blob: 84cb9080c27174447d16f88a59d5f2fb046d423c [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>
Olivier Houchard79321b92018-07-26 17:55:11 +0200150#include <fcntl.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200151#include <sys/types.h>
152
Willy Tarreau2dd0d472006-06-29 17:53:05 +0200153#include <common/compat.h>
154#include <common/config.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200155
Willy Tarreau7be79a42012-11-11 15:02:54 +0100156#include <types/global.h>
157
Willy Tarreau2a429502006-10-15 14:52:29 +0200158#include <proto/fd.h>
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200159#include <proto/log.h>
Willy Tarreauc6f4ce82009-06-10 11:09:37 +0200160#include <proto/port_range.h>
Willy Tarreaubaaee002006-06-26 02:48:02 +0200161
162struct fdtab *fdtab = NULL; /* array of all the file descriptors */
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200163unsigned long *polled_mask = NULL; /* Array for the polled_mask of each fd */
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200164struct fdinfo *fdinfo = NULL; /* less-often used infos for file descriptors */
Willy Tarreaubaaee002006-06-26 02:48:02 +0200165int totalconn; /* total # of terminated sessions */
166int actconn; /* # of active sessions */
167
Willy Tarreau4f60f162007-04-08 16:39:58 +0200168struct poller pollers[MAX_POLLERS];
169struct poller cur_poller;
170int nbpollers = 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200171
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100172volatile struct fdlist fd_cache ; // FD events cache
173volatile struct fdlist fd_cache_local[MAX_THREADS]; // FD events local for each thread
Olivier Houchard6b96f722018-04-25 16:58:25 +0200174volatile struct fdlist update_list; // Global update list
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100175
Christopher Faulet69553fe2018-01-15 11:57:03 +0100176unsigned long fd_cache_mask = 0; // Mask of threads with events in the cache
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200177
178THREAD_LOCAL int *fd_updt = NULL; // FD updates list
179THREAD_LOCAL int fd_nbupdt = 0; // number of updates in the list
Olivier Houchard79321b92018-07-26 17:55:11 +0200180THREAD_LOCAL int poller_rd_pipe = -1; // Pipe to wake the thread
181int poller_wr_pipe[MAX_THREADS]; // Pipe to wake the threads
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200182
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200183#define _GET_NEXT(fd, off) ((struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->next
184#define _GET_PREV(fd, off) ((struct fdlist_entry *)(void *)((char *)(&fdtab[fd]) + off))->prev
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100185/* adds fd <fd> to fd list <list> if it was not yet in it */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200186void fd_add_to_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100187{
188 int next;
189 int new;
190 int old;
191 int last;
192
193redo_next:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200194 next = _GET_NEXT(fd, off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100195 /* Check that we're not already in the cache, and if not, lock us. */
196 if (next >= -2)
197 goto done;
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200198 if (!HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100199 goto redo_next;
200 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100201
202 new = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100203redo_last:
204 /* First, insert in the linked list */
205 last = list->last;
206 old = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100207
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200208 _GET_PREV(fd, off) = -2;
Willy Tarreau11559a72018-02-05 17:52:24 +0100209 /* Make sure the "prev" store is visible before we update the last entry */
210 __ha_barrier_store();
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100211
Willy Tarreau11559a72018-02-05 17:52:24 +0100212 if (unlikely(last == -1)) {
213 /* list is empty, try to add ourselves alone so that list->last=fd */
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100214 if (unlikely(!HA_ATOMIC_CAS(&list->last, &old, new)))
215 goto redo_last;
216
217 /* list->first was necessary -1, we're guaranteed to be alone here */
218 list->first = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100219 } else {
Willy Tarreau11559a72018-02-05 17:52:24 +0100220 /* adding ourselves past the last element
221 * The CAS will only succeed if its next is -1,
222 * which means it's in the cache, and the last element.
223 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200224 if (unlikely(!HA_ATOMIC_CAS(&_GET_NEXT(last, off), &old, new)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100225 goto redo_last;
Willy Tarreau11559a72018-02-05 17:52:24 +0100226
227 /* Then, update the last entry */
228 list->last = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100229 }
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100230 __ha_barrier_store();
Willy Tarreau11559a72018-02-05 17:52:24 +0100231 /* since we're alone at the end of the list and still locked(-2),
232 * we know noone tried to add past us. Mark the end of list.
