;
1280
1281 if (!*args[cur_arg + 1]) {
1282 memprintf(err, "'%s' : missing value", args[cur_arg]);
1283 return ERR_ALERT | ERR_FATAL;
1284 }
1285
1286 val = atol(args[cur_arg + 1]);
1287 if (val <= 0) {
Willy Tarreaue2711c72019-02-27 15:39:41 +01001288 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001289 return ERR_ALERT | ERR_FATAL;
1290 }
1291
1292 list_for_each_entry(l, &conf->listeners, by_bind) {
1293 l->options |= LI_O_ACC_CIP;
1294 conf->ns_cip_magic = val;
1295 }
1296
1297 return 0;
1298}
1299
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001300/* parse the "backlog" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001301static 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 +02001302{
1303 struct listener *l;
1304 int val;
1305
1306 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001307 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001308 return ERR_ALERT | ERR_FATAL;
1309 }
1310
1311 val = atol(args[cur_arg + 1]);
Willy Tarreaue2711c72019-02-27 15:39:41 +01001312 if (val < 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001313 memprintf(err, "'%s' : invalid value %d, must be > 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001314 return ERR_ALERT | ERR_FATAL;
1315 }
1316
Willy Tarreau4348fad2012-09-20 16:48:07 +02001317 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001318 l->backlog = val;
1319
1320 return 0;
1321}
1322
1323/* parse the "id" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001324static 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 +02001325{
1326 struct eb32_node *node;
Willy Tarreau4348fad2012-09-20 16:48:07 +02001327 struct listener *l, *new;
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001328 char *error;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001329
Willy Tarreau4348fad2012-09-20 16:48:07 +02001330 if (conf->listeners.n != conf->listeners.p) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001331 memprintf(err, "'%s' can only be used with a single socket", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001332 return ERR_ALERT | ERR_FATAL;
1333 }
1334
1335 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001336 memprintf(err, "'%s' : expects an integer argument", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001337 return ERR_ALERT | ERR_FATAL;
1338 }
1339
Willy Tarreau4348fad2012-09-20 16:48:07 +02001340 new = LIST_NEXT(&conf->listeners, struct listener *, by_bind);
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001341 new->luid = strtol(args[cur_arg + 1], &error, 10);
1342 if (*error != '\0') {
1343 memprintf(err, "'%s' : expects an integer argument, found '%s'", args[cur_arg], args[cur_arg + 1]);
1344 return ERR_ALERT | ERR_FATAL;
1345 }
Willy Tarreau4348fad2012-09-20 16:48:07 +02001346 new->conf.id.key = new->luid;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001347
Willy Tarreau4348fad2012-09-20 16:48:07 +02001348 if (new->luid <= 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001349 memprintf(err, "'%s' : custom id has to be > 0", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001350 return ERR_ALERT | ERR_FATAL;
1351 }
1352
Willy Tarreau4348fad2012-09-20 16:48:07 +02001353 node = eb32_lookup(&px->conf.used_listener_id, new->luid);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001354 if (node) {
1355 l = container_of(node, struct listener, conf.id);
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001356 memprintf(err, "'%s' : custom id %d already used at %s:%d ('bind %s')",
1357 args[cur_arg], l->luid, l->bind_conf->file, l->bind_conf->line,
1358 l->bind_conf->arg);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001359 return ERR_ALERT | ERR_FATAL;
1360 }
1361
Willy Tarreau4348fad2012-09-20 16:48:07 +02001362 eb32_insert(&px->conf.used_listener_id, &new->conf.id);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001363 return 0;
1364}
1365
1366/* parse the "maxconn" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001367static 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 +02001368{
1369 struct listener *l;
1370 int val;
1371
1372 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001373 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001374 return ERR_ALERT | ERR_FATAL;
1375 }
1376
1377 val = atol(args[cur_arg + 1]);
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001378 if (val < 0) {
1379 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001380 return ERR_ALERT | ERR_FATAL;
1381 }
1382
