blob: adb0fd6e880d9131c26cc8c0c798feacafc0d12e [file] [log] [blame]
Willy Tarreau4f60f162007-04-08 16:39:58 +02001/*
2 * FD polling functions for linux epoll()
3 *
Willy Tarreaub7f694f2008-06-22 17:18:02 +02004 * Copyright 2000-2008 Willy Tarreau <w@1wt.eu>
Willy Tarreau4f60f162007-04-08 16:39:58 +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 *
11 */
12
13#include <unistd.h>
14#include <sys/time.h>
15#include <sys/types.h>
16
17#include <common/compat.h>
18#include <common/config.h>
Willy Tarreau58094f22007-04-10 01:33:20 +020019#include <common/standard.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020020#include <common/time.h>
Willy Tarreau1db37712007-06-03 17:16:49 +020021#include <common/tools.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020022
23#include <types/fd.h>
24#include <types/global.h>
25
26#include <proto/fd.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020027#include <proto/task.h>
28
29#if defined(USE_MY_EPOLL)
Willy Tarreau69801b82007-04-09 15:28:51 +020030#include <common/epoll.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020031#include <errno.h>
32#include <sys/syscall.h>
Willy Tarreaub40d4202007-04-15 23:57:41 +020033static _syscall1 (int, epoll_create, int, size);
34static _syscall4 (int, epoll_ctl, int, epfd, int, op, int, fd, struct epoll_event *, event);
35static _syscall4 (int, epoll_wait, int, epfd, struct epoll_event *, events, int, maxevents, int, timeout);
Willy Tarreau69801b82007-04-09 15:28:51 +020036#else
37#include <sys/epoll.h>
Willy Tarreau4f60f162007-04-08 16:39:58 +020038#endif
39
Willy Tarreau58094f22007-04-10 01:33:20 +020040/* This is what we store in a list. It consists in old values and fds to detect changes. */
41struct fd_chg {
42 unsigned int prev:2; // previous state mask. New one is in fd_evts.
43 unsigned int fd:30; // file descriptor
44};
Willy Tarreau4f60f162007-04-08 16:39:58 +020045
Willy Tarreau58094f22007-04-10 01:33:20 +020046static int nbchanges = 0; // number of changes pending
47static struct fd_chg *chg_list = NULL; // list of changes
48static struct fd_chg **chg_ptr = NULL; // per-fd changes
49
50/* Each 32-bit word contains 2-bit descriptors of the latest state for 16 FDs :
51 * desc = (u32 >> (2*fd)) & 3
52 * desc = 0 : FD not set
53 * 1 : WRITE not set, READ set
54 * 2 : WRITE set, READ not set
55 * 3 : WRITE set, READ set
56 */
57
58static uint32_t *fd_evts;
Willy Tarreau4f60f162007-04-08 16:39:58 +020059
60/* private data */
61static struct epoll_event *epoll_events;
62static int epoll_fd;
63
Willy Tarreau58094f22007-04-10 01:33:20 +020064/* This structure may be used for any purpose. Warning! do not use it in
65 * recursive functions !
66 */
67static struct epoll_event ev;
68
69/* converts a direction to a single bitmask.
70 * 0 => 1
71 * 1 => 2
72 */
73#define DIR2MSK(dir) ((dir) + 1)
74
75/* converts an FD to an fd_evts offset and to a bit shift */
76#define FD2OFS(fd) ((uint32_t)(fd) >> 4)
77#define FD2BIT(fd) (((uint32_t)(fd) & 0xF) << 1)
78#define FD2MSK(fd) (3 << FD2BIT(fd))
Willy Tarreau4f60f162007-04-08 16:39:58 +020079
80/*
Willy Tarreau58094f22007-04-10 01:33:20 +020081 * Returns non-zero if direction <dir> is already set for <fd>.
