libev/ev.pod

=encoding utf-8

=head1 NAME

libev - a high performance full-featured event loop written in C

=head1 SYNOPSIS

   #include <ev.h>

=head2 EXAMPLE PROGRAM

   // a single header file is required
   #include <ev.h>

   #include <stdio.h> // for puts

   // every watcher type has its own typedef'd struct
   // with the name ev_TYPE
   ev_io stdin_watcher;
   ev_timer timeout_watcher;

   // all watcher callbacks have a similar signature
   // this callback is called when data is readable on stdin
   static void
   stdin_cb (EV_P_ ev_io *w, int revents)
   {
     puts ("stdin ready");
     // for one-shot events, one must manually stop the watcher
     // with its corresponding stop function.
     ev_io_stop (EV_A_ w);

     // this causes all nested ev_run's to stop iterating
     ev_break (EV_A_ EVBREAK_ALL);
   }

   // another callback, this time for a time-out
   static void
   timeout_cb (EV_P_ ev_timer *w, int revents)
   {
     puts ("timeout");
     // this causes the innermost ev_run to stop iterating
     ev_break (EV_A_ EVBREAK_ONE);
   }

   int
   main (void)
   {
     // use the default event loop unless you have special needs
     struct ev_loop *loop = EV_DEFAULT;

     // initialise an io watcher, then start it
     // this one will watch for stdin to become readable
     ev_io_init (&stdin_watcher, stdin_cb, /*STDIN_FILENO*/ 0, EV_READ);
     ev_io_start (loop, &stdin_watcher);

     // initialise a timer watcher, then start it
     // simple non-repeating 5.5 second timeout
     ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
     ev_timer_start (loop, &timeout_watcher);

     // now wait for events to arrive
     ev_run (loop, 0);

     // break was called, so exit
     return 0;
   }

=head1 ABOUT THIS DOCUMENT

This document documents the libev software package.

The newest version of this document is also available as an html-formatted
web page you might find easier to navigate when reading it for the first
time: L<http://pod.tst.eu/http://cvs.schmorp.de/libev/ev.pod>.

While this document tries to be as complete as possible in documenting
libev, its usage and the rationale behind its design, it is not a tutorial
on event-based programming, nor will it introduce event-based programming
with libev.

Familiarity with event based programming techniques in general is assumed
throughout this document.

=head1 WHAT TO READ WHEN IN A HURRY

This manual tries to be very detailed, but unfortunately, this also makes
it very long. If you just want to know the basics of libev, I suggest
reading L</ANATOMY OF A WATCHER>, then the L</EXAMPLE PROGRAM> above and
look up the missing functions in L</GLOBAL FUNCTIONS> and the C<ev_io> and
C<ev_timer> sections in L</WATCHER TYPES>.

=head1 ABOUT LIBEV

Libev is an event loop: you register interest in certain events (such as a
file descriptor being readable or a timeout occurring), and it will manage
these event sources and provide your program with events.

To do this, it must take more or less complete control over your process
(or thread) by executing the I<event loop> handler, and will then
communicate events via a callback mechanism.

You register interest in certain events by registering so-called I<event
watchers>, which are relatively small C structures you initialise with the
details of the event, and then hand it over to libev by I<starting> the
watcher.

=head2 FEATURES

Libev supports C<select>, C<poll>, the Linux-specific aio and C<epoll>
interfaces, the BSD-specific C<kqueue> and the Solaris-specific event port
mechanisms for file descriptor events (C<ev_io>), the Linux C<inotify>
interface (for C<ev_stat>), Linux eventfd/signalfd (for faster and cleaner
inter-thread wakeup (C<ev_async>)/signal handling (C<ev_signal>)) relative
timers (C<ev_timer>), absolute timers with customised rescheduling
(C<ev_periodic>), synchronous signals (C<ev_signal>), process status
change events (C<ev_child>), and event watchers dealing with the event
loop mechanism itself (C<ev_idle>, C<ev_embed>, C<ev_prepare> and
C<ev_check> watchers) as well as file watchers (C<ev_stat>) and even
limited support for fork events (C<ev_fork>).

It also is quite fast (see this
L<benchmark|http://libev.schmorp.de/bench.html> comparing it to libevent
for example).

=head2 CONVENTIONS

Libev is very configurable. In this manual the default (and most common)
configuration will be described, which supports multiple event loops. For
more info about various configuration options please have a look at
B<EMBED> section in this manual. If libev was configured without support
for multiple event loops, then all functions taking an initial argument of
name C<loop> (which is always of type C<struct ev_loop *>) will not have
this argument.

=head2 TIME REPRESENTATION

Libev represents time as a single floating point number, representing
the (fractional) number of seconds since the (POSIX) epoch (in practice
somewhere near the beginning of 1970, details are complicated, don't
ask). This type is called C<ev_tstamp>, which is what you should use
too. It usually aliases to the C<double> type in C. When you need to do
any calculations on it, you should treat it as some floating point value.

Unlike the name component C<stamp> might indicate, it is also used for
time differences (e.g. delays) throughout libev.

=head1 ERROR HANDLING

Libev knows three classes of errors: operating system errors, usage errors
and internal errors (bugs).

When libev catches an operating system error it cannot handle (for example
a system call indicating a condition libev cannot fix), it calls the callback
set via C<ev_set_syserr_cb>, which is supposed to fix the problem or
abort. The default is to print a diagnostic message and to call C<abort
()>.

When libev detects a usage error such as a negative timer interval, then
it will print a diagnostic message and abort (via the C<assert> mechanism,
so C<NDEBUG> will disable this checking): these are programming errors in
the libev caller and need to be fixed there.

Via the C<EV_FREQUENT> macro you can compile in and/or enable extensive
consistency checking code inside libev that can be used to check for
internal inconsistencies, suually caused by application bugs.

Libev also has a few internal error-checking C<assert>ions. These do not
trigger under normal circumstances, as they indicate either a bug in libev
or worse.


=head1 GLOBAL FUNCTIONS

These functions can be called anytime, even before initialising the
library in any way.

=over 4

=item ev_tstamp ev_time ()

Returns the current time as libev would use it. Please note that the
C<ev_now> function is usually faster and also often returns the timestamp
you actually want to know. Also interesting is the combination of
C<ev_now_update> and C<ev_now>.

=item ev_sleep (ev_tstamp interval)

Sleep for the given interval: The current thread will be blocked
until either it is interrupted or the given time interval has
passed (approximately - it might return a bit earlier even if not
interrupted). Returns immediately if C<< interval <= 0 >>.

Basically this is a sub-second-resolution C<sleep ()>.

The range of the C<interval> is limited - libev only guarantees to work
with sleep times of up to one day (C<< interval <= 86400 >>).

=item int ev_version_major ()

=item int ev_version_minor ()

You can find out the major and minor ABI version numbers of the library
you linked against by calling the functions C<ev_version_major> and
C<ev_version_minor>. If you want, you can compare against the global
symbols C<EV_VERSION_MAJOR> and C<EV_VERSION_MINOR>, which specify the
version of the library your program was compiled against.

These version numbers refer to the ABI version of the library, not the
release version.

Usually, it's a good idea to terminate if the major versions mismatch,
as this indicates an incompatible change. Minor versions are usually
compatible to older versions, so a larger minor version alone is usually
not a problem.

Example: Make sure we haven't accidentally been linked against the wrong
version (note, however, that this will not detect other ABI mismatches,
such as LFS or reentrancy).

   assert (("libev version mismatch",
            ev_version_major () == EV_VERSION_MAJOR
            && ev_version_minor () >= EV_VERSION_MINOR));

=item unsigned int ev_supported_backends ()

Return the set of all backends (i.e. their corresponding C<EV_BACKEND_*>
value) compiled into this binary of libev (independent of their
availability on the system you are running on). See C<ev_default_loop> for
a description of the set values.

Example: make sure we have the epoll method, because yeah this is cool and
a must have and can we have a torrent of it please!!!11

   assert (("sorry, no epoll, no sex",
            ev_supported_backends () & EVBACKEND_EPOLL));

=item unsigned int ev_recommended_backends ()

Return the set of all backends compiled into this binary of libev and
also recommended for this platform, meaning it will work for most file
descriptor types. This set is often smaller than the one returned by
C<ev_supported_backends>, as for example kqueue is broken on most BSDs
and will not be auto-detected unless you explicitly request it (assuming
you know what you are doing). This is the set of backends that libev will
probe for if you specify no backends explicitly.

=item unsigned int ev_embeddable_backends ()

Returns the set of backends that are embeddable in other event loops. This
value is platform-specific but can include backends not available on the
current system. To find which embeddable backends might be supported on
the current system, you would need to look at C<ev_embeddable_backends ()
& ev_supported_backends ()>, likewise for recommended ones.

See the description of C<ev_embed> watchers for more info.

=item ev_set_allocator (void *(*cb)(void *ptr, long size) throw ())

Sets the allocation function to use (the prototype is similar - the
semantics are identical to the C<realloc> C89/SuS/POSIX function). It is
used to allocate and free memory (no surprises here). If it returns zero
when memory needs to be allocated (C<size != 0>), the library might abort
or take some potentially destructive action.

Since some systems (at least OpenBSD and Darwin) fail to implement
correct C<realloc> semantics, libev will use a wrapper around the system
C<realloc> and C<free> functions by default.

You could override this function in high-availability programs to, say,
free some memory if it cannot allocate memory, to use a special allocator,
or even to sleep a while and retry until some memory is available.

Example: The following is the C<realloc> function that libev itself uses
which should work with C<realloc> and C<free> functions of all kinds and
is probably a good basis for your own implementation.

   static void *
   ev_realloc_emul (void *ptr, long size) EV_NOEXCEPT
   {
     if (size)
       return realloc (ptr, size);

     free (ptr);
     return 0;
   }

Example: Replace the libev allocator with one that waits a bit and then
retries.

   static void *
   persistent_realloc (void *ptr, size_t size)
   {
     if (!size)
       {
         free (ptr);
         return 0;
       }

     for (;;)
       {
         void *newptr = realloc (ptr, size);

         if (newptr)
           return newptr;

         sleep (60);
       }
   }

   ...
   ev_set_allocator (persistent_realloc);

=item ev_set_syserr_cb (void (*cb)(const char *msg) throw ())

Set the callback function to call on a retryable system call error (such
as failed select, poll, epoll_wait). The message is a printable string
indicating the system call or subsystem causing the problem. If this
callback is set, then libev will expect it to remedy the situation, no
matter what, when it returns. That is, libev will generally retry the
requested operation, or, if the condition doesn't go away, do bad stuff
(such as abort).

Example: This is basically the same thing that libev does internally, too.

   static void
   fatal_error (const char *msg)
   {
     perror (msg);
     abort ();
   }

   ...
   ev_set_syserr_cb (fatal_error);

=item ev_feed_signal (int signum)

This function can be used to "simulate" a signal receive. It is completely
safe to call this function at any time, from any context, including signal
handlers or random threads.

Its main use is to customise signal handling in your process, especially
in the presence of threads. For example, you could block signals
by default in all threads (and specifying C<EVFLAG_NOSIGMASK> when
creating any loops), and in one thread, use C<sigwait> or any other
mechanism to wait for signals, then "deliver" them to libev by calling
C<ev_feed_signal>.

=back

=head1 FUNCTIONS CONTROLLING EVENT LOOPS

An event loop is described by a C<struct ev_loop *> (the C<struct> is
I<not> optional in this case unless libev 3 compatibility is disabled, as
libev 3 had an C<ev_loop> function colliding with the struct name).

The library knows two types of such loops, the I<default> loop, which
supports child process events, and dynamically created event loops which
do not.

=over 4

=item struct ev_loop *ev_default_loop (unsigned int flags)

This returns the "default" event loop object, which is what you should
normally use when you just need "the event loop". Event loop objects and
the C<flags> parameter are described in more detail in the entry for
C<ev_loop_new>.

If the default loop is already initialised then this function simply
returns it (and ignores the flags. If that is troubling you, check
C<ev_backend ()> afterwards). Otherwise it will create it with the given
flags, which should almost always be C<0>, unless the caller is also the
one calling C<ev_run> or otherwise qualifies as "the main program".

If you don't know what event loop to use, use the one returned from this
function (or via the C<EV_DEFAULT> macro).

Note that this function is I<not> thread-safe, so if you want to use it
from multiple threads, you have to employ some kind of mutex (note also
that this case is unlikely, as loops cannot be shared easily between
threads anyway).

The default loop is the only loop that can handle C<ev_child> watchers,
and to do this, it always registers a handler for C<SIGCHLD>. If this is
a problem for your application you can either create a dynamic loop with
C<ev_loop_new> which doesn't do that, or you can simply overwrite the
C<SIGCHLD> signal handler I<after> calling C<ev_default_init>.

Example: This is the most typical usage.

   if (!ev_default_loop (0))
     fatal ("could not initialise libev, bad $LIBEV_FLAGS in environment?");

Example: Restrict libev to the select and poll backends, and do not allow
environment settings to be taken into account:

   ev_default_loop (EVBACKEND_POLL | EVBACKEND_SELECT | EVFLAG_NOENV);

=item struct ev_loop *ev_loop_new (unsigned int flags)

This will create and initialise a new event loop object. If the loop
could not be initialised, returns false.

This function is thread-safe, and one common way to use libev with
threads is indeed to create one loop per thread, and using the default
loop in the "main" or "initial" thread.

The flags argument can be used to specify special behaviour or specific
backends to use, and is usually specified as C<0> (or C<EVFLAG_AUTO>).

The following flags are supported:

=over 4

=item C<EVFLAG_AUTO>

The default flags value. Use this if you have no clue (it's the right
thing, believe me).

=item C<EVFLAG_NOENV>

If this flag bit is or'ed into the flag value (or the program runs setuid
or setgid) then libev will I<not> look at the environment variable
C<LIBEV_FLAGS>. Otherwise (the default), this environment variable will
override the flags completely if it is found in the environment. This is
useful to try out specific backends to test their performance, to work
around bugs, or to make libev threadsafe (accessing environment variables
cannot be done in a threadsafe way, but usually it works if no other
thread modifies them).

=item C<EVFLAG_FORKCHECK>

Instead of calling C<ev_loop_fork> manually after a fork, you can also
make libev check for a fork in each iteration by enabling this flag.

This works by calling C<getpid ()> on every iteration of the loop,
and thus this might slow down your event loop if you do a lot of loop
iterations and little real work, but is usually not noticeable (on my
GNU/Linux system for example, C<getpid> is actually a simple 5-insn
sequence without a system call and thus I<very> fast, but my GNU/Linux
system also has C<pthread_atfork> which is even faster). (Update: glibc
versions 2.25 apparently removed the C<getpid> optimisation again).

The big advantage of this flag is that you can forget about fork (and
forget about forgetting to tell libev about forking, although you still
have to ignore C<SIGPIPE>) when you use this flag.

This flag setting cannot be overridden or specified in the C<LIBEV_FLAGS>
environment variable.

=item C<EVFLAG_NOINOTIFY>

When this flag is specified, then libev will not attempt to use the
I<inotify> API for its C<ev_stat> watchers. Apart from debugging and
testing, this flag can be useful to conserve inotify file descriptors, as
otherwise each loop using C<ev_stat> watchers consumes one inotify handle.

=item C<EVFLAG_SIGNALFD>

When this flag is specified, then libev will attempt to use the
I<signalfd> API for its C<ev_signal> (and C<ev_child>) watchers. This API
delivers signals synchronously, which makes it both faster and might make
it possible to get the queued signal data. It can also simplify signal
handling with threads, as long as you properly block signals in your
threads that are not interested in handling them.

Signalfd will not be used by default as this changes your signal mask, and
there are a lot of shoddy libraries and programs (glib's threadpool for
example) that can't properly initialise their signal masks.

=item C<EVFLAG_NOSIGMASK>

When this flag is specified, then libev will avoid to modify the signal
mask. Specifically, this means you have to make sure signals are unblocked
when you want to receive them.

This behaviour is useful when you want to do your own signal handling, or
want to handle signals only in specific threads and want to avoid libev
unblocking the signals.

It's also required by POSIX in a threaded program, as libev calls
C<sigprocmask>, whose behaviour is officially unspecified.

=item C<EVFLAG_NOTIMERFD>

When this flag is specified, the libev will avoid using a C<timerfd> to
detect time jumps. It will still be able to detect time jumps, but takes
longer and has a lower accuracy in doing so, but saves a file descriptor
per loop.

The current implementation only tries to use a C<timerfd> when the first
C<ev_periodic> watcher is started and falls back on other methods if it
cannot be created, but this behaviour might change in the future.

=item C<EVBACKEND_SELECT>  (value 1, portable select backend)

This is your standard select(2) backend. Not I<completely> standard, as
libev tries to roll its own fd_set with no limits on the number of fds,
but if that fails, expect a fairly low limit on the number of fds when
using this backend. It doesn't scale too well (O(highest_fd)), but its
usually the fastest backend for a low number of (low-numbered :) fds.

To get good performance out of this backend you need a high amount of
parallelism (most of the file descriptors should be busy). If you are
writing a server, you should C<accept ()> in a loop to accept as many
connections as possible during one iteration. You might also want to have
a look at C<ev_set_io_collect_interval ()> to increase the amount of
readiness notifications you get per iteration.

This backend maps C<EV_READ> to the C<readfds> set and C<EV_WRITE> to the
C<writefds> set (and to work around Microsoft Windows bugs, also onto the
C<exceptfds> set on that platform).

=item C<EVBACKEND_POLL>    (value 2, poll backend, available everywhere except on windows)

And this is your standard poll(2) backend. It's more complicated
than select, but handles sparse fds better and has no artificial
limit on the number of fds you can use (except it will slow down
considerably with a lot of inactive fds). It scales similarly to select,
i.e. O(total_fds). See the entry for C<EVBACKEND_SELECT>, above, for
performance tips.

This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and
C<EV_WRITE> to C<POLLOUT | POLLERR | POLLHUP>.

=item C<EVBACKEND_EPOLL>   (value 4, Linux)

Use the Linux-specific epoll(7) interface (for both pre- and post-2.6.9
kernels).

For few fds, this backend is a bit little slower than poll and select, but
it scales phenomenally better. While poll and select usually scale like
O(total_fds) where total_fds is the total number of fds (or the highest
fd), epoll scales either O(1) or O(active_fds).

The epoll mechanism deserves honorable mention as the most misdesigned
of the more advanced event mechanisms: mere annoyances include silently
dropping file descriptors, requiring a system call per change per file
descriptor (and unnecessary guessing of parameters), problems with dup,
returning before the timeout value, resulting in additional iterations
(and only giving 5ms accuracy while select on the same platform gives
0.1ms) and so on. The biggest issue is fork races, however - if a program
forks then I<both> parent and child process have to recreate the epoll
set, which can take considerable time (one syscall per file descriptor)
and is of course hard to detect.

Epoll is also notoriously buggy - embedding epoll fds I<should> work,
but of course I<doesn't>, and epoll just loves to report events for
totally I<different> file descriptors (even already closed ones, so
one cannot even remove them from the set) than registered in the set
(especially on SMP systems). Libev tries to counter these spurious
notifications by employing an additional generation counter and comparing
that against the events to filter out spurious ones, recreating the set
when required. Epoll also erroneously rounds down timeouts, but gives you
no way to know when and by how much, so sometimes you have to busy-wait
because epoll returns immediately despite a nonzero timeout. And last
not least, it also refuses to work with some file descriptors which work
perfectly fine with C<select> (files, many character devices...).

Epoll is truly the train wreck among event poll mechanisms, a frankenpoll,
cobbled together in a hurry, no thought to design or interaction with
others. Oh, the pain, will it ever stop...

While stopping, setting and starting an I/O watcher in the same iteration
will result in some caching, there is still a system call per such
incident (because the same I<file descriptor> could point to a different
I<file description> now), so its best to avoid that. Also, C<dup ()>'ed
file descriptors might not work very well if you register events for both
file descriptors.

Best performance from this backend is achieved by not unregistering all
watchers for a file descriptor until it has been closed, if possible,
i.e. keep at least one watcher active per fd at all times. Stopping and
starting a watcher (without re-setting it) also usually doesn't cause
extra overhead. A fork can both result in spurious notifications as well
as in libev having to destroy and recreate the epoll object, which can
take considerable time and thus should be avoided.

All this means that, in practice, C<EVBACKEND_SELECT> can be as fast or
faster than epoll for maybe up to a hundred file descriptors, depending on
the usage. So sad.

While nominally embeddable in other event loops, this feature is broken in
a lot of kernel revisions, but probably(!) works in current versions.

This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
C<EVBACKEND_POLL>.

=item C<EVBACKEND_LINUXAIO>   (value 64, Linux)

Use the Linux-specific Linux AIO (I<not> C<< aio(7) >> but C<<
io_submit(2) >>) event interface available in post-4.18 kernels (but libev
only tries to use it in 4.19+).

This is another Linux train wreck of an event interface.

If this backend works for you (as of this writing, it was very
experimental), it is the best event interface available on Linux and might
be well worth enabling it - if it isn't available in your kernel this will
be detected and this backend will be skipped.

This backend can batch oneshot requests and supports a user-space ring
buffer to receive events. It also doesn't suffer from most of the design
problems of epoll (such as not being able to remove event sources from
the epoll set), and generally sounds too good to be true. Because, this
being the Linux kernel, of course it suffers from a whole new set of
limitations, forcing you to fall back to epoll, inheriting all its design
issues.

For one, it is not easily embeddable (but probably could be done using
an event fd at some extra overhead). It also is subject to a system wide
limit that can be configured in F</proc/sys/fs/aio-max-nr>. If no AIO
requests are left, this backend will be skipped during initialisation, and
will switch to epoll when the loop is active.

Most problematic in practice, however, is that not all file descriptors
work with it. For example, in Linux 5.1, TCP sockets, pipes, event fds,
files, F</dev/null> and many others are supported, but ttys do not work
properly (a known bug that the kernel developers don't care about, see
L<https://lore.kernel.org/patchwork/patch/1047453/>), so this is not
(yet?) a generic event polling interface.

Overall, it seems the Linux developers just don't want it to have a
generic event handling mechanism other than C<select> or C<poll>.

To work around all these problem, the current version of libev uses its
epoll backend as a fallback for file descriptor types that do not work. Or
falls back completely to epoll if the kernel acts up.

This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
C<EVBACKEND_POLL>.

=item C<EVBACKEND_KQUEUE>  (value 8, most BSD clones)

Kqueue deserves special mention, as at the time this backend was
implemented, it was broken on all BSDs except NetBSD (usually it doesn't
work reliably with anything but sockets and pipes, except on Darwin,
where of course it's completely useless). Unlike epoll, however, whose
brokenness is by design, these kqueue bugs can be (and mostly have been)
fixed without API changes to existing programs. For this reason it's not
being "auto-detected" on all platforms unless you explicitly specify it
in the flags (i.e. using C<EVBACKEND_KQUEUE>) or libev was compiled on a
known-to-be-good (-enough) system like NetBSD.

You still can embed kqueue into a normal poll or select backend and use it
only for sockets (after having made sure that sockets work with kqueue on
the target platform). See C<ev_embed> watchers for more info.

It scales in the same way as the epoll backend, but the interface to the
kernel is more efficient (which says nothing about its actual speed, of
course). While stopping, setting and starting an I/O watcher does never
cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to
two event changes per incident. Support for C<fork ()> is very bad (you
might have to leak fds on fork, but it's more sane than epoll) and it
drops fds silently in similarly hard-to-detect cases.

This backend usually performs well under most conditions.

While nominally embeddable in other event loops, this doesn't work
everywhere, so you might need to test for this. And since it is broken
almost everywhere, you should only use it when you have a lot of sockets
(for which it usually works), by embedding it into another event loop
(e.g. C<EVBACKEND_SELECT> or C<EVBACKEND_POLL> (but C<poll> is of course
also broken on OS X)) and, did I mention it, using it only for sockets.

This backend maps C<EV_READ> into an C<EVFILT_READ> kevent with
C<NOTE_EOF>, and C<EV_WRITE> into an C<EVFILT_WRITE> kevent with
C<NOTE_EOF>.

=item C<EVBACKEND_DEVPOLL> (value 16, Solaris 8)

This is not implemented yet (and might never be, unless you send me an
implementation). According to reports, C</dev/poll> only supports sockets
and is not embeddable, which would limit the usefulness of this backend
immensely.

=item C<EVBACKEND_PORT>    (value 32, Solaris 10)

This uses the Solaris 10 event port mechanism. As with everything on Solaris,
it's really slow, but it still scales very well (O(active_fds)).

