blob: 3387653a582c60e16b902d1940c7340d06ebdddc [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 Tarreaufc8f6a82020-06-03 19:20:59 +020094#include <haproxy/port_range.h>
Willy Tarreaubc52bec2020-06-18 08:58:47 +020095#include <haproxy/tools.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020096
Willy Tarreaub2551052020-06-09 09:07:15 +020097
Willy Tarreaua1090a52021-04-10 16:58:13 +020098struct fdtab *fdtab __read_mostly = NULL; /* array of all the file descriptors */
99struct polled_mask *polled_mask __read_mostly = NULL; /* Array for the polled_mask of each fd */
100struct fdinfo *fdinfo __read_mostly = NULL; /* less-often used infos for file descriptors */
Willy Tarreaubaaee002006-06-26 02:48:02 +0200101int totalconn; /* total # of terminated sessions */
102int actconn; /* # of active sessions */
103
Willy Tarreaua1090a52021-04-10 16:58:13 +0200104struct poller pollers[MAX_POLLERS] __read_mostly;
105struct poller cur_poller __read_mostly;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200106int nbpollers = 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200107
Olivier Houchard6b96f722018-04-25 16:58:25 +0200108volatile struct fdlist update_list; // Global update list
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100109
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200110THREAD_LOCAL int *fd_updt = NULL; // FD updates list
111THREAD_LOCAL int fd_nbupdt = 0; // number of updates in the list
Olivier Houchard79321b92018-07-26 17:55:11 +0200112THREAD_LOCAL int poller_rd_pipe = -1; // Pipe to wake the thread
Willy Tarreaua1090a52021-04-10 16:58:13 +0200113int poller_wr_pipe[MAX_THREADS] __read_mostly; // Pipe to wake the threads
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200114
Olivier Houchard7c49d2e2019-04-16 18:37:05 +0200115volatile int ha_used_fds = 0; // Number of FD we're currently using
116
Willy Tarreau337fb712019-12-20 07:20:00 +0100117#define _GET_NEXT(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->next
118#define _GET_PREV(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->prev
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100119/* adds fd <fd> to fd list <list> if it was not yet in it */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200120void fd_add_to_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100121{
122 int next;
123 int new;
124 int old;
125 int last;
126
127redo_next:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200128 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100129 /* Check that we're not already in the cache, and if not, lock us. */
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100130 if (next > -2)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100131 goto done;
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100132 if (next == -2)
133 goto redo_next;
Olivier Houchardd3608792019-03-08 18:47:42 +0100134 if (!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100135 goto redo_next;
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100136 __ha_barrier_atomic_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100137
138 new = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100139redo_last:
140 /* First, insert in the linked list */
141 last = list->last;
142 old = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100143
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200144 _GET_PREV(fd, off) = -2;
Willy Tarreau11559a72018-02-05 17:52:24 +0100145 /* Make sure the "prev" store is visible before we update the last entry */
146 __ha_barrier_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100147
Willy Tarreau11559a72018-02-05 17:52:24 +0100148 if (unlikely(last == -1)) {
149 /* list is empty, try to add ourselves alone so that list->last=fd */
Olivier Houchardd3608792019-03-08 18:47:42 +0100150 if (unlikely(!_HA_ATOMIC_CAS(&list->last, &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100151 goto redo_last;
152
153 /* list->first was necessary -1, we're guaranteed to be alone here */
154 list->first = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100155 } else {
Willy Tarreau11559a72018-02-05 17:52:24 +0100156 /* adding ourselves past the last element
157 * The CAS will only succeed if its next is -1,
158 * which means it's in the cache, and the last element.
159 */
Olivier Houchardd3608792019-03-08 18:47:42 +0100160 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(last, off), &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100161 goto redo_last;
Willy Tarreau11559a72018-02-05 17:52:24 +0100162
163 /* Then, update the last entry */
164 list->last = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100165 }
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100166 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100167 /* since we're alone at the end of the list and still locked(-2),
Ilya Shipitsinb8888ab2021-01-06 21:20:16 +0500168 * we know no one tried to add past us. Mark the end of list.
