libev/ev_epoll.c

/*
 * libev epoll fd activity backend
 *
 * Copyright (c) 2007 Marc Alexander Lehmann <libev@schmorp.de>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are
 * met:
 *
 *     * Redistributions of source code must retain the above copyright
 *       notice, this list of conditions and the following disclaimer.
 *
 *     * Redistributions in binary form must reproduce the above
 *       copyright notice, this list of conditions and the following
 *       disclaimer in the documentation and/or other materials provided
 *       with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

/*
 * general notes about epoll:
 *
 * a) epoll silently removes fds from the fd set. as nothing tells us
 *    that an fd has been removed otherwise, we have to continually
 *    "rearm" fds that we suspect *might* have changed (same
 *    problem with kqueue, but much less costly there).
 * b) the fact that ADD != MOD creates a lot of extra syscalls due to a)
 *    and seems not to have any advantage.
 * c) the inability to handle fork or file descriptors (think dup)
 *    limits the applicability over poll, so this is not a generic
 *    poll replacement.
 *
 * lots of "weird code" and complication handling in this file is due
 * to these design problems with epoll, as we try very hard to avoid
 * epoll_ctl syscalls for common usage patterns.
 */

#include <sys/epoll.h>

static void
epoll_modify (EV_P_ int fd, int oev, int nev)
{
  struct epoll_event ev;

  /*
   * we handle EPOLL_CTL_DEL by ignoring it here
   * on the assumption that the fd is gone anyways
   * if that is wrong, we have to handle the spurious
   * event in epoll_poll.
   */
  if (!nev)
    return;

  ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */
  ev.events   = (nev & EV_READ  ? EPOLLIN  : 0)
              | (nev & EV_WRITE ? EPOLLOUT : 0);

  if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev)))
    return;

  if (expect_true (errno == ENOENT))
    {
      /* on ENOENT the fd went away, so try to do the right thing */
      if (!nev)
        return;

      if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev))
        return;
    }
  else if (expect_true (errno == EEXIST))
    {
      /* on EEXIST we ignored a previous DEL */
      if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev))
        return;
    }

  fd_kill (EV_A_ fd);
}

static void
epoll_poll (EV_P_ ev_tstamp timeout)
{
  int i;
  int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.));

  if (expect_false (eventcnt < 0))
    {
      if (errno != EINTR)
        syserr ("(libev) epoll_wait");

      return;
    }

  for (i = 0; i < eventcnt; ++i)
    {
      struct epoll_event *ev = epoll_events + i;

      int fd   = ev->data.u64;
      int got  = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
               | (ev->events & (EPOLLIN  | EPOLLERR | EPOLLHUP) ? EV_READ  : 0);
      int want = anfds [fd].events;

      if (expect_false (got & ~want))
        {
          /* we received an event but are not interested in it, try mod or del */
          ev->events = (want & EV_READ  ? EPOLLIN  : 0)
                     | (want & EV_WRITE ? EPOLLOUT : 0);

          epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev);
        }

      fd_event (EV_A_ fd, got);
    }

  /* if the receive array was full, increase its size */
  if (expect_false (eventcnt == epoll_eventmax))
    {
      ev_free (epoll_events);
      epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1);
      epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax);
    }
}

int inline_size
epoll_init (EV_P_ int flags)
{
  backend_fd = epoll_create (256);

  if (backend_fd < 0)
    return 0;

  fcntl (backend_fd, F_SETFD, FD_CLOEXEC);

  backend_fudge  = 0.; /* kernel sources seem to indicate this to be zero */
  backend_modify = epoll_modify;
  backend_poll   = epoll_poll;

  epoll_eventmax = 64; /* intiial number of events receivable per poll */
  epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax);

  return EVBACKEND_EPOLL;
}

void inline_size
epoll_destroy (EV_P)
{
  ev_free (epoll_events);
}

void inline_size
epoll_fork (EV_P)
{
  close (backend_fd);

  while ((backend_fd = epoll_create (256)) < 0)
    syserr ("(libev) epoll_create");

  fcntl (backend_fd, F_SETFD, FD_CLOEXEC);

