Mailing-List: contact modperl-help@perl.apache.org; run by ezmlm
Precedence: bulk
Received-SPF: pass (herse.apache.org: domain of torsten.foertsch@gmx.net
 designates 213.165.64.20 as permitted sender)
From: Torsten Foertsch <torsten.foertsch@gmx.net>
To: "Alex Beamish" <talexb@gmail.com>
Subject: Re: Forking to an interactive program under mod_perl
Date: Wed, 13 Dec 2006 11:44:08 +0100
User-Agent: KMail/1.9.1
Cc: "mod_perl list" <modperl@perl.apache.org>
References: <f9870c010612121318r48ba8a76i3e59ac850b97febb@mail.gmail.com>
 <57FDC457-C935-4C73-A1A8-6C9695672515@2xlp.com>
 <f9870c010612121703i2d7d43d3wb181907ae1301b9e@mail.gmail.com>
In-Reply-To: <f9870c010612121703i2d7d43d3wb181907ae1301b9e@mail.gmail.com>
MIME-Version: 1.0
Content-Type: multipart/signed;
  boundary="nextPart2672925.27GFXTykR6";
  protocol="application/pgp-signature";
  micalg=pgp-sha1
Content-Transfer-Encoding: 7bit
Message-Id: <200612131144.13082.torsten.foertsch@gmx.net>

--nextPart2672925.27GFXTykR6
Content-Type: text/plain;
  charset="iso-8859-15"
Content-Transfer-Encoding: quoted-printable
Content-Disposition: inline

On Wednesday 13 December 2006 02:03, Alex Beamish wrote:
> Interesting suggestions, thank you, and I'm coming around to the idea that
> a daemon will need to be used for the heavy lifting .. and then perhaps
> mod_perl can communicate with the daemon using named pipes ..
>
> The whole point of this exercise is to keep all of the page information in
> memory so that when Ghostscript gets another page request it can service =
it
> right away, rather having to reload everything all over again. Lower
> latency, at the cost of higher memory requirements. So that tells me I
> don't want to load Ghostscript each time a new request comes in.
>
> I'll deal with multiple documents with some combination of stale timers a=
nd
> LRU slots, but that's not really what I see as the most complicated or
> difficult part of this problem. For this particular application, my
> inactivity timer will probably by 10-15 minutes, and I'll expect to have
> 6-8 documents open at any given time, so it shouldn't be a big drain on
> memory. And I will probably be able to set something up that signals that=
 a
> document has been expired as well .. (this is just me thinking out loud) =
=2E.
>
> Thanks for your feedback .. I think named pipes is my next focus.

Alex,

one thing you might run into while trying to control gs is buffering. Most=
=20
programs buffer their output if STDOUT is not bound to a terminal. That mea=
ns=20
even if you have switched off output buffering in Perl ($|=3D1) you write a=
=20
command to the external program but won't get an answer until enough output=
=20
is generated to fill up a buffer.

This problem can be solved using pseudo terminals. IO::Pty is handy at this.

=46urther, I'd use a unix domain socket to talk to your gs-controller-daemo=
n.=20

apache + mod_perl
       ^
       |
       | IO::Socket::UNIX
       |
       v
 gs-controller
       ^
       |
       | IO::Pty
       |
       v
       gs

Let gs-controller listen on a socket. Each actual connection to a gs is=20
assigned a unique ID by the gs-controller. This ID you can put into your pa=
ge=20
as a session handle. So a subsequent request for the same gs-connection rea=
ds=20
this ID and transfers it to the gs-controller together with the gs-commands.
gs-controller then finds the actual PTY and submits the commands. Of course=
,=20
your gs-controller has to handle timeouts and errors.

If you need to spawn off the gs-controller daemon on demand from mod_perl=20
remember that

1) Apache opens a lot more file descriptors than STDIN, STDOUT and STDERR
2) use CORE::exit if you really want your program to exit (after a fork may=
be)
3) Apache signals it's process group if it wants to shut down or restart

Maybe some of the problems can be avoided using Apache2::SubProcess. I have=
=20
never used it. For me the following code does the job:

  sub sysclose {
    require 'syscall.ph';
    syscall &SYS_close, shift()+0;
  }

  sub RLIMIT_NOFILE {7;}                # linux specific
  sub getrlimit {
    require 'syscall.ph';
    my $x=3D"x"x8;			# result
    syscall &SYS_getrlimit, shift()+0, $x;
    unpack "ii", $x;
  }

  sub close_fd {
    my @l=3Dgetrlimit( RLIMIT_NOFILE );
    my $l;

    # close all files save STDIN, STDOUT and STDERR
    for( $l=3D3; $l<$l[0]; $l++ ) {
      sysclose $l;
    }
    sysclose 0;
    open STDIN, '/dev/null';
    sysclose 1;
    open STDOUT, '>/dev/null';
  }

  sub spawn {
    my ($daemon_should_survive_apache_restart, @args)=3D@_;

    local $SIG{CHLD}=3D'IGNORE';
    my $pid;
    # yes, even fork can fail
    select undef, undef, undef, .1 while( !defined($pid=3Dfork) );
    unless( $pid ) {		# child
      # 2nd fork to cut parent relationship with a mod_perl apache
      select undef, undef, undef, .1 while( !defined($pid=3Dfork) );
      if( $pid ) {
	CORE::exit 0;
      } else {
	close_fd;
	POSIX::setsid if( $daemon_should_survive_apache_restart );
	exec @args;
	CORE::exit 0;
      }
    }
    waitpid $pid, 0;		# avoid a zombie
  }

The spawned process has only 1 open file, STDERR. For mod_perl processes it=
=20
points to the ErrorLog. For processes spawned by CGI scripts (Apache 2.x) it
is a pipe to the Apache child. In this latter case you must open your own=20
STDERR.

Torsten

--nextPart2672925.27GFXTykR6
Content-Type: application/pgp-signature

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2 (GNU/Linux)

iD8DBQBFf9l9wicyCTir8T4RAjkxAKDIMDTQUPhh25guSC5eMvjvMC6d1wCfdLL2
NcaGMZXkOFFl3r5xBLSS2jY=
=NFde
-----END PGP SIGNATURE-----

--nextPart2672925.27GFXTykR6--