TWiki on FastCGI

I couldn't find much discussion of TWiki on FCGI. There is a beta project to make a standalone TWiki daemon which can also be run under FCGI, but I had already installed TWiki 4.2.0 so I was reticent to reinstall from a different branch.

I only really needed to speed up the view cgi, so it shouldn't be too hard surely? I already had mod_fcgid installed on my Apache2 server.

It turned out to really be very easy, and seems to be working fine so far. Here is my diff for twiki/bin/view:

--- view~       2008-01-22 14:18:52.000000000 +1100
+++ view        2008-07-20 18:40:33.000000000 +1000
@@ -27,6 +27,14 @@
     require 'setlib.cfg';
 }

+use FCGI;
 use TWiki::UI;
 use TWiki::UI::View;
-TWiki::UI::run( \&TWiki::UI::View::view, view => 1 );
+
+my $request = FCGI::Request;
+while ($request->Accept >= 0) {
+  eval {TWiki::UI::run( \&TWiki::UI::View::view, view => 1 );};
+  warn $@ if $@;
+  $request->Flush;
+  $request->Finish;
+}

And I added the following to the bin <Directory> section in my apache config:

<FilesMatch "^view$">
    SetHandler fcgid-script
</FilesMatch>

This seemed to work fine, but searches failed to spawn grep correctly. I think diffs would have also failed to spawn rcs. So I switched over to the pure perl versions by making the following settings in LocalSite.cfg:

$TWiki::cfg{StoreImpl} = 'RcsLite';
TWiki::cfg{RCS}{SearchAlgorithm} = 'TWiki::Store::SearchAlgorithms::PurePerl';

Working a treat so far - Memory use seems ok. It rose from about 14000k up to about 15200k, and then hovered around that level indefinately. I'll let you know if I see any memory leaks or wierd issues.

Update: I guess the pure perl search implementation isn't well used. It threw up a taint error when I tried to use it. No matter, the fix was as simple as replacing a horrible piece of string eval:

--- lib/TWiki/Store/SearchAlgorithms/PurePerl.pm~       2008-01-22 14:18:55.000000000 +1100
+++ lib/TWiki/Store/SearchAlgorithms/PurePerl.pm        2008-07-20 19:32:58.000000000 +1000
@@ -46,9 +46,14 @@
     # Convert GNU grep \< \> syntax to \b
     $searchString =~ s/(?]/\\b/g;
     $searchString =~ s/^(.*)$/\\b$1\\b/go if $options->{'wordboundaries'};
-    my $match_code = "return \$_[0] =~ m/$searchString/o";
-    $match_code .= 'i' unless ($options->{casesensitive});
-    my $doMatch = eval "sub { $match_code }";
+
+    my $doMatch;
+    if ($options->{casesensitive}) {
+      $doMatch = sub { $_[0] =~ m/$searchString/o };
+    } else {
+      $doMatch = sub { $_[0] =~ m/$searchString/oi };
+    }
+
   FILE:
     foreach my $file ( @$topics ) {
         next unless open(FILE, "<$sDir/$file.txt");

I'll have to track down how to submit TWiki bugs...

Update 2: Another search issue - in the persisted view, the search page never gets re-rendered (ie. after making one successful search, all future searches appear to have identical results).

I didn't have time to find if that was a problem in the view code or the pure pearl search, but it was easy enough to make sure a new cgi was spawned per search request by adding the following at the end of the VirtualHost:

<LocationMatch "WebSearch">
    SetHandler cgi-script
</LocationMatch>

Conveniently the FCGI script works fine as a regular one-shot cgi, and since LocationMatch is processed after FilesMatch by Apache, this overrides the fcgid handler setting.