Willy Tarreau4348fad2012-09-20 16:48:07 +02001383 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001384 l->maxconn = val;
1385
1386 return 0;
1387}
1388
1389/* parse the "name" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001390static 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 +02001391{
1392 struct listener *l;
1393
1394 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001395 memprintf(err, "'%s' : missing name", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001396 return ERR_ALERT | ERR_FATAL;
1397 }
1398
Willy Tarreau4348fad2012-09-20 16:48:07 +02001399 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001400 l->name = strdup(args[cur_arg + 1]);
1401
1402 return 0;
1403}
1404
1405/* parse the "nice" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001406static 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 +02001407{
1408 struct listener *l;
1409 int val;
1410
1411 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001412 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001413 return ERR_ALERT | ERR_FATAL;
1414 }
1415
1416 val = atol(args[cur_arg + 1]);
1417 if (val < -1024 || val > 1024) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001418 memprintf(err, "'%s' : invalid value %d, allowed range is -1024..1024", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001419 return ERR_ALERT | ERR_FATAL;
1420 }
1421
Willy Tarreau4348fad2012-09-20 16:48:07 +02001422 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001423 l->nice = val;
1424
1425 return 0;
1426}
1427
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001428/* parse the "process" bind keyword */
1429static int bind_parse_process(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1430{
Christopher Fauletc644fa92017-11-23 22:44:11 +01001431 char *slash;
1432 unsigned long proc = 0, thread = 0;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001433
Christopher Fauletc644fa92017-11-23 22:44:11 +01001434 if ((slash = strchr(args[cur_arg + 1], '/')) != NULL)
1435 *slash = 0;
1436
Willy Tarreauff9c9142019-02-07 10:39:36 +01001437 if (parse_process_number(args[cur_arg + 1], &proc, MAX_PROCS, NULL, err)) {
Christopher Fauletf1f0c5f2017-11-22 12:06:43 +01001438 memprintf(err, "'%s' : %s", args[cur_arg], *err);
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001439 return ERR_ALERT | ERR_FATAL;
1440 }
1441
Christopher Fauletc644fa92017-11-23 22:44:11 +01001442 if (slash) {
Willy Tarreauc9a82e42019-01-26 13:25:14 +01001443 if (parse_process_number(slash+1, &thread, MAX_THREADS, NULL, err)) {
Christopher Fauletc644fa92017-11-23 22:44:11 +01001444 memprintf(err, "'%s' : %s", args[cur_arg], *err);
1445 return ERR_ALERT | ERR_FATAL;
1446 }
1447 *slash = '/';
1448 }
1449
Willy Tarreaue26993c2020-09-03 07:18:55 +02001450 conf->settings.bind_proc |= proc;
1451 conf->settings.bind_thread |= thread;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001452 return 0;
1453}
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001454
Christopher Fauleta717b992018-04-10 14:43:00 +02001455/* parse the "proto" bind keyword */
1456static int bind_parse_proto(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1457{
1458 struct ist proto;
1459
1460 if (!*args[cur_arg + 1]) {
1461 memprintf(err, "'%s' : missing value", args[cur_arg]);
1462 return ERR_ALERT | ERR_FATAL;
1463 }
1464
1465 proto = ist2(args[cur_arg + 1], strlen(args[cur_arg + 1]));
1466 conf->mux_proto = get_mux_proto(proto);
1467 if (!conf->mux_proto) {
1468 memprintf(err, "'%s' : unknown MUX protocol '%s'", args[cur_arg], args[cur_arg+1]);
1469 return ERR_ALERT | ERR_FATAL;
1470 }
Christopher Fauleta717b992018-04-10 14:43:00 +02001471 return 0;
1472}
1473
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001474/* config parser for global "tune.listener.multi-queue", accepts "on" or "off" */
1475static int cfg_parse_tune_listener_mq(char **args, int section_type, struct proxy *curpx,
1476 struct proxy *defpx, const char *file, int line,
1477 char **err)
1478{
1479 if (too_many_args(1, args, err, NULL))
1480 return -1;
1481
1482 if (strcmp(args[1], "on") == 0)
1483 global.tune.options |= GTUNE_LISTENER_MQ;
1484 else if (strcmp(args[1], "off") == 0)
1485 global.tune.options &= ~GTUNE_LISTENER_MQ;
1486 else {
1487 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
1488 return -1;
1489 }
1490 return 0;
1491}
1492
Willy Tarreau61612d42012-04-19 18:42:05 +02001493/* Note: must not be declared <const> as its list will be overwritten.
