blob: d6c140c0ce057778b1d36ab0de00a836d4b921bb [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 Tarreauf817e9f2014-01-10 16:58:45 +010011 * This code implements an events cache for file descriptors. It remembers the
12 * readiness of a file descriptor after a return from poll() and the fact that
13 * an I/O attempt failed on EAGAIN. Events in the cache which are still marked
14 * ready and active are processed just as if they were reported by poll().
Willy Tarreau7be79a42012-11-11 15:02:54 +010015 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010016 * This serves multiple purposes. First, it significantly improves performance
17 * by avoiding to subscribe to polling unless absolutely necessary, so most
18 * events are processed without polling at all, especially send() which
19 * benefits from the socket buffers. Second, it is the only way to support
20 * edge-triggered pollers (eg: EPOLL_ET). And third, it enables I/O operations
21 * that are backed by invisible buffers. For example, SSL is able to read a
22 * whole socket buffer and not deliver it to the application buffer because
23 * it's full. Unfortunately, it won't be reported by a poller anymore until
24 * some new activity happens. The only way to call it again thus is to keep
25 * this readiness information in the cache and to access it without polling
26 * once the FD is enabled again.
Willy Tarreau7be79a42012-11-11 15:02:54 +010027 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010028 * One interesting feature of the cache is that it maintains the principle
29 * of speculative I/O introduced in haproxy 1.3 : the first time an event is
30 * enabled, the FD is considered as ready so that the I/O attempt is performed
31 * via the cache without polling. And the polling happens only when EAGAIN is
32 * first met. This avoids polling for HTTP requests, especially when the
33 * defer-accept mode is used. It also avoids polling for sending short data
34 * such as requests to servers or short responses to clients.
35 *
36 * The cache consists in a list of active events and a list of updates.
37 * Active events are events that are expected to come and that we must report
38 * to the application until it asks to stop or asks to poll. Updates are new
39 * requests for changing an FD state. Updates are the only way to create new
40 * events. This is important because it means that the number of cached events
41 * cannot increase between updates and will only grow one at a time while
42 * processing updates. All updates must always be processed, though events
43 * might be processed by small batches if required.
Willy Tarreau7be79a42012-11-11 15:02:54 +010044 *
45 * There is no direct link between the FD and the updates list. There is only a
46 * bit in the fdtab[] to indicate than a file descriptor is already present in
47 * the updates list. Once an fd is present in the updates list, it will have to
48 * be considered even if its changes are reverted in the middle or if the fd is
49 * replaced.
50 *
51 * It is important to understand that as long as all expected events are
52 * processed, they might starve the polled events, especially because polled
Willy Tarreauf817e9f2014-01-10 16:58:45 +010053 * I/O starvation quickly induces more cached I/O. One solution to this
Willy Tarreau7be79a42012-11-11 15:02:54 +010054 * consists in only processing a part of the events at once, but one drawback
Willy Tarreauf817e9f2014-01-10 16:58:45 +010055 * is that unhandled events will still wake the poller up. Using an edge-
56 * triggered poller such as EPOLL_ET will solve this issue though.
Willy Tarreau7be79a42012-11-11 15:02:54 +010057 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010058 * Since we do not want to scan all the FD list to find cached I/O events,
59 * we store them in a list consisting in a linear array holding only the FD
60 * indexes right now. Note that a closed FD cannot exist in the cache, because
61 * it is closed by fd_delete() which in turn calls fd_release_cache_entry()
62 * which always removes it from the list.
Willy Tarreau7be79a42012-11-11 15:02:54 +010063 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010064 * The FD array has to hold a back reference to the cache. This reference is
65 * always valid unless the FD is not in the cache and is not updated, in which
66 * case the reference points to index 0.