233 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200234 _GET_PREV(fd, off) = last;
235 _GET_NEXT(fd, off) = -1;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100236 __ha_barrier_store();
237done:
238 return;
239}
240
241/* removes fd <fd> from fd list <list> */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200242void fd_rm_from_fd_list(volatile struct fdlist *list, int fd, int off)
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100243{
244#if defined(HA_HAVE_CAS_DW) || defined(HA_CAS_IS_8B)
245 volatile struct fdlist_entry cur_list, next_list;
246#endif
247 int old;
248 int new = -2;
249 int prev;
250 int next;
251 int last;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100252lock_self:
253#if (defined(HA_CAS_IS_8B) || defined(HA_HAVE_CAS_DW))
254 next_list.next = next_list.prev = -2;
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200255 cur_list = *(volatile struct fdlist_entry *)(((char *)&fdtab[fd]) + off);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100256 /* First, attempt to lock our own entries */
257 do {
258 /* The FD is not in the FD cache, give up */
259 if (unlikely(cur_list.next <= -3))
260 return;
261 if (unlikely(cur_list.prev == -2 || cur_list.next == -2))
262 goto lock_self;
263 } while (
264#ifdef HA_CAS_IS_8B
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200265 unlikely(!HA_ATOMIC_CAS(((void **)(void *)&_GET_NEXT(fd, off)), ((void **)(void *)&cur_list), (*(void **)(void *)&next_list))))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100266#else
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200267 unlikely(!__ha_cas_dw((void *)&_GET_NEXT(fd, off), (void *)&cur_list, (void *)&next_list)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100268#endif
269 ;
270 next = cur_list.next;
271 prev = cur_list.prev;
272
273#else
274lock_self_next:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200275 next = ({ volatile int *next = &_GET_NEXT(fd, off); *next; });
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100276 if (next == -2)
277 goto lock_self_next;
278 if (next <= -3)
279 goto done;
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200280 if (unlikely(!HA_ATOMIC_CAS(&_GET_NEXT(fd, off), &next, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100281 goto lock_self_next;
282lock_self_prev:
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200283 prev = ({ volatile int *prev = &_GET_PREV(fd, off); *prev; });
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100284 if (prev == -2)
285 goto lock_self_prev;
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200286 if (unlikely(!HA_ATOMIC_CAS(&_GET_PREV(fd, off), &prev, -2)))
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100287 goto lock_self_prev;
288#endif
289 __ha_barrier_store();
290
291 /* Now, lock the entries of our neighbours */
292 if (likely(prev != -1)) {
293redo_prev:
294 old = fd;
295
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200296 if (unlikely(!HA_ATOMIC_CAS(&_GET_NEXT(prev, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100297 if (unlikely(old == -2)) {
298 /* Neighbour already locked, give up and
299 * retry again once he's done
300 */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200301 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100302 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200303 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100304 __ha_barrier_store();
305 goto lock_self;
306 }
307 goto redo_prev;
308 }
309 }
310 if (likely(next != -1)) {
311redo_next:
312 old = fd;
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200313 if (unlikely(!HA_ATOMIC_CAS(&_GET_PREV(next, off), &old, new))) {
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100314 if (unlikely(old == -2)) {
315 /* Neighbour already locked, give up and
316 * retry again once he's done
317 */
318 if (prev != -1) {
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200319 _GET_NEXT(prev, off) = fd;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100320 __ha_barrier_store();
321 }
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200322 _GET_PREV(fd, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100323 __ha_barrier_store();
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200324 _GET_NEXT(fd, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100325 __ha_barrier_store();
326 goto lock_self;
327 }
328 goto redo_next;
329 }
330 }
331 if (list->first == fd)
332 list->first = next;
333 __ha_barrier_store();
334 last = list->last;
335 while (unlikely(last == fd && (!HA_ATOMIC_CAS(&list->last, &last, prev))))
336 __ha_compiler_barrier();
337 /* Make sure we let other threads know we're no longer in cache,
338 * before releasing our neighbours.
