blob: d026b15eec360c3d348a66076bfe8e37088d285c [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 Tarreau2dd0d472006-06-29 17:53:05 +020090#include <common/compat.h>
91#include <common/config.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020092
Willy Tarreau7be79a42012-11-11 15:02:54 +010093#include <types/global.h>
94
Willy Tarreau2a429502006-10-15 14:52:29 +020095#include <proto/fd.h>
Christopher Fauletd4604ad2017-05-29 10:40:41 +020096#include <proto/log.h>
Willy Tarreauc6f4ce82009-06-10 11:09:37 +020097#include <proto/port_range.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020098
99struct fdtab *fdtab = NULL; /* array of all the file descriptors */
Olivier Houchard53055052019-07-25 14:00:18 +0000100struct polled_mask *polled_mask = NULL; /* Array for the polled_mask of each fd */
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200101struct fdinfo *fdinfo = 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 Tarreau4f60f162007-04-08 16:39:58 +0200105struct poller pollers[MAX_POLLERS];
106struct poller cur_poller;
107int 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
114int poller_wr_pipe[MAX_THREADS]; // 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
117
Willy Tarreau337fb712019-12-20 07:20:00 +0100118#define _GET_NEXT(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->next
119#define _GET_PREV(fd, off) ((volatile struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->prev
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100120/* adds fd <fd> to fd list <list> if it was not yet in it */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200121void fd_add_to_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100122{
123 int next;
124 int new;
125 int old;
126 int last;
127
128redo_next:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200129 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100130 /* Check that we're not already in the cache, and if not, lock us. */
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100131 if (next > -2)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100132 goto done;
Olivier Houchardfc51f0f52019-12-19 18:33:08 +0100133 if (next == -2)
134 goto redo_next;
Olivier Houchardd3608792019-03-08 18:47:42 +0100135 if (!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100136 goto redo_next;
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100137 __ha_barrier_atomic_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100138
139 new = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100140redo_last:
141 /* First, insert in the linked list */
142 last = list->last;
143 old = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100144
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200145 _GET_PREV(fd, off) = -2;
Willy Tarreau11559a72018-02-05 17:52:24 +0100146 /* Make sure the "prev" store is visible before we update the last entry */
147 __ha_barrier_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100148
Willy Tarreau11559a72018-02-05 17:52:24 +0100149 if (unlikely(last == -1)) {
150 /* list is empty, try to add ourselves alone so that list->last=fd */
Olivier Houchardd3608792019-03-08 18:47:42 +0100151 if (unlikely(!_HA_ATOMIC_CAS(&list->last, &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100152 goto redo_last;
153
154 /* list->first was necessary -1, we're guaranteed to be alone here */
155 list->first = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100156 } else {
Willy Tarreau11559a72018-02-05 17:52:24 +0100157 /* adding ourselves past the last element
158 * The CAS will only succeed if its next is -1,
159 * which means it's in the cache, and the last element.
160 */
Olivier Houchardd3608792019-03-08 18:47:42 +0100161 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(last, off), &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100162 goto redo_last;
Willy Tarreau11559a72018-02-05 17:52:24 +0100163
164 /* Then, update the last entry */
165 list->last = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100166 }
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100167 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100168 /* since we're alone at the end of the list and still locked(-2),
169 * we know noone tried to add past us. Mark the end of list.
