Home / Documentation / 2.0 / User's guide / | ||||
Writing mod_perl Handlers and Scripts | ||||
|
||
This chapter covers the mod_perl coding specifics, different from normal Perl coding. Most other perl coding issues are covered in the perl manpages and rich literature.
mod_perl 2.0 has all its methods spread across many modules. In order
to use these methods the modules containing them have to be loaded
first. If you don't do that mod_perl will complain that it can't find
the methods in question. The module
ModPerl::MethodLookup
can
be used to find out which modules need to be used.
In addition to function handlers method handlers can be used. Method
handlers are useful when you want to write code that takes advantage
of inheritance. To make the handler act as a method under mod_perl 2,
use the method
attribute.
See the Perl attributes manpage for details on the attributes
syntax (perldoc attributes
).
For example:
package Bird::Eagle; @ISA = qw(Bird); sub handler : method { my ($class_or_object, $r) = @_; ...; } sub new { bless {}, __PACKAGE__ }
and then register it as:
PerlResponseHandler Bird::Eagle
When mod_perl sees that the handler has a method attribute, it passes two arguments to it: the calling object or a class, depending on how it was called, and the request object, as shown above.
If Class->method
syntax is used for a Perl*Handler
, e.g.:
PerlResponseHandler Bird::Eagle->handler;
the :method
attribute is not required.
In the preceding configuration example, the handler()
method will
be called as a class (static) method.
Also, you can use objects created at startup to call methods. For example:
<Perl> use Bird::Eagle; $Bird::Global::object = Bird::Eagle->new(); </Perl> ... PerlResponseHandler $Bird::Global::object->handler
In this example, the handler()
method will be called as an instance
method on the global object $Bird::Global::object
.
It's possible to arrange for cleanups to happen at the end of various
phases. One can't rely on END
blocks to do the job, since these
don't get executed until the interpreter quits, with
an exception to the
Registry handlers.
Module authors needing to run cleanups after each HTTP request, should
use
PerlCleanupHandler
.
Module authors needing to run cleanups at other times can always
register a cleanup callback via
cleanup_register
on the pool object of choice. Here are some examples of its usage:
To run something at the server shutdown and restart use a cleanup
handler registered on
server_shutdown_cleanup_register()
in startup.pl:
#PerlPostConfigRequire startup.pl use Apache2::ServerUtil (); use APR::Pool (); warn "parent pid is $$\n"; Apache2::ServerUtil::server_shutdown_cleanup_register((\&cleanup); sub cleanup { warn "server cleanup in $$\n" }
This is usually useful when some server-wide cleanup should be performed when the server is stopped or restarted.
To run a cleanup at the end of each connection phase, assign a cleanup callback to the connection pool object:
use Apache2::Connection (); use APR::Pool (); my $pool = $c->pool; $pool->cleanup_register(\&my_cleanup); sub my_cleanup { ... }
You can also create your own pool object, register a cleanup callback and it'll be called when the object is destroyed:
use APR::Pool (); { my @args = 1..3; my $pool = APR::Pool->new; $pool->cleanup_register(\&cleanup, \@args); } sub cleanup { my @args = @{ +shift }; warn "cleanup was called with args: @args"; }
In this example the cleanup callback gets called, when $pool
goes
out of scope and gets destroyed. This is very similar to OO DESTROY
method.
mod_perl sets the following environment variables:
$ENV{MOD_PERL}
- is set to the mod_perl version the server is
running under. e.g.:
mod_perl/2.000002
If $ENV{MOD_PERL}
doesn't exist, most likely you are not running
under mod_perl.
die "I refuse to work without mod_perl!" unless exists $ENV{MOD_PERL};
However to check which version is used it's better to use the following technique:
use mod_perl; use constant MP2 => ( exists $ENV{MOD_PERL_API_VERSION} and $ENV{MOD_PERL_API_VERSION} >= 2 ); # die "I want mod_perl 2.0!" unless MP2;
mod_perl passes (exports) the following shell environment variables (if they are set) :
PATH
- Executables search path.
TZ
- Time Zone.
Any of these environment variables can be accessed via %ENV
.