While this backend scales well, it requires one system call per active
file descriptor per loop iteration. For small and medium numbers of file
descriptors a "slow" C<EVBACKEND_SELECT> or C<EVBACKEND_POLL> backend
might perform better.

On the positive side, this backend actually performed fully to
specification in all tests and is fully embeddable, which is a rare feat
among the OS-specific backends (I vastly prefer correctness over speed
hacks).

On the negative side, the interface is I<bizarre> - so bizarre that
even sun itself gets it wrong in their code examples: The event polling
function sometimes returns events to the caller even though an error
occurred, but with no indication whether it has done so or not (yes, it's
even documented that way) - deadly for edge-triggered interfaces where you
absolutely have to know whether an event occurred or not because you have
to re-arm the watcher.

Fortunately libev seems to be able to work around these idiocies.

This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
C<EVBACKEND_POLL>.

=item C<EVBACKEND_ALL>

Try all backends (even potentially broken ones that wouldn't be tried
with C<EVFLAG_AUTO>). Since this is a mask, you can do stuff such as
C<EVBACKEND_ALL & ~EVBACKEND_KQUEUE>.

It is definitely not recommended to use this flag, use whatever
C<ev_recommended_backends ()> returns, or simply do not specify a backend
at all.

=item C<EVBACKEND_MASK>

Not a backend at all, but a mask to select all backend bits from a
C<flags> value, in case you want to mask out any backends from a flags
value (e.g. when modifying the C<LIBEV_FLAGS> environment variable).

=back

If one or more of the backend flags are or'ed into the flags value,
then only these backends will be tried (in the reverse order as listed
here). If none are specified, all backends in C<ev_recommended_backends
()> will be tried.

Example: Try to create a event loop that uses epoll and nothing else.

   struct ev_loop *epoller = ev_loop_new (EVBACKEND_EPOLL | EVFLAG_NOENV);
   if (!epoller)
     fatal ("no epoll found here, maybe it hides under your chair");

Example: Use whatever libev has to offer, but make sure that kqueue is
used if available.

   struct ev_loop *loop = ev_loop_new (ev_recommended_backends () | EVBACKEND_KQUEUE);

Example: Similarly, on linux, you mgiht want to take advantage of the
linux aio backend if possible, but fall back to something else if that
isn't available.

   struct ev_loop *loop = ev_loop_new (ev_recommended_backends () | EVBACKEND_LINUXAIO);

=item ev_loop_destroy (loop)

Destroys an event loop object (frees all memory and kernel state
etc.). None of the active event watchers will be stopped in the normal
sense, so e.g. C<ev_is_active> might still return true. It is your
responsibility to either stop all watchers cleanly yourself I<before>
calling this function, or cope with the fact afterwards (which is usually
the easiest thing, you can just ignore the watchers and/or C<free ()> them
for example).

Note that certain global state, such as signal state (and installed signal
handlers), will not be freed by this function, and related watchers (such
as signal and child watchers) would need to be stopped manually.

This function is normally used on loop objects allocated by
C<ev_loop_new>, but it can also be used on the default loop returned by
C<ev_default_loop>, in which case it is not thread-safe.

Note that it is not advisable to call this function on the default loop
except in the rare occasion where you really need to free its resources.
If you need dynamically allocated loops it is better to use C<ev_loop_new>
and C<ev_loop_destroy>.

=item ev_loop_fork (loop)

This function sets a flag that causes subsequent C<ev_run> iterations
to reinitialise the kernel state for backends that have one. Despite
the name, you can call it anytime you are allowed to start or stop
watchers (except inside an C<ev_prepare> callback), but it makes most
sense after forking, in the child process. You I<must> call it (or use
C<EVFLAG_FORKCHECK>) in the child before resuming or calling C<ev_run>.

In addition, if you want to reuse a loop (via this function or
C<EVFLAG_FORKCHECK>), you I<also> have to ignore C<SIGPIPE>.

Again, you I<have> to call it on I<any> loop that you want to re-use after
a fork, I<even if you do not plan to use the loop in the parent>. This is
because some kernel interfaces *cough* I<kqueue> *cough* do funny things
during fork.

On the other hand, you only need to call this function in the child
process if and only if you want to use the event loop in the child. If
you just fork+exec or create a new loop in the child, you don't have to
call it at all (in fact, C<epoll> is so badly broken that it makes a
difference, but libev will usually detect this case on its own and do a
costly reset of the backend).

The function itself is quite fast and it's usually not a problem to call
it just in case after a fork.

Example: Automate calling C<ev_loop_fork> on the default loop when
using pthreads.

   static void
   post_fork_child (void)
   {
     ev_loop_fork (EV_DEFAULT);
   }

   ...
   pthread_atfork (0, 0, post_fork_child);

=item int ev_is_default_loop (loop)

Returns true when the given loop is, in fact, the default loop, and false
otherwise.

=item unsigned int ev_iteration (loop)

Returns the current iteration count for the event loop, which is identical
to the number of times libev did poll for new events. It starts at C<0>
and happily wraps around with enough iterations.

This value can sometimes be useful as a generation counter of sorts (it
"ticks" the number of loop iterations), as it roughly corresponds with
C<ev_prepare> and C<ev_check> calls - and is incremented between the
prepare and check phases.

=item unsigned int ev_depth (loop)

Returns the number of times C<ev_run> was entered minus the number of
times C<ev_run> was exited normally, in other words, the recursion depth.

Outside C<ev_run>, this number is zero. In a callback, this number is
C<1>, unless C<ev_run> was invoked recursively (or from another thread),
in which case it is higher.

Leaving C<ev_run> abnormally (setjmp/longjmp, cancelling the thread,
throwing an exception etc.), doesn't count as "exit" - consider this
as a hint to avoid such ungentleman-like behaviour unless it's really
convenient, in which case it is fully supported.

=item unsigned int ev_backend (loop)

Returns one of the C<EVBACKEND_*> flags indicating the event backend in
use.

=item ev_tstamp ev_now (loop)

Returns the current "event loop time", which is the time the event loop
received events and started processing them. This timestamp does not
change as long as callbacks are being processed, and this is also the base
time used for relative timers. You can treat it as the timestamp of the
event occurring (or more correctly, libev finding out about it).

=item ev_now_update (loop)

Establishes the current time by querying the kernel, updating the time
returned by C<ev_now ()> in the progress. This is a costly operation and
is usually done automatically within C<ev_run ()>.

This function is rarely useful, but when some event callback runs for a
very long time without entering the event loop, updating libev's idea of
the current time is a good idea.

See also L</The special problem of time updates> in the C<ev_timer> section.

=item ev_suspend (loop)

=item ev_resume (loop)

These two functions suspend and resume an event loop, for use when the
loop is not used for a while and timeouts should not be processed.

A typical use case would be an interactive program such as a game:  When
the user presses C<^Z> to suspend the game and resumes it an hour later it
would be best to handle timeouts as if no time had actually passed while
the program was suspended. This can be achieved by calling C<ev_suspend>
in your C<SIGTSTP> handler, sending yourself a C<SIGSTOP> and calling
C<ev_resume> directly afterwards to resume timer processing.

Effectively, all C<ev_timer> watchers will be delayed by the time spend
between C<ev_suspend> and C<ev_resume>, and all C<ev_periodic> watchers
will be rescheduled (that is, they will lose any events that would have
occurred while suspended).

After calling C<ev_suspend> you B<must not> call I<any> function on the
given loop other than C<ev_resume>, and you B<must not> call C<ev_resume>
without a previous call to C<ev_suspend>.

Calling C<ev_suspend>/C<ev_resume> has the side effect of updating the
event loop time (see C<ev_now_update>).

=item bool ev_run (loop, int flags)

Finally, this is it, the event handler. This function usually is called
after you have initialised all your watchers and you want to start
handling events. It will ask the operating system for any new events, call
the watcher callbacks, and then repeat the whole process indefinitely: This
is why event loops are called I<loops>.

If the flags argument is specified as C<0>, it will keep handling events
until either no event watchers are active anymore or C<ev_break> was
called.

The return value is false if there are no more active watchers (which
usually means "all jobs done" or "deadlock"), and true in all other cases
(which usually means " you should call C<ev_run> again").

Please note that an explicit C<ev_break> is usually better than
relying on all watchers to be stopped when deciding when a program has
finished (especially in interactive programs), but having a program
that automatically loops as long as it has to and no longer by virtue
of relying on its watchers stopping correctly, that is truly a thing of
beauty.

This function is I<mostly> exception-safe - you can break out of a
C<ev_run> call by calling C<longjmp> in a callback, throwing a C++
exception and so on. This does not decrement the C<ev_depth> value, nor
will it clear any outstanding C<EVBREAK_ONE> breaks.

A flags value of C<EVRUN_NOWAIT> will look for new events, will handle
those events and any already outstanding ones, but will not wait and
block your process in case there are no events and will return after one
iteration of the loop. This is sometimes useful to poll and handle new
events while doing lengthy calculations, to keep the program responsive.

A flags value of C<EVRUN_ONCE> will look for new events (waiting if
necessary) and will handle those and any already outstanding ones. It
will block your process until at least one new event arrives (which could
be an event internal to libev itself, so there is no guarantee that a
user-registered callback will be called), and will return after one
iteration of the loop.

This is useful if you are waiting for some external event in conjunction
with something not expressible using other libev watchers (i.e. "roll your
own C<ev_run>"). However, a pair of C<ev_prepare>/C<ev_check> watchers is
usually a better approach for this kind of thing.

Here are the gory details of what C<ev_run> does (this is for your
understanding, not a guarantee that things will work exactly like this in
future versions):

   - Increment loop depth.
   - Reset the ev_break status.
   - Before the first iteration, call any pending watchers.
   LOOP:
   - If EVFLAG_FORKCHECK was used, check for a fork.
   - If a fork was detected (by any means), queue and call all fork watchers.
   - Queue and call all prepare watchers.
   - If ev_break was called, goto FINISH.
   - If we have been forked, detach and recreate the kernel state
     as to not disturb the other process.
   - Update the kernel state with all outstanding changes.
   - Update the "event loop time" (ev_now ()).
   - Calculate for how long to sleep or block, if at all
     (active idle watchers, EVRUN_NOWAIT or not having
     any active watchers at all will result in not sleeping).
   - Sleep if the I/O and timer collect interval say so.
   - Increment loop iteration counter.
   - Block the process, waiting for any events.
   - Queue all outstanding I/O (fd) events.
   - Update the "event loop time" (ev_now ()), and do time jump adjustments.
   - Queue all expired timers.
   - Queue all expired periodics.
   - Queue all idle watchers with priority higher than that of pending events.
   - Queue all check watchers.
   - Call all queued watchers in reverse order (i.e. check watchers first).
     Signals and child watchers are implemented as I/O watchers, and will
     be handled here by queueing them when their watcher gets executed.
   - If ev_break has been called, or EVRUN_ONCE or EVRUN_NOWAIT
     were used, or there are no active watchers, goto FINISH, otherwise
     continue with step LOOP.
   FINISH:
   - Reset the ev_break status iff it was EVBREAK_ONE.
   - Decrement the loop depth.
   - Return.

Example: Queue some jobs and then loop until no events are outstanding
anymore.

   ... queue jobs here, make sure they register event watchers as long
   ... as they still have work to do (even an idle watcher will do..)
   ev_run (my_loop, 0);
   ... jobs done or somebody called break. yeah!

=item ev_break (loop, how)

Can be used to make a call to C<ev_run> return early (but only after it
has processed all outstanding events). The C<how> argument must be either
C<EVBREAK_ONE>, which will make the innermost C<ev_run> call return, or
C<EVBREAK_ALL>, which will make all nested C<ev_run> calls return.

This "break state" will be cleared on the next call to C<ev_run>.

It is safe to call C<ev_break> from outside any C<ev_run> calls, too, in
which case it will have no effect.

=item ev_ref (loop)

=item ev_unref (loop)

Ref/unref can be used to add or remove a reference count on the event
loop: Every watcher keeps one reference, and as long as the reference
count is nonzero, C<ev_run> will not return on its own.

This is useful when you have a watcher that you never intend to
unregister, but that nevertheless should not keep C<ev_run> from
returning. In such a case, call C<ev_unref> after starting, and C<ev_ref>
before stopping it.

As an example, libev itself uses this for its internal signal pipe: It
is not visible to the libev user and should not keep C<ev_run> from
exiting if no event watchers registered by it are active. It is also an
excellent way to do this for generic recurring timers or from within
third-party libraries. Just remember to I<unref after start> and I<ref
before stop> (but only if the watcher wasn't active before, or was active
before, respectively. Note also that libev might stop watchers itself
(e.g. non-repeating timers) in which case you have to C<ev_ref>
in the callback).

Example: Create a signal watcher, but keep it from keeping C<ev_run>
running when nothing else is active.

   ev_signal exitsig;
   ev_signal_init (&exitsig, sig_cb, SIGINT);
   ev_signal_start (loop, &exitsig);
   ev_unref (loop);

Example: For some weird reason, unregister the above signal handler again.

   ev_ref (loop);
   ev_signal_stop (loop, &exitsig);

=item ev_set_io_collect_interval (loop, ev_tstamp interval)

=item ev_set_timeout_collect_interval (loop, ev_tstamp interval)

These advanced functions influence the time that libev will spend waiting
for events. Both time intervals are by default C<0>, meaning that libev
will try to invoke timer/periodic callbacks and I/O callbacks with minimum
latency.

Setting these to a higher value (the C<interval> I<must> be >= C<0>)
allows libev to delay invocation of I/O and timer/periodic callbacks
to increase efficiency of loop iterations (or to increase power-saving
opportunities).

The idea is that sometimes your program runs just fast enough to handle
one (or very few) event(s) per loop iteration. While this makes the
program responsive, it also wastes a lot of CPU time to poll for new
events, especially with backends like C<select ()> which have a high
overhead for the actual polling but can deliver many events at once.

By setting a higher I<io collect interval> you allow libev to spend more
time collecting I/O events, so you can handle more events per iteration,
at the cost of increasing latency. Timeouts (both C<ev_periodic> and
C<ev_timer>) will not be affected. Setting this to a non-null value will
introduce an additional C<ev_sleep ()> call into most loop iterations. The
sleep time ensures that libev will not poll for I/O events more often then
once per this interval, on average (as long as the host time resolution is
good enough).

Likewise, by setting a higher I<timeout collect interval> you allow libev
to spend more time collecting timeouts, at the expense of increased
latency/jitter/inexactness (the watcher callback will be called
later). C<ev_io> watchers will not be affected. Setting this to a non-null
value will not introduce any overhead in libev.

Many (busy) programs can usually benefit by setting the I/O collect
interval to a value near C<0.1> or so, which is often enough for
interactive servers (of course not for games), likewise for timeouts. It
usually doesn't make much sense to set it to a lower value than C<0.01>,
as this approaches the timing granularity of most systems. Note that if
you do transactions with the outside world and you can't increase the
parallelity, then this setting will limit your transaction rate (if you
need to poll once per transaction and the I/O collect interval is 0.01,
then you can't do more than 100 transactions per second).

Setting the I<timeout collect interval> can improve the opportunity for
saving power, as the program will "bundle" timer callback invocations that
are "near" in time together, by delaying some, thus reducing the number of
times the process sleeps and wakes up again. Another useful technique to
reduce iterations/wake-ups is to use C<ev_periodic> watchers and make sure
they fire on, say, one-second boundaries only.

Example: we only need 0.1s timeout granularity, and we wish not to poll
more often than 100 times per second:

   ev_set_timeout_collect_interval (EV_DEFAULT_UC_ 0.1);
   ev_set_io_collect_interval (EV_DEFAULT_UC_ 0.01);

=item ev_invoke_pending (loop)

This call will simply invoke all pending watchers while resetting their
pending state. Normally, C<ev_run> does this automatically when required,
but when overriding the invoke callback this call comes handy. This
function can be invoked from a watcher - this can be useful for example
when you want to do some lengthy calculation and want to pass further
event handling to another thread (you still have to make sure only one
thread executes within C<ev_invoke_pending> or C<ev_run> of course).

=item int ev_pending_count (loop)

Returns the number of pending watchers - zero indicates that no watchers
are pending.

=item ev_set_invoke_pending_cb (loop, void (*invoke_pending_cb)(EV_P))

This overrides the invoke pending functionality of the loop: Instead of
invoking all pending watchers when there are any, C<ev_run> will call
this callback instead. This is useful, for example, when you want to
invoke the actual watchers inside another context (another thread etc.).

If you want to reset the callback, use C<ev_invoke_pending> as new
callback.

=item ev_set_loop_release_cb (loop, void (*release)(EV_P) throw (), void (*acquire)(EV_P) throw ())

Sometimes you want to share the same loop between multiple threads. This
can be done relatively simply by putting mutex_lock/unlock calls around
each call to a libev function.

However, C<ev_run> can run an indefinite time, so it is not feasible
to wait for it to return. One way around this is to wake up the event
loop via C<ev_break> and C<ev_async_send>, another way is to set these
I<release> and I<acquire> callbacks on the loop.

When set, then C<release> will be called just before the thread is
suspended waiting for new events, and C<acquire> is called just
afterwards.

Ideally, C<release> will just call your mutex_unlock function, and
C<acquire> will just call the mutex_lock function again.

While event loop modifications are allowed between invocations of
C<release> and C<acquire> (that's their only purpose after all), no
modifications done will affect the event loop, i.e. adding watchers will
have no effect on the set of file descriptors being watched, or the time
waited. Use an C<ev_async> watcher to wake up C<ev_run> when you want it
to take note of any changes you made.

In theory, threads executing C<ev_run> will be async-cancel safe between
invocations of C<release> and C<acquire>.

See also the locking example in the C<THREADS> section later in this
document.

=item ev_set_userdata (loop, void *data)

=item void *ev_userdata (loop)

Set and retrieve a single C<void *> associated with a loop. When
C<ev_set_userdata> has never been called, then C<ev_userdata> returns
C<0>.

These two functions can be used to associate arbitrary data with a loop,
and are intended solely for the C<invoke_pending_cb>, C<release> and
C<acquire> callbacks described above, but of course can be (ab-)used for
any other purpose as well.

=item ev_verify (loop)

This function only does something when C<EV_VERIFY> support has been
compiled in, which is the default for non-minimal builds. It tries to go
through all internal structures and checks them for validity. If anything
is found to be inconsistent, it will print an error message to standard
error and call C<abort ()>.

This can be used to catch bugs inside libev itself: under normal
circumstances, this function will never abort as of course libev keeps its
data structures consistent.

=back


=head1 ANATOMY OF A WATCHER

In the following description, uppercase C<TYPE> in names stands for the
watcher type, e.g. C<ev_TYPE_start> can mean C<ev_timer_start> for timer
watchers and C<ev_io_start> for I/O watchers.

A watcher is an opaque structure that you allocate and register to record
your interest in some event. To make a concrete example, imagine you want
to wait for STDIN to become readable, you would create an C<ev_io> watcher
for that:

   static void my_cb (struct ev_loop *loop, ev_io *w, int revents)
   {
     ev_io_stop (w);
     ev_break (loop, EVBREAK_ALL);
   }

   struct ev_loop *loop = ev_default_loop (0);

   ev_io stdin_watcher;

   ev_init (&stdin_watcher, my_cb);
   ev_io_set (&stdin_watcher, STDIN_FILENO, EV_READ);
   ev_io_start (loop, &stdin_watcher);

   ev_run (loop, 0);

As you can see, you are responsible for allocating the memory for your
watcher structures (and it is I<usually> a bad idea to do this on the
stack).

Each watcher has an associated watcher structure (called C<struct ev_TYPE>
or simply C<ev_TYPE>, as typedefs are provided for all watcher structs).

Each watcher structure must be initialised by a call to C<ev_init (watcher
*, callback)>, which expects a callback to be provided. This callback is
invoked each time the event occurs (or, in the case of I/O watchers, each
time the event loop detects that the file descriptor given is readable
and/or writable).

Each watcher type further has its own C<< ev_TYPE_set (watcher *, ...) >>
macro to configure it, with arguments specific to the watcher type. There
is also a macro to combine initialisation and setting in one call: C<<
ev_TYPE_init (watcher *, callback, ...) >>.

To make the watcher actually watch out for events, you have to start it
with a watcher-specific start function (C<< ev_TYPE_start (loop, watcher
*) >>), and you can stop watching for events at any time by calling the
corresponding stop function (C<< ev_TYPE_stop (loop, watcher *) >>.

As long as your watcher is active (has been started but not stopped) you
must not touch the values stored in it. Most specifically you must never
reinitialise it or call its C<ev_TYPE_set> macro.

Each and every callback receives the event loop pointer as first, the
registered watcher structure as second, and a bitset of received events as
third argument.

The received events usually include a single bit per event type received
(you can receive multiple events at the same time). The possible bit masks
are:

=over 4

=item C<EV_READ>

=item C<EV_WRITE>

The file descriptor in the C<ev_io> watcher has become readable and/or
writable.

=item C<EV_TIMER>

The C<ev_timer> watcher has timed out.

=item C<EV_PERIODIC>

The C<ev_periodic> watcher has timed out.

=item C<EV_SIGNAL>

The signal specified in the C<ev_signal> watcher has been received by a thread.

=item C<EV_CHILD>

The pid specified in the C<ev_child> watcher has received a status change.

=item C<EV_STAT>

The path specified in the C<ev_stat> watcher changed its attributes somehow.

=item C<EV_IDLE>

The C<ev_idle> watcher has determined that you have nothing better to do.

=item C<EV_PREPARE>

=item C<EV_CHECK>

All C<ev_prepare> watchers are invoked just I<before> C<ev_run> starts to
gather new events, and all C<ev_check> watchers are queued (not invoked)
just after C<ev_run> has gathered them, but before it queues any callbacks
for any received events. That means C<ev_prepare> watchers are the last
watchers invoked before the event loop sleeps or polls for new events, and
C<ev_check> watchers will be invoked before any other watchers of the same
or lower priority within an event loop iteration.

Callbacks of both watcher types can start and stop as many watchers as
they want, and all of them will be taken into account (for example, a
C<ev_prepare> watcher might start an idle watcher to keep C<ev_run> from
blocking).

=item C<EV_EMBED>

The embedded event loop specified in the C<ev_embed> watcher needs attention.

=item C<EV_FORK>

The event loop has been resumed in the child process after fork (see
C<ev_fork>).

=item C<EV_CLEANUP>

The event loop is about to be destroyed (see C<ev_cleanup>).

=item C<EV_ASYNC>

The given async watcher has been asynchronously notified (see C<ev_async>).

=item C<EV_CUSTOM>

Not ever sent (or otherwise used) by libev itself, but can be freely used
by libev users to signal watchers (e.g. via C<ev_feed_event>).

=item C<EV_ERROR>

An unspecified error has occurred, the watcher has been stopped. This might
happen because the watcher could not be properly started because libev
ran out of memory, a file descriptor was found to be closed or any other
problem. Libev considers these application bugs.

You best act on it by reporting the problem and somehow coping with the
watcher being stopped. Note that well-written programs should not receive
an error ever, so when your watcher receives it, this usually indicates a
bug in your program.

Libev will usually signal a few "dummy" events together with an error, for
example it might indicate that a fd is readable or writable, and if your
callbacks is well-written it can just attempt the operation and cope with
the error from read() or write(). This will not work in multi-threaded
programs, though, as the fd could already be closed and reused for another
thing, so beware.

=back

=head2 GENERIC WATCHER FUNCTIONS

=over 4

=item C<ev_init> (ev_TYPE *watcher, callback)

This macro initialises the generic portion of a watcher. The contents
of the watcher object can be arbitrary (so C<malloc> will do). Only
the generic parts of the watcher are initialised, you I<need> to call
the type-specific C<ev_TYPE_set> macro afterwards to initialise the
type-specific parts. For each type there is also a C<ev_TYPE_init> macro
which rolls both calls into one.

You can reinitialise a watcher at any time as long as it has been stopped
(or never started) and there are no pending events outstanding.

The callback is always of type C<void (*)(struct ev_loop *loop, ev_TYPE *watcher,
int revents)>.

Example: Initialise an C<ev_io> watcher in two steps.

   ev_io w;
   ev_init (&w, my_cb);
   ev_io_set (&w, STDIN_FILENO, EV_READ);

=item C<ev_TYPE_set> (ev_TYPE *watcher, [args])

This macro initialises the type-specific parts of a watcher. You need to
call C<ev_init> at least once before you call this macro, but you can
call C<ev_TYPE_set> any number of times. You must not, however, call this
macro on a watcher that is active (it can be pending, however, which is a
difference to the C<ev_init> macro).