Willy Tarreau11559a72018-02-05 17:52:24 +0100169 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200170 _GET_PREV(fd, off) = last;
171 _GET_NEXT(fd, off) = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100172 __ha_barrier_store();
173done:
174 return;
175}
176
177/* removes fd <fd> from fd list <list> */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200178void fd_rm_from_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100179{
180#if defined(HA_HAVE_CAS_DW) || defined(HA_CAS_IS_8B)
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100181 volatile union {
182 struct fdlist_entry ent;
183 uint64_t u64;
184 uint32_t u32[2];
185 } cur_list, next_list;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100186#endif
187 int old;
188 int new = -2;
189 int prev;
190 int next;
191 int last;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100192lock_self:
193#if (defined(HA_CAS_IS_8B) || defined(HA_HAVE_CAS_DW))
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100194 next_list.ent.next = next_list.ent.prev = -2;
195 cur_list.ent = *(volatile struct fdlist_entry *)(((char *)&fdtab[fd]) + off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100196 /* First, attempt to lock our own entries */
197 do {
198 /* The FD is not in the FD cache, give up */
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100199 if (unlikely(cur_list.ent.next <= -3))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100200 return;
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100201 if (unlikely(cur_list.ent.prev == -2 || cur_list.ent.next == -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100202 goto lock_self;
203 } while (
204#ifdef HA_CAS_IS_8B
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100205 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 +0100206#else
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100207 unlikely(!_HA_ATOMIC_DWCAS(((long *)&_GET_NEXT(fd, off)), (uint32_t *)&cur_list.u32, &next_list.u32))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100208#endif
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100209 );
210 next = cur_list.ent.next;
211 prev = cur_list.ent.prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100212
213#else
214lock_self_next:
Willy Tarreau337fb712019-12-20 07:20:00 +0100215 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100216 if (next == -2)
217 goto lock_self_next;
218 if (next <= -3)
219 goto done;
Olivier Houchardd3608792019-03-08 18:47:42 +0100220 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100221 goto lock_self_next;
222lock_self_prev:
Willy Tarreau337fb712019-12-20 07:20:00 +0100223 prev = _GET_PREV(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100224 if (prev == -2)
225 goto lock_self_prev;
Olivier Houchardd3608792019-03-08 18:47:42 +0100226 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(fd, off), &prev, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100227 goto lock_self_prev;
228#endif
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100229 __ha_barrier_atomic_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100230
231 /* Now, lock the entries of our neighbours */
232 if (likely(prev != -1)) {
233redo_prev:
234 old = fd;
235
Olivier Houchardd3608792019-03-08 18:47:42 +0100236 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(prev, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100237 if (unlikely(old == -2)) {
238 /* Neighbour already locked, give up and
239 * retry again once he's done
240 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200241 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100242 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200243 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100244 __ha_barrier_store();
245 goto lock_self;
246 }
247 goto redo_prev;
248 }
249 }
250 if (likely(next != -1)) {
251redo_next:
252 old = fd;
Olivier Houchardd3608792019-03-08 18:47:42 +0100253 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(next, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100254 if (unlikely(old == -2)) {
255 /* Neighbour already locked, give up and
256 * retry again once he's done
257 */
258 if (prev != -1) {
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200259 _GET_NEXT(prev, off) = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100260 __ha_barrier_store();
261 }
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200262 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100263 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200264 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100265 __ha_barrier_store();
266 goto lock_self;
267 }
268 goto redo_next;
269 }
270 }
271 if (list->first == fd)
272 list->first = next;
273 __ha_barrier_store();
274 last = list->last;
Olivier Houchardd3608792019-03-08 18:47:42 +0100275 while (unlikely(last == fd && (!_HA_ATOMIC_CAS(&list->last, &last, prev))))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100276 __ha_compiler_barrier();
277 /* Make sure we let other threads know we're no longer in cache,
278 * before releasing our neighbours.