If the code needs to behave differently depending on whether it's
running under one of the threaded MPMs, or not, the class method
Apache2::MPM->is_threaded
can be used. For example:
use Apache2::MPM (); if (Apache2::MPM->is_threaded) { require APR::OS; my $tid = APR::OS::current_thread_id(); print "current thread id: $tid (pid: $$)"; } else { print "current process id: $$"; }
This code prints the current thread id if running under a threaded MPM, otherwise it prints the process id.
If you write a CPAN module it's a bad idea to write code that won't run under all MPMs, and developers should strive to write a code that works with all mpms. However it's perfectly fine to perform different things under different mpms.
If you don't develop CPAN modules, it's perfectly fine to develop your project to be run under a specific MPM.
use Apache2::MPM (); my $mpm = lc Apache2::MPM->show; if ($mpm eq 'prefork') { # prefork-specific code } elsif ($mpm eq 'worker') { # worker-specific code } elsif ($mpm eq 'winnt') { # winnt-specific code } else { # others... }
META: need to port Apache2::Reload notes from the guide here. but the gist is:
PerlModule Apache2::Reload PerlInitHandler Apache2::Reload #PerlPreConnectionHandler Apache2::Reload PerlSetVar ReloadAll Off PerlSetVar ReloadModules "ModPerl::* Apache2::*"
Use:
PerlInitHandler Apache2::Reload
if you need to debug HTTP protocol handlers. Use:
PerlPreConnectionHandler Apache2::Reload
for any handlers.
Though notice that we have started to practice the following style in our modules:
package Apache2::Whatever; use strict; use warnings FATAL => 'all';
FATAL => 'all'
escalates all warnings into fatal errors. So
when Apache2::Whatever
is modified and reloaded by Apache2::Reload
the request is aborted. Therefore if you follow this very healthy
style and want to use Apache2::Reload
, flex the strictness by
changing it to:
use warnings FATAL => 'all'; no warnings 'redefine';
but you probably still want to get the redefine warnings, but downgrade them to be non-fatal. The following will do the trick:
use warnings FATAL => 'all'; no warnings 'redefine'; use warnings 'redefine';
Perl 5.8.0 allows to do all this in one line:
use warnings FATAL => 'all', NONFATAL => 'redefine';
but if your code may be used with older perl versions, you probably don't want to use this new functionality.
Refer to the perllexwarn manpage for more information.
In the following sections we discuss the specifics of Apache behavior relevant to mod_perl developers.
The best approach for generating HTTP response headers is by using the
mod_perl API. Some common
headers have dedicated methods, others are set by manipulating the
headers_out
table directly.
For example to set the Content-type header you should call
$r->content_type
:
use Apache2::RequestRec (); $r->content_type('text/html');
To set
a custom header
My-Header you should call:
use Apache2::RequestRec (); use APR::Table; $r->headers_out->set(My-Header => "SomeValue");
If you are inside a registry script you can still access the
Apache2::RequestRec
object.
Howerever you can choose a slower method of generating headers by just
printing them out before printing any response. This will work only if
PerlOptions +ParseHeaders
is
in effect. For example:
print "Content-type: text/html\n"; print "My-Header: SomeValue\n"; print "\n";
This method is slower since Apache needs to parse the text to identify certain headers it needs to know about. It also has several limitations which we will now discuss.
When using this approach you must make sure that the STDOUT
filehandle is not set to flush the data after each print (which is set
by the value of a special perl variable $|
). Here we assume that
STDOUT is the currently select()
ed filehandle and $|
affects it.
For example this code won't work:
local $| = 1; print "Content-type: text/html\n"; print "My-Header: SomeValue\n"; print "\n";
Having a true $|
causes the first print() call to flush its data
immediately, which is sent to the internal HTTP header parser, which
will fail since it won't see the terminating "\n\n"
. One solution
is to make sure that STDOUT won't flush immediately, like so:
local $| = 0; print "Content-type: text/html\n"; print "My-Header: SomeValue\n"; print "\n";
Notice that we local()
ize that change, so it won't affect any other code.
If you send headers line by line and their total length is bigger than 8k, you will have the header parser problem again, since mod_perl will flush data when the 8k buffer gets full. In which case the solution is not to print the headers one by one, but to buffer them all in a variable and then print the whole set at once.