Although some watcher types do not have type-specific arguments
(e.g. C<ev_prepare>) you still need to call its C<set> macro.

See C<ev_init>, above, for an example.

=item C<ev_TYPE_init> (ev_TYPE *watcher, callback, [args])

This convenience macro rolls both C<ev_init> and C<ev_TYPE_set> macro
calls into a single call. This is the most convenient method to initialise
a watcher. The same limitations apply, of course.

Example: Initialise and set an C<ev_io> watcher in one step.

   ev_io_init (&w, my_cb, STDIN_FILENO, EV_READ);

=item C<ev_TYPE_start> (loop, ev_TYPE *watcher)

Starts (activates) the given watcher. Only active watchers will receive
events. If the watcher is already active nothing will happen.

Example: Start the C<ev_io> watcher that is being abused as example in this
whole section.

   ev_io_start (EV_DEFAULT_UC, &w);

=item C<ev_TYPE_stop> (loop, ev_TYPE *watcher)

Stops the given watcher if active, and clears the pending status (whether
the watcher was active or not).

It is possible that stopped watchers are pending - for example,
non-repeating timers are being stopped when they become pending - but
calling C<ev_TYPE_stop> ensures that the watcher is neither active nor
pending. If you want to free or reuse the memory used by the watcher it is
therefore a good idea to always call its C<ev_TYPE_stop> function.

=item bool ev_is_active (ev_TYPE *watcher)

Returns a true value iff the watcher is active (i.e. it has been started
and not yet been stopped). As long as a watcher is active you must not modify
it.

=item bool ev_is_pending (ev_TYPE *watcher)

Returns a true value iff the watcher is pending, (i.e. it has outstanding
events but its callback has not yet been invoked). As long as a watcher
is pending (but not active) you must not call an init function on it (but
C<ev_TYPE_set> is safe), you must not change its priority, and you must
make sure the watcher is available to libev (e.g. you cannot C<free ()>
it).

=item callback ev_cb (ev_TYPE *watcher)

Returns the callback currently set on the watcher.

=item ev_set_cb (ev_TYPE *watcher, callback)

Change the callback. You can change the callback at virtually any time
(modulo threads).

=item ev_set_priority (ev_TYPE *watcher, int priority)

=item int ev_priority (ev_TYPE *watcher)

Set and query the priority of the watcher. The priority is a small
integer between C<EV_MAXPRI> (default: C<2>) and C<EV_MINPRI>
(default: C<-2>). Pending watchers with higher priority will be invoked
before watchers with lower priority, but priority will not keep watchers
from being executed (except for C<ev_idle> watchers).

If you need to suppress invocation when higher priority events are pending
you need to look at C<ev_idle> watchers, which provide this functionality.

You I<must not> change the priority of a watcher as long as it is active or
pending.

Setting a priority outside the range of C<EV_MINPRI> to C<EV_MAXPRI> is
fine, as long as you do not mind that the priority value you query might
or might not have been clamped to the valid range.

The default priority used by watchers when no priority has been set is
always C<0>, which is supposed to not be too high and not be too low :).

See L</WATCHER PRIORITY MODELS>, below, for a more thorough treatment of
priorities.

=item ev_invoke (loop, ev_TYPE *watcher, int revents)

Invoke the C<watcher> with the given C<loop> and C<revents>. Neither
C<loop> nor C<revents> need to be valid as long as the watcher callback
can deal with that fact, as both are simply passed through to the
callback.

=item int ev_clear_pending (loop, ev_TYPE *watcher)

If the watcher is pending, this function clears its pending status and
returns its C<revents> bitset (as if its callback was invoked). If the
watcher isn't pending it does nothing and returns C<0>.

Sometimes it can be useful to "poll" a watcher instead of waiting for its
callback to be invoked, which can be accomplished with this function.

=item ev_feed_event (loop, ev_TYPE *watcher, int revents)

Feeds the given event set into the event loop, as if the specified event
had happened for the specified watcher (which must be a pointer to an
initialised but not necessarily started event watcher). Obviously you must
not free the watcher as long as it has pending events.

Stopping the watcher, letting libev invoke it, or calling
C<ev_clear_pending> will clear the pending event, even if the watcher was
not started in the first place.

See also C<ev_feed_fd_event> and C<ev_feed_signal_event> for related
functions that do not need a watcher.

=back

See also the L</ASSOCIATING CUSTOM DATA WITH A WATCHER> and L</BUILDING YOUR
OWN COMPOSITE WATCHERS> idioms.

=head2 WATCHER STATES

There are various watcher states mentioned throughout this manual -
active, pending and so on. In this section these states and the rules to
transition between them will be described in more detail - and while these
rules might look complicated, they usually do "the right thing".

=over 4

=item initialised

Before a watcher can be registered with the event loop it has to be
initialised. This can be done with a call to C<ev_TYPE_init>, or calls to
C<ev_init> followed by the watcher-specific C<ev_TYPE_set> function.

In this state it is simply some block of memory that is suitable for
use in an event loop. It can be moved around, freed, reused etc. at
will - as long as you either keep the memory contents intact, or call
C<ev_TYPE_init> again.

=item started/running/active

Once a watcher has been started with a call to C<ev_TYPE_start> it becomes
property of the event loop, and is actively waiting for events. While in
this state it cannot be accessed (except in a few documented ways), moved,
freed or anything else - the only legal thing is to keep a pointer to it,
and call libev functions on it that are documented to work on active watchers.

=item pending

If a watcher is active and libev determines that an event it is interested
in has occurred (such as a timer expiring), it will become pending. It will
stay in this pending state until either it is stopped or its callback is
about to be invoked, so it is not normally pending inside the watcher
callback.

The watcher might or might not be active while it is pending (for example,
an expired non-repeating timer can be pending but no longer active). If it
is stopped, it can be freely accessed (e.g. by calling C<ev_TYPE_set>),
but it is still property of the event loop at this time, so cannot be
moved, freed or reused. And if it is active the rules described in the
previous item still apply.

It is also possible to feed an event on a watcher that is not active (e.g.
via C<ev_feed_event>), in which case it becomes pending without being
active.

=item stopped

A watcher can be stopped implicitly by libev (in which case it might still
be pending), or explicitly by calling its C<ev_TYPE_stop> function. The
latter will clear any pending state the watcher might be in, regardless
of whether it was active or not, so stopping a watcher explicitly before
freeing it is often a good idea.

While stopped (and not pending) the watcher is essentially in the
initialised state, that is, it can be reused, moved, modified in any way
you wish (but when you trash the memory block, you need to C<ev_TYPE_init>
it again).

=back

=head2 WATCHER PRIORITY MODELS

Many event loops support I<watcher priorities>, which are usually small
integers that influence the ordering of event callback invocation
between watchers in some way, all else being equal.

In libev, watcher priorities can be set using C<ev_set_priority>. See its
description for the more technical details such as the actual priority
range.

There are two common ways how these these priorities are being interpreted
by event loops:

In the more common lock-out model, higher priorities "lock out" invocation
of lower priority watchers, which means as long as higher priority
watchers receive events, lower priority watchers are not being invoked.

The less common only-for-ordering model uses priorities solely to order
callback invocation within a single event loop iteration: Higher priority
watchers are invoked before lower priority ones, but they all get invoked
before polling for new events.

Libev uses the second (only-for-ordering) model for all its watchers
except for idle watchers (which use the lock-out model).

The rationale behind this is that implementing the lock-out model for
watchers is not well supported by most kernel interfaces, and most event
libraries will just poll for the same events again and again as long as
their callbacks have not been executed, which is very inefficient in the
common case of one high-priority watcher locking out a mass of lower
priority ones.

Static (ordering) priorities are most useful when you have two or more
watchers handling the same resource: a typical usage example is having an
C<ev_io> watcher to receive data, and an associated C<ev_timer> to handle
timeouts. Under load, data might be received while the program handles
other jobs, but since timers normally get invoked first, the timeout
handler will be executed before checking for data. In that case, giving
the timer a lower priority than the I/O watcher ensures that I/O will be
handled first even under adverse conditions (which is usually, but not
always, what you want).

Since idle watchers use the "lock-out" model, meaning that idle watchers
will only be executed when no same or higher priority watchers have
received events, they can be used to implement the "lock-out" model when
required.

For example, to emulate how many other event libraries handle priorities,
you can associate an C<ev_idle> watcher to each such watcher, and in
the normal watcher callback, you just start the idle watcher. The real
processing is done in the idle watcher callback. This causes libev to
continuously poll and process kernel event data for the watcher, but when
the lock-out case is known to be rare (which in turn is rare :), this is
workable.

Usually, however, the lock-out model implemented that way will perform
miserably under the type of load it was designed to handle. In that case,
it might be preferable to stop the real watcher before starting the
idle watcher, so the kernel will not have to process the event in case
the actual processing will be delayed for considerable time.

Here is an example of an I/O watcher that should run at a strictly lower
priority than the default, and which should only process data when no
other events are pending:

   ev_idle idle; // actual processing watcher
   ev_io io;     // actual event watcher

   static void
   io_cb (EV_P_ ev_io *w, int revents)
   {
     // stop the I/O watcher, we received the event, but
     // are not yet ready to handle it.
     ev_io_stop (EV_A_ w);

     // start the idle watcher to handle the actual event.
     // it will not be executed as long as other watchers
     // with the default priority are receiving events.
     ev_idle_start (EV_A_ &idle);
   }

   static void
   idle_cb (EV_P_ ev_idle *w, int revents)
   {
     // actual processing
     read (STDIN_FILENO, ...);

     // have to start the I/O watcher again, as
     // we have handled the event
     ev_io_start (EV_P_ &io);
   }

   // initialisation
   ev_idle_init (&idle, idle_cb);
   ev_io_init (&io, io_cb, STDIN_FILENO, EV_READ);
   ev_io_start (EV_DEFAULT_ &io);

In the "real" world, it might also be beneficial to start a timer, so that
low-priority connections can not be locked out forever under load. This
enables your program to keep a lower latency for important connections
during short periods of high load, while not completely locking out less
important ones.


=head1 WATCHER TYPES

This section describes each watcher in detail, but will not repeat
information given in the last section. Any initialisation/set macros,
functions and members specific to the watcher type are explained.

Most members are additionally marked with either I<[read-only]>, meaning
that, while the watcher is active, you can look at the member and expect
some sensible content, but you must not modify it (you can modify it while
the watcher is stopped to your hearts content), or I<[read-write]>, which
means you can expect it to have some sensible content while the watcher
is active, but you can also modify it. Modifying it may not do something
sensible or take immediate effect (or do anything at all), but libev will
not crash or malfunction in any way.

In any case, the documentation for each member will explain what the
effects are, and if there are any additional access restrictions.

=head2 C<ev_io> - is this file descriptor readable or writable?

I/O watchers check whether a file descriptor is readable or writable
in each iteration of the event loop, or, more precisely, when reading
would not block the process and writing would at least be able to write
some data. This behaviour is called level-triggering because you keep
receiving events as long as the condition persists. Remember you can stop
the watcher if you don't want to act on the event and neither want to
receive future events.

In general you can register as many read and/or write event watchers per
fd as you want (as long as you don't confuse yourself). Setting all file
descriptors to non-blocking mode is also usually a good idea (but not
required if you know what you are doing).

Another thing you have to watch out for is that it is quite easy to
receive "spurious" readiness notifications, that is, your callback might
be called with C<EV_READ> but a subsequent C<read>(2) will actually block
because there is no data. It is very easy to get into this situation even
with a relatively standard program structure. Thus it is best to always
use non-blocking I/O: An extra C<read>(2) returning C<EAGAIN> is far
preferable to a program hanging until some data arrives.

If you cannot run the fd in non-blocking mode (for example you should
not play around with an Xlib connection), then you have to separately
re-test whether a file descriptor is really ready with a known-to-be good
interface such as poll (fortunately in the case of Xlib, it already does
this on its own, so its quite safe to use). Some people additionally
use C<SIGALRM> and an interval timer, just to be sure you won't block
indefinitely.

But really, best use non-blocking mode.

=head3 The special problem of disappearing file descriptors

Some backends (e.g. kqueue, epoll, linuxaio) need to be told about closing
a file descriptor (either due to calling C<close> explicitly or any other
means, such as C<dup2>). The reason is that you register interest in some
file descriptor, but when it goes away, the operating system will silently
drop this interest. If another file descriptor with the same number then
is registered with libev, there is no efficient way to see that this is,
in fact, a different file descriptor.

To avoid having to explicitly tell libev about such cases, libev follows
the following policy:  Each time C<ev_io_set> is being called, libev
will assume that this is potentially a new file descriptor, otherwise
it is assumed that the file descriptor stays the same. That means that
you I<have> to call C<ev_io_set> (or C<ev_io_init>) when you change the
descriptor even if the file descriptor number itself did not change.

This is how one would do it normally anyway, the important point is that
the libev application should not optimise around libev but should leave
optimisations to libev.

=head3 The special problem of dup'ed file descriptors

Some backends (e.g. epoll), cannot register events for file descriptors,
but only events for the underlying file descriptions. That means when you
have C<dup ()>'ed file descriptors or weirder constellations, and register
events for them, only one file descriptor might actually receive events.

There is no workaround possible except not registering events
for potentially C<dup ()>'ed file descriptors, or to resort to
C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.

=head3 The special problem of files

Many people try to use C<select> (or libev) on file descriptors
representing files, and expect it to become ready when their program
doesn't block on disk accesses (which can take a long time on their own).

However, this cannot ever work in the "expected" way - you get a readiness
notification as soon as the kernel knows whether and how much data is
there, and in the case of open files, that's always the case, so you
always get a readiness notification instantly, and your read (or possibly
write) will still block on the disk I/O.

Another way to view it is that in the case of sockets, pipes, character
devices and so on, there is another party (the sender) that delivers data
on its own, but in the case of files, there is no such thing: the disk
will not send data on its own, simply because it doesn't know what you
wish to read - you would first have to request some data.

Since files are typically not-so-well supported by advanced notification
mechanism, libev tries hard to emulate POSIX behaviour with respect
to files, even though you should not use it. The reason for this is
convenience: sometimes you want to watch STDIN or STDOUT, which is
usually a tty, often a pipe, but also sometimes files or special devices
(for example, C<epoll> on Linux works with F</dev/random> but not with
F</dev/urandom>), and even though the file might better be served with
asynchronous I/O instead of with non-blocking I/O, it is still useful when
it "just works" instead of freezing.

So avoid file descriptors pointing to files when you know it (e.g. use
libeio), but use them when it is convenient, e.g. for STDIN/STDOUT, or
when you rarely read from a file instead of from a socket, and want to
reuse the same code path.

=head3 The special problem of fork

Some backends (epoll, kqueue, linuxaio, iouring) do not support C<fork ()>
at all or exhibit useless behaviour. Libev fully supports fork, but needs
to be told about it in the child if you want to continue to use it in the
child.

To support fork in your child processes, you have to call C<ev_loop_fork
()> after a fork in the child, enable C<EVFLAG_FORKCHECK>, or resort to
C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.

=head3 The special problem of SIGPIPE

While not really specific to libev, it is easy to forget about C<SIGPIPE>:
when writing to a pipe whose other end has been closed, your program gets
sent a SIGPIPE, which, by default, aborts your program. For most programs
this is sensible behaviour, for daemons, this is usually undesirable.

So when you encounter spurious, unexplained daemon exits, make sure you
ignore SIGPIPE (and maybe make sure you log the exit status of your daemon
somewhere, as that would have given you a big clue).

=head3 The special problem of accept()ing when you can't

Many implementations of the POSIX C<accept> function (for example,
found in post-2004 Linux) have the peculiar behaviour of not removing a
connection from the pending queue in all error cases.

For example, larger servers often run out of file descriptors (because
of resource limits), causing C<accept> to fail with C<ENFILE> but not
rejecting the connection, leading to libev signalling readiness on
the next iteration again (the connection still exists after all), and
typically causing the program to loop at 100% CPU usage.

Unfortunately, the set of errors that cause this issue differs between
operating systems, there is usually little the app can do to remedy the
situation, and no known thread-safe method of removing the connection to
cope with overload is known (to me).

One of the easiest ways to handle this situation is to just ignore it
- when the program encounters an overload, it will just loop until the
situation is over. While this is a form of busy waiting, no OS offers an
event-based way to handle this situation, so it's the best one can do.

A better way to handle the situation is to log any errors other than
C<EAGAIN> and C<EWOULDBLOCK>, making sure not to flood the log with such
messages, and continue as usual, which at least gives the user an idea of
what could be wrong ("raise the ulimit!"). For extra points one could stop
the C<ev_io> watcher on the listening fd "for a while", which reduces CPU
usage.

If your program is single-threaded, then you could also keep a dummy file
descriptor for overload situations (e.g. by opening F</dev/null>), and
when you run into C<ENFILE> or C<EMFILE>, close it, run C<accept>,
close that fd, and create a new dummy fd. This will gracefully refuse
clients under typical overload conditions.

The last way to handle it is to simply log the error and C<exit>, as
is often done with C<malloc> failures, but this results in an easy
opportunity for a DoS attack.

=head3 Watcher-Specific Functions

=over 4

=item ev_io_init (ev_io *, callback, int fd, int events)

=item ev_io_set (ev_io *, int fd, int events)

Configures an C<ev_io> watcher. The C<fd> is the file descriptor to
receive events for and C<events> is either C<EV_READ>, C<EV_WRITE> or
C<EV_READ | EV_WRITE>, to express the desire to receive the given events.

=item ev_io_modify (ev_io *, int events)

Similar to C<ev_io_set>, but only changes the event mask. Using this might
be faster with some backends, as libev can assume that the C<fd> still
refers to the same underlying file description, something it cannot do
when using C<ev_io_set>.

=item int fd [no-modify]

The file descriptor being watched. While it can be read at any time, you
must not modify this member even when the watcher is stopped - always use
C<ev_io_set> for that.

=item int events [no-modify]

The set of events being watched, among other flags. This field is a
bit set - to test for C<EV_READ>, use C<< w->events & EV_READ >>, and
similarly for C<EV_WRITE>.

As with C<fd>, you must not modify this member even when the watcher is
stopped, always use C<ev_io_set> or C<ev_io_modify> for that.

=back

=head3 Examples

Example: Call C<stdin_readable_cb> when STDIN_FILENO has become, well
readable, but only once. Since it is likely line-buffered, you could
attempt to read a whole line in the callback.

   static void
   stdin_readable_cb (struct ev_loop *loop, ev_io *w, int revents)
   {
      ev_io_stop (loop, w);
     .. read from stdin here (or from w->fd) and handle any I/O errors
   }

   ...
   struct ev_loop *loop = ev_default_init (0);
   ev_io stdin_readable;
   ev_io_init (&stdin_readable, stdin_readable_cb, STDIN_FILENO, EV_READ);
   ev_io_start (loop, &stdin_readable);
   ev_run (loop, 0);


=head2 C<ev_timer> - relative and optionally repeating timeouts

Timer watchers are simple relative timers that generate an event after a
given time, and optionally repeating in regular intervals after that.

The timers are based on real time, that is, if you register an event that
times out after an hour and you reset your system clock to January last
year, it will still time out after (roughly) one hour. "Roughly" because
detecting time jumps is hard, and some inaccuracies are unavoidable (the
monotonic clock option helps a lot here).

The callback is guaranteed to be invoked only I<after> its timeout has
passed (not I<at>, so on systems with very low-resolution clocks this
might introduce a small delay, see "the special problem of being too
early", below). If multiple timers become ready during the same loop
iteration then the ones with earlier time-out values are invoked before
ones of the same priority with later time-out values (but this is no
longer true when a callback calls C<ev_run> recursively).

=head3 Be smart about timeouts

Many real-world problems involve some kind of timeout, usually for error
recovery. A typical example is an HTTP request - if the other side hangs,
you want to raise some error after a while.

What follows are some ways to handle this problem, from obvious and
inefficient to smart and efficient.

In the following, a 60 second activity timeout is assumed - a timeout that
gets reset to 60 seconds each time there is activity (e.g. each time some
data or other life sign was received).

=over 4

=item 1. Use a timer and stop, reinitialise and start it on activity.

This is the most obvious, but not the most simple way: In the beginning,
start the watcher:

   ev_timer_init (timer, callback, 60., 0.);
   ev_timer_start (loop, timer);

Then, each time there is some activity, C<ev_timer_stop> it, initialise it
and start it again:

   ev_timer_stop (loop, timer);
   ev_timer_set (timer, 60., 0.);
   ev_timer_start (loop, timer);

This is relatively simple to implement, but means that each time there is
some activity, libev will first have to remove the timer from its internal
data structure and then add it again. Libev tries to be fast, but it's
still not a constant-time operation.

=item 2. Use a timer and re-start it with C<ev_timer_again> inactivity.

This is the easiest way, and involves using C<ev_timer_again> instead of
C<ev_timer_start>.

To implement this, configure an C<ev_timer> with a C<repeat> value
of C<60> and then call C<ev_timer_again> at start and each time you
successfully read or write some data. If you go into an idle state where
you do not expect data to travel on the socket, you can C<ev_timer_stop>
the timer, and C<ev_timer_again> will automatically restart it if need be.

That means you can ignore both the C<ev_timer_start> function and the
C<after> argument to C<ev_timer_set>, and only ever use the C<repeat>
member and C<ev_timer_again>.

At start:

   ev_init (timer, callback);
   timer->repeat = 60.;
   ev_timer_again (loop, timer);

Each time there is some activity:

   ev_timer_again (loop, timer);

It is even possible to change the time-out on the fly, regardless of
whether the watcher is active or not:

   timer->repeat = 30.;
   ev_timer_again (loop, timer);

This is slightly more efficient then stopping/starting the timer each time
you want to modify its timeout value, as libev does not have to completely
remove and re-insert the timer from/into its internal data structure.

It is, however, even simpler than the "obvious" way to do it.

=item 3. Let the timer time out, but then re-arm it as required.

This method is more tricky, but usually most efficient: Most timeouts are
relatively long compared to the intervals between other activity - in
our example, within 60 seconds, there are usually many I/O events with
associated activity resets.

In this case, it would be more efficient to leave the C<ev_timer> alone,
but remember the time of last activity, and check for a real timeout only
within the callback:

   ev_tstamp timeout = 60.;
   ev_tstamp last_activity; // time of last activity
   ev_timer timer;

   static void
   callback (EV_P_ ev_timer *w, int revents)
   {
     // calculate when the timeout would happen
     ev_tstamp after = last_activity - ev_now (EV_A) + timeout;

     // if negative, it means we the timeout already occurred
     if (after < 0.)
       {
         // timeout occurred, take action
       }
     else
       {
         // callback was invoked, but there was some recent 
         // activity. simply restart the timer to time out
         // after "after" seconds, which is the earliest time
         // the timeout can occur.
         ev_timer_set (w, after, 0.);
         ev_timer_start (EV_A_ w);
       }
   }

To summarise the callback: first calculate in how many seconds the
timeout will occur (by calculating the absolute time when it would occur,
C<last_activity + timeout>, and subtracting the current time, C<ev_now
(EV_A)> from that).

If this value is negative, then we are already past the timeout, i.e. we
timed out, and need to do whatever is needed in this case.

Otherwise, we now the earliest time at which the timeout would trigger,
and simply start the timer with this timeout value.

In other words, each time the callback is invoked it will check whether
the timeout occurred. If not, it will simply reschedule itself to check
again at the earliest time it could time out. Rinse. Repeat.

This scheme causes more callback invocations (about one every 60 seconds
minus half the average time between activity), but virtually no calls to
libev to change the timeout.

To start the machinery, simply initialise the watcher and set
C<last_activity> to the current time (meaning there was some activity just
now), then call the callback, which will "do the right thing" and start
the timer:

   last_activity = ev_now (EV_A);
   ev_init (&timer, callback);
   callback (EV_A_ &timer, 0);

When there is some activity, simply store the current time in
C<last_activity>, no libev calls at all:

   if (activity detected)
     last_activity = ev_now (EV_A);

When your timeout value changes, then the timeout can be changed by simply
providing a new value, stopping the timer and calling the callback, which
will again do the right thing (for example, time out immediately :).

   timeout = new_value;
   ev_timer_stop (EV_A_ &timer);
   callback (EV_A_ &timer, 0);

This technique is slightly more complex, but in most cases where the
time-out is unlikely to be triggered, much more efficient.

=item 4. Wee, just use a double-linked list for your timeouts.

If there is not one request, but many thousands (millions...), all
employing some kind of timeout with the same timeout value, then one can
do even better:

When starting the timeout, calculate the timeout value and put the timeout
at the I<end> of the list.

