blob: fe712a9d768ca4c48289d1f20376ec0a3833b247 [file] [log] [blame]
Willy Tarreaubaaee002006-06-26 02:48:02 +02001/*
2 * File descriptors management functions.
3 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +01004 * Copyright 2000-2014 Willy Tarreau <w@1wt.eu>
Willy Tarreaubaaee002006-06-26 02:48:02 +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 *
Willy Tarreau7be79a42012-11-11 15:02:54 +010011 * There is no direct link between the FD and the updates list. There is only a
12 * bit in the fdtab[] to indicate than a file descriptor is already present in
13 * the updates list. Once an fd is present in the updates list, it will have to
14 * be considered even if its changes are reverted in the middle or if the fd is
15 * replaced.
16 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010017 * The event state for an FD, as found in fdtab[].state, is maintained for each
18 * direction. The state field is built this way, with R bits in the low nibble
19 * and W bits in the high nibble for ease of access and debugging :
Willy Tarreau7be79a42012-11-11 15:02:54 +010020 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010021 * 7 6 5 4 3 2 1 0
Willy Tarreau5bee3e22019-09-04 09:52:57 +020022 * [ 0 | 0 | RW | AW | 0 | 0 | RR | AR ]
Willy Tarreauf817e9f2014-01-10 16:58:45 +010023 *
24 * A* = active *R = read
Willy Tarreau5bee3e22019-09-04 09:52:57 +020025 * R* = ready *W = write
Willy Tarreauf817e9f2014-01-10 16:58:45 +010026 *
27 * An FD is marked "active" when there is a desire to use it.
Willy Tarreauf817e9f2014-01-10 16:58:45 +010028 * An FD is marked "ready" when it has not faced a new EAGAIN since last wake-up
Willy Tarreau5bee3e22019-09-04 09:52:57 +020029 * (it is a cache of the last EAGAIN regardless of polling changes). Each poller
30 * has its own "polled" state for the same fd, as stored in the polled_mask.
Willy Tarreau7be79a42012-11-11 15:02:54 +010031 *
Willy Tarreau5bee3e22019-09-04 09:52:57 +020032 * We have 4 possible states for each direction based on these 2 flags :
Willy Tarreau7be79a42012-11-11 15:02:54 +010033 *
Willy Tarreau5bee3e22019-09-04 09:52:57 +020034 * +---+---+----------+---------------------------------------------+
35 * | R | A | State | Description |
36 * +---+---+----------+---------------------------------------------+
37 * | 0 | 0 | DISABLED | No activity desired, not ready. |
38 * | 0 | 1 | ACTIVE | Activity desired. |
39 * | 1 | 0 | STOPPED | End of activity. |
40 * | 1 | 1 | READY | Activity desired and reported. |
41 * +---+---+----------+---------------------------------------------+
Willy Tarreau7be79a42012-11-11 15:02:54 +010042 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010043 * The transitions are pretty simple :
44 * - fd_want_*() : set flag A
45 * - fd_stop_*() : clear flag A
46 * - fd_cant_*() : clear flag R (when facing EAGAIN)
47 * - fd_may_*() : set flag R (upon return from poll())
Willy Tarreauf817e9f2014-01-10 16:58:45 +010048 *
Willy Tarreau5bee3e22019-09-04 09:52:57 +020049 * Each poller then computes its own polled state :
50 * if (A) { if (!R) P := 1 } else { P := 0 }
Willy Tarreauf817e9f2014-01-10 16:58:45 +010051 *
Willy Tarreau5bee3e22019-09-04 09:52:57 +020052 * The state transitions look like the diagram below.