Notice that you don't have any of these problems with mod_cgi, because it ignores any of the flush attempts by Perl. mod_cgi simply opens a pipe to the external process and reads any output sent from that process at once.
If you use $r
to set headers as explained at the beginning of this
section, you won't encounter any of these problems.
Finally, if you don't want Apache to send its own headers and you want
to send your own set of headers (non-parsed headers handlers) use the
$r->assbackwards
method. Notice that registry handlers will do that for you if the
script's name start with the nph-
prefix.
Apache 2.0 doesn't provide a method to force HTTP response headers
sending (what used to be done by send_http_header()
in Apache
1.3). HTTP response headers are sent as soon as the first bits of the
response body are seen by the special core output filter that
generates these headers. When the response handler sends the first
chunks of body it may be cached by the mod_perl internal buffer or
even by some of the output filters. The response handler needs to
flush the output in order to tell all the components participating in
the sending of the response to pass the data out.
For example if the handler needs to perform a relatively long-running operation (e.g. a slow db lookup) and the client may timeout if it receives nothing right away, you may want to start the handler by setting the Content-Type header, following by an immediate flush:
sub handler { my $r = shift; $r->content_type('text/html'); $r->rflush; # send the headers out $r->print(long_operation()); return Apache2::Const::OK; }
If this doesn't work, check whether you have configured any third-party output filters for the resource in question. Improperly written filter may ignore the command to flush the data.
In mod_perl 2.0 a response body can be sent only during the response phase. Any attempts to do that in the earlier phases will fail with an appropriate explanation logged into the error_log file.
This happens due to the Apache 2.0 HTTP architecture specifics. One of the issues is that the HTTP response filters are not setup before the response phase.
3rd party Apache 2 modules should avoid using code relying on
signals. This is because typical signal use is not thread-safe and
modules which rely on signals may not work portably. Certain signals
may still work for non-threaded mpms. For example alarm()
can be
used under prefork MPM, but it won't work on any other MPM. Moreover
the Apache developers don'tq guarantee that the signals that currently
happen to work will continue to do so in the future Apache
releases. So use them at your own risk.
It should be possible to rework the code using signals to use an
alternative solution, which works under threads. For example if you
were using alarm()
to trap potentially long running I/O, you can
modify the I/O logic for select/poll usage (or if you use APR I/O then
set timeouts on the apr pipes or sockets). For example, Apache 1.3 on
Unix made blocking I/O calls and relied on the parent process to send
the SIGALRM signal to break it out of the I/O after a timeout expired.
With Apache 2.0, APR support for timeouts on I/O operations is used so
that signals or other thread-unsafe mechanisms are not necessary.
CPU timeout handling is another example. It can be accomplished by modifying the computation logic to explicitly check for the timeout at intervals.
Talking about alarm()
under prefork mpm, POSIX signals seem to
work, but require Perl 5.8.x+. For example:
use POSIX qw(SIGALRM); my $mask = POSIX::SigSet->new( SIGALRM ); my $action = POSIX::SigAction->new(sub { die "alarm" }, $mask); my $oldaction = POSIX::SigAction->new(); POSIX::sigaction(SIGALRM, $action, $oldaction ); eval { alarm 2; sleep 10 # some real code should be here alarm 0; }; POSIX::sigaction(SIGALRM, $oldaction); # restore original warn "got alarm" if $@ and $@ =~ /alarm/;
For more details see: http://search.cpan.org/dist/perl/ext/POSIX/POSIX.pod#POSIX::SigAction.
One could use the $SIG{ALRM}
technique, working for 5.6.x+, but it
works only under DSO modperl build. Moreover starting from 5.8.0
Perl delays signal delivery, making signals safe. This change may
break previously working code. For more information please see:
http://search.cpan.org/dist/perl/pod/perl58delta.pod#Safe_Signals and
http://search.cpan.org/dist/perl/pod/perlipc.pod#Deferred_Signals_%28Safe_Signals%29.