339 */
340 __ha_barrier_store();
341 if (likely(prev != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200342 _GET_NEXT(prev, off) = next;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100343 __ha_barrier_store();
344 if (likely(next != -1))
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200345 _GET_PREV(next, off) = prev;
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100346 __ha_barrier_store();
347 /* Ok, now we're out of the fd cache */
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200348 _GET_NEXT(fd, off) = -(next + 4);
Willy Tarreau4cc67a22018-02-05 17:14:55 +0100349 __ha_barrier_store();
350done:
351 return;
352}
353
Olivier Houchard6a2cf872018-04-25 15:10:30 +0200354#undef _GET_NEXT
355#undef _GET_PREV
356
Willy Tarreau173d9952018-01-26 21:48:23 +0100357/* Deletes an FD from the fdsets.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200358 * The file descriptor is also closed.
359 */
Olivier Houchard1fc05162017-04-06 01:05:05 +0200360static void fd_dodelete(int fd, int do_close)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200361{
Willy Tarreau87d54a92018-10-15 09:44:46 +0200362 unsigned long locked = atleast2(fdtab[fd].thread_mask);
363
364 if (locked)
365 HA_SPIN_LOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreauad38ace2013-12-15 14:19:38 +0100366 if (fdtab[fd].linger_risk) {
367 /* this is generally set when connecting to servers */
368 setsockopt(fd, SOL_SOCKET, SO_LINGER,
369 (struct linger *) &nolinger, sizeof(struct linger));
370 }
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100371 if (cur_poller.clo)
372 cur_poller.clo(fd);
373
Willy Tarreau899d9572014-01-25 19:20:35 +0100374 fd_release_cache_entry(fd);
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100375 fdtab[fd].state = 0;
Willy Tarreau6ea20b12012-11-11 16:05:19 +0100376
Willy Tarreau8d5d77e2009-10-18 07:25:52 +0200377 port_range_release_port(fdinfo[fd].port_range, fdinfo[fd].local_port);
378 fdinfo[fd].port_range = NULL;
Willy Tarreaudb3b3262012-07-05 23:19:22 +0200379 fdtab[fd].owner = NULL;
Willy Tarreauf65610a2017-10-31 16:06:06 +0100380 fdtab[fd].thread_mask = 0;
Willy Tarreauc9c83782018-01-17 18:44:46 +0100381 if (do_close) {
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200382 polled_mask[fd] = 0;
Christopher Fauletd531f882017-06-01 16:55:03 +0200383 close(fd);
Willy Tarreauc9c83782018-01-17 18:44:46 +0100384 }
Willy Tarreau87d54a92018-10-15 09:44:46 +0200385 if (locked)
386 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreaubaaee002006-06-26 02:48:02 +0200387}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200388
Willy Tarreau173d9952018-01-26 21:48:23 +0100389/* Deletes an FD from the fdsets.
Olivier Houchard1fc05162017-04-06 01:05:05 +0200390 * The file descriptor is also closed.
391 */
392void fd_delete(int fd)
393{
394 fd_dodelete(fd, 1);
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 kept open.
