luna - libev/ev.pod

   1=encoding utf-8
   2
   3=head1 NAME
   4
   5libev - a high performance full-featured event loop written in C
   6
   7=head1 SYNOPSIS
   8
   9   #include <ev.h>
  10
  11=head2 EXAMPLE PROGRAM
  12
  13   // a single header file is required
  14   #include <ev.h>
  15
  16   #include <stdio.h> // for puts
  17
  18   // every watcher type has its own typedef'd struct
  19   // with the name ev_TYPE
  20   ev_io stdin_watcher;
  21   ev_timer timeout_watcher;
  22
  23   // all watcher callbacks have a similar signature
  24   // this callback is called when data is readable on stdin
  25   static void
  26   stdin_cb (EV_P_ ev_io *w, int revents)
  27   {
  28     puts ("stdin ready");
  29     // for one-shot events, one must manually stop the watcher
  30     // with its corresponding stop function.
  31     ev_io_stop (EV_A_ w);
  32
  33     // this causes all nested ev_run's to stop iterating
  34     ev_break (EV_A_ EVBREAK_ALL);
  35   }
  36
  37   // another callback, this time for a time-out
  38   static void
  39   timeout_cb (EV_P_ ev_timer *w, int revents)
  40   {
  41     puts ("timeout");
  42     // this causes the innermost ev_run to stop iterating
  43     ev_break (EV_A_ EVBREAK_ONE);
  44   }
  45
  46   int
  47   main (void)
  48   {
  49     // use the default event loop unless you have special needs
  50     struct ev_loop *loop = EV_DEFAULT;
  51
  52     // initialise an io watcher, then start it
  53     // this one will watch for stdin to become readable
  54     ev_io_init (&stdin_watcher, stdin_cb, /*STDIN_FILENO*/ 0, EV_READ);
  55     ev_io_start (loop, &stdin_watcher);
  56
  57     // initialise a timer watcher, then start it
  58     // simple non-repeating 5.5 second timeout
  59     ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
  60     ev_timer_start (loop, &timeout_watcher);
  61
  62     // now wait for events to arrive
  63     ev_run (loop, 0);
  64
  65     // break was called, so exit
  66     return 0;
  67   }
  68
  69=head1 ABOUT THIS DOCUMENT
  70
  71This document documents the libev software package.
  72
  73The newest version of this document is also available as an html-formatted
  74web page you might find easier to navigate when reading it for the first
  75time: L<http://pod.tst.eu/http://cvs.schmorp.de/libev/ev.pod>.
  76
  77While this document tries to be as complete as possible in documenting
  78libev, its usage and the rationale behind its design, it is not a tutorial
  79on event-based programming, nor will it introduce event-based programming
  80with libev.
  81
  82Familiarity with event based programming techniques in general is assumed
  83throughout this document.
  84
  85=head1 WHAT TO READ WHEN IN A HURRY
  86
  87This manual tries to be very detailed, but unfortunately, this also makes
  88it very long. If you just want to know the basics of libev, I suggest
  89reading L</ANATOMY OF A WATCHER>, then the L</EXAMPLE PROGRAM> above and
  90look up the missing functions in L</GLOBAL FUNCTIONS> and the C<ev_io> and
  91C<ev_timer> sections in L</WATCHER TYPES>.
  92
  93=head1 ABOUT LIBEV
  94
  95Libev is an event loop: you register interest in certain events (such as a
  96file descriptor being readable or a timeout occurring), and it will manage
  97these event sources and provide your program with events.
  98
  99To do this, it must take more or less complete control over your process
 100(or thread) by executing the I<event loop> handler, and will then
 101communicate events via a callback mechanism.
 102
 103You register interest in certain events by registering so-called I<event
 104watchers>, which are relatively small C structures you initialise with the
 105details of the event, and then hand it over to libev by I<starting> the
 106watcher.
 107
 108=head2 FEATURES
 109
 110Libev supports C<select>, C<poll>, the Linux-specific C<epoll>, the
 111BSD-specific C<kqueue> and the Solaris-specific event port mechanisms
 112for file descriptor events (C<ev_io>), the Linux C<inotify> interface
 113(for C<ev_stat>), Linux eventfd/signalfd (for faster and cleaner
 114inter-thread wakeup (C<ev_async>)/signal handling (C<ev_signal>)) relative
 115timers (C<ev_timer>), absolute timers with customised rescheduling
 116(C<ev_periodic>), synchronous signals (C<ev_signal>), process status
 117change events (C<ev_child>), and event watchers dealing with the event
 118loop mechanism itself (C<ev_idle>, C<ev_embed>, C<ev_prepare> and
 119C<ev_check> watchers) as well as file watchers (C<ev_stat>) and even
 120limited support for fork events (C<ev_fork>).
 121
 122It also is quite fast (see this
 123L<benchmark|http://libev.schmorp.de/bench.html> comparing it to libevent
 124for example).
 125
 126=head2 CONVENTIONS
 127
 128Libev is very configurable. In this manual the default (and most common)
 129configuration will be described, which supports multiple event loops. For
 130more info about various configuration options please have a look at
 131B<EMBED> section in this manual. If libev was configured without support
 132for multiple event loops, then all functions taking an initial argument of
 133name C<loop> (which is always of type C<struct ev_loop *>) will not have
 134this argument.
 135
 136=head2 TIME REPRESENTATION
 137
 138Libev represents time as a single floating point number, representing
 139the (fractional) number of seconds since the (POSIX) epoch (in practice
 140somewhere near the beginning of 1970, details are complicated, don't
 141ask). This type is called C<ev_tstamp>, which is what you should use
 142too. It usually aliases to the C<double> type in C. When you need to do
 143any calculations on it, you should treat it as some floating point value.
 144
 145Unlike the name component C<stamp> might indicate, it is also used for
 146time differences (e.g. delays) throughout libev.
 147
 148=head1 ERROR HANDLING
 149
 150Libev knows three classes of errors: operating system errors, usage errors
 151and internal errors (bugs).
 152
 153When libev catches an operating system error it cannot handle (for example
 154a system call indicating a condition libev cannot fix), it calls the callback
 155set via C<ev_set_syserr_cb>, which is supposed to fix the problem or
 156abort. The default is to print a diagnostic message and to call C<abort
 157()>.
 158
 159When libev detects a usage error such as a negative timer interval, then
 160it will print a diagnostic message and abort (via the C<assert> mechanism,
 161so C<NDEBUG> will disable this checking): these are programming errors in
 162the libev caller and need to be fixed there.
 163
 164Libev also has a few internal error-checking C<assert>ions, and also has
 165extensive consistency checking code. These do not trigger under normal
 166circumstances, as they indicate either a bug in libev or worse.
 167
 168
 169=head1 GLOBAL FUNCTIONS
 170
 171These functions can be called anytime, even before initialising the
 172library in any way.
 173
 174=over 4
 175
 176=item ev_tstamp ev_time ()
 177
 178Returns the current time as libev would use it. Please note that the
 179C<ev_now> function is usually faster and also often returns the timestamp
 180you actually want to know. Also interesting is the combination of
 181C<ev_now_update> and C<ev_now>.
 182
 183=item ev_sleep (ev_tstamp interval)
 184
 185Sleep for the given interval: The current thread will be blocked
 186until either it is interrupted or the given time interval has
 187passed (approximately - it might return a bit earlier even if not
 188interrupted). Returns immediately if C<< interval <= 0 >>.
 189
 190Basically this is a sub-second-resolution C<sleep ()>.
 191
 192The range of the C<interval> is limited - libev only guarantees to work
 193with sleep times of up to one day (C<< interval <= 86400 >>).
 194
 195=item int ev_version_major ()
 196
 197=item int ev_version_minor ()
 198
 199You can find out the major and minor ABI version numbers of the library
 200you linked against by calling the functions C<ev_version_major> and
 201C<ev_version_minor>. If you want, you can compare against the global
 202symbols C<EV_VERSION_MAJOR> and C<EV_VERSION_MINOR>, which specify the
 203version of the library your program was compiled against.
 204
 205These version numbers refer to the ABI version of the library, not the
 206release version.
 207
 208Usually, it's a good idea to terminate if the major versions mismatch,
 209as this indicates an incompatible change. Minor versions are usually
 210compatible to older versions, so a larger minor version alone is usually
 211not a problem.
 212
 213Example: Make sure we haven't accidentally been linked against the wrong
 214version (note, however, that this will not detect other ABI mismatches,
 215such as LFS or reentrancy).
 216
 217   assert (("libev version mismatch",
 218            ev_version_major () == EV_VERSION_MAJOR
 219            && ev_version_minor () >= EV_VERSION_MINOR));
 220
 221=item unsigned int ev_supported_backends ()
 222
 223Return the set of all backends (i.e. their corresponding C<EV_BACKEND_*>
 224value) compiled into this binary of libev (independent of their
 225availability on the system you are running on). See C<ev_default_loop> for
 226a description of the set values.
 227
 228Example: make sure we have the epoll method, because yeah this is cool and
 229a must have and can we have a torrent of it please!!!11
 230
 231   assert (("sorry, no epoll, no sex",
 232            ev_supported_backends () & EVBACKEND_EPOLL));
 233
 234=item unsigned int ev_recommended_backends ()
 235
 236Return the set of all backends compiled into this binary of libev and
 237also recommended for this platform, meaning it will work for most file
 238descriptor types. This set is often smaller than the one returned by
 239C<ev_supported_backends>, as for example kqueue is broken on most BSDs
 240and will not be auto-detected unless you explicitly request it (assuming
 241you know what you are doing). This is the set of backends that libev will
 242probe for if you specify no backends explicitly.
 243
 244=item unsigned int ev_embeddable_backends ()
 245
 246Returns the set of backends that are embeddable in other event loops. This
 247value is platform-specific but can include backends not available on the
 248current system. To find which embeddable backends might be supported on
 249the current system, you would need to look at C<ev_embeddable_backends ()
 250& ev_supported_backends ()>, likewise for recommended ones.
 251
 252See the description of C<ev_embed> watchers for more info.
 253
 254=item ev_set_allocator (void *(*cb)(void *ptr, long size) throw ())
 255
 256Sets the allocation function to use (the prototype is similar - the
 257semantics are identical to the C<realloc> C89/SuS/POSIX function). It is
 258used to allocate and free memory (no surprises here). If it returns zero
 259when memory needs to be allocated (C<size != 0>), the library might abort
 260or take some potentially destructive action.
 261
 262Since some systems (at least OpenBSD and Darwin) fail to implement
 263correct C<realloc> semantics, libev will use a wrapper around the system
 264C<realloc> and C<free> functions by default.
 265
 266You could override this function in high-availability programs to, say,
 267free some memory if it cannot allocate memory, to use a special allocator,
 268or even to sleep a while and retry until some memory is available.
 269
 270Example: Replace the libev allocator with one that waits a bit and then
 271retries (example requires a standards-compliant C<realloc>).
 272
 273   static void *
 274   persistent_realloc (void *ptr, size_t size)
 275   {
 276     for (;;)
 277       {
 278         void *newptr = realloc (ptr, size);
 279
 280         if (newptr)
 281           return newptr;
 282
 283         sleep (60);
 284       }
 285   }
 286
 287   ...
 288   ev_set_allocator (persistent_realloc);
 289
 290=item ev_set_syserr_cb (void (*cb)(const char *msg) throw ())
 291
 292Set the callback function to call on a retryable system call error (such
 293as failed select, poll, epoll_wait). The message is a printable string
 294indicating the system call or subsystem causing the problem. If this
 295callback is set, then libev will expect it to remedy the situation, no
 296matter what, when it returns. That is, libev will generally retry the
 297requested operation, or, if the condition doesn't go away, do bad stuff
 298(such as abort).
 299
 300Example: This is basically the same thing that libev does internally, too.
 301
 302   static void
 303   fatal_error (const char *msg)
 304   {
 305     perror (msg);
 306     abort ();
 307   }
 308
 309   ...
 310   ev_set_syserr_cb (fatal_error);
 311
 312=item ev_feed_signal (int signum)
 313
 314This function can be used to "simulate" a signal receive. It is completely
 315safe to call this function at any time, from any context, including signal
 316handlers or random threads.
 317
 318Its main use is to customise signal handling in your process, especially
 319in the presence of threads. For example, you could block signals
 320by default in all threads (and specifying C<EVFLAG_NOSIGMASK> when
 321creating any loops), and in one thread, use C<sigwait> or any other
 322mechanism to wait for signals, then "deliver" them to libev by calling
 323C<ev_feed_signal>.
 324
 325=back
 326
 327=head1 FUNCTIONS CONTROLLING EVENT LOOPS
 328
 329An event loop is described by a C<struct ev_loop *> (the C<struct> is
 330I<not> optional in this case unless libev 3 compatibility is disabled, as
 331libev 3 had an C<ev_loop> function colliding with the struct name).
 332
 333The library knows two types of such loops, the I<default> loop, which
 334supports child process events, and dynamically created event loops which
 335do not.
 336
 337=over 4
 338
 339=item struct ev_loop *ev_default_loop (unsigned int flags)
 340
 341This returns the "default" event loop object, which is what you should
 342normally use when you just need "the event loop". Event loop objects and
 343the C<flags> parameter are described in more detail in the entry for
 344C<ev_loop_new>.
 345
 346If the default loop is already initialised then this function simply
 347returns it (and ignores the flags. If that is troubling you, check
 348C<ev_backend ()> afterwards). Otherwise it will create it with the given
 349flags, which should almost always be C<0>, unless the caller is also the
 350one calling C<ev_run> or otherwise qualifies as "the main program".
 351
 352If you don't know what event loop to use, use the one returned from this
 353function (or via the C<EV_DEFAULT> macro).
 354
 355Note that this function is I<not> thread-safe, so if you want to use it
 356from multiple threads, you have to employ some kind of mutex (note also
 357that this case is unlikely, as loops cannot be shared easily between
 358threads anyway).
 359
 360The default loop is the only loop that can handle C<ev_child> watchers,
 361and to do this, it always registers a handler for C<SIGCHLD>. If this is
 362a problem for your application you can either create a dynamic loop with
 363C<ev_loop_new> which doesn't do that, or you can simply overwrite the
 364C<SIGCHLD> signal handler I<after> calling C<ev_default_init>.
 365
 366Example: This is the most typical usage.
 367
 368   if (!ev_default_loop (0))
 369     fatal ("could not initialise libev, bad $LIBEV_FLAGS in environment?");
 370
 371Example: Restrict libev to the select and poll backends, and do not allow
 372environment settings to be taken into account:
 373
 374   ev_default_loop (EVBACKEND_POLL | EVBACKEND_SELECT | EVFLAG_NOENV);
 375
 376=item struct ev_loop *ev_loop_new (unsigned int flags)
 377
 378This will create and initialise a new event loop object. If the loop
 379could not be initialised, returns false.
 380
 381This function is thread-safe, and one common way to use libev with
 382threads is indeed to create one loop per thread, and using the default
 383loop in the "main" or "initial" thread.
 384
 385The flags argument can be used to specify special behaviour or specific
 386backends to use, and is usually specified as C<0> (or C<EVFLAG_AUTO>).
 387
 388The following flags are supported:
 389
 390=over 4
 391
 392=item C<EVFLAG_AUTO>
 393
 394The default flags value. Use this if you have no clue (it's the right
 395thing, believe me).
 396
 397=item C<EVFLAG_NOENV>
 398
 399If this flag bit is or'ed into the flag value (or the program runs setuid
 400or setgid) then libev will I<not> look at the environment variable
 401C<LIBEV_FLAGS>. Otherwise (the default), this environment variable will
 402override the flags completely if it is found in the environment. This is
 403useful to try out specific backends to test their performance, to work
 404around bugs, or to make libev threadsafe (accessing environment variables
 405cannot be done in a threadsafe way, but usually it works if no other
 406thread modifies them).
 407
 408=item C<EVFLAG_FORKCHECK>
 409
 410Instead of calling C<ev_loop_fork> manually after a fork, you can also
 411make libev check for a fork in each iteration by enabling this flag.
 412
 413This works by calling C<getpid ()> on every iteration of the loop,
 414and thus this might slow down your event loop if you do a lot of loop
 415iterations and little real work, but is usually not noticeable (on my
 416GNU/Linux system for example, C<getpid> is actually a simple 5-insn sequence
 417without a system call and thus I<very> fast, but my GNU/Linux system also has
 418C<pthread_atfork> which is even faster).
 419
 420The big advantage of this flag is that you can forget about fork (and
 421forget about forgetting to tell libev about forking, although you still
 422have to ignore C<SIGPIPE>) when you use this flag.
 423
 424This flag setting cannot be overridden or specified in the C<LIBEV_FLAGS>
 425environment variable.
 426
 427=item C<EVFLAG_NOINOTIFY>
 428
 429When this flag is specified, then libev will not attempt to use the
 430I<inotify> API for its C<ev_stat> watchers. Apart from debugging and
 431testing, this flag can be useful to conserve inotify file descriptors, as
 432otherwise each loop using C<ev_stat> watchers consumes one inotify handle.
 433
 434=item C<EVFLAG_SIGNALFD>
 435
 436When this flag is specified, then libev will attempt to use the
 437I<signalfd> API for its C<ev_signal> (and C<ev_child>) watchers. This API
 438delivers signals synchronously, which makes it both faster and might make
 439it possible to get the queued signal data. It can also simplify signal
 440handling with threads, as long as you properly block signals in your
 441threads that are not interested in handling them.
 442
 443Signalfd will not be used by default as this changes your signal mask, and
 444there are a lot of shoddy libraries and programs (glib's threadpool for
 445example) that can't properly initialise their signal masks.
 446
 447=item C<EVFLAG_NOSIGMASK>
 448
 449When this flag is specified, then libev will avoid to modify the signal
 450mask. Specifically, this means you have to make sure signals are unblocked
 451when you want to receive them.
 452
 453This behaviour is useful when you want to do your own signal handling, or
 454want to handle signals only in specific threads and want to avoid libev
 455unblocking the signals.
 456
 457It's also required by POSIX in a threaded program, as libev calls
 458C<sigprocmask>, whose behaviour is officially unspecified.
 459
 460This flag's behaviour will become the default in future versions of libev.
 461
 462=item C<EVBACKEND_SELECT>  (value 1, portable select backend)
 463
 464This is your standard select(2) backend. Not I<completely> standard, as
 465libev tries to roll its own fd_set with no limits on the number of fds,
 466but if that fails, expect a fairly low limit on the number of fds when
 467using this backend. It doesn't scale too well (O(highest_fd)), but its
 468usually the fastest backend for a low number of (low-numbered :) fds.
 469
 470To get good performance out of this backend you need a high amount of
 471parallelism (most of the file descriptors should be busy). If you are
 472writing a server, you should C<accept ()> in a loop to accept as many
 473connections as possible during one iteration. You might also want to have
 474a look at C<ev_set_io_collect_interval ()> to increase the amount of
 475readiness notifications you get per iteration.
 476
 477This backend maps C<EV_READ> to the C<readfds> set and C<EV_WRITE> to the
 478C<writefds> set (and to work around Microsoft Windows bugs, also onto the
 479C<exceptfds> set on that platform).
 480
 481=item C<EVBACKEND_POLL>    (value 2, poll backend, available everywhere except on windows)
 482
 483And this is your standard poll(2) backend. It's more complicated
 484than select, but handles sparse fds better and has no artificial
 485limit on the number of fds you can use (except it will slow down
 486considerably with a lot of inactive fds). It scales similarly to select,
 487i.e. O(total_fds). See the entry for C<EVBACKEND_SELECT>, above, for
 488performance tips.
 489
 490This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and
 491C<EV_WRITE> to C<POLLOUT | POLLERR | POLLHUP>.
 492
 493=item C<EVBACKEND_EPOLL>   (value 4, Linux)
 494
 495Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9
 496kernels).
 497
 498For few fds, this backend is a bit little slower than poll and select, but
 499it scales phenomenally better. While poll and select usually scale like
 500O(total_fds) where total_fds is the total number of fds (or the highest
 501fd), epoll scales either O(1) or O(active_fds).
 502
 503The epoll mechanism deserves honorable mention as the most misdesigned
 504of the more advanced event mechanisms: mere annoyances include silently
 505dropping file descriptors, requiring a system call per change per file
 506descriptor (and unnecessary guessing of parameters), problems with dup,
 507returning before the timeout value, resulting in additional iterations
 508(and only giving 5ms accuracy while select on the same platform gives
 5090.1ms) and so on. The biggest issue is fork races, however - if a program
 510forks then I<both> parent and child process have to recreate the epoll
 511set, which can take considerable time (one syscall per file descriptor)
 512and is of course hard to detect.
 513
 514Epoll is also notoriously buggy - embedding epoll fds I<should> work,
 515but of course I<doesn't>, and epoll just loves to report events for
 516totally I<different> file descriptors (even already closed ones, so
 517one cannot even remove them from the set) than registered in the set
 518(especially on SMP systems). Libev tries to counter these spurious
 519notifications by employing an additional generation counter and comparing
 520that against the events to filter out spurious ones, recreating the set
 521when required. Epoll also erroneously rounds down timeouts, but gives you
 522no way to know when and by how much, so sometimes you have to busy-wait
 523because epoll returns immediately despite a nonzero timeout. And last
 524not least, it also refuses to work with some file descriptors which work
 525perfectly fine with C<select> (files, many character devices...).
 526
 527Epoll is truly the train wreck among event poll mechanisms, a frankenpoll,
 528cobbled together in a hurry, no thought to design or interaction with
 529others. Oh, the pain, will it ever stop...
 530
 531While stopping, setting and starting an I/O watcher in the same iteration
 532will result in some caching, there is still a system call per such
 533incident (because the same I<file descriptor> could point to a different
 534I<file description> now), so its best to avoid that. Also, C<dup ()>'ed
 535file descriptors might not work very well if you register events for both
 536file descriptors.
 537
 538Best performance from this backend is achieved by not unregistering all
 539watchers for a file descriptor until it has been closed, if possible,
 540i.e. keep at least one watcher active per fd at all times. Stopping and
 541starting a watcher (without re-setting it) also usually doesn't cause
 542extra overhead. A fork can both result in spurious notifications as well
 543as in libev having to destroy and recreate the epoll object, which can
 544take considerable time and thus should be avoided.
 545
 546All this means that, in practice, C<EVBACKEND_SELECT> can be as fast or
 547faster than epoll for maybe up to a hundred file descriptors, depending on
 548the usage. So sad.
 549
 550While nominally embeddable in other event loops, this feature is broken in
 551all kernel versions tested so far.
 552
 553This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
 554C<EVBACKEND_POLL>.
 555
 556=item C<EVBACKEND_KQUEUE>  (value 8, most BSD clones)
 557
 558Kqueue deserves special mention, as at the time of this writing, it
 559was broken on all BSDs except NetBSD (usually it doesn't work reliably
 560with anything but sockets and pipes, except on Darwin, where of course
 561it's completely useless). Unlike epoll, however, whose brokenness
 562is by design, these kqueue bugs can (and eventually will) be fixed
 563without API changes to existing programs. For this reason it's not being
 564"auto-detected" unless you explicitly specify it in the flags (i.e. using
 565C<EVBACKEND_KQUEUE>) or libev was compiled on a known-to-be-good (-enough)
 566system like NetBSD.
 567
 568You still can embed kqueue into a normal poll or select backend and use it
 569only for sockets (after having made sure that sockets work with kqueue on
 570the target platform). See C<ev_embed> watchers for more info.
 571
 572It scales in the same way as the epoll backend, but the interface to the
 573kernel is more efficient (which says nothing about its actual speed, of
 574course). While stopping, setting and starting an I/O watcher does never
 575cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to
 576two event changes per incident. Support for C<fork ()> is very bad (you
 577might have to leak fd's on fork, but it's more sane than epoll) and it
 578drops fds silently in similarly hard-to-detect cases.
 579
 580This backend usually performs well under most conditions.
 581
 582While nominally embeddable in other event loops, this doesn't work
 583everywhere, so you might need to test for this. And since it is broken
 584almost everywhere, you should only use it when you have a lot of sockets
 585(for which it usually works), by embedding it into another event loop
 586(e.g. C<EVBACKEND_SELECT> or C<EVBACKEND_POLL> (but C<poll> is of course
 587also broken on OS X)) and, did I mention it, using it only for sockets.
 588
 589This backend maps C<EV_READ> into an C<EVFILT_READ> kevent with
 590C<NOTE_EOF>, and C<EV_WRITE> into an C<EVFILT_WRITE> kevent with
 591C<NOTE_EOF>.
 592
 593=item C<EVBACKEND_DEVPOLL> (value 16, Solaris 8)
 594
 595This is not implemented yet (and might never be, unless you send me an
 596implementation). According to reports, C</dev/poll> only supports sockets
 597and is not embeddable, which would limit the usefulness of this backend
 598immensely.
 599
 600=item C<EVBACKEND_PORT>    (value 32, Solaris 10)
 601
 602This uses the Solaris 10 event port mechanism. As with everything on Solaris,
 603it's really slow, but it still scales very well (O(active_fds)).
 604
 605While this backend scales well, it requires one system call per active
 606file descriptor per loop iteration. For small and medium numbers of file
 607descriptors a "slow" C<EVBACKEND_SELECT> or C<EVBACKEND_POLL> backend
 608might perform better.
 609
 610On the positive side, this backend actually performed fully to
 611specification in all tests and is fully embeddable, which is a rare feat
 612among the OS-specific backends (I vastly prefer correctness over speed
 613hacks).
 614
 615On the negative side, the interface is I<bizarre> - so bizarre that
 616even sun itself gets it wrong in their code examples: The event polling
 617function sometimes returns events to the caller even though an error
 618occurred, but with no indication whether it has done so or not (yes, it's
 619even documented that way) - deadly for edge-triggered interfaces where you
 620absolutely have to know whether an event occurred or not because you have
 621to re-arm the watcher.
 622
 623Fortunately libev seems to be able to work around these idiocies.
 624
 625This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
 626C<EVBACKEND_POLL>.
 627
 628=item C<EVBACKEND_ALL>
 629
 630Try all backends (even potentially broken ones that wouldn't be tried
 631with C<EVFLAG_AUTO>). Since this is a mask, you can do stuff such as
 632C<EVBACKEND_ALL & ~EVBACKEND_KQUEUE>.
 633
 634It is definitely not recommended to use this flag, use whatever
 635C<ev_recommended_backends ()> returns, or simply do not specify a backend
 636at all.
 637
 638=item C<EVBACKEND_MASK>
 639
 640Not a backend at all, but a mask to select all backend bits from a
 641C<flags> value, in case you want to mask out any backends from a flags
 642value (e.g. when modifying the C<LIBEV_FLAGS> environment variable).
 643
 644=back
 645
 646If one or more of the backend flags are or'ed into the flags value,
 647then only these backends will be tried (in the reverse order as listed
 648here). If none are specified, all backends in C<ev_recommended_backends
 649()> will be tried.
 650
 651Example: Try to create a event loop that uses epoll and nothing else.
 652
 653   struct ev_loop *epoller = ev_loop_new (EVBACKEND_EPOLL | EVFLAG_NOENV);
 654   if (!epoller)
 655     fatal ("no epoll found here, maybe it hides under your chair");
 656
 657Example: Use whatever libev has to offer, but make sure that kqueue is
 658used if available.
 659
 660   struct ev_loop *loop = ev_loop_new (ev_recommended_backends () | EVBACKEND_KQUEUE);
 661
 662=item ev_loop_destroy (loop)
 663
 664Destroys an event loop object (frees all memory and kernel state
 665etc.). None of the active event watchers will be stopped in the normal
 666sense, so e.g. C<ev_is_active> might still return true. It is your
 667responsibility to either stop all watchers cleanly yourself I<before>
 668calling this function, or cope with the fact afterwards (which is usually
 669the easiest thing, you can just ignore the watchers and/or C<free ()> them
 670for example).
