1 files changed, 91 insertions, 0 deletions
diff --git a/doc/plugins/contrib/cvs/discussion.mdwn b/doc/plugins/contrib/cvs/discussion.mdwn
new file mode 100644
index 000000000..e1fa6e428
--- /dev/null
+++ b/doc/plugins/contrib/cvs/discussion.mdwn
@@ -0,0 +1,91 @@
+I've started reviewing this, and the main thing I don't like is the
+post-commit wrapper wrapper that ikiwiki-makerepo is patched to set up.
+That just seems unnecessarily complicated. Why can't ikiwiki itself detect
+the "cvs add <directory>" call and avoid doing anything in that case?
+--[[Joey]] 
+
+> The wrapper wrapper does three things:
+>
+> 7. It ignores `cvs add <directory>`, since this is a weird CVS
+> behavior that ikiwiki gets confused by and doesn't need to act on.
+> 7. It prevents `cvs` locking against itself: `cvs commit` takes a
+> write lock and runs the post-commit hook, which runs `cvs update`,
+> which wants a read lock and sleeps forever -- unless the post-commit
+> hook runs in the background so the commit can "finish".
+> 7. It fails silently if the ikiwiki post-commit hook is missing.
+> CVS doesn't have any magic post-commit filenames, so hooks have to
+> be configured explicitly. I don't think a commit will actually fail
+> if a configured post-commit hook is missing (though I can't test
+> this at the moment).
+>
+> Thing 1 can probably be handled within ikiwiki, if that seems less
+> gross to you.
+
+>> It seems like it might be. You can use a `getopt` hook to check
+>> `@ARGV` to see how it was called. --[[Joey]] 
+
+>>> This does the trick iff the post-commit wrapper passes its args
+>>> along. Committed on my branch. This seems potentially dangerous,
+>>> since the args passed to ikiwiki are influenced by web commits.
+>>> I don't see an exploit, but for paranoia's sake, maybe the wrapper
+>>> should only be built with execv() if the cvs plugin is loaded?
+>>> --[[schmonz]]
+
+>>>> Hadn't considered that. While in wrapper mode the normal getopt is not
+>>>> done, plugin getopt still runs, and so any unsafe options that 
+>>>> other plugins support could be a problem if another user runs
+>>>> the setuid wrapper and passes those options through. --[[Joey]]
+
+>>>>> I've tried compiling the argument check into the wrapper as
+>>>>> the first thing main() does, and was surprised to find that
+>>>>> this doesn't prevent the `cvs add <dir>` deadlock in a web
+>>>>> commit. I was convinced this'd be a reasonable solution,
+>>>>> especially if conditionalized on the cvs plugin being loaded,
+>>>>> but it doesn't work. And I stuck debug printfs at the beginning
+>>>>> of all the rcs_foo() subs, and whatever `cvs add <dir>` is
+>>>>> doing to ikiwiki isn't visible to my plugin, because none of
+>>>>> those subs are getting called. Nuts. Can you think of anything
+>>>>> else that might solve the problem, or should I go back to
+>>>>> generating a minimal wrapper wrapper that checks for just
+>>>>> this one thing? --[[schmonz]]
+
+>>>>>> I don't see how there could possibly be a difference between
+>>>>>> ikiwiki's C wrapper and your shell wrapper wrapper here. --[[Joey]]
+
+> Thing 2 I'm less sure of. (I'd like to see the web UI return
+> immediately on save anyway, to a temporary "rebuilding, please wait
+> if you feel like knowing when it's done" page, but this problem
+> with CVS happens with any kind of commit, and could conceivably
+> happen with some other VCS.)
+
+>> None of the other VCSes let a write lock block a read lock, apparently.
+>> 
+>> Anyway, re the backgrounding, when committing via the web, the
+>> post-commit hook doesn't run anyway; the rendering is done via the
+>> ikiwiki CGI. It would certianly be nice if it popped up a quick "working"
+>> page and replaced it with the updated page when done, but that's
+>> unrelated;  the post-commit
+>> hook only does rendering when committing using the VCS directly. The
+>> backgrounding you do actually seems safe enough -- but tacking
+>> on a " &" to the ikiwiki wrapper call doesn't need a wrapper script,
+>> does it? --[[Joey]]
+
+>>> Nope, it works fine to append it to the `CVSROOT/loginfo` line.
+>>> Fixed on my branch. --[[schmonz]]
+
+> Thing 3 I think I did in order to squelch the error messages that
+> were bollixing up the CGI. It was easy to do this in the wrapper
+> wrapper, but if that's going away, it can be done just as easily
+> with output redirection in `CVSROOT/loginfo`.
+>
+> --[[schmonz]]
+
+>> If the error messages screw up the CGI they must go to stdout.
+>> I thought we had stderr even in the the CVS dark ages. ;-) --[[Joey]] 
+
+>>> Some messages go to stderr, but definitely not all. That's why
+>>> I wound up reaching for IPC::Cmd, to execute the command line
+>>> safely while shutting CVS up. Anyway, I've tested what happens
+>>> if a configured post-commit hook is missing, and it seems fine,
+>>> probably also thanks to IPC::Cmd.
+>>> --[[schmonz]]