Willy Tarreauf817e9f2014-01-10 16:58:45 +010053 *
Willy Tarreau5bee3e22019-09-04 09:52:57 +020054 * may +----------+
55 * ,----| DISABLED | (READY=0, ACTIVE=0)
56 * | +----------+
57 * | want | ^
58 * | | |
59 * | v | stop
60 * | +----------+
61 * | | ACTIVE | (READY=0, ACTIVE=1)
62 * | +----------+
63 * | | ^
64 * | may | |
Thayne McCombs8f0cc5c2021-01-07 21:35:52 -070065 * | v | EAGAIN (can't)
Willy Tarreau5bee3e22019-09-04 09:52:57 +020066 * | +--------+
67 * | | READY | (READY=1, ACTIVE=1)
68 * | +--------+
69 * | stop | ^
70 * | | |
71 * | v | want
72 * | +---------+
73 * `--->| STOPPED | (READY=1, ACTIVE=0)
74 * +---------+
Willy Tarreaubaaee002006-06-26 02:48:02 +020075 */
76
Willy Tarreau2ff76222007-04-09 19:29:56 +020077#include <stdio.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020078#include <string.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020079#include <unistd.h>
Olivier Houchard79321b92018-07-26 17:55:11 +020080#include <fcntl.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020081#include <sys/types.h>
Willy Tarreau2d7f81b2019-02-21 22:19:17 +010082#include <sys/resource.h>
Willy Tarreau931d8b72019-08-27 11:08:17 +020083#include <sys/uio.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020084
Willy Tarreaue5733232019-05-22 19:24:06 +020085#if defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +010086#include <poll.h>
87#include <errno.h>
88#endif
89
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020090#include <haproxy/api.h>
Willy Tarreaubc52bec2020-06-18 08:58:47 +020091#include <haproxy/cfgparse.h>
Willy Tarreau0f6ffd62020-06-03 19:33:00 +020092#include <haproxy/fd.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020093#include <haproxy/global.h>
Willy Tarreauc28aab02021-05-08 20:35:03 +020094#include <haproxy/log.h>
Willy Tarreaufc8f6a82020-06-03 19:20:59 +020095#include <haproxy/port_range.h>
Willy Tarreaubc52bec2020-06-18 08:58:47 +020096#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020097
Willy Tarreaub2551052020-06-09 09:07:15 +020098
Willy Tarreaua1090a52021-04-10 16:58:13 +020099struct fdtab *fdtab __read_mostly = NULL; /* array of all the file descriptors */
100struct polled_mask *polled_mask __read_mostly = NULL; /* Array for the polled_mask of each fd */
101struct fdinfo *fdinfo __read_mostly = NULL; /* less-often used infos for file descriptors */
Willy Tarreaubaaee002006-06-26 02:48:02 +0200102int totalconn; /* total # of terminated sessions */
103int actconn; /* # of active sessions */
104
Willy Tarreaua1090a52021-04-10 16:58:13 +0200105struct poller pollers[MAX_POLLERS] __read_mostly;
106struct poller cur_poller __read_mostly;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200107int nbpollers = 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200108
Olivier Houchard6b96f722018-04-25 16:58:25 +0200109volatile struct fdlist update_list; // Global update list
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100110
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200111THREAD_LOCAL int *fd_updt = NULL; // FD updates list
112THREAD_LOCAL int fd_nbupdt = 0; // number of updates in the list
Olivier Houchard79321b92018-07-26 17:55:11 +0200113THREAD_LOCAL int poller_rd_pipe = -1; // Pipe to wake the thread
Willy Tarreaua1090a52021-04-10 16:58:13 +0200114int poller_wr_pipe[MAX_THREADS] __read_mostly; // Pipe to wake the threads
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200115
Olivier Houchard7c49d2e2019-04-16 18:37:05 +0200116volatile int ha_used_fds = 0; // Number of FD we're currently using
Willy Tarreau8c3c0962022-01-27 16:10:48 +0100117static struct fdtab *fdtab_addr; /* address of the allocated area containing fdtab */
Olivier Houchard7c49d2e2019-04-16 18:37:05 +0200118
Willy Tarreau337fb712019-12-20 07:20:00 +0100119#define _GET_NEXT(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->next
120#define _GET_PREV(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->prev
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100121/* adds fd <fd> to fd list <list> if it was not yet in it */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200122void fd_add_to_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100123{
124 int next;
125 int new;
126 int old;
127 int last;
128
129redo_next:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200130 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100131 /* Check that we're not already in the cache, and if not, lock us. */
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100132 if (next > -2)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100133 goto done;
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100134 if (next == -2)
135 goto redo_next;
Olivier Houchardd3608792019-03-08 18:47:42 +0100136 if (!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100137 goto redo_next;
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100138 __ha_barrier_atomic_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100139
140 new = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100141redo_last:
142 /* First, insert in the linked list */
143 last = list->last;
144 old = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100145
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200146 _GET_PREV(fd, off) = -2;
Willy Tarreau11559a72018-02-05 17:52:24 +0100147 /* Make sure the "prev" store is visible before we update the last entry */
148 __ha_barrier_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100149
Willy Tarreau11559a72018-02-05 17:52:24 +0100150 if (unlikely(last == -1)) {
151 /* list is empty, try to add ourselves alone so that list->last=fd */
Olivier Houchardd3608792019-03-08 18:47:42 +0100152 if (unlikely(!_HA_ATOMIC_CAS(&list->last, &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100153 goto redo_last;
154
155 /* list->first was necessary -1, we're guaranteed to be alone here */
156 list->first = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100157 } else {
Willy Tarreau11559a72018-02-05 17:52:24 +0100158 /* adding ourselves past the last element
159 * The CAS will only succeed if its next is -1,
160 * which means it's in the cache, and the last element.
161 */
Olivier Houchardd3608792019-03-08 18:47:42 +0100162 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(last, off), &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100163 goto redo_last;
Willy Tarreau11559a72018-02-05 17:52:24 +0100164
165 /* Then, update the last entry */
166 list->last = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100167 }
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100168 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100169 /* since we're alone at the end of the list and still locked(-2),
Ilya Shipitsinb8888ab2021-01-06 21:20:16 +0500170 * we know no one tried to add past us. Mark the end of list.