Willy Tarreau7be79a42012-11-11 15:02:54 +010067 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010068 * The event state for an FD, as found in fdtab[].state, is maintained for each
69 * direction. The state field is built this way, with R bits in the low nibble
70 * and W bits in the high nibble for ease of access and debugging :
Willy Tarreau7be79a42012-11-11 15:02:54 +010071 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010072 * 7 6 5 4 3 2 1 0
73 * [ 0 | PW | RW | AW | 0 | PR | RR | AR ]
74 *
75 * A* = active *R = read
76 * P* = polled *W = write
77 * R* = ready
78 *
79 * An FD is marked "active" when there is a desire to use it.
80 * An FD is marked "polled" when it is registered in the polling.
81 * An FD is marked "ready" when it has not faced a new EAGAIN since last wake-up
82 * (it is a cache of the last EAGAIN regardless of polling changes).
Willy Tarreau7be79a42012-11-11 15:02:54 +010083 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010084 * We have 8 possible states for each direction based on these 3 flags :
Willy Tarreau7be79a42012-11-11 15:02:54 +010085 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010086 * +---+---+---+----------+---------------------------------------------+
87 * | P | R | A | State | Description |
88 * +---+---+---+----------+---------------------------------------------+
89 * | 0 | 0 | 0 | DISABLED | No activity desired, not ready. |
90 * | 0 | 0 | 1 | MUSTPOLL | Activity desired via polling. |
91 * | 0 | 1 | 0 | STOPPED | End of activity without polling. |
92 * | 0 | 1 | 1 | ACTIVE | Activity desired without polling. |
93 * | 1 | 0 | 0 | ABORT | Aborted poll(). Not frequently seen. |
94 * | 1 | 0 | 1 | POLLED | FD is being polled. |
95 * | 1 | 1 | 0 | PAUSED | FD was paused while ready (eg: buffer full) |
96 * | 1 | 1 | 1 | READY | FD was marked ready by poll() |
97 * +---+---+---+----------+---------------------------------------------+
Willy Tarreau7be79a42012-11-11 15:02:54 +010098 *
Willy Tarreauf817e9f2014-01-10 16:58:45 +010099 * The transitions are pretty simple :
100 * - fd_want_*() : set flag A
101 * - fd_stop_*() : clear flag A
102 * - fd_cant_*() : clear flag R (when facing EAGAIN)
103 * - fd_may_*() : set flag R (upon return from poll())
104 * - sync() : if (A) { if (!R) P := 1 } else { P := 0 }
105 *
106 * The PAUSED, ABORT and MUSTPOLL states are transient for level-trigerred
107 * pollers and are fixed by the sync() which happens at the beginning of the
108 * poller. For event-triggered pollers, only the MUSTPOLL state will be
109 * transient and ABORT will lead to PAUSED. The ACTIVE state is the only stable
110 * one which has P != A.
111 *
112 * The READY state is a bit special as activity on the FD might be notified
113 * both by the poller or by the cache. But it is needed for some multi-layer
114 * protocols (eg: SSL) where connection activity is not 100% linked to FD
115 * activity. Also some pollers might prefer to implement it as ACTIVE if
116 * enabling/disabling the FD is cheap. The READY and ACTIVE states are the
117 * two states for which a cache entry is allocated.