1494 * Please take care of keeping this list alphabetically sorted.
1495 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001496static struct sample_fetch_kw_list smp_kws = {ILH, {
Thierry FOURNIER07ee64e2015-07-06 23:43:03 +02001497 { "dst_conn", smp_fetch_dconn, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
1498 { "so_id", smp_fetch_so_id, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
Jerome Magnineb421b22020-03-27 22:08:40 +01001499 { "so_name", smp_fetch_so_name, 0, NULL, SMP_T_STR, SMP_USE_FTEND, },
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001500 { /* END */ },
1501}};
1502
Willy Tarreau0108d902018-11-25 19:14:37 +01001503INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
1504
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001505/* Note: must not be declared <const> as its list will be overwritten.
1506 * Please take care of keeping this list alphabetically sorted.
1507 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001508static struct acl_kw_list acl_kws = {ILH, {
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001509 { /* END */ },
Willy Tarreau645513a2010-05-24 20:55:15 +02001510}};
1511
Willy Tarreau0108d902018-11-25 19:14:37 +01001512INITCALL1(STG_REGISTER, acl_register_keywords, &acl_kws);
1513
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001514/* Note: must not be declared <const> as its list will be overwritten.
1515 * Please take care of keeping this list alphabetically sorted, doing so helps
1516 * all code contributors.
1517 * Optional keywords are also declared with a NULL ->parse() function so that
1518 * the config parser can report an appropriate error when a known keyword was
1519 * not enabled.
1520 */
Willy Tarreau51fb7652012-09-18 18:24:39 +02001521static struct bind_kw_list bind_kws = { "ALL", { }, {
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001522 { "accept-netscaler-cip", bind_parse_accept_netscaler_cip, 1 }, /* enable NetScaler Client IP insertion protocol */
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001523 { "accept-proxy", bind_parse_accept_proxy, 0 }, /* enable PROXY protocol */
1524 { "backlog", bind_parse_backlog, 1 }, /* set backlog of listening socket */
1525 { "id", bind_parse_id, 1 }, /* set id of listening socket */
1526 { "maxconn", bind_parse_maxconn, 1 }, /* set maxconn of listening socket */
1527 { "name", bind_parse_name, 1 }, /* set name of listening socket */
1528 { "nice", bind_parse_nice, 1 }, /* set nice of listening socket */
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001529 { "process", bind_parse_process, 1 }, /* set list of allowed process for this socket */
Christopher Fauleta717b992018-04-10 14:43:00 +02001530 { "proto", bind_parse_proto, 1 }, /* set the proto to use for all incoming connections */
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001531 { /* END */ },
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001532}};
1533
Willy Tarreau0108d902018-11-25 19:14:37 +01001534INITCALL1(STG_REGISTER, bind_register_keywords, &bind_kws);
1535
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001536/* config keyword parsers */
1537static struct cfg_kw_list cfg_kws = {ILH, {
1538 { CFG_GLOBAL, "tune.listener.multi-queue", cfg_parse_tune_listener_mq },
1539 { 0, NULL, NULL }
1540}};
1541
1542INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
1543
Willy Tarreau645513a2010-05-24 20:55:15 +02001544/*
1545 * Local variables:
1546 * c-indent-level: 8
1547 * c-basic-offset: 8
1548 * End:
1549 */