blob: 85191f50bac7a01138d7399b5ad121ba566c6d2c [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 Tarreau6ae1ba62014-05-07 19:01:58 +020013#include <ctype.h>
Willy Tarreaubbebbbf2012-05-07 21:22:09 +020014#include <errno.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020015#include <stdio.h>
16#include <string.h>
Willy Tarreau95ccdde2014-02-01 09:28:36 +010017#include <unistd.h>
18#include <fcntl.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020019
Willy Tarreaudcc048a2020-06-04 19:11:43 +020020#include <haproxy/acl.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020021#include <haproxy/api.h>
Willy Tarreau6be78492020-06-05 00:00:29 +020022#include <haproxy/cfgparse.h>
Willy Tarreau7ea393d2020-06-04 18:02:10 +020023#include <haproxy/connection.h>
Willy Tarreau8d366972020-05-27 16:10:29 +020024#include <haproxy/errors.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020025#include <haproxy/fd.h>
26#include <haproxy/freq_ctr.h>
Willy Tarreauf268ee82020-06-04 17:05:57 +020027#include <haproxy/global.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020028#include <haproxy/list.h>
Willy Tarreau213e9902020-06-04 14:58:24 +020029#include <haproxy/listener.h>
Willy Tarreauaeed4a82020-06-04 22:01:04 +020030#include <haproxy/log.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020031#include <haproxy/protocol-t.h>
32#include <haproxy/protocol.h>
33#include <haproxy/sample.h>
Willy Tarreaudfd3de82020-06-04 23:46:14 +020034#include <haproxy/stream.h>
Willy Tarreaucea0e1b2020-06-04 17:25:40 +020035#include <haproxy/task.h>
Willy Tarreau92b4f132020-06-01 11:05:15 +020036#include <haproxy/time.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020037#include <haproxy/tools.h>
Willy Tarreaudd815982007-10-16 12:25:14 +020038
Willy Tarreaub648d632007-10-28 22:13:50 +010039
Willy Tarreau26982662012-09-12 23:17:10 +020040/* List head of all known bind keywords */
41static struct bind_kw_list bind_keywords = {
42 .list = LIST_HEAD_INIT(bind_keywords.list)
43};
44
Willy Tarreaua1d97f82019-12-10 11:18:41 +010045/* list of the temporarily limited listeners because of lack of resource */
46static struct mt_list global_listener_queue = MT_LIST_HEAD_INIT(global_listener_queue);
47static struct task *global_listener_queue_task;
48static struct task *manage_global_listener_queue(struct task *t, void *context, unsigned short state);
49
50
Willy Tarreau1efafce2019-01-27 15:37:19 +010051#if defined(USE_THREAD)
52
53struct accept_queue_ring accept_queue_rings[MAX_THREADS] __attribute__((aligned(64))) = { };
54
55/* dequeue and process a pending connection from the local accept queue (single
Willy Tarreau83efc322020-10-14 17:37:17 +020056 * consumer). Returns the accepted connection or NULL if none was found.
Willy Tarreau1efafce2019-01-27 15:37:19 +010057 */
Willy Tarreau83efc322020-10-14 17:37:17 +020058struct connection *accept_queue_pop_sc(struct accept_queue_ring *ring)
Willy Tarreau1efafce2019-01-27 15:37:19 +010059{
Willy Tarreau1efafce2019-01-27 15:37:19 +010060 unsigned int pos, next;
Willy Tarreau83efc322020-10-14 17:37:17 +020061 struct connection *ptr;
62 struct connection **e;
Willy Tarreau1efafce2019-01-27 15:37:19 +010063
64 pos = ring->head;
65
66 if (pos == ring->tail)
Willy Tarreau83efc322020-10-14 17:37:17 +020067 return NULL;
Willy Tarreau1efafce2019-01-27 15:37:19 +010068
69 next = pos + 1;
70 if (next >= ACCEPT_QUEUE_SIZE)
71 next = 0;
72
73 e = &ring->entry[pos];
74
75 /* wait for the producer to update the listener's pointer */
76 while (1) {
Willy Tarreau83efc322020-10-14 17:37:17 +020077 ptr = *e;
Willy Tarreau1efafce2019-01-27 15:37:19 +010078 __ha_barrier_load();
79 if (ptr)
80 break;
81 pl_cpu_relax();
82 }
83
Willy Tarreau1efafce2019-01-27 15:37:19 +010084 /* release the entry */
Willy Tarreau83efc322020-10-14 17:37:17 +020085 *e = NULL;
Willy Tarreau1efafce2019-01-27 15:37:19 +010086
87 __ha_barrier_store();
88 ring->head = next;
Willy Tarreau83efc322020-10-14 17:37:17 +020089 return ptr;
Willy Tarreau1efafce2019-01-27 15:37:19 +010090}
91
92
Willy Tarreau83efc322020-10-14 17:37:17 +020093/* tries to push a new accepted connection <conn> into ring <ring>. Returns
94 * non-zero if it succeeds, or zero if the ring is full. Supports multiple
95 * producers.
Willy Tarreau1efafce2019-01-27 15:37:19 +010096 */
Willy Tarreau83efc322020-10-14 17:37:17 +020097int accept_queue_push_mp(struct accept_queue_ring *ring, struct connection *conn)
Willy Tarreau1efafce2019-01-27 15:37:19 +010098{
Willy Tarreau1efafce2019-01-27 15:37:19 +010099 unsigned int pos, next;
100
101 pos = ring->tail;
102 do {
103 next = pos + 1;
104 if (next >= ACCEPT_QUEUE_SIZE)
105 next = 0;
106 if (next == ring->head)
107 return 0; // ring full
Olivier Houchard64213e92019-03-08 18:52:57 +0100108 } while (unlikely(!_HA_ATOMIC_CAS(&ring->tail, &pos, next)));
Willy Tarreau1efafce2019-01-27 15:37:19 +0100109
Willy Tarreau83efc322020-10-14 17:37:17 +0200110 ring->entry[pos] = conn;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100111 __ha_barrier_store();
Willy Tarreau1efafce2019-01-27 15:37:19 +0100112 return 1;
113}
114
115/* proceed with accepting new connections */
116static struct task *accept_queue_process(struct task *t, void *context, unsigned short state)
117{
118 struct accept_queue_ring *ring = context;
Willy Tarreau83efc322020-10-14 17:37:17 +0200119 struct connection *conn;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100120 struct listener *li;
Christopher Faulet102854c2019-04-30 12:17:13 +0200121 unsigned int max_accept;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100122 int ret;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100123
Christopher Faulet102854c2019-04-30 12:17:13 +0200124 /* if global.tune.maxaccept is -1, then max_accept is UINT_MAX. It
125 * is not really illimited, but it is probably enough.
126 */
127 max_accept = global.tune.maxaccept ? global.tune.maxaccept : 64;
128 for (; max_accept; max_accept--) {
Willy Tarreau83efc322020-10-14 17:37:17 +0200129 conn = accept_queue_pop_sc(ring);
130 if (!conn)
Willy Tarreau1efafce2019-01-27 15:37:19 +0100131 break;
132
Willy Tarreau83efc322020-10-14 17:37:17 +0200133 li = __objt_listener(conn->target);
Olivier Houchard64213e92019-03-08 18:52:57 +0100134 _HA_ATOMIC_ADD(&li->thr_conn[tid], 1);
Willy Tarreau83efc322020-10-14 17:37:17 +0200135 ret = li->accept(conn);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100136 if (ret <= 0) {
137 /* connection was terminated by the application */
138 continue;
139 }
140
141 /* increase the per-process number of cumulated sessions, this
142 * may only be done once l->accept() has accepted the connection.
143 */
144 if (!(li->options & LI_O_UNLIMITED)) {
145 HA_ATOMIC_UPDATE_MAX(&global.sps_max,
146 update_freq_ctr(&global.sess_per_sec, 1));
147 if (li->bind_conf && li->bind_conf->is_ssl) {
148 HA_ATOMIC_UPDATE_MAX(&global.ssl_max,
149 update_freq_ctr(&global.ssl_per_sec, 1));
150 }
151 }
152 }
153
154 /* ran out of budget ? Let's come here ASAP */
Christopher Faulet102854c2019-04-30 12:17:13 +0200155 if (!max_accept)
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200156 tasklet_wakeup(ring->tasklet);
Willy Tarreau1efafce2019-01-27 15:37:19 +0100157
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200158 return NULL;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100159}
160
161/* Initializes the accept-queues. Returns 0 on success, otherwise ERR_* flags */
162static int accept_queue_init()
163{
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200164 struct tasklet *t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100165 int i;
166
167 for (i = 0; i < global.nbthread; i++) {
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200168 t = tasklet_new();
Willy Tarreau1efafce2019-01-27 15:37:19 +0100169 if (!t) {
170 ha_alert("Out of memory while initializing accept queue for thread %d\n", i);
171 return ERR_FATAL|ERR_ABORT;
172 }
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200173 t->tid = i;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100174 t->process = accept_queue_process;
175 t->context = &accept_queue_rings[i];
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200176 accept_queue_rings[i].tasklet = t;
Willy Tarreau1efafce2019-01-27 15:37:19 +0100177 }
178 return 0;
179}
180
181REGISTER_CONFIG_POSTPARSER("multi-threaded accept queue", accept_queue_init);
182
183#endif // USE_THREAD
184
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200185/* adjust the listener's state and its proxy's listener counters if needed.
186 * It must be called under the listener's lock, but uses atomic ops to change
187 * the proxy's counters so that the proxy lock is not needed.