118 *
119 * The state transitions look like the diagram below. Only the 4 right states
120 * have polling enabled :
121 *
122 * (POLLED=0) (POLLED=1)
123 *
124 * +----------+ sync +-------+
125 * | DISABLED | <----- | ABORT | (READY=0, ACTIVE=0)
126 * +----------+ +-------+
127 * clr | ^ set | ^
128 * | | | |
129 * v | set v | clr
130 * +----------+ sync +--------+
131 * | MUSTPOLL | -----> | POLLED | (READY=0, ACTIVE=1)
132 * +----------+ +--------+
133 * ^ poll | ^
134 * | | |
135 * | EAGAIN v | EAGAIN
136 * +--------+ +-------+
137 * | ACTIVE | | READY | (READY=1, ACTIVE=1)
138 * +--------+ +-------+
139 * clr | ^ set | ^
140 * | | | |
141 * v | set v | clr
142 * +---------+ sync +--------+
143 * | STOPPED | <------ | PAUSED | (READY=1, ACTIVE=0)
144 * +---------+ +--------+
Willy Tarreaubaaee002006-06-26 02:48:02 +0200145 */
146
Willy Tarreau2ff76222007-04-09 19:29:56 +0200147#include <stdio.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +0200148#include <string.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200149#include <unistd.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200150#include <sys/types.h>
151
Willy Tarreau2dd0d472006-06-29 17:53:05 +0200152#include <common/compat.h>
153#include <common/config.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200154
Willy Tarreau7be79a42012-11-11 15:02:54 +0100155#include <types/global.h>
156
Willy Tarreau2a429502006-10-15 14:52:29 +0200157#include <proto/fd.h>
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200158#include <proto/log.h>
Willy Tarreauc6f4ce82009-06-10 11:09:37 +0200159#include <proto/port_range.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200160
161struct fdtab *fdtab = NULL; /* array of all the file descriptors */
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200162struct fdinfo *fdinfo = NULL; /* less-often used infos for file descriptors */
Willy Tarreaubaaee002006-06-26 02:48:02 +0200163int totalconn; /* total # of terminated sessions */
164int actconn; /* # of active sessions */
165
Willy Tarreau4f60f162007-04-08 16:39:58 +0200166struct poller pollers[MAX_POLLERS];
167struct poller cur_poller;
168int nbpollers = 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200169
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100170volatile struct fdlist fd_cache ; // FD events cache
171volatile struct fdlist fd_cache_local[MAX_THREADS]; // FD events local for each thread
172
Christopher Faulet69553fe2018-01-15 11:57:03 +0100173unsigned long fd_cache_mask = 0; // Mask of threads with events in the cache
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200174
175THREAD_LOCAL int *fd_updt = NULL; // FD updates list
176THREAD_LOCAL int fd_nbupdt = 0; // number of updates in the list
177
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100178/* adds fd <fd> to fd list <list> if it was not yet in it */
179void fd_add_to_fd_list(volatile struct fdlist *list, int fd)
180{
181 int next;
182 int new;
183 int old;
184 int last;
185
186redo_next:
187 next = fdtab[fd].cache.next;
188 /* Check that we're not already in the cache, and if not, lock us. */
189 if (next >= -2)
190 goto done;
191 if (!HA_ATOMIC_CAS(&fdtab[fd].cache.next, &next, -2))
192 goto redo_next;
193 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100194
195 new = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100196redo_last:
197 /* First, insert in the linked list */
198 last = list->last;
199 old = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100200
Willy Tarreau11559a72018-02-05 17:52:24 +0100201 fdtab[fd].cache.prev = last;
202 /* Make sure the "prev" store is visible before we update the last entry */
203 __ha_barrier_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100204
Willy Tarreau11559a72018-02-05 17:52:24 +0100205 if (unlikely(last == -1)) {
206 /* list is empty, try to add ourselves alone so that list->last=fd */
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100207 if (unlikely(!HA_ATOMIC_CAS(&list->last, &old, new)))
208 goto redo_last;
209
210 /* list->first was necessary -1, we're guaranteed to be alone here */
211 list->first = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100212 } else {
Willy Tarreau11559a72018-02-05 17:52:24 +0100213 /* adding ourselves past the last element
214 * The CAS will only succeed if its next is -1,
215 * which means it's in the cache, and the last element.
216 */
217 if (unlikely(!HA_ATOMIC_CAS(&fdtab[last].cache.next, &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100218 goto redo_last;
Willy Tarreau11559a72018-02-05 17:52:24 +0100219
220 /* Then, update the last entry */
221 list->last = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100222 }
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100223 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100224 /* since we're alone at the end of the list and still locked(-2),
225 * we know noone tried to add past us. Mark the end of list.