Willy Tarreau11559a72018-02-05 17:52:24 +0100171 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200172 _GET_PREV(fd, off) = last;
173 _GET_NEXT(fd, off) = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100174 __ha_barrier_store();
175done:
176 return;
177}
178
179/* removes fd <fd> from fd list <list> */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200180void fd_rm_from_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100181{
182#if defined(HA_HAVE_CAS_DW) || defined(HA_CAS_IS_8B)
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100183 volatile union {
184 struct fdlist_entry ent;
185 uint64_t u64;
186 uint32_t u32[2];
187 } cur_list, next_list;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100188#endif
189 int old;
190 int new = -2;
191 int prev;
192 int next;
193 int last;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100194lock_self:
195#if (defined(HA_CAS_IS_8B) || defined(HA_HAVE_CAS_DW))
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100196 next_list.ent.next = next_list.ent.prev = -2;
197 cur_list.ent = *(volatile struct fdlist_entry *)(((char *)&fdtab[fd]) + off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100198 /* First, attempt to lock our own entries */
199 do {
200 /* The FD is not in the FD cache, give up */
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100201 if (unlikely(cur_list.ent.next <= -3))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100202 return;
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100203 if (unlikely(cur_list.ent.prev == -2 || cur_list.ent.next == -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100204 goto lock_self;
205 } while (
206#ifdef HA_CAS_IS_8B
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100207 unlikely(!_HA_ATOMIC_CAS(((uint64_t *)&_GET_NEXT(fd, off)), (uint64_t *)&cur_list.u64, next_list.u64))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100208#else
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100209 unlikely(!_HA_ATOMIC_DWCAS(((long *)&_GET_NEXT(fd, off)), (uint32_t *)&cur_list.u32, &next_list.u32))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100210#endif
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100211 );
212 next = cur_list.ent.next;
213 prev = cur_list.ent.prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100214
215#else
216lock_self_next:
Willy Tarreau337fb712019-12-20 07:20:00 +0100217 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100218 if (next == -2)
219 goto lock_self_next;
220 if (next <= -3)
221 goto done;
Olivier Houchardd3608792019-03-08 18:47:42 +0100222 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100223 goto lock_self_next;
224lock_self_prev:
Willy Tarreau337fb712019-12-20 07:20:00 +0100225 prev = _GET_PREV(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100226 if (prev == -2)
227 goto lock_self_prev;
Olivier Houchardd3608792019-03-08 18:47:42 +0100228 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(fd, off), &prev, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100229 goto lock_self_prev;
230#endif
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100231 __ha_barrier_atomic_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100232
233 /* Now, lock the entries of our neighbours */
234 if (likely(prev != -1)) {
235redo_prev:
236 old = fd;
237
Olivier Houchardd3608792019-03-08 18:47:42 +0100238 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(prev, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100239 if (unlikely(old == -2)) {
240 /* Neighbour already locked, give up and
241 * retry again once he's done
242 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200243 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100244 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200245 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100246 __ha_barrier_store();
247 goto lock_self;
248 }
249 goto redo_prev;
250 }
251 }
252 if (likely(next != -1)) {
253redo_next:
254 old = fd;
Olivier Houchardd3608792019-03-08 18:47:42 +0100255 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(next, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100256 if (unlikely(old == -2)) {
257 /* Neighbour already locked, give up and
258 * retry again once he's done
259 */
260 if (prev != -1) {
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200261 _GET_NEXT(prev, off) = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100262 __ha_barrier_store();
263 }
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200264 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100265 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200266 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100267 __ha_barrier_store();
268 goto lock_self;
269 }
270 goto redo_next;
271 }
272 }
273 if (list->first == fd)
274 list->first = next;
275 __ha_barrier_store();
276 last = list->last;
Olivier Houchardd3608792019-03-08 18:47:42 +0100277 while (unlikely(last == fd && (!_HA_ATOMIC_CAS(&list->last, &last, prev))))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100278 __ha_compiler_barrier();
279 /* Make sure we let other threads know we're no longer in cache,
280 * before releasing our neighbours.
281 */
282 __ha_barrier_store();
283 if (likely(prev != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200284 _GET_NEXT(prev, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100285 __ha_barrier_store();
286 if (likely(next != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200287 _GET_PREV(next, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100288 __ha_barrier_store();
289 /* Ok, now we're out of the fd cache */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200290 _GET_NEXT(fd, off) = -(next + 4);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100291 __ha_barrier_store();
292done:
293 return;
294}
295
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200296#undef _GET_NEXT
297#undef _GET_PREV
298
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100299/* deletes the FD once nobody uses it anymore, as detected by the caller by its
300 * thread_mask being zero and its running mask turning to zero. There is no
301 * protection against concurrent accesses, it's up to the caller to make sure
302 * only the last thread will call it. This is only for internal use, please use
303 * fd_delete() instead.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200304 */
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100305void _fd_delete_orphan(int fd)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200306{
Willy Tarreaub41a6e92021-04-06 17:49:19 +0200307 if (fdtab[fd].state & FD_LINGER_RISK) {
Willy Tarreauad38ace2013-12-15 14:19:38 +0100308 /* this is generally set when connecting to servers */
Ilya Shipitsinb7e43f02020-04-02 15:02:08 +0500309 DISGUISE(setsockopt(fd, SOL_SOCKET, SO_LINGER,
310 (struct linger *) &nolinger, sizeof(struct linger)));
Willy Tarreauad38ace2013-12-15 14:19:38 +0100311 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100312 if (cur_poller.clo)
313 cur_poller.clo(fd);
Willy Tarreau2d423292021-03-24 15:34:25 +0100314
315 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
Olivier Houchardc22580c2019-08-05 18:51:52 +0200316 polled_mask[fd].poll_recv = polled_mask[fd].poll_send = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100317
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100318 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100319
Willy Tarreau38e8a1c2020-06-23 10:04:54 +0200320#ifdef DEBUG_FD
321 fdtab[fd].event_count = 0;
322#endif
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200323 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200324 fdtab[fd].owner = NULL;
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100325 /* perform the close() call last as it's what unlocks the instant reuse
326 * of this FD by any other thread.
