Event - Event loop processing

 use Event qw(loop unloop);

 # initialize application
 Event->flavor(attribute => value, ...);

 my $ret = loop();

 # and some callback will call
 unloop('ok');

The Event module provide a central facility to watch for various types of events and invoke a callback when these events occur. The idea is to delay the handling of events so that they may be dispatched in priority order when it is safe for callbacks to execute.

Events (in the ordinary sense of the word) are detected by watchers, which reify them as events (in the special Event module sense). For clarity, the former type of events may be called source events, and the latter target events. Source events, such as signals arriving, happen whether or not they are being watched. If a source event occurs which a watcher is actively watching then the watcher generates a corresponding target event. Target events are only created by watchers. If several watchers are interested in the same source event then each will generate their own target event. Hence, any particular source event may result in zero, one, two, or any number of target events: the same as the number of watchers which were actively watching for it.

Target events are queued to be processed in priority order (priority being determined by the creating watcher) and in FIFO order among events of the same priority. Queued (pending) events can, in some cases, be cancelled before being processed. A queued event is processed by being passed to the callback function (or method on a particular object or class) which was specified to the watcher.

A watcher, once created, operates autonomously without the Event user having to retain any reference to it. However, keeping a reference makes it possible to modify most of the watcher's characteristics. A watcher can be switched between active and inactive states. When inactive, it does not generate target events.

Some types of source event are not reified as target events immediately. Signals received, for example, are counted initially. The counted signals are reified at certain execution points. Hence, signal events may be processed out of order, and if handled carelessly, on the wrong side of a state change in event handling. A useful way to view this is that occurrence of the source event is not actually the arrival of the signal but is triggered by the counting of the signal.

Reification can be forced when necessary. The schedule on which some other events are created is non-obvious. This is especially the case with watchers that watch for a condition rather than an event. In some cases, target events are generated on a schedule that depends on the operation of the event loop.

Events (the occurrence of such) are noticed and queued by 'event watchers'. The creation and configuration of event watchers is the primary topic of the rest of this document.

The following functions control or interrogate the event loop as a whole:

$result = loop([$timeout])

Will enter a loop that calls one_event() until unloop() is called. The argument passed to unloop() is the return value of loop(). Loops can be nested.

unloop($result)

Make the inner-most loop() return with CW$result.

unloop_all($result)

Cause all pending loop()s to return immediately. This is not implemented with CWdie. It is works as if CWunloop($result) were called for all nested loops.

sweep([$max_prio])

Queue all pending events and dispatch any with priority strictly less than CW$max_prio (the highest priority is 0). The default is to process all events except idle events. (While idle events are ignored by sweep, idle watchers are not ignored. If you want to avoid triggering an idle watcher then set CWmax to CWundef or CWstop() it.)

one_event([$timeout])

If any events are outstanding then invoke the corresponding callback of the highest priority event. If there are no events available, block forever or until CW$timeout. Use of this API is not recommended because it is not efficient and does not trap exceptions. However, you might wish to understand how it works:

1

Queue asyncronous events (signals, etc). That is, previously recorded events are reified.

2

If there are any events with priority 5 or less (see StarvePrio) then service the next one and return.

3

Calculate the maximum wait time (minimum time till the next timer expiration) and pass control to the poll/select system call. Upon return, queue all pending events.

4

Queue asyncronous events again.

5

If there are any events then service the next one and return.

6

Service the next idle watcher. StarvePrio is the priority level for which events are dispatched during step 2. It cannot be changed without a recompile. In the rare case that an event is always pending at step 2 then I/O watchers will starve. However, this is highly unlikely since async watchers should never queue events so rapidly.

all_watchers()

Returns a list of all watchers (including stopped watchers).

all_running()

Returns a list of all watchers with actively running callbacks. Watchers are returned in order of most recent to least recent.

all_idle()

Returns a list of all the idle watchers. If the event queue is very busy, all the idle watchers will sit on the idle queue waiting to run. However, be aware that if an idle watcher has the CWmax attribute set then it will queue a normal event when its CWmax wait time is exceeded.

queue_pending()

Examines asynchronous source events (timers & signals) and reifies them as target events. CWqueue_pending() is only called implicitly by CWsweep() and CWone_event(). Otherwise, CWqueue_pending() is not called implicitly. NOTE: Signal watchers generate target events according to which watchers are active at the time that CWqueue_pending() is called rather than according to the time the signal is received. This is best explained by example. See the file CWdemo/queue_pending.t.

