X-Git-Url: http://git.vanrenterghem.biz/git.ikiwiki.info.git/blobdiff_plain/4b80bb5f9c43919fb86fd4c144280ac9246d20c8..182043667058d302d33aefa12dcb8678799d7fc7:/doc/post-commit/discussion.mdwn?ds=inline

diff --git a/doc/post-commit/discussion.mdwn b/doc/post-commit/discussion.mdwn
index c15396978..fc0a27ee4 100644
--- a/doc/post-commit/discussion.mdwn
+++ b/doc/post-commit/discussion.mdwn
@@ -48,13 +48,76 @@ Thanks
 
 -----
 
-I'm working on an [[rcs]] plugin for CVS, adapted from `svn.pm`, in order to integrate ikiwiki at sites where that's all they've got. What's working so far: web commit (post-commit hook and all), diff, add (under certain conditions), and remove. What's not working: with rcs_add(), iff any of the new page's parent dirs aren't already under CVS control and the post-commit hook is enabled, the browser and ikiwiki stall for several seconds trying to add it, then time out. (If I kill ikiwiki when this is happening, it cvs adds the topmost parent that needed adding; if I wait for timeout, it doesn't. I think.) If I disable the post-commit hook and do the same kind of thing, the page is created and saved.
-
-In case you're lucky enough not to know, cvs adds on directories are weird -- they operate immediately against the repository, unlike file adds:
+I'm working on an [[rcs]] plugin for CVS, adapted from `svn.pm`, in order
+to integrate ikiwiki at sites where that's all they've got. What's working
+so far: web commit (post-commit hook and all), diff, add (under certain
+conditions), and remove. What's not working: with rcs_add(), iff any of the
+new page's parent dirs aren't already under CVS control and the post-commit
+hook is enabled, the browser and ikiwiki stall for several seconds trying
+to add it, then time out. (If I kill ikiwiki when this is happening, it cvs
+adds the topmost parent that needed adding; if I wait for timeout, it
+doesn't. I think.) If I disable the post-commit hook and do the same kind
+of thing, the page is created and saved.
+
+In case you're lucky enough not to know, cvs adds on directories are weird
+-- they operate immediately against the repository, unlike file adds:
 
     $ cvs add randomdir
     Directory /Users/schmonz/Documents/cvswiki/repository/ikiwiki/randomdir added to the repository
 
-I was able to work out that when I'm seeing this page save misbehavior, my plugin is somewhere inside `system("cvs", "-Q", "add", "$file")`, which was never returning. If I changed it to anything other than cvs it iterated correctly over all the parent dirs which needed to be added to CVS, in the proper order. (cvs add isn't recursive, sadly.)
+I was able to work out that when I'm seeing this page save misbehavior, my
+plugin is somewhere inside `system("cvs", "-Q", "add", "$file")`, which was
+never returning. If I changed it to anything other than cvs it iterated
+correctly over all the parent dirs which needed to be added to CVS, in the
+proper order. (cvs add isn't recursive, sadly.)
 
 Can you offer an educated guess what's going wrong here? --[[Schmonz]]
+
+> Got `rcs_recentchanges` working, believe it or not, thanks to [cvsps](http://www.cobite.com/cvsps/). If I can figure out this interaction between the post-commit hook and `cvs add` on directories, the CVS plugin is mostly done. Could it be a locking issue? Where should I be looking? Any suggestions appreciated. --[[Schmonz]]
+
+>> Okay, it is definitely a locking issue. First, on the conjecture that
+>> `cvs add <directory>` was triggering the post-commit hook and confusing
+>> ikiwiki, I wrapped the ikiwiki post-commit binary with a shell script
+>> that exited 0 if the triggering file was a directory. The first half of
+>> the conjecture was correct -- my wrapper got triggered -- but the web
+>> add of `one/two/three.mdwn` (where `one` and `two` weren't existing
+>> CVS-controlled dirs) remained hung as before. There were two ikiwiki
+>> processes running. On a whim, I killed the one with the higher PID; `cvs
+>> add one` immediately completed successfully, then back to a hang and two
+>> ikiwiki processes. I killed the newer one again and then `cvs add
+>> one/two` and `cvs add one/two/three.mdwn` completed and the web add was
+>> successful. --[[Schmonz]]
+
+>>> Aaaaaand I was wrong about the second half of the conjecture being
+>>> wrong. The wrapper script wasn't correctly identifying directories;
+>>> with that fixed, everything works. I've created a
+>>> [[rcs/cvs]] page. Thanks for listening. :-)
+>>> --[[Schmonz]]
+
+>> Here is a comment I committed to my laptop from Madrid Airport before
+>> your most recent updates, in case it's still useful:
+>>
+>> Locking certianly seems likely to be a problem. ikiwiki calls `rcs_add`
+>> *before* disabling the post-commit plugin, since all over VCS allow
+>> adding something in a staged manner. You can see this in, for example,
+>> `editpage.pm` lines 391+.
+>> 
+>> So I guess what happens is that ikiwiki has taken the wiki lock, calls
+>> `rcs_add`, which does a `cvs add`, which runs the post commit hook,
+>> since it is not disabled -- which blocks waiting for the wiki lock.
+>>
+>> I guess you can fix this in either of three ways: Modify lots of places
+>> in ikiwiki to disable the post commit hook before calling `rcs_add`,
+>> or make cvs's `rcs_add` temporarily disable the commit hook and
+>> re-enable it (but only if it was not already disabled, somehow),
+>> or make cvs's `rcs_add` only make note that it needs to call `cvs add`
+>> later, and do so at `rcs_commit`. The last of these seems easist,
+>> especially since ikiwiki always commits after an add, in the same
+>> process, so you could just use a temporary list of things to add.
+>> --[[Joey]] 
+
+>>> Thanks for the comments. Attempting to set up a wiki on a different system with a different version of `cvs`, I've encountered a new locking problem within CVS: `cvs commit` takes a write lock, post-commit ikiwiki calls `rcs_update()`, `cvs update` wants a read lock and blocks. The easiest fix I can think of is to make `cvs commit` return and relinquish its lock -- so instead of my wrapper script `exec`ing ikiwiki's post-commit hook, I amp it off and exit 0. Seems to do the trick and, if I grok ikiwiki's behavior here, is not dangerous. (Beats me why my development `cvs` doesn't behave the same WRT locking.)
+
+>>> I was all set to take your third suggestion, but now that there's more than one CVS oddity fixed trivially in a wrapper script, I think I prefer doing it that way.
+
+>>> I'd be glad for the CVS plugin to be included in ikiwiki, if and when you deem it ready. Please let me know what needs to be done for that to happen. --[[Schmonz]]