327 */
Willy Tarreau63d8b602020-08-26 11:54:06 +0200328 close(fd);
Willy Tarreau4781b152021-04-06 13:53:36 +0200329 _HA_ATOMIC_DEC(&ha_used_fds);
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100330}
331
Olivier Houchardc23b3372021-03-25 01:38:54 +0100332#ifndef HA_HAVE_CAS_DW
333__decl_thread(__decl_rwlock(fd_mig_lock));
334#endif
335
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100336/* Deletes an FD from the fdsets. The file descriptor is also closed, possibly
337 * asynchronously. Only the owning thread may do this.
338 */
339void fd_delete(int fd)
340{
341 /* we must postpone removal of an FD that may currently be in use
Ilya Shipitsinb2be9a12021-04-24 13:25:42 +0500342 * by another thread. This can happen in the following two situations:
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100343 * - after a takeover, the owning thread closes the connection but
344 * the previous one just woke up from the poller and entered
345 * the FD handler iocb. That thread holds an entry in running_mask
346 * and requires removal protection.
347 * - multiple threads are accepting connections on a listener, and
348 * one of them (or even an separate one) decides to unbind the
349 * listener under the listener's lock while other ones still hold
350 * the running bit.
351 * In both situations the FD is marked as unused (thread_mask = 0) and
352 * will not take new bits in its running_mask so we have the guarantee
353 * that the last thread eliminating running_mask is the one allowed to
354 * safely delete the FD. Most of the time it will be the current thread.
355 */
356
357 HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
Olivier Houchardc23b3372021-03-25 01:38:54 +0100358#ifndef HA_HAVE_CAS_DW
359 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
360#endif
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100361 HA_ATOMIC_STORE(&fdtab[fd].thread_mask, 0);
Olivier Houchardc23b3372021-03-25 01:38:54 +0100362#ifndef HA_HAVE_CAS_DW
363 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
364#endif
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100365 if (fd_clr_running(fd) == 0)
366 _fd_delete_orphan(fd);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200367}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200368
Olivier Houchard88516642020-03-05 18:10:51 +0100369/*
370 * Take over a FD belonging to another thread.
371 * unexpected_conn is the expected owner of the fd.
372 * Returns 0 on success, and -1 on failure.
373 */
374int fd_takeover(int fd, void *expected_owner)
375{
Willy Tarreauc460c912020-06-18 07:28:09 +0200376 int ret = -1;
Olivier Houchard88516642020-03-05 18:10:51 +0100377
Willy Tarreauf1cad382020-06-18 08:14:59 +0200378#ifndef HA_HAVE_CAS_DW
Willy Tarreau1db42732021-04-06 11:44:07 +0200379 if (_HA_ATOMIC_OR_FETCH(&fdtab[fd].running_mask, tid_bit) == tid_bit) {
Willy Tarreauc460c912020-06-18 07:28:09 +0200380 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
381 if (fdtab[fd].owner == expected_owner) {
382 fdtab[fd].thread_mask = tid_bit;
383 ret = 0;
384 }
385 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
Olivier Houchard88516642020-03-05 18:10:51 +0100386 }
Olivier Houchard88516642020-03-05 18:10:51 +0100387#else
388 unsigned long old_masks[2];
389 unsigned long new_masks[2];
390
Olivier Houchard88516642020-03-05 18:10:51 +0100391 new_masks[0] = new_masks[1] = tid_bit;
Willy Tarreau42973632020-06-18 08:05:15 +0200392
Willy Tarreau1db42732021-04-06 11:44:07 +0200393 old_masks[0] = _HA_ATOMIC_OR_FETCH(&fdtab[fd].running_mask, tid_bit);
Willy Tarreau42973632020-06-18 08:05:15 +0200394 old_masks[1] = fdtab[fd].thread_mask;
395
Olivier Houchard88516642020-03-05 18:10:51 +0100396 /* protect ourself against a delete then an insert for the same fd,
397 * if it happens, then the owner will no longer be the expected
398 * connection.