188 */
Willy Tarreaua37b2442020-09-24 07:23:45 +0200189void listener_set_state(struct listener *l, enum li_state st)
190{
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200191 struct proxy *px = l->bind_conf->frontend;
192
193 if (px) {
194 /* from state */
195 switch (l->state) {
196 case LI_NEW: /* first call */
197 _HA_ATOMIC_ADD(&px->li_all, 1);
198 break;
199 case LI_INIT:
200 case LI_ASSIGNED:
201 break;
202 case LI_PAUSED:
203 _HA_ATOMIC_SUB(&px->li_paused, 1);
204 break;
205 case LI_LISTEN:
206 _HA_ATOMIC_SUB(&px->li_bound, 1);
207 break;
208 case LI_READY:
209 case LI_FULL:
210 case LI_LIMITED:
211 _HA_ATOMIC_SUB(&px->li_ready, 1);
212 break;
213 }
214
215 /* to state */
216 switch (st) {
217 case LI_NEW:
218 case LI_INIT:
219 case LI_ASSIGNED:
220 break;
221 case LI_PAUSED:
Willy Tarreau95a34602020-10-08 15:32:21 +0200222 BUG_ON(l->rx.fd == -1);
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200223 _HA_ATOMIC_ADD(&px->li_paused, 1);
224 break;
225 case LI_LISTEN:
Willy Tarreau95a34602020-10-08 15:32:21 +0200226 BUG_ON(l->rx.fd == -1);
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200227 _HA_ATOMIC_ADD(&px->li_bound, 1);
228 break;
229 case LI_READY:
230 case LI_FULL:
231 case LI_LIMITED:
Willy Tarreau95a34602020-10-08 15:32:21 +0200232 BUG_ON(l->rx.fd == -1);
Willy Tarreauefc0eec2020-09-24 07:27:06 +0200233 _HA_ATOMIC_ADD(&px->li_ready, 1);
234 break;
235 }
236 }
Willy Tarreaua37b2442020-09-24 07:23:45 +0200237 l->state = st;
238}
239
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100240/* This function adds the specified listener's file descriptor to the polling
241 * lists if it is in the LI_LISTEN state. The listener enters LI_READY or
Ilya Shipitsin6fb0f212020-04-02 15:25:26 +0500242 * LI_FULL state depending on its number of connections. In daemon mode, we
Willy Tarreauae302532014-05-07 19:22:24 +0200243 * also support binding only the relevant processes to their respective
244 * listeners. We don't do that in debug mode however.
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100245 */
Willy Tarreau7834a3f2020-09-25 16:40:18 +0200246void enable_listener(struct listener *listener)
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100247{
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100248 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaud6afb532020-10-09 10:35:40 +0200249
250 /* If this listener is supposed to be only in the master, close it in
251 * the workers. Conversely, if it's supposed to be only in the workers
252 * close it in the master.
253 */
Willy Tarreau18c20d22020-10-09 16:11:46 +0200254 if (!!master != !!(listener->rx.flags & RX_F_MWORKER))
Willy Tarreau75c98d12020-10-09 15:55:23 +0200255 do_unbind_listener(listener);
Willy Tarreaud6afb532020-10-09 10:35:40 +0200256
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100257 if (listener->state == LI_LISTEN) {
Willy Tarreau95a34602020-10-08 15:32:21 +0200258 BUG_ON(listener->rx.fd == -1);
William Lallemand095ba4c2017-06-01 17:38:50 +0200259 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau818a92e2020-09-03 07:50:19 +0200260 !(proc_mask(listener->rx.settings->bind_proc) & pid_bit)) {
Willy Tarreauae302532014-05-07 19:22:24 +0200261 /* we don't want to enable this listener and don't
262 * want any fd event to reach it.
263 */
Willy Tarreau75c98d12020-10-09 15:55:23 +0200264 do_unbind_listener(listener);
Willy Tarreauae302532014-05-07 19:22:24 +0200265 }
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100266 else if (!listener->maxconn || listener->nbconn < listener->maxconn) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200267 listener->rx.proto->enable(listener);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200268 listener_set_state(listener, LI_READY);
Willy Tarreauae302532014-05-07 19:22:24 +0200269 }
270 else {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200271 listener_set_state(listener, LI_FULL);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100272 }
273 }
Willy Tarreaud6afb532020-10-09 10:35:40 +0200274
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100275 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreaudabf2e22007-10-28 21:59:24 +0100276}
277
Willy Tarreaucaa7df12020-10-07 15:58:50 +0200278/*
279 * This function completely stops a listener. It will need to operate under the
280 * proxy's lock, the protocol's lock, and the listener's lock. The caller is
281 * responsible for indicating in lpx, lpr, lli whether the respective locks are
282 * already held (non-zero) or not (zero) so that the function picks the missing
283 * ones, in this order. The proxy's listeners count is updated and the proxy is
284 * disabled and woken up after the last one is gone.
285 */
286void stop_listener(struct listener *l, int lpx, int lpr, int lli)
287{
288 struct proxy *px = l->bind_conf->frontend;
Willy Tarreaucaa7df12020-10-07 15:58:50 +0200289
290 if (l->options & LI_O_NOSTOP) {
291 /* master-worker sockpairs are never closed but don't count as a
292 * job.
293 */
294 return;
295 }
296
Willy Tarreaucaa7df12020-10-07 15:58:50 +0200297 if (!lpx)
298 HA_SPIN_LOCK(PROXY_LOCK, &px->lock);
299
300 if (!lpr)
301 HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
302
303 if (!lli)
304 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
305
306 if (l->state > LI_INIT) {
Willy Tarreau75c98d12020-10-09 15:55:23 +0200307 do_unbind_listener(l);
Willy Tarreaucaa7df12020-10-07 15:58:50 +0200308
309 if (l->state >= LI_ASSIGNED)
310 __delete_listener(l);
311
Willy Tarreauacde1522020-10-07 16:31:39 +0200312 proxy_cond_disable(px);
Willy Tarreaucaa7df12020-10-07 15:58:50 +0200313 }
314
315 if (!lli)
316 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
317
318 if (!lpr)
319 HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
320
321 if (!lpx)
322 HA_SPIN_UNLOCK(PROXY_LOCK, &px->lock);
323}
324
Willy Tarreaue03204c2020-10-09 17:02:21 +0200325/* default function called to suspend a listener: it simply passes the call to
326 * the underlying receiver. This is find for most socket-based protocols. This
327 * must be called under the listener's lock. It will return non-zero on success,
328 * 0 on failure. If no receiver-level suspend is provided, the operation is
329 * assumed to succeed.
330 */
331int default_suspend_listener(struct listener *l)
332{
333 int ret = 1;
334
335 if (!l->rx.proto->rx_suspend)
336 return 1;
337
338 ret = l->rx.proto->rx_suspend(&l->rx);
339 return ret > 0 ? ret : 0;
340}
341
342
343/* Tries to resume a suspended listener, and returns non-zero on success or
344 * zero on failure. On certain errors, an alert or a warning might be displayed.
345 * It must be called with the listener's lock held. Depending on the listener's
346 * state and protocol, a listen() call might be used to resume operations, or a
347 * call to the receiver's resume() function might be used as well. This is
348 * suitable as a default function for TCP and UDP. This must be called with the
349 * listener's lock held.
350 */
351int default_resume_listener(struct listener *l)
352{
353 int ret = 1;
354
355 if (l->state == LI_ASSIGNED) {
356 char msg[100];
357 int err;
358
359 err = l->rx.proto->listen(l, msg, sizeof(msg));
360 if (err & ERR_ALERT)
361 ha_alert("Resuming listener: %s\n", msg);
362 else if (err & ERR_WARN)
363 ha_warning("Resuming listener: %s\n", msg);
364
365 if (err & (ERR_FATAL | ERR_ABORT)) {
366 ret = 0;
367 goto end;
368 }
369 }
370
371 if (l->state < LI_PAUSED) {
372 ret = 0;
373 goto end;
374 }
375
376 if (l->state == LI_PAUSED && l->rx.proto->rx_resume &&
377 l->rx.proto->rx_resume(&l->rx) <= 0)
378 ret = 0;
379 end:
380 return ret;
381}
382
383
Willy Tarreaube58c382011-07-24 18:28:10 +0200384/* This function tries to temporarily disable a listener, depending on the OS
385 * capabilities. Linux unbinds the listen socket after a SHUT_RD, and ignores
386 * SHUT_WR. Solaris refuses either shutdown(). OpenBSD ignores SHUT_RD but
387 * closes upon SHUT_WR and refuses to rebind. So a common validation path
388 * involves SHUT_WR && listen && SHUT_RD. In case of success, the FD's polling
389 * is disabled. It normally returns non-zero, unless an error is reported.
390 */
391int pause_listener(struct listener *l)
392{
Willy Tarreau58651b42020-09-24 16:03:29 +0200393 struct proxy *px = l->bind_conf->frontend;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200394 int ret = 1;
395
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100396 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200397
Willy Tarreau02e19752020-09-23 17:17:22 +0200398 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
399 !(proc_mask(l->rx.settings->bind_proc) & pid_bit))
400 goto end;
401
Willy Tarreau9b3a9322020-09-24 14:46:34 +0200402 if (l->state <= LI_PAUSED)
403 goto end;
404
Willy Tarreaue03204c2020-10-09 17:02:21 +0200405 if (l->rx.proto->suspend)
406 ret = l->rx.proto->suspend(l);
Willy Tarreaube58c382011-07-24 18:28:10 +0200407
Olivier Houchard859dc802019-08-08 15:47:21 +0200408 MT_LIST_DEL(&l->wait_queue);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200409
Willy Tarreaua37b2442020-09-24 07:23:45 +0200410 listener_set_state(l, LI_PAUSED);
Willy Tarreau58651b42020-09-24 16:03:29 +0200411
412 if (px && !px->li_ready) {
413 ha_warning("Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
414 send_log(px, LOG_WARNING, "Paused %s %s.\n", proxy_cap_str(px->cap), px->id);
415 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200416 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100417 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200418 return ret;
Willy Tarreaube58c382011-07-24 18:28:10 +0200419}
420
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200421/* This function tries to resume a temporarily disabled listener. Paused, full,
422 * limited and disabled listeners are handled, which means that this function
423 * may replace enable_listener(). The resulting state will either be LI_READY
424 * or LI_FULL. 0 is returned in case of failure to resume (eg: dead socket).