All watchers are constructed in one of the following ways:

  $w = Event->flavor( [attr1 => $value,]... );

  $w = Event::flavor($Class, [attr1 => $value,]...);

  $w = Event::flavor->new([attr1 => $value,]...);

Where flavor is substituted with the kind of watcher. Built-in types include idle, io, signal, timer, and var.

New watchers (hopefully) have reasonable defaults and can also be customized by passing extra attributes to the constructor. When created, watcher objects are started and are waiting for events (see CW$event->start below).

NetServer::Portal can display watchers in real-time, formatted similarly to the popular CWtop program. You may find this a useful aide for debugging.

Watchers are configured with attributes (also known as properties). For example:

   $watcher->cb(\&some_code);   # set callback

   warn $event->w->desc.": ".$event->hits." events happened; Wow!";

All watchers support the following attributes: cb, cbtime, debug, desc, prio, max_cb_tm, reentrant, and repeat. Watcher constructors accept the preceding and additionally: async and nice. Moreover, watchers also offer extra attributes according to their specialty.

The following methods are available for all watchers:

$watcher->start

Activate the watcher. Watchers refuse to CWstart() without sufficient configuration information to generate events. Constructors always invoke CWstart() unless the CWparked=>1 option is requested. You will need to set the parked option if you preallocate unconfigured watchers. Note: If there are any unreified asynchronous events that are of interest to the watcher, it will see these events even though they happened before it was started. This affects signal watchers, but there will only be existing unreified signal events if Event was already handling the signal, which it would only do if there were another active watcher for the same signal. If this situation might occur, and it would be a problem for the new watcher to see older events, call CWqueue_pending() immediately before starting the new watcher in order to reify any outstanding events. This explaination may be more clear if read along with CWdemo/queue_pending.t.

$watcher->again

This is the same as the CWstart except if a watcher has special repeat behavior. For example, repeating timers recalculate their alarm time using the CWinterval parameter.

$watcher->now

Cause the watcher to generate an event, even if it is stopped. The callback may or may not run immediately depending upon the event's priority. If you must unconditionally invoke the callback, consider something like

  $w->cb->($w);

$watcher->stop

Don't look for events any more. Running events are allowed to complete but pending events are cancelled. Note that a stopped watcher can be reactivated by calling the CWstart or CWagain methods. Watchers are stopped implicitly if their new configuration deprives them of the ability to generate events. For instance:

  my $io_watcher = Event->io(timeout => 1);  # started
  $io_watcher->timeout(undef);               # stopped implicitly
  $io_watcher->timeout(1);                   # still stopped
  $io_watcher->start;                        # restarted

$watcher->cancel

Stop and destroy CW$watcher. Running events are allowed to complete but pending events are cancelled. Cancelled watchers are no longer valid except for read-only operations. For example, prio() can return the watcher's priority, but start() will fail.

$watcher->is_cancelled

Reports whether the CW$watcher has been cancelled.

$watcher->is_active

Reports whether the CW$watcher has been started. The return value is not affected by suspend.

$watcher->is_running

Zero if the callback is not running. Otherwise, the number of levels that the callback has been entered. This can be greater than one if a CWreentrant callback invokes CWloop (or CWsweep, with lesser probability).

$watcher->is_suspended

Reports whether the CW$watcher is suspended. Suspension is a debugging feature; see the discussion of the suspend attribute below.

$watcher->pending

In scalar context, returns a boolean indicating whether this watcher has any events pending in the event queue. In array context, returns a list of all the watcher's pending events. Note that this does not check for unreified asynchronous events. Call CWqueue_pending() first if you want to see signals received since the last operation of the event loop.

idle

Extra attributes: min => CW$seconds, max => CW$seconds Watches for the Event system to be idle, i.e., to have no events pending. If the system is never idle, an event will be generated at least every CWmax seconds. While Event is idle, events will be generated not more often than CWmin seconds. If neither CWmin nor CWmax is specified, the watcher defaults to one-shot behaviour (CWrepeat false), otherwise it defaults to repeating. In either case, the default can be overidden by specifying a CWrepeat attribute. CWmin defaults to 0.01, and CWmax defaults to infinity.

var

Extra attributes: var => \$var, poll => 'rw' Var watchers generate events when the given variable is read from or written to, as specified by CWpoll. CWpoll defaults to w. As perl is a concise language, it is often difficult to predict when a variable will be read. For this reason, variable watchers should poll only for writes unless you know what you are doing.