399 */
400void fd_remove(int fd)
401{
402 fd_dodelete(fd, 0);
403}
404
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100405static inline void fdlist_process_cached_events(volatile struct fdlist *fdlist)
Willy Tarreau09f24562012-11-11 16:43:45 +0100406{
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100407 int fd, old_fd, e;
Willy Tarreau09f24562012-11-11 16:43:45 +0100408
Olivier Houchard12568362018-01-31 18:07:29 +0100409 for (old_fd = fd = fdlist->first; fd != -1; fd = fdtab[fd].cache.next) {
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100410 if (fd == -2) {
411 fd = old_fd;
412 continue;
413 } else if (fd <= -3)
414 fd = -fd - 4;
415 if (fd == -1)
416 break;
417 old_fd = fd;
418 if (!(fdtab[fd].thread_mask & tid_bit))
419 continue;
Olivier Houchard12568362018-01-31 18:07:29 +0100420 if (fdtab[fd].cache.next < -3)
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100421 continue;
Christopher Faulet69553fe2018-01-15 11:57:03 +0100422
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100423 HA_ATOMIC_OR(&fd_cache_mask, tid_bit);
Willy Tarreau87d54a92018-10-15 09:44:46 +0200424 if (atleast2(fdtab[fd].thread_mask) && HA_SPIN_TRYLOCK(FD_LOCK, &fdtab[fd].lock)) {
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100425 activity[tid].fd_lock++;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100426 continue;
Willy Tarreaud80cb4e2018-01-20 19:30:13 +0100427 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200428
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200429 e = fdtab[fd].state;
Willy Tarreau09f24562012-11-11 16:43:45 +0100430 fdtab[fd].ev &= FD_POLL_STICKY;
431
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100432 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 +0100433 fdtab[fd].ev |= FD_POLL_IN;
434
Willy Tarreauf817e9f2014-01-10 16:58:45 +0100435 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 +0100436 fdtab[fd].ev |= FD_POLL_OUT;
437
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200438 if (fdtab[fd].iocb && fdtab[fd].owner && fdtab[fd].ev) {
Willy Tarreau87d54a92018-10-15 09:44:46 +0200439 if (atleast2(fdtab[fd].thread_mask))
440 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Willy Tarreau09f24562012-11-11 16:43:45 +0100441 fdtab[fd].iocb(fd);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200442 }
443 else {
Willy Tarreau5be2f352014-11-19 19:43:05 +0100444 fd_release_cache_entry(fd);
Willy Tarreau87d54a92018-10-15 09:44:46 +0200445 if (atleast2(fdtab[fd].thread_mask))
446 HA_SPIN_UNLOCK(FD_LOCK, &fdtab[fd].lock);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200447 }
Willy Tarreau09f24562012-11-11 16:43:45 +0100448 }
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100449}
450
451/* Scan and process the cached events. This should be called right after
452 * the poller. The loop may cause new entries to be created, for example
453 * if a listener causes an accept() to initiate a new incoming connection
454 * wanting to attempt an recv().
455 */
456void fd_process_cached_events()
457{
458 HA_ATOMIC_AND(&fd_cache_mask, ~tid_bit);
459 fdlist_process_cached_events(&fd_cache_local[tid]);
460 fdlist_process_cached_events(&fd_cache);
Willy Tarreau09f24562012-11-11 16:43:45 +0100461}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200462
Willy Tarreau4f60f162007-04-08 16:39:58 +0200463/* disable the specified poller */
464void disable_poller(const char *poller_name)
465{
466 int p;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200467
Willy Tarreau4f60f162007-04-08 16:39:58 +0200468 for (p = 0; p < nbpollers; p++)
469 if (strcmp(pollers[p].name, poller_name) == 0)
470 pollers[p].pref = 0;
471}
Willy Tarreaubaaee002006-06-26 02:48:02 +0200472
Olivier Houchard79321b92018-07-26 17:55:11 +0200473void poller_pipe_io_handler(int fd)
474{
475 char buf[1024];
476 /* Flush the pipe */
477 while (read(fd, buf, sizeof(buf)) > 0);
478 fd_cant_recv(fd);
479}
480
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200481/* Initialize the pollers per thread */
482static int init_pollers_per_thread()
483{
Olivier Houchard79321b92018-07-26 17:55:11 +0200484 int mypipe[2];
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200485 if ((fd_updt = calloc(global.maxsock, sizeof(*fd_updt))) == NULL)
486 return 0;
Olivier Houchard79321b92018-07-26 17:55:11 +0200487 if (pipe(mypipe) < 0) {
488 free(fd_updt);
489 fd_updt = NULL;
490 return 0;
491 }
492 poller_rd_pipe = mypipe[0];
493 poller_wr_pipe[tid] = mypipe[1];
494 fcntl(poller_rd_pipe, F_SETFL, O_NONBLOCK);
495 fd_insert(poller_rd_pipe, poller_pipe_io_handler, poller_pipe_io_handler,
496 tid_bit);
497 fd_want_recv(poller_rd_pipe);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200498 return 1;
499}
500
501/* Deinitialize the pollers per thread */
502static void deinit_pollers_per_thread()
503{
504 free(fd_updt);
505 fd_updt = NULL;
William Lallemand808e1b72018-12-15 22:34:31 +0100506
507 /* rd and wr are init at the same place, but only rd is init to -1, so
508 we rely to rd to close. */
509 if (poller_rd_pipe > -1) {
510 close(poller_rd_pipe);
511 poller_rd_pipe = -1;
512 close(poller_wr_pipe[tid]);
513 poller_wr_pipe[tid] = -1;
514 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200515}
516
Willy Tarreaubaaee002006-06-26 02:48:02 +0200517/*
Willy Tarreau4f60f162007-04-08 16:39:58 +0200518 * Initialize the pollers till the best one is found.