Willy Tarreau4f60f162007-04-08 16:39:58 +020082 */
Willy Tarreau63455a92007-04-09 15:34:49 +020083REGPRM2 static int __fd_is_set(const int fd, int dir)
Willy Tarreau4f60f162007-04-08 16:39:58 +020084{
Willy Tarreau58094f22007-04-10 01:33:20 +020085 return (fd_evts[FD2OFS(fd)] >> FD2BIT(fd)) & DIR2MSK(dir);
Willy Tarreau4f60f162007-04-08 16:39:58 +020086}
87
Willy Tarreau58094f22007-04-10 01:33:20 +020088/*
89 * Adds, mods or deletes <fd> according to current status and to new desired
90 * mask <dmask> :
91 *
92 * 0 = nothing
93 * 1 = EPOLLIN
94 * 2 = EPOLLOUT
95 * 3 = EPOLLIN | EPOLLOUT
96 *
97 */
98static int dmsk2event[4] = { 0, EPOLLIN, EPOLLOUT, EPOLLIN | EPOLLOUT };
99
100
101REGPRM2 static void fd_flush_changes()
Willy Tarreau4f60f162007-04-08 16:39:58 +0200102{
Willy Tarreau58094f22007-04-10 01:33:20 +0200103 uint32_t ofs;
104 int opcode;
105 int prev, next;
106 int chg, fd;
107
108 for (chg = 0; chg < nbchanges; chg++) {
109 prev = chg_list[chg].prev;
110 fd = chg_list[chg].fd;
111 chg_ptr[fd] = NULL;
112
113 ofs = FD2OFS(fd);
114 next = (fd_evts[ofs] >> FD2BIT(fd)) & 3;
115
116 if (prev == next)
117 /* if the value was unchanged, do nothing */
118 continue;
119
120 ev.events = dmsk2event[next];
121 ev.data.fd = fd;
122
123 if (prev) {
124 if (!next) {
125 /* we want to delete it now */
126 opcode = EPOLL_CTL_DEL;
127 } else {
128 /* we want to switch it */
129 opcode = EPOLL_CTL_MOD;
130 }
131 } else {
132 /* the FD did not exist, let's add it */
133 opcode = EPOLL_CTL_ADD;
134 }
135
136 epoll_ctl(epoll_fd, opcode, fd, &ev);
137 }
138 nbchanges = 0;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200139}
140
Willy Tarreau58094f22007-04-10 01:33:20 +0200141REGPRM2 static void alloc_chg_list(const int fd, int old_evt)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200142{
Willy Tarreau58094f22007-04-10 01:33:20 +0200143 struct fd_chg *ptr;
144
Willy Tarreau70bcfb72008-01-27 02:21:53 +0100145 if (unlikely(chg_ptr[fd] != NULL))
Willy Tarreau58094f22007-04-10 01:33:20 +0200146 return;
147
148#if LIMIT_NUMBER_OF_CHANGES
149 if (nbchanges > 2)
150 fd_flush_changes();
151#endif
152
153 ptr = &chg_list[nbchanges++];
154 chg_ptr[fd] = ptr;
155 ptr->fd = fd;
156 ptr->prev = old_evt;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200157}
158
Willy Tarreau58094f22007-04-10 01:33:20 +0200159REGPRM2 static int __fd_set(const int fd, int dir)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200160{
Willy Tarreau58094f22007-04-10 01:33:20 +0200161 uint32_t ofs = FD2OFS(fd);
162 uint32_t dmsk = DIR2MSK(dir);
163 uint32_t old_evt;
164
165 old_evt = fd_evts[ofs] >> FD2BIT(fd);
166 old_evt &= 3;
167 if (unlikely(old_evt & dmsk))
168 return 0;
169
170 alloc_chg_list(fd, old_evt);
171 dmsk <<= FD2BIT(fd);
172 fd_evts[ofs] |= dmsk;
173 return 1;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200174}
175
Willy Tarreau58094f22007-04-10 01:33:20 +0200176REGPRM2 static int __fd_clr(const int fd, int dir)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200177{