226 */
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100227 fdtab[fd].cache.next = -1;
228 __ha_barrier_store();
229done:
230 return;
231}
232
233/* removes fd <fd> from fd list <list> */
234void fd_rm_from_fd_list(volatile struct fdlist *list, int fd)
235{
236#if defined(HA_HAVE_CAS_DW) || defined(HA_CAS_IS_8B)
237 volatile struct fdlist_entry cur_list, next_list;
238#endif
239 int old;
240 int new = -2;
241 int prev;
242 int next;
243 int last;
244
245lock_self:
246#if (defined(HA_CAS_IS_8B) || defined(HA_HAVE_CAS_DW))
247 next_list.next = next_list.prev = -2;
248 cur_list = fdtab[fd].cache;
249 /* First, attempt to lock our own entries */
250 do {
251 /* The FD is not in the FD cache, give up */
252 if (unlikely(cur_list.next <= -3))
253 return;
254 if (unlikely(cur_list.prev == -2 || cur_list.next == -2))
255 goto lock_self;
256 } while (
257#ifdef HA_CAS_IS_8B
258 unlikely(!HA_ATOMIC_CAS(((void **)(void *)&fdtab[fd].cache.next), ((void **)(void *)&cur_list), (*(void **)(void *)&next_list))))
259#else
260 unlikely(!__ha_cas_dw((void *)&fdtab[fd].cache.next, (void *)&cur_list, (void *)&next_list)))
261#endif
262 ;
263 next = cur_list.next;
264 prev = cur_list.prev;
265
266#else
267lock_self_next:
Willy Tarreau82106982018-02-06 12:00:27 +0100268 next = ({ volatile int *next = &fdtab[fd].cache.next; *next; });
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100269 if (next == -2)
270 goto lock_self_next;
271 if (next <= -3)
272 goto done;
273 if (unlikely(!HA_ATOMIC_CAS(&fdtab[fd].cache.next, &next, -2)))
274 goto lock_self_next;
275lock_self_prev:
Willy Tarreau82106982018-02-06 12:00:27 +0100276 prev = ({ volatile int *prev = &fdtab[fd].cache.prev; *prev; });
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100277 if (prev == -2)
278 goto lock_self_prev;
279 if (unlikely(!HA_ATOMIC_CAS(&fdtab[fd].cache.prev, &prev, -2)))
280 goto lock_self_prev;
281#endif
282 __ha_barrier_store();
283
284 /* Now, lock the entries of our neighbours */
285 if (likely(prev != -1)) {
286redo_prev:
287 old = fd;
288
289 if (unlikely(!HA_ATOMIC_CAS(&fdtab[prev].cache.next, &old, new))) {
290 if (unlikely(old == -2)) {
291 /* Neighbour already locked, give up and
292 * retry again once he's done
293 */
294 fdtab[fd].cache.prev = prev;
295 __ha_barrier_store();
296 fdtab[fd].cache.next = next;
297 __ha_barrier_store();
298 goto lock_self;
299 }
300 goto redo_prev;
301 }
302 }
303 if (likely(next != -1)) {
304redo_next:
305 old = fd;
306 if (unlikely(!HA_ATOMIC_CAS(&fdtab[next].cache.prev, &old, new))) {
307 if (unlikely(old == -2)) {
308 /* Neighbour already locked, give up and
309 * retry again once he's done
310 */
311 if (prev != -1) {
312 fdtab[prev].cache.next = fd;
313 __ha_barrier_store();
314 }
315 fdtab[fd].cache.prev = prev;
316 __ha_barrier_store();
317 fdtab[fd].cache.next = next;
318 __ha_barrier_store();
319 goto lock_self;
320 }
321 goto redo_next;
322 }
323 }
324 if (list->first == fd)
325 list->first = next;
326 __ha_barrier_store();
327 last = list->last;
328 while (unlikely(last == fd && (!HA_ATOMIC_CAS(&list->last, &last, prev))))
329 __ha_compiler_barrier();
330 /* Make sure we let other threads know we're no longer in cache,
331 * before releasing our neighbours.