399 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200400 if (fdtab[fd].owner == expected_owner) {
401 while (old_masks[0] == tid_bit && old_masks[1]) {
402 if (_HA_ATOMIC_DWCAS(&fdtab[fd].running_mask, &old_masks, &new_masks)) {
403 ret = 0;
404 break;
405 }
Olivier Houchard88516642020-03-05 18:10:51 +0100406 }
Willy Tarreauf1cad382020-06-18 08:14:59 +0200407 }
408#endif /* HW_HAVE_CAS_DW */
409
Olivier Houchard88516642020-03-05 18:10:51 +0100410 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
Willy Tarreauf1cad382020-06-18 08:14:59 +0200411
Olivier Houchardddc874c2020-06-17 20:34:05 +0200412 /* Make sure the FD doesn't have the active bit. It is possible that
413 * the fd is polled by the thread that used to own it, the new thread
414 * is supposed to call subscribe() later, to activate polling.
415 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200416 if (likely(ret == 0))
417 fd_stop_recv(fd);
418 return ret;
Olivier Houchard88516642020-03-05 18:10:51 +0100419}
420
Willy Tarreaudbe30602019-09-04 13:25:41 +0200421void updt_fd_polling(const int fd)
422{
Willy Tarreau5a7d6eb2020-11-26 22:25:10 +0100423 if (all_threads_mask == 1UL || (fdtab[fd].thread_mask & all_threads_mask) == tid_bit) {
Willy Tarreaudbe30602019-09-04 13:25:41 +0200424 if (HA_ATOMIC_BTS(&fdtab[fd].update_mask, tid))
425 return;
426
427 fd_updt[fd_nbupdt++] = fd;
428 } else {
429 unsigned long update_mask = fdtab[fd].update_mask;
430 do {
431 if (update_mask == fdtab[fd].thread_mask)
432 return;
Willy Tarreauf0158872020-09-25 12:18:53 +0200433 } while (!_HA_ATOMIC_CAS(&fdtab[fd].update_mask, &update_mask, fdtab[fd].thread_mask));
434
Willy Tarreaudbe30602019-09-04 13:25:41 +0200435 fd_add_to_fd_list(&update_list, fd, offsetof(struct fdtab, update));
Willy Tarreauf0158872020-09-25 12:18:53 +0200436
437 if (fd_active(fd) &&
438 !(fdtab[fd].thread_mask & tid_bit) &&
439 (fdtab[fd].thread_mask & ~tid_bit & all_threads_mask & ~sleeping_thread_mask) == 0) {
440 /* we need to wake up one thread to handle it immediately */
441 int thr = my_ffsl(fdtab[fd].thread_mask & ~tid_bit & all_threads_mask) - 1;
442
443 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
444 wake_thread(thr);
445 }
Willy Tarreaudbe30602019-09-04 13:25:41 +0200446 }
447}
448
Willy Tarreau931d8b72019-08-27 11:08:17 +0200449/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
450 * optionally followed by a newline if <nl> is non-null, to file descriptor
451 * <fd>. The message is sent atomically using writev(). It may be truncated to
452 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
453 * two lists, it's just a convenience to help the caller prepend some prefixes
454 * when necessary. It takes the fd's lock to make sure no other thread will
455 * write to the same fd in parallel. Returns the number of bytes sent, or <=0
456 * on failure. A limit to 31 total non-empty segments is enforced. The caller
457 * is responsible for taking care of making the fd non-blocking.
458 */
459ssize_t fd_write_frag_line(int fd, size_t maxlen, const struct ist pfx[], size_t npfx, const struct ist msg[], size_t nmsg, int nl)
460{
461 struct iovec iovec[32];
Willy Tarreau931d8b72019-08-27 11:08:17 +0200462 size_t sent = 0;
463 int vec = 0;
Willy Tarreaudf187872020-06-11 14:25:47 +0200464 int attempts = 0;
Willy Tarreau931d8b72019-08-27 11:08:17 +0200465
466 if (!maxlen)
467 maxlen = ~0;
468
469 /* keep one char for a possible trailing '\n' in any case */
470 maxlen--;
471
472 /* make an iovec from the concatenation of all parts of the original
473 * message. Skip empty fields and truncate the whole message to maxlen,
474 * leaving one spare iovec for the '\n'.
475 */
476 while (vec < (sizeof(iovec) / sizeof(iovec[0]) - 1)) {
477 if (!npfx) {
478 pfx = msg;
479 npfx = nmsg;
480 nmsg = 0;
481 if (!npfx)
482 break;
483 }
484
485 iovec[vec].iov_base = pfx->ptr;
486 iovec[vec].iov_len = MIN(maxlen, pfx->len);
487 maxlen -= iovec[vec].iov_len;
Willy Tarreau931d8b72019-08-27 11:08:17 +0200488 if (iovec[vec].iov_len)
489 vec++;
490 pfx++; npfx--;
491 };
492
493 if (nl) {
494 iovec[vec].iov_base = "\n";
495 iovec[vec].iov_len = 1;
496 vec++;
497 }
498
Willy Tarreaudf187872020-06-11 14:25:47 +0200499 /* make sure we never interleave writes and we never block. This means
500 * we prefer to fail on collision than to block. But we don't want to
501 * lose too many logs so we just perform a few lock attempts then give
502 * up.
