Re: [PATCH 0/2] gitweb: Add support for running gitweb as FastCGI script

On Mon, 10 May 2010, Peter Vereshagin wrote:

You can find current state of my take on gitweb output caching (based on / inspired by work by John 'Warthog9' Hawley) in my repository on repo.or.cz, in the 'gitweb/cache-kernel-pu' branch:

  http://repo.or.cz/w/git/jnareb-git.git  gitweb/cache-kernel-pu
 
You can find progress reports (and what current show-stoppers are) in
git mailing list archives.

Note that http://repo.or.cz does its own gitweb caching, IIRC by caching Perl data, and only for 'projects_list' page (the most costly one).

There was also "Gitweb caching" projects in GSoC 2008 by Lea Wiemann, which IIUC cached output of git commands. This project was, I think, completed but didn't get merged into git.

Why so old git?  Current version is git version 1.7.1

Hrmph.  Why not use "git diff --no-index <file1> <file2>" here?

The Perl-aware equivalent of '-p' option of GNU diff, i.e. showing in which function we are in hunk headers, would help here.

This 'exit' was here just in case there were some forgotten code below this line outside subroutines (that should not be run). It can be safely removed.

I have added my guess of in which subroutine this code is above.

Err... and gitweb works correctly with this change? This 'exit' was required for die_error to function like 'die' in that it finishes serving request, and should not continue subroutine it was called from.

I have changed this 'exit' to non-local goto to toplevel. It could be done instead by redefining 'exit' subroutine, like shown below, but I feel that would be hacky if you can change gitweb code (it is not black box you should not touch).

This is quite nice idea to replace 'exit' by subroutine that does non-local jump to outside of application, at the end of request loop. Such "monkey patching" is the only solution if you can't or shouldn't modify application code (like FCGI::Spawn being generic solution).

Here

   $spawn->{'callout'} = sub {
   	my $cgi_app = shift;
   	do $cgi_app;

        # this is needed for sane error handling
        die "Couldn't parse $cgi_app: $@" if $@;

   CALLED_OUT: 
   };

could be simply replaced by

  use CGI::Compile;

  # ...

  $spawn->{'callout'} = \&{CGI::Compile->compile}

or something like that. See CGI::Compile manpage and CGI::Compile source: http://cpansearch.perl.org/src/MIYAGAWA/CGI-Compile-0.11/lib/CGI/Compile.pm

BTW I wonder what are the consequences for performance on replacing 'exit' by non-local jump. It can degrade performance a bit for gitweb run as pure CGI (mod_cgi / mod_cgid), but should improve performance for mod_perl, at least if there are more connections... unless ModPerl::Registry does similar trick with exit().

Hmmm... is FCGI::Spawn really needed, or can it be replaced by simple PSGI wrapper using either Plack::App::CGIBin,

  use Plack::App::CGIBin;
  use Plack::Builder;

  my $app = Plack::App::CGIBin->new(root => "/path/to/cgi-bin")->to_app;
  builder {
        mount "/cgi-bin" => $app;
  };

or Plack::App::WrapCGI plus Plack::App::URLMap, the last indirectly via Plack::Builder DSL:

  use Plack::Builder;
  use Plack::App::WrapCGI;

  builder {
        mount "/foo" =>
                Plack::App::WrapCGI->new(script => "foo.cgi")->to_app;
        mount "/bar" =>
                Plack::App::WrapCGI->new(script => "bar.cgi")->to_app;
  };

Errr... excuse me, what you wanted to say in the paragraph above?

Gitweb doesn't use no POST requests: it is read-only web repository browser... well, except for the 'show_ctags' action.