Then use an C<ev_timer> to fire when the timeout at the I<beginning> of
the list is expected to fire (for example, using the technique #3).

When there is some activity, remove the timer from the list, recalculate
the timeout, append it to the end of the list again, and make sure to
update the C<ev_timer> if it was taken from the beginning of the list.

This way, one can manage an unlimited number of timeouts in O(1) time for
starting, stopping and updating the timers, at the expense of a major
complication, and having to use a constant timeout. The constant timeout
ensures that the list stays sorted.

=back

So which method the best?

Method #2 is a simple no-brain-required solution that is adequate in most
situations. Method #3 requires a bit more thinking, but handles many cases
better, and isn't very complicated either. In most case, choosing either
one is fine, with #3 being better in typical situations.

Method #1 is almost always a bad idea, and buys you nothing. Method #4 is
rather complicated, but extremely efficient, something that really pays
off after the first million or so of active timers, i.e. it's usually
overkill :)

=head3 The special problem of being too early

If you ask a timer to call your callback after three seconds, then
you expect it to be invoked after three seconds - but of course, this
cannot be guaranteed to infinite precision. Less obviously, it cannot be
guaranteed to any precision by libev - imagine somebody suspending the
process with a STOP signal for a few hours for example.

So, libev tries to invoke your callback as soon as possible I<after> the
delay has occurred, but cannot guarantee this.

A less obvious failure mode is calling your callback too early: many event
loops compare timestamps with a "elapsed delay >= requested delay", but
this can cause your callback to be invoked much earlier than you would
expect.

To see why, imagine a system with a clock that only offers full second
resolution (think windows if you can't come up with a broken enough OS
yourself). If you schedule a one-second timer at the time 500.9, then the
event loop will schedule your timeout to elapse at a system time of 500
(500.9 truncated to the resolution) + 1, or 501.

If an event library looks at the timeout 0.1s later, it will see "501 >=
501" and invoke the callback 0.1s after it was started, even though a
one-second delay was requested - this is being "too early", despite best
intentions.

This is the reason why libev will never invoke the callback if the elapsed
delay equals the requested delay, but only when the elapsed delay is
larger than the requested delay. In the example above, libev would only invoke
the callback at system time 502, or 1.1s after the timer was started.

So, while libev cannot guarantee that your callback will be invoked
exactly when requested, it I<can> and I<does> guarantee that the requested
delay has actually elapsed, or in other words, it always errs on the "too
late" side of things.

=head3 The special problem of time updates

Establishing the current time is a costly operation (it usually takes
at least one system call): EV therefore updates its idea of the current
time only before and after C<ev_run> collects new events, which causes a
growing difference between C<ev_now ()> and C<ev_time ()> when handling
lots of events in one iteration.

The relative timeouts are calculated relative to the C<ev_now ()>
time. This is usually the right thing as this timestamp refers to the time
of the event triggering whatever timeout you are modifying/starting. If
you suspect event processing to be delayed and you I<need> to base the
timeout on the current time, use something like the following to adjust
for it:

   ev_timer_set (&timer, after + (ev_time () - ev_now ()), 0.);

If the event loop is suspended for a long time, you can also force an
update of the time returned by C<ev_now ()> by calling C<ev_now_update
()>, although that will push the event time of all outstanding events
further into the future.

=head3 The special problem of unsynchronised clocks

Modern systems have a variety of clocks - libev itself uses the normal
"wall clock" clock and, if available, the monotonic clock (to avoid time
jumps).

Neither of these clocks is synchronised with each other or any other clock
on the system, so C<ev_time ()> might return a considerably different time
than C<gettimeofday ()> or C<time ()>. On a GNU/Linux system, for example,
a call to C<gettimeofday> might return a second count that is one higher
than a directly following call to C<time>.

The moral of this is to only compare libev-related timestamps with
C<ev_time ()> and C<ev_now ()>, at least if you want better precision than
a second or so.

One more problem arises due to this lack of synchronisation: if libev uses
the system monotonic clock and you compare timestamps from C<ev_time>
or C<ev_now> from when you started your timer and when your callback is
invoked, you will find that sometimes the callback is a bit "early".

This is because C<ev_timer>s work in real time, not wall clock time, so
libev makes sure your callback is not invoked before the delay happened,
I<measured according to the real time>, not the system clock.

If your timeouts are based on a physical timescale (e.g. "time out this
connection after 100 seconds") then this shouldn't bother you as it is
exactly the right behaviour.

If you want to compare wall clock/system timestamps to your timers, then
you need to use C<ev_periodic>s, as these are based on the wall clock
time, where your comparisons will always generate correct results.

=head3 The special problems of suspended animation

When you leave the server world it is quite customary to hit machines that
can suspend/hibernate - what happens to the clocks during such a suspend?

Some quick tests made with a Linux 2.6.28 indicate that a suspend freezes
all processes, while the clocks (C<times>, C<CLOCK_MONOTONIC>) continue
to run until the system is suspended, but they will not advance while the
system is suspended. That means, on resume, it will be as if the program
was frozen for a few seconds, but the suspend time will not be counted
towards C<ev_timer> when a monotonic clock source is used. The real time
clock advanced as expected, but if it is used as sole clocksource, then a
long suspend would be detected as a time jump by libev, and timers would
be adjusted accordingly.

I would not be surprised to see different behaviour in different between
operating systems, OS versions or even different hardware.

The other form of suspend (job control, or sending a SIGSTOP) will see a
time jump in the monotonic clocks and the realtime clock. If the program
is suspended for a very long time, and monotonic clock sources are in use,
then you can expect C<ev_timer>s to expire as the full suspension time
will be counted towards the timers. When no monotonic clock source is in
use, then libev will again assume a timejump and adjust accordingly.

It might be beneficial for this latter case to call C<ev_suspend>
and C<ev_resume> in code that handles C<SIGTSTP>, to at least get
deterministic behaviour in this case (you can do nothing against
C<SIGSTOP>).

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_timer_init (ev_timer *, callback, ev_tstamp after, ev_tstamp repeat)

=item ev_timer_set (ev_timer *, ev_tstamp after, ev_tstamp repeat)

Configure the timer to trigger after C<after> seconds (fractional and
negative values are supported). If C<repeat> is C<0.>, then it will
automatically be stopped once the timeout is reached. If it is positive,
then the timer will automatically be configured to trigger again C<repeat>
seconds later, again, and again, until stopped manually.

The timer itself will do a best-effort at avoiding drift, that is, if
you configure a timer to trigger every 10 seconds, then it will normally
trigger at exactly 10 second intervals. If, however, your program cannot
keep up with the timer (because it takes longer than those 10 seconds to
do stuff) the timer will not fire more than once per event loop iteration.

=item ev_timer_again (loop, ev_timer *)

This will act as if the timer timed out, and restarts it again if it is
repeating. It basically works like calling C<ev_timer_stop>, updating the
timeout to the C<repeat> value and calling C<ev_timer_start>.

The exact semantics are as in the following rules, all of which will be
applied to the watcher:

=over 4

=item If the timer is pending, the pending status is always cleared.

=item If the timer is started but non-repeating, stop it (as if it timed
out, without invoking it).

=item If the timer is repeating, make the C<repeat> value the new timeout
and start the timer, if necessary.

=back

This sounds a bit complicated, see L</Be smart about timeouts>, above, for a
usage example.

=item ev_tstamp ev_timer_remaining (loop, ev_timer *)

Returns the remaining time until a timer fires. If the timer is active,
then this time is relative to the current event loop time, otherwise it's
the timeout value currently configured.

That is, after an C<ev_timer_set (w, 5, 7)>, C<ev_timer_remaining> returns
C<5>. When the timer is started and one second passes, C<ev_timer_remaining>
will return C<4>. When the timer expires and is restarted, it will return
roughly C<7> (likely slightly less as callback invocation takes some time,
too), and so on.

=item ev_tstamp repeat [read-write]

The current C<repeat> value. Will be used each time the watcher times out
or C<ev_timer_again> is called, and determines the next timeout (if any),
which is also when any modifications are taken into account.

=back

=head3 Examples

Example: Create a timer that fires after 60 seconds.

   static void
   one_minute_cb (struct ev_loop *loop, ev_timer *w, int revents)
   {
     .. one minute over, w is actually stopped right here
   }

   ev_timer mytimer;
   ev_timer_init (&mytimer, one_minute_cb, 60., 0.);
   ev_timer_start (loop, &mytimer);

Example: Create a timeout timer that times out after 10 seconds of
inactivity.

   static void
   timeout_cb (struct ev_loop *loop, ev_timer *w, int revents)
   {
     .. ten seconds without any activity
   }

   ev_timer mytimer;
   ev_timer_init (&mytimer, timeout_cb, 0., 10.); /* note, only repeat used */
   ev_timer_again (&mytimer); /* start timer */
   ev_run (loop, 0);

   // and in some piece of code that gets executed on any "activity":
   // reset the timeout to start ticking again at 10 seconds
   ev_timer_again (&mytimer);


=head2 C<ev_periodic> - to cron or not to cron?

Periodic watchers are also timers of a kind, but they are very versatile
(and unfortunately a bit complex).

Unlike C<ev_timer>, periodic watchers are not based on real time (or
relative time, the physical time that passes) but on wall clock time
(absolute time, the thing you can read on your calendar or clock). The
difference is that wall clock time can run faster or slower than real
time, and time jumps are not uncommon (e.g. when you adjust your
wrist-watch).

You can tell a periodic watcher to trigger after some specific point
in time: for example, if you tell a periodic watcher to trigger "in 10
seconds" (by specifying e.g. C<ev_now () + 10.>, that is, an absolute time
not a delay) and then reset your system clock to January of the previous
year, then it will take a year or more to trigger the event (unlike an
C<ev_timer>, which would still trigger roughly 10 seconds after starting
it, as it uses a relative timeout).

C<ev_periodic> watchers can also be used to implement vastly more complex
timers, such as triggering an event on each "midnight, local time", or
other complicated rules. This cannot easily be done with C<ev_timer>
watchers, as those cannot react to time jumps.

As with timers, the callback is guaranteed to be invoked only when the
point in time where it is supposed to trigger has passed. If multiple
timers become ready during the same loop iteration then the ones with
earlier time-out values are invoked before ones with later time-out values
(but this is no longer true when a callback calls C<ev_run> recursively).

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_periodic_init (ev_periodic *, callback, ev_tstamp offset, ev_tstamp interval, reschedule_cb)

=item ev_periodic_set (ev_periodic *, ev_tstamp offset, ev_tstamp interval, reschedule_cb)

Lots of arguments, let's sort it out... There are basically three modes of
operation, and we will explain them from simplest to most complex:

=over 4

=item * absolute timer (offset = absolute time, interval = 0, reschedule_cb = 0)

In this configuration the watcher triggers an event after the wall clock
time C<offset> has passed. It will not repeat and will not adjust when a
time jump occurs, that is, if it is to be run at January 1st 2011 then it
will be stopped and invoked when the system clock reaches or surpasses
this point in time.

=item * repeating interval timer (offset = offset within interval, interval > 0, reschedule_cb = 0)

In this mode the watcher will always be scheduled to time out at the next
C<offset + N * interval> time (for some integer N, which can also be
negative) and then repeat, regardless of any time jumps. The C<offset>
argument is merely an offset into the C<interval> periods.

This can be used to create timers that do not drift with respect to the
system clock, for example, here is an C<ev_periodic> that triggers each
hour, on the hour (with respect to UTC):

   ev_periodic_set (&periodic, 0., 3600., 0);

This doesn't mean there will always be 3600 seconds in between triggers,
but only that the callback will be called when the system time shows a
full hour (UTC), or more correctly, when the system time is evenly divisible
by 3600.

Another way to think about it (for the mathematically inclined) is that
C<ev_periodic> will try to run the callback in this mode at the next possible
time where C<time = offset (mod interval)>, regardless of any time jumps.

The C<interval> I<MUST> be positive, and for numerical stability, the
interval value should be higher than C<1/8192> (which is around 100
microseconds) and C<offset> should be higher than C<0> and should have
at most a similar magnitude as the current time (say, within a factor of
ten). Typical values for offset are, in fact, C<0> or something between
C<0> and C<interval>, which is also the recommended range.

Note also that there is an upper limit to how often a timer can fire (CPU
speed for example), so if C<interval> is very small then timing stability
will of course deteriorate. Libev itself tries to be exact to be about one
millisecond (if the OS supports it and the machine is fast enough).

=item * manual reschedule mode (offset ignored, interval ignored, reschedule_cb = callback)

In this mode the values for C<interval> and C<offset> are both being
ignored. Instead, each time the periodic watcher gets scheduled, the
reschedule callback will be called with the watcher as first, and the
current time as second argument.

NOTE: I<This callback MUST NOT stop or destroy any periodic watcher, ever,
or make ANY other event loop modifications whatsoever, unless explicitly
allowed by documentation here>.

If you need to stop it, return C<now + 1e30> (or so, fudge fudge) and stop
it afterwards (e.g. by starting an C<ev_prepare> watcher, which is the
only event loop modification you are allowed to do).

The callback prototype is C<ev_tstamp (*reschedule_cb)(ev_periodic
*w, ev_tstamp now)>, e.g.:

   static ev_tstamp
   my_rescheduler (ev_periodic *w, ev_tstamp now)
   {
     return now + 60.;
   }

It must return the next time to trigger, based on the passed time value
(that is, the lowest time value larger than to the second argument). It
will usually be called just before the callback will be triggered, but
might be called at other times, too.

NOTE: I<< This callback must always return a time that is higher than or
equal to the passed C<now> value >>.

This can be used to create very complex timers, such as a timer that
triggers on "next midnight, local time". To do this, you would calculate
the next midnight after C<now> and return the timestamp value for
this. Here is a (completely untested, no error checking) example on how to
do this:

   #include <time.h>

   static ev_tstamp
   my_rescheduler (ev_periodic *w, ev_tstamp now)
   {
     time_t tnow = (time_t)now;
     struct tm tm;
     localtime_r (&tnow, &tm);

     tm.tm_sec = tm.tm_min = tm.tm_hour = 0; // midnight current day
     ++tm.tm_mday; // midnight next day

     return mktime (&tm);
   }

Note: this code might run into trouble on days that have more then two
midnights (beginning and end).

=back

=item ev_periodic_again (loop, ev_periodic *)

Simply stops and restarts the periodic watcher again. This is only useful
when you changed some parameters or the reschedule callback would return
a different time than the last time it was called (e.g. in a crond like
program when the crontabs have changed).

=item ev_tstamp ev_periodic_at (ev_periodic *)

When active, returns the absolute time that the watcher is supposed
to trigger next. This is not the same as the C<offset> argument to
C<ev_periodic_set>, but indeed works even in interval and manual
rescheduling modes.

=item ev_tstamp offset [read-write]

When repeating, this contains the offset value, otherwise this is the
absolute point in time (the C<offset> value passed to C<ev_periodic_set>,
although libev might modify this value for better numerical stability).

Can be modified any time, but changes only take effect when the periodic
timer fires or C<ev_periodic_again> is being called.

=item ev_tstamp interval [read-write]

The current interval value. Can be modified any time, but changes only
take effect when the periodic timer fires or C<ev_periodic_again> is being
called.

=item ev_tstamp (*reschedule_cb)(ev_periodic *w, ev_tstamp now) [read-write]

The current reschedule callback, or C<0>, if this functionality is
switched off. Can be changed any time, but changes only take effect when
the periodic timer fires or C<ev_periodic_again> is being called.

=back

=head3 Examples

Example: Call a callback every hour, or, more precisely, whenever the
system time is divisible by 3600. The callback invocation times have
potentially a lot of jitter, but good long-term stability.

   static void
   clock_cb (struct ev_loop *loop, ev_periodic *w, int revents)
   {
     ... its now a full hour (UTC, or TAI or whatever your clock follows)
   }

   ev_periodic hourly_tick;
   ev_periodic_init (&hourly_tick, clock_cb, 0., 3600., 0);
   ev_periodic_start (loop, &hourly_tick);

Example: The same as above, but use a reschedule callback to do it:

   #include <math.h>

   static ev_tstamp
   my_scheduler_cb (ev_periodic *w, ev_tstamp now)
   {
     return now + (3600. - fmod (now, 3600.));
   }

   ev_periodic_init (&hourly_tick, clock_cb, 0., 0., my_scheduler_cb);

Example: Call a callback every hour, starting now:

   ev_periodic hourly_tick;
   ev_periodic_init (&hourly_tick, clock_cb,
                     fmod (ev_now (loop), 3600.), 3600., 0);
   ev_periodic_start (loop, &hourly_tick);


=head2 C<ev_signal> - signal me when a signal gets signalled!

Signal watchers will trigger an event when the process receives a specific
signal one or more times. Even though signals are very asynchronous, libev
will try its best to deliver signals synchronously, i.e. as part of the
normal event processing, like any other event.

If you want signals to be delivered truly asynchronously, just use
C<sigaction> as you would do without libev and forget about sharing
the signal. You can even use C<ev_async> from a signal handler to
synchronously wake up an event loop.

You can configure as many watchers as you like for the same signal, but
only within the same loop, i.e. you can watch for C<SIGINT> in your
default loop and for C<SIGIO> in another loop, but you cannot watch for
C<SIGINT> in both the default loop and another loop at the same time. At
the moment, C<SIGCHLD> is permanently tied to the default loop.

Only after the first watcher for a signal is started will libev actually
register something with the kernel. It thus coexists with your own signal
handlers as long as you don't register any with libev for the same signal.

If possible and supported, libev will install its handlers with
C<SA_RESTART> (or equivalent) behaviour enabled, so system calls should
not be unduly interrupted. If you have a problem with system calls getting
interrupted by signals you can block all signals in an C<ev_check> watcher
and unblock them in an C<ev_prepare> watcher.

=head3 The special problem of inheritance over fork/execve/pthread_create

Both the signal mask (C<sigprocmask>) and the signal disposition
(C<sigaction>) are unspecified after starting a signal watcher (and after
stopping it again), that is, libev might or might not block the signal,
and might or might not set or restore the installed signal handler (but
see C<EVFLAG_NOSIGMASK>).

While this does not matter for the signal disposition (libev never
sets signals to C<SIG_IGN>, so handlers will be reset to C<SIG_DFL> on
C<execve>), this matters for the signal mask: many programs do not expect
certain signals to be blocked.

This means that before calling C<exec> (from the child) you should reset
the signal mask to whatever "default" you expect (all clear is a good
choice usually).

The simplest way to ensure that the signal mask is reset in the child is
to install a fork handler with C<pthread_atfork> that resets it. That will
catch fork calls done by libraries (such as the libc) as well.

In current versions of libev, the signal will not be blocked indefinitely
unless you use the C<signalfd> API (C<EV_SIGNALFD>). While this reduces
the window of opportunity for problems, it will not go away, as libev
I<has> to modify the signal mask, at least temporarily.

So I can't stress this enough: I<If you do not reset your signal mask when
you expect it to be empty, you have a race condition in your code>. This
is not a libev-specific thing, this is true for most event libraries.

=head3 The special problem of threads signal handling

POSIX threads has problematic signal handling semantics, specifically,
a lot of functionality (sigfd, sigwait etc.) only really works if all
threads in a process block signals, which is hard to achieve.

When you want to use sigwait (or mix libev signal handling with your own
for the same signals), you can tackle this problem by globally blocking
all signals before creating any threads (or creating them with a fully set
sigprocmask) and also specifying the C<EVFLAG_NOSIGMASK> when creating
loops. Then designate one thread as "signal receiver thread" which handles
these signals. You can pass on any signals that libev might be interested
in by calling C<ev_feed_signal>.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_signal_init (ev_signal *, callback, int signum)

=item ev_signal_set (ev_signal *, int signum)

Configures the watcher to trigger on the given signal number (usually one
of the C<SIGxxx> constants).

=item int signum [read-only]

The signal the watcher watches out for.

=back

=head3 Examples

Example: Try to exit cleanly on SIGINT.

   static void
   sigint_cb (struct ev_loop *loop, ev_signal *w, int revents)
   {
     ev_break (loop, EVBREAK_ALL);
   }

   ev_signal signal_watcher;
   ev_signal_init (&signal_watcher, sigint_cb, SIGINT);
   ev_signal_start (loop, &signal_watcher);


=head2 C<ev_child> - watch out for process status changes

Child watchers trigger when your process receives a SIGCHLD in response to
some child status changes (most typically when a child of yours dies or
exits). It is permissible to install a child watcher I<after> the child
has been forked (which implies it might have already exited), as long
as the event loop isn't entered (or is continued from a watcher), i.e.,
forking and then immediately registering a watcher for the child is fine,
but forking and registering a watcher a few event loop iterations later or
in the next callback invocation is not.

Only the default event loop is capable of handling signals, and therefore
you can only register child watchers in the default event loop.

Due to some design glitches inside libev, child watchers will always be
handled at maximum priority (their priority is set to C<EV_MAXPRI> by
libev)

=head3 Process Interaction

Libev grabs C<SIGCHLD> as soon as the default event loop is
initialised. This is necessary to guarantee proper behaviour even if the
first child watcher is started after the child exits. The occurrence
of C<SIGCHLD> is recorded asynchronously, but child reaping is done
synchronously as part of the event loop processing. Libev always reaps all
children, even ones not watched.

=head3 Overriding the Built-In Processing

Libev offers no special support for overriding the built-in child
processing, but if your application collides with libev's default child
handler, you can override it easily by installing your own handler for
C<SIGCHLD> after initialising the default loop, and making sure the
default loop never gets destroyed. You are encouraged, however, to use an
event-based approach to child reaping and thus use libev's support for
that, so other libev users can use C<ev_child> watchers freely.

=head3 Stopping the Child Watcher

Currently, the child watcher never gets stopped, even when the
child terminates, so normally one needs to stop the watcher in the
callback. Future versions of libev might stop the watcher automatically
when a child exit is detected (calling C<ev_child_stop> twice is not a
problem).

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_child_init (ev_child *, callback, int pid, int trace)

=item ev_child_set (ev_child *, int pid, int trace)

Configures the watcher to wait for status changes of process C<pid> (or
I<any> process if C<pid> is specified as C<0>). The callback can look
at the C<rstatus> member of the C<ev_child> watcher structure to see
the status word (use the macros from C<sys/wait.h> and see your systems
C<waitpid> documentation). The C<rpid> member contains the pid of the
process causing the status change. C<trace> must be either C<0> (only
activate the watcher when the process terminates) or C<1> (additionally
activate the watcher when the process is stopped or continued).

=item int pid [read-only]

The process id this watcher watches out for, or C<0>, meaning any process id.

=item int rpid [read-write]

The process id that detected a status change.

=item int rstatus [read-write]

The process exit/trace status caused by C<rpid> (see your systems
C<waitpid> and C<sys/wait.h> documentation for details).

=back

=head3 Examples

Example: C<fork()> a new process and install a child handler to wait for
its completion.

   ev_child cw;

   static void
   child_cb (EV_P_ ev_child *w, int revents)
   {
     ev_child_stop (EV_A_ w);
     printf ("process %d exited with status %x\n", w->rpid, w->rstatus);
   }

   pid_t pid = fork ();

   if (pid < 0)
     // error
   else if (pid == 0)
     {
       // the forked child executes here
       exit (1);
     }
   else
     {
       ev_child_init (&cw, child_cb, pid, 0);
       ev_child_start (EV_DEFAULT_ &cw);
     }


=head2 C<ev_stat> - did the file attributes just change?

This watches a file system path for attribute changes. That is, it calls
C<stat> on that path in regular intervals (or when the OS says it changed)
and sees if it changed compared to the last time, invoking the callback
if it did. Starting the watcher C<stat>'s the file, so only changes that
happen after the watcher has been started will be reported.

The path does not need to exist: changing from "path exists" to "path does
not exist" is a status change like any other. The condition "path does not
exist" (or more correctly "path cannot be stat'ed") is signified by the
C<st_nlink> field being zero (which is otherwise always forced to be at
least one) and all the other fields of the stat buffer having unspecified
contents.

The path I<must not> end in a slash or contain special components such as
C<.> or C<..>. The path I<should> be absolute: If it is relative and
your working directory changes, then the behaviour is undefined.

Since there is no portable change notification interface available, the
portable implementation simply calls C<stat(2)> regularly on the path
to see if it changed somehow. You can specify a recommended polling
interval for this case. If you specify a polling interval of C<0> (highly
recommended!) then a I<suitable, unspecified default> value will be used
(which you can expect to be around five seconds, although this might
change dynamically). Libev will also impose a minimum interval which is
currently around C<0.1>, but that's usually overkill.

This watcher type is not meant for massive numbers of stat watchers,
as even with OS-supported change notifications, this can be
resource-intensive.

