Guard/Guard.pm

=head1 NAME

Guard - safe cleanup blocks

=head1 SYNOPSIS

   use Guard;
   
   # temporarily chdir to "/etc" directory, but make sure
   # to go back to "/" no matter how myfun exits:
   sub myfun {
      scope_guard { chdir "/" };
      chdir "/etc";
   
      code_that_might_die_or_does_other_fun_stuff;
   }

   # create an object that, when the last reference to it is gone,
   # invokes the given codeblock:
   my $guard = guard { print "destroyed!\n" };
   undef $guard; # probably destroyed here

=head1 DESCRIPTION

This module implements so-called "guards". A guard is something (usually
an object) that "guards" a resource, ensuring that it is cleaned up when
expected.

Specifically, this module supports two different types of guards: guard
objects, which execute a given code block when destroyed, and scoped
guards, which are tied to the scope exit.

=head1 FUNCTIONS

This module currently exports the C<scope_guard> and C<guard> functions by
default.

=over 4

=cut

package Guard;

no warnings;

BEGIN {
   $VERSION = '1.021';
   @ISA = qw(Exporter);
   @EXPORT = qw(guard scope_guard);

   require Exporter;

   require XSLoader;
   XSLoader::load Guard, $VERSION;
}

our $DIED = sub { warn "$@" };

=item scope_guard BLOCK

Registers a block that is executed when the current scope (block,
function, method, eval etc.) is exited.

See the EXCEPTIONS section for an explanation of how exceptions
(i.e. C<die>) are handled inside guard blocks.

The description below sounds a bit complicated, but that's just because
C<scope_guard> tries to get even corner cases "right": the goal is to
provide you with a rock solid clean up tool.

The behaviour is similar to this code fragment:

   eval ... code following scope_guard ...
   {
      local $@;
      eval BLOCK;
      eval { $Guard::DIED->() } if $@;
   }
   die if $@;

Except it is much faster, and the whole thing gets executed even when the
BLOCK calls C<exit>, C<goto>, C<last> or escapes via other means.

If multiple BLOCKs are registered to the same scope, they will be executed
in reverse order. Other scope-related things such as C<local> are managed
via the same mechanism, so variables C<local>ised I<after> calling
C<scope_guard> will be restored when the guard runs.

Example: temporarily change the timezone for the current process,
ensuring it will be reset when the C<if> scope is exited:

   use Guard;
   use POSIX ();

   if ($need_to_switch_tz) {
      # make sure we call tzset after $ENV{TZ} has been restored
      scope_guard { POSIX::tzset };

      # localise after the scope_guard, so it gets undone in time
      local $ENV{TZ} = "Europe/London";
      POSIX::tzset;

      # do something with the new timezone
   }

=item my $guard = guard BLOCK

Behaves the same as C<scope_guard>, except that instead of executing
the block on scope exit, it returns an object whose lifetime determines
when the BLOCK gets executed: when the last reference to the object gets
destroyed, the BLOCK gets executed as with C<scope_guard>.

See the EXCEPTIONS section for an explanation of how exceptions
(i.e. C<die>) are handled inside guard blocks.

Example: acquire a Coro::Semaphore for a second by registering a
timer. The timer callback references the guard used to unlock it
again. (Please ignore the fact that C<Coro::Semaphore> has a C<guard>
method that does this already):

   use Guard;
   use Coro::AnyEvent;
   use Coro::Semaphore;

   my $sem = new Coro::Semaphore;

   sub lock_for_a_second {
      $sem->down;
      my $guard = guard { $sem->up };

      Coro::AnyEvent::sleep 1;

      # $sem->up gets executed when returning
   }

The advantage of doing this with a guard instead of simply calling C<<
$sem->down >> in the callback is that you can opt not to create the timer,
or your code can throw an exception before it can create the timer (or
the thread gets canceled), or you can create multiple timers or other
event watchers and only when the last one gets executed will the lock be
unlocked. Using the C<guard>, you do not have to worry about catching all
the places where you have to unlock the semaphore.

=item $guard->cancel

Calling this function will "disable" the guard object returned by the
C<guard> function, i.e. it will free the BLOCK originally passed to
C<guard >and will arrange for the BLOCK not to be executed.

This can be useful when you use C<guard> to create a cleanup handler to be
called under fatal conditions and later decide it is no longer needed.

=cut

1;

=back

=head1 EXCEPTIONS

Guard blocks should not normally throw exceptions (that is, C<die>). After
all, they are usually used to clean up after such exceptions. However,
if something truly exceptional is happening, a guard block should of
course be allowed to die. Also, programming errors are a large source of
exceptions, and the programmer certainly wants to know about those.

