blob: 79a9ec861ae71fce282eb6c2ce205032498f00d5 [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 | |
65 * | v | EAGAIN (cant)
66 * | +--------+
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 Tarreaubaaee002006-06-26 02:48:02 +020098struct fdtab *fdtab = NULL; /* array of all the file descriptors */
Olivier Houchard53055052019-07-25 14:00:18 +000099struct polled_mask *polled_mask = NULL; /* Array for the polled_mask of each fd */
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200100struct fdinfo *fdinfo = 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 Tarreau4f60f162007-04-08 16:39:58 +0200104struct poller pollers[MAX_POLLERS];
105struct poller cur_poller;
106int 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
113int poller_wr_pipe[MAX_THREADS]; // 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),
168 * we know noone tried to add past us. Mark the end of list.
169 */
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 Tarreau173d9952018-01-26 21:48:23 +0100297/* Deletes an FD from the fdsets.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200298 * The file descriptor is also closed.
299 */
Willy Tarreau63d8b602020-08-26 11:54:06 +0200300void fd_delete(int fd)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200301{
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100302 int locked = fdtab[fd].running_mask != tid_bit;
Willy Tarreau87d54a92018-10-15 09:44:46 +0200303
Olivier Houchard88516642020-03-05 18:10:51 +0100304 /* We're just trying to protect against a concurrent fd_insert()
Olivier Houchardf86a1062020-06-17 20:28:27 +0200305 * here, not against fd_takeover(), because either we're called
Olivier Houchard88516642020-03-05 18:10:51 +0100306 * directly from the iocb(), and we're already locked, or we're
307 * called from the mux tasklet, but then the mux is responsible for
308 * making sure the tasklet does nothing, and the connection is never
309 * destroyed.
310 */
Willy Tarreau87d54a92018-10-15 09:44:46 +0200311 if (locked)
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100312 fd_set_running_excl(fd);
313
Willy Tarreauad38ace2013-12-15 14:19:38 +0100314 if (fdtab[fd].linger_risk) {
315 /* this is generally set when connecting to servers */
Ilya Shipitsinb7e43f02020-04-02 15:02:08 +0500316 DISGUISE(setsockopt(fd, SOL_SOCKET, SO_LINGER,
317 (struct linger *) &nolinger, sizeof(struct linger)));
Willy Tarreauad38ace2013-12-15 14:19:38 +0100318 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100319 if (cur_poller.clo)
320 cur_poller.clo(fd);
Olivier Houchardc22580c2019-08-05 18:51:52 +0200321 polled_mask[fd].poll_recv = polled_mask[fd].poll_send = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100322
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100323 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100324
Willy Tarreau38e8a1c2020-06-23 10:04:54 +0200325#ifdef DEBUG_FD
326 fdtab[fd].event_count = 0;
327#endif
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200328 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
329 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200330 fdtab[fd].owner = NULL;
Willy Tarreauf65610a2017-10-31 16:06:06 +0100331 fdtab[fd].thread_mask = 0;
Willy Tarreaubb1caff2020-08-19 10:00:57 +0200332 fdtab[fd].exported = 0;
Willy Tarreau63d8b602020-08-26 11:54:06 +0200333 close(fd);
334 _HA_ATOMIC_SUB(&ha_used_fds, 1);
Willy Tarreau87d54a92018-10-15 09:44:46 +0200335 if (locked)
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100336 fd_clr_running(fd);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200337}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200338
Olivier Houchard88516642020-03-05 18:10:51 +0100339#ifndef HA_HAVE_CAS_DW
Willy Tarreauaf613e82020-06-05 08:40:51 +0200340__decl_thread(__decl_rwlock(fd_mig_lock));
Olivier Houchard88516642020-03-05 18:10:51 +0100341#endif
342
343/*
344 * Take over a FD belonging to another thread.
345 * unexpected_conn is the expected owner of the fd.
346 * Returns 0 on success, and -1 on failure.