332 */
333 __ha_barrier_store();
334 if (likely(prev != -1))
335 fdtab[prev].cache.next = next;
336 __ha_barrier_store();
337 if (likely(next != -1))
338 fdtab[next].cache.prev = prev;
339 __ha_barrier_store();
340 /* Ok, now we're out of the fd cache */
341 fdtab[fd].cache.next = -(next + 4);
342 __ha_barrier_store();
343done:
344 return;
345}
346
Willy Tarreau173d9952018-01-26 21:48:23 +0100347/* Deletes an FD from the fdsets.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200348 * The file descriptor is also closed.
349 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200350static void fd_dodelete(int fd, int do_close)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200351{
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100352 HA_SPIN_LOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreauad38ace2013-12-15 14:19:38 +0100353 if (fdtab[fd].linger_risk) {
354 /* this is generally set when connecting to servers */
355 setsockopt(fd, SOL_SOCKET, SO_LINGER,
356 (struct linger *) &nolinger, sizeof(struct linger));
357 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100358 if (cur_poller.clo)
359 cur_poller.clo(fd);
360
Willy Tarreau899d9572014-01-25 19:20:35 +0100361 fd_release_cache_entry(fd);
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100362 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100363
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200364 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
365 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200366 fdtab[fd].owner = NULL;
Willy Tarreauebc78d72018-01-20 23:53:50 +0100367 fdtab[fd].update_mask &= ~tid_bit;
Willy Tarreauf65610a2017-10-31 16:06:06 +0100368 fdtab[fd].thread_mask = 0;
Willy Tarreauc9c83782018-01-17 18:44:46 +0100369 if (do_close) {
370 fdtab[fd].polled_mask = 0;
Christopher Fauletd531f882017-06-01 16:55:03 +0200371 close(fd);
Willy Tarreauc9c83782018-01-17 18:44:46 +0100372 }
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100373 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200374}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200375
Willy Tarreau173d9952018-01-26 21:48:23 +0100376/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200377 * The file descriptor is also closed.
378 */
379void fd_delete(int fd)
380{
381 fd_dodelete(fd, 1);
382}
383
Willy Tarreau173d9952018-01-26 21:48:23 +0100384/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200385 * The file descriptor is kept open.
386 */
387void fd_remove(int fd)
388{
389 fd_dodelete(fd, 0);
390}
391
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100392static inline void fdlist_process_cached_events(volatile struct fdlist *fdlist)
Willy Tarreau09f24562012-11-11 16:43:45 +0100393{
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100394 int fd, old_fd, e;
Willy Tarreau09f24562012-11-11 16:43:45 +0100395
Olivier Houchard12568362018-01-31 18:07:29 +0100396 for (old_fd = fd = fdlist->first; fd != -1; fd = fdtab[fd].cache.next) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100397 if (fd == -2) {
398 fd = old_fd;
399 continue;
400 } else if (fd <= -3)
401 fd = -fd - 4;
402 if (fd == -1)
403 break;
404 old_fd = fd;
405 if (!(fdtab[fd].thread_mask & tid_bit))
406 continue;
Olivier Houchard12568362018-01-31 18:07:29 +0100407 if (fdtab[fd].cache.next < -3)
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100408 continue;
Christopher Faulet69553fe2018-01-15 11:57:03 +0100409
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100410 HA_ATOMIC_OR(&fd_cache_mask, tid_bit);
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100411 if (HA_SPIN_TRYLOCK(FD_LOCK, &fdtab[fd].lock)) {
412 activity[tid].fd_lock++;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100413 continue;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100414 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200415
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200416 e = fdtab[fd].state;
Willy Tarreau09f24562012-11-11 16:43:45 +0100417 fdtab[fd].ev &= FD_POLL_STICKY;
418
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100419 if ((e & (FD_EV_READY_R | FD_EV_ACTIVE_R)) == (FD_EV_READY_R | FD_EV_ACTIVE_R))
Willy Tarreau09f24562012-11-11 16:43:45 +0100420 fdtab[fd].ev |= FD_POLL_IN;
421
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100422 if ((e & (FD_EV_READY_W | FD_EV_ACTIVE_W)) == (FD_EV_READY_W | FD_EV_ACTIVE_W))
Willy Tarreau09f24562012-11-11 16:43:45 +0100423 fdtab[fd].ev |= FD_POLL_OUT;
424