 671
 672Note that certain global state, such as signal state (and installed signal
 673handlers), will not be freed by this function, and related watchers (such
 674as signal and child watchers) would need to be stopped manually.
 675
 676This function is normally used on loop objects allocated by
 677C<ev_loop_new>, but it can also be used on the default loop returned by
 678C<ev_default_loop>, in which case it is not thread-safe.
 679
 680Note that it is not advisable to call this function on the default loop
 681except in the rare occasion where you really need to free its resources.
 682If you need dynamically allocated loops it is better to use C<ev_loop_new>
 683and C<ev_loop_destroy>.
 684
 685=item ev_loop_fork (loop)
 686
 687This function sets a flag that causes subsequent C<ev_run> iterations
 688to reinitialise the kernel state for backends that have one. Despite
 689the name, you can call it anytime you are allowed to start or stop
 690watchers (except inside an C<ev_prepare> callback), but it makes most
 691sense after forking, in the child process. You I<must> call it (or use
 692C<EVFLAG_FORKCHECK>) in the child before resuming or calling C<ev_run>.
 693
 694In addition, if you want to reuse a loop (via this function or
 695C<EVFLAG_FORKCHECK>), you I<also> have to ignore C<SIGPIPE>.
 696
 697Again, you I<have> to call it on I<any> loop that you want to re-use after
 698a fork, I<even if you do not plan to use the loop in the parent>. This is
 699because some kernel interfaces *cough* I<kqueue> *cough* do funny things
 700during fork.
 701
 702On the other hand, you only need to call this function in the child
 703process if and only if you want to use the event loop in the child. If
 704you just fork+exec or create a new loop in the child, you don't have to
 705call it at all (in fact, C<epoll> is so badly broken that it makes a
 706difference, but libev will usually detect this case on its own and do a
 707costly reset of the backend).
 708
 709The function itself is quite fast and it's usually not a problem to call
 710it just in case after a fork.
 711
 712Example: Automate calling C<ev_loop_fork> on the default loop when
 713using pthreads.
 714
 715   static void
 716   post_fork_child (void)
 717   {
 718     ev_loop_fork (EV_DEFAULT);
 719   }
 720
 721   ...
 722   pthread_atfork (0, 0, post_fork_child);
 723
 724=item int ev_is_default_loop (loop)
 725
 726Returns true when the given loop is, in fact, the default loop, and false
 727otherwise.
 728
 729=item unsigned int ev_iteration (loop)
 730
 731Returns the current iteration count for the event loop, which is identical
 732to the number of times libev did poll for new events. It starts at C<0>
 733and happily wraps around with enough iterations.
 734
 735This value can sometimes be useful as a generation counter of sorts (it
 736"ticks" the number of loop iterations), as it roughly corresponds with
 737C<ev_prepare> and C<ev_check> calls - and is incremented between the
 738prepare and check phases.
 739
 740=item unsigned int ev_depth (loop)
 741
 742Returns the number of times C<ev_run> was entered minus the number of
 743times C<ev_run> was exited normally, in other words, the recursion depth.
 744
 745Outside C<ev_run>, this number is zero. In a callback, this number is
 746C<1>, unless C<ev_run> was invoked recursively (or from another thread),
 747in which case it is higher.
 748
 749Leaving C<ev_run> abnormally (setjmp/longjmp, cancelling the thread,
 750throwing an exception etc.), doesn't count as "exit" - consider this
 751as a hint to avoid such ungentleman-like behaviour unless it's really
 752convenient, in which case it is fully supported.
 753
 754=item unsigned int ev_backend (loop)
 755
 756Returns one of the C<EVBACKEND_*> flags indicating the event backend in
 757use.
 758
 759=item ev_tstamp ev_now (loop)
 760
 761Returns the current "event loop time", which is the time the event loop
 762received events and started processing them. This timestamp does not
 763change as long as callbacks are being processed, and this is also the base
 764time used for relative timers. You can treat it as the timestamp of the
 765event occurring (or more correctly, libev finding out about it).
 766
 767=item ev_now_update (loop)
 768
 769Establishes the current time by querying the kernel, updating the time
 770returned by C<ev_now ()> in the progress. This is a costly operation and
 771is usually done automatically within C<ev_run ()>.
 772
 773This function is rarely useful, but when some event callback runs for a
 774very long time without entering the event loop, updating libev's idea of
 775the current time is a good idea.
 776
 777See also L</The special problem of time updates> in the C<ev_timer> section.
 778
 779=item ev_suspend (loop)
 780
 781=item ev_resume (loop)
 782
 783These two functions suspend and resume an event loop, for use when the
 784loop is not used for a while and timeouts should not be processed.
 785
 786A typical use case would be an interactive program such as a game:  When
 787the user presses C<^Z> to suspend the game and resumes it an hour later it
 788would be best to handle timeouts as if no time had actually passed while
 789the program was suspended. This can be achieved by calling C<ev_suspend>
 790in your C<SIGTSTP> handler, sending yourself a C<SIGSTOP> and calling
 791C<ev_resume> directly afterwards to resume timer processing.
 792
 793Effectively, all C<ev_timer> watchers will be delayed by the time spend
 794between C<ev_suspend> and C<ev_resume>, and all C<ev_periodic> watchers
 795will be rescheduled (that is, they will lose any events that would have
 796occurred while suspended).
 797
 798After calling C<ev_suspend> you B<must not> call I<any> function on the
 799given loop other than C<ev_resume>, and you B<must not> call C<ev_resume>
 800without a previous call to C<ev_suspend>.
 801
 802Calling C<ev_suspend>/C<ev_resume> has the side effect of updating the
 803event loop time (see C<ev_now_update>).
 804
 805=item bool ev_run (loop, int flags)
 806
 807Finally, this is it, the event handler. This function usually is called
 808after you have initialised all your watchers and you want to start
 809handling events. It will ask the operating system for any new events, call
 810the watcher callbacks, and then repeat the whole process indefinitely: This
 811is why event loops are called I<loops>.
 812
 813If the flags argument is specified as C<0>, it will keep handling events
 814until either no event watchers are active anymore or C<ev_break> was
 815called.
 816
 817The return value is false if there are no more active watchers (which
 818usually means "all jobs done" or "deadlock"), and true in all other cases
 819(which usually means " you should call C<ev_run> again").
 820
 821Please note that an explicit C<ev_break> is usually better than
 822relying on all watchers to be stopped when deciding when a program has
 823finished (especially in interactive programs), but having a program
 824that automatically loops as long as it has to and no longer by virtue
 825of relying on its watchers stopping correctly, that is truly a thing of
 826beauty.
 827
 828This function is I<mostly> exception-safe - you can break out of a
 829C<ev_run> call by calling C<longjmp> in a callback, throwing a C++
 830exception and so on. This does not decrement the C<ev_depth> value, nor
 831will it clear any outstanding C<EVBREAK_ONE> breaks.
 832
 833A flags value of C<EVRUN_NOWAIT> will look for new events, will handle
 834those events and any already outstanding ones, but will not wait and
 835block your process in case there are no events and will return after one
 836iteration of the loop. This is sometimes useful to poll and handle new
 837events while doing lengthy calculations, to keep the program responsive.
 838
 839A flags value of C<EVRUN_ONCE> will look for new events (waiting if
 840necessary) and will handle those and any already outstanding ones. It
 841will block your process until at least one new event arrives (which could
 842be an event internal to libev itself, so there is no guarantee that a
 843user-registered callback will be called), and will return after one
 844iteration of the loop.
 845
 846This is useful if you are waiting for some external event in conjunction
 847with something not expressible using other libev watchers (i.e. "roll your
 848own C<ev_run>"). However, a pair of C<ev_prepare>/C<ev_check> watchers is
 849usually a better approach for this kind of thing.
 850
 851Here are the gory details of what C<ev_run> does (this is for your
 852understanding, not a guarantee that things will work exactly like this in
 853future versions):
 854
 855   - Increment loop depth.
 856   - Reset the ev_break status.
 857   - Before the first iteration, call any pending watchers.
 858   LOOP:
 859   - If EVFLAG_FORKCHECK was used, check for a fork.
 860   - If a fork was detected (by any means), queue and call all fork watchers.
 861   - Queue and call all prepare watchers.
 862   - If ev_break was called, goto FINISH.
 863   - If we have been forked, detach and recreate the kernel state
 864     as to not disturb the other process.
 865   - Update the kernel state with all outstanding changes.
 866   - Update the "event loop time" (ev_now ()).
 867   - Calculate for how long to sleep or block, if at all
 868     (active idle watchers, EVRUN_NOWAIT or not having
 869     any active watchers at all will result in not sleeping).
 870   - Sleep if the I/O and timer collect interval say so.
 871   - Increment loop iteration counter.
 872   - Block the process, waiting for any events.
 873   - Queue all outstanding I/O (fd) events.
 874   - Update the "event loop time" (ev_now ()), and do time jump adjustments.
 875   - Queue all expired timers.
 876   - Queue all expired periodics.
 877   - Queue all idle watchers with priority higher than that of pending events.
 878   - Queue all check watchers.
 879   - Call all queued watchers in reverse order (i.e. check watchers first).
 880     Signals and child watchers are implemented as I/O watchers, and will
 881     be handled here by queueing them when their watcher gets executed.
 882   - If ev_break has been called, or EVRUN_ONCE or EVRUN_NOWAIT
 883     were used, or there are no active watchers, goto FINISH, otherwise
 884     continue with step LOOP.
 885   FINISH:
 886   - Reset the ev_break status iff it was EVBREAK_ONE.
 887   - Decrement the loop depth.
 888   - Return.
 889
 890Example: Queue some jobs and then loop until no events are outstanding
 891anymore.
 892
 893   ... queue jobs here, make sure they register event watchers as long
 894   ... as they still have work to do (even an idle watcher will do..)
 895   ev_run (my_loop, 0);
 896   ... jobs done or somebody called break. yeah!
 897
 898=item ev_break (loop, how)
 899
 900Can be used to make a call to C<ev_run> return early (but only after it
 901has processed all outstanding events). The C<how> argument must be either
 902C<EVBREAK_ONE>, which will make the innermost C<ev_run> call return, or
 903C<EVBREAK_ALL>, which will make all nested C<ev_run> calls return.
 904
 905This "break state" will be cleared on the next call to C<ev_run>.
 906
 907It is safe to call C<ev_break> from outside any C<ev_run> calls, too, in
 908which case it will have no effect.
 909
 910=item ev_ref (loop)
 911
 912=item ev_unref (loop)
 913
 914Ref/unref can be used to add or remove a reference count on the event
 915loop: Every watcher keeps one reference, and as long as the reference
 916count is nonzero, C<ev_run> will not return on its own.
 917
 918This is useful when you have a watcher that you never intend to
 919unregister, but that nevertheless should not keep C<ev_run> from
 920returning. In such a case, call C<ev_unref> after starting, and C<ev_ref>
 921before stopping it.
 922
 923As an example, libev itself uses this for its internal signal pipe: It
 924is not visible to the libev user and should not keep C<ev_run> from
 925exiting if no event watchers registered by it are active. It is also an
 926excellent way to do this for generic recurring timers or from within
 927third-party libraries. Just remember to I<unref after start> and I<ref
 928before stop> (but only if the watcher wasn't active before, or was active
 929before, respectively. Note also that libev might stop watchers itself
 930(e.g. non-repeating timers) in which case you have to C<ev_ref>
 931in the callback).
 932
 933Example: Create a signal watcher, but keep it from keeping C<ev_run>
 934running when nothing else is active.
 935
 936   ev_signal exitsig;
 937   ev_signal_init (&exitsig, sig_cb, SIGINT);
 938   ev_signal_start (loop, &exitsig);
 939   ev_unref (loop);
 940
 941Example: For some weird reason, unregister the above signal handler again.
 942
 943   ev_ref (loop);
 944   ev_signal_stop (loop, &exitsig);
 945
 946=item ev_set_io_collect_interval (loop, ev_tstamp interval)
 947
 948=item ev_set_timeout_collect_interval (loop, ev_tstamp interval)
 949
 950These advanced functions influence the time that libev will spend waiting
 951for events. Both time intervals are by default C<0>, meaning that libev
 952will try to invoke timer/periodic callbacks and I/O callbacks with minimum
 953latency.
 954
 955Setting these to a higher value (the C<interval> I<must> be >= C<0>)
 956allows libev to delay invocation of I/O and timer/periodic callbacks
 957to increase efficiency of loop iterations (or to increase power-saving
 958opportunities).
 959
 960The idea is that sometimes your program runs just fast enough to handle
 961one (or very few) event(s) per loop iteration. While this makes the
 962program responsive, it also wastes a lot of CPU time to poll for new
 963events, especially with backends like C<select ()> which have a high
 964overhead for the actual polling but can deliver many events at once.
 965
 966By setting a higher I<io collect interval> you allow libev to spend more
 967time collecting I/O events, so you can handle more events per iteration,
 968at the cost of increasing latency. Timeouts (both C<ev_periodic> and
 969C<ev_timer>) will not be affected. Setting this to a non-null value will
 970introduce an additional C<ev_sleep ()> call into most loop iterations. The
 971sleep time ensures that libev will not poll for I/O events more often then
 972once per this interval, on average (as long as the host time resolution is
 973good enough).
 974
 975Likewise, by setting a higher I<timeout collect interval> you allow libev
 976to spend more time collecting timeouts, at the expense of increased
 977latency/jitter/inexactness (the watcher callback will be called
 978later). C<ev_io> watchers will not be affected. Setting this to a non-null
 979value will not introduce any overhead in libev.
 980
 981Many (busy) programs can usually benefit by setting the I/O collect
 982interval to a value near C<0.1> or so, which is often enough for
 983interactive servers (of course not for games), likewise for timeouts. It
 984usually doesn't make much sense to set it to a lower value than C<0.01>,
 985as this approaches the timing granularity of most systems. Note that if
 986you do transactions with the outside world and you can't increase the
 987parallelity, then this setting will limit your transaction rate (if you
 988need to poll once per transaction and the I/O collect interval is 0.01,
 989then you can't do more than 100 transactions per second).
 990
 991Setting the I<timeout collect interval> can improve the opportunity for
 992saving power, as the program will "bundle" timer callback invocations that
 993are "near" in time together, by delaying some, thus reducing the number of
 994times the process sleeps and wakes up again. Another useful technique to
 995reduce iterations/wake-ups is to use C<ev_periodic> watchers and make sure
 996they fire on, say, one-second boundaries only.
 997
 998Example: we only need 0.1s timeout granularity, and we wish not to poll
 999more often than 100 times per second:
1000
1001   ev_set_timeout_collect_interval (EV_DEFAULT_UC_ 0.1);
1002   ev_set_io_collect_interval (EV_DEFAULT_UC_ 0.01);
1003
1004=item ev_invoke_pending (loop)
1005
1006This call will simply invoke all pending watchers while resetting their
1007pending state. Normally, C<ev_run> does this automatically when required,
1008but when overriding the invoke callback this call comes handy. This
1009function can be invoked from a watcher - this can be useful for example
1010when you want to do some lengthy calculation and want to pass further
1011event handling to another thread (you still have to make sure only one
1012thread executes within C<ev_invoke_pending> or C<ev_run> of course).
1013
1014=item int ev_pending_count (loop)
1015
1016Returns the number of pending watchers - zero indicates that no watchers
1017are pending.
1018
1019=item ev_set_invoke_pending_cb (loop, void (*invoke_pending_cb)(EV_P))
1020
1021This overrides the invoke pending functionality of the loop: Instead of
1022invoking all pending watchers when there are any, C<ev_run> will call
1023this callback instead. This is useful, for example, when you want to
1024invoke the actual watchers inside another context (another thread etc.).
1025
1026If you want to reset the callback, use C<ev_invoke_pending> as new
1027callback.
1028
1029=item ev_set_loop_release_cb (loop, void (*release)(EV_P) throw (), void (*acquire)(EV_P) throw ())
1030
1031Sometimes you want to share the same loop between multiple threads. This
1032can be done relatively simply by putting mutex_lock/unlock calls around
1033each call to a libev function.
1034
1035However, C<ev_run> can run an indefinite time, so it is not feasible
1036to wait for it to return. One way around this is to wake up the event
1037loop via C<ev_break> and C<ev_async_send>, another way is to set these
1038I<release> and I<acquire> callbacks on the loop.
1039
1040When set, then C<release> will be called just before the thread is
1041suspended waiting for new events, and C<acquire> is called just
1042afterwards.
1043
1044Ideally, C<release> will just call your mutex_unlock function, and
1045C<acquire> will just call the mutex_lock function again.
1046
1047While event loop modifications are allowed between invocations of
1048C<release> and C<acquire> (that's their only purpose after all), no
1049modifications done will affect the event loop, i.e. adding watchers will
1050have no effect on the set of file descriptors being watched, or the time
1051waited. Use an C<ev_async> watcher to wake up C<ev_run> when you want it
1052to take note of any changes you made.
1053
1054In theory, threads executing C<ev_run> will be async-cancel safe between
1055invocations of C<release> and C<acquire>.
1056
1057See also the locking example in the C<THREADS> section later in this
1058document.
1059
1060=item ev_set_userdata (loop, void *data)
1061
1062=item void *ev_userdata (loop)
1063
1064Set and retrieve a single C<void *> associated with a loop. When
1065C<ev_set_userdata> has never been called, then C<ev_userdata> returns
1066C<0>.
1067
1068These two functions can be used to associate arbitrary data with a loop,
1069and are intended solely for the C<invoke_pending_cb>, C<release> and
1070C<acquire> callbacks described above, but of course can be (ab-)used for
1071any other purpose as well.
1072
1073=item ev_verify (loop)
1074
1075This function only does something when C<EV_VERIFY> support has been
1076compiled in, which is the default for non-minimal builds. It tries to go
1077through all internal structures and checks them for validity. If anything
1078is found to be inconsistent, it will print an error message to standard
1079error and call C<abort ()>.
1080
1081This can be used to catch bugs inside libev itself: under normal
1082circumstances, this function will never abort as of course libev keeps its
1083data structures consistent.
1084
1085=back
1086
1087
1088=head1 ANATOMY OF A WATCHER
1089
1090In the following description, uppercase C<TYPE> in names stands for the
1091watcher type, e.g. C<ev_TYPE_start> can mean C<ev_timer_start> for timer
1092watchers and C<ev_io_start> for I/O watchers.
1093
1094A watcher is an opaque structure that you allocate and register to record
1095your interest in some event. To make a concrete example, imagine you want
1096to wait for STDIN to become readable, you would create an C<ev_io> watcher
1097for that:
1098
1099   static void my_cb (struct ev_loop *loop, ev_io *w, int revents)
1100   {
1101     ev_io_stop (w);
1102     ev_break (loop, EVBREAK_ALL);
1103   }
1104
1105   struct ev_loop *loop = ev_default_loop (0);
1106
1107   ev_io stdin_watcher;
1108
1109   ev_init (&stdin_watcher, my_cb);
1110   ev_io_set (&stdin_watcher, STDIN_FILENO, EV_READ);
1111   ev_io_start (loop, &stdin_watcher);
1112
1113   ev_run (loop, 0);
1114
1115As you can see, you are responsible for allocating the memory for your
1116watcher structures (and it is I<usually> a bad idea to do this on the
1117stack).
1118
1119Each watcher has an associated watcher structure (called C<struct ev_TYPE>
1120or simply C<ev_TYPE>, as typedefs are provided for all watcher structs).
1121
1122Each watcher structure must be initialised by a call to C<ev_init (watcher
1123*, callback)>, which expects a callback to be provided. This callback is
1124invoked each time the event occurs (or, in the case of I/O watchers, each
1125time the event loop detects that the file descriptor given is readable
1126and/or writable).
1127
1128Each watcher type further has its own C<< ev_TYPE_set (watcher *, ...) >>
1129macro to configure it, with arguments specific to the watcher type. There
1130is also a macro to combine initialisation and setting in one call: C<<
1131ev_TYPE_init (watcher *, callback, ...) >>.
1132
1133To make the watcher actually watch out for events, you have to start it
1134with a watcher-specific start function (C<< ev_TYPE_start (loop, watcher
1135*) >>), and you can stop watching for events at any time by calling the
1136corresponding stop function (C<< ev_TYPE_stop (loop, watcher *) >>.
1137
1138As long as your watcher is active (has been started but not stopped) you
1139must not touch the values stored in it. Most specifically you must never
1140reinitialise it or call its C<ev_TYPE_set> macro.
1141
1142Each and every callback receives the event loop pointer as first, the
1143registered watcher structure as second, and a bitset of received events as
1144third argument.
1145
1146The received events usually include a single bit per event type received
1147(you can receive multiple events at the same time). The possible bit masks
1148are:
1149
1150=over 4
1151
1152=item C<EV_READ>
1153
1154=item C<EV_WRITE>
1155
1156The file descriptor in the C<ev_io> watcher has become readable and/or
1157writable.
1158
1159=item C<EV_TIMER>
1160
1161The C<ev_timer> watcher has timed out.
1162
1163=item C<EV_PERIODIC>
1164
1165The C<ev_periodic> watcher has timed out.
1166
1167=item C<EV_SIGNAL>
1168
1169The signal specified in the C<ev_signal> watcher has been received by a thread.
1170
1171=item C<EV_CHILD>
1172
1173The pid specified in the C<ev_child> watcher has received a status change.
1174
1175=item C<EV_STAT>
1176
1177The path specified in the C<ev_stat> watcher changed its attributes somehow.
1178
1179=item C<EV_IDLE>
1180
1181The C<ev_idle> watcher has determined that you have nothing better to do.
1182
1183=item C<EV_PREPARE>
1184
1185=item C<EV_CHECK>
1186
1187All C<ev_prepare> watchers are invoked just I<before> C<ev_run> starts to
1188gather new events, and all C<ev_check> watchers are queued (not invoked)
1189just after C<ev_run> has gathered them, but before it queues any callbacks
1190for any received events. That means C<ev_prepare> watchers are the last
1191watchers invoked before the event loop sleeps or polls for new events, and
1192C<ev_check> watchers will be invoked before any other watchers of the same
1193or lower priority within an event loop iteration.
1194
1195Callbacks of both watcher types can start and stop as many watchers as
1196they want, and all of them will be taken into account (for example, a
1197C<ev_prepare> watcher might start an idle watcher to keep C<ev_run> from
1198blocking).
1199
1200=item C<EV_EMBED>
1201
1202The embedded event loop specified in the C<ev_embed> watcher needs attention.
1203
1204=item C<EV_FORK>
1205
1206The event loop has been resumed in the child process after fork (see
1207C<ev_fork>).
1208
1209=item C<EV_CLEANUP>
1210
1211The event loop is about to be destroyed (see C<ev_cleanup>).
1212
1213=item C<EV_ASYNC>
1214
1215The given async watcher has been asynchronously notified (see C<ev_async>).
1216
1217=item C<EV_CUSTOM>
1218
1219Not ever sent (or otherwise used) by libev itself, but can be freely used
1220by libev users to signal watchers (e.g. via C<ev_feed_event>).
1221
1222=item C<EV_ERROR>
1223
1224An unspecified error has occurred, the watcher has been stopped. This might
1225happen because the watcher could not be properly started because libev
1226ran out of memory, a file descriptor was found to be closed or any other
1227problem. Libev considers these application bugs.
1228
1229You best act on it by reporting the problem and somehow coping with the
1230watcher being stopped. Note that well-written programs should not receive
1231an error ever, so when your watcher receives it, this usually indicates a
1232bug in your program.
1233
1234Libev will usually signal a few "dummy" events together with an error, for
1235example it might indicate that a fd is readable or writable, and if your
1236callbacks is well-written it can just attempt the operation and cope with
1237the error from read() or write(). This will not work in multi-threaded
1238programs, though, as the fd could already be closed and reused for another
1239thing, so beware.
1240
1241=back
1242
1243=head2 GENERIC WATCHER FUNCTIONS
1244
1245=over 4
1246
1247=item C<ev_init> (ev_TYPE *watcher, callback)
1248
1249This macro initialises the generic portion of a watcher. The contents
1250of the watcher object can be arbitrary (so C<malloc> will do). Only
1251the generic parts of the watcher are initialised, you I<need> to call
1252the type-specific C<ev_TYPE_set> macro afterwards to initialise the
1253type-specific parts. For each type there is also a C<ev_TYPE_init> macro
1254which rolls both calls into one.
1255
1256You can reinitialise a watcher at any time as long as it has been stopped
1257(or never started) and there are no pending events outstanding.
1258
1259The callback is always of type C<void (*)(struct ev_loop *loop, ev_TYPE *watcher,
1260int revents)>.
1261
1262Example: Initialise an C<ev_io> watcher in two steps.
1263
1264   ev_io w;
1265   ev_init (&w, my_cb);
1266   ev_io_set (&w, STDIN_FILENO, EV_READ);
1267
1268=item C<ev_TYPE_set> (ev_TYPE *watcher, [args])
1269
1270This macro initialises the type-specific parts of a watcher. You need to
1271call C<ev_init> at least once before you call this macro, but you can
1272call C<ev_TYPE_set> any number of times. You must not, however, call this
1273macro on a watcher that is active (it can be pending, however, which is a
1274difference to the C<ev_init> macro).
1275
1276Although some watcher types do not have type-specific arguments
1277(e.g. C<ev_prepare>) you still need to call its C<set> macro.
1278
1279See C<ev_init>, above, for an example.
1280
1281=item C<ev_TYPE_init> (ev_TYPE *watcher, callback, [args])
1282
1283This convenience macro rolls both C<ev_init> and C<ev_TYPE_set> macro
1284calls into a single call. This is the most convenient method to initialise
1285a watcher. The same limitations apply, of course.
1286
1287Example: Initialise and set an C<ev_io> watcher in one step.
1288
1289   ev_io_init (&w, my_cb, STDIN_FILENO, EV_READ);
1290
1291=item C<ev_TYPE_start> (loop, ev_TYPE *watcher)
1292
1293Starts (activates) the given watcher. Only active watchers will receive
1294events. If the watcher is already active nothing will happen.
1295
1296Example: Start the C<ev_io> watcher that is being abused as example in this
1297whole section.
1298
1299   ev_io_start (EV_DEFAULT_UC, &w);
1300
1301=item C<ev_TYPE_stop> (loop, ev_TYPE *watcher)
1302
1303Stops the given watcher if active, and clears the pending status (whether
1304the watcher was active or not).
1305
1306It is possible that stopped watchers are pending - for example,
1307non-repeating timers are being stopped when they become pending - but
1308calling C<ev_TYPE_stop> ensures that the watcher is neither active nor
1309pending. If you want to free or reuse the memory used by the watcher it is
1310therefore a good idea to always call its C<ev_TYPE_stop> function.
1311
1312=item bool ev_is_active (ev_TYPE *watcher)
1313
1314Returns a true value iff the watcher is active (i.e. it has been started
1315and not yet been stopped). As long as a watcher is active you must not modify
1316it.
1317
1318=item bool ev_is_pending (ev_TYPE *watcher)
1319
1320Returns a true value iff the watcher is pending, (i.e. it has outstanding
1321events but its callback has not yet been invoked). As long as a watcher
1322is pending (but not active) you must not call an init function on it (but
1323C<ev_TYPE_set> is safe), you must not change its priority, and you must
1324make sure the watcher is available to libev (e.g. you cannot C<free ()>
1325it).
1326
1327=item callback ev_cb (ev_TYPE *watcher)
1328
1329Returns the callback currently set on the watcher.
1330
1331=item ev_set_cb (ev_TYPE *watcher, callback)
1332
1333Change the callback. You can change the callback at virtually any time
1334(modulo threads).