279 */
280 __ha_barrier_store();
281 if (likely(prev != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200282 _GET_NEXT(prev, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100283 __ha_barrier_store();
284 if (likely(next != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200285 _GET_PREV(next, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100286 __ha_barrier_store();
287 /* Ok, now we're out of the fd cache */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200288 _GET_NEXT(fd, off) = -(next + 4);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100289 __ha_barrier_store();
290done:
291 return;
292}
293
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200294#undef _GET_NEXT
295#undef _GET_PREV
296
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100297/* deletes the FD once nobody uses it anymore, as detected by the caller by its
298 * thread_mask being zero and its running mask turning to zero. There is no
299 * protection against concurrent accesses, it's up to the caller to make sure
300 * only the last thread will call it. This is only for internal use, please use
301 * fd_delete() instead.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200302 */
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100303void _fd_delete_orphan(int fd)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200304{
Willy Tarreaub41a6e92021-04-06 17:49:19 +0200305 if (fdtab[fd].state & FD_LINGER_RISK) {
Willy Tarreauad38ace2013-12-15 14:19:38 +0100306 /* this is generally set when connecting to servers */
Ilya Shipitsinb7e43f02020-04-02 15:02:08 +0500307 DISGUISE(setsockopt(fd, SOL_SOCKET, SO_LINGER,
308 (struct linger *) &nolinger, sizeof(struct linger)));
Willy Tarreauad38ace2013-12-15 14:19:38 +0100309 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100310 if (cur_poller.clo)
311 cur_poller.clo(fd);
Willy Tarreau2d423292021-03-24 15:34:25 +0100312
313 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
Olivier Houchardc22580c2019-08-05 18:51:52 +0200314 polled_mask[fd].poll_recv = polled_mask[fd].poll_send = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100315
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100316 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100317
Willy Tarreau38e8a1c2020-06-23 10:04:54 +0200318#ifdef DEBUG_FD
319 fdtab[fd].event_count = 0;
320#endif
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200321 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200322 fdtab[fd].owner = NULL;
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100323 /* perform the close() call last as it's what unlocks the instant reuse
324 * of this FD by any other thread.
325 */
Willy Tarreau63d8b602020-08-26 11:54:06 +0200326 close(fd);
Willy Tarreau4781b152021-04-06 13:53:36 +0200327 _HA_ATOMIC_DEC(&ha_used_fds);
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100328}
329
Olivier Houchardc23b3372021-03-25 01:38:54 +0100330#ifndef HA_HAVE_CAS_DW
331__decl_thread(__decl_rwlock(fd_mig_lock));
332#endif
333
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100334/* Deletes an FD from the fdsets. The file descriptor is also closed, possibly
335 * asynchronously. Only the owning thread may do this.
336 */
337void fd_delete(int fd)
338{
339 /* we must postpone removal of an FD that may currently be in use
Ilya Shipitsinb2be9a12021-04-24 13:25:42 +0500340 * by another thread. This can happen in the following two situations:
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100341 * - after a takeover, the owning thread closes the connection but
342 * the previous one just woke up from the poller and entered
343 * the FD handler iocb. That thread holds an entry in running_mask
344 * and requires removal protection.
345 * - multiple threads are accepting connections on a listener, and
346 * one of them (or even an separate one) decides to unbind the
347 * listener under the listener's lock while other ones still hold
348 * the running bit.
349 * In both situations the FD is marked as unused (thread_mask = 0) and
350 * will not take new bits in its running_mask so we have the guarantee
351 * that the last thread eliminating running_mask is the one allowed to
352 * safely delete the FD. Most of the time it will be the current thread.
353 */
354
355 HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
Olivier Houchardc23b3372021-03-25 01:38:54 +0100356#ifndef HA_HAVE_CAS_DW
357 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
358#endif
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100359 HA_ATOMIC_STORE(&fdtab[fd].thread_mask, 0);
Olivier Houchardc23b3372021-03-25 01:38:54 +0100360#ifndef HA_HAVE_CAS_DW
361 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
362#endif
Willy Tarreau2c3f9812021-03-24 10:51:32 +0100363 if (fd_clr_running(fd) == 0)
364 _fd_delete_orphan(fd);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200365}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200366
Olivier Houchard88516642020-03-05 18:10:51 +0100367/*
368 * Take over a FD belonging to another thread.
369 * unexpected_conn is the expected owner of the fd.
370 * Returns 0 on success, and -1 on failure.