At the time of this writing, the only OS-specific interface implemented
is the Linux inotify interface (implementing kqueue support is left as an
exercise for the reader. Note, however, that the author sees no way of
implementing C<ev_stat> semantics with kqueue, except as a hint).

=head3 ABI Issues (Largefile Support)

Libev by default (unless the user overrides this) uses the default
compilation environment, which means that on systems with large file
support disabled by default, you get the 32 bit version of the stat
structure. When using the library from programs that change the ABI to
use 64 bit file offsets the programs will fail. In that case you have to
compile libev with the same flags to get binary compatibility. This is
obviously the case with any flags that change the ABI, but the problem is
most noticeably displayed with ev_stat and large file support.

The solution for this is to lobby your distribution maker to make large
file interfaces available by default (as e.g. FreeBSD does) and not
optional. Libev cannot simply switch on large file support because it has
to exchange stat structures with application programs compiled using the
default compilation environment.

=head3 Inotify and Kqueue

When C<inotify (7)> support has been compiled into libev and present at
runtime, it will be used to speed up change detection where possible. The
inotify descriptor will be created lazily when the first C<ev_stat>
watcher is being started.

Inotify presence does not change the semantics of C<ev_stat> watchers
except that changes might be detected earlier, and in some cases, to avoid
making regular C<stat> calls. Even in the presence of inotify support
there are many cases where libev has to resort to regular C<stat> polling,
but as long as kernel 2.6.25 or newer is used (2.6.24 and older have too
many bugs), the path exists (i.e. stat succeeds), and the path resides on
a local filesystem (libev currently assumes only ext2/3, jfs, reiserfs and
xfs are fully working) libev usually gets away without polling.

There is no support for kqueue, as apparently it cannot be used to
implement this functionality, due to the requirement of having a file
descriptor open on the object at all times, and detecting renames, unlinks
etc. is difficult.

=head3 C<stat ()> is a synchronous operation

Libev doesn't normally do any kind of I/O itself, and so is not blocking
the process. The exception are C<ev_stat> watchers - those call C<stat
()>, which is a synchronous operation.

For local paths, this usually doesn't matter: unless the system is very
busy or the intervals between stat's are large, a stat call will be fast,
as the path data is usually in memory already (except when starting the
watcher).

For networked file systems, calling C<stat ()> can block an indefinite
time due to network issues, and even under good conditions, a stat call
often takes multiple milliseconds.

Therefore, it is best to avoid using C<ev_stat> watchers on networked
paths, although this is fully supported by libev.

=head3 The special problem of stat time resolution

The C<stat ()> system call only supports full-second resolution portably,
and even on systems where the resolution is higher, most file systems
still only support whole seconds.

That means that, if the time is the only thing that changes, you can
easily miss updates: on the first update, C<ev_stat> detects a change and
calls your callback, which does something. When there is another update
within the same second, C<ev_stat> will be unable to detect unless the
stat data does change in other ways (e.g. file size).

The solution to this is to delay acting on a change for slightly more
than a second (or till slightly after the next full second boundary), using
a roughly one-second-delay C<ev_timer> (e.g. C<ev_timer_set (w, 0., 1.02);
ev_timer_again (loop, w)>).

The C<.02> offset is added to work around small timing inconsistencies
of some operating systems (where the second counter of the current time
might be be delayed. One such system is the Linux kernel, where a call to
C<gettimeofday> might return a timestamp with a full second later than
a subsequent C<time> call - if the equivalent of C<time ()> is used to
update file times then there will be a small window where the kernel uses
the previous second to update file times but libev might already execute
the timer callback).

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_stat_init (ev_stat *, callback, const char *path, ev_tstamp interval)

=item ev_stat_set (ev_stat *, const char *path, ev_tstamp interval)

Configures the watcher to wait for status changes of the given
C<path>. The C<interval> is a hint on how quickly a change is expected to
be detected and should normally be specified as C<0> to let libev choose
a suitable value. The memory pointed to by C<path> must point to the same
path for as long as the watcher is active.

The callback will receive an C<EV_STAT> event when a change was detected,
relative to the attributes at the time the watcher was started (or the
last change was detected).

=item ev_stat_stat (loop, ev_stat *)

Updates the stat buffer immediately with new values. If you change the
watched path in your callback, you could call this function to avoid
detecting this change (while introducing a race condition if you are not
the only one changing the path). Can also be useful simply to find out the
new values.

=item ev_statdata attr [read-only]

The most-recently detected attributes of the file. Although the type is
C<ev_statdata>, this is usually the (or one of the) C<struct stat> types
suitable for your system, but you can only rely on the POSIX-standardised
members to be present. If the C<st_nlink> member is C<0>, then there was
some error while C<stat>ing the file.

=item ev_statdata prev [read-only]

The previous attributes of the file. The callback gets invoked whenever
C<prev> != C<attr>, or, more precisely, one or more of these members
differ: C<st_dev>, C<st_ino>, C<st_mode>, C<st_nlink>, C<st_uid>,
C<st_gid>, C<st_rdev>, C<st_size>, C<st_atime>, C<st_mtime>, C<st_ctime>.

=item ev_tstamp interval [read-only]

The specified interval.

=item const char *path [read-only]

The file system path that is being watched.

=back

=head3 Examples

Example: Watch C</etc/passwd> for attribute changes.

   static void
   passwd_cb (struct ev_loop *loop, ev_stat *w, int revents)
   {
     /* /etc/passwd changed in some way */
     if (w->attr.st_nlink)
       {
         printf ("passwd current size  %ld\n", (long)w->attr.st_size);
         printf ("passwd current atime %ld\n", (long)w->attr.st_mtime);
         printf ("passwd current mtime %ld\n", (long)w->attr.st_mtime);
       }
     else
       /* you shalt not abuse printf for puts */
       puts ("wow, /etc/passwd is not there, expect problems. "
             "if this is windows, they already arrived\n");
   }

   ...
   ev_stat passwd;

   ev_stat_init (&passwd, passwd_cb, "/etc/passwd", 0.);
   ev_stat_start (loop, &passwd);

Example: Like above, but additionally use a one-second delay so we do not
miss updates (however, frequent updates will delay processing, too, so
one might do the work both on C<ev_stat> callback invocation I<and> on
C<ev_timer> callback invocation).

   static ev_stat passwd;
   static ev_timer timer;

   static void
   timer_cb (EV_P_ ev_timer *w, int revents)
   {
     ev_timer_stop (EV_A_ w);

     /* now it's one second after the most recent passwd change */
   }

   static void
   stat_cb (EV_P_ ev_stat *w, int revents)
   {
     /* reset the one-second timer */
     ev_timer_again (EV_A_ &timer);
   }

   ...
   ev_stat_init (&passwd, stat_cb, "/etc/passwd", 0.);
   ev_stat_start (loop, &passwd);
   ev_timer_init (&timer, timer_cb, 0., 1.02);


=head2 C<ev_idle> - when you've got nothing better to do...

Idle watchers trigger events when no other events of the same or higher
priority are pending (prepare, check and other idle watchers do not count
as receiving "events").

That is, as long as your process is busy handling sockets or timeouts
(or even signals, imagine) of the same or higher priority it will not be
triggered. But when your process is idle (or only lower-priority watchers
are pending), the idle watchers are being called once per event loop
iteration - until stopped, that is, or your process receives more events
and becomes busy again with higher priority stuff.

The most noteworthy effect is that as long as any idle watchers are
active, the process will not block when waiting for new events.

Apart from keeping your process non-blocking (which is a useful
effect on its own sometimes), idle watchers are a good place to do
"pseudo-background processing", or delay processing stuff to after the
event loop has handled all outstanding events.

=head3 Abusing an C<ev_idle> watcher for its side-effect

As long as there is at least one active idle watcher, libev will never
sleep unnecessarily. Or in other words, it will loop as fast as possible.
For this to work, the idle watcher doesn't need to be invoked at all - the
lowest priority will do.

This mode of operation can be useful together with an C<ev_check> watcher,
to do something on each event loop iteration - for example to balance load
between different connections.

See L</Abusing an ev_check watcher for its side-effect> for a longer
example.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_idle_init (ev_idle *, callback)

Initialises and configures the idle watcher - it has no parameters of any
kind. There is a C<ev_idle_set> macro, but using it is utterly pointless,
believe me.

=back

=head3 Examples

Example: Dynamically allocate an C<ev_idle> watcher, start it, and in the
callback, free it. Also, use no error checking, as usual.

   static void
   idle_cb (struct ev_loop *loop, ev_idle *w, int revents)
   {
     // stop the watcher
     ev_idle_stop (loop, w);

     // now we can free it
     free (w);

     // now do something you wanted to do when the program has
     // no longer anything immediate to do.
   }

   ev_idle *idle_watcher = malloc (sizeof (ev_idle));
   ev_idle_init (idle_watcher, idle_cb);
   ev_idle_start (loop, idle_watcher);


=head2 C<ev_prepare> and C<ev_check> - customise your event loop!

Prepare and check watchers are often (but not always) used in pairs:
prepare watchers get invoked before the process blocks and check watchers
afterwards.

You I<must not> call C<ev_run> (or similar functions that enter the
current event loop) or C<ev_loop_fork> from either C<ev_prepare> or
C<ev_check> watchers. Other loops than the current one are fine,
however. The rationale behind this is that you do not need to check
for recursion in those watchers, i.e. the sequence will always be
C<ev_prepare>, blocking, C<ev_check> so if you have one watcher of each
kind they will always be called in pairs bracketing the blocking call.

Their main purpose is to integrate other event mechanisms into libev and
their use is somewhat advanced. They could be used, for example, to track
variable changes, implement your own watchers, integrate net-snmp or a
coroutine library and lots more. They are also occasionally useful if
you cache some data and want to flush it before blocking (for example,
in X programs you might want to do an C<XFlush ()> in an C<ev_prepare>
watcher).

This is done by examining in each prepare call which file descriptors
need to be watched by the other library, registering C<ev_io> watchers
for them and starting an C<ev_timer> watcher for any timeouts (many
libraries provide exactly this functionality). Then, in the check watcher,
you check for any events that occurred (by checking the pending status
of all watchers and stopping them) and call back into the library. The
I/O and timer callbacks will never actually be called (but must be valid
nevertheless, because you never know, you know?).

As another example, the Perl Coro module uses these hooks to integrate
coroutines into libev programs, by yielding to other active coroutines
during each prepare and only letting the process block if no coroutines
are ready to run (it's actually more complicated: it only runs coroutines
with priority higher than or equal to the event loop and one coroutine
of lower priority, but only once, using idle watchers to keep the event
loop from blocking if lower-priority coroutines are active, thus mapping
low-priority coroutines to idle/background tasks).

When used for this purpose, it is recommended to give C<ev_check> watchers
highest (C<EV_MAXPRI>) priority, to ensure that they are being run before
any other watchers after the poll (this doesn't matter for C<ev_prepare>
watchers).

Also, C<ev_check> watchers (and C<ev_prepare> watchers, too) should not
activate ("feed") events into libev. While libev fully supports this, they
might get executed before other C<ev_check> watchers did their job. As
C<ev_check> watchers are often used to embed other (non-libev) event
loops those other event loops might be in an unusable state until their
C<ev_check> watcher ran (always remind yourself to coexist peacefully with
others).

=head3 Abusing an C<ev_check> watcher for its side-effect

C<ev_check> (and less often also C<ev_prepare>) watchers can also be
useful because they are called once per event loop iteration. For
example, if you want to handle a large number of connections fairly, you
normally only do a bit of work for each active connection, and if there
is more work to do, you wait for the next event loop iteration, so other
connections have a chance of making progress.

Using an C<ev_check> watcher is almost enough: it will be called on the
next event loop iteration. However, that isn't as soon as possible -
without external events, your C<ev_check> watcher will not be invoked.

This is where C<ev_idle> watchers come in handy - all you need is a
single global idle watcher that is active as long as you have one active
C<ev_check> watcher. The C<ev_idle> watcher makes sure the event loop
will not sleep, and the C<ev_check> watcher makes sure a callback gets
invoked. Neither watcher alone can do that.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_prepare_init (ev_prepare *, callback)

=item ev_check_init (ev_check *, callback)

Initialises and configures the prepare or check watcher - they have no
parameters of any kind. There are C<ev_prepare_set> and C<ev_check_set>
macros, but using them is utterly, utterly, utterly and completely
pointless.

=back

=head3 Examples

There are a number of principal ways to embed other event loops or modules
into libev. Here are some ideas on how to include libadns into libev
(there is a Perl module named C<EV::ADNS> that does this, which you could
use as a working example. Another Perl module named C<EV::Glib> embeds a
Glib main context into libev, and finally, C<Glib::EV> embeds EV into the
Glib event loop).

Method 1: Add IO watchers and a timeout watcher in a prepare handler,
and in a check watcher, destroy them and call into libadns. What follows
is pseudo-code only of course. This requires you to either use a low
priority for the check watcher or use C<ev_clear_pending> explicitly, as
the callbacks for the IO/timeout watchers might not have been called yet.

   static ev_io iow [nfd];
   static ev_timer tw;

   static void
   io_cb (struct ev_loop *loop, ev_io *w, int revents)
   {
   }

   // create io watchers for each fd and a timer before blocking
   static void
   adns_prepare_cb (struct ev_loop *loop, ev_prepare *w, int revents)
   {
     int timeout = 3600000;
     struct pollfd fds [nfd];
     // actual code will need to loop here and realloc etc.
     adns_beforepoll (ads, fds, &nfd, &timeout, timeval_from (ev_time ()));

     /* the callback is illegal, but won't be called as we stop during check */
     ev_timer_init (&tw, 0, timeout * 1e-3, 0.);
     ev_timer_start (loop, &tw);

     // create one ev_io per pollfd
     for (int i = 0; i < nfd; ++i)
       {
         ev_io_init (iow + i, io_cb, fds [i].fd,
           ((fds [i].events & POLLIN ? EV_READ : 0)
            | (fds [i].events & POLLOUT ? EV_WRITE : 0)));

         fds [i].revents = 0;
         ev_io_start (loop, iow + i);
       }
   }

   // stop all watchers after blocking
   static void
   adns_check_cb (struct ev_loop *loop, ev_check *w, int revents)
   {
     ev_timer_stop (loop, &tw);

     for (int i = 0; i < nfd; ++i)
       {
         // set the relevant poll flags
         // could also call adns_processreadable etc. here
         struct pollfd *fd = fds + i;
         int revents = ev_clear_pending (iow + i);
         if (revents & EV_READ ) fd->revents |= fd->events & POLLIN;
         if (revents & EV_WRITE) fd->revents |= fd->events & POLLOUT;

         // now stop the watcher
         ev_io_stop (loop, iow + i);
       }

     adns_afterpoll (adns, fds, nfd, timeval_from (ev_now (loop));
   }

Method 2: This would be just like method 1, but you run C<adns_afterpoll>
in the prepare watcher and would dispose of the check watcher.

Method 3: If the module to be embedded supports explicit event
notification (libadns does), you can also make use of the actual watcher
callbacks, and only destroy/create the watchers in the prepare watcher.

   static void
   timer_cb (EV_P_ ev_timer *w, int revents)
   {
     adns_state ads = (adns_state)w->data;
     update_now (EV_A);

     adns_processtimeouts (ads, &tv_now);
   }

   static void
   io_cb (EV_P_ ev_io *w, int revents)
   {
     adns_state ads = (adns_state)w->data;
     update_now (EV_A);

     if (revents & EV_READ ) adns_processreadable  (ads, w->fd, &tv_now);
     if (revents & EV_WRITE) adns_processwriteable (ads, w->fd, &tv_now);
   }

   // do not ever call adns_afterpoll

Method 4: Do not use a prepare or check watcher because the module you
want to embed is not flexible enough to support it. Instead, you can
override their poll function. The drawback with this solution is that the
main loop is now no longer controllable by EV. The C<Glib::EV> module uses
this approach, effectively embedding EV as a client into the horrible
libglib event loop.

   static gint
   event_poll_func (GPollFD *fds, guint nfds, gint timeout)
   {
     int got_events = 0;

     for (n = 0; n < nfds; ++n)
       // create/start io watcher that sets the relevant bits in fds[n] and increment got_events

     if (timeout >= 0)
       // create/start timer

     // poll
     ev_run (EV_A_ 0);

     // stop timer again
     if (timeout >= 0)
       ev_timer_stop (EV_A_ &to);

     // stop io watchers again - their callbacks should have set
     for (n = 0; n < nfds; ++n)
       ev_io_stop (EV_A_ iow [n]);

     return got_events;
   }


=head2 C<ev_embed> - when one backend isn't enough...

This is a rather advanced watcher type that lets you embed one event loop
into another (currently only C<ev_io> events are supported in the embedded
loop, other types of watchers might be handled in a delayed or incorrect
fashion and must not be used).

There are primarily two reasons you would want that: work around bugs and
prioritise I/O.

As an example for a bug workaround, the kqueue backend might only support
sockets on some platform, so it is unusable as generic backend, but you
still want to make use of it because you have many sockets and it scales
so nicely. In this case, you would create a kqueue-based loop and embed
it into your default loop (which might use e.g. poll). Overall operation
will be a bit slower because first libev has to call C<poll> and then
C<kevent>, but at least you can use both mechanisms for what they are
best: C<kqueue> for scalable sockets and C<poll> if you want it to work :)

As for prioritising I/O: under rare circumstances you have the case where
some fds have to be watched and handled very quickly (with low latency),
and even priorities and idle watchers might have too much overhead. In
this case you would put all the high priority stuff in one loop and all
the rest in a second one, and embed the second one in the first.

As long as the watcher is active, the callback will be invoked every
time there might be events pending in the embedded loop. The callback
must then call C<ev_embed_sweep (mainloop, watcher)> to make a single
sweep and invoke their callbacks (the callback doesn't need to invoke the
C<ev_embed_sweep> function directly, it could also start an idle watcher
to give the embedded loop strictly lower priority for example).

You can also set the callback to C<0>, in which case the embed watcher
will automatically execute the embedded loop sweep whenever necessary.

Fork detection will be handled transparently while the C<ev_embed> watcher
is active, i.e., the embedded loop will automatically be forked when the
embedding loop forks. In other cases, the user is responsible for calling
C<ev_loop_fork> on the embedded loop.

Unfortunately, not all backends are embeddable: only the ones returned by
C<ev_embeddable_backends> are, which, unfortunately, does not include any
portable one.

So when you want to use this feature you will always have to be prepared
that you cannot get an embeddable loop. The recommended way to get around
this is to have a separate variables for your embeddable loop, try to
create it, and if that fails, use the normal loop for everything.

=head3 C<ev_embed> and fork

While the C<ev_embed> watcher is running, forks in the embedding loop will
automatically be applied to the embedded loop as well, so no special
fork handling is required in that case. When the watcher is not running,
however, it is still the task of the libev user to call C<ev_loop_fork ()>
as applicable.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_embed_init (ev_embed *, callback, struct ev_loop *embedded_loop)

=item ev_embed_set (ev_embed *, struct ev_loop *embedded_loop)

Configures the watcher to embed the given loop, which must be
embeddable. If the callback is C<0>, then C<ev_embed_sweep> will be
invoked automatically, otherwise it is the responsibility of the callback
to invoke it (it will continue to be called until the sweep has been done,
if you do not want that, you need to temporarily stop the embed watcher).

=item ev_embed_sweep (loop, ev_embed *)

Make a single, non-blocking sweep over the embedded loop. This works
similarly to C<ev_run (embedded_loop, EVRUN_NOWAIT)>, but in the most
appropriate way for embedded loops.

=item struct ev_loop *other [read-only]

The embedded event loop.

=back

=head3 Examples

Example: Try to get an embeddable event loop and embed it into the default
event loop. If that is not possible, use the default loop. The default
loop is stored in C<loop_hi>, while the embeddable loop is stored in
C<loop_lo> (which is C<loop_hi> in the case no embeddable loop can be
used).

   struct ev_loop *loop_hi = ev_default_init (0);
   struct ev_loop *loop_lo = 0;
   ev_embed embed;

   // see if there is a chance of getting one that works
   // (remember that a flags value of 0 means autodetection)
   loop_lo = ev_embeddable_backends () & ev_recommended_backends ()
     ? ev_loop_new (ev_embeddable_backends () & ev_recommended_backends ())
     : 0;

   // if we got one, then embed it, otherwise default to loop_hi
   if (loop_lo)
     {
       ev_embed_init (&embed, 0, loop_lo);
       ev_embed_start (loop_hi, &embed);
     }
   else
     loop_lo = loop_hi;

Example: Check if kqueue is available but not recommended and create
a kqueue backend for use with sockets (which usually work with any
kqueue implementation). Store the kqueue/socket-only event loop in
C<loop_socket>. (One might optionally use C<EVFLAG_NOENV>, too).

   struct ev_loop *loop = ev_default_init (0);
   struct ev_loop *loop_socket = 0;
   ev_embed embed;

   if (ev_supported_backends () & ~ev_recommended_backends () & EVBACKEND_KQUEUE)
     if ((loop_socket = ev_loop_new (EVBACKEND_KQUEUE))
       {
         ev_embed_init (&embed, 0, loop_socket);
         ev_embed_start (loop, &embed);
       }

   if (!loop_socket)
     loop_socket = loop;

   // now use loop_socket for all sockets, and loop for everything else


=head2 C<ev_fork> - the audacity to resume the event loop after a fork

Fork watchers are called when a C<fork ()> was detected (usually because
whoever is a good citizen cared to tell libev about it by calling
C<ev_loop_fork>). The invocation is done before the event loop blocks next
and before C<ev_check> watchers are being called, and only in the child
after the fork. If whoever good citizen calling C<ev_default_fork> cheats
and calls it in the wrong process, the fork handlers will be invoked, too,
of course.

=head3 The special problem of life after fork - how is it possible?

Most uses of C<fork ()> consist of forking, then some simple calls to set
up/change the process environment, followed by a call to C<exec()>. This
sequence should be handled by libev without any problems.

This changes when the application actually wants to do event handling
in the child, or both parent in child, in effect "continuing" after the
fork.

The default mode of operation (for libev, with application help to detect
forks) is to duplicate all the state in the child, as would be expected
when I<either> the parent I<or> the child process continues.

When both processes want to continue using libev, then this is usually the
wrong result. In that case, usually one process (typically the parent) is
supposed to continue with all watchers in place as before, while the other
process typically wants to start fresh, i.e. without any active watchers.

The cleanest and most efficient way to achieve that with libev is to
simply create a new event loop, which of course will be "empty", and
use that for new watchers. This has the advantage of not touching more
memory than necessary, and thus avoiding the copy-on-write, and the
disadvantage of having to use multiple event loops (which do not support
signal watchers).

When this is not possible, or you want to use the default loop for
other reasons, then in the process that wants to start "fresh", call
C<ev_loop_destroy (EV_DEFAULT)> followed by C<ev_default_loop (...)>.
Destroying the default loop will "orphan" (not stop) all registered
watchers, so you have to be careful not to execute code that modifies
those watchers. Note also that in that case, you have to re-register any
signal watchers.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_fork_init (ev_fork *, callback)

Initialises and configures the fork watcher - it has no parameters of any
kind. There is a C<ev_fork_set> macro, but using it is utterly pointless,
really.

=back


=head2 C<ev_cleanup> - even the best things end

Cleanup watchers are called just before the event loop is being destroyed
by a call to C<ev_loop_destroy>.

While there is no guarantee that the event loop gets destroyed, cleanup
watchers provide a convenient method to install cleanup hooks for your
program, worker threads and so on - you just to make sure to destroy the
loop when you want them to be invoked.

Cleanup watchers are invoked in the same way as any other watcher. Unlike
all other watchers, they do not keep a reference to the event loop (which
makes a lot of sense if you think about it). Like all other watchers, you
can call libev functions in the callback, except C<ev_cleanup_start>.

=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_cleanup_init (ev_cleanup *, callback)

Initialises and configures the cleanup watcher - it has no parameters of
any kind. There is a C<ev_cleanup_set> macro, but using it is utterly
pointless, I assure you.

=back

Example: Register an atexit handler to destroy the default loop, so any
cleanup functions are called.

   static void
   program_exits (void)
   {
     ev_loop_destroy (EV_DEFAULT_UC);
   }

   ...
   atexit (program_exits);


=head2 C<ev_async> - how to wake up an event loop

In general, you cannot use an C<ev_loop> from multiple threads or other
asynchronous sources such as signal handlers (as opposed to multiple event
loops - those are of course safe to use in different threads).