  fd_rearm_all (EV_A);
}

Revision:	1.32
Committed:	Thu Dec 13 16:52:50 2007 UTC (16 years, 5 months ago) by root
Content type:	text/plain
Branch:	MAIN
CVS Tags:	rel-2_0, rel-1_86, rel-1_85
Changes since 1.31:	+70 -16 lines
Log Message:	* empty log message *
#	User	Rev	Content
1	root	1.5	/*
2	root	1.9	* libev epoll fd activity backend
3			*
4	root	1.5	* Copyright (c) 2007 Marc Alexander Lehmann <libev@schmorp.de>
5			* All rights reserved.
6			*
7			* Redistribution and use in source and binary forms, with or without
8			* modification, are permitted provided that the following conditions are
9			* met:
10			*
11			* * Redistributions of source code must retain the above copyright
12			* notice, this list of conditions and the following disclaimer.
13			*
14			* * Redistributions in binary form must reproduce the above
15			* copyright notice, this list of conditions and the following
16			* disclaimer in the documentation and/or other materials provided
17			* with the distribution.
18			*
19			* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20			* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21			* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22			* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23			* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24			* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
25			* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26			* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27			* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28			* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29			* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30			*/
31
32	root	1.32	/*
33			* general notes about epoll:
34			*
35			* a) epoll silently removes fds from the fd set. as nothing tells us
36			* that an fd has been removed otherwise, we have to continually
37			* "rearm" fds that we suspect might have changed (same
38			* problem with kqueue, but much less costly there).
39			* b) the fact that ADD != MOD creates a lot of extra syscalls due to a)
40			* and seems not to have any advantage.
41			* c) the inability to handle fork or file descriptors (think dup)
42			* limits the applicability over poll, so this is not a generic
43			* poll replacement.
44			*
45			* lots of "weird code" and complication handling in this file is due
46			* to these design problems with epoll, as we try very hard to avoid
47			* epoll_ctl syscalls for common usage patterns.
48			*/
49
50	root	1.1	#include <sys/epoll.h>
51
52			static void
53	root	1.13	epoll_modify (EV_P_ int fd, int oev, int nev)
54	root	1.1	{
55	root	1.32	struct epoll_event ev;
56
57			/*
58			* we handle EPOLL_CTL_DEL by ignoring it here
59			* on the assumption that the fd is gone anyways
60			* if that is wrong, we have to handle the spurious
61			* event in epoll_poll.
62			*/
63			if (!nev)
64			return;
65	root	1.1
66	root	1.11	ev.data.u64 = fd; /* use u64 to fully initialise the struct, for nicer strace etc. */
67	root	1.32	ev.events = (nev & EV_READ ? EPOLLIN : 0)
68			\| (nev & EV_WRITE ? EPOLLOUT : 0);
69
70			if (expect_true (!epoll_ctl (backend_fd, oev ? EPOLL_CTL_MOD : EPOLL_CTL_ADD, fd, &ev)))
71			return;
72
73			if (expect_true (errno == ENOENT))
74			{
75			/* on ENOENT the fd went away, so try to do the right thing */
76			if (!nev)
77			return;
78
79			if (!epoll_ctl (backend_fd, EPOLL_CTL_ADD, fd, &ev))
80			return;
81			}
82			else if (expect_true (errno == EEXIST))
83			{
84			/* on EEXIST we ignored a previous DEL */
85			if (!epoll_ctl (backend_fd, EPOLL_CTL_MOD, fd, &ev))
86			return;
87			}
88
89			fd_kill (EV_A_ fd);
90	root	1.1	}
91
92	root	1.6	static void
93	root	1.13	epoll_poll (EV_P_ ev_tstamp timeout)
94	root	1.1	{
95	root	1.20	int i;
96	root	1.26	int eventcnt = epoll_wait (backend_fd, epoll_events, epoll_eventmax, (int)ceil (timeout * 1000.));
97	root	1.1
98	root	1.29	if (expect_false (eventcnt < 0))
99	root	1.20	{
100			if (errno != EINTR)
101	root	1.21	syserr ("(libev) epoll_wait");
102	root	1.20
103			return;
104			}
105	root	1.1
106			for (i = 0; i < eventcnt; ++i)
107	root	1.32	{
108			struct epoll_event *ev = epoll_events + i;
109
110			int fd = ev->data.u64;
111			int got = (ev->events & (EPOLLOUT \| EPOLLERR \| EPOLLHUP) ? EV_WRITE : 0)
112			\| (ev->events & (EPOLLIN \| EPOLLERR \| EPOLLHUP) ? EV_READ : 0);
113			int want = anfds [fd].events;
114
115			if (expect_false (got & ~want))
116			{
117			/* we received an event but are not interested in it, try mod or del */
118			ev->events = (want & EV_READ ? EPOLLIN : 0)
119			\| (want & EV_WRITE ? EPOLLOUT : 0);
120
121			epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev);
122			}
123
124			fd_event (EV_A_ fd, got);
125			}
126	root	1.1
127			/* if the receive array was full, increase its size */
128	root	1.14	if (expect_false (eventcnt == epoll_eventmax))
129	root	1.1	{
130	root	1.20	ev_free (epoll_events);
131	root	1.28	epoll_eventmax = array_nextsize (sizeof (struct epoll_event), epoll_eventmax, epoll_eventmax + 1);
132	root	1.23	epoll_events = (struct epoll_event )ev_malloc (sizeof (struct epoll_event) epoll_eventmax);
133	root	1.1	}
134			}
135
136	root	1.27	int inline_size
137	root	1.13	epoll_init (EV_P_ int flags)
138	root	1.1	{
139	root	1.26	backend_fd = epoll_create (256);
140	root	1.1
141	root	1.26	if (backend_fd < 0)
142	root	1.13	return 0;
143	root	1.3
144	root	1.26	fcntl (backend_fd, F_SETFD, FD_CLOEXEC);
145	root	1.1
146	root	1.31	backend_fudge = 0.; /* kernel sources seem to indicate this to be zero */
147	root	1.25	backend_modify = epoll_modify;
148			backend_poll = epoll_poll;
149	root	1.1
150	root	1.14	epoll_eventmax = 64; /* intiial number of events receivable per poll */
151	root	1.23	epoll_events = (struct epoll_event )ev_malloc (sizeof (struct epoll_event) epoll_eventmax);
152	root	1.13
153	root	1.24	return EVBACKEND_EPOLL;
154	root	1.1	}
155	root	1.6
156	root	1.27	void inline_size
157	root	1.15	epoll_destroy (EV_P)
158			{
159	root	1.20	ev_free (epoll_events);
160	root	1.15	}
161
162	root	1.27	void inline_size
163	root	1.15	epoll_fork (EV_P)
164			{
165	root	1.26	close (backend_fd);
166	root	1.21
167	root	1.26	while ((backend_fd = epoll_create (256)) < 0)
168	root	1.22	syserr ("(libev) epoll_create");
169	root	1.21
170	root	1.26	fcntl (backend_fd, F_SETFD, FD_CLOEXEC);
171	root	1.15
172	root	1.16	fd_rearm_all (EV_A);
173	root	1.15	}
174