347 */
348int fd_takeover(int fd, void *expected_owner)
349{
Willy Tarreauc460c912020-06-18 07:28:09 +0200350 int ret = -1;
Olivier Houchard88516642020-03-05 18:10:51 +0100351
Willy Tarreauf1cad382020-06-18 08:14:59 +0200352#ifndef HA_HAVE_CAS_DW
Willy Tarreauc460c912020-06-18 07:28:09 +0200353 if (_HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit) == tid_bit) {
354 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
355 if (fdtab[fd].owner == expected_owner) {
356 fdtab[fd].thread_mask = tid_bit;
357 ret = 0;
358 }
359 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
Olivier Houchard88516642020-03-05 18:10:51 +0100360 }
Olivier Houchard88516642020-03-05 18:10:51 +0100361#else
362 unsigned long old_masks[2];
363 unsigned long new_masks[2];
364
Olivier Houchard88516642020-03-05 18:10:51 +0100365 new_masks[0] = new_masks[1] = tid_bit;
Willy Tarreau42973632020-06-18 08:05:15 +0200366
367 old_masks[0] = _HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
368 old_masks[1] = fdtab[fd].thread_mask;
369
Olivier Houchard88516642020-03-05 18:10:51 +0100370 /* protect ourself against a delete then an insert for the same fd,
371 * if it happens, then the owner will no longer be the expected
372 * connection.
373 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200374 if (fdtab[fd].owner == expected_owner) {
375 while (old_masks[0] == tid_bit && old_masks[1]) {
376 if (_HA_ATOMIC_DWCAS(&fdtab[fd].running_mask, &old_masks, &new_masks)) {
377 ret = 0;
378 break;
379 }
Olivier Houchard88516642020-03-05 18:10:51 +0100380 }
Willy Tarreauf1cad382020-06-18 08:14:59 +0200381 }
382#endif /* HW_HAVE_CAS_DW */
383
Olivier Houchard88516642020-03-05 18:10:51 +0100384 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
Willy Tarreauf1cad382020-06-18 08:14:59 +0200385
Olivier Houchardddc874c2020-06-17 20:34:05 +0200386 /* Make sure the FD doesn't have the active bit. It is possible that
387 * the fd is polled by the thread that used to own it, the new thread
388 * is supposed to call subscribe() later, to activate polling.
389 */
Willy Tarreauf1cad382020-06-18 08:14:59 +0200390 if (likely(ret == 0))
391 fd_stop_recv(fd);
392 return ret;
Olivier Houchard88516642020-03-05 18:10:51 +0100393}
394
Willy Tarreaudbe30602019-09-04 13:25:41 +0200395void updt_fd_polling(const int fd)
396{
397 if ((fdtab[fd].thread_mask & all_threads_mask) == tid_bit) {
398
399 /* note: we don't have a test-and-set yet in hathreads */
400
401 if (HA_ATOMIC_BTS(&fdtab[fd].update_mask, tid))
402 return;
403
404 fd_updt[fd_nbupdt++] = fd;
405 } else {
406 unsigned long update_mask = fdtab[fd].update_mask;
407 do {
408 if (update_mask == fdtab[fd].thread_mask)
409 return;
Willy Tarreauf0158872020-09-25 12:18:53 +0200410 } while (!_HA_ATOMIC_CAS(&fdtab[fd].update_mask, &update_mask, fdtab[fd].thread_mask));
411
Willy Tarreaudbe30602019-09-04 13:25:41 +0200412 fd_add_to_fd_list(&update_list, fd, offsetof(struct fdtab, update));
Willy Tarreauf0158872020-09-25 12:18:53 +0200413
414 if (fd_active(fd) &&
415 !(fdtab[fd].thread_mask & tid_bit) &&
416 (fdtab[fd].thread_mask & ~tid_bit & all_threads_mask & ~sleeping_thread_mask) == 0) {
417 /* we need to wake up one thread to handle it immediately */
418 int thr = my_ffsl(fdtab[fd].thread_mask & ~tid_bit & all_threads_mask) - 1;
419
420 _HA_ATOMIC_AND(&sleeping_thread_mask, ~(1UL << thr));
421 wake_thread(thr);
422 }
Willy Tarreaudbe30602019-09-04 13:25:41 +0200423 }
424}
425
Olivier Houchard7fa55622020-02-27 17:25:43 +0100426__decl_spinlock(log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200427/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
428 * optionally followed by a newline if <nl> is non-null, to file descriptor
429 * <fd>. The message is sent atomically using writev(). It may be truncated to
430 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
431 * two lists, it's just a convenience to help the caller prepend some prefixes
432 * when necessary. It takes the fd's lock to make sure no other thread will
433 * write to the same fd in parallel. Returns the number of bytes sent, or <=0
434 * on failure. A limit to 31 total non-empty segments is enforced. The caller
435 * is responsible for taking care of making the fd non-blocking.