1335
1336=item ev_set_priority (ev_TYPE *watcher, int priority)
1337
1338=item int ev_priority (ev_TYPE *watcher)
1339
1340Set and query the priority of the watcher. The priority is a small
1341integer between C<EV_MAXPRI> (default: C<2>) and C<EV_MINPRI>
1342(default: C<-2>). Pending watchers with higher priority will be invoked
1343before watchers with lower priority, but priority will not keep watchers
1344from being executed (except for C<ev_idle> watchers).
1345
1346If you need to suppress invocation when higher priority events are pending
1347you need to look at C<ev_idle> watchers, which provide this functionality.
1348
1349You I<must not> change the priority of a watcher as long as it is active or
1350pending.
1351
1352Setting a priority outside the range of C<EV_MINPRI> to C<EV_MAXPRI> is
1353fine, as long as you do not mind that the priority value you query might
1354or might not have been clamped to the valid range.
1355
1356The default priority used by watchers when no priority has been set is
1357always C<0>, which is supposed to not be too high and not be too low :).
1358
1359See L</WATCHER PRIORITY MODELS>, below, for a more thorough treatment of
1360priorities.
1361
1362=item ev_invoke (loop, ev_TYPE *watcher, int revents)
1363
1364Invoke the C<watcher> with the given C<loop> and C<revents>. Neither
1365C<loop> nor C<revents> need to be valid as long as the watcher callback
1366can deal with that fact, as both are simply passed through to the
1367callback.
1368
1369=item int ev_clear_pending (loop, ev_TYPE *watcher)
1370
1371If the watcher is pending, this function clears its pending status and
1372returns its C<revents> bitset (as if its callback was invoked). If the
1373watcher isn't pending it does nothing and returns C<0>.
1374
1375Sometimes it can be useful to "poll" a watcher instead of waiting for its
1376callback to be invoked, which can be accomplished with this function.
1377
1378=item ev_feed_event (loop, ev_TYPE *watcher, int revents)
1379
1380Feeds the given event set into the event loop, as if the specified event
1381had happened for the specified watcher (which must be a pointer to an
1382initialised but not necessarily started event watcher). Obviously you must
1383not free the watcher as long as it has pending events.
1384
1385Stopping the watcher, letting libev invoke it, or calling
1386C<ev_clear_pending> will clear the pending event, even if the watcher was
1387not started in the first place.
1388
1389See also C<ev_feed_fd_event> and C<ev_feed_signal_event> for related
1390functions that do not need a watcher.
1391
1392=back
1393
1394See also the L</ASSOCIATING CUSTOM DATA WITH A WATCHER> and L</BUILDING YOUR
1395OWN COMPOSITE WATCHERS> idioms.
1396
1397=head2 WATCHER STATES
1398
1399There are various watcher states mentioned throughout this manual -
1400active, pending and so on. In this section these states and the rules to
1401transition between them will be described in more detail - and while these
1402rules might look complicated, they usually do "the right thing".
1403
1404=over 4
1405
1406=item initialised
1407
1408Before a watcher can be registered with the event loop it has to be
1409initialised. This can be done with a call to C<ev_TYPE_init>, or calls to
1410C<ev_init> followed by the watcher-specific C<ev_TYPE_set> function.
1411
1412In this state it is simply some block of memory that is suitable for
1413use in an event loop. It can be moved around, freed, reused etc. at
1414will - as long as you either keep the memory contents intact, or call
1415C<ev_TYPE_init> again.
1416
1417=item started/running/active
1418
1419Once a watcher has been started with a call to C<ev_TYPE_start> it becomes
1420property of the event loop, and is actively waiting for events. While in
1421this state it cannot be accessed (except in a few documented ways), moved,
1422freed or anything else - the only legal thing is to keep a pointer to it,
1423and call libev functions on it that are documented to work on active watchers.
1424
1425=item pending
1426
1427If a watcher is active and libev determines that an event it is interested
1428in has occurred (such as a timer expiring), it will become pending. It will
1429stay in this pending state until either it is stopped or its callback is
1430about to be invoked, so it is not normally pending inside the watcher
1431callback.
1432
1433The watcher might or might not be active while it is pending (for example,
1434an expired non-repeating timer can be pending but no longer active). If it
1435is stopped, it can be freely accessed (e.g. by calling C<ev_TYPE_set>),
1436but it is still property of the event loop at this time, so cannot be
1437moved, freed or reused. And if it is active the rules described in the
1438previous item still apply.
1439
1440It is also possible to feed an event on a watcher that is not active (e.g.
1441via C<ev_feed_event>), in which case it becomes pending without being
1442active.
1443
1444=item stopped
1445
1446A watcher can be stopped implicitly by libev (in which case it might still
1447be pending), or explicitly by calling its C<ev_TYPE_stop> function. The
1448latter will clear any pending state the watcher might be in, regardless
1449of whether it was active or not, so stopping a watcher explicitly before
1450freeing it is often a good idea.
1451
1452While stopped (and not pending) the watcher is essentially in the
1453initialised state, that is, it can be reused, moved, modified in any way
1454you wish (but when you trash the memory block, you need to C<ev_TYPE_init>
1455it again).
1456
1457=back
1458
1459=head2 WATCHER PRIORITY MODELS
1460
1461Many event loops support I<watcher priorities>, which are usually small
1462integers that influence the ordering of event callback invocation
1463between watchers in some way, all else being equal.
1464
1465In libev, Watcher priorities can be set using C<ev_set_priority>. See its
1466description for the more technical details such as the actual priority
1467range.
1468
1469There are two common ways how these these priorities are being interpreted
1470by event loops:
1471
1472In the more common lock-out model, higher priorities "lock out" invocation
1473of lower priority watchers, which means as long as higher priority
1474watchers receive events, lower priority watchers are not being invoked.
1475
1476The less common only-for-ordering model uses priorities solely to order
1477callback invocation within a single event loop iteration: Higher priority
1478watchers are invoked before lower priority ones, but they all get invoked
1479before polling for new events.
1480
1481Libev uses the second (only-for-ordering) model for all its watchers
1482except for idle watchers (which use the lock-out model).
1483
1484The rationale behind this is that implementing the lock-out model for
1485watchers is not well supported by most kernel interfaces, and most event
1486libraries will just poll for the same events again and again as long as
1487their callbacks have not been executed, which is very inefficient in the
1488common case of one high-priority watcher locking out a mass of lower
1489priority ones.
1490
1491Static (ordering) priorities are most useful when you have two or more
1492watchers handling the same resource: a typical usage example is having an
1493C<ev_io> watcher to receive data, and an associated C<ev_timer> to handle
1494timeouts. Under load, data might be received while the program handles
1495other jobs, but since timers normally get invoked first, the timeout
1496handler will be executed before checking for data. In that case, giving
1497the timer a lower priority than the I/O watcher ensures that I/O will be
1498handled first even under adverse conditions (which is usually, but not
1499always, what you want).
1500
1501Since idle watchers use the "lock-out" model, meaning that idle watchers
1502will only be executed when no same or higher priority watchers have
1503received events, they can be used to implement the "lock-out" model when
1504required.
1505
1506For example, to emulate how many other event libraries handle priorities,
1507you can associate an C<ev_idle> watcher to each such watcher, and in
1508the normal watcher callback, you just start the idle watcher. The real
1509processing is done in the idle watcher callback. This causes libev to
1510continuously poll and process kernel event data for the watcher, but when
1511the lock-out case is known to be rare (which in turn is rare :), this is
1512workable.
1513
1514Usually, however, the lock-out model implemented that way will perform
1515miserably under the type of load it was designed to handle. In that case,
1516it might be preferable to stop the real watcher before starting the
1517idle watcher, so the kernel will not have to process the event in case
1518the actual processing will be delayed for considerable time.
1519
1520Here is an example of an I/O watcher that should run at a strictly lower
1521priority than the default, and which should only process data when no
1522other events are pending:
1523
1524   ev_idle idle; // actual processing watcher
1525   ev_io io;     // actual event watcher
1526
1527   static void
1528   io_cb (EV_P_ ev_io *w, int revents)
1529   {
1530     // stop the I/O watcher, we received the event, but
1531     // are not yet ready to handle it.
1532     ev_io_stop (EV_A_ w);
1533
1534     // start the idle watcher to handle the actual event.
1535     // it will not be executed as long as other watchers
1536     // with the default priority are receiving events.
1537     ev_idle_start (EV_A_ &idle);
1538   }
1539
1540   static void
1541   idle_cb (EV_P_ ev_idle *w, int revents)
1542   {
1543     // actual processing
1544     read (STDIN_FILENO, ...);
1545
1546     // have to start the I/O watcher again, as
1547     // we have handled the event
1548     ev_io_start (EV_P_ &io);
1549   }
1550
1551   // initialisation
1552   ev_idle_init (&idle, idle_cb);
1553   ev_io_init (&io, io_cb, STDIN_FILENO, EV_READ);
1554   ev_io_start (EV_DEFAULT_ &io);
1555
1556In the "real" world, it might also be beneficial to start a timer, so that
1557low-priority connections can not be locked out forever under load. This
1558enables your program to keep a lower latency for important connections
1559during short periods of high load, while not completely locking out less
1560important ones.
1561
1562
1563=head1 WATCHER TYPES
1564
1565This section describes each watcher in detail, but will not repeat
1566information given in the last section. Any initialisation/set macros,
1567functions and members specific to the watcher type are explained.
1568
1569Members are additionally marked with either I<[read-only]>, meaning that,
1570while the watcher is active, you can look at the member and expect some
1571sensible content, but you must not modify it (you can modify it while the
1572watcher is stopped to your hearts content), or I<[read-write]>, which
1573means you can expect it to have some sensible content while the watcher
1574is active, but you can also modify it. Modifying it may not do something
1575sensible or take immediate effect (or do anything at all), but libev will
1576not crash or malfunction in any way.
1577
1578
1579=head2 C<ev_io> - is this file descriptor readable or writable?
1580
1581I/O watchers check whether a file descriptor is readable or writable
1582in each iteration of the event loop, or, more precisely, when reading
1583would not block the process and writing would at least be able to write
1584some data. This behaviour is called level-triggering because you keep
1585receiving events as long as the condition persists. Remember you can stop
1586the watcher if you don't want to act on the event and neither want to
1587receive future events.
1588
1589In general you can register as many read and/or write event watchers per
1590fd as you want (as long as you don't confuse yourself). Setting all file
1591descriptors to non-blocking mode is also usually a good idea (but not
1592required if you know what you are doing).
1593
1594Another thing you have to watch out for is that it is quite easy to
1595receive "spurious" readiness notifications, that is, your callback might
1596be called with C<EV_READ> but a subsequent C<read>(2) will actually block
1597because there is no data. It is very easy to get into this situation even
1598with a relatively standard program structure. Thus it is best to always
1599use non-blocking I/O: An extra C<read>(2) returning C<EAGAIN> is far
1600preferable to a program hanging until some data arrives.
1601
1602If you cannot run the fd in non-blocking mode (for example you should
1603not play around with an Xlib connection), then you have to separately
1604re-test whether a file descriptor is really ready with a known-to-be good
1605interface such as poll (fortunately in the case of Xlib, it already does
1606this on its own, so its quite safe to use). Some people additionally
1607use C<SIGALRM> and an interval timer, just to be sure you won't block
1608indefinitely.
1609
1610But really, best use non-blocking mode.
1611
1612=head3 The special problem of disappearing file descriptors
1613
1614Some backends (e.g. kqueue, epoll) need to be told about closing a file
1615descriptor (either due to calling C<close> explicitly or any other means,
1616such as C<dup2>). The reason is that you register interest in some file
1617descriptor, but when it goes away, the operating system will silently drop
1618this interest. If another file descriptor with the same number then is
1619registered with libev, there is no efficient way to see that this is, in
1620fact, a different file descriptor.
1621
1622To avoid having to explicitly tell libev about such cases, libev follows
1623the following policy:  Each time C<ev_io_set> is being called, libev
1624will assume that this is potentially a new file descriptor, otherwise
1625it is assumed that the file descriptor stays the same. That means that
1626you I<have> to call C<ev_io_set> (or C<ev_io_init>) when you change the
1627descriptor even if the file descriptor number itself did not change.
1628
1629This is how one would do it normally anyway, the important point is that
1630the libev application should not optimise around libev but should leave
1631optimisations to libev.
1632
1633=head3 The special problem of dup'ed file descriptors
1634
1635Some backends (e.g. epoll), cannot register events for file descriptors,
1636but only events for the underlying file descriptions. That means when you
1637have C<dup ()>'ed file descriptors or weirder constellations, and register
1638events for them, only one file descriptor might actually receive events.
1639
1640There is no workaround possible except not registering events
1641for potentially C<dup ()>'ed file descriptors, or to resort to
1642C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.
1643
1644=head3 The special problem of files
1645
1646Many people try to use C<select> (or libev) on file descriptors
1647representing files, and expect it to become ready when their program
1648doesn't block on disk accesses (which can take a long time on their own).
1649
1650However, this cannot ever work in the "expected" way - you get a readiness
1651notification as soon as the kernel knows whether and how much data is
1652there, and in the case of open files, that's always the case, so you
1653always get a readiness notification instantly, and your read (or possibly
1654write) will still block on the disk I/O.
1655
1656Another way to view it is that in the case of sockets, pipes, character
1657devices and so on, there is another party (the sender) that delivers data
1658on its own, but in the case of files, there is no such thing: the disk
1659will not send data on its own, simply because it doesn't know what you
1660wish to read - you would first have to request some data.
1661
1662Since files are typically not-so-well supported by advanced notification
1663mechanism, libev tries hard to emulate POSIX behaviour with respect
1664to files, even though you should not use it. The reason for this is
1665convenience: sometimes you want to watch STDIN or STDOUT, which is
1666usually a tty, often a pipe, but also sometimes files or special devices
1667(for example, C<epoll> on Linux works with F</dev/random> but not with
1668F</dev/urandom>), and even though the file might better be served with
1669asynchronous I/O instead of with non-blocking I/O, it is still useful when
1670it "just works" instead of freezing.
1671
1672So avoid file descriptors pointing to files when you know it (e.g. use
1673libeio), but use them when it is convenient, e.g. for STDIN/STDOUT, or
1674when you rarely read from a file instead of from a socket, and want to
1675reuse the same code path.
1676
1677=head3 The special problem of fork
1678
1679Some backends (epoll, kqueue) do not support C<fork ()> at all or exhibit
1680useless behaviour. Libev fully supports fork, but needs to be told about
1681it in the child if you want to continue to use it in the child.
1682
1683To support fork in your child processes, you have to call C<ev_loop_fork
1684()> after a fork in the child, enable C<EVFLAG_FORKCHECK>, or resort to
1685C<EVBACKEND_SELECT> or C<EVBACKEND_POLL>.
1686
1687=head3 The special problem of SIGPIPE
1688
1689While not really specific to libev, it is easy to forget about C<SIGPIPE>:
1690when writing to a pipe whose other end has been closed, your program gets
1691sent a SIGPIPE, which, by default, aborts your program. For most programs
1692this is sensible behaviour, for daemons, this is usually undesirable.
1693
1694So when you encounter spurious, unexplained daemon exits, make sure you
1695ignore SIGPIPE (and maybe make sure you log the exit status of your daemon
1696somewhere, as that would have given you a big clue).
1697
1698=head3 The special problem of accept()ing when you can't
1699
1700Many implementations of the POSIX C<accept> function (for example,
1701found in post-2004 Linux) have the peculiar behaviour of not removing a
1702connection from the pending queue in all error cases.
1703
1704For example, larger servers often run out of file descriptors (because
1705of resource limits), causing C<accept> to fail with C<ENFILE> but not
1706rejecting the connection, leading to libev signalling readiness on
1707the next iteration again (the connection still exists after all), and
1708typically causing the program to loop at 100% CPU usage.
1709
1710Unfortunately, the set of errors that cause this issue differs between
1711operating systems, there is usually little the app can do to remedy the
1712situation, and no known thread-safe method of removing the connection to
1713cope with overload is known (to me).
1714
1715One of the easiest ways to handle this situation is to just ignore it
1716- when the program encounters an overload, it will just loop until the
1717situation is over. While this is a form of busy waiting, no OS offers an
1718event-based way to handle this situation, so it's the best one can do.
1719
1720A better way to handle the situation is to log any errors other than
1721C<EAGAIN> and C<EWOULDBLOCK>, making sure not to flood the log with such
1722messages, and continue as usual, which at least gives the user an idea of
1723what could be wrong ("raise the ulimit!"). For extra points one could stop
1724the C<ev_io> watcher on the listening fd "for a while", which reduces CPU
1725usage.
1726
1727If your program is single-threaded, then you could also keep a dummy file
1728descriptor for overload situations (e.g. by opening F</dev/null>), and
1729when you run into C<ENFILE> or C<EMFILE>, close it, run C<accept>,
1730close that fd, and create a new dummy fd. This will gracefully refuse
1731clients under typical overload conditions.
1732
1733The last way to handle it is to simply log the error and C<exit>, as
1734is often done with C<malloc> failures, but this results in an easy
1735opportunity for a DoS attack.
1736
1737=head3 Watcher-Specific Functions
1738
1739=over 4
1740
1741=item ev_io_init (ev_io *, callback, int fd, int events)
1742
1743=item ev_io_set (ev_io *, int fd, int events)
1744
1745Configures an C<ev_io> watcher. The C<fd> is the file descriptor to
1746receive events for and C<events> is either C<EV_READ>, C<EV_WRITE> or
1747C<EV_READ | EV_WRITE>, to express the desire to receive the given events.
1748
1749=item int fd [read-only]
1750
1751The file descriptor being watched.
1752
1753=item int events [read-only]
1754
1755The events being watched.
1756
1757=back
1758
1759=head3 Examples
1760
1761Example: Call C<stdin_readable_cb> when STDIN_FILENO has become, well
1762readable, but only once. Since it is likely line-buffered, you could
1763attempt to read a whole line in the callback.
1764
1765   static void
1766   stdin_readable_cb (struct ev_loop *loop, ev_io *w, int revents)
1767   {
1768      ev_io_stop (loop, w);
1769     .. read from stdin here (or from w->fd) and handle any I/O errors
1770   }
1771
1772   ...
1773   struct ev_loop *loop = ev_default_init (0);
1774   ev_io stdin_readable;
1775   ev_io_init (&stdin_readable, stdin_readable_cb, STDIN_FILENO, EV_READ);
1776   ev_io_start (loop, &stdin_readable);
1777   ev_run (loop, 0);
1778
1779
1780=head2 C<ev_timer> - relative and optionally repeating timeouts
1781
1782Timer watchers are simple relative timers that generate an event after a
1783given time, and optionally repeating in regular intervals after that.
1784
1785The timers are based on real time, that is, if you register an event that
1786times out after an hour and you reset your system clock to January last
1787year, it will still time out after (roughly) one hour. "Roughly" because
1788detecting time jumps is hard, and some inaccuracies are unavoidable (the
1789monotonic clock option helps a lot here).
1790
1791The callback is guaranteed to be invoked only I<after> its timeout has
1792passed (not I<at>, so on systems with very low-resolution clocks this
1793might introduce a small delay, see "the special problem of being too
1794early", below). If multiple timers become ready during the same loop
1795iteration then the ones with earlier time-out values are invoked before
1796ones of the same priority with later time-out values (but this is no
1797longer true when a callback calls C<ev_run> recursively).
1798
1799=head3 Be smart about timeouts
1800
1801Many real-world problems involve some kind of timeout, usually for error
1802recovery. A typical example is an HTTP request - if the other side hangs,
1803you want to raise some error after a while.
1804
1805What follows are some ways to handle this problem, from obvious and
1806inefficient to smart and efficient.
1807
1808In the following, a 60 second activity timeout is assumed - a timeout that
1809gets reset to 60 seconds each time there is activity (e.g. each time some
1810data or other life sign was received).
1811
1812=over 4
1813
1814=item 1. Use a timer and stop, reinitialise and start it on activity.
1815
1816This is the most obvious, but not the most simple way: In the beginning,
1817start the watcher:
1818
1819   ev_timer_init (timer, callback, 60., 0.);
1820   ev_timer_start (loop, timer);
1821
1822Then, each time there is some activity, C<ev_timer_stop> it, initialise it
1823and start it again:
1824
1825   ev_timer_stop (loop, timer);
1826   ev_timer_set (timer, 60., 0.);
1827   ev_timer_start (loop, timer);
1828
1829This is relatively simple to implement, but means that each time there is
1830some activity, libev will first have to remove the timer from its internal
1831data structure and then add it again. Libev tries to be fast, but it's
1832still not a constant-time operation.
1833
1834=item 2. Use a timer and re-start it with C<ev_timer_again> inactivity.
1835
1836This is the easiest way, and involves using C<ev_timer_again> instead of
1837C<ev_timer_start>.
1838
1839To implement this, configure an C<ev_timer> with a C<repeat> value
1840of C<60> and then call C<ev_timer_again> at start and each time you
1841successfully read or write some data. If you go into an idle state where
1842you do not expect data to travel on the socket, you can C<ev_timer_stop>
1843the timer, and C<ev_timer_again> will automatically restart it if need be.
1844
1845That means you can ignore both the C<ev_timer_start> function and the
1846C<after> argument to C<ev_timer_set>, and only ever use the C<repeat>
1847member and C<ev_timer_again>.
1848
1849At start:
1850
1851   ev_init (timer, callback);
1852   timer->repeat = 60.;
1853   ev_timer_again (loop, timer);
1854
1855Each time there is some activity:
1856
1857   ev_timer_again (loop, timer);
1858
1859It is even possible to change the time-out on the fly, regardless of
1860whether the watcher is active or not:
1861
1862   timer->repeat = 30.;
1863   ev_timer_again (loop, timer);
1864
1865This is slightly more efficient then stopping/starting the timer each time
1866you want to modify its timeout value, as libev does not have to completely
1867remove and re-insert the timer from/into its internal data structure.
1868
1869It is, however, even simpler than the "obvious" way to do it.
1870
1871=item 3. Let the timer time out, but then re-arm it as required.
1872
1873This method is more tricky, but usually most efficient: Most timeouts are
1874relatively long compared to the intervals between other activity - in
1875our example, within 60 seconds, there are usually many I/O events with
1876associated activity resets.
1877
1878In this case, it would be more efficient to leave the C<ev_timer> alone,
1879but remember the time of last activity, and check for a real timeout only
1880within the callback:
1881
1882   ev_tstamp timeout = 60.;
1883   ev_tstamp last_activity; // time of last activity
1884   ev_timer timer;
1885
1886   static void
1887   callback (EV_P_ ev_timer *w, int revents)
1888   {
1889     // calculate when the timeout would happen
1890     ev_tstamp after = last_activity - ev_now (EV_A) + timeout;
1891
1892     // if negative, it means we the timeout already occurred
1893     if (after < 0.)
1894       {
1895         // timeout occurred, take action
1896       }
1897     else
1898       {
1899         // callback was invoked, but there was some recent 
1900         // activity. simply restart the timer to time out
1901         // after "after" seconds, which is the earliest time
1902         // the timeout can occur.
1903         ev_timer_set (w, after, 0.);
1904         ev_timer_start (EV_A_ w);
1905       }
1906   }
1907
1908To summarise the callback: first calculate in how many seconds the
1909timeout will occur (by calculating the absolute time when it would occur,
1910C<last_activity + timeout>, and subtracting the current time, C<ev_now
1911(EV_A)> from that).
1912
1913If this value is negative, then we are already past the timeout, i.e. we
1914timed out, and need to do whatever is needed in this case.
1915
1916Otherwise, we now the earliest time at which the timeout would trigger,
1917and simply start the timer with this timeout value.
1918
1919In other words, each time the callback is invoked it will check whether
1920the timeout occurred. If not, it will simply reschedule itself to check
1921again at the earliest time it could time out. Rinse. Repeat.
1922
1923This scheme causes more callback invocations (about one every 60 seconds
1924minus half the average time between activity), but virtually no calls to
1925libev to change the timeout.
1926
1927To start the machinery, simply initialise the watcher and set
1928C<last_activity> to the current time (meaning there was some activity just
1929now), then call the callback, which will "do the right thing" and start
1930the timer:
1931
1932   last_activity = ev_now (EV_A);
1933   ev_init (&timer, callback);
1934   callback (EV_A_ &timer, 0);
1935
1936When there is some activity, simply store the current time in
1937C<last_activity>, no libev calls at all:
1938
1939   if (activity detected)
1940     last_activity = ev_now (EV_A);
1941
1942When your timeout value changes, then the timeout can be changed by simply
1943providing a new value, stopping the timer and calling the callback, which
1944will again do the right thing (for example, time out immediately :).
1945
1946   timeout = new_value;
1947   ev_timer_stop (EV_A_ &timer);
1948   callback (EV_A_ &timer, 0);
1949
1950This technique is slightly more complex, but in most cases where the
1951time-out is unlikely to be triggered, much more efficient.
1952
1953=item 4. Wee, just use a double-linked list for your timeouts.
1954
1955If there is not one request, but many thousands (millions...), all
1956employing some kind of timeout with the same timeout value, then one can
1957do even better:
1958
1959When starting the timeout, calculate the timeout value and put the timeout
1960at the I<end> of the list.
1961
1962Then use an C<ev_timer> to fire when the timeout at the I<beginning> of
1963the list is expected to fire (for example, using the technique #3).
1964
1965When there is some activity, remove the timer from the list, recalculate
1966the timeout, append it to the end of the list again, and make sure to
1967update the C<ev_timer> if it was taken from the beginning of the list.
1968
1969This way, one can manage an unlimited number of timeouts in O(1) time for
1970starting, stopping and updating the timers, at the expense of a major
1971complication, and having to use a constant timeout. The constant timeout
1972ensures that the list stays sorted.
1973
1974=back
1975
1976So which method the best?
1977
1978Method #2 is a simple no-brain-required solution that is adequate in most
1979situations. Method #3 requires a bit more thinking, but handles many cases
1980better, and isn't very complicated either. In most case, choosing either
1981one is fine, with #3 being better in typical situations.
1982
1983Method #1 is almost always a bad idea, and buys you nothing. Method #4 is
1984rather complicated, but extremely efficient, something that really pays
1985off after the first million or so of active timers, i.e. it's usually
1986overkill :)
1987
1988=head3 The special problem of being too early
1989
1990If you ask a timer to call your callback after three seconds, then
1991you expect it to be invoked after three seconds - but of course, this
1992cannot be guaranteed to infinite precision. Less obviously, it cannot be
1993guaranteed to any precision by libev - imagine somebody suspending the
1994process with a STOP signal for a few hours for example.
1995
1996So, libev tries to invoke your callback as soon as possible I<after> the
1997delay has occurred, but cannot guarantee this.
1998
1999A less obvious failure mode is calling your callback too early: many event
2000loops compare timestamps with a "elapsed delay >= requested delay", but
2001this can cause your callback to be invoked much earlier than you would
2002expect.
2003
2004To see why, imagine a system with a clock that only offers full second
2005resolution (think windows if you can't come up with a broken enough OS
2006yourself). If you schedule a one-second timer at the time 500.9, then the
2007event loop will schedule your timeout to elapse at a system time of 500
2008(500.9 truncated to the resolution) + 1, or 501.
2009
2010If an event library looks at the timeout 0.1s later, it will see "501 >=
2011501" and invoke the callback 0.1s after it was started, even though a
2012one-second delay was requested - this is being "too early", despite best
2013intentions.
2014
2015This is the reason why libev will never invoke the callback if the elapsed
2016delay equals the requested delay, but only when the elapsed delay is
2017larger than the requested delay. In the example above, libev would only invoke
2018the callback at system time 502, or 1.1s after the timer was started.