371 */
372int fd_takeover(int fd, void *expected_owner)
373{
Willy Tarreauc460c912020-06-18 07:28:09 +0200374 int ret = -1;
Olivier Houchard88516642020-03-05 18:10:51 +0100375
Willy Tarreauf1cad382020-06-18 08:14:59 +0200376#ifndef HA_HAVE_CAS_DW
Willy Tarreau1db42732021-04-06 11:44:07 +0200377 if (_HA_ATOMIC_OR_FETCH(&fdtab[fd].running_mask, tid_bit) == tid_bit) {
Willy Tarreauc460c912020-06-18 07:28:09 +0200378 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
379 if (fdtab[fd].owner == expected_owner) {
380 fdtab[fd].thread_mask = tid_bit;
381 ret = 0;
382 }
383 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
Olivier Houchard88516642020-03-05 18:10:51 +0100384 }
Olivier Houchard88516642020-03-05 18:10:51 +0100385#else
386 unsigned long old_masks[2];
387 unsigned long new_masks[2];
388
Olivier Houchard88516642020-03-05 18:10:51 +0100389 new_masks[0] = new_masks[1] = tid_bit;
Willy Tarreau42973632020-06-18 08:05:15 +0200390
Willy Tarreau1db42732021-04-06 11:44:07 +0200391 old_masks[0] = _HA_ATOMIC_OR_FETCH(&fdtab[fd].running_mask, tid_bit);
Willy Tarreau42973632020-06-18 08:05:15 +0200392 old_masks[1] = fdtab[fd].thread_mask;
393
Olivier Houchard88516642020-03-05 18:10:51 +0100394 /* protect ourself against a delete then an insert for the same fd,
395 * if it happens, then the owner will no longer be the expected
396 * connection.
397 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200398 if (fdtab[fd].owner == expected_owner) {
399 while (old_masks[0] == tid_bit && old_masks[1]) {
400 if (_HA_ATOMIC_DWCAS(&fdtab[fd].running_mask, &old_masks, &new_masks)) {
401 ret = 0;
402 break;
403 }
Olivier Houchard88516642020-03-05 18:10:51 +0100404 }
Willy Tarreauf1cad382020-06-18 08:14:59 +0200405 }
406#endif /* HW_HAVE_CAS_DW */
407
Olivier Houchard88516642020-03-05 18:10:51 +0100408 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
Willy Tarreauf1cad382020-06-18 08:14:59 +0200409
Olivier Houchardddc874c2020-06-17 20:34:05 +0200410 /* Make sure the FD doesn't have the active bit. It is possible that
411 * the fd is polled by the thread that used to own it, the new thread
412 * is supposed to call subscribe() later, to activate polling.
413 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200414 if (likely(ret == 0))
415 fd_stop_recv(fd);
416 return ret;
Olivier Houchard88516642020-03-05 18:10:51 +0100417}
418
Willy Tarreaudbe30602019-09-04 13:25:41 +0200419void updt_fd_polling(const int fd)
420{
Willy Tarreau5a7d6eb2020-11-26 22:25:10 +0100421 if (all_threads_mask == 1UL || (fdtab[fd].thread_mask & all_threads_mask) == tid_bit) {
Willy Tarreaudbe30602019-09-04 13:25:41 +0200422 if (HA_ATOMIC_BTS(&fdtab[fd].update_mask, tid))
423 return;
424
425 fd_updt[fd_nbupdt++] = fd;
426 } else {
427 unsigned long update_mask = fdtab[fd].update_mask;
428 do {
429 if (update_mask == fdtab[fd].thread_mask)
430 return;
Willy Tarreauf0158872020-09-25 12:18:53 +0200431 } while (!_HA_ATOMIC_CAS(&fdtab[fd].update_mask, &update_mask, fdtab[fd].thread_mask));
432
Willy Tarreaudbe30602019-09-04 13:25:41 +0200433 fd_add_to_fd_list(&update_list, fd, offsetof(struct fdtab, update));
Willy Tarreauf0158872020-09-25 12:18:53 +0200434
435 if (fd_active(fd) &&
436 !(fdtab[fd].thread_mask & tid_bit) &&
437 (fdtab[fd].thread_mask & ~tid_bit & all_threads_mask & ~sleeping_thread_mask) == 0) {
438 /* we need to wake up one thread to handle it immediately */
439 int thr = my_ffsl(fdtab[fd].thread_mask & ~tid_bit & all_threads_mask) - 1;
440
441 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
442 wake_thread(thr);
443 }
Willy Tarreaudbe30602019-09-04 13:25:41 +0200444 }
445}
446
Willy Tarreau931d8b72019-08-27 11:08:17 +0200447/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
448 * optionally followed by a newline if <nl> is non-null, to file descriptor
449 * <fd>. The message is sent atomically using writev(). It may be truncated to
450 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
451 * two lists, it's just a convenience to help the caller prepend some prefixes
452 * when necessary. It takes the fd's lock to make sure no other thread will
453 * write to the same fd in parallel. Returns the number of bytes sent, or <=0
454 * on failure. A limit to 31 total non-empty segments is enforced. The caller
455 * is responsible for taking care of making the fd non-blocking.
