AnyEvent::Impl::IOAsync - AnyEvent adaptor for IO::Async
use AnyEvent; use IO::ASync::Loop; use AnyEvent::Impl::IOAsync; my $loop = new IO::Async::Loop; AnyEvent::Impl::IOAsync::set_loop $loop;
This module provides support for IO::Async as AnyEvent backend. Due to the rather sad state of IO::Async, support is only available partially (only timers and I/O watchers are supported, signals and child watchers are emulated by AnyEvent itself (fighting with IO::Async, so you cannot use both), idle watchers are being emulated, I/O watchers need to dup their fh.
There have been multiple attempts at providing an AnyEvent interface to IO::Async, and the effort is ongoing. Supporting IO::Async is hard. Here's why:
- IO::Async integration cannot be automatic - no default loop
IO::Async doesn't offer an interface suitable for independent usage of event sources: there is no standard way to share the main event loop between modules - modules have to somehow agree on how to do this.
For AnyEvent to work with IO::Async, the IO::Async main program has to call
IO::Async::Loopobject that AnyEvent is to use, see the SYNOPSIS section for an example.
It is possible to get a copy of the loop by reading
$AnyEvent::Impl::IOAsync::LOOP, so AnyEvent could be used as a central place to store the default IO::Async::Loop object, also for other modules, if there is any need for that.
- Broken child watchers
IO::Async is the only module that requires you to call a special function before even forking your child program, while the AnyEvent API itself (which works with other event loops) works as long as it has been initialised (AnyEvent's own pure perl implementation doesn't even suffer from these limitations and just works).
Worse, IO::Async does not let you install multiple child watchers, does not let you watch for any child, and apparently an interface to unregister child watchers has been forgotten as well.
As a result, AnyEvent::Impl::IOAsync has to fall back on it's own child management, which makes it impossible to watch for child processes via both AnyEvent and IO::Async. Hooking and Patching IO::Async has been considered, but is considerable work.
- No support for multiple watchers per event
In most (all? documentation?) cases you cannot have multiple watchers for the same event (what's the point of having all these fancy notifier classes when you cannot have multiple notifiers for the same event? That's like only allowing one timer per second or so...).
This makes signal watchers almost useless (You could just hook them yourself, you can't share any of them, as would make sense for e.g. SIGTERM, SIGTSTP, SIGPWR, SIGUSR1 etc.).
As a result, AnyEvent falls back to it's own signal handling (it is pointless to somehow share the IO::Async watcher, as it doesn't matter if AnyEvent blocks the signal via IO::Async or directly, and AnyEvents signal handling is race-free).
For I/O watchers, AnyEvent has to dup() every file handle, as IO::Async fails to support the same or different file handles pointing to the same fd (this is at least documented, but why not fix it instead?).
Apart from these fatal flaws, there are a number of unpleasent properties that just need some mentioning:
- Confusing and misleading name
Another rather negative point about this module family is its name, which is deeply confusing: Despite the "async" in the name, IO::Async only does synchronous I/O, there is nothing "asynchronous" about it whatsoever (when I first heard about it, I thought, "wow, a second async I/O module, what does it do compared to IO::AIO", and was somehow set back when I learned that the only "async" aspect of it is the name).
- Inconsistent, incomplete and convoluted API
Implementing AnyEvent's rather simple timers on top of IO::Async's timers was a nightmare (try implementing a timer with configurable interval and delay value...).
How to actually get I/O events in IO::Async::Handle is undocumented: read events are apparently automatic, for write events, you have to explicitly request
want_writeready, and specifying
want_readreadyis apparently a usage bug (it doesn't exist). All this must be deduced from reading the sources.
You can't stop child watchers. Even reading the sources I found no way to stop them. It must have been forgotten.
The method naming is chaotic:
watch_childcreates a child watcher, but
watch_iois an internal method;
detach_signalremoves a signal watcher, but
detach_childforks a subprocess and so on).
IO::Async has weird checks - passing in a callable reference is sometimes forbidden (of course, this is checked on every invocation, not when the callback is registered, so you have no idea where in your code you passed it in), as the code checks explicitly for code references, disallowing callable objects.
- Unpleasant surprises on GNU/Linux
When you develop your program on FreeBSD and run it on GNU/Linux, you might have unpleasant surprises, as IO::Async::Loop will by default use IO::Async::Loop::Epoll, which is incompatible with
fork, so your network server will run into spurious and very hard to debug problems under heavy load, as IO::Async forks a lot of processes, e.g. for DNS resolution. It would be better if IO::Async would only load "safe" backends by default (or fix the epoll backend to work in the presence of fork, which admittedly is hard - EV does it for you, and also does not use unsafe backends by default).
- Exiting considered harmful
(in cleanup) Can't call method "parent" on an undefined value at IO/Async/Loop.pm line 297 during global destruction.
IO::Async just hates global destruction. Calling
exitwill easily give you one such line per watcher.
The problem is that IO::Async::Loop is itself not warning-free, but actually enables warnings for itself.
(Ok, the real bug is of course perl's broken mark & sweep garbage collector that corrupts data structures).
On the positive side, performance with IO::Async is quite good even in my very demanding eyes.
Marc Lehmann <email@example.com> http://home.schmorp.de/