503 */
504
Willy Tarreau1673c4a2021-04-07 17:36:57 +0200505 while (HA_ATOMIC_BTS(&fdtab[fd].state, FD_EXCL_SYSCALL_BIT)) {
Willy Tarreaudf187872020-06-11 14:25:47 +0200506 if (++attempts >= 200) {
507 /* so that the caller knows the message couldn't be delivered */
508 sent = -1;
509 errno = EAGAIN;
510 goto leave;
511 }
512 ha_thread_relax();
513 }
514
Willy Tarreau0cc61282021-04-06 17:57:12 +0200515 if (unlikely(!(fdtab[fd].state & FD_INITIALIZED))) {
516 HA_ATOMIC_OR(&fdtab[fd].state, FD_INITIALIZED);
Willy Tarreau7e9776a2019-08-30 14:41:47 +0200517 if (!isatty(fd))
518 fcntl(fd, F_SETFL, O_NONBLOCK);
519 }
Willy Tarreau931d8b72019-08-27 11:08:17 +0200520 sent = writev(fd, iovec, vec);
Willy Tarreau1673c4a2021-04-07 17:36:57 +0200521 HA_ATOMIC_BTR(&fdtab[fd].state, FD_EXCL_SYSCALL_BIT);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200522
Willy Tarreaudf187872020-06-11 14:25:47 +0200523 leave:
Willy Tarreau931d8b72019-08-27 11:08:17 +0200524 /* sent > 0 if the message was delivered */
525 return sent;
526}
527
Olivier Houchard2292edf2019-02-25 14:26:54 +0100528#if defined(USE_CLOSEFROM)
529void my_closefrom(int start)
530{
531 closefrom(start);
532}
533
Willy Tarreaue5733232019-05-22 19:24:06 +0200534#elif defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100535/* This is a portable implementation of closefrom(). It closes all open file
536 * descriptors starting at <start> and above. It relies on the fact that poll()
537 * will return POLLNVAL for each invalid (hence close) file descriptor passed
538 * in argument in order to skip them. It acts with batches of FDs and will
539 * typically perform one poll() call per 1024 FDs so the overhead is low in
540 * case all FDs have to be closed.
541 */
542void my_closefrom(int start)
543{
544 struct pollfd poll_events[1024];
545 struct rlimit limit;
546 int nbfds, fd, ret, idx;
547 int step, next;
548
549 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
550 step = nbfds = limit.rlim_cur;
551 else
552 step = nbfds = 0;
553
554 if (nbfds <= 0) {
555 /* set safe limit */
556 nbfds = 1024;
557 step = 256;
558 }
559
560 if (step > sizeof(poll_events) / sizeof(poll_events[0]))
561 step = sizeof(poll_events) / sizeof(poll_events[0]);
562
563 while (start < nbfds) {
564 next = (start / step + 1) * step;
565
566 for (fd = start; fd < next && fd < nbfds; fd++) {
567 poll_events[fd - start].fd = fd;
568 poll_events[fd - start].events = 0;
569 }
570
571 do {
572 ret = poll(poll_events, fd - start, 0);
573 if (ret >= 0)
574 break;
575 } while (errno == EAGAIN || errno == EINTR || errno == ENOMEM);
576
Willy Tarreaub8e602c2019-02-22 09:07:42 +0100577 if (ret)
578 ret = fd - start;
579
Willy Tarreau9188ac62019-02-21 22:12:47 +0100580 for (idx = 0; idx < ret; idx++) {
581 if (poll_events[idx].revents & POLLNVAL)
582 continue; /* already closed */
583
584 fd = poll_events[idx].fd;
585 close(fd);
586 }
587 start = next;
588 }
589}
590
Willy Tarreaue5733232019-05-22 19:24:06 +0200591#else // defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100592
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100593/* This is a portable implementation of closefrom(). It closes all open file
594 * descriptors starting at <start> and above. This is a naive version for use
595 * when the operating system provides no alternative.
596 */
597void my_closefrom(int start)
598{
599 struct rlimit limit;
600 int nbfds;
601
602 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
603 nbfds = limit.rlim_cur;
604 else
605 nbfds = 0;
606
607 if (nbfds <= 0)
608 nbfds = 1024; /* safe limit */
609
610 while (start < nbfds)
611 close(start++);
612}
Willy Tarreaue5733232019-05-22 19:24:06 +0200613#endif // defined(USE_POLL)
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100614
Willy Tarreau4f60f162007-04-08 16:39:58 +0200615/* disable the specified poller */
616void disable_poller(const char *poller_name)
617{
618 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200619
Willy Tarreau4f60f162007-04-08 16:39:58 +0200620 for (p = 0; p < nbpollers; p++)
621 if (strcmp(pollers[p].name, poller_name) == 0)
622 pollers[p].pref = 0;
623}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200624
Olivier Houchard79321b92018-07-26 17:55:11 +0200625void poller_pipe_io_handler(int fd)
626{
627 char buf[1024];
628 /* Flush the pipe */
629 while (read(fd, buf, sizeof(buf)) > 0);
630 fd_cant_recv(fd);
631}
632
Willy Tarreau082b6282019-05-22 14:42:12 +0200633/* allocate the per-thread fd_updt thus needs to be called early after
634 * thread creation.