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200425 if (fdtab[fd].iocb && fdtab[fd].owner && fdtab[fd].ev) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100426 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreau09f24562012-11-11 16:43:45 +0100427 fdtab[fd].iocb(fd);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200428 }
429 else {
Willy Tarreau5be2f352014-11-19 19:43:05 +0100430 fd_release_cache_entry(fd);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100431 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200432 }
Willy Tarreau09f24562012-11-11 16:43:45 +0100433 }
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100434}
435
436/* Scan and process the cached events. This should be called right after
437 * the poller. The loop may cause new entries to be created, for example
438 * if a listener causes an accept() to initiate a new incoming connection
439 * wanting to attempt an recv().
440 */
441void fd_process_cached_events()
442{
443 HA_ATOMIC_AND(&fd_cache_mask, ~tid_bit);
444 fdlist_process_cached_events(&fd_cache_local[tid]);
445 fdlist_process_cached_events(&fd_cache);
Willy Tarreau09f24562012-11-11 16:43:45 +0100446}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200447
Willy Tarreau4f60f162007-04-08 16:39:58 +0200448/* disable the specified poller */
449void disable_poller(const char *poller_name)
450{
451 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200452
Willy Tarreau4f60f162007-04-08 16:39:58 +0200453 for (p = 0; p < nbpollers; p++)
454 if (strcmp(pollers[p].name, poller_name) == 0)
455 pollers[p].pref = 0;
456}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200457
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200458/* Initialize the pollers per thread */
459static int init_pollers_per_thread()
460{
461 if ((fd_updt = calloc(global.maxsock, sizeof(*fd_updt))) == NULL)
462 return 0;
463 return 1;
464}
465
466/* Deinitialize the pollers per thread */
467static void deinit_pollers_per_thread()
468{
469 free(fd_updt);
470 fd_updt = NULL;
471}
472
Willy Tarreaubaaee002006-06-26 02:48:02 +0200473/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200474 * Initialize the pollers till the best one is found.
475 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200476 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200477int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200478{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200479 int p;
480 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200481
Christopher Faulet63fe6522017-08-31 17:52:09 +0200482 if ((fdtab = calloc(global.maxsock, sizeof(struct fdtab))) == NULL)
483 goto fail_tab;
484
485 if ((fdinfo = calloc(global.maxsock, sizeof(struct fdinfo))) == NULL)
486 goto fail_info;
487
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100488 fd_cache.first = fd_cache.last = -1;
Christopher Fauletcd7879a2017-10-27 13:53:47 +0200489 hap_register_per_thread_init(init_pollers_per_thread);
490 hap_register_per_thread_deinit(deinit_pollers_per_thread);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200491
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100492 for (p = 0; p < global.maxsock; p++) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100493 HA_SPIN_INIT(&fdtab[p].lock);
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100494 /* Mark the fd as out of the fd cache */
Olivier Houchard12568362018-01-31 18:07:29 +0100495 fdtab[p].cache.next = -3;
496 fdtab[p].cache.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100497 }
498 for (p = 0; p < global.nbthread; p++)
499 fd_cache_local[p].first = fd_cache_local[p].last = -1;
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200500
Willy Tarreau4f60f162007-04-08 16:39:58 +0200501 do {
502 bp = NULL;
503 for (p = 0; p < nbpollers; p++)
504 if (!bp || (pollers[p].pref > bp->pref))
505 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200506
Willy Tarreau4f60f162007-04-08 16:39:58 +0200507 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200508 break;
509
Willy Tarreau4f60f162007-04-08 16:39:58 +0200510 if (bp->init(bp)) {
511 memcpy(&cur_poller, bp, sizeof(*bp));
512 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200513 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200514 } while (!bp || bp->pref == 0);
515 return 0;
Willy Tarreau7be79a42012-11-11 15:02:54 +0100516
Willy Tarreau16f649c2014-01-25 19:10:48 +0100517 fail_cache:
Christopher Faulet63fe6522017-08-31 17:52:09 +0200518 free(fdinfo);
519 fail_info:
520 free(fdtab);
521 fail_tab:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100522 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200523}
524
Willy Tarreaubaaee002006-06-26 02:48:02 +0200525/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200526 * Deinitialize the pollers.