Willy Tarreauae302532014-05-07 19:22:24 +0200425 * Listeners bound to a different process are not woken up unless we're in
Willy Tarreauaf2fd582015-04-14 12:07:16 +0200426 * foreground mode, and are ignored. If the listener was only in the assigned
427 * state, it's totally rebound. This can happen if a pause() has completely
428 * stopped it. If the resume fails, 0 is returned and an error might be
429 * displayed.
Willy Tarreaube58c382011-07-24 18:28:10 +0200430 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100431int resume_listener(struct listener *l)
Willy Tarreaube58c382011-07-24 18:28:10 +0200432{
Willy Tarreau58651b42020-09-24 16:03:29 +0200433 struct proxy *px = l->bind_conf->frontend;
434 int was_paused = px && px->li_paused;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200435 int ret = 1;
436
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100437 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200438
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200439 /* check that another thread didn't to the job in parallel (e.g. at the
440 * end of listen_accept() while we'd come from dequeue_all_listeners().
441 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200442 if (MT_LIST_ADDED(&l->wait_queue))
Willy Tarreauf2cb1692019-07-11 10:08:31 +0200443 goto end;
444
William Lallemand095ba4c2017-06-01 17:38:50 +0200445 if ((global.mode & (MODE_DAEMON | MODE_MWORKER)) &&
Willy Tarreau818a92e2020-09-03 07:50:19 +0200446 !(proc_mask(l->rx.settings->bind_proc) & pid_bit))
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200447 goto end;
Willy Tarreau3569df32017-03-15 12:47:46 +0100448
Willy Tarreau5d7f9ce2020-09-24 18:54:11 +0200449 if (l->state == LI_READY)
450 goto end;
Willy Tarreaube58c382011-07-24 18:28:10 +0200451
Willy Tarreaue03204c2020-10-09 17:02:21 +0200452 if (l->rx.proto->resume)
453 ret = l->rx.proto->resume(l);
Willy Tarreaube58c382011-07-24 18:28:10 +0200454
Willy Tarreaua8cf66b2019-02-27 16:49:00 +0100455 if (l->maxconn && l->nbconn >= l->maxconn) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200456 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200457 listener_set_state(l, LI_FULL);
Willy Tarreau58651b42020-09-24 16:03:29 +0200458 goto done;
Willy Tarreaube58c382011-07-24 18:28:10 +0200459 }
460
Willy Tarreau4b51f422020-09-25 20:32:28 +0200461 l->rx.proto->enable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200462 listener_set_state(l, LI_READY);
Willy Tarreau58651b42020-09-24 16:03:29 +0200463
464 done:
465 if (was_paused && !px->li_paused) {
466 ha_warning("Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
467 send_log(px, LOG_WARNING, "Resumed %s %s.\n", proxy_cap_str(px->cap), px->id);
468 }
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200469 end:
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100470 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200471 return ret;
472}
473
Willy Tarreau87b09662015-04-03 00:22:06 +0200474/* Marks a ready listener as full so that the stream code tries to re-enable
Willy Tarreau62793712011-07-24 19:23:38 +0200475 * it upon next close() using resume_listener().
476 */
Christopher Faulet5580ba22017-08-28 15:29:20 +0200477static void listener_full(struct listener *l)
Willy Tarreau62793712011-07-24 19:23:38 +0200478{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100479 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200480 if (l->state >= LI_READY) {
Olivier Houchard859dc802019-08-08 15:47:21 +0200481 MT_LIST_DEL(&l->wait_queue);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100482 if (l->state != LI_FULL) {
Willy Tarreau4b51f422020-09-25 20:32:28 +0200483 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200484 listener_set_state(l, LI_FULL);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100485 }
Willy Tarreau62793712011-07-24 19:23:38 +0200486 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100487 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreau62793712011-07-24 19:23:38 +0200488}
489
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200490/* Marks a ready listener as limited so that we only try to re-enable it when
491 * resources are free again. It will be queued into the specified queue.
492 */
Olivier Houchard859dc802019-08-08 15:47:21 +0200493static void limit_listener(struct listener *l, struct mt_list *list)
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200494{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100495 HA_SPIN_LOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200496 if (l->state == LI_READY) {
Willy Tarreaude4db172020-07-10 08:10:29 +0200497 MT_LIST_TRY_ADDQ(list, &l->wait_queue);
Willy Tarreau4b51f422020-09-25 20:32:28 +0200498 l->rx.proto->disable(l);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200499 listener_set_state(l, LI_LIMITED);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200500 }
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100501 HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200502}
503
Willy Tarreau241797a2019-12-10 14:10:52 +0100504/* Dequeues all listeners waiting for a resource the global wait queue */
505void dequeue_all_listeners()
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200506{
Willy Tarreau01abd022019-02-28 10:27:18 +0100507 struct listener *listener;
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200508
Willy Tarreau241797a2019-12-10 14:10:52 +0100509 while ((listener = MT_LIST_POP(&global_listener_queue, struct listener *, wait_queue))) {
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200510 /* This cannot fail because the listeners are by definition in
Willy Tarreau01abd022019-02-28 10:27:18 +0100511 * the LI_LIMITED state.
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200512 */
Willy Tarreau01abd022019-02-28 10:27:18 +0100513 resume_listener(listener);
Willy Tarreaue6ca1fc2011-07-24 22:03:52 +0200514 }
515}
516
Willy Tarreau241797a2019-12-10 14:10:52 +0100517/* Dequeues all listeners waiting for a resource in proxy <px>'s queue */
518void dequeue_proxy_listeners(struct proxy *px)
519{
520 struct listener *listener;
521
522 while ((listener = MT_LIST_POP(&px->listener_queue, struct listener *, wait_queue))) {
523 /* This cannot fail because the listeners are by definition in
524 * the LI_LIMITED state.
525 */
526 resume_listener(listener);
527 }
528}
529
Willy Tarreau7b2febd2020-10-09 17:18:29 +0200530
531/* default function used to unbind a listener. This is for use by standard
532 * protocols working on top of accepted sockets. The receiver's rx_unbind()
533 * will automatically be used after the listener is disabled if the socket is
534 * still bound. This must be used under the listener's lock.
Christopher Faulet510c0d62018-03-16 10:04:47 +0100535 */
Willy Tarreau7b2febd2020-10-09 17:18:29 +0200536void default_unbind_listener(struct listener *listener)
Willy Tarreaub648d632007-10-28 22:13:50 +0100537{
Willy Tarreau87acd4e2020-10-08 15:36:46 +0200538 if (listener->state <= LI_ASSIGNED)
539 goto out_close;
540
541 if (listener->rx.fd == -1) {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200542 listener_set_state(listener, LI_ASSIGNED);
Willy Tarreau87acd4e2020-10-08 15:36:46 +0200543 goto out_close;
544 }
545
Willy Tarreauf58b8db2020-10-09 16:32:08 +0200546 if (listener->state >= LI_READY) {
547 listener->rx.proto->disable(listener);
548 if (listener->rx.flags & RX_F_BOUND)
Willy Tarreau87acd4e2020-10-08 15:36:46 +0200549 listener_set_state(listener, LI_LISTEN);
Willy Tarreaub6607bf2020-09-23 16:24:23 +0200550 }
551
Willy Tarreau87acd4e2020-10-08 15:36:46 +0200552 out_close:
Willy Tarreauf58b8db2020-10-09 16:32:08 +0200553 if (listener->rx.flags & RX_F_BOUND)
554 listener->rx.proto->rx_unbind(&listener->rx);
Willy Tarreau7b2febd2020-10-09 17:18:29 +0200555}
556
557/* This function closes the listening socket for the specified listener,
558 * provided that it's already in a listening state. The protocol's unbind()
559 * is called to put the listener into LI_ASSIGNED or LI_LISTEN and handle
560 * the unbinding tasks. The listener enters then the LI_ASSIGNED state if
561 * the receiver is unbound. Must be called with the lock held.