Willy Tarreau58094f22007-04-10 01:33:20 +0200178 uint32_t ofs = FD2OFS(fd);
179 uint32_t dmsk = DIR2MSK(dir);
180 uint32_t old_evt;
181
182 old_evt = fd_evts[ofs] >> FD2BIT(fd);
183 old_evt &= 3;
184 if (unlikely(!(old_evt & dmsk)))
185 return 0;
186
187 alloc_chg_list(fd, old_evt);
188 dmsk <<= FD2BIT(fd);
189 fd_evts[ofs] &= ~dmsk;
190 return 1;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200191}
192
Willy Tarreau58094f22007-04-10 01:33:20 +0200193REGPRM1 static void __fd_rem(int fd)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200194{
Willy Tarreau58094f22007-04-10 01:33:20 +0200195 uint32_t ofs = FD2OFS(fd);
196
197 if (unlikely(!((fd_evts[ofs] >> FD2BIT(fd)) & 3)))
198 return;
199
200 alloc_chg_list(fd, 0);
201 fd_evts[ofs] &= ~FD2MSK(fd);
202 return;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200203}
204
Willy Tarreau58094f22007-04-10 01:33:20 +0200205/*
206 * On valid epoll() implementations, a call to close() automatically removes
207 * the fds. This means that the FD will appear as previously unset.
208 */
209REGPRM1 static void __fd_clo(int fd)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200210{
Willy Tarreau58094f22007-04-10 01:33:20 +0200211 struct fd_chg *ptr;
212 fd_evts[FD2OFS(fd)] &= ~FD2MSK(fd);
213 ptr = chg_ptr[fd];
214 if (ptr) {
215 ptr->prev = 0;
216 chg_ptr[fd] = NULL;
217 }
218 return;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200219}
220
Willy Tarreau4f60f162007-04-08 16:39:58 +0200221/*
222 * epoll() poller
223 */
Willy Tarreaud825eef2007-05-12 22:35:00 +0200224REGPRM2 static void _do_poll(struct poller *p, struct timeval *exp)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200225{
226 int status;
227 int fd;
Willy Tarreau58094f22007-04-10 01:33:20 +0200228 int count;
Willy Tarreaud825eef2007-05-12 22:35:00 +0200229 int wait_time;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200230
Willy Tarreau58094f22007-04-10 01:33:20 +0200231 if (likely(nbchanges))
232 fd_flush_changes();
Willy Tarreau4f60f162007-04-08 16:39:58 +0200233
Willy Tarreau4f60f162007-04-08 16:39:58 +0200234 /* now let's wait for events */
Willy Tarreau3a628112008-06-13 21:06:56 +0200235 if (run_queue)
236 wait_time = 0;
237 else if (tv_iseternity(exp))
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200238 wait_time = MAX_DELAY_MS;
Willy Tarreaubdefc512007-05-14 02:02:04 +0200239 else if (tv_isge(&now, exp))
240 wait_time = 0;
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200241 else {
Willy Tarreaubdefc512007-05-14 02:02:04 +0200242 wait_time = __tv_ms_elapsed(&now, exp) + 1;
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200243 if (wait_time > MAX_DELAY_MS)
244 wait_time = MAX_DELAY_MS;
245 }
Willy Tarreaud825eef2007-05-12 22:35:00 +0200246
Willy Tarreau1db37712007-06-03 17:16:49 +0200247 fd = MIN(maxfd, global.tune.maxpollevents);
248 status = epoll_wait(epoll_fd, epoll_events, fd, wait_time);