635 */
636static int alloc_pollers_per_thread()
637{
638 fd_updt = calloc(global.maxsock, sizeof(*fd_updt));
639 return fd_updt != NULL;
640}
641
642/* Initialize the pollers per thread.*/
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200643static int init_pollers_per_thread()
644{
Olivier Houchard79321b92018-07-26 17:55:11 +0200645 int mypipe[2];
Willy Tarreau082b6282019-05-22 14:42:12 +0200646
647 if (pipe(mypipe) < 0)
Olivier Houchard79321b92018-07-26 17:55:11 +0200648 return 0;
Willy Tarreau082b6282019-05-22 14:42:12 +0200649
Olivier Houchard79321b92018-07-26 17:55:11 +0200650 poller_rd_pipe = mypipe[0];
651 poller_wr_pipe[tid] = mypipe[1];
652 fcntl(poller_rd_pipe, F_SETFL, O_NONBLOCK);
653 fd_insert(poller_rd_pipe, poller_pipe_io_handler, poller_pipe_io_handler,
654 tid_bit);
655 fd_want_recv(poller_rd_pipe);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200656 return 1;
657}
658
659/* Deinitialize the pollers per thread */
660static void deinit_pollers_per_thread()
661{
William Lallemand808e1b72018-12-15 22:34:31 +0100662 /* rd and wr are init at the same place, but only rd is init to -1, so
663 we rely to rd to close. */
664 if (poller_rd_pipe > -1) {
665 close(poller_rd_pipe);
666 poller_rd_pipe = -1;
667 close(poller_wr_pipe[tid]);
668 poller_wr_pipe[tid] = -1;
669 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200670}
671
Willy Tarreau082b6282019-05-22 14:42:12 +0200672/* Release the pollers per thread, to be called late */
673static void free_pollers_per_thread()
674{
Willy Tarreau61cfdf42021-02-20 10:46:51 +0100675 ha_free(&fd_updt);
Willy Tarreau082b6282019-05-22 14:42:12 +0200676}
677
Willy Tarreaubaaee002006-06-26 02:48:02 +0200678/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200679 * Initialize the pollers till the best one is found.
680 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200681 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200682int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200683{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200684 int p;
685 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200686
Willy Tarreau8c3c0962022-01-27 16:10:48 +0100687 if ((fdtab_addr = calloc(global.maxsock, sizeof(*fdtab) + 64)) == NULL) {
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200688 ha_alert("Not enough memory to allocate %d entries for fdtab!\n", global.maxsock);
Christopher Faulet63fe6522017-08-31 17:52:09 +0200689 goto fail_tab;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200690 }
Christopher Faulet63fe6522017-08-31 17:52:09 +0200691
Willy Tarreau8c3c0962022-01-27 16:10:48 +0100692 /* always provide an aligned fdtab */
693 fdtab = (struct fdtab*)((((size_t)fdtab_addr) + 63) & -(size_t)64);
694
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200695 if ((polled_mask = calloc(global.maxsock, sizeof(*polled_mask))) == NULL) {
696 ha_alert("Not enough memory to allocate %d entries for polled_mask!\n", global.maxsock);
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200697 goto fail_polledmask;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200698 }
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000699
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200700 if ((fdinfo = calloc(global.maxsock, sizeof(*fdinfo))) == NULL) {
701 ha_alert("Not enough memory to allocate %d entries for fdinfo!\n", global.maxsock);
Christopher Faulet63fe6522017-08-31 17:52:09 +0200702 goto fail_info;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200703 }
Christopher Faulet63fe6522017-08-31 17:52:09 +0200704
Olivier Houchard6b96f722018-04-25 16:58:25 +0200705 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200706
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100707 for (p = 0; p < global.maxsock; p++) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100708 /* Mark the fd as out of the fd cache */
Olivier Houchard6b96f722018-04-25 16:58:25 +0200709 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100710 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200711
Willy Tarreau4f60f162007-04-08 16:39:58 +0200712 do {
713 bp = NULL;
714 for (p = 0; p < nbpollers; p++)
715 if (!bp || (pollers[p].pref > bp->pref))
716 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200717
Willy Tarreau4f60f162007-04-08 16:39:58 +0200718 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200719 break;
720
Willy Tarreau4f60f162007-04-08 16:39:58 +0200721 if (bp->init(bp)) {
722 memcpy(&cur_poller, bp, sizeof(*bp));
723 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200724 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200725 } while (!bp || bp->pref == 0);
Willy Tarreau7be79a42012-11-11 15:02:54 +0100726
Christopher Faulet63fe6522017-08-31 17:52:09 +0200727 free(fdinfo);
728 fail_info:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200729 free(polled_mask);
730 fail_polledmask:
Willy Tarreau8c3c0962022-01-27 16:10:48 +0100731 free(fdtab_addr);
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000732 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100733 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200734}
735
Willy Tarreaubaaee002006-06-26 02:48:02 +0200736/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200737 * Deinitialize the pollers.
