blob: c7bc54efafa661c0305dd7c62a09ad1be7487f14 [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 Tarreaubaaee002006-06-26 02:48:02 +020090
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020091#include <haproxy/api.h>
Willy Tarreau7be79a42012-11-11 15:02:54 +010092#include <types/global.h>
93
Willy Tarreau2a429502006-10-15 14:52:29 +020094#include <proto/fd.h>
Christopher Fauletd4604ad2017-05-29 10:40:41 +020095#include <proto/log.h>
Willy Tarreauc6f4ce82009-06-10 11:09:37 +020096#include <proto/port_range.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +020097
98struct 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 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200300static void fd_dodelete(int fd, int do_close)
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()
305 * here, not against fd_takeother(), because either we're called
306 * 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 Tarreau8d5d77e2009-10-18 07:25:52 +0200325 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
326 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200327 fdtab[fd].owner = NULL;
Willy Tarreauf65610a2017-10-31 16:06:06 +0100328 fdtab[fd].thread_mask = 0;
Willy Tarreauc9c83782018-01-17 18:44:46 +0100329 if (do_close) {
Christopher Fauletd531f882017-06-01 16:55:03 +0200330 close(fd);
Olivier Houchard7c49d2e2019-04-16 18:37:05 +0200331 _HA_ATOMIC_SUB(&ha_used_fds, 1);
Willy Tarreauc9c83782018-01-17 18:44:46 +0100332 }
Willy Tarreau87d54a92018-10-15 09:44:46 +0200333 if (locked)
Olivier Houcharda7bf5732020-02-27 17:26:13 +0100334 fd_clr_running(fd);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200335}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200336
Olivier Houchard88516642020-03-05 18:10:51 +0100337#ifndef HA_HAVE_CAS_DW
Willy Tarreauaf613e82020-06-05 08:40:51 +0200338__decl_thread(__decl_rwlock(fd_mig_lock));
Olivier Houchard88516642020-03-05 18:10:51 +0100339#endif
340
341/*
342 * Take over a FD belonging to another thread.
343 * unexpected_conn is the expected owner of the fd.
344 * Returns 0 on success, and -1 on failure.
345 */
346int fd_takeover(int fd, void *expected_owner)
347{
348#ifndef HA_HAVE_CAS_DW
349 int ret;
350
351 HA_RWLOCK_WRLOCK(OTHER_LOCK, &fd_mig_lock);
352 _HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
353 if (fdtab[fd].running_mask != tid_bit || fdtab[fd].owner != expected_owner) {
354 ret = -1;
355 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
356 goto end;
357 }
358 fdtab[fd].thread_mask = tid_bit;
359 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
360 ret = 0;
361end:
362 HA_RWLOCK_WRUNLOCK(OTHER_LOCK, &fd_mig_lock);
363 /* Make sure the FD doesn't have the active bit. It is possible that
364 * the fd is polled by the thread that used to own it, the new thread
365 * is supposed to call subscribe() later, to activate polling.
366 */
367 fd_stop_recv(fd);
368 return ret;
369#else
370 unsigned long old_masks[2];
371 unsigned long new_masks[2];
372
373 old_masks[0] = tid_bit;
374 old_masks[1] = fdtab[fd].thread_mask;
375 new_masks[0] = new_masks[1] = tid_bit;
376 /* protect ourself against a delete then an insert for the same fd,
377 * if it happens, then the owner will no longer be the expected
378 * connection.
379 */
380 _HA_ATOMIC_OR(&fdtab[fd].running_mask, tid_bit);
381 if (fdtab[fd].owner != expected_owner) {
382 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
383 return -1;
384 }
385 do {
386 if (old_masks[0] != tid_bit || !old_masks[1]) {
387 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
388 return -1;
389 }
390 } while (!(_HA_ATOMIC_DWCAS(&fdtab[fd].running_mask, &old_masks,
391 &new_masks)));
392 _HA_ATOMIC_AND(&fdtab[fd].running_mask, ~tid_bit);
393 return 0;
394#endif /* HW_HAVE_CAS_DW */
395}
396
Willy Tarreau173d9952018-01-26 21:48:23 +0100397/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200398 * The file descriptor is also closed.
399 */
400void fd_delete(int fd)
401{
402 fd_dodelete(fd, 1);
403}
404
Willy Tarreau173d9952018-01-26 21:48:23 +0100405/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200406 * The file descriptor is kept open.