562 */
563void do_unbind_listener(struct listener *listener)
564{
565 MT_LIST_DEL(&listener->wait_queue);
566
567 if (listener->rx.proto->unbind)
568 listener->rx.proto->unbind(listener);
Willy Tarreau374e9af2020-10-09 15:47:17 +0200569
Willy Tarreauf58b8db2020-10-09 16:32:08 +0200570 /* we may have to downgrade the listener if the rx was closed */
571 if (!(listener->rx.flags & RX_F_BOUND) && listener->state > LI_ASSIGNED)
Willy Tarreau374e9af2020-10-09 15:47:17 +0200572 listener_set_state(listener, LI_ASSIGNED);
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100573}
574
Olivier Houchard1fc05162017-04-06 01:05:05 +0200575/* This function closes the listening socket for the specified listener,
576 * provided that it's already in a listening state. The listener enters the
Willy Tarreau75c98d12020-10-09 15:55:23 +0200577 * LI_ASSIGNED state, except if the FD is not closed, in which case it may
578 * remain in LI_LISTEN. This function is intended to be used as a generic
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100579 * function for standard protocols.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200580 */
Willy Tarreaubbd09b92017-11-05 11:38:44 +0100581void unbind_listener(struct listener *listener)
Olivier Houchard1fc05162017-04-06 01:05:05 +0200582{
Christopher Faulet510c0d62018-03-16 10:04:47 +0100583 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau75c98d12020-10-09 15:55:23 +0200584 do_unbind_listener(listener);
Christopher Faulet510c0d62018-03-16 10:04:47 +0100585 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Olivier Houchard1fc05162017-04-06 01:05:05 +0200586}
587
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200588/* creates one or multiple listeners for bind_conf <bc> on sockaddr <ss> on port
589 * range <portl> to <porth>, and possibly attached to fd <fd> (or -1 for auto
Willy Tarreau9b3178d2020-09-16 17:58:55 +0200590 * allocation). The address family is taken from ss->ss_family, and the protocol
Willy Tarreaud2fb99f2020-10-15 21:22:29 +0200591 * passed in <proto> must be usable on this family. The protocol's default iocb
592 * is automatically preset as the receivers' iocb. The number of jobs and
Willy Tarreau9b3178d2020-09-16 17:58:55 +0200593 * listeners is automatically increased by the number of listeners created. It
594 * returns non-zero on success, zero on error with the error message set in <err>.
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200595 */
596int create_listeners(struct bind_conf *bc, const struct sockaddr_storage *ss,
Willy Tarreau9b3178d2020-09-16 17:58:55 +0200597 int portl, int porth, int fd, struct protocol *proto, char **err)
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200598{
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200599 struct listener *l;
600 int port;
601
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200602 for (port = portl; port <= porth; port++) {
603 l = calloc(1, sizeof(*l));
604 if (!l) {
605 memprintf(err, "out of memory");
606 return 0;
607 }
608 l->obj_type = OBJ_TYPE_LISTENER;
609 LIST_ADDQ(&bc->frontend->conf.listeners, &l->by_fe);
610 LIST_ADDQ(&bc->listeners, &l->by_bind);
611 l->bind_conf = bc;
Willy Tarreau0fce6bc2020-09-03 07:46:06 +0200612 l->rx.settings = &bc->settings;
Willy Tarreaueef45422020-09-03 10:05:03 +0200613 l->rx.owner = l;
Willy Tarreaud2fb99f2020-10-15 21:22:29 +0200614 l->rx.iocb = proto->default_iocb;
Willy Tarreau38ba6472020-08-27 08:16:52 +0200615 l->rx.fd = fd;
Willy Tarreau37159062020-08-27 07:48:42 +0200616 memcpy(&l->rx.addr, ss, sizeof(*ss));
Olivier Houchard859dc802019-08-08 15:47:21 +0200617 MT_LIST_INIT(&l->wait_queue);
Willy Tarreaua37b2442020-09-24 07:23:45 +0200618 listener_set_state(l, LI_INIT);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200619
620 proto->add(l, port);
621
Willy Tarreau909c23b2020-09-15 13:50:58 +0200622 if (fd != -1)
Willy Tarreau43046fa2020-09-01 15:41:59 +0200623 l->rx.flags |= RX_F_INHERITED;
William Lallemand75ea0a02017-11-15 19:02:58 +0100624
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100625 HA_SPIN_INIT(&l->lock);
Olivier Houchard64213e92019-03-08 18:52:57 +0100626 _HA_ATOMIC_ADD(&jobs, 1);
627 _HA_ATOMIC_ADD(&listeners, 1);
Willy Tarreau0de59fd2017-09-15 08:10:44 +0200628 }
629 return 1;
630}
631
Willy Tarreau1a64d162007-10-28 22:26:05 +0100632/* Delete a listener from its protocol's list of listeners. The listener's
633 * state is automatically updated from LI_ASSIGNED to LI_INIT. The protocol's
Willy Tarreau2cc5bae2017-09-15 08:18:11 +0200634 * number of listeners is updated, as well as the global number of listeners
635 * and jobs. Note that the listener must have previously been unbound. This
Willy Tarreaub4c083f2020-10-07 15:36:16 +0200636 * is a low-level function expected to be called with the proto_lock and the
637 * listener's lock held.
Willy Tarreau1a64d162007-10-28 22:26:05 +0100638 */
Willy Tarreaub4c083f2020-10-07 15:36:16 +0200639void __delete_listener(struct listener *listener)
Willy Tarreau1a64d162007-10-28 22:26:05 +0100640{
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100641 if (listener->state == LI_ASSIGNED) {
Willy Tarreaua37b2442020-09-24 07:23:45 +0200642 listener_set_state(listener, LI_INIT);
Willy Tarreaub7436612020-08-28 19:51:44 +0200643 LIST_DEL(&listener->rx.proto_list);
Willy Tarreaud7f331c2020-09-25 17:01:43 +0200644 listener->rx.proto->nb_receivers--;
Olivier Houchard64213e92019-03-08 18:52:57 +0100645 _HA_ATOMIC_SUB(&jobs, 1);
646 _HA_ATOMIC_SUB(&listeners, 1);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100647 }
Willy Tarreaub4c083f2020-10-07 15:36:16 +0200648}
649
650/* Delete a listener from its protocol's list of listeners (please check
651 * __delete_listener() above). The proto_lock and the listener's lock will
652 * be grabbed in this order.
653 */
654void delete_listener(struct listener *listener)
655{
656 HA_SPIN_LOCK(PROTO_LOCK, &proto_lock);
657 HA_SPIN_LOCK(LISTENER_LOCK, &listener->lock);
658 __delete_listener(listener);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100659 HA_SPIN_UNLOCK(LISTENER_LOCK, &listener->lock);
Willy Tarreau6ee9f8d2019-08-26 10:55:52 +0200660 HA_SPIN_UNLOCK(PROTO_LOCK, &proto_lock);
Willy Tarreau1a64d162007-10-28 22:26:05 +0100661}
662
Willy Tarreaue2711c72019-02-27 15:39:41 +0100663/* Returns a suitable value for a listener's backlog. It uses the listener's,
664 * otherwise the frontend's backlog, otherwise the listener's maxconn,
665 * otherwise the frontend's maxconn, otherwise 1024.
666 */
667int listener_backlog(const struct listener *l)
668{
669 if (l->backlog)
670 return l->backlog;
671
672 if (l->bind_conf->frontend->backlog)
673 return l->bind_conf->frontend->backlog;
674
675 if (l->maxconn)
676 return l->maxconn;
677
678 if (l->bind_conf->frontend->maxconn)
679 return l->bind_conf->frontend->maxconn;
680
681 return 1024;
682}
683
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200684/* This function is called on a read event from a listening socket, corresponding
685 * to an accept. It tries to accept as many connections as possible, and for each
686 * calls the listener's accept handler (generally the frontend's accept handler).
687 */
Willy Tarreaua74cb382020-10-15 21:29:49 +0200688void listener_accept(struct listener *l)
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200689{
Willy Tarreau83efc322020-10-14 17:37:17 +0200690 struct connection *cli_conn;
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100691 struct proxy *p;
Christopher Faulet102854c2019-04-30 12:17:13 +0200692 unsigned int max_accept;
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100693 int next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100694 int next_feconn = 0;
695 int next_actconn = 0;
Willy Tarreaubb660302014-05-07 19:47:02 +0200696 int expire;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200697 int ret;
698
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100699 p = l->bind_conf->frontend;
Christopher Faulet102854c2019-04-30 12:17:13 +0200700
701 /* if l->maxaccept is -1, then max_accept is UINT_MAX. It is not really
702 * illimited, but it is probably enough.
703 */
Olivier Houchardd16a9df2019-02-25 16:18:16 +0100704 max_accept = l->maxaccept ? l->maxaccept : 1;
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200705
Willy Tarreau93e7c002013-10-07 18:51:07 +0200706 if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
707 int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
Willy Tarreau93e7c002013-10-07 18:51:07 +0200708
709 if (unlikely(!max)) {
710 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200711 expire = tick_add(now_ms, next_event_delay(&global.sess_per_sec, global.sps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100712 goto limit_global;
Willy Tarreau93e7c002013-10-07 18:51:07 +0200713 }
714
715 if (max_accept > max)
716 max_accept = max;
717 }
718
719 if (!(l->options & LI_O_UNLIMITED) && global.cps_lim) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200720 int max = freq_ctr_remain(&global.conn_per_sec, global.cps_lim, 0);
721
722 if (unlikely(!max)) {
723 /* frontend accept rate limit was reached */
Willy Tarreau93e7c002013-10-07 18:51:07 +0200724 expire = tick_add(now_ms, next_event_delay(&global.conn_per_sec, global.cps_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100725 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200726 }
727
728 if (max_accept > max)
729 max_accept = max;
730 }
Willy Tarreaue43d5322013-10-07 20:01:52 +0200731#ifdef USE_OPENSSL
732 if (!(l->options & LI_O_UNLIMITED) && global.ssl_lim && l->bind_conf && l->bind_conf->is_ssl) {
733 int max = freq_ctr_remain(&global.ssl_per_sec, global.ssl_lim, 0);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200734
Willy Tarreaue43d5322013-10-07 20:01:52 +0200735 if (unlikely(!max)) {
736 /* frontend accept rate limit was reached */
Willy Tarreaue43d5322013-10-07 20:01:52 +0200737 expire = tick_add(now_ms, next_event_delay(&global.ssl_per_sec, global.ssl_lim, 0));
Willy Tarreau0591bf72019-12-10 12:01:21 +0100738 goto limit_global;
Willy Tarreaue43d5322013-10-07 20:01:52 +0200739 }
740
741 if (max_accept > max)
742 max_accept = max;
743 }
744#endif
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200745 if (p && p->fe_sps_lim) {
746 int max = freq_ctr_remain(&p->fe_sess_per_sec, p->fe_sps_lim, 0);
747
748 if (unlikely(!max)) {
749 /* frontend accept rate limit was reached */
Willy Tarreau0591bf72019-12-10 12:01:21 +0100750 expire = tick_add(now_ms, next_event_delay(&p->fe_sess_per_sec, p->fe_sps_lim, 0));
751 goto limit_proxy;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200752 }
753
754 if (max_accept > max)
755 max_accept = max;
756 }
757
758 /* Note: if we fail to allocate a connection because of configured
759 * limits, we'll schedule a new attempt worst 1 second later in the
760 * worst case. If we fail due to system limits or temporary resource
761 * shortage, we try again 100ms later in the worst case.