436 */
437ssize_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)
438{
439 struct iovec iovec[32];
440 size_t totlen = 0;
441 size_t sent = 0;
442 int vec = 0;
Willy Tarreaudf187872020-06-11 14:25:47 +0200443 int attempts = 0;
Willy Tarreau931d8b72019-08-27 11:08:17 +0200444
445 if (!maxlen)
446 maxlen = ~0;
447
448 /* keep one char for a possible trailing '\n' in any case */
449 maxlen--;
450
451 /* make an iovec from the concatenation of all parts of the original
452 * message. Skip empty fields and truncate the whole message to maxlen,
453 * leaving one spare iovec for the '\n'.
454 */
455 while (vec < (sizeof(iovec) / sizeof(iovec[0]) - 1)) {
456 if (!npfx) {
457 pfx = msg;
458 npfx = nmsg;
459 nmsg = 0;
460 if (!npfx)
461 break;
462 }
463
464 iovec[vec].iov_base = pfx->ptr;
465 iovec[vec].iov_len = MIN(maxlen, pfx->len);
466 maxlen -= iovec[vec].iov_len;
467 totlen += iovec[vec].iov_len;
468 if (iovec[vec].iov_len)
469 vec++;
470 pfx++; npfx--;
471 };
472
473 if (nl) {
474 iovec[vec].iov_base = "\n";
475 iovec[vec].iov_len = 1;
476 vec++;
477 }
478
Willy Tarreaudf187872020-06-11 14:25:47 +0200479 /* make sure we never interleave writes and we never block. This means
480 * we prefer to fail on collision than to block. But we don't want to
481 * lose too many logs so we just perform a few lock attempts then give
482 * up.
483 */
484
485 while (HA_SPIN_TRYLOCK(OTHER_LOCK, &log_lock) != 0) {
486 if (++attempts >= 200) {
487 /* so that the caller knows the message couldn't be delivered */
488 sent = -1;
489 errno = EAGAIN;
490 goto leave;
491 }
492 ha_thread_relax();
493 }
494
Willy Tarreau7e9776a2019-08-30 14:41:47 +0200495 if (unlikely(!fdtab[fd].initialized)) {
496 fdtab[fd].initialized = 1;
497 if (!isatty(fd))
498 fcntl(fd, F_SETFL, O_NONBLOCK);
499 }
Willy Tarreau931d8b72019-08-27 11:08:17 +0200500 sent = writev(fd, iovec, vec);
Olivier Houchard7fa55622020-02-27 17:25:43 +0100501 HA_SPIN_UNLOCK(OTHER_LOCK, &log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200502
Willy Tarreaudf187872020-06-11 14:25:47 +0200503 leave:
Willy Tarreau931d8b72019-08-27 11:08:17 +0200504 /* sent > 0 if the message was delivered */
505 return sent;
506}
507
Olivier Houchard2292edf2019-02-25 14:26:54 +0100508#if defined(USE_CLOSEFROM)
509void my_closefrom(int start)
510{
511 closefrom(start);
512}
513
Willy Tarreaue5733232019-05-22 19:24:06 +0200514#elif defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100515/* This is a portable implementation of closefrom(). It closes all open file
516 * descriptors starting at <start> and above. It relies on the fact that poll()
517 * will return POLLNVAL for each invalid (hence close) file descriptor passed
518 * in argument in order to skip them. It acts with batches of FDs and will
519 * typically perform one poll() call per 1024 FDs so the overhead is low in
520 * case all FDs have to be closed.