170 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200171 _GET_PREV(fd, off) = last;
172 _GET_NEXT(fd, off) = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100173 __ha_barrier_store();
174done:
175 return;
176}
177
178/* removes fd <fd> from fd list <list> */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200179void fd_rm_from_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100180{
181#if defined(HA_HAVE_CAS_DW) || defined(HA_CAS_IS_8B)
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100182 volatile union {
183 struct fdlist_entry ent;
184 uint64_t u64;
185 uint32_t u32[2];
186 } cur_list, next_list;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100187#endif
188 int old;
189 int new = -2;
190 int prev;
191 int next;
192 int last;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100193lock_self:
194#if (defined(HA_CAS_IS_8B) || defined(HA_HAVE_CAS_DW))
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100195 next_list.ent.next = next_list.ent.prev = -2;
196 cur_list.ent = *(volatile struct fdlist_entry *)(((char *)&fdtab[fd]) + off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100197 /* First, attempt to lock our own entries */
198 do {
199 /* The FD is not in the FD cache, give up */
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100200 if (unlikely(cur_list.ent.next <= -3))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100201 return;
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100202 if (unlikely(cur_list.ent.prev == -2 || cur_list.ent.next == -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100203 goto lock_self;
204 } while (
205#ifdef HA_CAS_IS_8B
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100206 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 +0100207#else
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100208 unlikely(!_HA_ATOMIC_DWCAS(((long *)&_GET_NEXT(fd, off)), (uint32_t *)&cur_list.u32, &next_list.u32))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100209#endif
Willy Tarreau2b9f0662020-02-25 09:25:53 +0100210 );
211 next = cur_list.ent.next;
212 prev = cur_list.ent.prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100213
214#else
215lock_self_next:
Willy Tarreau337fb712019-12-20 07:20:00 +0100216 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100217 if (next == -2)
218 goto lock_self_next;
219 if (next <= -3)
220 goto done;
Olivier Houchardd3608792019-03-08 18:47:42 +0100221 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100222 goto lock_self_next;
223lock_self_prev:
Willy Tarreau337fb712019-12-20 07:20:00 +0100224 prev = _GET_PREV(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100225 if (prev == -2)
226 goto lock_self_prev;
Olivier Houchardd3608792019-03-08 18:47:42 +0100227 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(fd, off), &prev, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100228 goto lock_self_prev;
229#endif
Olivier Houchardd2b5d162019-03-08 13:47:21 +0100230 __ha_barrier_atomic_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100231
232 /* Now, lock the entries of our neighbours */
233 if (likely(prev != -1)) {
234redo_prev:
235 old = fd;
236
Olivier Houchardd3608792019-03-08 18:47:42 +0100237 if (unlikely(!_HA_ATOMIC_CAS(&_GET_NEXT(prev, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100238 if (unlikely(old == -2)) {
239 /* Neighbour already locked, give up and
240 * retry again once he's done
241 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200242 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100243 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200244 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100245 __ha_barrier_store();
246 goto lock_self;
247 }
248 goto redo_prev;
249 }
250 }
251 if (likely(next != -1)) {
252redo_next:
253 old = fd;
Olivier Houchardd3608792019-03-08 18:47:42 +0100254 if (unlikely(!_HA_ATOMIC_CAS(&_GET_PREV(next, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100255 if (unlikely(old == -2)) {
256 /* Neighbour already locked, give up and
257 * retry again once he's done
258 */
259 if (prev != -1) {
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200260 _GET_NEXT(prev, off) = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100261 __ha_barrier_store();
262 }
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200263 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100264 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200265 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100266 __ha_barrier_store();
267 goto lock_self;
268 }
269 goto redo_next;
270 }
271 }
272 if (list->first == fd)
273 list->first = next;
274 __ha_barrier_store();
275 last = list->last;
Olivier Houchardd3608792019-03-08 18:47:42 +0100276 while (unlikely(last == fd && (!_HA_ATOMIC_CAS(&list->last, &last, prev))))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100277 __ha_compiler_barrier();
278 /* Make sure we let other threads know we're no longer in cache,
279 * before releasing our neighbours.
280 */
281 __ha_barrier_store();
282 if (likely(prev != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200283 _GET_NEXT(prev, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100284 __ha_barrier_store();
285 if (likely(next != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200286 _GET_PREV(next, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100287 __ha_barrier_store();
288 /* Ok, now we're out of the fd cache */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200289 _GET_NEXT(fd, off) = -(next + 4);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100290 __ha_barrier_store();
291done:
292 return;
293}
294
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200295#undef _GET_NEXT
296#undef _GET_PREV
297
Willy Tarreau173d9952018-01-26 21:48:23 +0100298/* Deletes an FD from the fdsets.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200299 * The file descriptor is also closed.
300 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200301static void fd_dodelete(int fd, int do_close)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200302{
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100303 int locked = fdtab[fd].running_mask != tid_bit;
Willy Tarreau87d54a92018-10-15 09:44:46 +0200304
Olivier Houchard88516642020-03-05 18:10:51 +0100305 /* We're just trying to protect against a concurrent fd_insert()
306 * here, not against fd_takeother(), because either we're called
307 * directly from the iocb(), and we're already locked, or we're
308 * called from the mux tasklet, but then the mux is responsible for
309 * making sure the tasklet does nothing, and the connection is never
310 * destroyed.