456 */
457ssize_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)
458{
459 struct iovec iovec[32];
460 size_t totlen = 0;
461 size_t sent = 0;
462 int vec = 0;
Willy Tarreaudf187872020-06-11 14:25:47 +0200463 int attempts = 0;
Willy Tarreau931d8b72019-08-27 11:08:17 +0200464
465 if (!maxlen)
466 maxlen = ~0;
467
468 /* keep one char for a possible trailing '\n' in any case */
469 maxlen--;
470
471 /* make an iovec from the concatenation of all parts of the original
472 * message. Skip empty fields and truncate the whole message to maxlen,
473 * leaving one spare iovec for the '\n'.
474 */
475 while (vec < (sizeof(iovec) / sizeof(iovec[0]) - 1)) {
476 if (!npfx) {
477 pfx = msg;
478 npfx = nmsg;
479 nmsg = 0;
480 if (!npfx)
481 break;
482 }
483
484 iovec[vec].iov_base = pfx->ptr;
485 iovec[vec].iov_len = MIN(maxlen, pfx->len);
486 maxlen -= iovec[vec].iov_len;
487 totlen += iovec[vec].iov_len;
488 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 Tarreau7c9f7562020-10-13 15:45:07 +0200687 if ((fdtab = calloc(global.maxsock, sizeof(*fdtab))) == NULL) {
688 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 Tarreau7c9f7562020-10-13 15:45:07 +0200692 if ((polled_mask = calloc(global.maxsock, sizeof(*polled_mask))) == NULL) {
693 ha_alert("Not enough memory to allocate %d entries for polled_mask!\n", global.maxsock);
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200694 goto fail_polledmask;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200695 }
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000696
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200697 if ((fdinfo = calloc(global.maxsock, sizeof(*fdinfo))) == NULL) {
698 ha_alert("Not enough memory to allocate %d entries for fdinfo!\n", global.maxsock);
Christopher Faulet63fe6522017-08-31 17:52:09 +0200699 goto fail_info;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200700 }
Christopher Faulet63fe6522017-08-31 17:52:09 +0200701
Olivier Houchard6b96f722018-04-25 16:58:25 +0200702 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200703
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100704 for (p = 0; p < global.maxsock; p++) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100705 /* Mark the fd as out of the fd cache */
Olivier Houchard6b96f722018-04-25 16:58:25 +0200706 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100707 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200708
Willy Tarreau4f60f162007-04-08 16:39:58 +0200709 do {
710 bp = NULL;
711 for (p = 0; p < nbpollers; p++)
712 if (!bp || (pollers[p].pref > bp->pref))
713 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200714
Willy Tarreau4f60f162007-04-08 16:39:58 +0200715 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200716 break;
717
Willy Tarreau4f60f162007-04-08 16:39:58 +0200718 if (bp->init(bp)) {
719 memcpy(&cur_poller, bp, sizeof(*bp));
720 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200721 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200722 } while (!bp || bp->pref == 0);
Willy Tarreau7be79a42012-11-11 15:02:54 +0100723
Christopher Faulet63fe6522017-08-31 17:52:09 +0200724 free(fdinfo);
725 fail_info:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200726 free(polled_mask);
727 fail_polledmask:
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000728 free(fdtab);
729 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100730 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200731}
732
Willy Tarreaubaaee002006-06-26 02:48:02 +0200733/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200734 * Deinitialize the pollers.