521 */
522void my_closefrom(int start)
523{
524 struct pollfd poll_events[1024];
525 struct rlimit limit;
526 int nbfds, fd, ret, idx;
527 int step, next;
528
529 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
530 step = nbfds = limit.rlim_cur;
531 else
532 step = nbfds = 0;
533
534 if (nbfds <= 0) {
535 /* set safe limit */
536 nbfds = 1024;
537 step = 256;
538 }
539
540 if (step > sizeof(poll_events) / sizeof(poll_events[0]))
541 step = sizeof(poll_events) / sizeof(poll_events[0]);
542
543 while (start < nbfds) {
544 next = (start / step + 1) * step;
545
546 for (fd = start; fd < next && fd < nbfds; fd++) {
547 poll_events[fd - start].fd = fd;
548 poll_events[fd - start].events = 0;
549 }
550
551 do {
552 ret = poll(poll_events, fd - start, 0);
553 if (ret >= 0)
554 break;
555 } while (errno == EAGAIN || errno == EINTR || errno == ENOMEM);
556
Willy Tarreaub8e602c2019-02-22 09:07:42 +0100557 if (ret)
558 ret = fd - start;
559
Willy Tarreau9188ac62019-02-21 22:12:47 +0100560 for (idx = 0; idx < ret; idx++) {
561 if (poll_events[idx].revents & POLLNVAL)
562 continue; /* already closed */
563
564 fd = poll_events[idx].fd;
565 close(fd);
566 }
567 start = next;
568 }
569}
570
Willy Tarreaue5733232019-05-22 19:24:06 +0200571#else // defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100572
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100573/* This is a portable implementation of closefrom(). It closes all open file
574 * descriptors starting at <start> and above. This is a naive version for use
575 * when the operating system provides no alternative.
576 */
577void my_closefrom(int start)
578{
579 struct rlimit limit;
580 int nbfds;
581
582 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
583 nbfds = limit.rlim_cur;
584 else
585 nbfds = 0;
586
587 if (nbfds <= 0)
588 nbfds = 1024; /* safe limit */
589
590 while (start < nbfds)
591 close(start++);
592}
Willy Tarreaue5733232019-05-22 19:24:06 +0200593#endif // defined(USE_POLL)
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100594
Willy Tarreau4f60f162007-04-08 16:39:58 +0200595/* disable the specified poller */
596void disable_poller(const char *poller_name)
597{
598 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200599
Willy Tarreau4f60f162007-04-08 16:39:58 +0200600 for (p = 0; p < nbpollers; p++)
601 if (strcmp(pollers[p].name, poller_name) == 0)
602 pollers[p].pref = 0;
603}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200604
Olivier Houchard79321b92018-07-26 17:55:11 +0200605void poller_pipe_io_handler(int fd)
606{
607 char buf[1024];
608 /* Flush the pipe */
609 while (read(fd, buf, sizeof(buf)) > 0);
610 fd_cant_recv(fd);
611}
612
Willy Tarreau082b6282019-05-22 14:42:12 +0200613/* allocate the per-thread fd_updt thus needs to be called early after
614 * thread creation.