311 */
Willy Tarreau87d54a92018-10-15 09:44:46 +0200312 if (locked)
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100313 fd_set_running_excl(fd);
314
Willy Tarreauad38ace2013-12-15 14:19:38 +0100315 if (fdtab[fd].linger_risk) {
316 /* this is generally set when connecting to servers */
317 setsockopt(fd, SOL_SOCKET, SO_LINGER,
318 (struct linger *) &nolinger, sizeof(struct linger));
319 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100320 if (cur_poller.clo)
321 cur_poller.clo(fd);
Olivier Houchardc22580c2019-08-05 18:51:52 +0200322 polled_mask[fd].poll_recv = polled_mask[fd].poll_send = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100323
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100324 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100325
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200326 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
327 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200328 fdtab[fd].owner = NULL;
Willy Tarreauf65610a2017-10-31 16:06:06 +0100329 fdtab[fd].thread_mask = 0;
Willy Tarreauc9c83782018-01-17 18:44:46 +0100330 if (do_close) {
Christopher Fauletd531f882017-06-01 16:55:03 +0200331 close(fd);
Olivier Houchard7c49d2e2019-04-16 18:37:05 +0200332 _HA_ATOMIC_SUB(&ha_used_fds, 1);
Willy Tarreauc9c83782018-01-17 18:44:46 +0100333 }
Willy Tarreau87d54a92018-10-15 09:44:46 +0200334 if (locked)
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100335 fd_clr_running(fd);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200336}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200337
Olivier Houchard88516642020-03-05 18:10:51 +0100338#ifndef HA_HAVE_CAS_DW
Olivier Houcharde4ba0d42020-03-20 11:03:38 +0100339__decl_hathreads(__decl_rwlock(fd_mig_lock));
Olivier Houchard88516642020-03-05 18:10:51 +0100340#endif
341
342/*
343 * Take over a FD belonging to another thread.
344 * unexpected_conn is the expected owner of the fd.
345 * Returns 0 on success, and -1 on failure.
346 */
347int fd_takeover(int fd, void *expected_owner)
348{
349#ifndef HA_HAVE_CAS_DW
350 int ret;
351
352 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
353 _HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
354 if (fdtab[fd].running_mask != tid_bit || fdtab[fd].owner != expected_owner) {
355 ret = -1;
356 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
357 goto end;
358 }
359 fdtab[fd].thread_mask = tid_bit;
360 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
361 ret = 0;
362end:
363 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
364 /* Make sure the FD doesn't have the active bit. It is possible that
365 * the fd is polled by the thread that used to own it, the new thread
366 * is supposed to call subscribe() later, to activate polling.
367 */
368 fd_stop_recv(fd);
369 return ret;
370#else
371 unsigned long old_masks[2];
372 unsigned long new_masks[2];
373
374 old_masks[0] = tid_bit;
375 old_masks[1] = fdtab[fd].thread_mask;
376 new_masks[0] = new_masks[1] = tid_bit;
377 /* protect ourself against a delete then an insert for the same fd,
378 * if it happens, then the owner will no longer be the expected
379 * connection.
380 */
381 _HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
382 if (fdtab[fd].owner != expected_owner) {
383 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
384 return -1;
385 }
386 do {
387 if (old_masks[0] != tid_bit || !old_masks[1]) {
388 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
389 return -1;
390 }
391 } while (!(_HA_ATOMIC_DWCAS(&fdtab[fd].running_mask, &old_masks,
392 &new_masks)));
393 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
394 return 0;
395#endif /* HW_HAVE_CAS_DW */
396}
397
Willy Tarreau173d9952018-01-26 21:48:23 +0100398/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200399 * The file descriptor is also closed.
400 */
401void fd_delete(int fd)
402{
403 fd_dodelete(fd, 1);
404}
405
Willy Tarreau173d9952018-01-26 21:48:23 +0100406/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200407 * The file descriptor is kept open.
408 */
409void fd_remove(int fd)
410{
411 fd_dodelete(fd, 0);
412}
413
Willy Tarreaudbe30602019-09-04 13:25:41 +0200414void updt_fd_polling(const int fd)
415{
416 if ((fdtab[fd].thread_mask & all_threads_mask) == tid_bit) {
417
418 /* note: we don't have a test-and-set yet in hathreads */
419
420 if (HA_ATOMIC_BTS(&fdtab[fd].update_mask, tid))
421 return;
422
423 fd_updt[fd_nbupdt++] = fd;
424 } else {
425 unsigned long update_mask = fdtab[fd].update_mask;
426 do {
427 if (update_mask == fdtab[fd].thread_mask)
428 return;
429 } while (!_HA_ATOMIC_CAS(&fdtab[fd].update_mask, &update_mask,
430 fdtab[fd].thread_mask));
431 fd_add_to_fd_list(&update_list, fd, offsetof(struct fdtab, update));
432 }
433}
434
Olivier Houchard7fa55622020-02-27 17:25:43 +0100435__decl_spinlock(log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200436/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
437 * optionally followed by a newline if <nl> is non-null, to file descriptor
438 * <fd>. The message is sent atomically using writev(). It may be truncated to
439 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
440 * two lists, it's just a convenience to help the caller prepend some prefixes
441 * when necessary. It takes the fd's lock to make sure no other thread will
442 * write to the same fd in parallel. Returns the number of bytes sent, or <=0
443 * on failure. A limit to 31 total non-empty segments is enforced. The caller
444 * is responsible for taking care of making the fd non-blocking.