timer

Extra attributes: at => CW$time, after => CW$sec, interval => CW$sec, hard => CW$bool Generate events at particular times. The CW$time and CW$sec are in seconds. Fractional seconds may be used if Time::HiRes is available. CWat and CWafter are mutually exclusive. CWat or CWafter specify the initial time that the event will trigger. Subsequent timer events occur at intervals specified by CWinterval or CWafter (in that order of preference) if either was supplied. The timer defaults to one-shot behaviour if CWinterval was not specified, or repeating behaviour if CWinterval was specified; in either case this can be overridden by providing CWrepeat explicitly. You need to know the time at the start of today if you are trying to set timers to trigger at day relative times. You can find it with:

  use Time::Local;
  my $TodaySeconds = int timelocal(0,0,0,(localtime)[3,4,5]);

This calculation may seem a little heavy weight. If you want to use UTC rather than local time then you can use this instead:

  my $TodaySeconds = time - time % 86400;

Beware that, due to lags in the event loop, the CWinterval timeout may already be in the past. If the CWhard flag is set, the event will be queued for execution relative to the last time the callback was invoked. However, if CWhard is false the new timeout will be calculated relative to the current time. CWhard defaults to false.

io

Extra attributes: fd => CW$fd, poll => 'rwe' [timeout => CW$seconds, hard => CW$bool, timeout_cb => \&code] The callback is invoked when the file descriptor, CWfd, has data to be read, written, or pending exceptions. CWfd can be a GLOB, an IO::Handle object, or a file number (file descriptor). CWpoll defaults to r. Note that it is your option whether to have multiple watchers per file handle or to use a single watcher for all event conditions. If CWtimeout is set, events are also generated regularly if no actual I/O event occurs. If CWtimeout_cb is set then timeouts use this alternate callback instead of the main callback.

signal

Extra attribute: signal => CW$str Generates events based on signal arrival. The events are not actually generated immediately when the signal arrives: signals received are counted and reified by CWqueue_pending() or implicitly by CWone_event(). Several signals of the same type may be merged into a single event. In such cases, the number of signals represented by a single event is stored in the hits attribute.

Priority is used to sort the event queue. Meaningful priorities range from -1 to 6 inclusive. Lower numbers mean higher priority (-1 is the highest priority and 6 is the lowest). If multiple events get queued, the ones with the highest priority are serviced first. Events with equal priority are serviced in first-in-first-out order.

  use Event qw(PRIO_HIGH PRIO_NORMAL);   # some constants

  LEVELS: -1      0      1      2      3      4      5      6
          ----------------------+-------------+---------------
                            PRIO_HIGH     PRIO_NORMAL

A negative priority causes the callback to be invoked immediately upon event occurrence. Use this with caution. While it may seem advantageous to use negative priorities, they bypass the whole point of having an event queue.

Each watcher has a default priority, assigned by its constructor:

  io       PRIO_NORMAL
  signal   PRIO_HIGH
  timer    PRIO_NORMAL
  var      PRIO_NORMAL

Default priorities are stored in ${Event::${type}::DefaultPriority}. If the default priority is not satisfactory for your purposes, the constructor options CWnice, CWasync, or CWprio can be used to adjust it. CWnice specifies an offset from the default priority; CWasync forces the priority to -1; and CWprio assigns a given priority of your choice. If more than one of these options are given then CWprio overrides CWasync overrides CWnice.

These options are only supported as constructor arguments. See the discussion of the timer watcher. If CW$bool then the watcher priority is set to -1. Offset from the default priority. By default, watcher constructors automatically invoke the CWstart() method. If you don't want the watcher started then request CWparked=>1.

The expiration time in the same units as the system clock. For a timer, CWat will usually be in the future.

cb => \&code

The function or method to call when an event is dispatched. The callback is invoked with CW$event as its only argument. Perhaps you are wondering what happens if something goes wrong and an untrapped CWdie occurs within your callback? CW$Event::DIED is just for this purpose. See the full description of CWDIED below. When the callback was invoked most recently. The CWdata() method associates arbitrary data with a watcher. This method is not intended for implementers of watchers. If you are subclassing or implementing a watcher, consider the CWprivate() method. Debugging can be activated globally or per watcher. When debugging is enabled for a particular watcher, CW$Event::DebugLevel is treated as two levels higher. Levels of 1, 2, 3, or 4 give progressively more diagnostics on STDERR. An identifying name. If this is not passed explicitly to the constructor, it will be initialized with a string that attempts to identify the location in the source code where the watcher was constructed. This attribute can accept either a perl-esque filehandle or a system call derived file descriptor number. Determines how repeating timers (or timeouts) are recalculated. The timer is restarted either before or after the callback depending on whether it is true or false, respectively. In long-running callbacks this can make a significant difference. How long between repeating timeouts. The CWat attribute is recalculated using CWinterval upon callback return. The maximum number of seconds to wait before triggering the callback. Similar to a CWtimeout. The maximum number of seconds to spend in a callback. If a callback uses more time then it is aborted. Defaults to 1 sec. This feature is normally disabled. See Event::Stats. Enforce a minimum number of seconds between triggering events. Determines which kinds of events are of interest. This attribute can be set with either strings or bit constants. The bit constants are available via 'use Event::Watcher qw(R W E T);'.