762 */
Christopher Faulet102854c2019-04-30 12:17:13 +0200763 for (; max_accept; next_conn = next_feconn = next_actconn = 0, max_accept--) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +0200764 unsigned int count;
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200765 int status;
Willy Tarreau0aa5a5b2020-10-16 17:43:04 +0200766 __decl_thread(unsigned long mask);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200767
Willy Tarreau82c97892019-02-27 19:32:32 +0100768 /* pre-increase the number of connections without going too far.
769 * We process the listener, then the proxy, then the process.
770 * We know which ones to unroll based on the next_xxx value.
771 */
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100772 do {
773 count = l->nbconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100774 if (unlikely(l->maxconn && count >= l->maxconn)) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100775 /* the listener was marked full or another
776 * thread is going to do it.
777 */
778 next_conn = 0;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100779 listener_full(l);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100780 goto end;
781 }
782 next_conn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000783 } while (!_HA_ATOMIC_CAS(&l->nbconn, (int *)(&count), next_conn));
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100784
Willy Tarreau82c97892019-02-27 19:32:32 +0100785 if (p) {
786 do {
787 count = p->feconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100788 if (unlikely(count >= p->maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100789 /* the frontend was marked full or another
790 * thread is going to do it.
791 */
792 next_feconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100793 expire = TICK_ETERNITY;
794 goto limit_proxy;
Willy Tarreau82c97892019-02-27 19:32:32 +0100795 }
796 next_feconn = count + 1;
Olivier Houchard64213e92019-03-08 18:52:57 +0100797 } while (!_HA_ATOMIC_CAS(&p->feconn, &count, next_feconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200798 }
799
Willy Tarreau82c97892019-02-27 19:32:32 +0100800 if (!(l->options & LI_O_UNLIMITED)) {
801 do {
802 count = actconn;
Willy Tarreau93604ed2019-11-15 10:20:07 +0100803 if (unlikely(count >= global.maxconn)) {
Willy Tarreau82c97892019-02-27 19:32:32 +0100804 /* the process was marked full or another
805 * thread is going to do it.
806 */
807 next_actconn = 0;
Willy Tarreau0591bf72019-12-10 12:01:21 +0100808 expire = tick_add(now_ms, 1000); /* try again in 1 second */
809 goto limit_global;
Willy Tarreau82c97892019-02-27 19:32:32 +0100810 }
811 next_actconn = count + 1;
David Carlier56716622019-03-27 16:08:42 +0000812 } while (!_HA_ATOMIC_CAS(&actconn, (int *)(&count), next_actconn));
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200813 }
814
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200815 cli_conn = l->rx.proto->accept_conn(l, &status);
816 if (!cli_conn) {
817 switch (status) {
818 case CO_AC_DONE:
819 goto end;
Willy Tarreau818dca52014-01-31 19:40:19 +0100820
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200821 case CO_AC_RETRY: /* likely a signal */
Olivier Houchard64213e92019-03-08 18:52:57 +0100822 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100823 if (p)
Olivier Houchard64213e92019-03-08 18:52:57 +0100824 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +0100825 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +0100826 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreaua593ec52014-01-20 21:21:30 +0100827 continue;
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200828
829 case CO_AC_YIELD:
Willy Tarreau92079932019-12-10 09:30:05 +0100830 max_accept = 0;
831 goto end;
William Lallemandd9138002018-11-27 12:02:39 +0100832
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200833 default:
834 goto transient_error;
Willy Tarreau83efc322020-10-14 17:37:17 +0200835 }
836 }
837
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100838 /* The connection was accepted, it must be counted as such */
839 if (l->counters)
840 HA_ATOMIC_UPDATE_MAX(&l->counters->conn_max, next_conn);
841
Willy Tarreau82c97892019-02-27 19:32:32 +0100842 if (p)
843 HA_ATOMIC_UPDATE_MAX(&p->fe_counters.conn_max, next_feconn);
844
845 proxy_inc_fe_conn_ctr(l, p);
846
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100847 if (!(l->options & LI_O_UNLIMITED)) {
848 count = update_freq_ctr(&global.conn_per_sec, 1);
849 HA_ATOMIC_UPDATE_MAX(&global.cps_max, count);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100850 }
851
Willy Tarreau64a9c052019-04-12 15:27:17 +0200852 _HA_ATOMIC_ADD(&activity[tid].accepted, 1);
853
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200854 if (unlikely(cli_conn->handle.fd >= global.maxsock)) {
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200855 send_log(p, LOG_EMERG,
856 "Proxy %s reached the configured maximum connection limit. Please check the global 'maxconn' value.\n",
857 p->id);
Willy Tarreau83efc322020-10-14 17:37:17 +0200858 conn_free(cli_conn);
Willy Tarreau9378bbe2020-10-15 10:09:31 +0200859 close(cli_conn->handle.fd);
Willy Tarreau0591bf72019-12-10 12:01:21 +0100860 expire = tick_add(now_ms, 1000); /* try again in 1 second */
861 goto limit_global;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200862 }
863
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100864 /* past this point, l->accept() will automatically decrement
Willy Tarreau82c97892019-02-27 19:32:32 +0100865 * l->nbconn, feconn and actconn once done. Setting next_*conn=0
866 * allows the error path not to rollback on nbconn. It's more
867 * convenient than duplicating all exit labels.
Willy Tarreau3f0d02b2019-02-25 19:23:37 +0100868 */
869 next_conn = 0;
Willy Tarreau82c97892019-02-27 19:32:32 +0100870 next_feconn = 0;
871 next_actconn = 0;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200872
Willy Tarreau83efc322020-10-14 17:37:17 +0200873
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100874#if defined(USE_THREAD)
Willy Tarreau818a92e2020-09-03 07:50:19 +0200875 mask = thread_mask(l->rx.settings->bind_thread) & all_threads_mask;
Willy Tarreaua7da5e82020-03-12 17:33:29 +0100876 if (atleast2(mask) && (global.tune.options & GTUNE_LISTENER_MQ) && !stopping) {
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100877 struct accept_queue_ring *ring;
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100878 unsigned int t, t0, t1, t2;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100879
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100880 /* The principle is that we have two running indexes,
881 * each visiting in turn all threads bound to this
882 * listener. The connection will be assigned to the one
883 * with the least connections, and the other one will
884 * be updated. This provides a good fairness on short
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100885 * connections (round robin) and on long ones (conn
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100886 * count), without ever missing any idle thread.
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100887 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100888
889 /* keep a copy for the final update. thr_idx is composite
890 * and made of (t2<<16) + t1.
891 */
Willy Tarreau0cf33172019-03-06 15:26:33 +0100892 t0 = l->thr_idx;
Willy Tarreaufc630bd2019-03-04 19:57:34 +0100893 do {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100894 unsigned long m1, m2;
895 int q1, q2;
896
897 t2 = t1 = t0;
898 t2 >>= 16;
899 t1 &= 0xFFFF;
900
901 /* t1 walks low to high bits ;
902 * t2 walks high to low.
903 */
904 m1 = mask >> t1;
905 m2 = mask & (t2 ? nbits(t2 + 1) : ~0UL);
906
Willy Tarreau85d04242019-04-16 18:09:13 +0200907 if (unlikely(!(m1 & 1))) {
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100908 m1 &= ~1UL;
909 if (!m1) {
910 m1 = mask;
911 t1 = 0;
912 }
913 t1 += my_ffsl(m1) - 1;
914 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100915
Willy Tarreau85d04242019-04-16 18:09:13 +0200916 if (unlikely(!(m2 & (1UL << t2)) || t1 == t2)) {
917 /* highest bit not set */
918 if (!m2)
919 m2 = mask;
920
921 t2 = my_flsl(m2) - 1;
922 }
923
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100924 /* now we have two distinct thread IDs belonging to the mask */
925 q1 = accept_queue_rings[t1].tail - accept_queue_rings[t1].head + ACCEPT_QUEUE_SIZE;
926 if (q1 >= ACCEPT_QUEUE_SIZE)
927 q1 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100928
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100929 q2 = accept_queue_rings[t2].tail - accept_queue_rings[t2].head + ACCEPT_QUEUE_SIZE;
930 if (q2 >= ACCEPT_QUEUE_SIZE)
931 q2 -= ACCEPT_QUEUE_SIZE;
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100932
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100933 /* we have 3 possibilities now :
934 * q1 < q2 : t1 is less loaded than t2, so we pick it
935 * and update t2 (since t1 might still be
936 * lower than another thread)
937 * q1 > q2 : t2 is less loaded than t1, so we pick it
938 * and update t1 (since t2 might still be
939 * lower than another thread)
940 * q1 = q2 : both are equally loaded, thus we pick t1
941 * and update t1 as it will become more loaded
942 * than t2.