Sometimes, however, you need to wake up an event loop you do not control,
for example because it belongs to another thread. This is what C<ev_async>
watchers do: as long as the C<ev_async> watcher is active, you can signal
it by calling C<ev_async_send>, which is thread- and signal safe.

This functionality is very similar to C<ev_signal> watchers, as signals,
too, are asynchronous in nature, and signals, too, will be compressed
(i.e. the number of callback invocations may be less than the number of
C<ev_async_send> calls). In fact, you could use signal watchers as a kind
of "global async watchers" by using a watcher on an otherwise unused
signal, and C<ev_feed_signal> to signal this watcher from another thread,
even without knowing which loop owns the signal.

=head3 Queueing

C<ev_async> does not support queueing of data in any way. The reason
is that the author does not know of a simple (or any) algorithm for a
multiple-writer-single-reader queue that works in all cases and doesn't
need elaborate support such as pthreads or unportable memory access
semantics.

That means that if you want to queue data, you have to provide your own
queue. But at least I can tell you how to implement locking around your
queue:

=over 4

=item queueing from a signal handler context

To implement race-free queueing, you simply add to the queue in the signal
handler but you block the signal handler in the watcher callback. Here is
an example that does that for some fictitious SIGUSR1 handler:

   static ev_async mysig;

   static void
   sigusr1_handler (void)
   {
     sometype data;

     // no locking etc.
     queue_put (data);
     ev_async_send (EV_DEFAULT_ &mysig);
   }

   static void
   mysig_cb (EV_P_ ev_async *w, int revents)
   {
     sometype data;
     sigset_t block, prev;

     sigemptyset (&block);
     sigaddset (&block, SIGUSR1);
     sigprocmask (SIG_BLOCK, &block, &prev);

     while (queue_get (&data))
       process (data);

     if (sigismember (&prev, SIGUSR1)
       sigprocmask (SIG_UNBLOCK, &block, 0);
   }

(Note: pthreads in theory requires you to use C<pthread_setmask>
instead of C<sigprocmask> when you use threads, but libev doesn't do it
either...).

=item queueing from a thread context

The strategy for threads is different, as you cannot (easily) block
threads but you can easily preempt them, so to queue safely you need to
employ a traditional mutex lock, such as in this pthread example:

   static ev_async mysig;
   static pthread_mutex_t mymutex = PTHREAD_MUTEX_INITIALIZER;

   static void
   otherthread (void)
   {
     // only need to lock the actual queueing operation
     pthread_mutex_lock (&mymutex);
     queue_put (data);
     pthread_mutex_unlock (&mymutex);

     ev_async_send (EV_DEFAULT_ &mysig);
   }

   static void
   mysig_cb (EV_P_ ev_async *w, int revents)
   {
     pthread_mutex_lock (&mymutex);

     while (queue_get (&data))
       process (data);

     pthread_mutex_unlock (&mymutex);
   }

=back


=head3 Watcher-Specific Functions and Data Members

=over 4

=item ev_async_init (ev_async *, callback)

Initialises and configures the async watcher - it has no parameters of any
kind. There is a C<ev_async_set> macro, but using it is utterly pointless,
trust me.

=item ev_async_send (loop, ev_async *)

Sends/signals/activates the given C<ev_async> watcher, that is, feeds
an C<EV_ASYNC> event on the watcher into the event loop, and instantly
returns.

Unlike C<ev_feed_event>, this call is safe to do from other threads,
signal or similar contexts (see the discussion of C<EV_ATOMIC_T> in the
embedding section below on what exactly this means).

Note that, as with other watchers in libev, multiple events might get
compressed into a single callback invocation (another way to look at
this is that C<ev_async> watchers are level-triggered: they are set on
C<ev_async_send>, reset when the event loop detects that).

This call incurs the overhead of at most one extra system call per event
loop iteration, if the event loop is blocked, and no syscall at all if
the event loop (or your program) is processing events. That means that
repeated calls are basically free (there is no need to avoid calls for
performance reasons) and that the overhead becomes smaller (typically
zero) under load.

=item bool = ev_async_pending (ev_async *)

Returns a non-zero value when C<ev_async_send> has been called on the
watcher but the event has not yet been processed (or even noted) by the
event loop.

C<ev_async_send> sets a flag in the watcher and wakes up the loop. When
the loop iterates next and checks for the watcher to have become active,
it will reset the flag again. C<ev_async_pending> can be used to very
quickly check whether invoking the loop might be a good idea.

Not that this does I<not> check whether the watcher itself is pending,
only whether it has been requested to make this watcher pending: there
is a time window between the event loop checking and resetting the async
notification, and the callback being invoked.

=back


=head1 OTHER FUNCTIONS

There are some other functions of possible interest. Described. Here. Now.

=over 4

=item ev_once (loop, int fd, int events, ev_tstamp timeout, callback, arg)

This function combines a simple timer and an I/O watcher, calls your
callback on whichever event happens first and automatically stops both
watchers. This is useful if you want to wait for a single event on an fd
or timeout without having to allocate/configure/start/stop/free one or
more watchers yourself.

If C<fd> is less than 0, then no I/O watcher will be started and the
C<events> argument is being ignored. Otherwise, an C<ev_io> watcher for
the given C<fd> and C<events> set will be created and started.

If C<timeout> is less than 0, then no timeout watcher will be
started. Otherwise an C<ev_timer> watcher with after = C<timeout> (and
repeat = 0) will be started. C<0> is a valid timeout.

The callback has the type C<void (*cb)(int revents, void *arg)> and is
passed an C<revents> set like normal event callbacks (a combination of
C<EV_ERROR>, C<EV_READ>, C<EV_WRITE> or C<EV_TIMER>) and the C<arg>
value passed to C<ev_once>. Note that it is possible to receive I<both>
a timeout and an io event at the same time - you probably should give io
events precedence.

Example: wait up to ten seconds for data to appear on STDIN_FILENO.

   static void stdin_ready (int revents, void *arg)
   {
     if (revents & EV_READ)
       /* stdin might have data for us, joy! */;
     else if (revents & EV_TIMER)
       /* doh, nothing entered */;
   }

   ev_once (STDIN_FILENO, EV_READ, 10., stdin_ready, 0);

=item ev_feed_fd_event (loop, int fd, int revents)

Feed an event on the given fd, as if a file descriptor backend detected
the given events.

=item ev_feed_signal_event (loop, int signum)

Feed an event as if the given signal occurred. See also C<ev_feed_signal>,
which is async-safe.

=back


=head1 COMMON OR USEFUL IDIOMS (OR BOTH)

This section explains some common idioms that are not immediately
obvious. Note that examples are sprinkled over the whole manual, and this
section only contains stuff that wouldn't fit anywhere else.

=head2 ASSOCIATING CUSTOM DATA WITH A WATCHER

Each watcher has, by default, a C<void *data> member that you can read
or modify at any time: libev will completely ignore it. This can be used
to associate arbitrary data with your watcher. If you need more data and
don't want to allocate memory separately and store a pointer to it in that
data member, you can also "subclass" the watcher type and provide your own
data:

   struct my_io
   {
     ev_io io;
     int otherfd;
     void *somedata;
     struct whatever *mostinteresting;
   };

   ...
   struct my_io w;
   ev_io_init (&w.io, my_cb, fd, EV_READ);

And since your callback will be called with a pointer to the watcher, you
can cast it back to your own type:

   static void my_cb (struct ev_loop *loop, ev_io *w_, int revents)
   {
     struct my_io *w = (struct my_io *)w_;
     ...
   }

More interesting and less C-conformant ways of casting your callback
function type instead have been omitted.

=head2 BUILDING YOUR OWN COMPOSITE WATCHERS

Another common scenario is to use some data structure with multiple
embedded watchers, in effect creating your own watcher that combines
multiple libev event sources into one "super-watcher":

   struct my_biggy
   {
     int some_data;
     ev_timer t1;
     ev_timer t2;
   }

In this case getting the pointer to C<my_biggy> is a bit more
complicated: Either you store the address of your C<my_biggy> struct in
the C<data> member of the watcher (for woozies or C++ coders), or you need
to use some pointer arithmetic using C<offsetof> inside your watchers (for
real programmers):

   #include <stddef.h>

   static void
   t1_cb (EV_P_ ev_timer *w, int revents)
   {
     struct my_biggy big = (struct my_biggy *)
       (((char *)w) - offsetof (struct my_biggy, t1));
   }

   static void
   t2_cb (EV_P_ ev_timer *w, int revents)
   {
     struct my_biggy big = (struct my_biggy *)
       (((char *)w) - offsetof (struct my_biggy, t2));
   }

=head2 AVOIDING FINISHING BEFORE RETURNING

Often you have structures like this in event-based programs:

  callback ()
  {
    free (request);
  }

  request = start_new_request (..., callback);

The intent is to start some "lengthy" operation. The C<request> could be
used to cancel the operation, or do other things with it.

It's not uncommon to have code paths in C<start_new_request> that
immediately invoke the callback, for example, to report errors. Or you add
some caching layer that finds that it can skip the lengthy aspects of the
operation and simply invoke the callback with the result.

The problem here is that this will happen I<before> C<start_new_request>
has returned, so C<request> is not set.

Even if you pass the request by some safer means to the callback, you
might want to do something to the request after starting it, such as
canceling it, which probably isn't working so well when the callback has
already been invoked.

A common way around all these issues is to make sure that
C<start_new_request> I<always> returns before the callback is invoked. If
C<start_new_request> immediately knows the result, it can artificially
delay invoking the callback by using a C<prepare> or C<idle> watcher for
example, or more sneakily, by reusing an existing (stopped) watcher and
pushing it into the pending queue:

   ev_set_cb (watcher, callback);
   ev_feed_event (EV_A_ watcher, 0);

This way, C<start_new_request> can safely return before the callback is
invoked, while not delaying callback invocation too much.

=head2 MODEL/NESTED EVENT LOOP INVOCATIONS AND EXIT CONDITIONS

Often (especially in GUI toolkits) there are places where you have
I<modal> interaction, which is most easily implemented by recursively
invoking C<ev_run>.

This brings the problem of exiting - a callback might want to finish the
main C<ev_run> call, but not the nested one (e.g. user clicked "Quit", but
a modal "Are you sure?" dialog is still waiting), or just the nested one
and not the main one (e.g. user clocked "Ok" in a modal dialog), or some
other combination: In these cases, a simple C<ev_break> will not work.

The solution is to maintain "break this loop" variable for each C<ev_run>
invocation, and use a loop around C<ev_run> until the condition is
triggered, using C<EVRUN_ONCE>:

   // main loop
   int exit_main_loop = 0;

   while (!exit_main_loop)
     ev_run (EV_DEFAULT_ EVRUN_ONCE);

   // in a modal watcher
   int exit_nested_loop = 0;

   while (!exit_nested_loop)
     ev_run (EV_A_ EVRUN_ONCE);

To exit from any of these loops, just set the corresponding exit variable:

   // exit modal loop
   exit_nested_loop = 1;

   // exit main program, after modal loop is finished
   exit_main_loop = 1;

   // exit both
   exit_main_loop = exit_nested_loop = 1;

=head2 THREAD LOCKING EXAMPLE

Here is a fictitious example of how to run an event loop in a different
thread from where callbacks are being invoked and watchers are
created/added/removed.

For a real-world example, see the C<EV::Loop::Async> perl module,
which uses exactly this technique (which is suited for many high-level
languages).

The example uses a pthread mutex to protect the loop data, a condition
variable to wait for callback invocations, an async watcher to notify the
event loop thread and an unspecified mechanism to wake up the main thread.

First, you need to associate some data with the event loop:

   typedef struct {
     mutex_t lock; /* global loop lock */
     ev_async async_w;
     thread_t tid;
     cond_t invoke_cv;
   } userdata;

   void prepare_loop (EV_P)
   {
      // for simplicity, we use a static userdata struct.
      static userdata u;

      ev_async_init (&u->async_w, async_cb);
      ev_async_start (EV_A_ &u->async_w);

      pthread_mutex_init (&u->lock, 0);
      pthread_cond_init (&u->invoke_cv, 0);

      // now associate this with the loop
      ev_set_userdata (EV_A_ u);
      ev_set_invoke_pending_cb (EV_A_ l_invoke);
      ev_set_loop_release_cb (EV_A_ l_release, l_acquire);

      // then create the thread running ev_run
      pthread_create (&u->tid, 0, l_run, EV_A);
   }

The callback for the C<ev_async> watcher does nothing: the watcher is used
solely to wake up the event loop so it takes notice of any new watchers
that might have been added:

   static void
   async_cb (EV_P_ ev_async *w, int revents)
   {
      // just used for the side effects
   }

The C<l_release> and C<l_acquire> callbacks simply unlock/lock the mutex
protecting the loop data, respectively.

   static void
   l_release (EV_P)
   {
     userdata *u = ev_userdata (EV_A);
     pthread_mutex_unlock (&u->lock);
   }

   static void
   l_acquire (EV_P)
   {
     userdata *u = ev_userdata (EV_A);
     pthread_mutex_lock (&u->lock);
   }

The event loop thread first acquires the mutex, and then jumps straight
into C<ev_run>:

   void *
   l_run (void *thr_arg)
   {
     struct ev_loop *loop = (struct ev_loop *)thr_arg;

     l_acquire (EV_A);
     pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
     ev_run (EV_A_ 0);
     l_release (EV_A);

     return 0;
   }

Instead of invoking all pending watchers, the C<l_invoke> callback will
signal the main thread via some unspecified mechanism (signals? pipe
writes? C<Async::Interrupt>?) and then waits until all pending watchers
have been called (in a while loop because a) spurious wakeups are possible
and b) skipping inter-thread-communication when there are no pending
watchers is very beneficial):

   static void
   l_invoke (EV_P)
   {
     userdata *u = ev_userdata (EV_A);

     while (ev_pending_count (EV_A))
       {
         wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
         pthread_cond_wait (&u->invoke_cv, &u->lock);
       }
   }

Now, whenever the main thread gets told to invoke pending watchers, it
will grab the lock, call C<ev_invoke_pending> and then signal the loop
thread to continue:

   static void
   real_invoke_pending (EV_P)
   {
     userdata *u = ev_userdata (EV_A);

     pthread_mutex_lock (&u->lock);
     ev_invoke_pending (EV_A);
     pthread_cond_signal (&u->invoke_cv);
     pthread_mutex_unlock (&u->lock);
   }

Whenever you want to start/stop a watcher or do other modifications to an
event loop, you will now have to lock:

   ev_timer timeout_watcher;
   userdata *u = ev_userdata (EV_A);

   ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);

   pthread_mutex_lock (&u->lock);
   ev_timer_start (EV_A_ &timeout_watcher);
   ev_async_send (EV_A_ &u->async_w);
   pthread_mutex_unlock (&u->lock);

Note that sending the C<ev_async> watcher is required because otherwise
an event loop currently blocking in the kernel will have no knowledge
about the newly added timer. By waking up the loop it will pick up any new
watchers in the next event loop iteration.

=head2 THREADS, COROUTINES, CONTINUATIONS, QUEUES... INSTEAD OF CALLBACKS

While the overhead of a callback that e.g. schedules a thread is small, it
is still an overhead. If you embed libev, and your main usage is with some
kind of threads or coroutines, you might want to customise libev so that
doesn't need callbacks anymore.

Imagine you have coroutines that you can switch to using a function
C<switch_to (coro)>, that libev runs in a coroutine called C<libev_coro>
and that due to some magic, the currently active coroutine is stored in a
global called C<current_coro>. Then you can build your own "wait for libev
event" primitive by changing C<EV_CB_DECLARE> and C<EV_CB_INVOKE> (note
the differing C<;> conventions):

   #define EV_CB_DECLARE(type)   struct my_coro *cb;
   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)

That means instead of having a C callback function, you store the
coroutine to switch to in each watcher, and instead of having libev call
your callback, you instead have it switch to that coroutine.

A coroutine might now wait for an event with a function called
C<wait_for_event>. (the watcher needs to be started, as always, but it doesn't
matter when, or whether the watcher is active or not when this function is
called):

   void
   wait_for_event (ev_watcher *w)
   {
     ev_set_cb (w, current_coro);
     switch_to (libev_coro);
   }

That basically suspends the coroutine inside C<wait_for_event> and
continues the libev coroutine, which, when appropriate, switches back to
this or any other coroutine.

You can do similar tricks if you have, say, threads with an event queue -
instead of storing a coroutine, you store the queue object and instead of
switching to a coroutine, you push the watcher onto the queue and notify
any waiters.

To embed libev, see L</EMBEDDING>, but in short, it's easiest to create two
files, F<my_ev.h> and F<my_ev.c> that include the respective libev files:

   // my_ev.h
   #define EV_CB_DECLARE(type)   struct my_coro *cb;
   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)
   #include "../libev/ev.h"

   // my_ev.c
   #define EV_H "my_ev.h"
   #include "../libev/ev.c"

And then use F<my_ev.h> when you would normally use F<ev.h>, and compile
F<my_ev.c> into your project. When properly specifying include paths, you
can even use F<ev.h> as header file name directly.


=head1 LIBEVENT EMULATION

Libev offers a compatibility emulation layer for libevent. It cannot
emulate the internals of libevent, so here are some usage hints:

=over 4

=item * Only the libevent-1.4.1-beta API is being emulated.

This was the newest libevent version available when libev was implemented,
and is still mostly unchanged in 2010.

=item * Use it by including <event.h>, as usual.

=item * The following members are fully supported: ev_base, ev_callback,
ev_arg, ev_fd, ev_res, ev_events.

=item * Avoid using ev_flags and the EVLIST_*-macros, while it is
maintained by libev, it does not work exactly the same way as in libevent (consider
it a private API).

=item * Priorities are not currently supported. Initialising priorities
will fail and all watchers will have the same priority, even though there
is an ev_pri field.

=item * In libevent, the last base created gets the signals, in libev, the
base that registered the signal gets the signals.

=item * Other members are not supported.

=item * The libev emulation is I<not> ABI compatible to libevent, you need
to use the libev header file and library.

=back

=head1 C++ SUPPORT

=head2 C API

The normal C API should work fine when used from C++: both ev.h and the
libev sources can be compiled as C++. Therefore, code that uses the C API
will work fine.

Proper exception specifications might have to be added to callbacks passed
to libev: exceptions may be thrown only from watcher callbacks, all other
callbacks (allocator, syserr, loop acquire/release and periodic reschedule
callbacks) must not throw exceptions, and might need a C<noexcept>
specification. If you have code that needs to be compiled as both C and
C++ you can use the C<EV_NOEXCEPT> macro for this:

   static void
   fatal_error (const char *msg) EV_NOEXCEPT
   {
     perror (msg);
     abort ();
   }

   ...
   ev_set_syserr_cb (fatal_error);

The only API functions that can currently throw exceptions are C<ev_run>,
C<ev_invoke>, C<ev_invoke_pending> and C<ev_loop_destroy> (the latter
because it runs cleanup watchers).

Throwing exceptions in watcher callbacks is only supported if libev itself
is compiled with a C++ compiler or your C and C++ environments allow
throwing exceptions through C libraries (most do).

=head2 C++ API

Libev comes with some simplistic wrapper classes for C++ that mainly allow
you to use some convenience methods to start/stop watchers and also change
the callback model to a model using method callbacks on objects.

To use it,

   #include <ev++.h>

This automatically includes F<ev.h> and puts all of its definitions (many
of them macros) into the global namespace. All C++ specific things are
put into the C<ev> namespace. It should support all the same embedding
options as F<ev.h>, most notably C<EV_MULTIPLICITY>.

Care has been taken to keep the overhead low. The only data member the C++
classes add (compared to plain C-style watchers) is the event loop pointer
that the watcher is associated with (or no additional members at all if
you disable C<EV_MULTIPLICITY> when embedding libev).

Currently, functions, static and non-static member functions and classes
with C<operator ()> can be used as callbacks. Other types should be easy
to add as long as they only need one additional pointer for context. If
you need support for other types of functors please contact the author
(preferably after implementing it).

For all this to work, your C++ compiler either has to use the same calling
conventions as your C compiler (for static member functions), or you have
to embed libev and compile libev itself as C++.

Here is a list of things available in the C<ev> namespace:

=over 4

=item C<ev::READ>, C<ev::WRITE> etc.

These are just enum values with the same values as the C<EV_READ> etc.
macros from F<ev.h>.

=item C<ev::tstamp>, C<ev::now>

Aliases to the same types/functions as with the C<ev_> prefix.

=item C<ev::io>, C<ev::timer>, C<ev::periodic>, C<ev::idle>, C<ev::sig> etc.

For each C<ev_TYPE> watcher in F<ev.h> there is a corresponding class of
the same name in the C<ev> namespace, with the exception of C<ev_signal>
which is called C<ev::sig> to avoid clashes with the C<signal> macro
defined by many implementations.

All of those classes have these methods:

=over 4

=item ev::TYPE::TYPE ()

=item ev::TYPE::TYPE (loop)

=item ev::TYPE::~TYPE

The constructor (optionally) takes an event loop to associate the watcher
with. If it is omitted, it will use C<EV_DEFAULT>.

The constructor calls C<ev_init> for you, which means you have to call the
C<set> method before starting it.

It will not set a callback, however: You have to call the templated C<set>
method to set a callback before you can start the watcher.

(The reason why you have to use a method is a limitation in C++ which does
not allow explicit template arguments for constructors).

The destructor automatically stops the watcher if it is active.

=item w->set<class, &class::method> (object *)

This method sets the callback method to call. The method has to have a
signature of C<void (*)(ev_TYPE &, int)>, it receives the watcher as
first argument and the C<revents> as second. The object must be given as
parameter and is stored in the C<data> member of the watcher.

This method synthesizes efficient thunking code to call your method from
the C callback that libev requires. If your compiler can inline your
callback (i.e. it is visible to it at the place of the C<set> call and
your compiler is good :), then the method will be fully inlined into the
thunking function, making it as fast as a direct C callback.

Example: simple class declaration and watcher initialisation

   struct myclass
   {
     void io_cb (ev::io &w, int revents) { }
   }

   myclass obj;
   ev::io iow;
   iow.set <myclass, &myclass::io_cb> (&obj);

=item w->set (object *)

This is a variation of a method callback - leaving out the method to call
will default the method to C<operator ()>, which makes it possible to use
functor objects without having to manually specify the C<operator ()> all
the time. Incidentally, you can then also leave out the template argument
list.

The C<operator ()> method prototype must be C<void operator ()(watcher &w,
int revents)>.

See the method-C<set> above for more details.

Example: use a functor object as callback.

   struct myfunctor
   {
     void operator() (ev::io &w, int revents)
     {
       ...
     }
   }

   myfunctor f;

   ev::io w;
   w.set (&f);

=item w->set<function> (void *data = 0)

Also sets a callback, but uses a static method or plain function as
callback. The optional C<data> argument will be stored in the watcher's
C<data> member and is free for you to use.

The prototype of the C<function> must be C<void (*)(ev::TYPE &w, int)>.

See the method-C<set> above for more details.

Example: Use a plain function as callback.

   static void io_cb (ev::io &w, int revents) { }
   iow.set <io_cb> ();

=item w->set (loop)

Associates a different C<struct ev_loop> with this watcher. You can only
do this when the watcher is inactive (and not pending either).

=item w->set ([arguments])

Basically the same as C<ev_TYPE_set> (except for C<ev::embed> watchers>),
with the same arguments. Either this method or a suitable start method
must be called at least once. Unlike the C counterpart, an active watcher
gets automatically stopped and restarted when reconfiguring it with this
method.

For C<ev::embed> watchers this method is called C<set_embed>, to avoid
clashing with the C<set (loop)> method.

=item w->start ()

Starts the watcher. Note that there is no C<loop> argument, as the
constructor already stores the event loop.

=item w->start ([arguments])

Instead of calling C<set> and C<start> methods separately, it is often
convenient to wrap them in one call. Uses the same type of arguments as
the configure C<set> method of the watcher.

=item w->stop ()

Stops the watcher if it is active. Again, no C<loop> argument.

=item w->again () (C<ev::timer>, C<ev::periodic> only)

For C<ev::timer> and C<ev::periodic>, this invokes the corresponding
C<ev_TYPE_again> function.

=item w->sweep () (C<ev::embed> only)

Invokes C<ev_embed_sweep>.

=item w->update () (C<ev::stat> only)

Invokes C<ev_stat_stat>.