Since in most cases, the block executing when the guard gets executed does
not know or does not care about the guard blocks, it makes little sense to
let containing code handle the exception.

Therefore, whenever a guard block throws an exception, it will be caught
by Guard, followed by calling the code reference stored in C<$Guard::DIED>
(with C<$@> set to the actual exception), which is similar to how most
event loops handle this case.

The default for C<$Guard::DIED> is to call C<warn "$@">, i.e. the error is
printed as a warning and the program continues.

The C<$@> variable will be restored to its value before the guard call in
all cases, so guards will not disturb C<$@> in any way.

The code reference stored in C<$Guard::DIED> should not die (behaviour is
not guaranteed, but right now, the exception will simply be ignored).

=head1 AUTHOR

 Marc Lehmann <schmorp@schmorp.de>
 http://home.schmorp.de/

=head1 THANKS

Thanks to Marco Maisenhelder, who reminded me of the C<$Guard::DIED>
solution to the problem of exceptions.

=head1 SEE ALSO

L<Scope::Guard> and L<Sub::ScopeFinalizer>, which actually implement
dynamic guards only, not scoped guards, and have a lot higher CPU, memory
and typing overhead.

L<Hook::Scope>, which has apparently never been finished and can corrupt
memory when used.

=cut

Revision:	1.22
Committed:	Sun Jul 26 08:07:11 2009 UTC (14 years, 9 months ago) by root
Branch:	MAIN
Changes since 1.21:	+24 -28 lines
Log Message:	* empty log message *
#	User	Rev	Content
1	root	1.1	=head1 NAME
2
3			Guard - safe cleanup blocks
4
5			=head1 SYNOPSIS
6
7	root	1.9	use Guard;
8
9			# temporarily chdir to "/etc" directory, but make sure
10			# to go back to "/" no matter how myfun exits:
11			sub myfun {
12			scope_guard { chdir "/" };
13			chdir "/etc";
14
15	root	1.15	code_that_might_die_or_does_other_fun_stuff;
16	root	1.9	}
17	root	1.4
18	root	1.20	# create an object that, when the last reference to it is gone,
19			# invokes the given codeblock:
20			my $guard = guard { print "destroyed!\n" };
21			undef $guard; # probably destroyed here
22
23	root	1.1	=head1 DESCRIPTION
24
25			This module implements so-called "guards". A guard is something (usually
26			an object) that "guards" a resource, ensuring that it is cleaned up when
27			expected.
28
29			Specifically, this module supports two different types of guards: guard
30			objects, which execute a given code block when destroyed, and scoped
31			guards, which are tied to the scope exit.
32
33	root	1.2	=head1 FUNCTIONS
34
35			This module currently exports the C<scope_guard> and C<guard> functions by
36			default.
37
38	root	1.1	=over 4
39
40			=cut
41
42			package Guard;
43
44	root	1.13	no warnings;
45
46	root	1.1	BEGIN {
47	root	1.21	$VERSION = '1.021';
48	root	1.1	@ISA = qw(Exporter);
49	root	1.3	@EXPORT = qw(guard scope_guard);
50	root	1.1
51			require Exporter;
52
53			require XSLoader;
54			XSLoader::load Guard, $VERSION;
55			}
56
57			our $DIED = sub { warn "$@" };
58
59			=item scope_guard BLOCK
60
61			Registers a block that is executed when the current scope (block,
62			function, method, eval etc.) is exited.
63
64	root	1.8	See the EXCEPTIONS section for an explanation of how exceptions
65			(i.e. C<die>) are handled inside guard blocks.
66
67	root	1.2	The description below sounds a bit complicated, but that's just because
68			C<scope_guard> tries to get even corner cases "right": the goal is to
69			provide you with a rock solid clean up tool.
70
71	root	1.8	The behaviour is similar to this code fragment:
72	root	1.1
73			eval ... code following scope_guard ...
74			{
75			local $@;
76			eval BLOCK;
77			eval { $Guard::DIED->() } if $@;
78			}
79	root	1.2	die if $@;
80	root	1.1
81			Except it is much faster, and the whole thing gets executed even when the
82			BLOCK calls C<exit>, C<goto>, C<last> or escapes via other means.
83
84	root	1.4	If multiple BLOCKs are registered to the same scope, they will be executed
85	root	1.8	in reverse order. Other scope-related things such as C<local> are managed
86			via the same mechanism, so variables C<local>ised I<after> calling
87			C<scope_guard> will be restored when the guard runs.
88	root	1.1
89	root	1.4	Example: temporarily change the timezone for the current process,
90			ensuring it will be reset when the C<if> scope is exited:
91
92			use Guard;
93			use POSIX ();
94
95			if ($need_to_switch_tz) {
96			# make sure we call tzset after $ENV{TZ} has been restored