407 */
408void fd_remove(int fd)
409{
410 fd_dodelete(fd, 0);
411}
412
Willy Tarreaudbe30602019-09-04 13:25:41 +0200413void updt_fd_polling(const int fd)
414{
415 if ((fdtab[fd].thread_mask & all_threads_mask) == tid_bit) {
416
417 /* note: we don't have a test-and-set yet in hathreads */
418
419 if (HA_ATOMIC_BTS(&fdtab[fd].update_mask, tid))
420 return;
421
422 fd_updt[fd_nbupdt++] = fd;
423 } else {
424 unsigned long update_mask = fdtab[fd].update_mask;
425 do {
426 if (update_mask == fdtab[fd].thread_mask)
427 return;
428 } while (!_HA_ATOMIC_CAS(&fdtab[fd].update_mask, &update_mask,
429 fdtab[fd].thread_mask));
430 fd_add_to_fd_list(&update_list, fd, offsetof(struct fdtab, update));
431 }
432}
433
Olivier Houchard7fa55622020-02-27 17:25:43 +0100434__decl_spinlock(log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200435/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
436 * optionally followed by a newline if <nl> is non-null, to file descriptor
437 * <fd>. The message is sent atomically using writev(). It may be truncated to
438 * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
439 * two lists, it's just a convenience to help the caller prepend some prefixes
440 * when necessary. It takes the fd's lock to make sure no other thread will
441 * write to the same fd in parallel. Returns the number of bytes sent, or <=0
442 * on failure. A limit to 31 total non-empty segments is enforced. The caller
443 * is responsible for taking care of making the fd non-blocking.
444 */
445ssize_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)
446{
447 struct iovec iovec[32];
448 size_t totlen = 0;
449 size_t sent = 0;
450 int vec = 0;
451
452 if (!maxlen)
453 maxlen = ~0;
454
455 /* keep one char for a possible trailing '\n' in any case */
456 maxlen--;
457
458 /* make an iovec from the concatenation of all parts of the original
459 * message. Skip empty fields and truncate the whole message to maxlen,
460 * leaving one spare iovec for the '\n'.
461 */
462 while (vec < (sizeof(iovec) / sizeof(iovec[0]) - 1)) {
463 if (!npfx) {
464 pfx = msg;
465 npfx = nmsg;
466 nmsg = 0;
467 if (!npfx)
468 break;
469 }
470
471 iovec[vec].iov_base = pfx->ptr;
472 iovec[vec].iov_len = MIN(maxlen, pfx->len);
473 maxlen -= iovec[vec].iov_len;
474 totlen += iovec[vec].iov_len;
475 if (iovec[vec].iov_len)
476 vec++;
477 pfx++; npfx--;
478 };
479
480 if (nl) {
481 iovec[vec].iov_base = "\n";
482 iovec[vec].iov_len = 1;
483 vec++;
484 }
485
Willy Tarreau7e9776a2019-08-30 14:41:47 +0200486 if (unlikely(!fdtab[fd].initialized)) {
487 fdtab[fd].initialized = 1;
488 if (!isatty(fd))
489 fcntl(fd, F_SETFL, O_NONBLOCK);
490 }
491
Olivier Houchard7fa55622020-02-27 17:25:43 +0100492 HA_SPIN_LOCK(OTHER_LOCK, &log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200493 sent = writev(fd, iovec, vec);
Olivier Houchard7fa55622020-02-27 17:25:43 +0100494 HA_SPIN_UNLOCK(OTHER_LOCK, &log_lock);
Willy Tarreau931d8b72019-08-27 11:08:17 +0200495
496 /* sent > 0 if the message was delivered */
497 return sent;
498}
499
Olivier Houchard2292edf2019-02-25 14:26:54 +0100500#if defined(USE_CLOSEFROM)
501void my_closefrom(int start)
502{
503 closefrom(start);
504}
505
Willy Tarreaue5733232019-05-22 19:24:06 +0200506#elif defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100507/* This is a portable implementation of closefrom(). It closes all open file
508 * descriptors starting at <start> and above. It relies on the fact that poll()
509 * will return POLLNVAL for each invalid (hence close) file descriptor passed
510 * in argument in order to skip them. It acts with batches of FDs and will
511 * typically perform one poll() call per 1024 FDs so the overhead is low in
512 * case all FDs have to be closed.
