[Linux-cluster] IMAP server clustering ...

[linux-cluster at spam.dragonhold.org]

On Mon, Nov 01, 2004 at 10:25:03AM -0700, Michael Gale wrote:
> 
> I was reading up on Courier IMAP which use dot-locking with support NFS 
> mounted maildirs.
> 
> So would that application not take care of the locking ?

No, it's operating at the wrong level.

(starting from nothing cached)

Think about it this way - you create a new file on the disk (say the lock file).  The other
machine then tries to access the directory.  It scans down from the root of the partition
(successfully, since nothing has changed), and gets to the directory.  This finds the
lockfile.

So far so good.

Now the 1st machine deletes the lockfile.  However, the 2nd machine still has this cached as
locked - and therefore doesn't notice.


--

Other example.

Both machines read the directory (and it's not locked).  Next machine 1 locks it.  Even if
reiserfs writes this lock back to disk (which it will eventually), the 1st machine doesn't
know, since it still has a cached version of the directory which shows that the file doesn't
exist.  Now both can lock (successfully as far as they are concerned).


--

Final example.

Machines 1&2 both have the lock.  One deletes a file, and updates the disk.  The 2nd adds a
file, and then updates the directory with it's version (which still has the first file in
it).

This means you've got the file pointing to the blocks where it exists, but the blocks have
been freed.

If you do that with another directory (create a new IMAP folder) rather than file, and it
gets even worse - the machine that didn't create it won't know that those inodes are a
directory, so will happily then write a file over it.

--

If this doesn't make sense (quite possible, I've not worked through the examples properly),
just work it through on paper.  Remember that the machines have no reason to doubt their
cached copy of the data, and they will cache as much as possible.

Go through what could happen from a starting point of the disk & caches agreeing,
remembering that not only is read data cached, data is not written back out immediately.


Graham