Willy Tarreaub0b37bc2008-06-23 14:00:57 +0200249 tv_update_date(wait_time, status);
Willy Tarreau4f60f162007-04-08 16:39:58 +0200250
251 for (count = 0; count < status; count++) {
252 fd = epoll_events[count].data.fd;
253
Willy Tarreau58094f22007-04-10 01:33:20 +0200254 if ((fd_evts[FD2OFS(fd)] >> FD2BIT(fd)) & DIR2MSK(DIR_RD)) {
Willy Tarreau4f60f162007-04-08 16:39:58 +0200255 if (fdtab[fd].state == FD_STCLOSE)
256 continue;
257 if (epoll_events[count].events & ( EPOLLIN | EPOLLERR | EPOLLHUP ))
258 fdtab[fd].cb[DIR_RD].f(fd);
259 }
260
Willy Tarreau58094f22007-04-10 01:33:20 +0200261 if ((fd_evts[FD2OFS(fd)] >> FD2BIT(fd)) & DIR2MSK(DIR_WR)) {
Willy Tarreau4f60f162007-04-08 16:39:58 +0200262 if (fdtab[fd].state == FD_STCLOSE)
263 continue;
264 if (epoll_events[count].events & ( EPOLLOUT | EPOLLERR | EPOLLHUP ))
265 fdtab[fd].cb[DIR_WR].f(fd);
266 }
267 }
268}
269
270/*
Willy Tarreaue54e9172007-04-09 09:23:31 +0200271 * Initialization of the epoll() poller.
272 * Returns 0 in case of failure, non-zero in case of success. If it fails, it
273 * disables the poller by setting its pref to 0.
274 */
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200275REGPRM1 static int _do_init(struct poller *p)
Willy Tarreaue54e9172007-04-09 09:23:31 +0200276{
Willy Tarreau58094f22007-04-10 01:33:20 +0200277 __label__ fail_chg_ptr, fail_chg_list, fail_fdevt, fail_ee, fail_fd;
Willy Tarreaue54e9172007-04-09 09:23:31 +0200278 int fd_set_bytes;
279
280 p->private = NULL;
Willy Tarreau58094f22007-04-10 01:33:20 +0200281 fd_set_bytes = 4 * (global.maxsock + 15) / 16;
Willy Tarreaue54e9172007-04-09 09:23:31 +0200282
283 epoll_fd = epoll_create(global.maxsock + 1);
284 if (epoll_fd < 0)
285 goto fail_fd;
286
287 epoll_events = (struct epoll_event*)
Willy Tarreau1db37712007-06-03 17:16:49 +0200288 calloc(1, sizeof(struct epoll_event) * global.tune.maxpollevents);
Willy Tarreaue54e9172007-04-09 09:23:31 +0200289
290 if (epoll_events == NULL)
291 goto fail_ee;
292
Willy Tarreau58094f22007-04-10 01:33:20 +0200293 if ((fd_evts = (uint32_t *)calloc(1, fd_set_bytes)) == NULL)
294 goto fail_fdevt;
Willy Tarreaue54e9172007-04-09 09:23:31 +0200295
Willy Tarreau58094f22007-04-10 01:33:20 +0200296 chg_list = (struct fd_chg *)calloc(1, sizeof(struct fd_chg) * global.maxsock);
297 if (chg_list == NULL)
298 goto fail_chg_list;
Willy Tarreaue54e9172007-04-09 09:23:31 +0200299
Willy Tarreau58094f22007-04-10 01:33:20 +0200300 chg_ptr = (struct fd_chg **)calloc(1, sizeof(struct fd_chg*) * global.maxsock);
301 if (chg_ptr == NULL)
302 goto fail_chg_ptr;
Willy Tarreaue54e9172007-04-09 09:23:31 +0200303
304 return 1;
305
Willy Tarreau58094f22007-04-10 01:33:20 +0200306 fail_chg_ptr:
307 free(chg_list);
308 fail_chg_list:
309 free(fd_evts);
310 fail_fdevt:
Willy Tarreaue54e9172007-04-09 09:23:31 +0200311 free(epoll_events);
312 fail_ee:
313 close(epoll_fd);
314 epoll_fd = 0;
315 fail_fd:
316 p->pref = 0;
317 return 0;
318}
319
320/*
321 * Termination of the epoll() poller.