513 */
514void my_closefrom(int start)
515{
516 struct pollfd poll_events[1024];
517 struct rlimit limit;
518 int nbfds, fd, ret, idx;
519 int step, next;
520
521 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
522 step = nbfds = limit.rlim_cur;
523 else
524 step = nbfds = 0;
525
526 if (nbfds <= 0) {
527 /* set safe limit */
528 nbfds = 1024;
529 step = 256;
530 }
531
532 if (step > sizeof(poll_events) / sizeof(poll_events[0]))
533 step = sizeof(poll_events) / sizeof(poll_events[0]);
534
535 while (start < nbfds) {
536 next = (start / step + 1) * step;
537
538 for (fd = start; fd < next && fd < nbfds; fd++) {
539 poll_events[fd - start].fd = fd;
540 poll_events[fd - start].events = 0;
541 }
542
543 do {
544 ret = poll(poll_events, fd - start, 0);
545 if (ret >= 0)
546 break;
547 } while (errno == EAGAIN || errno == EINTR || errno == ENOMEM);
548
Willy Tarreaub8e602c2019-02-22 09:07:42 +0100549 if (ret)
550 ret = fd - start;
551
Willy Tarreau9188ac62019-02-21 22:12:47 +0100552 for (idx = 0; idx < ret; idx++) {
553 if (poll_events[idx].revents & POLLNVAL)
554 continue; /* already closed */
555
556 fd = poll_events[idx].fd;
557 close(fd);
558 }
559 start = next;
560 }
561}
562
Willy Tarreaue5733232019-05-22 19:24:06 +0200563#else // defined(USE_POLL)
Willy Tarreau9188ac62019-02-21 22:12:47 +0100564
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100565/* This is a portable implementation of closefrom(). It closes all open file
566 * descriptors starting at <start> and above. This is a naive version for use
567 * when the operating system provides no alternative.
568 */
569void my_closefrom(int start)
570{
571 struct rlimit limit;
572 int nbfds;
573
574 if (getrlimit(RLIMIT_NOFILE, &limit) == 0)
575 nbfds = limit.rlim_cur;
576 else
577 nbfds = 0;
578
579 if (nbfds <= 0)
580 nbfds = 1024; /* safe limit */
581
582 while (start < nbfds)
583 close(start++);
584}
Willy Tarreaue5733232019-05-22 19:24:06 +0200585#endif // defined(USE_POLL)
Willy Tarreau2d7f81b2019-02-21 22:19:17 +0100586
Willy Tarreau4f60f162007-04-08 16:39:58 +0200587/* disable the specified poller */
588void disable_poller(const char *poller_name)
589{
590 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200591
Willy Tarreau4f60f162007-04-08 16:39:58 +0200592 for (p = 0; p < nbpollers; p++)
593 if (strcmp(pollers[p].name, poller_name) == 0)
594 pollers[p].pref = 0;
595}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200596
Olivier Houchard79321b92018-07-26 17:55:11 +0200597void poller_pipe_io_handler(int fd)
598{
599 char buf[1024];
600 /* Flush the pipe */
601 while (read(fd, buf, sizeof(buf)) > 0);
602 fd_cant_recv(fd);
603}
604
Willy Tarreau082b6282019-05-22 14:42:12 +0200605/* allocate the per-thread fd_updt thus needs to be called early after
606 * thread creation.
607 */
608static int alloc_pollers_per_thread()
609{
610 fd_updt = calloc(global.maxsock, sizeof(*fd_updt));
611 return fd_updt != NULL;
612}
613
614/* Initialize the pollers per thread.*/
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200615static int init_pollers_per_thread()
616{
Olivier Houchard79321b92018-07-26 17:55:11 +0200617 int mypipe[2];
Willy Tarreau082b6282019-05-22 14:42:12 +0200618
619 if (pipe(mypipe) < 0)
Olivier Houchard79321b92018-07-26 17:55:11 +0200620 return 0;
Willy Tarreau082b6282019-05-22 14:42:12 +0200621
Olivier Houchard79321b92018-07-26 17:55:11 +0200622 poller_rd_pipe = mypipe[0];
623 poller_wr_pipe[tid] = mypipe[1];
624 fcntl(poller_rd_pipe, F_SETFL, O_NONBLOCK);
625 fd_insert(poller_rd_pipe, poller_pipe_io_handler, poller_pipe_io_handler,
626 tid_bit);
627 fd_want_recv(poller_rd_pipe);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200628 return 1;
629}
630
631/* Deinitialize the pollers per thread */
632static void deinit_pollers_per_thread()
633{
William Lallemand808e1b72018-12-15 22:34:31 +0100634 /* rd and wr are init at the same place, but only rd is init to -1, so
635 we rely to rd to close. */
636 if (poller_rd_pipe > -1) {
637 close(poller_rd_pipe);
638 poller_rd_pipe = -1;
639 close(poller_wr_pipe[tid]);
640 poller_wr_pipe[tid] = -1;
641 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200642}
643
Willy Tarreau082b6282019-05-22 14:42:12 +0200644/* Release the pollers per thread, to be called late */
645static void free_pollers_per_thread()
646{
647 free(fd_updt);
648 fd_updt = NULL;
649}
650
Willy Tarreaubaaee002006-06-26 02:48:02 +0200651/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200652 * Initialize the pollers till the best one is found.