For example if you had the alarm code:
eval { local $SIG{ALRM} = sub { die "alarm" }; alarm 3; sleep 10; # in reality some real code should be here alarm 0; }; die "the operation was aborted" if $@ and $@ =~ /alarm/;
It may not work anymore. Starting from 5.8.1 it's possible to circumvent the safeness of signals, by setting:
$ENV{PERL_SIGNALS} = "unsafe";
as soon as you start your program (e.g. in the case of mod_perl in startup.pl). As of this writing, this workaround fails on MacOSX, POSIX signals must be used instead.
For more information please refer to: http://search.cpan.org/dist/perl/pod/perl581delta.pod#Unsafe_signals_again_available and http://search.cpan.org/dist/perl/pod/perlrun.pod#PERL_SIGNALS.
Though if you use perl 5.8.x+ it's preferrable to use the POSIX API technique explained earlier in this section.
In the following sections we discuss the specifics of Perl behavior under mod_perl.
BEGIN
BlocksPerl executes BEGIN
blocks as soon as possible, at the time of
compiling the code. The same is true under mod_perl. However, since
mod_perl normally only compiles scripts and modules once, either in
the parent server (at the server startup) or once per-child (on the
first request using a module), BEGIN
blocks in that code will only
be run once. As the perlmod
manpage explains, once a BEGIN
block has run, it is immediately undefined. In the mod_perl
environment, this means that BEGIN
blocks will not be run during
the response to an incoming request unless that request happens to be
the one that causes the compilation of the code, i.e. if it wasn't
loaded yet.
BEGIN
blocks in modules and files pulled in via require()
or
use()
will be executed:
Only once, if pulled in by the parent process at the server startup.
Once per each child process or Perl interpreter if not pulled in by the parent process.
An additional time, once per each child process or Perl interpreter if
the module is reloaded off disk again via
Apache2::Reload
.
Unpredictable if you fiddle with %INC
yourself.
The BEGIN
blocks behavior is different in
ModPerl::Registry
and
ModPerl::PerlRun
handlers, and their subclasses.
CHECK
and INIT
BlocksCHECK
and INIT
blocks run when the source code compilation is
complete, but before the program starts. CHECK
can mean
"checkpoint" or "double-check" or even just "stop". INIT
stands for
"initialization". The difference is subtle; CHECK
blocks are run
just after the compilation ends, INIT
just before the runtime
begins. (Hence the -c
command-line perl option runs CHECK
blocks
but not INIT
blocks.)
Perl only calls these blocks during perl_parse(), which mod_perl
calls once at startup time. Under threaded mpm, these blocks will be
called once per parent perl interpreter startup
. Therefore
CHECK
and INIT
blocks don't work after the server is started,
for the same reason these code samples don't work:
% perl -e 'eval qq(CHECK { print "ok\n" })' % perl -e 'eval qq(INIT { print "ok\n" })'
END
BlocksAs the perlmod
manpage explains, an END
block is executed as
late as possible, that is, when the interpreter exits. So for example
mod_cgi will run its END
blocks on each invocation, since on every
invocation it starts a new interpreter and then kills it when the
request processing is done.
In the mod_perl environment, the interpreter does not exit after
serving a single request (unless it is configured to do so) and hence
it will run its END
blocks only when it exits, which usually
happens during the server shutdown, but may also happen earlier than
that (e.g. a process exits because it has served a
MaxRequestsPerChild
number of requests).
mod_perl does make a special case
for scripts running under
ModPerl::Registry
and friends.
The Cleaning up section explains how to deal with cleanups for non-Registry handlers.
ModPerl::Global
API:
special_list_register
,
special_list_call
and
special_list_clear
,
internally used by registry handlers, can be used to run END
blocks
at arbitrary times.
mod_perl 2.0 provides two types of SetHandler
handlers:
modperl
and
perl-script
.
Remember that the SetHandler
directive is only relevant for the
response phase handlers, it neither needed nor affects non-response
phases.
Under the handler:
SetHandler perl-script
several special global Perl variables are saved before the handler is
called and restored afterwards. This includes: %ENV
, @INC
,
$/
, STDOUT
's $|
and END
blocks array (PL_endav
).
Under:
SetHandler modperl
nothing is restored, so you should be especially careful to remember localize all special Perl variables so the local changes won't affect other handlers.
exit
In the normal Perl code exit() is used to stop the program flow and exit the Perl interpreter. However under mod_perl we only want the stop the program flow without killing the Perl interpreter.