  string constant description
  ------ -------- ---------------
   'r'     R      read
   'w'     W      write
   'e'     E      exception
   't'     T      timeout

Thus, both of these statements enable interest in read:

  $w->poll($w->poll . 'r');
  $w->poll($w->poll | R);

A given type of watcher may support all or a subset of the available events. Changes the watcher's priority to the given level. Events generated by a watcher usually inherit the priority of the watcher. Use the CWprivate() method to associate arbitrary data with a watcher. This method is intended for implementers of watchers or watcher subclasses. Each caller's package accesses its own private attribute. By default, callbacks are allowed to invoke CWsweep or CWloop which in turn may invoke the same callback again recursively. This can be useful but can also be confusing. Moreover, if you keep reentering callbacks you will quickly run out of stack space. Disable this feature per watcher by setting reentrant to false. This will cause the watcher to be suspended during recursive calls to CWsweep or CWloop. The repeat flag controls whether the callback should either be one-shot or continue waiting for new events. The default setting depends on the type of watcher. io, signal, and var default to true. The callback is invoked after the specified signal is received. The CW$str string should be something like 'INT' or 'QUIT'. Also see the documentation for CW%SIG. A given signal can be handled by CW%SIG or Event, but not both at the same time. Event handles the signal as long as there is at least one active watcher. If all watchers for the signal are cancelled or stopped then Event sets the signal handler to SIG_DFL. Stop looking for events. Running events are allowed to complete, but queued events are cancelled. Suspend is for debugging. If you suspend all watchers in an application then you can examine the complete state unchanged for as long as you like without worrying about timer expirations. If you actually wish to stop a watcher then use the CWstop() method. The number of seconds before a watcher times out.

timeout_cb => \&code

This is an optional attribute for use when it is desired that timeouts be serviced in a separate code path than normal events. When this attribute is unset, timeouts are serviced by CWcb. A reference to the variable being watched.

CWgot is available in the callback of watchers with CWpoll. CWgot is in the same format as CWpoll except that it gives what kind of event actually happened. In contrast, CWpoll is just an indication of interest. Signals in quick succession can be clumped into a single event. The number of signals clumped together is indicated by this attribute. This is always one for event types which don't clump. Be aware that this priority can differ from the watcher's priority. For instance, the watcher's priority may have changed since the event was generated. Moreover, the C extension API offers the freedom to queue events of arbitrary priority. This method return the event's watcher. It is read-only.

Enables progressively more debugging output. Meaningful levels range from 1 (least output) to 5 (most output). Also see CWdebug. When CWloop or CWsweep is called, an exception context is established for the duration of event processing. If an exception is detected then CW$Event::DIED is invoked. The default hook uses CWwarn to output the exception. After the DIED handler completes, event processing continues as if nothing happened. If you'd like more detailed output you can install the verbose handler:

  $Event::DIED = \&Event::verbose_exception_handler;

  $Event::DIED->($event, $@);

  $Event::DIED = sub {
    Event::verbose_exception_handler(@_);
    Event::unloop_all();
  };

* Event->add_hooks(key => sub { ... }, ...);

The bulk of Event's implementation is in C for maximum performance. The CWadd_hooks method allows insertion of perl code at key points in the optimized event processing core. While flexible, this can hurt performance *significantly*. If you want customization *and* performance, please see the C API. Currently support hooks are detailed as follows:

  hook          purpose
  ------------- ----------------------------------------------
  prepare       returns minimum time to block (timeable)
  check         assess state after normal return from select/poll
  asynccheck    check for signals, etc
  callback      invoked before each event callback

Event also has a direct API for callbacks written exclusively in C. See Event::MakeMaker.

Event loops and threads are two different solutions to the same problem: asynchronous processing. Event loops have been around since the beginning of computing. They are well understood and proven to be a good solution for many applications.