2019
2020So, while libev cannot guarantee that your callback will be invoked
2021exactly when requested, it I<can> and I<does> guarantee that the requested
2022delay has actually elapsed, or in other words, it always errs on the "too
2023late" side of things.
2024
2025=head3 The special problem of time updates
2026
2027Establishing the current time is a costly operation (it usually takes
2028at least one system call): EV therefore updates its idea of the current
2029time only before and after C<ev_run> collects new events, which causes a
2030growing difference between C<ev_now ()> and C<ev_time ()> when handling
2031lots of events in one iteration.
2032
2033The relative timeouts are calculated relative to the C<ev_now ()>
2034time. This is usually the right thing as this timestamp refers to the time
2035of the event triggering whatever timeout you are modifying/starting. If
2036you suspect event processing to be delayed and you I<need> to base the
2037timeout on the current time, use something like the following to adjust
2038for it:
2039
2040   ev_timer_set (&timer, after + (ev_time () - ev_now ()), 0.);
2041
2042If the event loop is suspended for a long time, you can also force an
2043update of the time returned by C<ev_now ()> by calling C<ev_now_update
2044()>, although that will push the event time of all outstanding events
2045further into the future.
2046
2047=head3 The special problem of unsynchronised clocks
2048
2049Modern systems have a variety of clocks - libev itself uses the normal
2050"wall clock" clock and, if available, the monotonic clock (to avoid time
2051jumps).
2052
2053Neither of these clocks is synchronised with each other or any other clock
2054on the system, so C<ev_time ()> might return a considerably different time
2055than C<gettimeofday ()> or C<time ()>. On a GNU/Linux system, for example,
2056a call to C<gettimeofday> might return a second count that is one higher
2057than a directly following call to C<time>.
2058
2059The moral of this is to only compare libev-related timestamps with
2060C<ev_time ()> and C<ev_now ()>, at least if you want better precision than
2061a second or so.
2062
2063One more problem arises due to this lack of synchronisation: if libev uses
2064the system monotonic clock and you compare timestamps from C<ev_time>
2065or C<ev_now> from when you started your timer and when your callback is
2066invoked, you will find that sometimes the callback is a bit "early".
2067
2068This is because C<ev_timer>s work in real time, not wall clock time, so
2069libev makes sure your callback is not invoked before the delay happened,
2070I<measured according to the real time>, not the system clock.
2071
2072If your timeouts are based on a physical timescale (e.g. "time out this
2073connection after 100 seconds") then this shouldn't bother you as it is
2074exactly the right behaviour.
2075
2076If you want to compare wall clock/system timestamps to your timers, then
2077you need to use C<ev_periodic>s, as these are based on the wall clock
2078time, where your comparisons will always generate correct results.
2079
2080=head3 The special problems of suspended animation
2081
2082When you leave the server world it is quite customary to hit machines that
2083can suspend/hibernate - what happens to the clocks during such a suspend?
2084
2085Some quick tests made with a Linux 2.6.28 indicate that a suspend freezes
2086all processes, while the clocks (C<times>, C<CLOCK_MONOTONIC>) continue
2087to run until the system is suspended, but they will not advance while the
2088system is suspended. That means, on resume, it will be as if the program
2089was frozen for a few seconds, but the suspend time will not be counted
2090towards C<ev_timer> when a monotonic clock source is used. The real time
2091clock advanced as expected, but if it is used as sole clocksource, then a
2092long suspend would be detected as a time jump by libev, and timers would
2093be adjusted accordingly.
2094
2095I would not be surprised to see different behaviour in different between
2096operating systems, OS versions or even different hardware.
2097
2098The other form of suspend (job control, or sending a SIGSTOP) will see a
2099time jump in the monotonic clocks and the realtime clock. If the program
2100is suspended for a very long time, and monotonic clock sources are in use,
2101then you can expect C<ev_timer>s to expire as the full suspension time
2102will be counted towards the timers. When no monotonic clock source is in
2103use, then libev will again assume a timejump and adjust accordingly.
2104
2105It might be beneficial for this latter case to call C<ev_suspend>
2106and C<ev_resume> in code that handles C<SIGTSTP>, to at least get
2107deterministic behaviour in this case (you can do nothing against
2108C<SIGSTOP>).
2109
2110=head3 Watcher-Specific Functions and Data Members
2111
2112=over 4
2113
2114=item ev_timer_init (ev_timer *, callback, ev_tstamp after, ev_tstamp repeat)
2115
2116=item ev_timer_set (ev_timer *, ev_tstamp after, ev_tstamp repeat)
2117
2118Configure the timer to trigger after C<after> seconds. If C<repeat>
2119is C<0.>, then it will automatically be stopped once the timeout is
2120reached. If it is positive, then the timer will automatically be
2121configured to trigger again C<repeat> seconds later, again, and again,
2122until stopped manually.
2123
2124The timer itself will do a best-effort at avoiding drift, that is, if
2125you configure a timer to trigger every 10 seconds, then it will normally
2126trigger at exactly 10 second intervals. If, however, your program cannot
2127keep up with the timer (because it takes longer than those 10 seconds to
2128do stuff) the timer will not fire more than once per event loop iteration.
2129
2130=item ev_timer_again (loop, ev_timer *)
2131
2132This will act as if the timer timed out, and restarts it again if it is
2133repeating. It basically works like calling C<ev_timer_stop>, updating the
2134timeout to the C<repeat> value and calling C<ev_timer_start>.
2135
2136The exact semantics are as in the following rules, all of which will be
2137applied to the watcher:
2138
2139=over 4
2140
2141=item If the timer is pending, the pending status is always cleared.
2142
2143=item If the timer is started but non-repeating, stop it (as if it timed
2144out, without invoking it).
2145
2146=item If the timer is repeating, make the C<repeat> value the new timeout
2147and start the timer, if necessary.
2148
2149=back
2150
2151This sounds a bit complicated, see L</Be smart about timeouts>, above, for a
2152usage example.
2153
2154=item ev_tstamp ev_timer_remaining (loop, ev_timer *)
2155
2156Returns the remaining time until a timer fires. If the timer is active,
2157then this time is relative to the current event loop time, otherwise it's
2158the timeout value currently configured.
2159
2160That is, after an C<ev_timer_set (w, 5, 7)>, C<ev_timer_remaining> returns
2161C<5>. When the timer is started and one second passes, C<ev_timer_remaining>
2162will return C<4>. When the timer expires and is restarted, it will return
2163roughly C<7> (likely slightly less as callback invocation takes some time,
2164too), and so on.
2165
2166=item ev_tstamp repeat [read-write]
2167
2168The current C<repeat> value. Will be used each time the watcher times out
2169or C<ev_timer_again> is called, and determines the next timeout (if any),
2170which is also when any modifications are taken into account.
2171
2172=back
2173
2174=head3 Examples
2175
2176Example: Create a timer that fires after 60 seconds.
2177
2178   static void
2179   one_minute_cb (struct ev_loop *loop, ev_timer *w, int revents)
2180   {
2181     .. one minute over, w is actually stopped right here
2182   }
2183
2184   ev_timer mytimer;
2185   ev_timer_init (&mytimer, one_minute_cb, 60., 0.);
2186   ev_timer_start (loop, &mytimer);
2187
2188Example: Create a timeout timer that times out after 10 seconds of
2189inactivity.
2190
2191   static void
2192   timeout_cb (struct ev_loop *loop, ev_timer *w, int revents)
2193   {
2194     .. ten seconds without any activity
2195   }
2196
2197   ev_timer mytimer;
2198   ev_timer_init (&mytimer, timeout_cb, 0., 10.); /* note, only repeat used */
2199   ev_timer_again (&mytimer); /* start timer */
2200   ev_run (loop, 0);
2201
2202   // and in some piece of code that gets executed on any "activity":
2203   // reset the timeout to start ticking again at 10 seconds
2204   ev_timer_again (&mytimer);
2205
2206
2207=head2 C<ev_periodic> - to cron or not to cron?
2208
2209Periodic watchers are also timers of a kind, but they are very versatile
2210(and unfortunately a bit complex).
2211
2212Unlike C<ev_timer>, periodic watchers are not based on real time (or
2213relative time, the physical time that passes) but on wall clock time
2214(absolute time, the thing you can read on your calender or clock). The
2215difference is that wall clock time can run faster or slower than real
2216time, and time jumps are not uncommon (e.g. when you adjust your
2217wrist-watch).
2218
2219You can tell a periodic watcher to trigger after some specific point
2220in time: for example, if you tell a periodic watcher to trigger "in 10
2221seconds" (by specifying e.g. C<ev_now () + 10.>, that is, an absolute time
2222not a delay) and then reset your system clock to January of the previous
2223year, then it will take a year or more to trigger the event (unlike an
2224C<ev_timer>, which would still trigger roughly 10 seconds after starting
2225it, as it uses a relative timeout).
2226
2227C<ev_periodic> watchers can also be used to implement vastly more complex
2228timers, such as triggering an event on each "midnight, local time", or
2229other complicated rules. This cannot be done with C<ev_timer> watchers, as
2230those cannot react to time jumps.
2231
2232As with timers, the callback is guaranteed to be invoked only when the
2233point in time where it is supposed to trigger has passed. If multiple
2234timers become ready during the same loop iteration then the ones with
2235earlier time-out values are invoked before ones with later time-out values
2236(but this is no longer true when a callback calls C<ev_run> recursively).
2237
2238=head3 Watcher-Specific Functions and Data Members
2239
2240=over 4
2241
2242=item ev_periodic_init (ev_periodic *, callback, ev_tstamp offset, ev_tstamp interval, reschedule_cb)
2243
2244=item ev_periodic_set (ev_periodic *, ev_tstamp offset, ev_tstamp interval, reschedule_cb)
2245
2246Lots of arguments, let's sort it out... There are basically three modes of
2247operation, and we will explain them from simplest to most complex:
2248
2249=over 4
2250
2251=item * absolute timer (offset = absolute time, interval = 0, reschedule_cb = 0)
2252
2253In this configuration the watcher triggers an event after the wall clock
2254time C<offset> has passed. It will not repeat and will not adjust when a
2255time jump occurs, that is, if it is to be run at January 1st 2011 then it
2256will be stopped and invoked when the system clock reaches or surpasses
2257this point in time.
2258
2259=item * repeating interval timer (offset = offset within interval, interval > 0, reschedule_cb = 0)
2260
2261In this mode the watcher will always be scheduled to time out at the next
2262C<offset + N * interval> time (for some integer N, which can also be
2263negative) and then repeat, regardless of any time jumps. The C<offset>
2264argument is merely an offset into the C<interval> periods.
2265
2266This can be used to create timers that do not drift with respect to the
2267system clock, for example, here is an C<ev_periodic> that triggers each
2268hour, on the hour (with respect to UTC):
2269
2270   ev_periodic_set (&periodic, 0., 3600., 0);
2271
2272This doesn't mean there will always be 3600 seconds in between triggers,
2273but only that the callback will be called when the system time shows a
2274full hour (UTC), or more correctly, when the system time is evenly divisible
2275by 3600.
2276
2277Another way to think about it (for the mathematically inclined) is that
2278C<ev_periodic> will try to run the callback in this mode at the next possible
2279time where C<time = offset (mod interval)>, regardless of any time jumps.
2280
2281The C<interval> I<MUST> be positive, and for numerical stability, the
2282interval value should be higher than C<1/8192> (which is around 100
2283microseconds) and C<offset> should be higher than C<0> and should have
2284at most a similar magnitude as the current time (say, within a factor of
2285ten). Typical values for offset are, in fact, C<0> or something between
2286C<0> and C<interval>, which is also the recommended range.
2287
2288Note also that there is an upper limit to how often a timer can fire (CPU
2289speed for example), so if C<interval> is very small then timing stability
2290will of course deteriorate. Libev itself tries to be exact to be about one
2291millisecond (if the OS supports it and the machine is fast enough).
2292
2293=item * manual reschedule mode (offset ignored, interval ignored, reschedule_cb = callback)
2294
2295In this mode the values for C<interval> and C<offset> are both being
2296ignored. Instead, each time the periodic watcher gets scheduled, the
2297reschedule callback will be called with the watcher as first, and the
2298current time as second argument.
2299
2300NOTE: I<This callback MUST NOT stop or destroy any periodic watcher, ever,
2301or make ANY other event loop modifications whatsoever, unless explicitly
2302allowed by documentation here>.
2303
2304If you need to stop it, return C<now + 1e30> (or so, fudge fudge) and stop
2305it afterwards (e.g. by starting an C<ev_prepare> watcher, which is the
2306only event loop modification you are allowed to do).
2307
2308The callback prototype is C<ev_tstamp (*reschedule_cb)(ev_periodic
2309*w, ev_tstamp now)>, e.g.:
2310
2311   static ev_tstamp
2312   my_rescheduler (ev_periodic *w, ev_tstamp now)
2313   {
2314     return now + 60.;
2315   }
2316
2317It must return the next time to trigger, based on the passed time value
2318(that is, the lowest time value larger than to the second argument). It
2319will usually be called just before the callback will be triggered, but
2320might be called at other times, too.
2321
2322NOTE: I<< This callback must always return a time that is higher than or
2323equal to the passed C<now> value >>.
2324
2325This can be used to create very complex timers, such as a timer that
2326triggers on "next midnight, local time". To do this, you would calculate the
2327next midnight after C<now> and return the timestamp value for this. How
2328you do this is, again, up to you (but it is not trivial, which is the main
2329reason I omitted it as an example).
2330
2331=back
2332
2333=item ev_periodic_again (loop, ev_periodic *)
2334
2335Simply stops and restarts the periodic watcher again. This is only useful
2336when you changed some parameters or the reschedule callback would return
2337a different time than the last time it was called (e.g. in a crond like
2338program when the crontabs have changed).
2339
2340=item ev_tstamp ev_periodic_at (ev_periodic *)
2341
2342When active, returns the absolute time that the watcher is supposed
2343to trigger next. This is not the same as the C<offset> argument to
2344C<ev_periodic_set>, but indeed works even in interval and manual
2345rescheduling modes.
2346
2347=item ev_tstamp offset [read-write]
2348
2349When repeating, this contains the offset value, otherwise this is the
2350absolute point in time (the C<offset> value passed to C<ev_periodic_set>,
2351although libev might modify this value for better numerical stability).
2352
2353Can be modified any time, but changes only take effect when the periodic
2354timer fires or C<ev_periodic_again> is being called.
2355
2356=item ev_tstamp interval [read-write]
2357
2358The current interval value. Can be modified any time, but changes only
2359take effect when the periodic timer fires or C<ev_periodic_again> is being
2360called.
2361
2362=item ev_tstamp (*reschedule_cb)(ev_periodic *w, ev_tstamp now) [read-write]
2363
2364The current reschedule callback, or C<0>, if this functionality is
2365switched off. Can be changed any time, but changes only take effect when
2366the periodic timer fires or C<ev_periodic_again> is being called.
2367
2368=back
2369
2370=head3 Examples
2371
2372Example: Call a callback every hour, or, more precisely, whenever the
2373system time is divisible by 3600. The callback invocation times have
2374potentially a lot of jitter, but good long-term stability.
2375
2376   static void
2377   clock_cb (struct ev_loop *loop, ev_periodic *w, int revents)
2378   {
2379     ... its now a full hour (UTC, or TAI or whatever your clock follows)
2380   }
2381
2382   ev_periodic hourly_tick;
2383   ev_periodic_init (&hourly_tick, clock_cb, 0., 3600., 0);
2384   ev_periodic_start (loop, &hourly_tick);
2385
2386Example: The same as above, but use a reschedule callback to do it:
2387
2388   #include <math.h>
2389
2390   static ev_tstamp
2391   my_scheduler_cb (ev_periodic *w, ev_tstamp now)
2392   {
2393     return now + (3600. - fmod (now, 3600.));
2394   }
2395
2396   ev_periodic_init (&hourly_tick, clock_cb, 0., 0., my_scheduler_cb);
2397
2398Example: Call a callback every hour, starting now:
2399
2400   ev_periodic hourly_tick;
2401   ev_periodic_init (&hourly_tick, clock_cb,
2402                     fmod (ev_now (loop), 3600.), 3600., 0);
2403   ev_periodic_start (loop, &hourly_tick);
2404
2405
2406=head2 C<ev_signal> - signal me when a signal gets signalled!
2407
2408Signal watchers will trigger an event when the process receives a specific
2409signal one or more times. Even though signals are very asynchronous, libev
2410will try its best to deliver signals synchronously, i.e. as part of the
2411normal event processing, like any other event.
2412
2413If you want signals to be delivered truly asynchronously, just use
2414C<sigaction> as you would do without libev and forget about sharing
2415the signal. You can even use C<ev_async> from a signal handler to
2416synchronously wake up an event loop.
2417
2418You can configure as many watchers as you like for the same signal, but
2419only within the same loop, i.e. you can watch for C<SIGINT> in your
2420default loop and for C<SIGIO> in another loop, but you cannot watch for
2421C<SIGINT> in both the default loop and another loop at the same time. At
2422the moment, C<SIGCHLD> is permanently tied to the default loop.
2423
2424Only after the first watcher for a signal is started will libev actually
2425register something with the kernel. It thus coexists with your own signal
2426handlers as long as you don't register any with libev for the same signal.
2427
2428If possible and supported, libev will install its handlers with
2429C<SA_RESTART> (or equivalent) behaviour enabled, so system calls should
2430not be unduly interrupted. If you have a problem with system calls getting
2431interrupted by signals you can block all signals in an C<ev_check> watcher
2432and unblock them in an C<ev_prepare> watcher.
2433
2434=head3 The special problem of inheritance over fork/execve/pthread_create
2435
2436Both the signal mask (C<sigprocmask>) and the signal disposition
2437(C<sigaction>) are unspecified after starting a signal watcher (and after
2438stopping it again), that is, libev might or might not block the signal,
2439and might or might not set or restore the installed signal handler (but
2440see C<EVFLAG_NOSIGMASK>).
2441
2442While this does not matter for the signal disposition (libev never
2443sets signals to C<SIG_IGN>, so handlers will be reset to C<SIG_DFL> on
2444C<execve>), this matters for the signal mask: many programs do not expect
2445certain signals to be blocked.
2446
2447This means that before calling C<exec> (from the child) you should reset
2448the signal mask to whatever "default" you expect (all clear is a good
2449choice usually).
2450
2451The simplest way to ensure that the signal mask is reset in the child is
2452to install a fork handler with C<pthread_atfork> that resets it. That will
2453catch fork calls done by libraries (such as the libc) as well.
2454
2455In current versions of libev, the signal will not be blocked indefinitely
2456unless you use the C<signalfd> API (C<EV_SIGNALFD>). While this reduces
2457the window of opportunity for problems, it will not go away, as libev
2458I<has> to modify the signal mask, at least temporarily.
2459
2460So I can't stress this enough: I<If you do not reset your signal mask when
2461you expect it to be empty, you have a race condition in your code>. This
2462is not a libev-specific thing, this is true for most event libraries.
2463
2464=head3 The special problem of threads signal handling
2465
2466POSIX threads has problematic signal handling semantics, specifically,
2467a lot of functionality (sigfd, sigwait etc.) only really works if all
2468threads in a process block signals, which is hard to achieve.
2469
2470When you want to use sigwait (or mix libev signal handling with your own
2471for the same signals), you can tackle this problem by globally blocking
2472all signals before creating any threads (or creating them with a fully set
2473sigprocmask) and also specifying the C<EVFLAG_NOSIGMASK> when creating
2474loops. Then designate one thread as "signal receiver thread" which handles
2475these signals. You can pass on any signals that libev might be interested
2476in by calling C<ev_feed_signal>.
2477
2478=head3 Watcher-Specific Functions and Data Members
2479
2480=over 4
2481
2482=item ev_signal_init (ev_signal *, callback, int signum)
2483
2484=item ev_signal_set (ev_signal *, int signum)
2485
2486Configures the watcher to trigger on the given signal number (usually one
2487of the C<SIGxxx> constants).
2488
2489=item int signum [read-only]
2490
2491The signal the watcher watches out for.
2492
2493=back
2494
2495=head3 Examples
2496
2497Example: Try to exit cleanly on SIGINT.
2498
2499   static void
2500   sigint_cb (struct ev_loop *loop, ev_signal *w, int revents)
2501   {
2502     ev_break (loop, EVBREAK_ALL);
2503   }
2504
2505   ev_signal signal_watcher;
2506   ev_signal_init (&signal_watcher, sigint_cb, SIGINT);
2507   ev_signal_start (loop, &signal_watcher);
2508
2509
2510=head2 C<ev_child> - watch out for process status changes
2511
2512Child watchers trigger when your process receives a SIGCHLD in response to
2513some child status changes (most typically when a child of yours dies or
2514exits). It is permissible to install a child watcher I<after> the child
2515has been forked (which implies it might have already exited), as long
2516as the event loop isn't entered (or is continued from a watcher), i.e.,
2517forking and then immediately registering a watcher for the child is fine,
2518but forking and registering a watcher a few event loop iterations later or
2519in the next callback invocation is not.
2520
2521Only the default event loop is capable of handling signals, and therefore
2522you can only register child watchers in the default event loop.
2523
2524Due to some design glitches inside libev, child watchers will always be
2525handled at maximum priority (their priority is set to C<EV_MAXPRI> by
2526libev)
2527
2528=head3 Process Interaction
2529
2530Libev grabs C<SIGCHLD> as soon as the default event loop is
2531initialised. This is necessary to guarantee proper behaviour even if the
2532first child watcher is started after the child exits. The occurrence
2533of C<SIGCHLD> is recorded asynchronously, but child reaping is done
2534synchronously as part of the event loop processing. Libev always reaps all
2535children, even ones not watched.
2536
2537=head3 Overriding the Built-In Processing
2538
2539Libev offers no special support for overriding the built-in child
2540processing, but if your application collides with libev's default child
2541handler, you can override it easily by installing your own handler for
2542C<SIGCHLD> after initialising the default loop, and making sure the
2543default loop never gets destroyed. You are encouraged, however, to use an
2544event-based approach to child reaping and thus use libev's support for
2545that, so other libev users can use C<ev_child> watchers freely.
2546
2547=head3 Stopping the Child Watcher
2548
2549Currently, the child watcher never gets stopped, even when the
2550child terminates, so normally one needs to stop the watcher in the
2551callback. Future versions of libev might stop the watcher automatically
2552when a child exit is detected (calling C<ev_child_stop> twice is not a
2553problem).
2554
2555=head3 Watcher-Specific Functions and Data Members
2556
2557=over 4
2558
2559=item ev_child_init (ev_child *, callback, int pid, int trace)
2560
2561=item ev_child_set (ev_child *, int pid, int trace)
2562
2563Configures the watcher to wait for status changes of process C<pid> (or
2564I<any> process if C<pid> is specified as C<0>). The callback can look
2565at the C<rstatus> member of the C<ev_child> watcher structure to see
2566the status word (use the macros from C<sys/wait.h> and see your systems
2567C<waitpid> documentation). The C<rpid> member contains the pid of the
2568process causing the status change. C<trace> must be either C<0> (only
2569activate the watcher when the process terminates) or C<1> (additionally
2570activate the watcher when the process is stopped or continued).
2571
2572=item int pid [read-only]
2573
2574The process id this watcher watches out for, or C<0>, meaning any process id.
2575
2576=item int rpid [read-write]
2577
2578The process id that detected a status change.
2579
2580=item int rstatus [read-write]
2581
2582The process exit/trace status caused by C<rpid> (see your systems
2583C<waitpid> and C<sys/wait.h> documentation for details).
2584
2585=back
2586
2587=head3 Examples
2588
2589Example: C<fork()> a new process and install a child handler to wait for
2590its completion.
2591
2592   ev_child cw;
2593
2594   static void
2595   child_cb (EV_P_ ev_child *w, int revents)
2596   {
2597     ev_child_stop (EV_A_ w);
2598     printf ("process %d exited with status %x\n", w->rpid, w->rstatus);
2599   }
2600
2601   pid_t pid = fork ();
2602
2603   if (pid < 0)
2604     // error
2605   else if (pid == 0)
2606     {
2607       // the forked child executes here
2608       exit (1);
2609     }
2610   else
2611     {
2612       ev_child_init (&cw, child_cb, pid, 0);
2613       ev_child_start (EV_DEFAULT_ &cw);
2614     }
2615
2616
2617=head2 C<ev_stat> - did the file attributes just change?
2618
2619This watches a file system path for attribute changes. That is, it calls
2620C<stat> on that path in regular intervals (or when the OS says it changed)
2621and sees if it changed compared to the last time, invoking the callback
2622if it did. Starting the watcher C<stat>'s the file, so only changes that
2623happen after the watcher has been started will be reported.
2624
2625The path does not need to exist: changing from "path exists" to "path does
2626not exist" is a status change like any other. The condition "path does not
2627exist" (or more correctly "path cannot be stat'ed") is signified by the
2628C<st_nlink> field being zero (which is otherwise always forced to be at
2629least one) and all the other fields of the stat buffer having unspecified
2630contents.
2631
2632The path I<must not> end in a slash or contain special components such as
2633C<.> or C<..>. The path I<should> be absolute: If it is relative and
2634your working directory changes, then the behaviour is undefined.
2635
2636Since there is no portable change notification interface available, the
2637portable implementation simply calls C<stat(2)> regularly on the path
2638to see if it changed somehow. You can specify a recommended polling
2639interval for this case. If you specify a polling interval of C<0> (highly
2640recommended!) then a I<suitable, unspecified default> value will be used
2641(which you can expect to be around five seconds, although this might
2642change dynamically). Libev will also impose a minimum interval which is
2643currently around C<0.1>, but that's usually overkill.
2644
2645This watcher type is not meant for massive numbers of stat watchers,
2646as even with OS-supported change notifications, this can be
2647resource-intensive.
2648
2649At the time of this writing, the only OS-specific interface implemented
2650is the Linux inotify interface (implementing kqueue support is left as an
2651exercise for the reader. Note, however, that the author sees no way of
2652implementing C<ev_stat> semantics with kqueue, except as a hint).
2653
2654=head3 ABI Issues (Largefile Support)
2655
2656Libev by default (unless the user overrides this) uses the default
2657compilation environment, which means that on systems with large file
2658support disabled by default, you get the 32 bit version of the stat
2659structure. When using the library from programs that change the ABI to
2660use 64 bit file offsets the programs will fail. In that case you have to
2661compile libev with the same flags to get binary compatibility. This is
2662obviously the case with any flags that change the ABI, but the problem is
2663most noticeably displayed with ev_stat and large file support.
2664
2665The solution for this is to lobby your distribution maker to make large
2666file interfaces available by default (as e.g. FreeBSD does) and not
2667optional. Libev cannot simply switch on large file support because it has
2668to exchange stat structures with application programs compiled using the
2669default compilation environment.
2670
2671=head3 Inotify and Kqueue
2672
2673When C<inotify (7)> support has been compiled into libev and present at
2674runtime, it will be used to speed up change detection where possible. The
2675inotify descriptor will be created lazily when the first C<ev_stat>
2676watcher is being started.
2677
2678Inotify presence does not change the semantics of C<ev_stat> watchers
2679except that changes might be detected earlier, and in some cases, to avoid
2680making regular C<stat> calls. Even in the presence of inotify support
2681there are many cases where libev has to resort to regular C<stat> polling,
2682but as long as kernel 2.6.25 or newer is used (2.6.24 and older have too
2683many bugs), the path exists (i.e. stat succeeds), and the path resides on
2684a local filesystem (libev currently assumes only ext2/3, jfs, reiserfs and
2685xfs are fully working) libev usually gets away without polling.
2686
2687There is no support for kqueue, as apparently it cannot be used to
2688implement this functionality, due to the requirement of having a file
2689descriptor open on the object at all times, and detecting renames, unlinks
2690etc. is difficult.