97			scope_guard { POSIX::tzset };
98	root	1.1
99	root	1.4	# localise after the scope_guard, so it gets undone in time
100			local $ENV{TZ} = "Europe/London";
101			POSIX::tzset;
102	root	1.1
103	root	1.4	# do something with the new timezone
104	root	1.1	}
105
106			=item my $guard = guard BLOCK
107
108			Behaves the same as C<scope_guard>, except that instead of executing
109			the block on scope exit, it returns an object whose lifetime determines
110			when the BLOCK gets executed: when the last reference to the object gets
111			destroyed, the BLOCK gets executed as with C<scope_guard>.
112
113	root	1.8	See the EXCEPTIONS section for an explanation of how exceptions
114			(i.e. C<die>) are handled inside guard blocks.
115	root	1.1
116			Example: acquire a Coro::Semaphore for a second by registering a
117	root	1.10	timer. The timer callback references the guard used to unlock it
118			again. (Please ignore the fact that C<Coro::Semaphore> has a C<guard>
119			method that does this already):
120	root	1.1
121	root	1.9	use Guard;
122	root	1.22	use Coro::AnyEvent;
123	root	1.1	use Coro::Semaphore;
124
125			my $sem = new Coro::Semaphore;
126
127	root	1.9	sub lock_for_a_second {
128	root	1.1	$sem->down;
129			my $guard = guard { $sem->up };
130
131	root	1.22	Coro::AnyEvent::sleep 1;
132
133			# $sem->up gets executed when returning
134	root	1.1	}
135
136			The advantage of doing this with a guard instead of simply calling C<<
137			$sem->down >> in the callback is that you can opt not to create the timer,
138	root	1.22	or your code can throw an exception before it can create the timer (or
139			the thread gets canceled), or you can create multiple timers or other
140			event watchers and only when the last one gets executed will the lock be
141			unlocked. Using the C<guard>, you do not have to worry about catching all
142			the places where you have to unlock the semaphore.
143	root	1.1
144	root	1.13	=item $guard->cancel
145	root	1.1
146			Calling this function will "disable" the guard object returned by the
147			C<guard> function, i.e. it will free the BLOCK originally passed to
148			C<guard >and will arrange for the BLOCK not to be executed.
149
150	root	1.22	This can be useful when you use C<guard> to create a cleanup handler to be
151			called under fatal conditions and later decide it is no longer needed.
152	root	1.1
153			=cut
154
155			1;
156
157			=back
158
159			=head1 EXCEPTIONS
160
161	root	1.5	Guard blocks should not normally throw exceptions (that is, C<die>). After
162	root	1.22	all, they are usually used to clean up after such exceptions. However,
163			if something truly exceptional is happening, a guard block should of
164			course be allowed to die. Also, programming errors are a large source of
165			exceptions, and the programmer certainly wants to know about those.
166	root	1.1
167	root	1.14	Since in most cases, the block executing when the guard gets executed does
168	root	1.1	not know or does not care about the guard blocks, it makes little sense to
169			let containing code handle the exception.
170
171	root	1.22	Therefore, whenever a guard block throws an exception, it will be caught
172			by Guard, followed by calling the code reference stored in C<$Guard::DIED>
173			(with C<$@> set to the actual exception), which is similar to how most
174			event loops handle this case.
175	root	1.1
176	root	1.22	The default for C<$Guard::DIED> is to call C<warn "$@">, i.e. the error is
177			printed as a warning and the program continues.
178	root	1.12
179			The C<$@> variable will be restored to its value before the guard call in
180			all cases, so guards will not disturb C<$@> in any way.
181
182	root	1.1	The code reference stored in C<$Guard::DIED> should not die (behaviour is
183			not guaranteed, but right now, the exception will simply be ignored).
184
185			=head1 AUTHOR
186
187			Marc Lehmann <schmorp@schmorp.de>
188			http://home.schmorp.de/
189
190			=head1 THANKS
191
192	root	1.6	Thanks to Marco Maisenhelder, who reminded me of the C<$Guard::DIED>
193			solution to the problem of exceptions.
194	root	1.1
195	root	1.16	=head1 SEE ALSO
196
197			L<Scope::Guard> and L<Sub::ScopeFinalizer>, which actually implement
198	root	1.22	dynamic guards only, not scoped guards, and have a lot higher CPU, memory
199			and typing overhead.
200	root	1.16
201	root	1.22	L<Hook::Scope>, which has apparently never been finished and can corrupt
202	root	1.16	memory when used.
203
204	root	1.1	=cut
205