NAME
Callback::Frame - Preserve error handlers and "local" variables across callbacks
SYNOPSIS
use Callback::Frame;
my $callback;
frame(name => "base frame",
code => sub {
$callback = frame(name => "frame #1",
code => sub {
die "some error";
});
},
catch => sub {
my $stack_trace = shift;
print $stack_trace;
## Also, $@ is set to "some error at ..."
}
)->();
$callback->();
This will print something like:
some error at synopsis.pl line 9.
----- Callback::Frame stack-trace -----
synopsis.pl:10 - frame #1
synopsis.pl:17 - base frame
BACKGROUND
When programming with callbacks in perl, you create anonymous functions with sub { ... }
. These functions are especially useful because when they are called they will preserve their surrounding lexical environment.
In other words, the following bit of code
my $callback;
{
my $var = 123;
$callback = sub { $var };
}
print $callback->();
will print 123
even though $var
is no longer in scope when the callback is invoked.
Sometimes people call these anonymous functions that reference variables in their surrounding lexical scope "closures". Whatever you call them, they are essential for convenient and efficient asynchronous programming.
DESCRIPTION
The problem that this module solves is that although closures preserve their lexical environment, they don't preserve their dynamic environment.
Consider the following piece of broken code:
use AnyEvent;
eval { ## broken!
$watcher = AE::timer 0.1, 0,
sub {
die "some error";
};
};
if ($@) { ## broken!
print STDERR "Oops: $@";
}
AE::cv->recv;
The intent behind the eval
above is obviously to catch any exceptions thrown inside the callback. However, this will not work because the eval
will only be in effect while installing the callback in the event loop, not while running it. When the event loop calls the callback, it will probably wrap its own eval
around the callback and you will see something like this:
EV: error in callback (ignoring): some error at broken.pl line 6.
(The above applies to EV which is a well-designed event loop. Other event loops may fail more catastrophically.)
So the root of the problem is that the dynamic environment has not been preserved. In this case it is the dynamic exception handlers that were not preserved. In some other cases we would like to preserve dynamically scoped (aka "local") variables (see below).
By the way, "lexical" and "dynamic" are the lisp terms and I use them because they actually make sense. When it applies to variables, perl confusingly calls dynamic scoping "local" scoping.
Here is how we could fix this code using Callback::Frame:
use AnyEvent;
use Callback::Frame;
frame(
code => sub {
$watcher = AE::timer 0.1, 0,
frame(code => sub {
die "some error";
});
}, catch => sub {
print STDERR "Oops: $@";
}
)->();
AE::cv->recv;
Now we see the desired error message:
Oops: some error at fixed.pl line 8.
We created two framed to accomplish this: A root frame to contain the exception handler, and another frame to use as a callback. Because the callback frame is created while the root frame is running, the callback will preserve the dynamic environment of the root frame.
USAGE
This module exports two functions, frame
and fub
.
frame
is the general interface and requires at least a code
argument which should be a coderef (a function or a closure). It will return another coderef that "wraps" the coderef you passed in. When this codref is run, it will re-instate the dynamic environment that was present when the frame was created, and then run the coderef that you passed in as code
.
IMPORTANT NOTE: All callbacks that may be invoked outside the dynamic environment of the current frame should be wrapped in a new frame
or fub
call so that the dynamic environment will be correctly re-applied when the callback is invoked.
fub
is a wrapper around frame and exists to simplify converting existing callbacks into Callback::Frame enabled code. For example, given the following AnyEvent statement:
$watcher = AE::io $sock, 0, sub { do_stuff() };
In order for the callback to have its dynamic environment maintained, you just need to change it to this:
$watcher = AE::io $sock, 0, fub { do_stuff() };
frame
and fub
both also accept catch
and local
parameters which are described in detail below.
If you wish to run a coderef inside an existing frame's dynamic environment, when creating a frame you can pass in an existing frame as the existing_frame
parameter. When this frame is executed, the code
of the frame will be run inside existing_frame
's dynamic environment. This is useful for throwing exceptions from within some given callback's environment (timeouts for example) and for extracting/setting a callback's local variables.
Libraries that wrap callbacks in frames can use the Callback::Frame::is_frame()
function to determine if a given callback is already wrapped in a frame. It returns true if the callback is wrapped in a frame and is therefore suitable for use with existing_frame
.
You should never need to, but the internal frame stack can be accessed at $Callback::Frame::top_of_stack
. When this variable is defined, a frame is currently being executed.
NESTING AND STACK-TRACES
Frames can be nested. When an exception is raised, the most deeply nested catch
handler is invoked. If this handler itself throws an error, the next most deeply nested handler is invoked with the new exception but the original stack trace. If the last catch
handler re-throws the error, the error will simply be thrown to your application from wherever the callback happened to be called (probably not what you want).
When a catch
handler is called, not only is $@
set, but also a stack-trace string is passed in as the first argument. All frames will be listed in this stack-trace, starting with the most deeply nested frame.