943 */
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100944
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100945 q1 += l->thr_conn[t1];
946 q2 += l->thr_conn[t2];
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100947
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100948 if (q1 - q2 < 0) {
949 t = t1;
950 t2 = t2 ? t2 - 1 : LONGBITS - 1;
951 }
952 else if (q1 - q2 > 0) {
953 t = t2;
954 t1++;
955 if (t1 >= LONGBITS)
956 t1 = 0;
957 }
958 else {
959 t = t1;
960 t1++;
961 if (t1 >= LONGBITS)
962 t1 = 0;
963 }
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100964
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100965 /* new value for thr_idx */
966 t1 += (t2 << 16);
Olivier Houchard64213e92019-03-08 18:52:57 +0100967 } while (unlikely(!_HA_ATOMIC_CAS(&l->thr_idx, &t0, t1)));
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100968
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100969 /* We successfully selected the best thread "t" for this
970 * connection. We use deferred accepts even if it's the
971 * local thread because tests show that it's the best
972 * performing model, likely due to better cache locality
973 * when processing this loop.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100974 */
Willy Tarreau0fe703b2019-03-05 08:46:28 +0100975 ring = &accept_queue_rings[t];
Willy Tarreau83efc322020-10-14 17:37:17 +0200976 if (accept_queue_push_mp(ring, cli_conn)) {
Olivier Houchard64213e92019-03-08 18:52:57 +0100977 _HA_ATOMIC_ADD(&activity[t].accq_pushed, 1);
Willy Tarreau2bd65a72019-09-24 06:55:18 +0200978 tasklet_wakeup(ring->tasklet);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100979 continue;
980 }
981 /* If the ring is full we do a synchronous accept on
982 * the local thread here.
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100983 */
Olivier Houchard64213e92019-03-08 18:52:57 +0100984 _HA_ATOMIC_ADD(&activity[t].accq_full, 1);
Willy Tarreaue0e9c482019-01-27 15:37:19 +0100985 }
986#endif // USE_THREAD
987
Olivier Houchard64213e92019-03-08 18:52:57 +0100988 _HA_ATOMIC_ADD(&l->thr_conn[tid], 1);
Willy Tarreau83efc322020-10-14 17:37:17 +0200989 ret = l->accept(cli_conn);
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200990 if (unlikely(ret <= 0)) {
Willy Tarreau87b09662015-04-03 00:22:06 +0200991 /* The connection was closed by stream_accept(). Either
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200992 * we just have to ignore it (ret == 0) or it's a critical
993 * error due to a resource shortage, and we must stop the
994 * listener (ret < 0).
995 */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +0200996 if (ret == 0) /* successful termination */
997 continue;
998
Willy Tarreaubb660302014-05-07 19:47:02 +0200999 goto transient_error;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001000 }
1001
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001002 /* increase the per-process number of cumulated sessions, this
1003 * may only be done once l->accept() has accepted the connection.
1004 */
Willy Tarreau93e7c002013-10-07 18:51:07 +02001005 if (!(l->options & LI_O_UNLIMITED)) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001006 count = update_freq_ctr(&global.sess_per_sec, 1);
1007 HA_ATOMIC_UPDATE_MAX(&global.sps_max, count);
Willy Tarreau93e7c002013-10-07 18:51:07 +02001008 }
Willy Tarreaue43d5322013-10-07 20:01:52 +02001009#ifdef USE_OPENSSL
1010 if (!(l->options & LI_O_UNLIMITED) && l->bind_conf && l->bind_conf->is_ssl) {
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001011 count = update_freq_ctr(&global.ssl_per_sec, 1);
1012 HA_ATOMIC_UPDATE_MAX(&global.ssl_max, count);
Willy Tarreaue43d5322013-10-07 20:01:52 +02001013 }
1014#endif
Willy Tarreau93e7c002013-10-07 18:51:07 +02001015
Willy Tarreau8d2c98b2020-05-01 09:51:11 +02001016 ti->flags &= ~TI_FL_STUCK; // this thread is still running
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001017 } /* end of for (max_accept--) */
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001018
Christopher Faulet8d8aa0d2017-05-30 15:36:50 +02001019 end:
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001020 if (next_conn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001021 _HA_ATOMIC_SUB(&l->nbconn, 1);
Willy Tarreau741b4d62019-02-25 15:02:04 +01001022
Willy Tarreau82c97892019-02-27 19:32:32 +01001023 if (p && next_feconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001024 _HA_ATOMIC_SUB(&p->feconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001025
1026 if (next_actconn)
Olivier Houchard64213e92019-03-08 18:52:57 +01001027 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001028
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001029 if ((l->state == LI_FULL && (!l->maxconn || l->nbconn < l->maxconn)) ||
Willy Tarreaucdcba112019-12-11 15:06:30 +01001030 (l->state == LI_LIMITED &&
1031 ((!p || p->feconn < p->maxconn) && (actconn < global.maxconn) &&
1032 (!tick_isset(global_listener_queue_task->expire) ||
1033 tick_is_expired(global_listener_queue_task->expire, now_ms))))) {
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001034 /* at least one thread has to this when quitting */
1035 resume_listener(l);
1036
1037 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001038 dequeue_all_listeners();
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001039
Olivier Houchard859dc802019-08-08 15:47:21 +02001040 if (p && !MT_LIST_ISEMPTY(&p->listener_queue) &&
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001041 (!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 +01001042 dequeue_proxy_listeners(p);
Willy Tarreau3f0d02b2019-02-25 19:23:37 +01001043 }
Willy Tarreau0591bf72019-12-10 12:01:21 +01001044 return;
1045
1046 transient_error:
1047 /* pause the listener for up to 100 ms */
1048 expire = tick_add(now_ms, 100);
1049
Willy Tarreau258b3512020-10-13 17:46:05 +02001050 /* This may be a shared socket that was paused by another process.
1051 * Let's put it to pause in this case.
1052 */
1053 if (l->rx.proto && l->rx.proto->rx_listening(&l->rx) == 0) {
1054 pause_listener(l);
1055 goto end;
1056 }
1057
Willy Tarreau0591bf72019-12-10 12:01:21 +01001058 limit_global:
1059 /* (re-)queue the listener to the global queue and set it to expire no
1060 * later than <expire> ahead. The listener turns to LI_LIMITED.
1061 */
1062 limit_listener(l, &global_listener_queue);
1063 task_schedule(global_listener_queue_task, expire);
1064 goto end;
1065
1066 limit_proxy:
1067 /* (re-)queue the listener to the proxy's queue and set it to expire no
1068 * later than <expire> ahead. The listener turns to LI_LIMITED.
1069 */
1070 limit_listener(l, &p->listener_queue);
Willy Tarreaueeea8082020-01-08 19:15:07 +01001071 if (p->task && tick_isset(expire))
1072 task_schedule(p->task, expire);
Willy Tarreau0591bf72019-12-10 12:01:21 +01001073 goto end;
Willy Tarreaubbebbbf2012-05-07 21:22:09 +02001074}
1075
Willy Tarreau05f50472017-09-15 09:19:58 +02001076/* Notify the listener that a connection initiated from it was released. This
1077 * is used to keep the connection count consistent and to possibly re-open
1078 * listening when it was limited.
1079 */
1080void listener_release(struct listener *l)
1081{
1082 struct proxy *fe = l->bind_conf->frontend;
1083
1084 if (!(l->options & LI_O_UNLIMITED))
Olivier Houchard64213e92019-03-08 18:52:57 +01001085 _HA_ATOMIC_SUB(&actconn, 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001086 if (fe)
Olivier Houchard64213e92019-03-08 18:52:57 +01001087 _HA_ATOMIC_SUB(&fe->feconn, 1);
1088 _HA_ATOMIC_SUB(&l->nbconn, 1);
1089 _HA_ATOMIC_SUB(&l->thr_conn[tid], 1);
Willy Tarreau82c97892019-02-27 19:32:32 +01001090
1091 if (l->state == LI_FULL || l->state == LI_LIMITED)
Willy Tarreau05f50472017-09-15 09:19:58 +02001092 resume_listener(l);
1093
1094 /* Dequeues all of the listeners waiting for a resource */
Willy Tarreau241797a2019-12-10 14:10:52 +01001095 dequeue_all_listeners();
Willy Tarreau05f50472017-09-15 09:19:58 +02001096
Olivier Houchard859dc802019-08-08 15:47:21 +02001097 if (!MT_LIST_ISEMPTY(&fe->listener_queue) &&
Willy Tarreau05f50472017-09-15 09:19:58 +02001098 (!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 +01001099 dequeue_proxy_listeners(fe);
Willy Tarreau05f50472017-09-15 09:19:58 +02001100}
1101
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001102/* Initializes the listener queues. Returns 0 on success, otherwise ERR_* flags */
1103static int listener_queue_init()
1104{
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001105 global_listener_queue_task = task_new(MAX_THREADS_MASK);
1106 if (!global_listener_queue_task) {
1107 ha_alert("Out of memory when initializing global listener queue\n");
1108 return ERR_FATAL|ERR_ABORT;
1109 }
1110 /* very simple initialization, users will queue the task if needed */
1111 global_listener_queue_task->context = NULL; /* not even a context! */
1112 global_listener_queue_task->process = manage_global_listener_queue;
1113
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001114 return 0;
1115}
1116
1117static void listener_queue_deinit()
1118{
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001119 task_destroy(global_listener_queue_task);
1120 global_listener_queue_task = NULL;
Willy Tarreauf2cb1692019-07-11 10:08:31 +02001121}
1122
1123REGISTER_CONFIG_POSTPARSER("multi-threaded listener queue", listener_queue_init);
1124REGISTER_POST_DEINIT(listener_queue_deinit);
1125
Willy Tarreaua1d97f82019-12-10 11:18:41 +01001126
1127/* This is the global management task for listeners. It enables listeners waiting
1128 * for global resources when there are enough free resource, or at least once in
1129 * a while. It is designed to be called as a task.