You should take no action if your code includes exit() calls and it's OK to continue using them. mod_perl worries to override the exit() function with its own version which stops the program flow, and performs all the necessary cleanups, but doesn't kill the server. This is done by overriding:
*CORE::GLOBAL::exit = \&ModPerl::Util::exit;
so if you mess up with *CORE::GLOBAL::exit
yourself you better know
what you are doing.
You can still call CORE::exit
to kill the interpreter, again if you
know what you are doing.
One caveat is when exit
is called inside eval
-- the ModPerl::Util::exit documentation explains how to
deal with this situation.
If you have a CGI script test.pl:
#!/usr/bin/perl print "Content-type: text/plain\n\n"; print "Hello";
a typical registry family handler turns it into something like:
package foo_bar_baz; sub handler { local $0 = "/full/path/to/test.pl"; #line 1 test.pl #!/usr/bin/perl print "Content-type: text/plain\n\n"; print "Hello"; }
Turning it into an almost full-fledged mod_perl handler. The only difference is that it handles the return status for you. (META: more details on return status needed.)
It then executes it as:
foo_bar_baz::handler($r);
passing the $r
object as the
only argument to the handler()
function.
Depending on the used registry handler the package is made of the file path, the uri or anything else. Check the handler's documentation to learn which method is used.
$r
ObjectAs explained in A Look Behind the Scenes
the $r
object is always passed to the registry script's special
function handler
as the first and the only argument, so you can get
this object by accessing @_
, since:
my $r = shift; print "Content-type: text/plain\n\n"; print "Hello";
is turned into:
sub handler { my $r = shift; print "Content-type: text/plain\n\n"; print "Hello"; }
behind the scenes. Now you can use $r
to call various mod_perl
methods, e.g. rewriting the script as:
my $r = shift; $r->content_type('text/plain'); $r->print();
If you are deep inside some code and can't get to the entry point to
reach for $r
, you can use
Apache2->request
.
The following sections discuss threading issues when running mod_perl under a threaded MPM.
The "only" thing you have to worry about your code is that it's thread-safe and that you don't use functions that affect all threads in the same process.
Perl 5.8.0 itself is thread-safe. That means that operations like
push()
, map()
, chomp()
, =
, /
, +=
, etc. are
thread-safe. Operations that involve system calls, may or may not be
thread-safe. It all depends on whether the underlying C libraries used
by the perl functions are thread-safe.
For example the function localtime()
is not thread-safe when the
implementation of asctime(3)
is not thread-safe. Other usually
problematic functions include readdir()
, srand()
, etc.
Another important issue that shouldn't be missed is what some people
refer to as thread-locality. Certain functions executed in a single
thread affect the whole process and therefore all other threads
running inside that process. For example if you chdir()
in one
thread, all other thread now see the current working directory of that
thread that chdir()
'ed to that directory. Other functions with
similar effects include umask()
, chroot()
, etc. Currently there
is no cure for this problem. You have to find these functions in your
code and replace them with alternative solutions which don't incur
this problem.
For more information refer to the perlthrtut (http://perldoc.perl.org/perlthrtut.html) manpage.
This is actually quite unrelated to mod_perl 2.0. You don't have to know much about Perl threads, other than Thread-environment Issues, to have your code properly work under threaded MPM mod_perl.
If you want to spawn your own threads, first of all study how the new ithreads Perl model works, by reading the perlthrtut, threads (http://search.cpan.org/search?query=threads) and threads::shared (http://search.cpan.org/search?query=threads%3A%3Ashared) manpages.
Artur Bergman wrote an article which explains how to port pure Perl
modules to work properly with Perl ithreads. Issues with chdir()
and other functions that rely on shared process' datastructures are
discussed. http://www.perl.com/lpt/a/2002/06/11/threads.html.
Global variables are only global to the interpreter in which they are
created. Other interpreters from other threads can't access that
variable. Though it's possible to make existing variables shared
between several threads running in the same process by using the
function threads::shared::share()
. New variables can be shared by
using the shared attribute when creating them. This feature is
documented in the threads::shared
(http://search.cpan.org/search?query=threads%3A%3Ashared) manpage.
|