If you want you can use simple frame names like "accepted"
but if you are recording error messages in a log you might find it useful to name your frames things like "accepted connection from $ip:$port at $time"
and "connecting to $host (timeout = $timeout seconds)"
.
All frames you omit the name from will be shown as "ANONYMOUS FRAME"
in stack-traces.
One important thing to keep in mind is that frames aren't necessarily always arranged in terms of a stack. Really, frames are more of a tree data structure (known in lisp as a "spaghetti stack"). This occurs most often when two asynchronous request frames are started up concurrently while the same frame is in effect. At this point the "stack" has essentially branched. If you are ever surprised by an exception handler being called twice, this is probably what is happening.
"LOCAL" VARIABLES
In the same way that using catch
as described above preserves the dynamic environment of error handlers, local
preserves the dynamic environment of variables. Of course, the scope of these bindings is not actually local in the real sense of the word, only in the perl sense.
Technically, local
maintains the dynamic environment of bindings. Although perl coders often ignore the distinction, lisp coders are careful to distinguish between variables and bindings. See, when a lexical binding is created, it is there "forever" -- or at least until it is no longer reachable by your program according to the rules of lexical scoping. Therefore, lexical variables are statically mapped to bindings and it is redundant to distinguish between lexical variables and bindings.
However, with dynamic variables the same variable can refer to different bindings at different times. That's why they are called "dynamic" variables and lexical variables are called "static" variables.
Because any code in any file, function, or package can access a dynamic variable, they are the opposite of local. They are global. However, the bindings are only global for a little while at a time. After a while they will go out of scope and then they are no longer visible at all. Or sometimes they will get "shadowed" by some other binding and will come back again later.
OK, to make this concrete, after running this bit of code the binding containing 2
is lost forever:
our $foo = 1;
my $cb;
{
local $foo = 2;
$cb = sub {
return $foo;
};
}
say $foo; # 1
say $cb->(); # 1 <- not 2!
say $foo; # 1
Here's a way to "fix" that using Callback::Frame:
our $foo = 1;
my $cb;
frame(local => __PACKAGE__ . '::foo',
code => sub {
$foo = 2;
$cb = frame(code => sub {
return $foo;
});
})->();
say $foo; # 1
say $cb->(); # 2 <- hooray!
say $foo; # 1
Don't be fooled into thinking that this is a lexical binding though. While the variable $foo
points to the binding containing 2
, all parts of the program will see this binding as demonstrated in the following example:
our $foo = 1;
my $cb;
sub global_foo_getter {
return $foo;
}
frame(local => __PACKAGE__ . '::foo',
code => sub {
$foo = 2;
$cb = frame(code => sub {
return global_foo_getter();
});
})->();
say $foo; # 1
say $cb->(); # 2 <- still 2
$foo; # 1
You can install multiple local variables in the same frame:
frame(local => __PACKAGE__ . '::foo',
local => 'main::bar',
... );
Note that if you have both catch
and local
elements in a frame, in the event of an error the local bindings will not be present inside the catch
handler (use a nested frame if you need this).
All variable names must be fully package qualified. The easiest way to do this for variables in your current package is to use the ugly __PACKAGE__ .
technique.
Objects stored in local bindings managed by Callback::Frame will not be destroyed until all references to the frame-wrapped callback that contains the binding, along with all references to any deeper frames, are freed.
SEE ALSO
The Callback::Frame github repo
The Callback::Frame philosophy is that we actually like callback style and just want to simplify the reseting of dynamic environments as necessary. This module should make adding error handling support to an existing asynchronous application as easy as possible (it shouldn't force you to pass extra parameters around). Finally, this module should make lifer easier because as a side effect of adding error checking it should also produce detailed and useful stack traces.
This module's catch
syntax is of course modeled after "normal language" style exception handling as implemented by Try::Tiny &c.
This module depends on Guard
to maintain the $Callback::Frame::active_frames
datastructure, and to ensure that even when exceptions are thrown, local
binding updates aren't lost.
AnyEvent::Debug is a very interesting and useful module that provides an interactive debugger for AnyEvent applications and uses some of the same techniques that Callback::Frame does. AnyEvent::Callback and AnyEvent::CallbackStack sort of solve the dynamic error handler problem. Unlike these modules, Callback::Frame is not related at all to AnyEvent, except that it happens to be useful in AnyEvent libraries and applications (among other things).
Miscellaneous other modules: IO::Lambda::Backtrace, POE::Filter::ErrorProof
Python Tornado's StackContext and async_callback
UNWIND-PROTECT vs. Continuations
BUGS
For now, local
bindings can only be created in the scalar namespace. Also, none of the other nifty things that local can do (like localising a hash table value) are supported yet.
The local
implementation is currently pretty inefficient. It evals a string that returns a sub whenever a frame is entered. Fortunately, perl's compiler is really fast. Also, this overhead is not there at all when you are only using the catch
functionality which I anticipate to be the most common usage pattern.
AUTHOR
Doug Hoyte, <doug@hcsw.org>
COPYRIGHT & LICENSE
Copyright 2012 Doug Hoyte.
This module is licensed under the same terms as perl itself.