519 * If none works, returns 0, otherwise 1.
Willy Tarreaubaaee002006-06-26 02:48:02 +0200520 */
Willy Tarreau4f60f162007-04-08 16:39:58 +0200521int init_pollers()
Willy Tarreaubaaee002006-06-26 02:48:02 +0200522{
Willy Tarreau4f60f162007-04-08 16:39:58 +0200523 int p;
524 struct poller *bp;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200525
Christopher Faulet63fe6522017-08-31 17:52:09 +0200526 if ((fdtab = calloc(global.maxsock, sizeof(struct fdtab))) == NULL)
527 goto fail_tab;
528
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200529 if ((polled_mask = calloc(global.maxsock, sizeof(unsigned long))) == NULL)
530 goto fail_polledmask;
Christopher Faulet63fe6522017-08-31 17:52:09 +0200531 if ((fdinfo = calloc(global.maxsock, sizeof(struct fdinfo))) == NULL)
532 goto fail_info;
533
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100534 fd_cache.first = fd_cache.last = -1;
Olivier Houchard6b96f722018-04-25 16:58:25 +0200535 update_list.first = update_list.last = -1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200536
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100537 for (p = 0; p < global.maxsock; p++) {
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100538 HA_SPIN_INIT(&fdtab[p].lock);
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100539 /* Mark the fd as out of the fd cache */
Olivier Houchard12568362018-01-31 18:07:29 +0100540 fdtab[p].cache.next = -3;
Olivier Houchard6b96f722018-04-25 16:58:25 +0200541 fdtab[p].update.next = -3;
Olivier Houchard4815c8c2018-01-24 18:17:56 +0100542 }
543 for (p = 0; p < global.nbthread; p++)
544 fd_cache_local[p].first = fd_cache_local[p].last = -1;
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200545
Willy Tarreau4f60f162007-04-08 16:39:58 +0200546 do {
547 bp = NULL;
548 for (p = 0; p < nbpollers; p++)
549 if (!bp || (pollers[p].pref > bp->pref))
550 bp = &pollers[p];
Willy Tarreaubaaee002006-06-26 02:48:02 +0200551
Willy Tarreau4f60f162007-04-08 16:39:58 +0200552 if (!bp || bp->pref == 0)
Willy Tarreaubaaee002006-06-26 02:48:02 +0200553 break;
554
Willy Tarreau4f60f162007-04-08 16:39:58 +0200555 if (bp->init(bp)) {
556 memcpy(&cur_poller, bp, sizeof(*bp));
557 return 1;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200558 }
Willy Tarreau4f60f162007-04-08 16:39:58 +0200559 } while (!bp || bp->pref == 0);
560 return 0;
Willy Tarreau7be79a42012-11-11 15:02:54 +0100561
Willy Tarreau16f649c2014-01-25 19:10:48 +0100562 fail_cache:
Christopher Faulet63fe6522017-08-31 17:52:09 +0200563 free(fdinfo);
564 fail_info:
565 free(fdtab);
566 fail_tab:
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200567 free(polled_mask);
568 fail_polledmask:
Willy Tarreau7be79a42012-11-11 15:02:54 +0100569 return 0;
Willy Tarreaubaaee002006-06-26 02:48:02 +0200570}
571
Willy Tarreaubaaee002006-06-26 02:48:02 +0200572/*
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200573 * Deinitialize the pollers.