2691
2692=head3 C<stat ()> is a synchronous operation
2693
2694Libev doesn't normally do any kind of I/O itself, and so is not blocking
2695the process. The exception are C<ev_stat> watchers - those call C<stat
2696()>, which is a synchronous operation.
2697
2698For local paths, this usually doesn't matter: unless the system is very
2699busy or the intervals between stat's are large, a stat call will be fast,
2700as the path data is usually in memory already (except when starting the
2701watcher).
2702
2703For networked file systems, calling C<stat ()> can block an indefinite
2704time due to network issues, and even under good conditions, a stat call
2705often takes multiple milliseconds.
2706
2707Therefore, it is best to avoid using C<ev_stat> watchers on networked
2708paths, although this is fully supported by libev.
2709
2710=head3 The special problem of stat time resolution
2711
2712The C<stat ()> system call only supports full-second resolution portably,
2713and even on systems where the resolution is higher, most file systems
2714still only support whole seconds.
2715
2716That means that, if the time is the only thing that changes, you can
2717easily miss updates: on the first update, C<ev_stat> detects a change and
2718calls your callback, which does something. When there is another update
2719within the same second, C<ev_stat> will be unable to detect unless the
2720stat data does change in other ways (e.g. file size).
2721
2722The solution to this is to delay acting on a change for slightly more
2723than a second (or till slightly after the next full second boundary), using
2724a roughly one-second-delay C<ev_timer> (e.g. C<ev_timer_set (w, 0., 1.02);
2725ev_timer_again (loop, w)>).
2726
2727The C<.02> offset is added to work around small timing inconsistencies
2728of some operating systems (where the second counter of the current time
2729might be be delayed. One such system is the Linux kernel, where a call to
2730C<gettimeofday> might return a timestamp with a full second later than
2731a subsequent C<time> call - if the equivalent of C<time ()> is used to
2732update file times then there will be a small window where the kernel uses
2733the previous second to update file times but libev might already execute
2734the timer callback).
2735
2736=head3 Watcher-Specific Functions and Data Members
2737
2738=over 4
2739
2740=item ev_stat_init (ev_stat *, callback, const char *path, ev_tstamp interval)
2741
2742=item ev_stat_set (ev_stat *, const char *path, ev_tstamp interval)
2743
2744Configures the watcher to wait for status changes of the given
2745C<path>. The C<interval> is a hint on how quickly a change is expected to
2746be detected and should normally be specified as C<0> to let libev choose
2747a suitable value. The memory pointed to by C<path> must point to the same
2748path for as long as the watcher is active.
2749
2750The callback will receive an C<EV_STAT> event when a change was detected,
2751relative to the attributes at the time the watcher was started (or the
2752last change was detected).
2753
2754=item ev_stat_stat (loop, ev_stat *)
2755
2756Updates the stat buffer immediately with new values. If you change the
2757watched path in your callback, you could call this function to avoid
2758detecting this change (while introducing a race condition if you are not
2759the only one changing the path). Can also be useful simply to find out the
2760new values.
2761
2762=item ev_statdata attr [read-only]
2763
2764The most-recently detected attributes of the file. Although the type is
2765C<ev_statdata>, this is usually the (or one of the) C<struct stat> types
2766suitable for your system, but you can only rely on the POSIX-standardised
2767members to be present. If the C<st_nlink> member is C<0>, then there was
2768some error while C<stat>ing the file.
2769
2770=item ev_statdata prev [read-only]
2771
2772The previous attributes of the file. The callback gets invoked whenever
2773C<prev> != C<attr>, or, more precisely, one or more of these members
2774differ: C<st_dev>, C<st_ino>, C<st_mode>, C<st_nlink>, C<st_uid>,
2775C<st_gid>, C<st_rdev>, C<st_size>, C<st_atime>, C<st_mtime>, C<st_ctime>.
2776
2777=item ev_tstamp interval [read-only]
2778
2779The specified interval.
2780
2781=item const char *path [read-only]
2782
2783The file system path that is being watched.
2784
2785=back
2786
2787=head3 Examples
2788
2789Example: Watch C</etc/passwd> for attribute changes.
2790
2791   static void
2792   passwd_cb (struct ev_loop *loop, ev_stat *w, int revents)
2793   {
2794     /* /etc/passwd changed in some way */
2795     if (w->attr.st_nlink)
2796       {
2797         printf ("passwd current size  %ld\n", (long)w->attr.st_size);
2798         printf ("passwd current atime %ld\n", (long)w->attr.st_mtime);
2799         printf ("passwd current mtime %ld\n", (long)w->attr.st_mtime);
2800       }
2801     else
2802       /* you shalt not abuse printf for puts */
2803       puts ("wow, /etc/passwd is not there, expect problems. "
2804             "if this is windows, they already arrived\n");
2805   }
2806
2807   ...
2808   ev_stat passwd;
2809
2810   ev_stat_init (&passwd, passwd_cb, "/etc/passwd", 0.);
2811   ev_stat_start (loop, &passwd);
2812
2813Example: Like above, but additionally use a one-second delay so we do not
2814miss updates (however, frequent updates will delay processing, too, so
2815one might do the work both on C<ev_stat> callback invocation I<and> on
2816C<ev_timer> callback invocation).
2817
2818   static ev_stat passwd;
2819   static ev_timer timer;
2820
2821   static void
2822   timer_cb (EV_P_ ev_timer *w, int revents)
2823   {
2824     ev_timer_stop (EV_A_ w);
2825
2826     /* now it's one second after the most recent passwd change */
2827   }
2828
2829   static void
2830   stat_cb (EV_P_ ev_stat *w, int revents)
2831   {
2832     /* reset the one-second timer */
2833     ev_timer_again (EV_A_ &timer);
2834   }
2835
2836   ...
2837   ev_stat_init (&passwd, stat_cb, "/etc/passwd", 0.);
2838   ev_stat_start (loop, &passwd);
2839   ev_timer_init (&timer, timer_cb, 0., 1.02);
2840
2841
2842=head2 C<ev_idle> - when you've got nothing better to do...
2843
2844Idle watchers trigger events when no other events of the same or higher
2845priority are pending (prepare, check and other idle watchers do not count
2846as receiving "events").
2847
2848That is, as long as your process is busy handling sockets or timeouts
2849(or even signals, imagine) of the same or higher priority it will not be
2850triggered. But when your process is idle (or only lower-priority watchers
2851are pending), the idle watchers are being called once per event loop
2852iteration - until stopped, that is, or your process receives more events
2853and becomes busy again with higher priority stuff.
2854
2855The most noteworthy effect is that as long as any idle watchers are
2856active, the process will not block when waiting for new events.
2857
2858Apart from keeping your process non-blocking (which is a useful
2859effect on its own sometimes), idle watchers are a good place to do
2860"pseudo-background processing", or delay processing stuff to after the
2861event loop has handled all outstanding events.
2862
2863=head3 Abusing an C<ev_idle> watcher for its side-effect
2864
2865As long as there is at least one active idle watcher, libev will never
2866sleep unnecessarily. Or in other words, it will loop as fast as possible.
2867For this to work, the idle watcher doesn't need to be invoked at all - the
2868lowest priority will do.
2869
2870This mode of operation can be useful together with an C<ev_check> watcher,
2871to do something on each event loop iteration - for example to balance load
2872between different connections.
2873
2874See L</Abusing an ev_check watcher for its side-effect> for a longer
2875example.
2876
2877=head3 Watcher-Specific Functions and Data Members
2878
2879=over 4
2880
2881=item ev_idle_init (ev_idle *, callback)
2882
2883Initialises and configures the idle watcher - it has no parameters of any
2884kind. There is a C<ev_idle_set> macro, but using it is utterly pointless,
2885believe me.
2886
2887=back
2888
2889=head3 Examples
2890
2891Example: Dynamically allocate an C<ev_idle> watcher, start it, and in the
2892callback, free it. Also, use no error checking, as usual.
2893
2894   static void
2895   idle_cb (struct ev_loop *loop, ev_idle *w, int revents)
2896   {
2897     // stop the watcher
2898     ev_idle_stop (loop, w);
2899
2900     // now we can free it
2901     free (w);
2902
2903     // now do something you wanted to do when the program has
2904     // no longer anything immediate to do.
2905   }
2906
2907   ev_idle *idle_watcher = malloc (sizeof (ev_idle));
2908   ev_idle_init (idle_watcher, idle_cb);
2909   ev_idle_start (loop, idle_watcher);
2910
2911
2912=head2 C<ev_prepare> and C<ev_check> - customise your event loop!
2913
2914Prepare and check watchers are often (but not always) used in pairs:
2915prepare watchers get invoked before the process blocks and check watchers
2916afterwards.
2917
2918You I<must not> call C<ev_run> (or similar functions that enter the
2919current event loop) or C<ev_loop_fork> from either C<ev_prepare> or
2920C<ev_check> watchers. Other loops than the current one are fine,
2921however. The rationale behind this is that you do not need to check
2922for recursion in those watchers, i.e. the sequence will always be
2923C<ev_prepare>, blocking, C<ev_check> so if you have one watcher of each
2924kind they will always be called in pairs bracketing the blocking call.
2925
2926Their main purpose is to integrate other event mechanisms into libev and
2927their use is somewhat advanced. They could be used, for example, to track
2928variable changes, implement your own watchers, integrate net-snmp or a
2929coroutine library and lots more. They are also occasionally useful if
2930you cache some data and want to flush it before blocking (for example,
2931in X programs you might want to do an C<XFlush ()> in an C<ev_prepare>
2932watcher).
2933
2934This is done by examining in each prepare call which file descriptors
2935need to be watched by the other library, registering C<ev_io> watchers
2936for them and starting an C<ev_timer> watcher for any timeouts (many
2937libraries provide exactly this functionality). Then, in the check watcher,
2938you check for any events that occurred (by checking the pending status
2939of all watchers and stopping them) and call back into the library. The
2940I/O and timer callbacks will never actually be called (but must be valid
2941nevertheless, because you never know, you know?).
2942
2943As another example, the Perl Coro module uses these hooks to integrate
2944coroutines into libev programs, by yielding to other active coroutines
2945during each prepare and only letting the process block if no coroutines
2946are ready to run (it's actually more complicated: it only runs coroutines
2947with priority higher than or equal to the event loop and one coroutine
2948of lower priority, but only once, using idle watchers to keep the event
2949loop from blocking if lower-priority coroutines are active, thus mapping
2950low-priority coroutines to idle/background tasks).
2951
2952When used for this purpose, it is recommended to give C<ev_check> watchers
2953highest (C<EV_MAXPRI>) priority, to ensure that they are being run before
2954any other watchers after the poll (this doesn't matter for C<ev_prepare>
2955watchers).
2956
2957Also, C<ev_check> watchers (and C<ev_prepare> watchers, too) should not
2958activate ("feed") events into libev. While libev fully supports this, they
2959might get executed before other C<ev_check> watchers did their job. As
2960C<ev_check> watchers are often used to embed other (non-libev) event
2961loops those other event loops might be in an unusable state until their
2962C<ev_check> watcher ran (always remind yourself to coexist peacefully with
2963others).
2964
2965=head3 Abusing an C<ev_check> watcher for its side-effect
2966
2967C<ev_check> (and less often also C<ev_prepare>) watchers can also be
2968useful because they are called once per event loop iteration. For
2969example, if you want to handle a large number of connections fairly, you
2970normally only do a bit of work for each active connection, and if there
2971is more work to do, you wait for the next event loop iteration, so other
2972connections have a chance of making progress.
2973
2974Using an C<ev_check> watcher is almost enough: it will be called on the
2975next event loop iteration. However, that isn't as soon as possible -
2976without external events, your C<ev_check> watcher will not be invoked.
2977
2978This is where C<ev_idle> watchers come in handy - all you need is a
2979single global idle watcher that is active as long as you have one active
2980C<ev_check> watcher. The C<ev_idle> watcher makes sure the event loop
2981will not sleep, and the C<ev_check> watcher makes sure a callback gets
2982invoked. Neither watcher alone can do that.
2983
2984=head3 Watcher-Specific Functions and Data Members
2985
2986=over 4
2987
2988=item ev_prepare_init (ev_prepare *, callback)
2989
2990=item ev_check_init (ev_check *, callback)
2991
2992Initialises and configures the prepare or check watcher - they have no
2993parameters of any kind. There are C<ev_prepare_set> and C<ev_check_set>
2994macros, but using them is utterly, utterly, utterly and completely
2995pointless.
2996
2997=back
2998
2999=head3 Examples
3000
3001There are a number of principal ways to embed other event loops or modules
3002into libev. Here are some ideas on how to include libadns into libev
3003(there is a Perl module named C<EV::ADNS> that does this, which you could
3004use as a working example. Another Perl module named C<EV::Glib> embeds a
3005Glib main context into libev, and finally, C<Glib::EV> embeds EV into the
3006Glib event loop).
3007
3008Method 1: Add IO watchers and a timeout watcher in a prepare handler,
3009and in a check watcher, destroy them and call into libadns. What follows
3010is pseudo-code only of course. This requires you to either use a low
3011priority for the check watcher or use C<ev_clear_pending> explicitly, as
3012the callbacks for the IO/timeout watchers might not have been called yet.
3013
3014   static ev_io iow [nfd];
3015   static ev_timer tw;
3016
3017   static void
3018   io_cb (struct ev_loop *loop, ev_io *w, int revents)
3019   {
3020   }
3021
3022   // create io watchers for each fd and a timer before blocking
3023   static void
3024   adns_prepare_cb (struct ev_loop *loop, ev_prepare *w, int revents)
3025   {
3026     int timeout = 3600000;
3027     struct pollfd fds [nfd];
3028     // actual code will need to loop here and realloc etc.
3029     adns_beforepoll (ads, fds, &nfd, &timeout, timeval_from (ev_time ()));
3030
3031     /* the callback is illegal, but won't be called as we stop during check */
3032     ev_timer_init (&tw, 0, timeout * 1e-3, 0.);
3033     ev_timer_start (loop, &tw);
3034
3035     // create one ev_io per pollfd
3036     for (int i = 0; i < nfd; ++i)
3037       {
3038         ev_io_init (iow + i, io_cb, fds [i].fd,
3039           ((fds [i].events & POLLIN ? EV_READ : 0)
3040            | (fds [i].events & POLLOUT ? EV_WRITE : 0)));
3041
3042         fds [i].revents = 0;
3043         ev_io_start (loop, iow + i);
3044       }
3045   }
3046
3047   // stop all watchers after blocking
3048   static void
3049   adns_check_cb (struct ev_loop *loop, ev_check *w, int revents)
3050   {
3051     ev_timer_stop (loop, &tw);
3052
3053     for (int i = 0; i < nfd; ++i)
3054       {
3055         // set the relevant poll flags
3056         // could also call adns_processreadable etc. here
3057         struct pollfd *fd = fds + i;
3058         int revents = ev_clear_pending (iow + i);
3059         if (revents & EV_READ ) fd->revents |= fd->events & POLLIN;
3060         if (revents & EV_WRITE) fd->revents |= fd->events & POLLOUT;
3061
3062         // now stop the watcher
3063         ev_io_stop (loop, iow + i);
3064       }
3065
3066     adns_afterpoll (adns, fds, nfd, timeval_from (ev_now (loop));
3067   }
3068
3069Method 2: This would be just like method 1, but you run C<adns_afterpoll>
3070in the prepare watcher and would dispose of the check watcher.
3071
3072Method 3: If the module to be embedded supports explicit event
3073notification (libadns does), you can also make use of the actual watcher
3074callbacks, and only destroy/create the watchers in the prepare watcher.
3075
3076   static void
3077   timer_cb (EV_P_ ev_timer *w, int revents)
3078   {
3079     adns_state ads = (adns_state)w->data;
3080     update_now (EV_A);
3081
3082     adns_processtimeouts (ads, &tv_now);
3083   }
3084
3085   static void
3086   io_cb (EV_P_ ev_io *w, int revents)
3087   {
3088     adns_state ads = (adns_state)w->data;
3089     update_now (EV_A);
3090
3091     if (revents & EV_READ ) adns_processreadable  (ads, w->fd, &tv_now);
3092     if (revents & EV_WRITE) adns_processwriteable (ads, w->fd, &tv_now);
3093   }
3094
3095   // do not ever call adns_afterpoll
3096
3097Method 4: Do not use a prepare or check watcher because the module you
3098want to embed is not flexible enough to support it. Instead, you can
3099override their poll function. The drawback with this solution is that the
3100main loop is now no longer controllable by EV. The C<Glib::EV> module uses
3101this approach, effectively embedding EV as a client into the horrible
3102libglib event loop.
3103
3104   static gint
3105   event_poll_func (GPollFD *fds, guint nfds, gint timeout)
3106   {
3107     int got_events = 0;
3108
3109     for (n = 0; n < nfds; ++n)
3110       // create/start io watcher that sets the relevant bits in fds[n] and increment got_events
3111
3112     if (timeout >= 0)
3113       // create/start timer
3114
3115     // poll
3116     ev_run (EV_A_ 0);
3117
3118     // stop timer again
3119     if (timeout >= 0)
3120       ev_timer_stop (EV_A_ &to);
3121
3122     // stop io watchers again - their callbacks should have set
3123     for (n = 0; n < nfds; ++n)
3124       ev_io_stop (EV_A_ iow [n]);
3125
3126     return got_events;
3127   }
3128
3129
3130=head2 C<ev_embed> - when one backend isn't enough...
3131
3132This is a rather advanced watcher type that lets you embed one event loop
3133into another (currently only C<ev_io> events are supported in the embedded
3134loop, other types of watchers might be handled in a delayed or incorrect
3135fashion and must not be used).
3136
3137There are primarily two reasons you would want that: work around bugs and
3138prioritise I/O.
3139
3140As an example for a bug workaround, the kqueue backend might only support
3141sockets on some platform, so it is unusable as generic backend, but you
3142still want to make use of it because you have many sockets and it scales
3143so nicely. In this case, you would create a kqueue-based loop and embed
3144it into your default loop (which might use e.g. poll). Overall operation
3145will be a bit slower because first libev has to call C<poll> and then
3146C<kevent>, but at least you can use both mechanisms for what they are
3147best: C<kqueue> for scalable sockets and C<poll> if you want it to work :)
3148
3149As for prioritising I/O: under rare circumstances you have the case where
3150some fds have to be watched and handled very quickly (with low latency),
3151and even priorities and idle watchers might have too much overhead. In
3152this case you would put all the high priority stuff in one loop and all
3153the rest in a second one, and embed the second one in the first.
3154
3155As long as the watcher is active, the callback will be invoked every
3156time there might be events pending in the embedded loop. The callback
3157must then call C<ev_embed_sweep (mainloop, watcher)> to make a single
3158sweep and invoke their callbacks (the callback doesn't need to invoke the
3159C<ev_embed_sweep> function directly, it could also start an idle watcher
3160to give the embedded loop strictly lower priority for example).
3161
3162You can also set the callback to C<0>, in which case the embed watcher
3163will automatically execute the embedded loop sweep whenever necessary.
3164
3165Fork detection will be handled transparently while the C<ev_embed> watcher
3166is active, i.e., the embedded loop will automatically be forked when the
3167embedding loop forks. In other cases, the user is responsible for calling
3168C<ev_loop_fork> on the embedded loop.
3169
3170Unfortunately, not all backends are embeddable: only the ones returned by
3171C<ev_embeddable_backends> are, which, unfortunately, does not include any
3172portable one.
3173
3174So when you want to use this feature you will always have to be prepared
3175that you cannot get an embeddable loop. The recommended way to get around
3176this is to have a separate variables for your embeddable loop, try to
3177create it, and if that fails, use the normal loop for everything.
3178
3179=head3 C<ev_embed> and fork
3180
3181While the C<ev_embed> watcher is running, forks in the embedding loop will
3182automatically be applied to the embedded loop as well, so no special
3183fork handling is required in that case. When the watcher is not running,
3184however, it is still the task of the libev user to call C<ev_loop_fork ()>
3185as applicable.
3186
3187=head3 Watcher-Specific Functions and Data Members
3188
3189=over 4
3190
3191=item ev_embed_init (ev_embed *, callback, struct ev_loop *embedded_loop)
3192
3193=item ev_embed_set (ev_embed *, struct ev_loop *embedded_loop)
3194
3195Configures the watcher to embed the given loop, which must be
3196embeddable. If the callback is C<0>, then C<ev_embed_sweep> will be
3197invoked automatically, otherwise it is the responsibility of the callback
3198to invoke it (it will continue to be called until the sweep has been done,
3199if you do not want that, you need to temporarily stop the embed watcher).
3200
3201=item ev_embed_sweep (loop, ev_embed *)
3202
3203Make a single, non-blocking sweep over the embedded loop. This works
3204similarly to C<ev_run (embedded_loop, EVRUN_NOWAIT)>, but in the most
3205appropriate way for embedded loops.
3206
3207=item struct ev_loop *other [read-only]
3208
3209The embedded event loop.
3210
3211=back
3212
3213=head3 Examples
3214
3215Example: Try to get an embeddable event loop and embed it into the default
3216event loop. If that is not possible, use the default loop. The default
3217loop is stored in C<loop_hi>, while the embeddable loop is stored in
3218C<loop_lo> (which is C<loop_hi> in the case no embeddable loop can be
3219used).
3220
3221   struct ev_loop *loop_hi = ev_default_init (0);
3222   struct ev_loop *loop_lo = 0;
3223   ev_embed embed;
3224
3225   // see if there is a chance of getting one that works
3226   // (remember that a flags value of 0 means autodetection)
3227   loop_lo = ev_embeddable_backends () & ev_recommended_backends ()
3228     ? ev_loop_new (ev_embeddable_backends () & ev_recommended_backends ())
3229     : 0;
3230
3231   // if we got one, then embed it, otherwise default to loop_hi
3232   if (loop_lo)
3233     {
3234       ev_embed_init (&embed, 0, loop_lo);
3235       ev_embed_start (loop_hi, &embed);
3236     }
3237   else
3238     loop_lo = loop_hi;
3239
3240Example: Check if kqueue is available but not recommended and create
3241a kqueue backend for use with sockets (which usually work with any
3242kqueue implementation). Store the kqueue/socket-only event loop in
3243C<loop_socket>. (One might optionally use C<EVFLAG_NOENV>, too).
3244
3245   struct ev_loop *loop = ev_default_init (0);
3246   struct ev_loop *loop_socket = 0;
3247   ev_embed embed;
3248
3249   if (ev_supported_backends () & ~ev_recommended_backends () & EVBACKEND_KQUEUE)
3250     if ((loop_socket = ev_loop_new (EVBACKEND_KQUEUE))
3251       {
3252         ev_embed_init (&embed, 0, loop_socket);
3253         ev_embed_start (loop, &embed);
3254       }
3255
3256   if (!loop_socket)
3257     loop_socket = loop;
3258
3259   // now use loop_socket for all sockets, and loop for everything else
3260
3261
3262=head2 C<ev_fork> - the audacity to resume the event loop after a fork
3263
3264Fork watchers are called when a C<fork ()> was detected (usually because
3265whoever is a good citizen cared to tell libev about it by calling
3266C<ev_loop_fork>). The invocation is done before the event loop blocks next
3267and before C<ev_check> watchers are being called, and only in the child
3268after the fork. If whoever good citizen calling C<ev_default_fork> cheats
3269and calls it in the wrong process, the fork handlers will be invoked, too,
3270of course.
3271
3272=head3 The special problem of life after fork - how is it possible?
3273
3274Most uses of C<fork ()> consist of forking, then some simple calls to set
3275up/change the process environment, followed by a call to C<exec()>. This
3276sequence should be handled by libev without any problems.
3277
3278This changes when the application actually wants to do event handling
3279in the child, or both parent in child, in effect "continuing" after the
3280fork.
3281
3282The default mode of operation (for libev, with application help to detect
3283forks) is to duplicate all the state in the child, as would be expected
3284when I<either> the parent I<or> the child process continues.
3285
3286When both processes want to continue using libev, then this is usually the
3287wrong result. In that case, usually one process (typically the parent) is
3288supposed to continue with all watchers in place as before, while the other
3289process typically wants to start fresh, i.e. without any active watchers.
3290
3291The cleanest and most efficient way to achieve that with libev is to
3292simply create a new event loop, which of course will be "empty", and
3293use that for new watchers. This has the advantage of not touching more
3294memory than necessary, and thus avoiding the copy-on-write, and the
3295disadvantage of having to use multiple event loops (which do not support
3296signal watchers).
3297
3298When this is not possible, or you want to use the default loop for
3299other reasons, then in the process that wants to start "fresh", call
3300C<ev_loop_destroy (EV_DEFAULT)> followed by C<ev_default_loop (...)>.
3301Destroying the default loop will "orphan" (not stop) all registered
3302watchers, so you have to be careful not to execute code that modifies
3303those watchers. Note also that in that case, you have to re-register any
3304signal watchers.
3305
3306=head3 Watcher-Specific Functions and Data Members
3307
3308=over 4
3309
3310=item ev_fork_init (ev_fork *, callback)
3311
3312Initialises and configures the fork watcher - it has no parameters of any
3313kind. There is a C<ev_fork_set> macro, but using it is utterly pointless,
3314really.
3315
3316=back
3317
3318
3319=head2 C<ev_cleanup> - even the best things end
3320
3321Cleanup watchers are called just before the event loop is being destroyed
3322by a call to C<ev_loop_destroy>.
3323
3324While there is no guarantee that the event loop gets destroyed, cleanup
3325watchers provide a convenient method to install cleanup hooks for your
3326program, worker threads and so on - you just to make sure to destroy the
3327loop when you want them to be invoked.
3328
3329Cleanup watchers are invoked in the same way as any other watcher. Unlike
3330all other watchers, they do not keep a reference to the event loop (which
3331makes a lot of sense if you think about it). Like all other watchers, you
3332can call libev functions in the callback, except C<ev_cleanup_start>.
3333
3334=head3 Watcher-Specific Functions and Data Members
3335
3336=over 4
3337
3338=item ev_cleanup_init (ev_cleanup *, callback)
3339
3340Initialises and configures the cleanup watcher - it has no parameters of
3341any kind. There is a C<ev_cleanup_set> macro, but using it is utterly
3342pointless, I assure you.
3343
3344=back
3345
3346Example: Register an atexit handler to destroy the default loop, so any
3347cleanup functions are called.
3348
3349   static void
3350   program_exits (void)
3351   {
3352     ev_loop_destroy (EV_DEFAULT_UC);
3353   }
3354
3355   ...
3356   atexit (program_exits);
3357
3358
3359=head2 C<ev_async> - how to wake up an event loop
3360
3361In general, you cannot use an C<ev_loop> from multiple threads or other
3362asynchronous sources such as signal handlers (as opposed to multiple event
3363loops - those are of course safe to use in different threads).
3364
3365Sometimes, however, you need to wake up an event loop you do not control,
3366for example because it belongs to another thread. This is what C<ev_async>
3367watchers do: as long as the C<ev_async> watcher is active, you can signal
3368it by calling C<ev_async_send>, which is thread- and signal safe.
3369
3370This functionality is very similar to C<ev_signal> watchers, as signals,
3371too, are asynchronous in nature, and signals, too, will be compressed
3372(i.e. the number of callback invocations may be less than the number of
3373C<ev_async_send> calls). In fact, you could use signal watchers as a kind
3374of "global async watchers" by using a watcher on an otherwise unused
3375signal, and C<ev_feed_signal> to signal this watcher from another thread,
3376even without knowing which loop owns the signal.
3377
3378=head3 Queueing
3379
3380C<ev_async> does not support queueing of data in any way. The reason
3381is that the author does not know of a simple (or any) algorithm for a
3382multiple-writer-single-reader queue that works in all cases and doesn't
3383need elaborate support such as pthreads or unportable memory access
3384semantics.