1130 */
1131static struct task *manage_global_listener_queue(struct task *t, void *context, unsigned short state)
1132{
1133 /* If there are still too many concurrent connections, let's wait for
1134 * some of them to go away. We don't need to re-arm the timer because
1135 * each of them will scan the queue anyway.
1136 */
1137 if (unlikely(actconn >= global.maxconn))
1138 goto out;
1139
1140 /* We should periodically try to enable listeners waiting for a global
1141 * resource here, because it is possible, though very unlikely, that
1142 * they have been blocked by a temporary lack of global resource such
1143 * as a file descriptor or memory and that the temporary condition has
1144 * disappeared.
1145 */
1146 dequeue_all_listeners();
1147
1148 out:
1149 t->expire = TICK_ETERNITY;
1150 task_queue(t);
1151 return t;
1152}
1153
Willy Tarreau26982662012-09-12 23:17:10 +02001154/*
1155 * Registers the bind keyword list <kwl> as a list of valid keywords for next
1156 * parsing sessions.
1157 */
1158void bind_register_keywords(struct bind_kw_list *kwl)
1159{
1160 LIST_ADDQ(&bind_keywords.list, &kwl->list);
1161}
1162
1163/* Return a pointer to the bind keyword <kw>, or NULL if not found. If the
1164 * keyword is found with a NULL ->parse() function, then an attempt is made to
1165 * find one with a valid ->parse() function. This way it is possible to declare
1166 * platform-dependant, known keywords as NULL, then only declare them as valid
1167 * if some options are met. Note that if the requested keyword contains an
1168 * opening parenthesis, everything from this point is ignored.
1169 */
1170struct bind_kw *bind_find_kw(const char *kw)
1171{
1172 int index;
1173 const char *kwend;
1174 struct bind_kw_list *kwl;
1175 struct bind_kw *ret = NULL;
1176
1177 kwend = strchr(kw, '(');
1178 if (!kwend)
1179 kwend = kw + strlen(kw);
1180
1181 list_for_each_entry(kwl, &bind_keywords.list, list) {
1182 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1183 if ((strncmp(kwl->kw[index].kw, kw, kwend - kw) == 0) &&
1184 kwl->kw[index].kw[kwend-kw] == 0) {
1185 if (kwl->kw[index].parse)
1186 return &kwl->kw[index]; /* found it !*/
1187 else
1188 ret = &kwl->kw[index]; /* may be OK */
1189 }
1190 }
1191 }
1192 return ret;
1193}
1194
Willy Tarreau8638f482012-09-18 18:01:17 +02001195/* Dumps all registered "bind" keywords to the <out> string pointer. The
1196 * unsupported keywords are only dumped if their supported form was not
1197 * found.
1198 */
1199void bind_dump_kws(char **out)
1200{
1201 struct bind_kw_list *kwl;
1202 int index;
1203
Christopher Faulet784063e2020-05-18 12:14:18 +02001204 if (!out)
1205 return;
1206
Willy Tarreau8638f482012-09-18 18:01:17 +02001207 *out = NULL;
1208 list_for_each_entry(kwl, &bind_keywords.list, list) {
1209 for (index = 0; kwl->kw[index].kw != NULL; index++) {
1210 if (kwl->kw[index].parse ||
1211 bind_find_kw(kwl->kw[index].kw) == &kwl->kw[index]) {
Willy Tarreau51fb7652012-09-18 18:24:39 +02001212 memprintf(out, "%s[%4s] %s%s%s\n", *out ? *out : "",
1213 kwl->scope,
Willy Tarreau8638f482012-09-18 18:01:17 +02001214 kwl->kw[index].kw,
Willy Tarreau51fb7652012-09-18 18:24:39 +02001215 kwl->kw[index].skip ? " <arg>" : "",
1216 kwl->kw[index].parse ? "" : " (not supported)");
Willy Tarreau8638f482012-09-18 18:01:17 +02001217 }
1218 }
1219 }
1220}
1221
Willy Tarreau645513a2010-05-24 20:55:15 +02001222/************************************************************************/
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001223/* All supported sample and ACL keywords must be declared here. */
Willy Tarreau645513a2010-05-24 20:55:15 +02001224/************************************************************************/
1225
Willy Tarreaua5e37562011-12-16 17:06:15 +01001226/* set temp integer to the number of connexions to the same listening socket */
Willy Tarreau645513a2010-05-24 20:55:15 +02001227static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001228smp_fetch_dconn(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau645513a2010-05-24 20:55:15 +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->nbconn;
Willy Tarreau645513a2010-05-24 20:55:15 +02001232 return 1;
1233}
1234
Willy Tarreaua5e37562011-12-16 17:06:15 +01001235/* set temp integer to the id of the socket (listener) */
Willy Tarreau645513a2010-05-24 20:55:15 +02001236static int
Thierry FOURNIER0786d052015-05-11 15:42:45 +02001237smp_fetch_so_id(const struct arg *args, struct sample *smp, const char *kw, void *private)
Willy Tarreau37406352012-04-23 16:16:37 +02001238{
Thierry FOURNIER8c542ca2015-08-19 09:00:18 +02001239 smp->data.type = SMP_T_SINT;
Thierry FOURNIER136f9d32015-08-19 09:07:19 +02001240 smp->data.u.sint = smp->sess->listener->luid;
Willy Tarreau645513a2010-05-24 20:55:15 +02001241 return 1;
1242}
Jerome Magnineb421b22020-03-27 22:08:40 +01001243static int
1244smp_fetch_so_name(const struct arg *args, struct sample *smp, const char *kw, void *private)
1245{
1246 smp->data.u.str.area = smp->sess->listener->name;
1247 if (!smp->data.u.str.area)
1248 return 0;
1249
1250 smp->data.type = SMP_T_STR;
1251 smp->flags = SMP_F_CONST;
1252 smp->data.u.str.data = strlen(smp->data.u.str.area);
1253 return 1;
1254}
Willy Tarreau645513a2010-05-24 20:55:15 +02001255
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001256/* parse the "accept-proxy" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001257static 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 +02001258{
1259 struct listener *l;
1260
Willy Tarreau4348fad2012-09-20 16:48:07 +02001261 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001262 l->options |= LI_O_ACC_PROXY;
1263
1264 return 0;
1265}
1266
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001267/* parse the "accept-netscaler-cip" bind keyword */
1268static int bind_parse_accept_netscaler_cip(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1269{
1270 struct listener *l;
1271 uint32_t val;
1272
1273 if (!*args[cur_arg + 1]) {
1274 memprintf(err, "'%s' : missing value", args[cur_arg]);
1275 return ERR_ALERT | ERR_FATAL;
1276 }
1277
1278 val = atol(args[cur_arg + 1]);
1279 if (val <= 0) {
Willy Tarreaue2711c72019-02-27 15:39:41 +01001280 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001281 return ERR_ALERT | ERR_FATAL;
1282 }
1283
1284 list_for_each_entry(l, &conf->listeners, by_bind) {
1285 l->options |= LI_O_ACC_CIP;
1286 conf->ns_cip_magic = val;
1287 }
1288
1289 return 0;
1290}
1291
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001292/* parse the "backlog" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001293static 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 +02001294{
1295 struct listener *l;
1296 int val;
1297
1298 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001299 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001300 return ERR_ALERT | ERR_FATAL;
1301 }
1302
1303 val = atol(args[cur_arg + 1]);
Willy Tarreaue2711c72019-02-27 15:39:41 +01001304 if (val < 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001305 memprintf(err, "'%s' : invalid value %d, must be > 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001306 return ERR_ALERT | ERR_FATAL;
1307 }
1308
Willy Tarreau4348fad2012-09-20 16:48:07 +02001309 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001310 l->backlog = val;
1311
1312 return 0;
1313}
1314
1315/* parse the "id" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001316static 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 +02001317{
1318 struct eb32_node *node;
Willy Tarreau4348fad2012-09-20 16:48:07 +02001319 struct listener *l, *new;
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001320 char *error;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001321
Willy Tarreau4348fad2012-09-20 16:48:07 +02001322 if (conf->listeners.n != conf->listeners.p) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001323 memprintf(err, "'%s' can only be used with a single socket", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001324 return ERR_ALERT | ERR_FATAL;
1325 }
1326
1327 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001328 memprintf(err, "'%s' : expects an integer argument", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001329 return ERR_ALERT | ERR_FATAL;
1330 }
1331
Willy Tarreau4348fad2012-09-20 16:48:07 +02001332 new = LIST_NEXT(&conf->listeners, struct listener *, by_bind);
Thierry Fourniere7fe8eb2016-02-26 08:45:58 +01001333 new->luid = strtol(args[cur_arg + 1], &error, 10);
1334 if (*error != '\0') {
1335 memprintf(err, "'%s' : expects an integer argument, found '%s'", args[cur_arg], args[cur_arg + 1]);
1336 return ERR_ALERT | ERR_FATAL;
1337 }
Willy Tarreau4348fad2012-09-20 16:48:07 +02001338 new->conf.id.key = new->luid;
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001339
Willy Tarreau4348fad2012-09-20 16:48:07 +02001340 if (new->luid <= 0) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001341 memprintf(err, "'%s' : custom id has to be > 0", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001342 return ERR_ALERT | ERR_FATAL;
1343 }
1344
Willy Tarreau4348fad2012-09-20 16:48:07 +02001345 node = eb32_lookup(&px->conf.used_listener_id, new->luid);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001346 if (node) {