445 */
446ssize_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)
447{
448 struct iovec iovec[32];
449 size_t totlen = 0;
450 size_t sent = 0;
451 int vec = 0;
452
453 if (!maxlen)
454 maxlen = ~0;
455
456 /* keep one char for a possible trailing '\n' in any case */
457 maxlen--;
458
459 /* make an iovec from the concatenation of all parts of the original
460 * message. Skip empty fields and truncate the whole message to maxlen,
461 * leaving one spare iovec for the '\n'.
462 */
463 while (vec < (sizeof(iovec) / sizeof(iovec[0]) - 1)) {
464 if (!npfx) {
465 pfx = msg;
466 npfx = nmsg;
467 nmsg = 0;
468 if (!npfx)
469 break;
470 }
471
472 iovec[vec].iov_base = pfx->ptr;
473 iovec[vec].iov_len = MIN(maxlen, pfx->len);
474 maxlen -= iovec[vec].iov_len;
475 totlen += iovec[vec].iov_len;
476 if (iovec[vec].iov_len)
477 vec++;
478 pfx++; npfx--;
479 };
480
481 if (nl) {
482 iovec[vec].iov_base = "\n";
483 iovec[vec].iov_len = 1;
484 vec++;
485 }
486
Willy Tarreau7e9776a2019-08-30 14:41:47 +0200487 if (unlikely(!fdtab[fd].initialized)) {
488 fdtab[fd].initialized = 1;
489 if (!isatty(fd))
490 fcntl(fd, F_SETFL, O_NONBLOCK);
491 }
492
Olivier Houchard7fa55622020-02-27 17:25:43 +0100493 HA_SPIN_LOCK(OTHER_LOCK, &log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200494 sent = writev(fd, iovec, vec);
Olivier Houchard7fa55622020-02-27 17:25:43 +0100495 HA_SPIN_UNLOCK(OTHER_LOCK, &log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200496
497 /* sent > 0 if the message was delivered */
498 return sent;
499}
500
Olivier Houchard2292edf2019-02-25 14:26:54 +0100501#if defined(USE_CLOSEFROM)
502void my_closefrom(int start)
503{
504 closefrom(start);
505}
506
Willy Tarreaue5733232019-05-22 19:24:06 +0200507#elif defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100508/* This is a portable implementation of closefrom(). It closes all open file
509 * descriptors starting at <start> and above. It relies on the fact that poll()
510 * will return POLLNVAL for each invalid (hence close) file descriptor passed
511 * in argument in order to skip them. It acts with batches of FDs and will
512 * typically perform one poll() call per 1024 FDs so the overhead is low in
513 * case all FDs have to be closed.
514 */
515void my_closefrom(int start)
516{
517 struct pollfd poll_events[1024];
518 struct rlimit limit;
519 int nbfds, fd, ret, idx;
520 int step, next;
521
522 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
523 step = nbfds = limit.rlim_cur;
524 else
525 step = nbfds = 0;
526
527 if (nbfds <= 0) {
528 /* set safe limit */
529 nbfds = 1024;
530 step = 256;
531 }
532
533 if (step > sizeof(poll_events) / sizeof(poll_events[0]))
534 step = sizeof(poll_events) / sizeof(poll_events[0]);
535
536 while (start < nbfds) {
537 next = (start / step + 1) * step;
538
539 for (fd = start; fd < next && fd < nbfds; fd++) {
540 poll_events[fd - start].fd = fd;
541 poll_events[fd - start].events = 0;
542 }
543
544 do {
545 ret = poll(poll_events, fd - start, 0);
546 if (ret >= 0)
547 break;
548 } while (errno == EAGAIN || errno == EINTR || errno == ENOMEM);
549
Willy Tarreaub8e602c2019-02-22 09:07:42 +0100550 if (ret)
551 ret = fd - start;
552
Willy Tarreau9188ac62019-02-21 22:12:47 +0100553 for (idx = 0; idx < ret; idx++) {
554 if (poll_events[idx].revents & POLLNVAL)
555 continue; /* already closed */
556
557 fd = poll_events[idx].fd;
558 close(fd);
559 }
560 start = next;
561 }
562}
563
Willy Tarreaue5733232019-05-22 19:24:06 +0200564#else // defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100565
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100566/* This is a portable implementation of closefrom(). It closes all open file
567 * descriptors starting at <start> and above. This is a naive version for use
568 * when the operating system provides no alternative.