=back

=back

Example: Define a class with two I/O and idle watchers, start the I/O
watchers in the constructor.

   class myclass
   {
     ev::io   io  ; void io_cb   (ev::io   &w, int revents);
     ev::io   io2 ; void io2_cb  (ev::io   &w, int revents);
     ev::idle idle; void idle_cb (ev::idle &w, int revents);

     myclass (int fd)
     {
       io  .set <myclass, &myclass::io_cb  > (this);
       io2 .set <myclass, &myclass::io2_cb > (this);
       idle.set <myclass, &myclass::idle_cb> (this);

       io.set (fd, ev::WRITE); // configure the watcher
       io.start ();            // start it whenever convenient

       io2.start (fd, ev::READ); // set + start in one call
     }
   };


=head1 OTHER LANGUAGE BINDINGS

Libev does not offer other language bindings itself, but bindings for a
number of languages exist in the form of third-party packages. If you know
any interesting language binding in addition to the ones listed here, drop
me a note.

=over 4

=item Perl

The EV module implements the full libev API and is actually used to test
libev. EV is developed together with libev. Apart from the EV core module,
there are additional modules that implement libev-compatible interfaces
to C<libadns> (C<EV::ADNS>, but C<AnyEvent::DNS> is preferred nowadays),
C<Net::SNMP> (C<Net::SNMP::EV>) and the C<libglib> event core (C<Glib::EV>
and C<EV::Glib>).

It can be found and installed via CPAN, its homepage is at
L<http://software.schmorp.de/pkg/EV>.

=item Python

Python bindings can be found at L<http://code.google.com/p/pyev/>. It
seems to be quite complete and well-documented.

=item Ruby

Tony Arcieri has written a ruby extension that offers access to a subset
of the libev API and adds file handle abstractions, asynchronous DNS and
more on top of it. It can be found via gem servers. Its homepage is at
L<http://rev.rubyforge.org/>.

Roger Pack reports that using the link order C<-lws2_32 -lmsvcrt-ruby-190>
makes rev work even on mingw.

=item Haskell

A haskell binding to libev is available at
L<http://hackage.haskell.org/cgi-bin/hackage-scripts/package/hlibev>.

=item D

Leandro Lucarella has written a D language binding (F<ev.d>) for libev, to
be found at L<http://www.llucax.com.ar/proj/ev.d/index.html>.

=item Ocaml

Erkki Seppala has written Ocaml bindings for libev, to be found at
L<http://modeemi.cs.tut.fi/~flux/software/ocaml-ev/>.

=item Lua

Brian Maher has written a partial interface to libev for lua (at the
time of this writing, only C<ev_io> and C<ev_timer>), to be found at
L<http://github.com/brimworks/lua-ev>.

=item Javascript

Node.js (L<http://nodejs.org>) uses libev as the underlying event library.

=item Others

There are others, and I stopped counting.

=back


=head1 MACRO MAGIC

Libev can be compiled with a variety of options, the most fundamental
of which is C<EV_MULTIPLICITY>. This option determines whether (most)
functions and callbacks have an initial C<struct ev_loop *> argument.

To make it easier to write programs that cope with either variant, the
following macros are defined:

=over 4

=item C<EV_A>, C<EV_A_>

This provides the loop I<argument> for functions, if one is required ("ev
loop argument"). The C<EV_A> form is used when this is the sole argument,
C<EV_A_> is used when other arguments are following. Example:

   ev_unref (EV_A);
   ev_timer_add (EV_A_ watcher);
   ev_run (EV_A_ 0);

It assumes the variable C<loop> of type C<struct ev_loop *> is in scope,
which is often provided by the following macro.

=item C<EV_P>, C<EV_P_>

This provides the loop I<parameter> for functions, if one is required ("ev
loop parameter"). The C<EV_P> form is used when this is the sole parameter,
C<EV_P_> is used when other parameters are following. Example:

   // this is how ev_unref is being declared
   static void ev_unref (EV_P);

   // this is how you can declare your typical callback
   static void cb (EV_P_ ev_timer *w, int revents)

It declares a parameter C<loop> of type C<struct ev_loop *>, quite
suitable for use with C<EV_A>.

=item C<EV_DEFAULT>, C<EV_DEFAULT_>

Similar to the other two macros, this gives you the value of the default
loop, if multiple loops are supported ("ev loop default"). The default loop
will be initialised if it isn't already initialised.

For non-multiplicity builds, these macros do nothing, so you always have
to initialise the loop somewhere.

=item C<EV_DEFAULT_UC>, C<EV_DEFAULT_UC_>

Usage identical to C<EV_DEFAULT> and C<EV_DEFAULT_>, but requires that the
default loop has been initialised (C<UC> == unchecked). Their behaviour
is undefined when the default loop has not been initialised by a previous
execution of C<EV_DEFAULT>, C<EV_DEFAULT_> or C<ev_default_init (...)>.

It is often prudent to use C<EV_DEFAULT> when initialising the first
watcher in a function but use C<EV_DEFAULT_UC> afterwards.

=back

Example: Declare and initialise a check watcher, utilising the above
macros so it will work regardless of whether multiple loops are supported
or not.

   static void
   check_cb (EV_P_ ev_timer *w, int revents)
   {
     ev_check_stop (EV_A_ w);
   }

   ev_check check;
   ev_check_init (&check, check_cb);
   ev_check_start (EV_DEFAULT_ &check);
   ev_run (EV_DEFAULT_ 0);

=head1 EMBEDDING

Libev can (and often is) directly embedded into host
applications. Examples of applications that embed it include the Deliantra
Game Server, the EV perl module, the GNU Virtual Private Ethernet (gvpe)
and rxvt-unicode.

The goal is to enable you to just copy the necessary files into your
source directory without having to change even a single line in them, so
you can easily upgrade by simply copying (or having a checked-out copy of
libev somewhere in your source tree).

=head2 FILESETS

Depending on what features you need you need to include one or more sets of files
in your application.

=head3 CORE EVENT LOOP

To include only the libev core (all the C<ev_*> functions), with manual
configuration (no autoconf):

   #define EV_STANDALONE 1
   #include "ev.c"

This will automatically include F<ev.h>, too, and should be done in a
single C source file only to provide the function implementations. To use
it, do the same for F<ev.h> in all files wishing to use this API (best
done by writing a wrapper around F<ev.h> that you can include instead and
where you can put other configuration options):

   #define EV_STANDALONE 1
   #include "ev.h"

Both header files and implementation files can be compiled with a C++
compiler (at least, that's a stated goal, and breakage will be treated
as a bug).

You need the following files in your source tree, or in a directory
in your include path (e.g. in libev/ when using -Ilibev):

   ev.h
   ev.c
   ev_vars.h
   ev_wrap.h

   ev_win32.c      required on win32 platforms only

   ev_select.c     only when select backend is enabled
   ev_poll.c       only when poll backend is enabled
   ev_epoll.c      only when the epoll backend is enabled
   ev_linuxaio.c   only when the linux aio backend is enabled
   ev_iouring.c    only when the linux io_uring backend is enabled
   ev_kqueue.c     only when the kqueue backend is enabled
   ev_port.c       only when the solaris port backend is enabled

F<ev.c> includes the backend files directly when enabled, so you only need
to compile this single file.

=head3 LIBEVENT COMPATIBILITY API

To include the libevent compatibility API, also include:

   #include "event.c"

in the file including F<ev.c>, and:

   #include "event.h"

in the files that want to use the libevent API. This also includes F<ev.h>.

You need the following additional files for this:

   event.h
   event.c

=head3 AUTOCONF SUPPORT

Instead of using C<EV_STANDALONE=1> and providing your configuration in
whatever way you want, you can also C<m4_include([libev.m4])> in your
F<configure.ac> and leave C<EV_STANDALONE> undefined. F<ev.c> will then
include F<config.h> and configure itself accordingly.

For this of course you need the m4 file:

   libev.m4

=head2 PREPROCESSOR SYMBOLS/MACROS

Libev can be configured via a variety of preprocessor symbols you have to
define before including (or compiling) any of its files. The default in
the absence of autoconf is documented for every option.

Symbols marked with "(h)" do not change the ABI, and can have different
values when compiling libev vs. including F<ev.h>, so it is permissible
to redefine them before including F<ev.h> without breaking compatibility
to a compiled library. All other symbols change the ABI, which means all
users of libev and the libev code itself must be compiled with compatible
settings.

=over 4

=item EV_COMPAT3 (h)

Backwards compatibility is a major concern for libev. This is why this
release of libev comes with wrappers for the functions and symbols that
have been renamed between libev version 3 and 4.

You can disable these wrappers (to test compatibility with future
versions) by defining C<EV_COMPAT3> to C<0> when compiling your
sources. This has the additional advantage that you can drop the C<struct>
from C<struct ev_loop> declarations, as libev will provide an C<ev_loop>
typedef in that case.

In some future version, the default for C<EV_COMPAT3> will become C<0>,
and in some even more future version the compatibility code will be
removed completely.

=item EV_STANDALONE (h)

Must always be C<1> if you do not use autoconf configuration, which
keeps libev from including F<config.h>, and it also defines dummy
implementations for some libevent functions (such as logging, which is not
supported). It will also not define any of the structs usually found in
F<event.h> that are not directly supported by the libev core alone.

In standalone mode, libev will still try to automatically deduce the
configuration, but has to be more conservative.

=item EV_USE_FLOOR

If defined to be C<1>, libev will use the C<floor ()> function for its
periodic reschedule calculations, otherwise libev will fall back on a
portable (slower) implementation. If you enable this, you usually have to
link against libm or something equivalent. Enabling this when the C<floor>
function is not available will fail, so the safe default is to not enable
this.

=item EV_USE_MONOTONIC

If defined to be C<1>, libev will try to detect the availability of the
monotonic clock option at both compile time and runtime. Otherwise no
use of the monotonic clock option will be attempted. If you enable this,
you usually have to link against librt or something similar. Enabling it
when the functionality isn't available is safe, though, although you have
to make sure you link against any libraries where the C<clock_gettime>
function is hiding in (often F<-lrt>). See also C<EV_USE_CLOCK_SYSCALL>.

=item EV_USE_REALTIME

If defined to be C<1>, libev will try to detect the availability of the
real-time clock option at compile time (and assume its availability
at runtime if successful). Otherwise no use of the real-time clock
option will be attempted. This effectively replaces C<gettimeofday>
by C<clock_get (CLOCK_REALTIME, ...)> and will not normally affect
correctness. See the note about libraries in the description of
C<EV_USE_MONOTONIC>, though. Defaults to the opposite value of
C<EV_USE_CLOCK_SYSCALL>.

=item EV_USE_CLOCK_SYSCALL

If defined to be C<1>, libev will try to use a direct syscall instead
of calling the system-provided C<clock_gettime> function. This option
exists because on GNU/Linux, C<clock_gettime> is in C<librt>, but C<librt>
unconditionally pulls in C<libpthread>, slowing down single-threaded
programs needlessly. Using a direct syscall is slightly slower (in
theory), because no optimised vdso implementation can be used, but avoids
the pthread dependency. Defaults to C<1> on GNU/Linux with glibc 2.x or
higher, as it simplifies linking (no need for C<-lrt>).

=item EV_USE_NANOSLEEP

If defined to be C<1>, libev will assume that C<nanosleep ()> is available
and will use it for delays. Otherwise it will use C<select ()>.

=item EV_USE_EVENTFD

If defined to be C<1>, then libev will assume that C<eventfd ()> is
available and will probe for kernel support at runtime. This will improve
C<ev_signal> and C<ev_async> performance and reduce resource consumption.
If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc
2.7 or newer, otherwise disabled.

=item EV_USE_SIGNALFD

If defined to be C<1>, then libev will assume that C<signalfd ()> is
available and will probe for kernel support at runtime. This enables
the use of EVFLAG_SIGNALFD for faster and simpler signal handling. If
undefined, it will be enabled if the headers indicate GNU/Linux + Glibc
2.7 or newer, otherwise disabled.

=item EV_USE_TIMERFD

If defined to be C<1>, then libev will assume that C<timerfd ()> is
available and will probe for kernel support at runtime. This allows
libev to detect time jumps accurately. If undefined, it will be enabled
if the headers indicate GNU/Linux + Glibc 2.8 or newer and define
C<TFD_TIMER_CANCEL_ON_SET>, otherwise disabled.

=item EV_USE_EVENTFD

If defined to be C<1>, then libev will assume that C<eventfd ()> is
available and will probe for kernel support at runtime. This will improve
C<ev_signal> and C<ev_async> performance and reduce resource consumption.
If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc
2.7 or newer, otherwise disabled.

=item EV_USE_SELECT

If undefined or defined to be C<1>, libev will compile in support for the
C<select>(2) backend. No attempt at auto-detection will be done: if no
other method takes over, select will be it. Otherwise the select backend
will not be compiled in.

=item EV_SELECT_USE_FD_SET

If defined to C<1>, then the select backend will use the system C<fd_set>
structure. This is useful if libev doesn't compile due to a missing
C<NFDBITS> or C<fd_mask> definition or it mis-guesses the bitset layout
on exotic systems. This usually limits the range of file descriptors to
some low limit such as 1024 or might have other limitations (winsocket
only allows 64 sockets). The C<FD_SETSIZE> macro, set before compilation,
configures the maximum size of the C<fd_set>.

=item EV_SELECT_IS_WINSOCKET

When defined to C<1>, the select backend will assume that
select/socket/connect etc. don't understand file descriptors but
wants osf handles on win32 (this is the case when the select to
be used is the winsock select). This means that it will call
C<_get_osfhandle> on the fd to convert it to an OS handle. Otherwise,
it is assumed that all these functions actually work on fds, even
on win32. Should not be defined on non-win32 platforms.

=item EV_FD_TO_WIN32_HANDLE(fd)

If C<EV_SELECT_IS_WINSOCKET> is enabled, then libev needs a way to map
file descriptors to socket handles. When not defining this symbol (the
default), then libev will call C<_get_osfhandle>, which is usually
correct. In some cases, programs use their own file descriptor management,
in which case they can provide this function to map fds to socket handles.

=item EV_WIN32_HANDLE_TO_FD(handle)

If C<EV_SELECT_IS_WINSOCKET> then libev maps handles to file descriptors
using the standard C<_open_osfhandle> function. For programs implementing
their own fd to handle mapping, overwriting this function makes it easier
to do so. This can be done by defining this macro to an appropriate value.

=item EV_WIN32_CLOSE_FD(fd)

If programs implement their own fd to handle mapping on win32, then this
macro can be used to override the C<close> function, useful to unregister
file descriptors again. Note that the replacement function has to close
the underlying OS handle.

=item EV_USE_WSASOCKET

If defined to be C<1>, libev will use C<WSASocket> to create its internal
communication socket, which works better in some environments. Otherwise,
the normal C<socket> function will be used, which works better in other
environments.

=item EV_USE_POLL

If defined to be C<1>, libev will compile in support for the C<poll>(2)
backend. Otherwise it will be enabled on non-win32 platforms. It
takes precedence over select.

=item EV_USE_EPOLL

If defined to be C<1>, libev will compile in support for the Linux
C<epoll>(7) backend. Its availability will be detected at runtime,
otherwise another method will be used as fallback. This is the preferred
backend for GNU/Linux systems. If undefined, it will be enabled if the
headers indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.

=item EV_USE_LINUXAIO

If defined to be C<1>, libev will compile in support for the Linux aio
backend (C<EV_USE_EPOLL> must also be enabled). If undefined, it will be
enabled on linux, otherwise disabled.

=item EV_USE_IOURING

If defined to be C<1>, libev will compile in support for the Linux
io_uring backend (C<EV_USE_EPOLL> must also be enabled). Due to it's
current limitations it has to be requested explicitly. If undefined, it
will be enabled on linux, otherwise disabled.

=item EV_USE_KQUEUE

If defined to be C<1>, libev will compile in support for the BSD style
C<kqueue>(2) backend. Its actual availability will be detected at runtime,
otherwise another method will be used as fallback. This is the preferred
backend for BSD and BSD-like systems, although on most BSDs kqueue only
supports some types of fds correctly (the only platform we found that
supports ptys for example was NetBSD), so kqueue might be compiled in, but
not be used unless explicitly requested. The best way to use it is to find
out whether kqueue supports your type of fd properly and use an embedded
kqueue loop.

=item EV_USE_PORT

If defined to be C<1>, libev will compile in support for the Solaris
10 port style backend. Its availability will be detected at runtime,
otherwise another method will be used as fallback. This is the preferred
backend for Solaris 10 systems.

=item EV_USE_DEVPOLL

Reserved for future expansion, works like the USE symbols above.

=item EV_USE_INOTIFY

If defined to be C<1>, libev will compile in support for the Linux inotify
interface to speed up C<ev_stat> watchers. Its actual availability will
be detected at runtime. If undefined, it will be enabled if the headers
indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.

=item EV_NO_SMP

If defined to be C<1>, libev will assume that memory is always coherent
between threads, that is, threads can be used, but threads never run on
different cpus (or different cpu cores). This reduces dependencies
and makes libev faster.

=item EV_NO_THREADS

If defined to be C<1>, libev will assume that it will never be called from
different threads (that includes signal handlers), which is a stronger
assumption than C<EV_NO_SMP>, above. This reduces dependencies and makes
libev faster.

=item EV_ATOMIC_T

Libev requires an integer type (suitable for storing C<0> or C<1>) whose
access is atomic with respect to other threads or signal contexts. No
such type is easily found in the C language, so you can provide your own
type that you know is safe for your purposes. It is used both for signal
handler "locking" as well as for signal and thread safety in C<ev_async>
watchers.

In the absence of this define, libev will use C<sig_atomic_t volatile>
(from F<signal.h>), which is usually good enough on most platforms.

=item EV_H (h)

The name of the F<ev.h> header file used to include it. The default if
undefined is C<"ev.h"> in F<event.h>, F<ev.c> and F<ev++.h>. This can be
used to virtually rename the F<ev.h> header file in case of conflicts.

=item EV_CONFIG_H (h)

If C<EV_STANDALONE> isn't C<1>, this variable can be used to override
F<ev.c>'s idea of where to find the F<config.h> file, similarly to
C<EV_H>, above.

=item EV_EVENT_H (h)

Similarly to C<EV_H>, this macro can be used to override F<event.c>'s idea
of how the F<event.h> header can be found, the default is C<"event.h">.

=item EV_PROTOTYPES (h)

If defined to be C<0>, then F<ev.h> will not define any function
prototypes, but still define all the structs and other symbols. This is
occasionally useful if you want to provide your own wrapper functions
around libev functions.

=item EV_MULTIPLICITY

If undefined or defined to C<1>, then all event-loop-specific functions
will have the C<struct ev_loop *> as first argument, and you can create
additional independent event loops. Otherwise there will be no support
for multiple event loops and there is no first event loop pointer
argument. Instead, all functions act on the single default loop.

Note that C<EV_DEFAULT> and C<EV_DEFAULT_> will no longer provide a
default loop when multiplicity is switched off - you always have to
initialise the loop manually in this case.

=item EV_MINPRI

=item EV_MAXPRI

The range of allowed priorities. C<EV_MINPRI> must be smaller or equal to
C<EV_MAXPRI>, but otherwise there are no non-obvious limitations. You can
provide for more priorities by overriding those symbols (usually defined
to be C<-2> and C<2>, respectively).

When doing priority-based operations, libev usually has to linearly search
all the priorities, so having many of them (hundreds) uses a lot of space
and time, so using the defaults of five priorities (-2 .. +2) is usually
fine.

If your embedding application does not need any priorities, defining these
both to C<0> will save some memory and CPU.

=item EV_PERIODIC_ENABLE, EV_IDLE_ENABLE, EV_EMBED_ENABLE, EV_STAT_ENABLE,
EV_PREPARE_ENABLE, EV_CHECK_ENABLE, EV_FORK_ENABLE, EV_SIGNAL_ENABLE,
EV_ASYNC_ENABLE, EV_CHILD_ENABLE.

If undefined or defined to be C<1> (and the platform supports it), then
the respective watcher type is supported. If defined to be C<0>, then it
is not. Disabling watcher types mainly saves code size.

=item EV_FEATURES

If you need to shave off some kilobytes of code at the expense of some
speed (but with the full API), you can define this symbol to request
certain subsets of functionality. The default is to enable all features
that can be enabled on the platform.

A typical way to use this symbol is to define it to C<0> (or to a bitset
with some broad features you want) and then selectively re-enable
additional parts you want, for example if you want everything minimal,
but multiple event loop support, async and child watchers and the poll
backend, use this:

   #define EV_FEATURES 0
   #define EV_MULTIPLICITY 1
   #define EV_USE_POLL 1
   #define EV_CHILD_ENABLE 1
   #define EV_ASYNC_ENABLE 1

The actual value is a bitset, it can be a combination of the following
values (by default, all of these are enabled):

=over 4

=item C<1> - faster/larger code

Use larger code to speed up some operations.

Currently this is used to override some inlining decisions (enlarging the
code size by roughly 30% on amd64).

When optimising for size, use of compiler flags such as C<-Os> with
gcc is recommended, as well as C<-DNDEBUG>, as libev contains a number of
assertions.

The default is off when C<__OPTIMIZE_SIZE__> is defined by your compiler
(e.g. gcc with C<-Os>).

=item C<2> - faster/larger data structures

Replaces the small 2-heap for timer management by a faster 4-heap, larger
hash table sizes and so on. This will usually further increase code size
and can additionally have an effect on the size of data structures at
runtime.

The default is off when C<__OPTIMIZE_SIZE__> is defined by your compiler
(e.g. gcc with C<-Os>).

=item C<4> - full API configuration

This enables priorities (sets C<EV_MAXPRI>=2 and C<EV_MINPRI>=-2), and
enables multiplicity (C<EV_MULTIPLICITY>=1).

=item C<8> - full API

This enables a lot of the "lesser used" API functions. See C<ev.h> for
details on which parts of the API are still available without this
feature, and do not complain if this subset changes over time.

=item C<16> - enable all optional watcher types

Enables all optional watcher types.  If you want to selectively enable
only some watcher types other than I/O and timers (e.g. prepare,
embed, async, child...) you can enable them manually by defining
C<EV_watchertype_ENABLE> to C<1> instead.

=item C<32> - enable all backends

This enables all backends - without this feature, you need to enable at
least one backend manually (C<EV_USE_SELECT> is a good choice).

=item C<64> - enable OS-specific "helper" APIs

Enable inotify, eventfd, signalfd and similar OS-specific helper APIs by
default.

=back

Compiling with C<gcc -Os -DEV_STANDALONE -DEV_USE_EPOLL=1 -DEV_FEATURES=0>
reduces the compiled size of libev from 24.7Kb code/2.8Kb data to 6.5Kb
code/0.3Kb data on my GNU/Linux amd64 system, while still giving you I/O
watchers, timers and monotonic clock support.

With an intelligent-enough linker (gcc+binutils are intelligent enough
when you use C<-Wl,--gc-sections -ffunction-sections>) functions unused by
your program might be left out as well - a binary starting a timer and an
I/O watcher then might come out at only 5Kb.

=item EV_API_STATIC

If this symbol is defined (by default it is not), then all identifiers
will have static linkage. This means that libev will not export any
identifiers, and you cannot link against libev anymore. This can be useful
when you embed libev, only want to use libev functions in a single file,
and do not want its identifiers to be visible.

To use this, define C<EV_API_STATIC> and include F<ev.c> in the file that
wants to use libev.

This option only works when libev is compiled with a C compiler, as C++
doesn't support the required declaration syntax.

=item EV_AVOID_STDIO

If this is set to C<1> at compiletime, then libev will avoid using stdio
functions (printf, scanf, perror etc.). This will increase the code size
somewhat, but if your program doesn't otherwise depend on stdio and your
libc allows it, this avoids linking in the stdio library which is quite
big.

Note that error messages might become less precise when this option is
enabled.

=item EV_NSIG

The highest supported signal number, +1 (or, the number of
signals): Normally, libev tries to deduce the maximum number of signals
automatically, but sometimes this fails, in which case it can be
specified. Also, using a lower number than detected (C<32> should be
good for about any system in existence) can save some memory, as libev
statically allocates some 12-24 bytes per signal number.

=item EV_PID_HASHSIZE

C<ev_child> watchers use a small hash table to distribute workload by
pid. The default size is C<16> (or C<1> with C<EV_FEATURES> disabled),
usually more than enough. If you need to manage thousands of children you
might want to increase this value (I<must> be a power of two).

=item EV_INOTIFY_HASHSIZE

C<ev_stat> watchers use a small hash table to distribute workload by
inotify watch id. The default size is C<16> (or C<1> with C<EV_FEATURES>
disabled), usually more than enough. If you need to manage thousands of
C<ev_stat> watchers you might want to increase this value (I<must> be a
power of two).