3385
3386That means that if you want to queue data, you have to provide your own
3387queue. But at least I can tell you how to implement locking around your
3388queue:
3389
3390=over 4
3391
3392=item queueing from a signal handler context
3393
3394To implement race-free queueing, you simply add to the queue in the signal
3395handler but you block the signal handler in the watcher callback. Here is
3396an example that does that for some fictitious SIGUSR1 handler:
3397
3398   static ev_async mysig;
3399
3400   static void
3401   sigusr1_handler (void)
3402   {
3403     sometype data;
3404
3405     // no locking etc.
3406     queue_put (data);
3407     ev_async_send (EV_DEFAULT_ &mysig);
3408   }
3409
3410   static void
3411   mysig_cb (EV_P_ ev_async *w, int revents)
3412   {
3413     sometype data;
3414     sigset_t block, prev;
3415
3416     sigemptyset (&block);
3417     sigaddset (&block, SIGUSR1);
3418     sigprocmask (SIG_BLOCK, &block, &prev);
3419
3420     while (queue_get (&data))
3421       process (data);
3422
3423     if (sigismember (&prev, SIGUSR1)
3424       sigprocmask (SIG_UNBLOCK, &block, 0);
3425   }
3426
3427(Note: pthreads in theory requires you to use C<pthread_setmask>
3428instead of C<sigprocmask> when you use threads, but libev doesn't do it
3429either...).
3430
3431=item queueing from a thread context
3432
3433The strategy for threads is different, as you cannot (easily) block
3434threads but you can easily preempt them, so to queue safely you need to
3435employ a traditional mutex lock, such as in this pthread example:
3436
3437   static ev_async mysig;
3438   static pthread_mutex_t mymutex = PTHREAD_MUTEX_INITIALIZER;
3439
3440   static void
3441   otherthread (void)
3442   {
3443     // only need to lock the actual queueing operation
3444     pthread_mutex_lock (&mymutex);
3445     queue_put (data);
3446     pthread_mutex_unlock (&mymutex);
3447
3448     ev_async_send (EV_DEFAULT_ &mysig);
3449   }
3450
3451   static void
3452   mysig_cb (EV_P_ ev_async *w, int revents)
3453   {
3454     pthread_mutex_lock (&mymutex);
3455
3456     while (queue_get (&data))
3457       process (data);
3458
3459     pthread_mutex_unlock (&mymutex);
3460   }
3461
3462=back
3463
3464
3465=head3 Watcher-Specific Functions and Data Members
3466
3467=over 4
3468
3469=item ev_async_init (ev_async *, callback)
3470
3471Initialises and configures the async watcher - it has no parameters of any
3472kind. There is a C<ev_async_set> macro, but using it is utterly pointless,
3473trust me.
3474
3475=item ev_async_send (loop, ev_async *)
3476
3477Sends/signals/activates the given C<ev_async> watcher, that is, feeds
3478an C<EV_ASYNC> event on the watcher into the event loop, and instantly
3479returns.
3480
3481Unlike C<ev_feed_event>, this call is safe to do from other threads,
3482signal or similar contexts (see the discussion of C<EV_ATOMIC_T> in the
3483embedding section below on what exactly this means).
3484
3485Note that, as with other watchers in libev, multiple events might get
3486compressed into a single callback invocation (another way to look at
3487this is that C<ev_async> watchers are level-triggered: they are set on
3488C<ev_async_send>, reset when the event loop detects that).
3489
3490This call incurs the overhead of at most one extra system call per event
3491loop iteration, if the event loop is blocked, and no syscall at all if
3492the event loop (or your program) is processing events. That means that
3493repeated calls are basically free (there is no need to avoid calls for
3494performance reasons) and that the overhead becomes smaller (typically
3495zero) under load.
3496
3497=item bool = ev_async_pending (ev_async *)
3498
3499Returns a non-zero value when C<ev_async_send> has been called on the
3500watcher but the event has not yet been processed (or even noted) by the
3501event loop.
3502
3503C<ev_async_send> sets a flag in the watcher and wakes up the loop. When
3504the loop iterates next and checks for the watcher to have become active,
3505it will reset the flag again. C<ev_async_pending> can be used to very
3506quickly check whether invoking the loop might be a good idea.
3507
3508Not that this does I<not> check whether the watcher itself is pending,
3509only whether it has been requested to make this watcher pending: there
3510is a time window between the event loop checking and resetting the async
3511notification, and the callback being invoked.
3512
3513=back
3514
3515
3516=head1 OTHER FUNCTIONS
3517
3518There are some other functions of possible interest. Described. Here. Now.
3519
3520=over 4
3521
3522=item ev_once (loop, int fd, int events, ev_tstamp timeout, callback)
3523
3524This function combines a simple timer and an I/O watcher, calls your
3525callback on whichever event happens first and automatically stops both
3526watchers. This is useful if you want to wait for a single event on an fd
3527or timeout without having to allocate/configure/start/stop/free one or
3528more watchers yourself.
3529
3530If C<fd> is less than 0, then no I/O watcher will be started and the
3531C<events> argument is being ignored. Otherwise, an C<ev_io> watcher for
3532the given C<fd> and C<events> set will be created and started.
3533
3534If C<timeout> is less than 0, then no timeout watcher will be
3535started. Otherwise an C<ev_timer> watcher with after = C<timeout> (and
3536repeat = 0) will be started. C<0> is a valid timeout.
3537
3538The callback has the type C<void (*cb)(int revents, void *arg)> and is
3539passed an C<revents> set like normal event callbacks (a combination of
3540C<EV_ERROR>, C<EV_READ>, C<EV_WRITE> or C<EV_TIMER>) and the C<arg>
3541value passed to C<ev_once>. Note that it is possible to receive I<both>
3542a timeout and an io event at the same time - you probably should give io
3543events precedence.
3544
3545Example: wait up to ten seconds for data to appear on STDIN_FILENO.
3546
3547   static void stdin_ready (int revents, void *arg)
3548   {
3549     if (revents & EV_READ)
3550       /* stdin might have data for us, joy! */;
3551     else if (revents & EV_TIMER)
3552       /* doh, nothing entered */;
3553   }
3554
3555   ev_once (STDIN_FILENO, EV_READ, 10., stdin_ready, 0);
3556
3557=item ev_feed_fd_event (loop, int fd, int revents)
3558
3559Feed an event on the given fd, as if a file descriptor backend detected
3560the given events.
3561
3562=item ev_feed_signal_event (loop, int signum)
3563
3564Feed an event as if the given signal occurred. See also C<ev_feed_signal>,
3565which is async-safe.
3566
3567=back
3568
3569
3570=head1 COMMON OR USEFUL IDIOMS (OR BOTH)
3571
3572This section explains some common idioms that are not immediately
3573obvious. Note that examples are sprinkled over the whole manual, and this
3574section only contains stuff that wouldn't fit anywhere else.
3575
3576=head2 ASSOCIATING CUSTOM DATA WITH A WATCHER
3577
3578Each watcher has, by default, a C<void *data> member that you can read
3579or modify at any time: libev will completely ignore it. This can be used
3580to associate arbitrary data with your watcher. If you need more data and
3581don't want to allocate memory separately and store a pointer to it in that
3582data member, you can also "subclass" the watcher type and provide your own
3583data:
3584
3585   struct my_io
3586   {
3587     ev_io io;
3588     int otherfd;
3589     void *somedata;
3590     struct whatever *mostinteresting;
3591   };
3592
3593   ...
3594   struct my_io w;
3595   ev_io_init (&w.io, my_cb, fd, EV_READ);
3596
3597And since your callback will be called with a pointer to the watcher, you
3598can cast it back to your own type:
3599
3600   static void my_cb (struct ev_loop *loop, ev_io *w_, int revents)
3601   {
3602     struct my_io *w = (struct my_io *)w_;
3603     ...
3604   }
3605
3606More interesting and less C-conformant ways of casting your callback
3607function type instead have been omitted.
3608
3609=head2 BUILDING YOUR OWN COMPOSITE WATCHERS
3610
3611Another common scenario is to use some data structure with multiple
3612embedded watchers, in effect creating your own watcher that combines
3613multiple libev event sources into one "super-watcher":
3614
3615   struct my_biggy
3616   {
3617     int some_data;
3618     ev_timer t1;
3619     ev_timer t2;
3620   }
3621
3622In this case getting the pointer to C<my_biggy> is a bit more
3623complicated: Either you store the address of your C<my_biggy> struct in
3624the C<data> member of the watcher (for woozies or C++ coders), or you need
3625to use some pointer arithmetic using C<offsetof> inside your watchers (for
3626real programmers):
3627
3628   #include <stddef.h>
3629
3630   static void
3631   t1_cb (EV_P_ ev_timer *w, int revents)
3632   {
3633     struct my_biggy big = (struct my_biggy *)
3634       (((char *)w) - offsetof (struct my_biggy, t1));
3635   }
3636
3637   static void
3638   t2_cb (EV_P_ ev_timer *w, int revents)
3639   {
3640     struct my_biggy big = (struct my_biggy *)
3641       (((char *)w) - offsetof (struct my_biggy, t2));
3642   }
3643
3644=head2 AVOIDING FINISHING BEFORE RETURNING
3645
3646Often you have structures like this in event-based programs:
3647
3648  callback ()
3649  {
3650    free (request);
3651  }
3652
3653  request = start_new_request (..., callback);
3654
3655The intent is to start some "lengthy" operation. The C<request> could be
3656used to cancel the operation, or do other things with it.
3657
3658It's not uncommon to have code paths in C<start_new_request> that
3659immediately invoke the callback, for example, to report errors. Or you add
3660some caching layer that finds that it can skip the lengthy aspects of the
3661operation and simply invoke the callback with the result.
3662
3663The problem here is that this will happen I<before> C<start_new_request>
3664has returned, so C<request> is not set.
3665
3666Even if you pass the request by some safer means to the callback, you
3667might want to do something to the request after starting it, such as
3668canceling it, which probably isn't working so well when the callback has
3669already been invoked.
3670
3671A common way around all these issues is to make sure that
3672C<start_new_request> I<always> returns before the callback is invoked. If
3673C<start_new_request> immediately knows the result, it can artificially
3674delay invoking the callback by using a C<prepare> or C<idle> watcher for
3675example, or more sneakily, by reusing an existing (stopped) watcher and
3676pushing it into the pending queue:
3677
3678   ev_set_cb (watcher, callback);
3679   ev_feed_event (EV_A_ watcher, 0);
3680
3681This way, C<start_new_request> can safely return before the callback is
3682invoked, while not delaying callback invocation too much.
3683
3684=head2 MODEL/NESTED EVENT LOOP INVOCATIONS AND EXIT CONDITIONS
3685
3686Often (especially in GUI toolkits) there are places where you have
3687I<modal> interaction, which is most easily implemented by recursively
3688invoking C<ev_run>.
3689
3690This brings the problem of exiting - a callback might want to finish the
3691main C<ev_run> call, but not the nested one (e.g. user clicked "Quit", but
3692a modal "Are you sure?" dialog is still waiting), or just the nested one
3693and not the main one (e.g. user clocked "Ok" in a modal dialog), or some
3694other combination: In these cases, a simple C<ev_break> will not work.
3695
3696The solution is to maintain "break this loop" variable for each C<ev_run>
3697invocation, and use a loop around C<ev_run> until the condition is
3698triggered, using C<EVRUN_ONCE>:
3699
3700   // main loop
3701   int exit_main_loop = 0;
3702
3703   while (!exit_main_loop)
3704     ev_run (EV_DEFAULT_ EVRUN_ONCE);
3705
3706   // in a modal watcher
3707   int exit_nested_loop = 0;
3708
3709   while (!exit_nested_loop)
3710     ev_run (EV_A_ EVRUN_ONCE);
3711
3712To exit from any of these loops, just set the corresponding exit variable:
3713
3714   // exit modal loop
3715   exit_nested_loop = 1;
3716
3717   // exit main program, after modal loop is finished
3718   exit_main_loop = 1;
3719
3720   // exit both
3721   exit_main_loop = exit_nested_loop = 1;
3722
3723=head2 THREAD LOCKING EXAMPLE
3724
3725Here is a fictitious example of how to run an event loop in a different
3726thread from where callbacks are being invoked and watchers are
3727created/added/removed.
3728
3729For a real-world example, see the C<EV::Loop::Async> perl module,
3730which uses exactly this technique (which is suited for many high-level
3731languages).
3732
3733The example uses a pthread mutex to protect the loop data, a condition
3734variable to wait for callback invocations, an async watcher to notify the
3735event loop thread and an unspecified mechanism to wake up the main thread.
3736
3737First, you need to associate some data with the event loop:
3738
3739   typedef struct {
3740     mutex_t lock; /* global loop lock */
3741     ev_async async_w;
3742     thread_t tid;
3743     cond_t invoke_cv;
3744   } userdata;
3745
3746   void prepare_loop (EV_P)
3747   {
3748      // for simplicity, we use a static userdata struct.
3749      static userdata u;
3750
3751      ev_async_init (&u->async_w, async_cb);
3752      ev_async_start (EV_A_ &u->async_w);
3753
3754      pthread_mutex_init (&u->lock, 0);
3755      pthread_cond_init (&u->invoke_cv, 0);
3756
3757      // now associate this with the loop
3758      ev_set_userdata (EV_A_ u);
3759      ev_set_invoke_pending_cb (EV_A_ l_invoke);
3760      ev_set_loop_release_cb (EV_A_ l_release, l_acquire);
3761
3762      // then create the thread running ev_run
3763      pthread_create (&u->tid, 0, l_run, EV_A);
3764   }
3765
3766The callback for the C<ev_async> watcher does nothing: the watcher is used
3767solely to wake up the event loop so it takes notice of any new watchers
3768that might have been added:
3769
3770   static void
3771   async_cb (EV_P_ ev_async *w, int revents)
3772   {
3773      // just used for the side effects
3774   }
3775
3776The C<l_release> and C<l_acquire> callbacks simply unlock/lock the mutex
3777protecting the loop data, respectively.
3778
3779   static void
3780   l_release (EV_P)
3781   {
3782     userdata *u = ev_userdata (EV_A);
3783     pthread_mutex_unlock (&u->lock);
3784   }
3785
3786   static void
3787   l_acquire (EV_P)
3788   {
3789     userdata *u = ev_userdata (EV_A);
3790     pthread_mutex_lock (&u->lock);
3791   }
3792
3793The event loop thread first acquires the mutex, and then jumps straight
3794into C<ev_run>:
3795
3796   void *
3797   l_run (void *thr_arg)
3798   {
3799     struct ev_loop *loop = (struct ev_loop *)thr_arg;
3800
3801     l_acquire (EV_A);
3802     pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
3803     ev_run (EV_A_ 0);
3804     l_release (EV_A);
3805
3806     return 0;
3807   }
3808
3809Instead of invoking all pending watchers, the C<l_invoke> callback will
3810signal the main thread via some unspecified mechanism (signals? pipe
3811writes? C<Async::Interrupt>?) and then waits until all pending watchers
3812have been called (in a while loop because a) spurious wakeups are possible
3813and b) skipping inter-thread-communication when there are no pending
3814watchers is very beneficial):
3815
3816   static void
3817   l_invoke (EV_P)
3818   {
3819     userdata *u = ev_userdata (EV_A);
3820
3821     while (ev_pending_count (EV_A))
3822       {
3823         wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
3824         pthread_cond_wait (&u->invoke_cv, &u->lock);
3825       }
3826   }
3827
3828Now, whenever the main thread gets told to invoke pending watchers, it
3829will grab the lock, call C<ev_invoke_pending> and then signal the loop
3830thread to continue:
3831
3832   static void
3833   real_invoke_pending (EV_P)
3834   {
3835     userdata *u = ev_userdata (EV_A);
3836
3837     pthread_mutex_lock (&u->lock);
3838     ev_invoke_pending (EV_A);
3839     pthread_cond_signal (&u->invoke_cv);
3840     pthread_mutex_unlock (&u->lock);
3841   }
3842
3843Whenever you want to start/stop a watcher or do other modifications to an
3844event loop, you will now have to lock:
3845
3846   ev_timer timeout_watcher;
3847   userdata *u = ev_userdata (EV_A);
3848
3849   ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
3850
3851   pthread_mutex_lock (&u->lock);
3852   ev_timer_start (EV_A_ &timeout_watcher);
3853   ev_async_send (EV_A_ &u->async_w);
3854   pthread_mutex_unlock (&u->lock);
3855
3856Note that sending the C<ev_async> watcher is required because otherwise
3857an event loop currently blocking in the kernel will have no knowledge
3858about the newly added timer. By waking up the loop it will pick up any new
3859watchers in the next event loop iteration.
3860
3861=head2 THREADS, COROUTINES, CONTINUATIONS, QUEUES... INSTEAD OF CALLBACKS
3862
3863While the overhead of a callback that e.g. schedules a thread is small, it
3864is still an overhead. If you embed libev, and your main usage is with some
3865kind of threads or coroutines, you might want to customise libev so that
3866doesn't need callbacks anymore.
3867
3868Imagine you have coroutines that you can switch to using a function
3869C<switch_to (coro)>, that libev runs in a coroutine called C<libev_coro>
3870and that due to some magic, the currently active coroutine is stored in a
3871global called C<current_coro>. Then you can build your own "wait for libev
3872event" primitive by changing C<EV_CB_DECLARE> and C<EV_CB_INVOKE> (note
3873the differing C<;> conventions):
3874
3875   #define EV_CB_DECLARE(type)   struct my_coro *cb;
3876   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)
3877
3878That means instead of having a C callback function, you store the
3879coroutine to switch to in each watcher, and instead of having libev call
3880your callback, you instead have it switch to that coroutine.
3881
3882A coroutine might now wait for an event with a function called
3883C<wait_for_event>. (the watcher needs to be started, as always, but it doesn't
3884matter when, or whether the watcher is active or not when this function is
3885called):
3886
3887   void
3888   wait_for_event (ev_watcher *w)
3889   {
3890     ev_set_cb (w, current_coro);
3891     switch_to (libev_coro);
3892   }
3893
3894That basically suspends the coroutine inside C<wait_for_event> and
3895continues the libev coroutine, which, when appropriate, switches back to
3896this or any other coroutine.
3897
3898You can do similar tricks if you have, say, threads with an event queue -
3899instead of storing a coroutine, you store the queue object and instead of
3900switching to a coroutine, you push the watcher onto the queue and notify
3901any waiters.
3902
3903To embed libev, see L</EMBEDDING>, but in short, it's easiest to create two
3904files, F<my_ev.h> and F<my_ev.c> that include the respective libev files:
3905
3906   // my_ev.h
3907   #define EV_CB_DECLARE(type)   struct my_coro *cb;
3908   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)
3909   #include "../libev/ev.h"
3910
3911   // my_ev.c
3912   #define EV_H "my_ev.h"
3913   #include "../libev/ev.c"
3914
3915And then use F<my_ev.h> when you would normally use F<ev.h>, and compile
3916F<my_ev.c> into your project. When properly specifying include paths, you
3917can even use F<ev.h> as header file name directly.
3918
3919
3920=head1 LIBEVENT EMULATION
3921
3922Libev offers a compatibility emulation layer for libevent. It cannot
3923emulate the internals of libevent, so here are some usage hints:
3924
3925=over 4
3926
3927=item * Only the libevent-1.4.1-beta API is being emulated.
3928
3929This was the newest libevent version available when libev was implemented,
3930and is still mostly unchanged in 2010.
3931
3932=item * Use it by including <event.h>, as usual.
3933
3934=item * The following members are fully supported: ev_base, ev_callback,
3935ev_arg, ev_fd, ev_res, ev_events.
3936
3937=item * Avoid using ev_flags and the EVLIST_*-macros, while it is
3938maintained by libev, it does not work exactly the same way as in libevent (consider
3939it a private API).
3940
3941=item * Priorities are not currently supported. Initialising priorities
3942will fail and all watchers will have the same priority, even though there
3943is an ev_pri field.
3944
3945=item * In libevent, the last base created gets the signals, in libev, the
3946base that registered the signal gets the signals.
3947
3948=item * Other members are not supported.
3949
3950=item * The libev emulation is I<not> ABI compatible to libevent, you need
3951to use the libev header file and library.
3952
3953=back
3954
3955=head1 C++ SUPPORT
3956
3957=head2 C API
3958
3959The normal C API should work fine when used from C++: both ev.h and the
3960libev sources can be compiled as C++. Therefore, code that uses the C API
3961will work fine.
3962
3963Proper exception specifications might have to be added to callbacks passed
3964to libev: exceptions may be thrown only from watcher callbacks, all
3965other callbacks (allocator, syserr, loop acquire/release and periodic
3966reschedule callbacks) must not throw exceptions, and might need a C<throw
3967()> specification. If you have code that needs to be compiled as both C
3968and C++ you can use the C<EV_THROW> macro for this:
3969
3970   static void
3971   fatal_error (const char *msg) EV_THROW
3972   {
3973     perror (msg);
3974     abort ();
3975   }
3976
3977   ...
3978   ev_set_syserr_cb (fatal_error);
3979
3980The only API functions that can currently throw exceptions are C<ev_run>,
3981C<ev_invoke>, C<ev_invoke_pending> and C<ev_loop_destroy> (the latter
3982because it runs cleanup watchers).
3983
3984Throwing exceptions in watcher callbacks is only supported if libev itself
3985is compiled with a C++ compiler or your C and C++ environments allow
3986throwing exceptions through C libraries (most do).
3987
3988=head2 C++ API
3989
3990Libev comes with some simplistic wrapper classes for C++ that mainly allow
3991you to use some convenience methods to start/stop watchers and also change
3992the callback model to a model using method callbacks on objects.
3993
3994To use it,
3995
3996   #include <ev++.h>
3997
3998This automatically includes F<ev.h> and puts all of its definitions (many
3999of them macros) into the global namespace. All C++ specific things are
4000put into the C<ev> namespace. It should support all the same embedding
4001options as F<ev.h>, most notably C<EV_MULTIPLICITY>.
4002
4003Care has been taken to keep the overhead low. The only data member the C++
4004classes add (compared to plain C-style watchers) is the event loop pointer
4005that the watcher is associated with (or no additional members at all if
4006you disable C<EV_MULTIPLICITY> when embedding libev).
4007
4008Currently, functions, static and non-static member functions and classes
4009with C<operator ()> can be used as callbacks. Other types should be easy
4010to add as long as they only need one additional pointer for context. If
4011you need support for other types of functors please contact the author
4012(preferably after implementing it).
4013
4014For all this to work, your C++ compiler either has to use the same calling
4015conventions as your C compiler (for static member functions), or you have
4016to embed libev and compile libev itself as C++.
4017
4018Here is a list of things available in the C<ev> namespace:
4019
4020=over 4
4021
4022=item C<ev::READ>, C<ev::WRITE> etc.
4023
4024These are just enum values with the same values as the C<EV_READ> etc.
4025macros from F<ev.h>.
4026
4027=item C<ev::tstamp>, C<ev::now>
4028
4029Aliases to the same types/functions as with the C<ev_> prefix.
4030
4031=item C<ev::io>, C<ev::timer>, C<ev::periodic>, C<ev::idle>, C<ev::sig> etc.
4032
4033For each C<ev_TYPE> watcher in F<ev.h> there is a corresponding class of
4034the same name in the C<ev> namespace, with the exception of C<ev_signal>
4035which is called C<ev::sig> to avoid clashes with the C<signal> macro
4036defined by many implementations.
4037
4038All of those classes have these methods:
4039
4040=over 4
4041
4042=item ev::TYPE::TYPE ()
4043
4044=item ev::TYPE::TYPE (loop)
4045
4046=item ev::TYPE::~TYPE
4047
4048The constructor (optionally) takes an event loop to associate the watcher
4049with. If it is omitted, it will use C<EV_DEFAULT>.
4050
4051The constructor calls C<ev_init> for you, which means you have to call the
4052C<set> method before starting it.
4053
4054It will not set a callback, however: You have to call the templated C<set>
4055method to set a callback before you can start the watcher.
4056
4057(The reason why you have to use a method is a limitation in C++ which does
4058not allow explicit template arguments for constructors).
4059
4060The destructor automatically stops the watcher if it is active.
4061
4062=item w->set<class, &class::method> (object *)
4063
4064This method sets the callback method to call. The method has to have a
4065signature of C<void (*)(ev_TYPE &, int)>, it receives the watcher as
4066first argument and the C<revents> as second. The object must be given as
4067parameter and is stored in the C<data> member of the watcher.
4068
4069This method synthesizes efficient thunking code to call your method from
4070the C callback that libev requires. If your compiler can inline your
4071callback (i.e. it is visible to it at the place of the C<set> call and
4072your compiler is good :), then the method will be fully inlined into the
4073thunking function, making it as fast as a direct C callback.
4074
4075Example: simple class declaration and watcher initialisation
4076
4077   struct myclass
4078   {
4079     void io_cb (ev::io &w, int revents) { }
4080   }
4081
4082   myclass obj;
4083   ev::io iow;
4084   iow.set <myclass, &myclass::io_cb> (&obj);
4085
4086=item w->set (object *)
4087
4088This is a variation of a method callback - leaving out the method to call
4089will default the method to C<operator ()>, which makes it possible to use
4090functor objects without having to manually specify the C<operator ()> all
4091the time. Incidentally, you can then also leave out the template argument
4092list.
4093
4094The C<operator ()> method prototype must be C<void operator ()(watcher &w,
4095int revents)>.
4096
4097See the method-C<set> above for more details.
4098
4099Example: use a functor object as callback.
4100
4101   struct myfunctor
4102   {
4103     void operator() (ev::io &w, int revents)
4104     {
4105       ...
4106     }
4107   }
4108
4109   myfunctor f;
4110
4111   ev::io w;
4112   w.set (&f);
4113
4114=item w->set<function> (void *data = 0)
4115
4116Also sets a callback, but uses a static method or plain function as
4117callback. The optional C<data> argument will be stored in the watcher's
4118C<data> member and is free for you to use.
4119
4120The prototype of the C<function> must be C<void (*)(ev::TYPE &w, int)>.
4121
4122See the method-C<set> above for more details.
4123
4124Example: Use a plain function as callback.
4125
4126   static void io_cb (ev::io &w, int revents) { }
4127   iow.set <io_cb> ();
4128
4129=item w->set (loop)
4130
4131Associates a different C<struct ev_loop> with this watcher. You can only
4132do this when the watcher is inactive (and not pending either).
4133
4134=item w->set ([arguments])
4135
4136Basically the same as C<ev_TYPE_set> (except for C<ev::embed> watchers>),
4137with the same arguments. Either this method or a suitable start method
4138must be called at least once. Unlike the C counterpart, an active watcher
4139gets automatically stopped and restarted when reconfiguring it with this
4140method.
4141
4142For C<ev::embed> watchers this method is called C<set_embed>, to avoid
4143clashing with the C<set (loop)> method.
4144
4145=item w->start ()
4146
4147Starts the watcher. Note that there is no C<loop> argument, as the
4148constructor already stores the event loop.
4149
4150=item w->start ([arguments])
4151
4152Instead of calling C<set> and C<start> methods separately, it is often
4153convenient to wrap them in one call. Uses the same type of arguments as
4154the configure C<set> method of the watcher.
4155
4156=item w->stop ()
4157
4158Stops the watcher if it is active. Again, no C<loop> argument.
4159
4160=item w->again () (C<ev::timer>, C<ev::periodic> only)
4161
4162For C<ev::timer> and C<ev::periodic>, this invokes the corresponding
4163C<ev_TYPE_again> function.