653 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200654 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200655int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200656{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200657 int p;
658 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200659
Christopher Faulet63fe6522017-08-31 17:52:09 +0200660 if ((fdtab = calloc(global.maxsock, sizeof(struct fdtab))) == NULL)
661 goto fail_tab;
662
Olivier Houchard53055052019-07-25 14:00:18 +0000663 if ((polled_mask = calloc(global.maxsock, sizeof(*polled_mask))) == NULL)
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200664 goto fail_polledmask;
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000665
Christopher Faulet63fe6522017-08-31 17:52:09 +0200666 if ((fdinfo = calloc(global.maxsock, sizeof(struct fdinfo))) == NULL)
667 goto fail_info;
668
Olivier Houchard6b96f722018-04-25 16:58:25 +0200669 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200670
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100671 for (p = 0; p < global.maxsock; p++) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100672 /* Mark the fd as out of the fd cache */
Olivier Houchard6b96f722018-04-25 16:58:25 +0200673 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100674 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200675
Willy Tarreau4f60f162007-04-08 16:39:58 +0200676 do {
677 bp = NULL;
678 for (p = 0; p < nbpollers; p++)
679 if (!bp || (pollers[p].pref > bp->pref))
680 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200681
Willy Tarreau4f60f162007-04-08 16:39:58 +0200682 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200683 break;
684
Willy Tarreau4f60f162007-04-08 16:39:58 +0200685 if (bp->init(bp)) {
686 memcpy(&cur_poller, bp, sizeof(*bp));
687 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200688 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200689 } while (!bp || bp->pref == 0);
Willy Tarreau7be79a42012-11-11 15:02:54 +0100690
Christopher Faulet63fe6522017-08-31 17:52:09 +0200691 free(fdinfo);
692 fail_info:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200693 free(polled_mask);
694 fail_polledmask:
Uman Shahzadda7eeed2019-01-17 08:21:39 +0000695 free(fdtab);
696 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100697 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200698}
699
Willy Tarreaubaaee002006-06-26 02:48:02 +0200700/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200701 * Deinitialize the pollers.
702 */
703void deinit_pollers() {
704
705 struct poller *bp;
706 int p;
707
708 for (p = 0; p < nbpollers; p++) {
709 bp = &pollers[p];
710
711 if (bp && bp->pref)
712 bp->term(bp);
713 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200714
Christopher Faulet63fe6522017-08-31 17:52:09 +0200715 free(fdinfo); fdinfo = NULL;
716 free(fdtab); fdtab = NULL;
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200717 free(polled_mask); polled_mask = NULL;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200718}
719
720/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200721 * Lists the known pollers on <out>.
722 * Should be performed only before initialization.
723 */
724int list_pollers(FILE *out)
725{
726 int p;
727 int last, next;
728 int usable;
729 struct poller *bp;
730
731 fprintf(out, "Available polling systems :\n");
732
733 usable = 0;
734 bp = NULL;
735 last = next = -1;
736 while (1) {
737 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200738 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100739 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200740 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100741 if (!bp || (pollers[p].pref > bp->pref))
742 bp = &pollers[p];
743 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200744 }
745
746 if (next == -1)
747 break;
748
749 for (p = 0; p < nbpollers; p++) {
750 if (pollers[p].pref == next) {
751 fprintf(out, " %10s : ", pollers[p].name);
752 if (pollers[p].pref == 0)
753 fprintf(out, "disabled, ");
754 else
755 fprintf(out, "pref=%3d, ", pollers[p].pref);
756 if (pollers[p].test(&pollers[p])) {
757 fprintf(out, " test result OK");
758 if (next > 0)
759 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100760 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200761 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100762 if (bp == &pollers[p])
763 bp = NULL;
764 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200765 fprintf(out, "\n");
766 }
767 }
768 last = next;
769 next = -1;
770 };
771 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
772 return 0;
773}
774
775/*
776 * Some pollers may lose their connection after a fork(). It may be necessary
777 * to create initialize part of them again. Returns 0 in case of failure,
778 * otherwise 1. The fork() function may be NULL if unused. In case of error,
779 * the the current poller is destroyed and the caller is responsible for trying
780 * another one by calling init_pollers() again.
781 */
782int fork_poller()
783{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200784 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100785 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200786 if (fdtab[fd].owner) {
787 fdtab[fd].cloned = 1;
788 }
789 }
790
Willy Tarreau2ff76222007-04-09 19:29:56 +0200791 if (cur_poller.fork) {
792 if (cur_poller.fork(&cur_poller))
793 return 1;
794 cur_poller.term(&cur_poller);
795 return 0;
796 }
797 return 1;
798}
799
Willy Tarreau082b6282019-05-22 14:42:12 +0200800REGISTER_PER_THREAD_ALLOC(alloc_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100801REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
802REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
Willy Tarreau082b6282019-05-22 14:42:12 +0200803REGISTER_PER_THREAD_FREE(free_pollers_per_thread);
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100804
Willy Tarreau2ff76222007-04-09 19:29:56 +0200805/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200806 * Local variables:
807 * c-indent-level: 8
808 * c-basic-offset: 8
809 * End:
810 */