569 */
570void my_closefrom(int start)
571{
572 struct rlimit limit;
573 int nbfds;
574
575 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
576 nbfds = limit.rlim_cur;
577 else
578 nbfds = 0;
579
580 if (nbfds <= 0)
581 nbfds = 1024; /* safe limit */
582
583 while (start < nbfds)
584 close(start++);
585}
Willy Tarreaue5733232019-05-22 19:24:06 +0200586#endif // defined(USE_POLL)
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100587
Willy Tarreau4f60f162007-04-08 16:39:58 +0200588/* disable the specified poller */
589void disable_poller(const char *poller_name)
590{
591 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200592
Willy Tarreau4f60f162007-04-08 16:39:58 +0200593 for (p = 0; p < nbpollers; p++)
594 if (strcmp(pollers[p].name, poller_name) == 0)
595 pollers[p].pref = 0;
596}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200597
Olivier Houchard79321b92018-07-26 17:55:11 +0200598void poller_pipe_io_handler(int fd)
599{
600 char buf[1024];
601 /* Flush the pipe */
602 while (read(fd, buf, sizeof(buf)) > 0);
603 fd_cant_recv(fd);
604}
605
Willy Tarreau082b6282019-05-22 14:42:12 +0200606/* allocate the per-thread fd_updt thus needs to be called early after
607 * thread creation.
608 */
609static int alloc_pollers_per_thread()
610{
611 fd_updt = calloc(global.maxsock, sizeof(*fd_updt));
612 return fd_updt != NULL;
613}
614
615/* Initialize the pollers per thread.*/
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200616static int init_pollers_per_thread()
617{
Olivier Houchard79321b92018-07-26 17:55:11 +0200618 int mypipe[2];
Willy Tarreau082b6282019-05-22 14:42:12 +0200619
620 if (pipe(mypipe) < 0)
Olivier Houchard79321b92018-07-26 17:55:11 +0200621 return 0;
Willy Tarreau082b6282019-05-22 14:42:12 +0200622
Olivier Houchard79321b92018-07-26 17:55:11 +0200623 poller_rd_pipe = mypipe[0];
624 poller_wr_pipe[tid] = mypipe[1];
625 fcntl(poller_rd_pipe, F_SETFL, O_NONBLOCK);
626 fd_insert(poller_rd_pipe, poller_pipe_io_handler, poller_pipe_io_handler,
627 tid_bit);
628 fd_want_recv(poller_rd_pipe);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200629 return 1;
630}
631
632/* Deinitialize the pollers per thread */
633static void deinit_pollers_per_thread()
634{
William Lallemand808e1b72018-12-15 22:34:31 +0100635 /* rd and wr are init at the same place, but only rd is init to -1, so
636 we rely to rd to close. */
637 if (poller_rd_pipe > -1) {
638 close(poller_rd_pipe);
639 poller_rd_pipe = -1;
640 close(poller_wr_pipe[tid]);
641 poller_wr_pipe[tid] = -1;
642 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200643}
644
Willy Tarreau082b6282019-05-22 14:42:12 +0200645/* Release the pollers per thread, to be called late */
646static void free_pollers_per_thread()
647{
648 free(fd_updt);
649 fd_updt = NULL;
650}
651
Willy Tarreaubaaee002006-06-26 02:48:02 +0200652/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200653 * Initialize the pollers till the best one is found.