4164
4165=item w->sweep () (C<ev::embed> only)
4166
4167Invokes C<ev_embed_sweep>.
4168
4169=item w->update () (C<ev::stat> only)
4170
4171Invokes C<ev_stat_stat>.
4172
4173=back
4174
4175=back
4176
4177Example: Define a class with two I/O and idle watchers, start the I/O
4178watchers in the constructor.
4179
4180   class myclass
4181   {
4182     ev::io   io  ; void io_cb   (ev::io   &w, int revents);
4183     ev::io   io2 ; void io2_cb  (ev::io   &w, int revents);
4184     ev::idle idle; void idle_cb (ev::idle &w, int revents);
4185
4186     myclass (int fd)
4187     {
4188       io  .set <myclass, &myclass::io_cb  > (this);
4189       io2 .set <myclass, &myclass::io2_cb > (this);
4190       idle.set <myclass, &myclass::idle_cb> (this);
4191
4192       io.set (fd, ev::WRITE); // configure the watcher
4193       io.start ();            // start it whenever convenient
4194
4195       io2.start (fd, ev::READ); // set + start in one call
4196     }
4197   };
4198
4199
4200=head1 OTHER LANGUAGE BINDINGS
4201
4202Libev does not offer other language bindings itself, but bindings for a
4203number of languages exist in the form of third-party packages. If you know
4204any interesting language binding in addition to the ones listed here, drop
4205me a note.
4206
4207=over 4
4208
4209=item Perl
4210
4211The EV module implements the full libev API and is actually used to test
4212libev. EV is developed together with libev. Apart from the EV core module,
4213there are additional modules that implement libev-compatible interfaces
4214to C<libadns> (C<EV::ADNS>, but C<AnyEvent::DNS> is preferred nowadays),
4215C<Net::SNMP> (C<Net::SNMP::EV>) and the C<libglib> event core (C<Glib::EV>
4216and C<EV::Glib>).
4217
4218It can be found and installed via CPAN, its homepage is at
4219L<http://software.schmorp.de/pkg/EV>.
4220
4221=item Python
4222
4223Python bindings can be found at L<http://code.google.com/p/pyev/>. It
4224seems to be quite complete and well-documented.
4225
4226=item Ruby
4227
4228Tony Arcieri has written a ruby extension that offers access to a subset
4229of the libev API and adds file handle abstractions, asynchronous DNS and
4230more on top of it. It can be found via gem servers. Its homepage is at
4231L<http://rev.rubyforge.org/>.
4232
4233Roger Pack reports that using the link order C<-lws2_32 -lmsvcrt-ruby-190>
4234makes rev work even on mingw.
4235
4236=item Haskell
4237
4238A haskell binding to libev is available at
4239L<http://hackage.haskell.org/cgi-bin/hackage-scripts/package/hlibev>.
4240
4241=item D
4242
4243Leandro Lucarella has written a D language binding (F<ev.d>) for libev, to
4244be found at L<http://www.llucax.com.ar/proj/ev.d/index.html>.
4245
4246=item Ocaml
4247
4248Erkki Seppala has written Ocaml bindings for libev, to be found at
4249L<http://modeemi.cs.tut.fi/~flux/software/ocaml-ev/>.
4250
4251=item Lua
4252
4253Brian Maher has written a partial interface to libev for lua (at the
4254time of this writing, only C<ev_io> and C<ev_timer>), to be found at
4255L<http://github.com/brimworks/lua-ev>.
4256
4257=item Javascript
4258
4259Node.js (L<http://nodejs.org>) uses libev as the underlying event library.
4260
4261=item Others
4262
4263There are others, and I stopped counting.
4264
4265=back
4266
4267
4268=head1 MACRO MAGIC
4269
4270Libev can be compiled with a variety of options, the most fundamental
4271of which is C<EV_MULTIPLICITY>. This option determines whether (most)
4272functions and callbacks have an initial C<struct ev_loop *> argument.
4273
4274To make it easier to write programs that cope with either variant, the
4275following macros are defined:
4276
4277=over 4
4278
4279=item C<EV_A>, C<EV_A_>
4280
4281This provides the loop I<argument> for functions, if one is required ("ev
4282loop argument"). The C<EV_A> form is used when this is the sole argument,
4283C<EV_A_> is used when other arguments are following. Example:
4284
4285   ev_unref (EV_A);
4286   ev_timer_add (EV_A_ watcher);
4287   ev_run (EV_A_ 0);
4288
4289It assumes the variable C<loop> of type C<struct ev_loop *> is in scope,
4290which is often provided by the following macro.
4291
4292=item C<EV_P>, C<EV_P_>
4293
4294This provides the loop I<parameter> for functions, if one is required ("ev
4295loop parameter"). The C<EV_P> form is used when this is the sole parameter,
4296C<EV_P_> is used when other parameters are following. Example:
4297
4298   // this is how ev_unref is being declared
4299   static void ev_unref (EV_P);
4300
4301   // this is how you can declare your typical callback
4302   static void cb (EV_P_ ev_timer *w, int revents)
4303
4304It declares a parameter C<loop> of type C<struct ev_loop *>, quite
4305suitable for use with C<EV_A>.
4306
4307=item C<EV_DEFAULT>, C<EV_DEFAULT_>
4308
4309Similar to the other two macros, this gives you the value of the default
4310loop, if multiple loops are supported ("ev loop default"). The default loop
4311will be initialised if it isn't already initialised.
4312
4313For non-multiplicity builds, these macros do nothing, so you always have
4314to initialise the loop somewhere.
4315
4316=item C<EV_DEFAULT_UC>, C<EV_DEFAULT_UC_>
4317
4318Usage identical to C<EV_DEFAULT> and C<EV_DEFAULT_>, but requires that the
4319default loop has been initialised (C<UC> == unchecked). Their behaviour
4320is undefined when the default loop has not been initialised by a previous
4321execution of C<EV_DEFAULT>, C<EV_DEFAULT_> or C<ev_default_init (...)>.
4322
4323It is often prudent to use C<EV_DEFAULT> when initialising the first
4324watcher in a function but use C<EV_DEFAULT_UC> afterwards.
4325
4326=back
4327
4328Example: Declare and initialise a check watcher, utilising the above
4329macros so it will work regardless of whether multiple loops are supported
4330or not.
4331
4332   static void
4333   check_cb (EV_P_ ev_timer *w, int revents)
4334   {
4335     ev_check_stop (EV_A_ w);
4336   }
4337
4338   ev_check check;
4339   ev_check_init (&check, check_cb);
4340   ev_check_start (EV_DEFAULT_ &check);
4341   ev_run (EV_DEFAULT_ 0);
4342
4343=head1 EMBEDDING
4344
4345Libev can (and often is) directly embedded into host
4346applications. Examples of applications that embed it include the Deliantra
4347Game Server, the EV perl module, the GNU Virtual Private Ethernet (gvpe)
4348and rxvt-unicode.
4349
4350The goal is to enable you to just copy the necessary files into your
4351source directory without having to change even a single line in them, so
4352you can easily upgrade by simply copying (or having a checked-out copy of
4353libev somewhere in your source tree).
4354
4355=head2 FILESETS
4356
4357Depending on what features you need you need to include one or more sets of files
4358in your application.
4359
4360=head3 CORE EVENT LOOP
4361
4362To include only the libev core (all the C<ev_*> functions), with manual
4363configuration (no autoconf):
4364
4365   #define EV_STANDALONE 1
4366   #include "ev.c"
4367
4368This will automatically include F<ev.h>, too, and should be done in a
4369single C source file only to provide the function implementations. To use
4370it, do the same for F<ev.h> in all files wishing to use this API (best
4371done by writing a wrapper around F<ev.h> that you can include instead and
4372where you can put other configuration options):
4373
4374   #define EV_STANDALONE 1
4375   #include "ev.h"
4376
4377Both header files and implementation files can be compiled with a C++
4378compiler (at least, that's a stated goal, and breakage will be treated
4379as a bug).
4380
4381You need the following files in your source tree, or in a directory
4382in your include path (e.g. in libev/ when using -Ilibev):
4383
4384   ev.h
4385   ev.c
4386   ev_vars.h
4387   ev_wrap.h
4388
4389   ev_win32.c      required on win32 platforms only
4390
4391   ev_select.c     only when select backend is enabled (which is enabled by default)
4392   ev_poll.c       only when poll backend is enabled (disabled by default)
4393   ev_epoll.c      only when the epoll backend is enabled (disabled by default)
4394   ev_kqueue.c     only when the kqueue backend is enabled (disabled by default)
4395   ev_port.c       only when the solaris port backend is enabled (disabled by default)
4396
4397F<ev.c> includes the backend files directly when enabled, so you only need
4398to compile this single file.
4399
4400=head3 LIBEVENT COMPATIBILITY API
4401
4402To include the libevent compatibility API, also include:
4403
4404   #include "event.c"
4405
4406in the file including F<ev.c>, and:
4407
4408   #include "event.h"
4409
4410in the files that want to use the libevent API. This also includes F<ev.h>.
4411
4412You need the following additional files for this:
4413
4414   event.h
4415   event.c
4416
4417=head3 AUTOCONF SUPPORT
4418
4419Instead of using C<EV_STANDALONE=1> and providing your configuration in
4420whatever way you want, you can also C<m4_include([libev.m4])> in your
4421F<configure.ac> and leave C<EV_STANDALONE> undefined. F<ev.c> will then
4422include F<config.h> and configure itself accordingly.
4423
4424For this of course you need the m4 file:
4425
4426   libev.m4
4427
4428=head2 PREPROCESSOR SYMBOLS/MACROS
4429
4430Libev can be configured via a variety of preprocessor symbols you have to
4431define before including (or compiling) any of its files. The default in
4432the absence of autoconf is documented for every option.
4433
4434Symbols marked with "(h)" do not change the ABI, and can have different
4435values when compiling libev vs. including F<ev.h>, so it is permissible
4436to redefine them before including F<ev.h> without breaking compatibility
4437to a compiled library. All other symbols change the ABI, which means all
4438users of libev and the libev code itself must be compiled with compatible
4439settings.
4440
4441=over 4
4442
4443=item EV_COMPAT3 (h)
4444
4445Backwards compatibility is a major concern for libev. This is why this
4446release of libev comes with wrappers for the functions and symbols that
4447have been renamed between libev version 3 and 4.
4448
4449You can disable these wrappers (to test compatibility with future
4450versions) by defining C<EV_COMPAT3> to C<0> when compiling your
4451sources. This has the additional advantage that you can drop the C<struct>
4452from C<struct ev_loop> declarations, as libev will provide an C<ev_loop>
4453typedef in that case.
4454
4455In some future version, the default for C<EV_COMPAT3> will become C<0>,
4456and in some even more future version the compatibility code will be
4457removed completely.
4458
4459=item EV_STANDALONE (h)
4460
4461Must always be C<1> if you do not use autoconf configuration, which
4462keeps libev from including F<config.h>, and it also defines dummy
4463implementations for some libevent functions (such as logging, which is not
4464supported). It will also not define any of the structs usually found in
4465F<event.h> that are not directly supported by the libev core alone.
4466
4467In standalone mode, libev will still try to automatically deduce the
4468configuration, but has to be more conservative.
4469
4470=item EV_USE_FLOOR
4471
4472If defined to be C<1>, libev will use the C<floor ()> function for its
4473periodic reschedule calculations, otherwise libev will fall back on a
4474portable (slower) implementation. If you enable this, you usually have to
4475link against libm or something equivalent. Enabling this when the C<floor>
4476function is not available will fail, so the safe default is to not enable
4477this.
4478
4479=item EV_USE_MONOTONIC
4480
4481If defined to be C<1>, libev will try to detect the availability of the
4482monotonic clock option at both compile time and runtime. Otherwise no
4483use of the monotonic clock option will be attempted. If you enable this,
4484you usually have to link against librt or something similar. Enabling it
4485when the functionality isn't available is safe, though, although you have
4486to make sure you link against any libraries where the C<clock_gettime>
4487function is hiding in (often F<-lrt>). See also C<EV_USE_CLOCK_SYSCALL>.
4488
4489=item EV_USE_REALTIME
4490
4491If defined to be C<1>, libev will try to detect the availability of the
4492real-time clock option at compile time (and assume its availability
4493at runtime if successful). Otherwise no use of the real-time clock
4494option will be attempted. This effectively replaces C<gettimeofday>
4495by C<clock_get (CLOCK_REALTIME, ...)> and will not normally affect
4496correctness. See the note about libraries in the description of
4497C<EV_USE_MONOTONIC>, though. Defaults to the opposite value of
4498C<EV_USE_CLOCK_SYSCALL>.
4499
4500=item EV_USE_CLOCK_SYSCALL
4501
4502If defined to be C<1>, libev will try to use a direct syscall instead
4503of calling the system-provided C<clock_gettime> function. This option
4504exists because on GNU/Linux, C<clock_gettime> is in C<librt>, but C<librt>
4505unconditionally pulls in C<libpthread>, slowing down single-threaded
4506programs needlessly. Using a direct syscall is slightly slower (in
4507theory), because no optimised vdso implementation can be used, but avoids
4508the pthread dependency. Defaults to C<1> on GNU/Linux with glibc 2.x or
4509higher, as it simplifies linking (no need for C<-lrt>).
4510
4511=item EV_USE_NANOSLEEP
4512
4513If defined to be C<1>, libev will assume that C<nanosleep ()> is available
4514and will use it for delays. Otherwise it will use C<select ()>.
4515
4516=item EV_USE_EVENTFD
4517
4518If defined to be C<1>, then libev will assume that C<eventfd ()> is
4519available and will probe for kernel support at runtime. This will improve
4520C<ev_signal> and C<ev_async> performance and reduce resource consumption.
4521If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc
45222.7 or newer, otherwise disabled.
4523
4524=item EV_USE_SELECT
4525
4526If undefined or defined to be C<1>, libev will compile in support for the
4527C<select>(2) backend. No attempt at auto-detection will be done: if no
4528other method takes over, select will be it. Otherwise the select backend
4529will not be compiled in.
4530
4531=item EV_SELECT_USE_FD_SET
4532
4533If defined to C<1>, then the select backend will use the system C<fd_set>
4534structure. This is useful if libev doesn't compile due to a missing
4535C<NFDBITS> or C<fd_mask> definition or it mis-guesses the bitset layout
4536on exotic systems. This usually limits the range of file descriptors to
4537some low limit such as 1024 or might have other limitations (winsocket
4538only allows 64 sockets). The C<FD_SETSIZE> macro, set before compilation,
4539configures the maximum size of the C<fd_set>.
4540
4541=item EV_SELECT_IS_WINSOCKET
4542
4543When defined to C<1>, the select backend will assume that
4544select/socket/connect etc. don't understand file descriptors but
4545wants osf handles on win32 (this is the case when the select to
4546be used is the winsock select). This means that it will call
4547C<_get_osfhandle> on the fd to convert it to an OS handle. Otherwise,
4548it is assumed that all these functions actually work on fds, even
4549on win32. Should not be defined on non-win32 platforms.
4550
4551=item EV_FD_TO_WIN32_HANDLE(fd)
4552
4553If C<EV_SELECT_IS_WINSOCKET> is enabled, then libev needs a way to map
4554file descriptors to socket handles. When not defining this symbol (the
4555default), then libev will call C<_get_osfhandle>, which is usually
4556correct. In some cases, programs use their own file descriptor management,
4557in which case they can provide this function to map fds to socket handles.
4558
4559=item EV_WIN32_HANDLE_TO_FD(handle)
4560
4561If C<EV_SELECT_IS_WINSOCKET> then libev maps handles to file descriptors
4562using the standard C<_open_osfhandle> function. For programs implementing
4563their own fd to handle mapping, overwriting this function makes it easier
4564to do so. This can be done by defining this macro to an appropriate value.
4565
4566=item EV_WIN32_CLOSE_FD(fd)
4567
4568If programs implement their own fd to handle mapping on win32, then this
4569macro can be used to override the C<close> function, useful to unregister
4570file descriptors again. Note that the replacement function has to close
4571the underlying OS handle.
4572
4573=item EV_USE_WSASOCKET
4574
4575If defined to be C<1>, libev will use C<WSASocket> to create its internal
4576communication socket, which works better in some environments. Otherwise,
4577the normal C<socket> function will be used, which works better in other
4578environments.
4579
4580=item EV_USE_POLL
4581
4582If defined to be C<1>, libev will compile in support for the C<poll>(2)
4583backend. Otherwise it will be enabled on non-win32 platforms. It
4584takes precedence over select.
4585
4586=item EV_USE_EPOLL
4587
4588If defined to be C<1>, libev will compile in support for the Linux
4589C<epoll>(7) backend. Its availability will be detected at runtime,
4590otherwise another method will be used as fallback. This is the preferred
4591backend for GNU/Linux systems. If undefined, it will be enabled if the
4592headers indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.
4593
4594=item EV_USE_KQUEUE
4595
4596If defined to be C<1>, libev will compile in support for the BSD style
4597C<kqueue>(2) backend. Its actual availability will be detected at runtime,
4598otherwise another method will be used as fallback. This is the preferred
4599backend for BSD and BSD-like systems, although on most BSDs kqueue only
4600supports some types of fds correctly (the only platform we found that
4601supports ptys for example was NetBSD), so kqueue might be compiled in, but
4602not be used unless explicitly requested. The best way to use it is to find
4603out whether kqueue supports your type of fd properly and use an embedded
4604kqueue loop.
4605
4606=item EV_USE_PORT
4607
4608If defined to be C<1>, libev will compile in support for the Solaris
460910 port style backend. Its availability will be detected at runtime,
4610otherwise another method will be used as fallback. This is the preferred
4611backend for Solaris 10 systems.
4612
4613=item EV_USE_DEVPOLL
4614
4615Reserved for future expansion, works like the USE symbols above.
4616
4617=item EV_USE_INOTIFY
4618
4619If defined to be C<1>, libev will compile in support for the Linux inotify
4620interface to speed up C<ev_stat> watchers. Its actual availability will
4621be detected at runtime. If undefined, it will be enabled if the headers
4622indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.
4623
4624=item EV_NO_SMP
4625
4626If defined to be C<1>, libev will assume that memory is always coherent
4627between threads, that is, threads can be used, but threads never run on
4628different cpus (or different cpu cores). This reduces dependencies
4629and makes libev faster.
4630
4631=item EV_NO_THREADS
4632
4633If defined to be C<1>, libev will assume that it will never be called from
4634different threads (that includes signal handlers), which is a stronger
4635assumption than C<EV_NO_SMP>, above. This reduces dependencies and makes
4636libev faster.
4637
4638=item EV_ATOMIC_T
4639
4640Libev requires an integer type (suitable for storing C<0> or C<1>) whose
4641access is atomic with respect to other threads or signal contexts. No
4642such type is easily found in the C language, so you can provide your own
4643type that you know is safe for your purposes. It is used both for signal
4644handler "locking" as well as for signal and thread safety in C<ev_async>
4645watchers.
4646
4647In the absence of this define, libev will use C<sig_atomic_t volatile>
4648(from F<signal.h>), which is usually good enough on most platforms.
4649
4650=item EV_H (h)
4651
4652The name of the F<ev.h> header file used to include it. The default if
4653undefined is C<"ev.h"> in F<event.h>, F<ev.c> and F<ev++.h>. This can be
4654used to virtually rename the F<ev.h> header file in case of conflicts.
4655
4656=item EV_CONFIG_H (h)
4657
4658If C<EV_STANDALONE> isn't C<1>, this variable can be used to override
4659F<ev.c>'s idea of where to find the F<config.h> file, similarly to
4660C<EV_H>, above.
4661
4662=item EV_EVENT_H (h)
4663
4664Similarly to C<EV_H>, this macro can be used to override F<event.c>'s idea
4665of how the F<event.h> header can be found, the default is C<"event.h">.
4666
4667=item EV_PROTOTYPES (h)
4668
4669If defined to be C<0>, then F<ev.h> will not define any function
4670prototypes, but still define all the structs and other symbols. This is
4671occasionally useful if you want to provide your own wrapper functions
4672around libev functions.
4673
4674=item EV_MULTIPLICITY
4675
4676If undefined or defined to C<1>, then all event-loop-specific functions
4677will have the C<struct ev_loop *> as first argument, and you can create
4678additional independent event loops. Otherwise there will be no support
4679for multiple event loops and there is no first event loop pointer
4680argument. Instead, all functions act on the single default loop.
4681
4682Note that C<EV_DEFAULT> and C<EV_DEFAULT_> will no longer provide a
4683default loop when multiplicity is switched off - you always have to
4684initialise the loop manually in this case.
4685
4686=item EV_MINPRI
4687
4688=item EV_MAXPRI
4689
4690The range of allowed priorities. C<EV_MINPRI> must be smaller or equal to
4691C<EV_MAXPRI>, but otherwise there are no non-obvious limitations. You can
4692provide for more priorities by overriding those symbols (usually defined
4693to be C<-2> and C<2>, respectively).
4694
4695When doing priority-based operations, libev usually has to linearly search
4696all the priorities, so having many of them (hundreds) uses a lot of space
4697and time, so using the defaults of five priorities (-2 .. +2) is usually
4698fine.
4699
4700If your embedding application does not need any priorities, defining these
4701both to C<0> will save some memory and CPU.
4702
4703=item EV_PERIODIC_ENABLE, EV_IDLE_ENABLE, EV_EMBED_ENABLE, EV_STAT_ENABLE,
4704EV_PREPARE_ENABLE, EV_CHECK_ENABLE, EV_FORK_ENABLE, EV_SIGNAL_ENABLE,
4705EV_ASYNC_ENABLE, EV_CHILD_ENABLE.
4706
4707If undefined or defined to be C<1> (and the platform supports it), then
4708the respective watcher type is supported. If defined to be C<0>, then it
4709is not. Disabling watcher types mainly saves code size.
4710
4711=item EV_FEATURES
4712
4713If you need to shave off some kilobytes of code at the expense of some
4714speed (but with the full API), you can define this symbol to request
4715certain subsets of functionality. The default is to enable all features
4716that can be enabled on the platform.
4717
4718A typical way to use this symbol is to define it to C<0> (or to a bitset
4719with some broad features you want) and then selectively re-enable
4720additional parts you want, for example if you want everything minimal,
4721but multiple event loop support, async and child watchers and the poll
4722backend, use this:
4723
4724   #define EV_FEATURES 0
4725   #define EV_MULTIPLICITY 1
4726   #define EV_USE_POLL 1
4727   #define EV_CHILD_ENABLE 1
4728   #define EV_ASYNC_ENABLE 1
4729
4730The actual value is a bitset, it can be a combination of the following
4731values (by default, all of these are enabled):
4732
4733=over 4
4734
4735=item C<1> - faster/larger code
4736
4737Use larger code to speed up some operations.
4738
4739Currently this is used to override some inlining decisions (enlarging the
4740code size by roughly 30% on amd64).
4741
4742When optimising for size, use of compiler flags such as C<-Os> with
4743gcc is recommended, as well as C<-DNDEBUG>, as libev contains a number of
4744assertions.
4745
4746The default is off when C<__OPTIMIZE_SIZE__> is defined by your compiler
4747(e.g. gcc with C<-Os>).
4748
4749=item C<2> - faster/larger data structures
4750
4751Replaces the small 2-heap for timer management by a faster 4-heap, larger
4752hash table sizes and so on. This will usually further increase code size
4753and can additionally have an effect on the size of data structures at
4754runtime.
4755
4756The default is off when C<__OPTIMIZE_SIZE__> is defined by your compiler
4757(e.g. gcc with C<-Os>).
4758
4759=item C<4> - full API configuration
4760
4761This enables priorities (sets C<EV_MAXPRI>=2 and C<EV_MINPRI>=-2), and
4762enables multiplicity (C<EV_MULTIPLICITY>=1).
4763
4764=item C<8> - full API
4765
4766This enables a lot of the "lesser used" API functions. See C<ev.h> for
4767details on which parts of the API are still available without this
4768feature, and do not complain if this subset changes over time.
4769
4770=item C<16> - enable all optional watcher types
4771
4772Enables all optional watcher types.  If you want to selectively enable
4773only some watcher types other than I/O and timers (e.g. prepare,
4774embed, async, child...) you can enable them manually by defining
4775C<EV_watchertype_ENABLE> to C<1> instead.
4776
4777=item C<32> - enable all backends
4778
4779This enables all backends - without this feature, you need to enable at
4780least one backend manually (C<EV_USE_SELECT> is a good choice).
4781
4782=item C<64> - enable OS-specific "helper" APIs
4783
4784Enable inotify, eventfd, signalfd and similar OS-specific helper APIs by
4785default.
4786
4787=back
4788
4789Compiling with C<gcc -Os -DEV_STANDALONE -DEV_USE_EPOLL=1 -DEV_FEATURES=0>
4790reduces the compiled size of libev from 24.7Kb code/2.8Kb data to 6.5Kb
4791code/0.3Kb data on my GNU/Linux amd64 system, while still giving you I/O
4792watchers, timers and monotonic clock support.
4793
4794With an intelligent-enough linker (gcc+binutils are intelligent enough
4795when you use C<-Wl,--gc-sections -ffunction-sections>) functions unused by
4796your program might be left out as well - a binary starting a timer and an
4797I/O watcher then might come out at only 5Kb.
4798
4799=item EV_API_STATIC
4800
4801If this symbol is defined (by default it is not), then all identifiers
4802will have static linkage. This means that libev will not export any
4803identifiers, and you cannot link against libev anymore. This can be useful
4804when you embed libev, only want to use libev functions in a single file,
4805and do not want its identifiers to be visible.
4806
4807To use this, define C<EV_API_STATIC> and include F<ev.c> in the file that
4808wants to use libev.
4809
4810This option only works when libev is compiled with a C compiler, as C++
4811doesn't support the required declaration syntax.
4812
4813=item EV_AVOID_STDIO
4814
4815If this is set to C<1> at compiletime, then libev will avoid using stdio
4816functions (printf, scanf, perror etc.). This will increase the code size
4817somewhat, but if your program doesn't otherwise depend on stdio and your
4818libc allows it, this avoids linking in the stdio library which is quite
4819big.
4820
4821Note that error messages might become less precise when this option is
4822enabled.
4823
4824=item EV_NSIG
4825
4826The highest supported signal number, +1 (or, the number of
4827signals): Normally, libev tries to deduce the maximum number of signals
4828automatically, but sometimes this fails, in which case it can be
4829specified. Also, using a lower number than detected (C<32> should be
4830good for about any system in existence) can save some memory, as libev
4831statically allocates some 12-24 bytes per signal number.
4832
4833=item EV_PID_HASHSIZE
4834
4835C<ev_child> watchers use a small hash table to distribute workload by
4836pid. The default size is C<16> (or C<1> with C<EV_FEATURES> disabled),
4837usually more than enough. If you need to manage thousands of children you
4838might want to increase this value (I<must> be a power of two).
4839
4840=item EV_INOTIFY_HASHSIZE
4841
4842C<ev_stat> watchers use a small hash table to distribute workload by
4843inotify watch id. The default size is C<16> (or C<1> with C<EV_FEATURES>
4844disabled), usually more than enough. If you need to manage thousands of
4845C<ev_stat> watchers you might want to increase this value (I<must> be a
4846power of two).
4847
4848=item EV_USE_4HEAP
4849
4850Heaps are not very cache-efficient. To improve the cache-efficiency of the
4851timer and periodics heaps, libev uses a 4-heap when this symbol is defined
4852to C<1>. The 4-heap uses more complicated (longer) code but has noticeably
4853faster performance with many (thousands) of watchers.
4854
4855The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
4856will be C<0>.