While event loops make use of basic operating system services, the bulk of their implementation is usually outside the kernel. While an event loop may appear to do many things in parallel, it does not, even on multiprocessor hardware. Actions are always dispatched sequentially. This implies that long running callbacks must be avoided because otherwise event processing is halted.

Event loops work well when actions are short and to the point. Long-running tasks must be broken into short steps and scheduled for execution. Some sort of a state machine is usually required. While a big, complex application server is usually simpler to implement in a multithreaded fashion, a web browser can easily get by without threads. Consider a JPEG file download and render. When some new bytes are available they are sorted to the right place on the screen. Only a little state must be kept to keep track of how much has been rendered and to process subsequent incoming bytes.

Threads can either substitute for an event loop or complement it. Threads are similar to processes in that the operating system manages task switching for you. However, the difference is that all threads share the same address space. This is good and bad. Higher performance can be achieved but since data is shared between threads, extreme care must be taken to access or modify global data. The operating system can switch threads at any moment or can execute multiple threads simultaneously. I hope this sounds dangerous! It is! Threads can introduce maddeningly complicated and hard to debug synchronization problems.

Threads are like rocket fuel. They are essential when you really need them but most applications would be better off with a simple event loop. Even if threads are genuinely needed, consider confining them to the parts of an application where truly scalable performance is really worth the difficulty of a multithreaded implementation. For example, most GUIs applications do not need threads and most scientific compute intensive problems can be isolated from event dispatching. On the other hand, high performance transaction servers generally do mandate a truly multithreaded approach.

Another consideration is that threads are not quite as widely available as event loops. While a few forward-thinking operating systems have offered threads since the beginning, their addition to many popular operating systems is much more recent and some still offer no threads support. If portability is a requirement, one must check that threads support is available and also carefully test a particular threads implementation to see whether it supports the features you need. It is likely that all platforms will have a solid implementation soon but at this point in history it is best to double check.

Many suggestions by Mark Mielke <Mark.Mielke.markm@nt.com>

The Java language is oriented to use non-preemptive threads, yet even Java uses an event-loop for Swing (AFAIK). That is one of the reasons I don't use Java for network-centric applications. My belief is that the benefit of multi-threading is the gain in performance on SMP hardware. In my view, non-preemptive threads (java green-threads) are usually poor design. I find them harder to work with, harder to debug, and slower for a rather marginal gain in readability. I really like working with a state machine. I find it leads to more stable and better code. It also has the benefit of abstracting away how concurrency is achieved.

Contributed by artur@vogon-solutions.com, 12 Jul 1999.

No support for epoll, or better, libevent.

The scope of events is pretty strange compared to most other perl objects. I'm not sure if this is a bug or a feature (OK, probably it was a mistake). We'll probably want to re-work things for Perl6.

The meaning of CW$io->timeout(0) might change. Use CWundef to unset the timeout.

There seems to be some sort of bug in the global destruction phase:

  Attempt to free unreferenced scalar during global destruction.
  Use of uninitialized value during global destruction.
  Explicit blessing to '' (assuming package main) during global
  destruction.

Even if this module does not end up being the One and True Event Loop, the author will insure that it is source compatible with its successor, or arrange for gradual migration. Back in the early days, the Event programming API was changing at every release. Care was taken to allow the old API to continue to work, and the transition was eased by printing out lots of warnings about the new usage. So you shouldn't sit on your hands in anticipation of the One and True Event Loop. Just start coding!

* Useful and Fun

Time::HiRes, NetServer::Portal, Time::Warp

* Message Passing

COPE, IPC::LDT, Event-tcp

* GUI

While Tk does not yet support Event, PerlQt does.

* C API

Inline

If you have insights or complaints then please subscribe to the mailing list! Send email to:

  perl-loop-subscribe@perl.org

Joshua N. Pritikin <jpritikin@pobox.com>

Initial 0.01 implementation by Graham Barr <gbarr@pobox.com>. Other contributors include at least those lists below and folks mentioned in the ChangeLog.

 Gisle Aas <gisle@aas.no>
 Uri Guttman <uri@sysarch.com>
 Nick Ing-Simmons <nick@ni-s.u-net.com> (Tk)
 Sarathy <gsar@engin.umich.edu>
 Jochen Stenzel <perl@jochen-stenzel.de>

Copyright © 1997 Joshua Nathaniel Pritikin & Graham Barr

Copyright © 1998, 1999, 2000, 2001, 2002, 2003, 2004 Joshua Nathaniel Pritikin

All rights reserved. This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.