735 */
736void deinit_pollers() {
737
738 struct poller *bp;
739 int p;
740
741 for (p = 0; p < nbpollers; p++) {
742 bp = &pollers[p];
743
744 if (bp && bp->pref)
745 bp->term(bp);
746 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200747
Willy Tarreau61cfdf42021-02-20 10:46:51 +0100748 ha_free(&fdinfo);
749 ha_free(&fdtab);
750 ha_free(&polled_mask);
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200751}
752
753/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200754 * Lists the known pollers on <out>.
755 * Should be performed only before initialization.
756 */
757int list_pollers(FILE *out)
758{
759 int p;
760 int last, next;
761 int usable;
762 struct poller *bp;
763
764 fprintf(out, "Available polling systems :\n");
765
766 usable = 0;
767 bp = NULL;
768 last = next = -1;
769 while (1) {
770 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200771 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100772 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200773 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100774 if (!bp || (pollers[p].pref > bp->pref))
775 bp = &pollers[p];
776 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200777 }
778
779 if (next == -1)
780 break;
781
782 for (p = 0; p < nbpollers; p++) {
783 if (pollers[p].pref == next) {
784 fprintf(out, " %10s : ", pollers[p].name);
785 if (pollers[p].pref == 0)
786 fprintf(out, "disabled, ");
787 else
788 fprintf(out, "pref=%3d, ", pollers[p].pref);
789 if (pollers[p].test(&pollers[p])) {
790 fprintf(out, " test result OK");
791 if (next > 0)
792 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100793 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200794 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100795 if (bp == &pollers[p])
796 bp = NULL;
797 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200798 fprintf(out, "\n");
799 }
800 }
801 last = next;
802 next = -1;
803 };
804 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
805 return 0;
806}
807
808/*
809 * Some pollers may lose their connection after a fork(). It may be necessary
810 * to create initialize part of them again. Returns 0 in case of failure,
811 * otherwise 1. The fork() function may be NULL if unused. In case of error,
812 * the the current poller is destroyed and the caller is responsible for trying
813 * another one by calling init_pollers() again.
814 */
815int fork_poller()
816{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200817 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100818 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200819 if (fdtab[fd].owner) {
Willy Tarreau030dae12021-04-06 17:53:33 +0200820 HA_ATOMIC_OR(&fdtab[fd].state, FD_CLONED);
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200821 }
822 }
823
Willy Tarreau2ff76222007-04-09 19:29:56 +0200824 if (cur_poller.fork) {
825 if (cur_poller.fork(&cur_poller))
826 return 1;
827 cur_poller.term(&cur_poller);
828 return 0;
829 }
830 return 1;
831}
832
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200833/* config parser for global "tune.fd.edge-triggered", accepts "on" or "off" */
834static int cfg_parse_tune_fd_edge_triggered(char **args, int section_type, struct proxy *curpx,
Willy Tarreau01825162021-03-09 09:53:46 +0100835 const struct proxy *defpx, const char *file, int line,
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200836 char **err)
837{
838 if (too_many_args(1, args, err, NULL))
839 return -1;
840
841 if (strcmp(args[1], "on") == 0)
842 global.tune.options |= GTUNE_FD_ET;
843 else if (strcmp(args[1], "off") == 0)
844 global.tune.options &= ~GTUNE_FD_ET;
845 else {
846 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
847 return -1;
848 }
849 return 0;
850}
851
852/* config keyword parsers */
853static struct cfg_kw_list cfg_kws = {ILH, {
854 { CFG_GLOBAL, "tune.fd.edge-triggered", cfg_parse_tune_fd_edge_triggered },
855 { 0, NULL, NULL }
856}};
857
858INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
859
Willy Tarreau082b6282019-05-22 14:42:12 +0200860REGISTER_PER_THREAD_ALLOC(alloc_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100861REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
862REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
Willy Tarreau082b6282019-05-22 14:42:12 +0200863REGISTER_PER_THREAD_FREE(free_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100864
Willy Tarreau2ff76222007-04-09 19:29:56 +0200865/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200866 * Local variables:
867 * c-indent-level: 8
868 * c-basic-offset: 8
869 * End:
870 */