615 */
616static int alloc_pollers_per_thread()
617{
618 fd_updt = calloc(global.maxsock, sizeof(*fd_updt));
619 return fd_updt != NULL;
620}
621
622/* Initialize the pollers per thread.*/
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200623static int init_pollers_per_thread()
624{
Olivier Houchard79321b92018-07-26 17:55:11 +0200625 int mypipe[2];
Willy Tarreau082b6282019-05-22 14:42:12 +0200626
627 if (pipe(mypipe) < 0)
Olivier Houchard79321b92018-07-26 17:55:11 +0200628 return 0;
Willy Tarreau082b6282019-05-22 14:42:12 +0200629
Olivier Houchard79321b92018-07-26 17:55:11 +0200630 poller_rd_pipe = mypipe[0];
631 poller_wr_pipe[tid] = mypipe[1];
632 fcntl(poller_rd_pipe, F_SETFL, O_NONBLOCK);
633 fd_insert(poller_rd_pipe, poller_pipe_io_handler, poller_pipe_io_handler,
634 tid_bit);
635 fd_want_recv(poller_rd_pipe);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200636 return 1;
637}
638
639/* Deinitialize the pollers per thread */
640static void deinit_pollers_per_thread()
641{
William Lallemand808e1b72018-12-15 22:34:31 +0100642 /* rd and wr are init at the same place, but only rd is init to -1, so
643 we rely to rd to close. */
644 if (poller_rd_pipe > -1) {
645 close(poller_rd_pipe);
646 poller_rd_pipe = -1;
647 close(poller_wr_pipe[tid]);
648 poller_wr_pipe[tid] = -1;
649 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200650}
651
Willy Tarreau082b6282019-05-22 14:42:12 +0200652/* Release the pollers per thread, to be called late */
653static void free_pollers_per_thread()
654{
655 free(fd_updt);
656 fd_updt = NULL;
657}
658
Willy Tarreaubaaee002006-06-26 02:48:02 +0200659/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200660 * Initialize the pollers till the best one is found.
661 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200662 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200663int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200664{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200665 int p;
666 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200667
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200668 if ((fdtab = calloc(global.maxsock, sizeof(*fdtab))) == NULL) {
669 ha_alert("Not enough memory to allocate %d entries for fdtab!\n", global.maxsock);
Christopher Faulet63fe6522017-08-31 17:52:09 +0200670 goto fail_tab;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200671 }
Christopher Faulet63fe6522017-08-31 17:52:09 +0200672
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200673 if ((polled_mask = calloc(global.maxsock, sizeof(*polled_mask))) == NULL) {
674 ha_alert("Not enough memory to allocate %d entries for polled_mask!\n", global.maxsock);
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200675 goto fail_polledmask;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200676 }
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000677
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200678 if ((fdinfo = calloc(global.maxsock, sizeof(*fdinfo))) == NULL) {
679 ha_alert("Not enough memory to allocate %d entries for fdinfo!\n", global.maxsock);
Christopher Faulet63fe6522017-08-31 17:52:09 +0200680 goto fail_info;
Willy Tarreau7c9f7562020-10-13 15:45:07 +0200681 }
Christopher Faulet63fe6522017-08-31 17:52:09 +0200682
Olivier Houchard6b96f722018-04-25 16:58:25 +0200683 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200684
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100685 for (p = 0; p < global.maxsock; p++) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100686 /* Mark the fd as out of the fd cache */
Olivier Houchard6b96f722018-04-25 16:58:25 +0200687 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100688 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200689
Willy Tarreau4f60f162007-04-08 16:39:58 +0200690 do {
691 bp = NULL;
692 for (p = 0; p < nbpollers; p++)
693 if (!bp || (pollers[p].pref > bp->pref))
694 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200695
Willy Tarreau4f60f162007-04-08 16:39:58 +0200696 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200697 break;
698
Willy Tarreau4f60f162007-04-08 16:39:58 +0200699 if (bp->init(bp)) {
700 memcpy(&cur_poller, bp, sizeof(*bp));
701 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200702 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200703 } while (!bp || bp->pref == 0);
Willy Tarreau7be79a42012-11-11 15:02:54 +0100704
Christopher Faulet63fe6522017-08-31 17:52:09 +0200705 free(fdinfo);
706 fail_info:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200707 free(polled_mask);
708 fail_polledmask:
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000709 free(fdtab);
710 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100711 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200712}
713
Willy Tarreaubaaee002006-06-26 02:48:02 +0200714/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200715 * Deinitialize the pollers.