527 */
528void deinit_pollers() {
529
530 struct poller *bp;
531 int p;
532
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200533 for (p = 0; p < global.maxsock; p++)
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100534 HA_SPIN_DESTROY(&fdtab[p].lock);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200535
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200536 for (p = 0; p < nbpollers; p++) {
537 bp = &pollers[p];
538
539 if (bp && bp->pref)
540 bp->term(bp);
541 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200542
Christopher Faulet63fe6522017-08-31 17:52:09 +0200543 free(fdinfo); fdinfo = NULL;
544 free(fdtab); fdtab = NULL;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200545}
546
547/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200548 * Lists the known pollers on <out>.
549 * Should be performed only before initialization.
550 */
551int list_pollers(FILE *out)
552{
553 int p;
554 int last, next;
555 int usable;
556 struct poller *bp;
557
558 fprintf(out, "Available polling systems :\n");
559
560 usable = 0;
561 bp = NULL;
562 last = next = -1;
563 while (1) {
564 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200565 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100566 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200567 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100568 if (!bp || (pollers[p].pref > bp->pref))
569 bp = &pollers[p];
570 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200571 }
572
573 if (next == -1)
574 break;
575
576 for (p = 0; p < nbpollers; p++) {
577 if (pollers[p].pref == next) {
578 fprintf(out, " %10s : ", pollers[p].name);
579 if (pollers[p].pref == 0)
580 fprintf(out, "disabled, ");
581 else
582 fprintf(out, "pref=%3d, ", pollers[p].pref);
583 if (pollers[p].test(&pollers[p])) {
584 fprintf(out, " test result OK");
585 if (next > 0)
586 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100587 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200588 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100589 if (bp == &pollers[p])
590 bp = NULL;
591 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200592 fprintf(out, "\n");
593 }
594 }
595 last = next;
596 next = -1;
597 };
598 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
599 return 0;
600}
601
602/*
603 * Some pollers may lose their connection after a fork(). It may be necessary
604 * to create initialize part of them again. Returns 0 in case of failure,
605 * otherwise 1. The fork() function may be NULL if unused. In case of error,
606 * the the current poller is destroyed and the caller is responsible for trying
607 * another one by calling init_pollers() again.
608 */
609int fork_poller()
610{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200611 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100612 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200613 if (fdtab[fd].owner) {
614 fdtab[fd].cloned = 1;
615 }
616 }
617
Willy Tarreau2ff76222007-04-09 19:29:56 +0200618 if (cur_poller.fork) {
619 if (cur_poller.fork(&cur_poller))
620 return 1;
621 cur_poller.term(&cur_poller);
622 return 0;
623 }
624 return 1;
625}
626
627/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200628 * Local variables:
629 * c-indent-level: 8
630 * c-basic-offset: 8
631 * End:
632 */