574 */
575void deinit_pollers() {
576
577 struct poller *bp;
578 int p;
579
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200580 for (p = 0; p < global.maxsock; p++)
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100581 HA_SPIN_DESTROY(&fdtab[p].lock);
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200582
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200583 for (p = 0; p < nbpollers; p++) {
584 bp = &pollers[p];
585
586 if (bp && bp->pref)
587 bp->term(bp);
588 }
Christopher Fauletd4604ad2017-05-29 10:40:41 +0200589
Christopher Faulet63fe6522017-08-31 17:52:09 +0200590 free(fdinfo); fdinfo = NULL;
591 free(fdtab); fdtab = NULL;
Olivier Houchardcb92f5c2018-04-26 14:23:07 +0200592 free(polled_mask); polled_mask = NULL;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200593}
594
595/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200596 * Lists the known pollers on <out>.
597 * Should be performed only before initialization.
598 */
599int list_pollers(FILE *out)
600{
601 int p;
602 int last, next;
603 int usable;
604 struct poller *bp;
605
606 fprintf(out, "Available polling systems :\n");
607
608 usable = 0;
609 bp = NULL;
610 last = next = -1;
611 while (1) {
612 for (p = 0; p < nbpollers; p++) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200613 if ((next < 0 || pollers[p].pref > next)
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100614 && (last < 0 || pollers[p].pref < last)) {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200615 next = pollers[p].pref;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100616 if (!bp || (pollers[p].pref > bp->pref))
617 bp = &pollers[p];
618 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200619 }
620
621 if (next == -1)
622 break;
623
624 for (p = 0; p < nbpollers; p++) {
625 if (pollers[p].pref == next) {
626 fprintf(out, " %10s : ", pollers[p].name);
627 if (pollers[p].pref == 0)
628 fprintf(out, "disabled, ");
629 else
630 fprintf(out, "pref=%3d, ", pollers[p].pref);
631 if (pollers[p].test(&pollers[p])) {
632 fprintf(out, " test result OK");
633 if (next > 0)
634 usable++;
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100635 } else {
Willy Tarreau2ff76222007-04-09 19:29:56 +0200636 fprintf(out, " test result FAILED");
Willy Tarreaue79c3b22010-11-19 10:20:36 +0100637 if (bp == &pollers[p])
638 bp = NULL;
639 }
Willy Tarreau2ff76222007-04-09 19:29:56 +0200640 fprintf(out, "\n");
641 }
642 }
643 last = next;
644 next = -1;
645 };
646 fprintf(out, "Total: %d (%d usable), will use %s.\n", nbpollers, usable, bp ? bp->name : "none");
647 return 0;
648}
649
650/*
651 * Some pollers may lose their connection after a fork(). It may be necessary
652 * to create initialize part of them again. Returns 0 in case of failure,
653 * otherwise 1. The fork() function may be NULL if unused. In case of error,
654 * the the current poller is destroyed and the caller is responsible for trying
655 * another one by calling init_pollers() again.
656 */
657int fork_poller()
658{
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200659 int fd;
Willy Tarreauce036bc2018-01-29 14:58:02 +0100660 for (fd = 0; fd < global.maxsock; fd++) {
Conrad Hoffmann041751c2014-05-20 14:28:24 +0200661 if (fdtab[fd].owner) {
662 fdtab[fd].cloned = 1;
663 }
664 }
665
Willy Tarreau2ff76222007-04-09 19:29:56 +0200666 if (cur_poller.fork) {
667 if (cur_poller.fork(&cur_poller))
668 return 1;
669 cur_poller.term(&cur_poller);
670 return 0;
671 }
672 return 1;
673}
674
Willy Tarreau172f5ce2018-11-26 11:21:50 +0100675REGISTER_PER_THREAD_INIT(init_pollers_per_thread);
676REGISTER_PER_THREAD_DEINIT(deinit_pollers_per_thread);
677
Willy Tarreau2ff76222007-04-09 19:29:56 +0200678/*
Willy Tarreaubaaee002006-06-26 02:48:02 +0200679 * Local variables:
680 * c-indent-level: 8
681 * c-basic-offset: 8
682 * End:
683 */