738 */
739void deinit_pollers() {
740
741 struct poller *bp;
742 int p;
743
744 for (p = 0; p < nbpollers; p++) {
745 bp = &pollers[p];
746
747 if (bp && bp->pref)
748 bp->term(bp);
749 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200750
Willy Tarreau61cfdf42021-02-20 10:46:51 +0100751 ha_free(&fdinfo);
Willy Tarreau8c3c0962022-01-27 16:10:48 +0100752 ha_free(&fdtab_addr);
Willy Tarreau61cfdf42021-02-20 10:46:51 +0100753 ha_free(&polled_mask);
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200754}
755
756/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200757 * Lists the known pollers on <out>.
758 * Should be performed only before initialization.
759 */
760int list_pollers(FILE *out)
761{
762 int p;
763 int last, next;
764 int usable;
765 struct poller *bp;
766
767 fprintf(out, "Available polling systems :\n");
768
769 usable = 0;
770 bp = NULL;
771 last = next = -1;
772 while (1) {
773 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200774 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100775 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200776 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100777 if (!bp || (pollers[p].pref > bp->pref))
778 bp = &pollers[p];
779 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200780 }
781
782 if (next == -1)
783 break;
784
785 for (p = 0; p < nbpollers; p++) {
786 if (pollers[p].pref == next) {
787 fprintf(out, " %10s : ", pollers[p].name);
788 if (pollers[p].pref == 0)
789 fprintf(out, "disabled, ");
790 else
791 fprintf(out, "pref=%3d, ", pollers[p].pref);
792 if (pollers[p].test(&pollers[p])) {
793 fprintf(out, " test result OK");
794 if (next > 0)
795 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100796 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200797 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100798 if (bp == &pollers[p])
799 bp = NULL;
800 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200801 fprintf(out, "\n");
802 }
803 }
804 last = next;
805 next = -1;
806 };
807 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
808 return 0;
809}
810
811/*
812 * Some pollers may lose their connection after a fork(). It may be necessary
813 * to create initialize part of them again. Returns 0 in case of failure,
814 * otherwise 1. The fork() function may be NULL if unused. In case of error,
815 * the the current poller is destroyed and the caller is responsible for trying
816 * another one by calling init_pollers() again.
817 */
818int fork_poller()
819{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200820 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100821 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200822 if (fdtab[fd].owner) {
Willy Tarreau030dae12021-04-06 17:53:33 +0200823 HA_ATOMIC_OR(&fdtab[fd].state, FD_CLONED);
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200824 }
825 }
826
Willy Tarreau2ff76222007-04-09 19:29:56 +0200827 if (cur_poller.fork) {
828 if (cur_poller.fork(&cur_poller))
829 return 1;
830 cur_poller.term(&cur_poller);
831 return 0;
832 }
833 return 1;
834}
835
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200836/* config parser for global "tune.fd.edge-triggered", accepts "on" or "off" */
837static int cfg_parse_tune_fd_edge_triggered(char **args, int section_type, struct proxy *curpx,
Willy Tarreau01825162021-03-09 09:53:46 +0100838 const struct proxy *defpx, const char *file, int line,
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200839 char **err)
840{
841 if (too_many_args(1, args, err, NULL))
842 return -1;
843
844 if (strcmp(args[1], "on") == 0)
845 global.tune.options |= GTUNE_FD_ET;
846 else if (strcmp(args[1], "off") == 0)
847 global.tune.options &= ~GTUNE_FD_ET;
848 else {
849 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
850 return -1;
851 }
852 return 0;
853}
854
855/* config keyword parsers */
856static struct cfg_kw_list cfg_kws = {ILH, {
Willy Tarreau9eec7e22021-05-08 11:06:32 +0200857 { CFG_GLOBAL, "tune.fd.edge-triggered", cfg_parse_tune_fd_edge_triggered, KWF_EXPERIMENTAL },
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200858 { 0, NULL, NULL }
859}};
860
861INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
862
Willy Tarreau082b6282019-05-22 14:42:12 +0200863REGISTER_PER_THREAD_ALLOC(alloc_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100864REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
865REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
Willy Tarreau082b6282019-05-22 14:42:12 +0200866REGISTER_PER_THREAD_FREE(free_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100867
Willy Tarreau2ff76222007-04-09 19:29:56 +0200868/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200869 * Local variables:
870 * c-indent-level: 8
871 * c-basic-offset: 8
872 * End:
873 */