1347 l = container_of(node, struct listener, conf.id);
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001348 memprintf(err, "'%s' : custom id %d already used at %s:%d ('bind %s')",
1349 args[cur_arg], l->luid, l->bind_conf->file, l->bind_conf->line,
1350 l->bind_conf->arg);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001351 return ERR_ALERT | ERR_FATAL;
1352 }
1353
Willy Tarreau4348fad2012-09-20 16:48:07 +02001354 eb32_insert(&px->conf.used_listener_id, &new->conf.id);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001355 return 0;
1356}
1357
1358/* parse the "maxconn" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001359static 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 +02001360{
1361 struct listener *l;
1362 int val;
1363
1364 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001365 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001366 return ERR_ALERT | ERR_FATAL;
1367 }
1368
1369 val = atol(args[cur_arg + 1]);
Willy Tarreaua8cf66b2019-02-27 16:49:00 +01001370 if (val < 0) {
1371 memprintf(err, "'%s' : invalid value %d, must be >= 0", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001372 return ERR_ALERT | ERR_FATAL;
1373 }
1374
Willy Tarreau4348fad2012-09-20 16:48:07 +02001375 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001376 l->maxconn = val;
1377
1378 return 0;
1379}
1380
1381/* parse the "name" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001382static 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 +02001383{
1384 struct listener *l;
1385
1386 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001387 memprintf(err, "'%s' : missing name", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001388 return ERR_ALERT | ERR_FATAL;
1389 }
1390
Willy Tarreau4348fad2012-09-20 16:48:07 +02001391 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001392 l->name = strdup(args[cur_arg + 1]);
1393
1394 return 0;
1395}
1396
1397/* parse the "nice" bind keyword */
Willy Tarreau4348fad2012-09-20 16:48:07 +02001398static 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 +02001399{
1400 struct listener *l;
1401 int val;
1402
1403 if (!*args[cur_arg + 1]) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001404 memprintf(err, "'%s' : missing value", args[cur_arg]);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001405 return ERR_ALERT | ERR_FATAL;
1406 }
1407
1408 val = atol(args[cur_arg + 1]);
1409 if (val < -1024 || val > 1024) {
Willy Tarreaueb6cead2012-09-20 19:43:14 +02001410 memprintf(err, "'%s' : invalid value %d, allowed range is -1024..1024", args[cur_arg], val);
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001411 return ERR_ALERT | ERR_FATAL;
1412 }
1413
Willy Tarreau4348fad2012-09-20 16:48:07 +02001414 list_for_each_entry(l, &conf->listeners, by_bind)
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001415 l->nice = val;
1416
1417 return 0;
1418}
1419
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001420/* parse the "process" bind keyword */
1421static int bind_parse_process(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1422{
Christopher Fauletc644fa92017-11-23 22:44:11 +01001423 char *slash;
1424 unsigned long proc = 0, thread = 0;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001425
Christopher Fauletc644fa92017-11-23 22:44:11 +01001426 if ((slash = strchr(args[cur_arg + 1], '/')) != NULL)
1427 *slash = 0;
1428
Willy Tarreauff9c9142019-02-07 10:39:36 +01001429 if (parse_process_number(args[cur_arg + 1], &proc, MAX_PROCS, NULL, err)) {
Christopher Fauletf1f0c5f2017-11-22 12:06:43 +01001430 memprintf(err, "'%s' : %s", args[cur_arg], *err);
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001431 return ERR_ALERT | ERR_FATAL;
1432 }
1433
Christopher Fauletc644fa92017-11-23 22:44:11 +01001434 if (slash) {
Willy Tarreauc9a82e42019-01-26 13:25:14 +01001435 if (parse_process_number(slash+1, &thread, MAX_THREADS, NULL, err)) {
Christopher Fauletc644fa92017-11-23 22:44:11 +01001436 memprintf(err, "'%s' : %s", args[cur_arg], *err);
1437 return ERR_ALERT | ERR_FATAL;
1438 }
1439 *slash = '/';
1440 }
1441
Willy Tarreaue26993c2020-09-03 07:18:55 +02001442 conf->settings.bind_proc |= proc;
1443 conf->settings.bind_thread |= thread;
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001444 return 0;
1445}
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001446
Christopher Fauleta717b992018-04-10 14:43:00 +02001447/* parse the "proto" bind keyword */
1448static int bind_parse_proto(char **args, int cur_arg, struct proxy *px, struct bind_conf *conf, char **err)
1449{
1450 struct ist proto;
1451
1452 if (!*args[cur_arg + 1]) {
1453 memprintf(err, "'%s' : missing value", args[cur_arg]);
1454 return ERR_ALERT | ERR_FATAL;
1455 }
1456
1457 proto = ist2(args[cur_arg + 1], strlen(args[cur_arg + 1]));
1458 conf->mux_proto = get_mux_proto(proto);
1459 if (!conf->mux_proto) {
1460 memprintf(err, "'%s' : unknown MUX protocol '%s'", args[cur_arg], args[cur_arg+1]);
1461 return ERR_ALERT | ERR_FATAL;
1462 }
Christopher Fauleta717b992018-04-10 14:43:00 +02001463 return 0;
1464}
1465
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001466/* config parser for global "tune.listener.multi-queue", accepts "on" or "off" */
1467static int cfg_parse_tune_listener_mq(char **args, int section_type, struct proxy *curpx,
1468 struct proxy *defpx, const char *file, int line,
1469 char **err)
1470{
1471 if (too_many_args(1, args, err, NULL))
1472 return -1;
1473
1474 if (strcmp(args[1], "on") == 0)
1475 global.tune.options |= GTUNE_LISTENER_MQ;
1476 else if (strcmp(args[1], "off") == 0)
1477 global.tune.options &= ~GTUNE_LISTENER_MQ;
1478 else {
1479 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
1480 return -1;
1481 }
1482 return 0;
1483}
1484
Willy Tarreau61612d42012-04-19 18:42:05 +02001485/* Note: must not be declared <const> as its list will be overwritten.
1486 * Please take care of keeping this list alphabetically sorted.
1487 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001488static struct sample_fetch_kw_list smp_kws = {ILH, {
Thierry FOURNIER07ee64e2015-07-06 23:43:03 +02001489 { "dst_conn", smp_fetch_dconn, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
1490 { "so_id", smp_fetch_so_id, 0, NULL, SMP_T_SINT, SMP_USE_FTEND, },
Jerome Magnineb421b22020-03-27 22:08:40 +01001491 { "so_name", smp_fetch_so_name, 0, NULL, SMP_T_STR, SMP_USE_FTEND, },
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001492 { /* END */ },
1493}};
1494
Willy Tarreau0108d902018-11-25 19:14:37 +01001495INITCALL1(STG_REGISTER, sample_register_fetches, &smp_kws);
1496
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001497/* Note: must not be declared <const> as its list will be overwritten.
1498 * Please take care of keeping this list alphabetically sorted.
1499 */
Willy Tarreaudc13c112013-06-21 23:16:39 +02001500static struct acl_kw_list acl_kws = {ILH, {
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001501 { /* END */ },
Willy Tarreau645513a2010-05-24 20:55:15 +02001502}};
1503
Willy Tarreau0108d902018-11-25 19:14:37 +01001504INITCALL1(STG_REGISTER, acl_register_keywords, &acl_kws);
1505
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001506/* Note: must not be declared <const> as its list will be overwritten.
1507 * Please take care of keeping this list alphabetically sorted, doing so helps
1508 * all code contributors.
1509 * Optional keywords are also declared with a NULL ->parse() function so that
1510 * the config parser can report an appropriate error when a known keyword was
1511 * not enabled.
1512 */
Willy Tarreau51fb7652012-09-18 18:24:39 +02001513static struct bind_kw_list bind_kws = { "ALL", { }, {
Bertrand Jacquin93b227d2016-06-04 15:11:10 +01001514 { "accept-netscaler-cip", bind_parse_accept_netscaler_cip, 1 }, /* enable NetScaler Client IP insertion protocol */
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001515 { "accept-proxy", bind_parse_accept_proxy, 0 }, /* enable PROXY protocol */
1516 { "backlog", bind_parse_backlog, 1 }, /* set backlog of listening socket */
1517 { "id", bind_parse_id, 1 }, /* set id of listening socket */
1518 { "maxconn", bind_parse_maxconn, 1 }, /* set maxconn of listening socket */
1519 { "name", bind_parse_name, 1 }, /* set name of listening socket */
1520 { "nice", bind_parse_nice, 1 }, /* set nice of listening socket */
Willy Tarreau6ae1ba62014-05-07 19:01:58 +02001521 { "process", bind_parse_process, 1 }, /* set list of allowed process for this socket */
Christopher Fauleta717b992018-04-10 14:43:00 +02001522 { "proto", bind_parse_proto, 1 }, /* set the proto to use for all incoming connections */
Willy Tarreau0ccb7442013-01-07 22:54:17 +01001523 { /* END */ },
Willy Tarreau3dcc3412012-09-18 17:17:28 +02001524}};
1525
Willy Tarreau0108d902018-11-25 19:14:37 +01001526INITCALL1(STG_REGISTER, bind_register_keywords, &bind_kws);
1527
Willy Tarreau7ac908b2019-02-27 12:02:18 +01001528/* config keyword parsers */
1529static struct cfg_kw_list cfg_kws = {ILH, {
1530 { CFG_GLOBAL, "tune.listener.multi-queue", cfg_parse_tune_listener_mq },
1531 { 0, NULL, NULL }
1532}};
1533
1534INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
1535
Willy Tarreau645513a2010-05-24 20:55:15 +02001536/*
1537 * Local variables:
1538 * c-indent-level: 8
1539 * c-basic-offset: 8
1540 * End:
1541 */