654 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200655 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200656int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200657{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200658 int p;
659 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200660
Christopher Faulet63fe6522017-08-31 17:52:09 +0200661 if ((fdtab = calloc(global.maxsock, sizeof(struct fdtab))) == NULL)
662 goto fail_tab;
663
Olivier Houchard53055052019-07-25 14:00:18 +0000664 if ((polled_mask = calloc(global.maxsock, sizeof(*polled_mask))) == NULL)
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200665 goto fail_polledmask;
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000666
Christopher Faulet63fe6522017-08-31 17:52:09 +0200667 if ((fdinfo = calloc(global.maxsock, sizeof(struct fdinfo))) == NULL)
668 goto fail_info;
669
Olivier Houchard6b96f722018-04-25 16:58:25 +0200670 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200671
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100672 for (p = 0; p < global.maxsock; p++) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100673 /* Mark the fd as out of the fd cache */
Olivier Houchard6b96f722018-04-25 16:58:25 +0200674 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100675 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200676
Willy Tarreau4f60f162007-04-08 16:39:58 +0200677 do {
678 bp = NULL;
679 for (p = 0; p < nbpollers; p++)
680 if (!bp || (pollers[p].pref > bp->pref))
681 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200682
Willy Tarreau4f60f162007-04-08 16:39:58 +0200683 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200684 break;
685
Willy Tarreau4f60f162007-04-08 16:39:58 +0200686 if (bp->init(bp)) {
687 memcpy(&cur_poller, bp, sizeof(*bp));
688 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200689 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200690 } while (!bp || bp->pref == 0);
Willy Tarreau7be79a42012-11-11 15:02:54 +0100691
Christopher Faulet63fe6522017-08-31 17:52:09 +0200692 free(fdinfo);
693 fail_info:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200694 free(polled_mask);
695 fail_polledmask:
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000696 free(fdtab);
697 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100698 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200699}
700
Willy Tarreaubaaee002006-06-26 02:48:02 +0200701/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200702 * Deinitialize the pollers.
703 */
704void deinit_pollers() {
705
706 struct poller *bp;
707 int p;
708
709 for (p = 0; p < nbpollers; p++) {
710 bp = &pollers[p];
711
712 if (bp && bp->pref)
713 bp->term(bp);
714 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200715
Christopher Faulet63fe6522017-08-31 17:52:09 +0200716 free(fdinfo); fdinfo = NULL;
717 free(fdtab); fdtab = NULL;
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200718 free(polled_mask); polled_mask = NULL;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200719}
720
721/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200722 * Lists the known pollers on <out>.
723 * Should be performed only before initialization.
724 */
725int list_pollers(FILE *out)
726{
727 int p;
728 int last, next;
729 int usable;
730 struct poller *bp;
731
732 fprintf(out, "Available polling systems :\n");
733
734 usable = 0;
735 bp = NULL;
736 last = next = -1;
737 while (1) {
738 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200739 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100740 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200741 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100742 if (!bp || (pollers[p].pref > bp->pref))
743 bp = &pollers[p];
744 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200745 }
746
747 if (next == -1)
748 break;
749
750 for (p = 0; p < nbpollers; p++) {
751 if (pollers[p].pref == next) {
752 fprintf(out, " %10s : ", pollers[p].name);
753 if (pollers[p].pref == 0)
754 fprintf(out, "disabled, ");
755 else
756 fprintf(out, "pref=%3d, ", pollers[p].pref);
757 if (pollers[p].test(&pollers[p])) {
758 fprintf(out, " test result OK");
759 if (next > 0)
760 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100761 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200762 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100763 if (bp == &pollers[p])
764 bp = NULL;
765 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200766 fprintf(out, "\n");
767 }
768 }
769 last = next;
770 next = -1;
771 };
772 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
773 return 0;
774}
775
776/*
777 * Some pollers may lose their connection after a fork(). It may be necessary
778 * to create initialize part of them again. Returns 0 in case of failure,
779 * otherwise 1. The fork() function may be NULL if unused. In case of error,
780 * the the current poller is destroyed and the caller is responsible for trying
781 * another one by calling init_pollers() again.
782 */
783int fork_poller()
784{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200785 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100786 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200787 if (fdtab[fd].owner) {
788 fdtab[fd].cloned = 1;
789 }
790 }
791
Willy Tarreau2ff76222007-04-09 19:29:56 +0200792 if (cur_poller.fork) {
793 if (cur_poller.fork(&cur_poller))
794 return 1;
795 cur_poller.term(&cur_poller);
796 return 0;
797 }
798 return 1;
799}
800
Willy Tarreau082b6282019-05-22 14:42:12 +0200801REGISTER_PER_THREAD_ALLOC(alloc_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100802REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
803REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
Willy Tarreau082b6282019-05-22 14:42:12 +0200804REGISTER_PER_THREAD_FREE(free_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100805
Willy Tarreau2ff76222007-04-09 19:29:56 +0200806/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200807 * Local variables:
808 * c-indent-level: 8
809 * c-basic-offset: 8
810 * End:
811 */