4857
4858=item EV_HEAP_CACHE_AT
4859
4860Heaps are not very cache-efficient. To improve the cache-efficiency of the
4861timer and periodics heaps, libev can cache the timestamp (I<at>) within
4862the heap structure (selected by defining C<EV_HEAP_CACHE_AT> to C<1>),
4863which uses 8-12 bytes more per watcher and a few hundred bytes more code,
4864but avoids random read accesses on heap changes. This improves performance
4865noticeably with many (hundreds) of watchers.
4866
4867The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
4868will be C<0>.
4869
4870=item EV_VERIFY
4871
4872Controls how much internal verification (see C<ev_verify ()>) will
4873be done: If set to C<0>, no internal verification code will be compiled
4874in. If set to C<1>, then verification code will be compiled in, but not
4875called. If set to C<2>, then the internal verification code will be
4876called once per loop, which can slow down libev. If set to C<3>, then the
4877verification code will be called very frequently, which will slow down
4878libev considerably.
4879
4880The default is C<1>, unless C<EV_FEATURES> overrides it, in which case it
4881will be C<0>.
4882
4883=item EV_COMMON
4884
4885By default, all watchers have a C<void *data> member. By redefining
4886this macro to something else you can include more and other types of
4887members. You have to define it each time you include one of the files,
4888though, and it must be identical each time.
4889
4890For example, the perl EV module uses something like this:
4891
4892   #define EV_COMMON                       \
4893     SV *self; /* contains this struct */  \
4894     SV *cb_sv, *fh /* note no trailing ";" */
4895
4896=item EV_CB_DECLARE (type)
4897
4898=item EV_CB_INVOKE (watcher, revents)
4899
4900=item ev_set_cb (ev, cb)
4901
4902Can be used to change the callback member declaration in each watcher,
4903and the way callbacks are invoked and set. Must expand to a struct member
4904definition and a statement, respectively. See the F<ev.h> header file for
4905their default definitions. One possible use for overriding these is to
4906avoid the C<struct ev_loop *> as first argument in all cases, or to use
4907method calls instead of plain function calls in C++.
4908
4909=back
4910
4911=head2 EXPORTED API SYMBOLS
4912
4913If you need to re-export the API (e.g. via a DLL) and you need a list of
4914exported symbols, you can use the provided F<Symbol.*> files which list
4915all public symbols, one per line:
4916
4917   Symbols.ev      for libev proper
4918   Symbols.event   for the libevent emulation
4919
4920This can also be used to rename all public symbols to avoid clashes with
4921multiple versions of libev linked together (which is obviously bad in
4922itself, but sometimes it is inconvenient to avoid this).
4923
4924A sed command like this will create wrapper C<#define>'s that you need to
4925include before including F<ev.h>:
4926
4927   <Symbols.ev sed -e "s/.*/#define & myprefix_&/" >wrap.h
4928
4929This would create a file F<wrap.h> which essentially looks like this:
4930
4931   #define ev_backend     myprefix_ev_backend
4932   #define ev_check_start myprefix_ev_check_start
4933   #define ev_check_stop  myprefix_ev_check_stop
4934   ...
4935
4936=head2 EXAMPLES
4937
4938For a real-world example of a program the includes libev
4939verbatim, you can have a look at the EV perl module
4940(L<http://software.schmorp.de/pkg/EV.html>). It has the libev files in
4941the F<libev/> subdirectory and includes them in the F<EV/EVAPI.h> (public
4942interface) and F<EV.xs> (implementation) files. Only the F<EV.xs> file
4943will be compiled. It is pretty complex because it provides its own header
4944file.
4945
4946The usage in rxvt-unicode is simpler. It has a F<ev_cpp.h> header file
4947that everybody includes and which overrides some configure choices:
4948
4949   #define EV_FEATURES 8
4950   #define EV_USE_SELECT 1
4951   #define EV_PREPARE_ENABLE 1
4952   #define EV_IDLE_ENABLE 1
4953   #define EV_SIGNAL_ENABLE 1
4954   #define EV_CHILD_ENABLE 1
4955   #define EV_USE_STDEXCEPT 0
4956   #define EV_CONFIG_H <config.h>
4957
4958   #include "ev++.h"
4959
4960And a F<ev_cpp.C> implementation file that contains libev proper and is compiled:
4961
4962   #include "ev_cpp.h"
4963   #include "ev.c"
4964
4965=head1 INTERACTION WITH OTHER PROGRAMS, LIBRARIES OR THE ENVIRONMENT
4966
4967=head2 THREADS AND COROUTINES
4968
4969=head3 THREADS
4970
4971All libev functions are reentrant and thread-safe unless explicitly
4972documented otherwise, but libev implements no locking itself. This means
4973that you can use as many loops as you want in parallel, as long as there
4974are no concurrent calls into any libev function with the same loop
4975parameter (C<ev_default_*> calls have an implicit default loop parameter,
4976of course): libev guarantees that different event loops share no data
4977structures that need any locking.
4978
4979Or to put it differently: calls with different loop parameters can be done
4980concurrently from multiple threads, calls with the same loop parameter
4981must be done serially (but can be done from different threads, as long as
4982only one thread ever is inside a call at any point in time, e.g. by using
4983a mutex per loop).
4984
4985Specifically to support threads (and signal handlers), libev implements
4986so-called C<ev_async> watchers, which allow some limited form of
4987concurrency on the same event loop, namely waking it up "from the
4988outside".
4989
4990If you want to know which design (one loop, locking, or multiple loops
4991without or something else still) is best for your problem, then I cannot
4992help you, but here is some generic advice:
4993
4994=over 4
4995
4996=item * most applications have a main thread: use the default libev loop
4997in that thread, or create a separate thread running only the default loop.
4998
4999This helps integrating other libraries or software modules that use libev
5000themselves and don't care/know about threading.
5001
5002=item * one loop per thread is usually a good model.
5003
5004Doing this is almost never wrong, sometimes a better-performance model
5005exists, but it is always a good start.
5006
5007=item * other models exist, such as the leader/follower pattern, where one
5008loop is handed through multiple threads in a kind of round-robin fashion.
5009
5010Choosing a model is hard - look around, learn, know that usually you can do
5011better than you currently do :-)
5012
5013=item * often you need to talk to some other thread which blocks in the
5014event loop.
5015
5016C<ev_async> watchers can be used to wake them up from other threads safely
5017(or from signal contexts...).
5018
5019An example use would be to communicate signals or other events that only
5020work in the default loop by registering the signal watcher with the
5021default loop and triggering an C<ev_async> watcher from the default loop
5022watcher callback into the event loop interested in the signal.
5023
5024=back
5025
5026See also L</THREAD LOCKING EXAMPLE>.
5027
5028=head3 COROUTINES
5029
5030Libev is very accommodating to coroutines ("cooperative threads"):
5031libev fully supports nesting calls to its functions from different
5032coroutines (e.g. you can call C<ev_run> on the same loop from two
5033different coroutines, and switch freely between both coroutines running
5034the loop, as long as you don't confuse yourself). The only exception is
5035that you must not do this from C<ev_periodic> reschedule callbacks.
5036
5037Care has been taken to ensure that libev does not keep local state inside
5038C<ev_run>, and other calls do not usually allow for coroutine switches as
5039they do not call any callbacks.
5040
5041=head2 COMPILER WARNINGS
5042
5043Depending on your compiler and compiler settings, you might get no or a
5044lot of warnings when compiling libev code. Some people are apparently
5045scared by this.
5046
5047However, these are unavoidable for many reasons. For one, each compiler
5048has different warnings, and each user has different tastes regarding
5049warning options. "Warn-free" code therefore cannot be a goal except when
5050targeting a specific compiler and compiler-version.
5051
5052Another reason is that some compiler warnings require elaborate
5053workarounds, or other changes to the code that make it less clear and less
5054maintainable.
5055
5056And of course, some compiler warnings are just plain stupid, or simply
5057wrong (because they don't actually warn about the condition their message
5058seems to warn about). For example, certain older gcc versions had some
5059warnings that resulted in an extreme number of false positives. These have
5060been fixed, but some people still insist on making code warn-free with
5061such buggy versions.
5062
5063While libev is written to generate as few warnings as possible,
5064"warn-free" code is not a goal, and it is recommended not to build libev
5065with any compiler warnings enabled unless you are prepared to cope with
5066them (e.g. by ignoring them). Remember that warnings are just that:
5067warnings, not errors, or proof of bugs.
5068
5069
5070=head2 VALGRIND
5071
5072Valgrind has a special section here because it is a popular tool that is
5073highly useful. Unfortunately, valgrind reports are very hard to interpret.
5074
5075If you think you found a bug (memory leak, uninitialised data access etc.)
5076in libev, then check twice: If valgrind reports something like:
5077
5078   ==2274==    definitely lost: 0 bytes in 0 blocks.
5079   ==2274==      possibly lost: 0 bytes in 0 blocks.
5080   ==2274==    still reachable: 256 bytes in 1 blocks.
5081
5082Then there is no memory leak, just as memory accounted to global variables
5083is not a memleak - the memory is still being referenced, and didn't leak.
5084
5085Similarly, under some circumstances, valgrind might report kernel bugs
5086as if it were a bug in libev (e.g. in realloc or in the poll backend,
5087although an acceptable workaround has been found here), or it might be
5088confused.
5089
5090Keep in mind that valgrind is a very good tool, but only a tool. Don't
5091make it into some kind of religion.
5092
5093If you are unsure about something, feel free to contact the mailing list
5094with the full valgrind report and an explanation on why you think this
5095is a bug in libev (best check the archives, too :). However, don't be
5096annoyed when you get a brisk "this is no bug" answer and take the chance
5097of learning how to interpret valgrind properly.
5098
5099If you need, for some reason, empty reports from valgrind for your project
5100I suggest using suppression lists.
5101
5102
5103=head1 PORTABILITY NOTES
5104
5105=head2 GNU/LINUX 32 BIT LIMITATIONS
5106
5107GNU/Linux is the only common platform that supports 64 bit file/large file
5108interfaces but I<disables> them by default.
5109
5110That means that libev compiled in the default environment doesn't support
5111files larger than 2GiB or so, which mainly affects C<ev_stat> watchers.
5112
5113Unfortunately, many programs try to work around this GNU/Linux issue
5114by enabling the large file API, which makes them incompatible with the
5115standard libev compiled for their system.
5116
5117Likewise, libev cannot enable the large file API itself as this would
5118suddenly make it incompatible to the default compile time environment,
5119i.e. all programs not using special compile switches.
5120
5121=head2 OS/X AND DARWIN BUGS
5122
5123The whole thing is a bug if you ask me - basically any system interface
5124you touch is broken, whether it is locales, poll, kqueue or even the
5125OpenGL drivers.
5126
5127=head3 C<kqueue> is buggy
5128
5129The kqueue syscall is broken in all known versions - most versions support
5130only sockets, many support pipes.
5131
5132Libev tries to work around this by not using C<kqueue> by default on this
5133rotten platform, but of course you can still ask for it when creating a
5134loop - embedding a socket-only kqueue loop into a select-based one is
5135probably going to work well.
5136
5137=head3 C<poll> is buggy
5138
5139Instead of fixing C<kqueue>, Apple replaced their (working) C<poll>
5140implementation by something calling C<kqueue> internally around the 10.5.6
5141release, so now C<kqueue> I<and> C<poll> are broken.
5142
5143Libev tries to work around this by not using C<poll> by default on
5144this rotten platform, but of course you can still ask for it when creating
5145a loop.
5146
5147=head3 C<select> is buggy
5148
5149All that's left is C<select>, and of course Apple found a way to fuck this
5150one up as well: On OS/X, C<select> actively limits the number of file
5151descriptors you can pass in to 1024 - your program suddenly crashes when
5152you use more.
5153
5154There is an undocumented "workaround" for this - defining
5155C<_DARWIN_UNLIMITED_SELECT>, which libev tries to use, so select I<should>
5156work on OS/X.
5157
5158=head2 SOLARIS PROBLEMS AND WORKAROUNDS
5159
5160=head3 C<errno> reentrancy
5161
5162The default compile environment on Solaris is unfortunately so
5163thread-unsafe that you can't even use components/libraries compiled
5164without C<-D_REENTRANT> in a threaded program, which, of course, isn't
5165defined by default. A valid, if stupid, implementation choice.
5166
5167If you want to use libev in threaded environments you have to make sure
5168it's compiled with C<_REENTRANT> defined.
5169
5170=head3 Event port backend
5171
5172The scalable event interface for Solaris is called "event
5173ports". Unfortunately, this mechanism is very buggy in all major
5174releases. If you run into high CPU usage, your program freezes or you get
5175a large number of spurious wakeups, make sure you have all the relevant
5176and latest kernel patches applied. No, I don't know which ones, but there
5177are multiple ones to apply, and afterwards, event ports actually work
5178great.
5179
5180If you can't get it to work, you can try running the program by setting
5181the environment variable C<LIBEV_FLAGS=3> to only allow C<poll> and
5182C<select> backends.
5183
5184=head2 AIX POLL BUG
5185
5186AIX unfortunately has a broken C<poll.h> header. Libev works around
5187this by trying to avoid the poll backend altogether (i.e. it's not even
5188compiled in), which normally isn't a big problem as C<select> works fine
5189with large bitsets on AIX, and AIX is dead anyway.
5190
5191=head2 WIN32 PLATFORM LIMITATIONS AND WORKAROUNDS
5192
5193=head3 General issues
5194
5195Win32 doesn't support any of the standards (e.g. POSIX) that libev
5196requires, and its I/O model is fundamentally incompatible with the POSIX
5197model. Libev still offers limited functionality on this platform in
5198the form of the C<EVBACKEND_SELECT> backend, and only supports socket
5199descriptors. This only applies when using Win32 natively, not when using
5200e.g. cygwin. Actually, it only applies to the microsofts own compilers,
5201as every compiler comes with a slightly differently broken/incompatible
5202environment.
5203
5204Lifting these limitations would basically require the full
5205re-implementation of the I/O system. If you are into this kind of thing,
5206then note that glib does exactly that for you in a very portable way (note
5207also that glib is the slowest event library known to man).
5208
5209There is no supported compilation method available on windows except
5210embedding it into other applications.
5211
5212Sensible signal handling is officially unsupported by Microsoft - libev
5213tries its best, but under most conditions, signals will simply not work.
5214
5215Not a libev limitation but worth mentioning: windows apparently doesn't
5216accept large writes: instead of resulting in a partial write, windows will
5217either accept everything or return C<ENOBUFS> if the buffer is too large,
5218so make sure you only write small amounts into your sockets (less than a
5219megabyte seems safe, but this apparently depends on the amount of memory
5220available).
5221
5222Due to the many, low, and arbitrary limits on the win32 platform and
5223the abysmal performance of winsockets, using a large number of sockets
5224is not recommended (and not reasonable). If your program needs to use
5225more than a hundred or so sockets, then likely it needs to use a totally
5226different implementation for windows, as libev offers the POSIX readiness
5227notification model, which cannot be implemented efficiently on windows
5228(due to Microsoft monopoly games).
5229
5230A typical way to use libev under windows is to embed it (see the embedding
5231section for details) and use the following F<evwrap.h> header file instead
5232of F<ev.h>:
5233
5234   #define EV_STANDALONE              /* keeps ev from requiring config.h */
5235   #define EV_SELECT_IS_WINSOCKET 1   /* configure libev for windows select */
5236
5237   #include "ev.h"
5238
5239And compile the following F<evwrap.c> file into your project (make sure
5240you do I<not> compile the F<ev.c> or any other embedded source files!):
5241
5242   #include "evwrap.h"
5243   #include "ev.c"
5244
5245=head3 The winsocket C<select> function
5246
5247The winsocket C<select> function doesn't follow POSIX in that it
5248requires socket I<handles> and not socket I<file descriptors> (it is
5249also extremely buggy). This makes select very inefficient, and also
5250requires a mapping from file descriptors to socket handles (the Microsoft
5251C runtime provides the function C<_open_osfhandle> for this). See the
5252discussion of the C<EV_SELECT_USE_FD_SET>, C<EV_SELECT_IS_WINSOCKET> and
5253C<EV_FD_TO_WIN32_HANDLE> preprocessor symbols for more info.
5254
5255The configuration for a "naked" win32 using the Microsoft runtime
5256libraries and raw winsocket select is:
5257
5258   #define EV_USE_SELECT 1
5259   #define EV_SELECT_IS_WINSOCKET 1   /* forces EV_SELECT_USE_FD_SET, too */
5260
5261Note that winsockets handling of fd sets is O(n), so you can easily get a
5262complexity in the O(n²) range when using win32.
5263
5264=head3 Limited number of file descriptors
5265
5266Windows has numerous arbitrary (and low) limits on things.
5267
5268Early versions of winsocket's select only supported waiting for a maximum
5269of C<64> handles (probably owning to the fact that all windows kernels
5270can only wait for C<64> things at the same time internally; Microsoft
5271recommends spawning a chain of threads and wait for 63 handles and the
5272previous thread in each. Sounds great!).
5273
5274Newer versions support more handles, but you need to define C<FD_SETSIZE>
5275to some high number (e.g. C<2048>) before compiling the winsocket select
5276call (which might be in libev or elsewhere, for example, perl and many
5277other interpreters do their own select emulation on windows).
5278
5279Another limit is the number of file descriptors in the Microsoft runtime
5280libraries, which by default is C<64> (there must be a hidden I<64>
5281fetish or something like this inside Microsoft). You can increase this
5282by calling C<_setmaxstdio>, which can increase this limit to C<2048>
5283(another arbitrary limit), but is broken in many versions of the Microsoft
5284runtime libraries. This might get you to about C<512> or C<2048> sockets
5285(depending on windows version and/or the phase of the moon). To get more,
5286you need to wrap all I/O functions and provide your own fd management, but
5287the cost of calling select (O(n²)) will likely make this unworkable.
5288
5289=head2 PORTABILITY REQUIREMENTS
5290
5291In addition to a working ISO-C implementation and of course the
5292backend-specific APIs, libev relies on a few additional extensions:
5293
5294=over 4
5295
5296=item C<void (*)(ev_watcher_type *, int revents)> must have compatible
5297calling conventions regardless of C<ev_watcher_type *>.
5298
5299Libev assumes not only that all watcher pointers have the same internal
5300structure (guaranteed by POSIX but not by ISO C for example), but it also
5301assumes that the same (machine) code can be used to call any watcher
5302callback: The watcher callbacks have different type signatures, but libev
5303calls them using an C<ev_watcher *> internally.
5304
5305=item pointer accesses must be thread-atomic
5306
5307Accessing a pointer value must be atomic, it must both be readable and
5308writable in one piece - this is the case on all current architectures.
5309
5310=item C<sig_atomic_t volatile> must be thread-atomic as well
5311
5312The type C<sig_atomic_t volatile> (or whatever is defined as
5313C<EV_ATOMIC_T>) must be atomic with respect to accesses from different
5314threads. This is not part of the specification for C<sig_atomic_t>, but is
5315believed to be sufficiently portable.
5316
5317=item C<sigprocmask> must work in a threaded environment
5318
5319Libev uses C<sigprocmask> to temporarily block signals. This is not
5320allowed in a threaded program (C<pthread_sigmask> has to be used). Typical
5321pthread implementations will either allow C<sigprocmask> in the "main
5322thread" or will block signals process-wide, both behaviours would
5323be compatible with libev. Interaction between C<sigprocmask> and
5324C<pthread_sigmask> could complicate things, however.
5325
5326The most portable way to handle signals is to block signals in all threads
5327except the initial one, and run the signal handling loop in the initial
5328thread as well.
5329
5330=item C<long> must be large enough for common memory allocation sizes
5331
5332To improve portability and simplify its API, libev uses C<long> internally
5333instead of C<size_t> when allocating its data structures. On non-POSIX
5334systems (Microsoft...) this might be unexpectedly low, but is still at
5335least 31 bits everywhere, which is enough for hundreds of millions of
5336watchers.
5337
5338=item C<double> must hold a time value in seconds with enough accuracy
5339
5340The type C<double> is used to represent timestamps. It is required to
5341have at least 51 bits of mantissa (and 9 bits of exponent), which is
5342good enough for at least into the year 4000 with millisecond accuracy
5343(the design goal for libev). This requirement is overfulfilled by
5344implementations using IEEE 754, which is basically all existing ones.
5345
5346With IEEE 754 doubles, you get microsecond accuracy until at least the
5347year 2255 (and millisecond accuracy till the year 287396 - by then, libev
5348is either obsolete or somebody patched it to use C<long double> or
5349something like that, just kidding).
5350
5351=back
5352
5353If you know of other additional requirements drop me a note.
5354
5355
5356=head1 ALGORITHMIC COMPLEXITIES
5357
5358In this section the complexities of (many of) the algorithms used inside
5359libev will be documented. For complexity discussions about backends see
5360the documentation for C<ev_default_init>.
5361
5362All of the following are about amortised time: If an array needs to be
5363extended, libev needs to realloc and move the whole array, but this
5364happens asymptotically rarer with higher number of elements, so O(1) might
5365mean that libev does a lengthy realloc operation in rare cases, but on
5366average it is much faster and asymptotically approaches constant time.
5367
5368=over 4
5369
5370=item Starting and stopping timer/periodic watchers: O(log skipped_other_timers)
5371
5372This means that, when you have a watcher that triggers in one hour and
5373there are 100 watchers that would trigger before that, then inserting will
5374have to skip roughly seven (C<ld 100>) of these watchers.
5375
5376=item Changing timer/periodic watchers (by autorepeat or calling again): O(log skipped_other_timers)
5377
5378That means that changing a timer costs less than removing/adding them,
5379as only the relative motion in the event queue has to be paid for.
5380
5381=item Starting io/check/prepare/idle/signal/child/fork/async watchers: O(1)
5382
5383These just add the watcher into an array or at the head of a list.
5384
5385=item Stopping check/prepare/idle/fork/async watchers: O(1)
5386
5387=item Stopping an io/signal/child watcher: O(number_of_watchers_for_this_(fd/signal/pid % EV_PID_HASHSIZE))
5388
5389These watchers are stored in lists, so they need to be walked to find the
5390correct watcher to remove. The lists are usually short (you don't usually
5391have many watchers waiting for the same fd or signal: one is typical, two
5392is rare).
5393
5394=item Finding the next timer in each loop iteration: O(1)
5395
5396By virtue of using a binary or 4-heap, the next timer is always found at a
5397fixed position in the storage array.
5398
5399=item Each change on a file descriptor per loop iteration: O(number_of_watchers_for_this_fd)
5400
5401A change means an I/O watcher gets started or stopped, which requires
5402libev to recalculate its status (and possibly tell the kernel, depending
5403on backend and whether C<ev_io_set> was used).
5404
5405=item Activating one watcher (putting it into the pending state): O(1)
5406
5407=item Priority handling: O(number_of_priorities)
5408
5409Priorities are implemented by allocating some space for each
5410priority. When doing priority-based operations, libev usually has to
5411linearly search all the priorities, but starting/stopping and activating
5412watchers becomes O(1) with respect to priority handling.
5413
5414=item Sending an ev_async: O(1)
5415
5416=item Processing ev_async_send: O(number_of_async_watchers)
5417
5418=item Processing signals: O(max_signal_number)
5419
5420Sending involves a system call I<iff> there were no other C<ev_async_send>
5421calls in the current loop iteration and the loop is currently
5422blocked. Checking for async and signal events involves iterating over all
5423running async watchers or all signal numbers.
5424
5425=back
5426
5427
5428=head1 PORTING FROM LIBEV 3.X TO 4.X
5429
5430The major version 4 introduced some incompatible changes to the API.
5431
5432At the moment, the C<ev.h> header file provides compatibility definitions
5433for all changes, so most programs should still compile. The compatibility
5434layer might be removed in later versions of libev, so better update to the
5435new API early than late.
5436
5437=over 4
5438
5439=item C<EV_COMPAT3> backwards compatibility mechanism
5440
5441The backward compatibility mechanism can be controlled by
5442C<EV_COMPAT3>. See L</"PREPROCESSOR SYMBOLS/MACROS"> in the L</EMBEDDING>
5443section.
5444
5445=item C<ev_default_destroy> and C<ev_default_fork> have been removed
5446
5447These calls can be replaced easily by their C<ev_loop_xxx> counterparts:
5448
5449   ev_loop_destroy (EV_DEFAULT_UC);
5450   ev_loop_fork (EV_DEFAULT);
5451
5452=item function/symbol renames
5453
5454A number of functions and symbols have been renamed:
5455
5456  ev_loop         => ev_run
5457  EVLOOP_NONBLOCK => EVRUN_NOWAIT
5458  EVLOOP_ONESHOT  => EVRUN_ONCE
5459
5460  ev_unloop       => ev_break
5461  EVUNLOOP_CANCEL => EVBREAK_CANCEL
5462  EVUNLOOP_ONE    => EVBREAK_ONE
5463  EVUNLOOP_ALL    => EVBREAK_ALL
5464
5465  EV_TIMEOUT      => EV_TIMER
5466
5467  ev_loop_count   => ev_iteration
5468  ev_loop_depth   => ev_depth
5469  ev_loop_verify  => ev_verify
5470
5471Most functions working on C<struct ev_loop> objects don't have an
5472C<ev_loop_> prefix, so it was removed; C<ev_loop>, C<ev_unloop> and
5473associated constants have been renamed to not collide with the C<struct
5474ev_loop> anymore and C<EV_TIMER> now follows the same naming scheme
5475as all other watcher types. Note that C<ev_loop_fork> is still called
5476C<ev_loop_fork> because it would otherwise clash with the C<ev_fork>
5477typedef.
5478
5479=item C<EV_MINIMAL> mechanism replaced by C<EV_FEATURES>
5480
5481The preprocessor symbol C<EV_MINIMAL> has been replaced by a different
5482mechanism, C<EV_FEATURES>. Programs using C<EV_MINIMAL> usually compile
5483and work, but the library code will of course be larger.
5484
5485=back
5486
5487
5488=head1 GLOSSARY
5489
5490=over 4
5491
5492=item active
5493
5494A watcher is active as long as it has been started and not yet stopped.
5495See L</WATCHER STATES> for details.
5496
5497=item application
5498
5499In this document, an application is whatever is using libev.
5500
5501=item backend
5502
5503The part of the code dealing with the operating system interfaces.
5504
5505=item callback
5506
5507The address of a function that is called when some event has been
5508detected. Callbacks are being passed the event loop, the watcher that
5509received the event, and the actual event bitset.
5510
5511=item callback/watcher invocation
5512
5513The act of calling the callback associated with a watcher.
5514
5515=item event
5516
5517A change of state of some external event, such as data now being available
5518for reading on a file descriptor, time having passed or simply not having
5519any other events happening anymore.
5520
5521In libev, events are represented as single bits (such as C<EV_READ> or
5522C<EV_TIMER>).
5523
5524=item event library
5525
5526A software package implementing an event model and loop.
5527
5528=item event loop
5529
5530An entity that handles and processes external events and converts them
5531into callback invocations.
5532
5533=item event model
5534
5535The model used to describe how an event loop handles and processes
5536watchers and events.
5537
5538=item pending
5539
5540A watcher is pending as soon as the corresponding event has been
5541detected. See L</WATCHER STATES> for details.
5542
5543=item real time
5544
5545The physical time that is observed. It is apparently strictly monotonic :)
5546
5547=item wall-clock time
5548
5549The time and date as shown on clocks. Unlike real time, it can actually
5550be wrong and jump forwards and backwards, e.g. when you adjust your
5551clock.
5552
5553=item watcher
5554
5555A data structure that describes interest in certain events. Watchers need
5556to be started (attached to an event loop) before they can receive events.
5557
5558=back
5559
5560=head1 AUTHOR
5561
5562Marc Lehmann <libev@schmorp.de>, with repeated corrections by Mikael
5563Magnusson and Emanuele Giaquinta, and minor corrections by many others.
5564