=item EV_USE_4HEAP

Heaps are not very cache-efficient. To improve the cache-efficiency of the
timer and periodics heaps, libev uses a 4-heap when this symbol is defined
to C<1>. The 4-heap uses more complicated (longer) code but has noticeably
faster performance with many (thousands) of watchers.

The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
will be C<0>.

=item EV_HEAP_CACHE_AT

Heaps are not very cache-efficient. To improve the cache-efficiency of the
timer and periodics heaps, libev can cache the timestamp (I<at>) within
the heap structure (selected by defining C<EV_HEAP_CACHE_AT> to C<1>),
which uses 8-12 bytes more per watcher and a few hundred bytes more code,
but avoids random read accesses on heap changes. This improves performance
noticeably with many (hundreds) of watchers.

The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
will be C<0>.

=item EV_VERIFY

Controls how much internal verification (see C<ev_verify ()>) will
be done: If set to C<0>, no internal verification code will be compiled
in. If set to C<1>, then verification code will be compiled in, but not
called. If set to C<2>, then the internal verification code will be
called once per loop, which can slow down libev. If set to C<3>, then the
verification code will be called very frequently, which will slow down
libev considerably.

Verification errors are reported via C's C<assert> mechanism, so if you
disable that (e.g. by defining C<NDEBUG>) then no errors will be reported.

The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
will be C<0>.

=item EV_COMMON

By default, all watchers have a C<void *data> member. By redefining
this macro to something else you can include more and other types of
members. You have to define it each time you include one of the files,
though, and it must be identical each time.

For example, the perl EV module uses something like this:

   #define EV_COMMON                       \
     SV *self; /* contains this struct */  \
     SV *cb_sv, *fh /* note no trailing ";" */

=item EV_CB_DECLARE (type)

=item EV_CB_INVOKE (watcher, revents)

=item ev_set_cb (ev, cb)

Can be used to change the callback member declaration in each watcher,
and the way callbacks are invoked and set. Must expand to a struct member
definition and a statement, respectively. See the F<ev.h> header file for
their default definitions. One possible use for overriding these is to
avoid the C<struct ev_loop *> as first argument in all cases, or to use
method calls instead of plain function calls in C++.

=back

=head2 EXPORTED API SYMBOLS

If you need to re-export the API (e.g. via a DLL) and you need a list of
exported symbols, you can use the provided F<Symbol.*> files which list
all public symbols, one per line:

   Symbols.ev      for libev proper
   Symbols.event   for the libevent emulation

This can also be used to rename all public symbols to avoid clashes with
multiple versions of libev linked together (which is obviously bad in
itself, but sometimes it is inconvenient to avoid this).

A sed command like this will create wrapper C<#define>'s that you need to
include before including F<ev.h>:

   <Symbols.ev sed -e "s/.*/#define & myprefix_&/" >wrap.h

This would create a file F<wrap.h> which essentially looks like this:

   #define ev_backend     myprefix_ev_backend
   #define ev_check_start myprefix_ev_check_start
   #define ev_check_stop  myprefix_ev_check_stop
   ...

=head2 EXAMPLES

For a real-world example of a program the includes libev
verbatim, you can have a look at the EV perl module
(L<http://software.schmorp.de/pkg/EV.html>). It has the libev files in
the F<libev/> subdirectory and includes them in the F<EV/EVAPI.h> (public
interface) and F<EV.xs> (implementation) files. Only the F<EV.xs> file
will be compiled. It is pretty complex because it provides its own header
file.

The usage in rxvt-unicode is simpler. It has a F<ev_cpp.h> header file
that everybody includes and which overrides some configure choices:

   #define EV_FEATURES 8
   #define EV_USE_SELECT 1
   #define EV_PREPARE_ENABLE 1
   #define EV_IDLE_ENABLE 1
   #define EV_SIGNAL_ENABLE 1
   #define EV_CHILD_ENABLE 1
   #define EV_USE_STDEXCEPT 0
   #define EV_CONFIG_H <config.h>

   #include "ev++.h"

And a F<ev_cpp.C> implementation file that contains libev proper and is compiled:

   #include "ev_cpp.h"
   #include "ev.c"

=head1 INTERACTION WITH OTHER PROGRAMS, LIBRARIES OR THE ENVIRONMENT

=head2 THREADS AND COROUTINES

=head3 THREADS

All libev functions are reentrant and thread-safe unless explicitly
documented otherwise, but libev implements no locking itself. This means
that you can use as many loops as you want in parallel, as long as there
are no concurrent calls into any libev function with the same loop
parameter (C<ev_default_*> calls have an implicit default loop parameter,
of course): libev guarantees that different event loops share no data
structures that need any locking.

Or to put it differently: calls with different loop parameters can be done
concurrently from multiple threads, calls with the same loop parameter
must be done serially (but can be done from different threads, as long as
only one thread ever is inside a call at any point in time, e.g. by using
a mutex per loop).

Specifically to support threads (and signal handlers), libev implements
so-called C<ev_async> watchers, which allow some limited form of
concurrency on the same event loop, namely waking it up "from the
outside".

If you want to know which design (one loop, locking, or multiple loops
without or something else still) is best for your problem, then I cannot
help you, but here is some generic advice:

=over 4

=item * most applications have a main thread: use the default libev loop
in that thread, or create a separate thread running only the default loop.

This helps integrating other libraries or software modules that use libev
themselves and don't care/know about threading.

=item * one loop per thread is usually a good model.

Doing this is almost never wrong, sometimes a better-performance model
exists, but it is always a good start.

=item * other models exist, such as the leader/follower pattern, where one
loop is handed through multiple threads in a kind of round-robin fashion.

Choosing a model is hard - look around, learn, know that usually you can do
better than you currently do :-)

=item * often you need to talk to some other thread which blocks in the
event loop.

C<ev_async> watchers can be used to wake them up from other threads safely
(or from signal contexts...).

An example use would be to communicate signals or other events that only
work in the default loop by registering the signal watcher with the
default loop and triggering an C<ev_async> watcher from the default loop
watcher callback into the event loop interested in the signal.

=back

See also L</THREAD LOCKING EXAMPLE>.

=head3 COROUTINES

Libev is very accommodating to coroutines ("cooperative threads"):
libev fully supports nesting calls to its functions from different
coroutines (e.g. you can call C<ev_run> on the same loop from two
different coroutines, and switch freely between both coroutines running
the loop, as long as you don't confuse yourself). The only exception is
that you must not do this from C<ev_periodic> reschedule callbacks.

Care has been taken to ensure that libev does not keep local state inside
C<ev_run>, and other calls do not usually allow for coroutine switches as
they do not call any callbacks.

=head2 COMPILER WARNINGS

Depending on your compiler and compiler settings, you might get no or a
lot of warnings when compiling libev code. Some people are apparently
scared by this.

However, these are unavoidable for many reasons. For one, each compiler
has different warnings, and each user has different tastes regarding
warning options. "Warn-free" code therefore cannot be a goal except when
targeting a specific compiler and compiler-version.

Another reason is that some compiler warnings require elaborate
workarounds, or other changes to the code that make it less clear and less
maintainable.

And of course, some compiler warnings are just plain stupid, or simply
wrong (because they don't actually warn about the condition their message
seems to warn about). For example, certain older gcc versions had some
warnings that resulted in an extreme number of false positives. These have
been fixed, but some people still insist on making code warn-free with
such buggy versions.

While libev is written to generate as few warnings as possible,
"warn-free" code is not a goal, and it is recommended not to build libev
with any compiler warnings enabled unless you are prepared to cope with
them (e.g. by ignoring them). Remember that warnings are just that:
warnings, not errors, or proof of bugs.


=head2 VALGRIND

Valgrind has a special section here because it is a popular tool that is
highly useful. Unfortunately, valgrind reports are very hard to interpret.

If you think you found a bug (memory leak, uninitialised data access etc.)
in libev, then check twice: If valgrind reports something like:

   ==2274==    definitely lost: 0 bytes in 0 blocks.
   ==2274==      possibly lost: 0 bytes in 0 blocks.
   ==2274==    still reachable: 256 bytes in 1 blocks.

Then there is no memory leak, just as memory accounted to global variables
is not a memleak - the memory is still being referenced, and didn't leak.

Similarly, under some circumstances, valgrind might report kernel bugs
as if it were a bug in libev (e.g. in realloc or in the poll backend,
although an acceptable workaround has been found here), or it might be
confused.

Keep in mind that valgrind is a very good tool, but only a tool. Don't
make it into some kind of religion.

If you are unsure about something, feel free to contact the mailing list
with the full valgrind report and an explanation on why you think this
is a bug in libev (best check the archives, too :). However, don't be
annoyed when you get a brisk "this is no bug" answer and take the chance
of learning how to interpret valgrind properly.

If you need, for some reason, empty reports from valgrind for your project
I suggest using suppression lists.


=head1 PORTABILITY NOTES

=head2 GNU/LINUX 32 BIT LIMITATIONS

GNU/Linux is the only common platform that supports 64 bit file/large file
interfaces but I<disables> them by default.

That means that libev compiled in the default environment doesn't support
files larger than 2GiB or so, which mainly affects C<ev_stat> watchers.

Unfortunately, many programs try to work around this GNU/Linux issue
by enabling the large file API, which makes them incompatible with the
standard libev compiled for their system.

Likewise, libev cannot enable the large file API itself as this would
suddenly make it incompatible to the default compile time environment,
i.e. all programs not using special compile switches.

=head2 OS/X AND DARWIN BUGS

The whole thing is a bug if you ask me - basically any system interface
you touch is broken, whether it is locales, poll, kqueue or even the
OpenGL drivers.

=head3 C<kqueue> is buggy

The kqueue syscall is broken in all known versions - most versions support
only sockets, many support pipes.

Libev tries to work around this by not using C<kqueue> by default on this
rotten platform, but of course you can still ask for it when creating a
loop - embedding a socket-only kqueue loop into a select-based one is
probably going to work well.

=head3 C<poll> is buggy

Instead of fixing C<kqueue>, Apple replaced their (working) C<poll>
implementation by something calling C<kqueue> internally around the 10.5.6
release, so now C<kqueue> I<and> C<poll> are broken.

Libev tries to work around this by not using C<poll> by default on
this rotten platform, but of course you can still ask for it when creating
a loop.

=head3 C<select> is buggy

All that's left is C<select>, and of course Apple found a way to fuck this
one up as well: On OS/X, C<select> actively limits the number of file
descriptors you can pass in to 1024 - your program suddenly crashes when
you use more.

There is an undocumented "workaround" for this - defining
C<_DARWIN_UNLIMITED_SELECT>, which libev tries to use, so select I<should>
work on OS/X.

=head2 SOLARIS PROBLEMS AND WORKAROUNDS

=head3 C<errno> reentrancy

The default compile environment on Solaris is unfortunately so
thread-unsafe that you can't even use components/libraries compiled
without C<-D_REENTRANT> in a threaded program, which, of course, isn't
defined by default. A valid, if stupid, implementation choice.

If you want to use libev in threaded environments you have to make sure
it's compiled with C<_REENTRANT> defined.

=head3 Event port backend

The scalable event interface for Solaris is called "event
ports". Unfortunately, this mechanism is very buggy in all major
releases. If you run into high CPU usage, your program freezes or you get
a large number of spurious wakeups, make sure you have all the relevant
and latest kernel patches applied. No, I don't know which ones, but there
are multiple ones to apply, and afterwards, event ports actually work
great.

If you can't get it to work, you can try running the program by setting
the environment variable C<LIBEV_FLAGS=3> to only allow C<poll> and
C<select> backends.

=head2 AIX POLL BUG

AIX unfortunately has a broken C<poll.h> header. Libev works around
this by trying to avoid the poll backend altogether (i.e. it's not even
compiled in), which normally isn't a big problem as C<select> works fine
with large bitsets on AIX, and AIX is dead anyway.

=head2 WIN32 PLATFORM LIMITATIONS AND WORKAROUNDS

=head3 General issues

Win32 doesn't support any of the standards (e.g. POSIX) that libev
requires, and its I/O model is fundamentally incompatible with the POSIX
model. Libev still offers limited functionality on this platform in
the form of the C<EVBACKEND_SELECT> backend, and only supports socket
descriptors. This only applies when using Win32 natively, not when using
e.g. cygwin. Actually, it only applies to the microsofts own compilers,
as every compiler comes with a slightly differently broken/incompatible
environment.

Lifting these limitations would basically require the full
re-implementation of the I/O system. If you are into this kind of thing,
then note that glib does exactly that for you in a very portable way (note
also that glib is the slowest event library known to man).

There is no supported compilation method available on windows except
embedding it into other applications.

Sensible signal handling is officially unsupported by Microsoft - libev
tries its best, but under most conditions, signals will simply not work.

Not a libev limitation but worth mentioning: windows apparently doesn't
accept large writes: instead of resulting in a partial write, windows will
either accept everything or return C<ENOBUFS> if the buffer is too large,
so make sure you only write small amounts into your sockets (less than a
megabyte seems safe, but this apparently depends on the amount of memory
available).

Due to the many, low, and arbitrary limits on the win32 platform and
the abysmal performance of winsockets, using a large number of sockets
is not recommended (and not reasonable). If your program needs to use
more than a hundred or so sockets, then likely it needs to use a totally
different implementation for windows, as libev offers the POSIX readiness
notification model, which cannot be implemented efficiently on windows
(due to Microsoft monopoly games).

A typical way to use libev under windows is to embed it (see the embedding
section for details) and use the following F<evwrap.h> header file instead
of F<ev.h>:

   #define EV_STANDALONE              /* keeps ev from requiring config.h */
   #define EV_SELECT_IS_WINSOCKET 1   /* configure libev for windows select */

   #include "ev.h"

And compile the following F<evwrap.c> file into your project (make sure
you do I<not> compile the F<ev.c> or any other embedded source files!):

   #include "evwrap.h"
   #include "ev.c"

=head3 The winsocket C<select> function

The winsocket C<select> function doesn't follow POSIX in that it
requires socket I<handles> and not socket I<file descriptors> (it is
also extremely buggy). This makes select very inefficient, and also
requires a mapping from file descriptors to socket handles (the Microsoft
C runtime provides the function C<_open_osfhandle> for this). See the
discussion of the C<EV_SELECT_USE_FD_SET>, C<EV_SELECT_IS_WINSOCKET> and
C<EV_FD_TO_WIN32_HANDLE> preprocessor symbols for more info.

The configuration for a "naked" win32 using the Microsoft runtime
libraries and raw winsocket select is:

   #define EV_USE_SELECT 1
   #define EV_SELECT_IS_WINSOCKET 1   /* forces EV_SELECT_USE_FD_SET, too */

Note that winsockets handling of fd sets is O(n), so you can easily get a
complexity in the O(n²) range when using win32.

=head3 Limited number of file descriptors

Windows has numerous arbitrary (and low) limits on things.

Early versions of winsocket's select only supported waiting for a maximum
of C<64> handles (probably owning to the fact that all windows kernels
can only wait for C<64> things at the same time internally; Microsoft
recommends spawning a chain of threads and wait for 63 handles and the
previous thread in each. Sounds great!).

Newer versions support more handles, but you need to define C<FD_SETSIZE>
to some high number (e.g. C<2048>) before compiling the winsocket select
call (which might be in libev or elsewhere, for example, perl and many
other interpreters do their own select emulation on windows).

Another limit is the number of file descriptors in the Microsoft runtime
libraries, which by default is C<64> (there must be a hidden I<64>
fetish or something like this inside Microsoft). You can increase this
by calling C<_setmaxstdio>, which can increase this limit to C<2048>
(another arbitrary limit), but is broken in many versions of the Microsoft
runtime libraries. This might get you to about C<512> or C<2048> sockets
(depending on windows version and/or the phase of the moon). To get more,
you need to wrap all I/O functions and provide your own fd management, but
the cost of calling select (O(n²)) will likely make this unworkable.

=head2 PORTABILITY REQUIREMENTS

In addition to a working ISO-C implementation and of course the
backend-specific APIs, libev relies on a few additional extensions:

=over 4

=item C<void (*)(ev_watcher_type *, int revents)> must have compatible
calling conventions regardless of C<ev_watcher_type *>.

Libev assumes not only that all watcher pointers have the same internal
structure (guaranteed by POSIX but not by ISO C for example), but it also
assumes that the same (machine) code can be used to call any watcher
callback: The watcher callbacks have different type signatures, but libev
calls them using an C<ev_watcher *> internally.

=item null pointers and integer zero are represented by 0 bytes

Libev uses C<memset> to initialise structs and arrays to C<0> bytes, and
relies on this setting pointers and integers to null.

=item pointer accesses must be thread-atomic

Accessing a pointer value must be atomic, it must both be readable and
writable in one piece - this is the case on all current architectures.

=item C<sig_atomic_t volatile> must be thread-atomic as well

The type C<sig_atomic_t volatile> (or whatever is defined as
C<EV_ATOMIC_T>) must be atomic with respect to accesses from different
threads. This is not part of the specification for C<sig_atomic_t>, but is
believed to be sufficiently portable.

=item C<sigprocmask> must work in a threaded environment

Libev uses C<sigprocmask> to temporarily block signals. This is not
allowed in a threaded program (C<pthread_sigmask> has to be used). Typical
pthread implementations will either allow C<sigprocmask> in the "main
thread" or will block signals process-wide, both behaviours would
be compatible with libev. Interaction between C<sigprocmask> and
C<pthread_sigmask> could complicate things, however.

The most portable way to handle signals is to block signals in all threads
except the initial one, and run the signal handling loop in the initial
thread as well.

=item C<long> must be large enough for common memory allocation sizes

To improve portability and simplify its API, libev uses C<long> internally
instead of C<size_t> when allocating its data structures. On non-POSIX
systems (Microsoft...) this might be unexpectedly low, but is still at
least 31 bits everywhere, which is enough for hundreds of millions of
watchers.

=item C<double> must hold a time value in seconds with enough accuracy

The type C<double> is used to represent timestamps. It is required to
have at least 51 bits of mantissa (and 9 bits of exponent), which is
good enough for at least into the year 4000 with millisecond accuracy
(the design goal for libev). This requirement is overfulfilled by
implementations using IEEE 754, which is basically all existing ones.

With IEEE 754 doubles, you get microsecond accuracy until at least the
year 2255 (and millisecond accuracy till the year 287396 - by then, libev
is either obsolete or somebody patched it to use C<long double> or
something like that, just kidding).

=back

If you know of other additional requirements drop me a note.


=head1 ALGORITHMIC COMPLEXITIES

In this section the complexities of (many of) the algorithms used inside
libev will be documented. For complexity discussions about backends see
the documentation for C<ev_default_init>.

All of the following are about amortised time: If an array needs to be
extended, libev needs to realloc and move the whole array, but this
happens asymptotically rarer with higher number of elements, so O(1) might
mean that libev does a lengthy realloc operation in rare cases, but on
average it is much faster and asymptotically approaches constant time.

=over 4

=item Starting and stopping timer/periodic watchers: O(log skipped_other_timers)

This means that, when you have a watcher that triggers in one hour and
there are 100 watchers that would trigger before that, then inserting will
have to skip roughly seven (C<ld 100>) of these watchers.

=item Changing timer/periodic watchers (by autorepeat or calling again): O(log skipped_other_timers)

That means that changing a timer costs less than removing/adding them,
as only the relative motion in the event queue has to be paid for.

=item Starting io/check/prepare/idle/signal/child/fork/async watchers: O(1)

These just add the watcher into an array or at the head of a list.

=item Stopping check/prepare/idle/fork/async watchers: O(1)

=item Stopping an io/signal/child watcher: O(number_of_watchers_for_this_(fd/signal/pid % EV_PID_HASHSIZE))

These watchers are stored in lists, so they need to be walked to find the
correct watcher to remove. The lists are usually short (you don't usually
have many watchers waiting for the same fd or signal: one is typical, two
is rare).

=item Finding the next timer in each loop iteration: O(1)

By virtue of using a binary or 4-heap, the next timer is always found at a
fixed position in the storage array.

=item Each change on a file descriptor per loop iteration: O(number_of_watchers_for_this_fd)

A change means an I/O watcher gets started or stopped, which requires
libev to recalculate its status (and possibly tell the kernel, depending
on backend and whether C<ev_io_set> was used).

=item Activating one watcher (putting it into the pending state): O(1)

=item Priority handling: O(number_of_priorities)

Priorities are implemented by allocating some space for each
priority. When doing priority-based operations, libev usually has to
linearly search all the priorities, but starting/stopping and activating
watchers becomes O(1) with respect to priority handling.

=item Sending an ev_async: O(1)

=item Processing ev_async_send: O(number_of_async_watchers)

=item Processing signals: O(max_signal_number)

Sending involves a system call I<iff> there were no other C<ev_async_send>
calls in the current loop iteration and the loop is currently
blocked. Checking for async and signal events involves iterating over all
running async watchers or all signal numbers.

=back


=head1 PORTING FROM LIBEV 3.X TO 4.X

The major version 4 introduced some incompatible changes to the API.

At the moment, the C<ev.h> header file provides compatibility definitions
for all changes, so most programs should still compile. The compatibility
layer might be removed in later versions of libev, so better update to the
new API early than late.

=over 4

=item C<EV_COMPAT3> backwards compatibility mechanism

The backward compatibility mechanism can be controlled by
C<EV_COMPAT3>. See L</"PREPROCESSOR SYMBOLS/MACROS"> in the L</EMBEDDING>
section.

=item C<ev_default_destroy> and C<ev_default_fork> have been removed

These calls can be replaced easily by their C<ev_loop_xxx> counterparts:

   ev_loop_destroy (EV_DEFAULT_UC);
   ev_loop_fork (EV_DEFAULT);

=item function/symbol renames

A number of functions and symbols have been renamed:

  ev_loop         => ev_run
  EVLOOP_NONBLOCK => EVRUN_NOWAIT
  EVLOOP_ONESHOT  => EVRUN_ONCE

  ev_unloop       => ev_break
  EVUNLOOP_CANCEL => EVBREAK_CANCEL
  EVUNLOOP_ONE    => EVBREAK_ONE
  EVUNLOOP_ALL    => EVBREAK_ALL

  EV_TIMEOUT      => EV_TIMER

  ev_loop_count   => ev_iteration
  ev_loop_depth   => ev_depth
  ev_loop_verify  => ev_verify

Most functions working on C<struct ev_loop> objects don't have an
C<ev_loop_> prefix, so it was removed; C<ev_loop>, C<ev_unloop> and
associated constants have been renamed to not collide with the C<struct
ev_loop> anymore and C<EV_TIMER> now follows the same naming scheme
as all other watcher types. Note that C<ev_loop_fork> is still called
C<ev_loop_fork> because it would otherwise clash with the C<ev_fork>
typedef.

=item C<EV_MINIMAL> mechanism replaced by C<EV_FEATURES>

The preprocessor symbol C<EV_MINIMAL> has been replaced by a different
mechanism, C<EV_FEATURES>. Programs using C<EV_MINIMAL> usually compile
and work, but the library code will of course be larger.

=back


=head1 GLOSSARY

=over 4

=item active

A watcher is active as long as it has been started and not yet stopped.
See L</WATCHER STATES> for details.

=item application

In this document, an application is whatever is using libev.

=item backend

The part of the code dealing with the operating system interfaces.

=item callback

The address of a function that is called when some event has been
detected. Callbacks are being passed the event loop, the watcher that
received the event, and the actual event bitset.

=item callback/watcher invocation

The act of calling the callback associated with a watcher.

=item event

A change of state of some external event, such as data now being available
for reading on a file descriptor, time having passed or simply not having
any other events happening anymore.

In libev, events are represented as single bits (such as C<EV_READ> or
C<EV_TIMER>).

=item event library

A software package implementing an event model and loop.

=item event loop

An entity that handles and processes external events and converts them
into callback invocations.

=item event model

The model used to describe how an event loop handles and processes
watchers and events.

=item pending

A watcher is pending as soon as the corresponding event has been
detected. See L</WATCHER STATES> for details.

=item real time

The physical time that is observed. It is apparently strictly monotonic :)

=item wall-clock time

The time and date as shown on clocks. Unlike real time, it can actually
be wrong and jump forwards and backwards, e.g. when you adjust your
clock.

=item watcher

A data structure that describes interest in certain events. Watchers need
to be started (attached to an event loop) before they can receive events.

=back

=head1 AUTHOR

Marc Lehmann <libev@schmorp.de>, with repeated corrections by Mikael
Magnusson and Emanuele Giaquinta, and minor corrections by many others.

Revision:	1.459
Committed:	Wed Jan 22 01:50:42 2020 UTC (4 years, 3 months ago) by root
Branch:	MAIN
Changes since 1.458:	+24 -8 lines
Log Message:	* empty log message *