322 * Memory is released and the poller is marked as unselectable.
323 */
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200324REGPRM1 static void _do_term(struct poller *p)
Willy Tarreaue54e9172007-04-09 09:23:31 +0200325{
Willy Tarreau58094f22007-04-10 01:33:20 +0200326 fd_flush_changes();
Willy Tarreaue54e9172007-04-09 09:23:31 +0200327
Willy Tarreau58094f22007-04-10 01:33:20 +0200328 if (chg_ptr)
329 free(chg_ptr);
330 if (chg_list)
331 free(chg_list);
332 if (fd_evts)
333 free(fd_evts);
Willy Tarreaue54e9172007-04-09 09:23:31 +0200334 if (epoll_events)
335 free(epoll_events);
336
337 close(epoll_fd);
338 epoll_fd = 0;
339
Willy Tarreau58094f22007-04-10 01:33:20 +0200340 chg_ptr = NULL;
341 chg_list = NULL;
342 fd_evts = NULL;
343 epoll_events = NULL;
344
Willy Tarreaue54e9172007-04-09 09:23:31 +0200345 p->private = NULL;
346 p->pref = 0;
347}
348
349/*
Willy Tarreau2ff76222007-04-09 19:29:56 +0200350 * Check that the poller works.
351 * Returns 1 if OK, otherwise 0.
352 */
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200353REGPRM1 static int _do_test(struct poller *p)
Willy Tarreau2ff76222007-04-09 19:29:56 +0200354{
355 int fd;
356
357 fd = epoll_create(global.maxsock + 1);
358 if (fd < 0)
359 return 0;
360 close(fd);
361 return 1;
362}
363
364/*
Willy Tarreaufb8983f2007-06-03 16:40:44 +0200365 * Recreate the epoll file descriptor after a fork(). Returns 1 if OK,
366 * otherwise 0. It will ensure that all processes will not share their
367 * epoll_fd. Some side effects were encountered because of this, such
368 * as epoll_wait() returning an FD which was previously deleted.
369 */
370REGPRM1 static int _do_fork(struct poller *p)
371{
372 close(epoll_fd);
373 epoll_fd = epoll_create(global.maxsock + 1);
374 if (epoll_fd < 0)
375 return 0;
376 return 1;
377}
378
379/*
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200380 * It is a constructor, which means that it will automatically be called before
381 * main(). This is GCC-specific but it works at least since 2.95.
382 * Special care must be taken so that it does not need any uninitialized data.
Willy Tarreau4f60f162007-04-08 16:39:58 +0200383 */
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200384__attribute__((constructor))
385static void _do_register(void)
Willy Tarreau4f60f162007-04-08 16:39:58 +0200386{
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200387 struct poller *p;
388
389 if (nbpollers >= MAX_POLLERS)
390 return;
391 p = &pollers[nbpollers++];
392
Willy Tarreau4f60f162007-04-08 16:39:58 +0200393 p->name = "epoll";
394 p->pref = 300;
395 p->private = NULL;
396
Willy Tarreauef1d1f82007-04-16 00:25:25 +0200397 p->test = _do_test;
398 p->init = _do_init;
399 p->term = _do_term;
400 p->poll = _do_poll;
Willy Tarreaufb8983f2007-06-03 16:40:44 +0200401 p->fork = _do_fork;
Willy Tarreau58094f22007-04-10 01:33:20 +0200402
403 p->is_set = __fd_is_set;
404 p->cond_s = p->set = __fd_set;
405 p->cond_c = p->clr = __fd_clr;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200406 p->rem = __fd_rem;
407 p->clo = __fd_clo;
Willy Tarreau4f60f162007-04-08 16:39:58 +0200408}
409
410
411/*
412 * Local variables:
413 * c-indent-level: 8
414 * c-basic-offset: 8
415 * End:
416 */