716 */
717void deinit_pollers() {
718
719 struct poller *bp;
720 int p;
721
722 for (p = 0; p < nbpollers; p++) {
723 bp = &pollers[p];
724
725 if (bp && bp->pref)
726 bp->term(bp);
727 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200728
Christopher Faulet63fe6522017-08-31 17:52:09 +0200729 free(fdinfo); fdinfo = NULL;
730 free(fdtab); fdtab = NULL;
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200731 free(polled_mask); polled_mask = NULL;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200732}
733
734/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200735 * Lists the known pollers on <out>.
736 * Should be performed only before initialization.
737 */
738int list_pollers(FILE *out)
739{
740 int p;
741 int last, next;
742 int usable;
743 struct poller *bp;
744
745 fprintf(out, "Available polling systems :\n");
746
747 usable = 0;
748 bp = NULL;
749 last = next = -1;
750 while (1) {
751 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200752 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100753 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200754 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100755 if (!bp || (pollers[p].pref > bp->pref))
756 bp = &pollers[p];
757 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200758 }
759
760 if (next == -1)
761 break;
762
763 for (p = 0; p < nbpollers; p++) {
764 if (pollers[p].pref == next) {
765 fprintf(out, " %10s : ", pollers[p].name);
766 if (pollers[p].pref == 0)
767 fprintf(out, "disabled, ");
768 else
769 fprintf(out, "pref=%3d, ", pollers[p].pref);
770 if (pollers[p].test(&pollers[p])) {
771 fprintf(out, " test result OK");
772 if (next > 0)
773 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100774 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200775 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100776 if (bp == &pollers[p])
777 bp = NULL;
778 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200779 fprintf(out, "\n");
780 }
781 }
782 last = next;
783 next = -1;
784 };
785 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
786 return 0;
787}
788
789/*
790 * Some pollers may lose their connection after a fork(). It may be necessary
791 * to create initialize part of them again. Returns 0 in case of failure,
792 * otherwise 1. The fork() function may be NULL if unused. In case of error,
793 * the the current poller is destroyed and the caller is responsible for trying
794 * another one by calling init_pollers() again.
795 */
796int fork_poller()
797{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200798 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100799 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200800 if (fdtab[fd].owner) {
801 fdtab[fd].cloned = 1;
802 }
803 }
804
Willy Tarreau2ff76222007-04-09 19:29:56 +0200805 if (cur_poller.fork) {
806 if (cur_poller.fork(&cur_poller))
807 return 1;
808 cur_poller.term(&cur_poller);
809 return 0;
810 }
811 return 1;
812}
813
Willy Tarreaubc52bec2020-06-18 08:58:47 +0200814/* config parser for global "tune.fd.edge-triggered", accepts "on" or "off" */
815static int cfg_parse_tune_fd_edge_triggered(char **args, int section_type, struct proxy *curpx,
816 struct proxy *defpx, const char *file, int line,
817 char **err)
818{
819 if (too_many_args(1, args, err, NULL))
820 return -1;
821
822 if (strcmp(args[1], "on") == 0)
823 global.tune.options |= GTUNE_FD_ET;
824 else if (strcmp(args[1], "off") == 0)
825 global.tune.options &= ~GTUNE_FD_ET;
826 else {
827 memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
828 return -1;
829 }
830 return 0;
831}
832
833/* config keyword parsers */
834static struct cfg_kw_list cfg_kws = {ILH, {
835 { CFG_GLOBAL, "tune.fd.edge-triggered", cfg_parse_tune_fd_edge_triggered },
836 { 0, NULL, NULL }
837}};
838
839INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);
840
Willy Tarreau082b6282019-05-22 14:42:12 +0200841REGISTER_PER_THREAD_ALLOC(alloc_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100842REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
843REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
Willy Tarreau082b6282019-05-22 14:42:12 +0200844REGISTER_PER_THREAD_FREE(free_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100845
Willy Tarreau2ff76222007-04-09 19:29:56 +0200846/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200847 * Local variables:
848 * c-indent-level: 8
849 * c-basic-offset: 8
850 * End:
851 */