On Sat, 16 Aug 2003, Jeremy Fitzhardinge wrote:
> On Tue, 2003-08-12 at 00:16, Maneesh Soni wrote:
> > IMHO, it will be nice to have active maintainer for autofs4, particularly for
> > keeping autofs4 in sync with VFS layer changes. This is necessary as I found
> > autofs4 interacting with VFS (dcache and vfsmount) in a way no other file
> > system does.
> >
> > The problem is with is_tree_busy() routine where we check and depend on
> > dentry ref count to find busy dentries. Though it is done under dcache_lock
> > but I don't think taking dcache_lock will guarantee consistent reference count
> > for dentries. We have other people doing dget() or dput() without taking
> > dcache_lock. Further, introduction of RCU based dcache lookup in 2.5.61 it is
> > more susceptible to races.
>
> I've actually been thinking about this for the last week or so. I think
> the correct solution is to make is_tree_busy a three-phase affair:
> 1. get all the vfsmnt structures and inc their refcounts
> 2. take dcache_lock and do the core is_tree_busy logic
> 3. traverse the list of vfsmnts we got in 1 and release them
>
> The current code gets itself into a tangle because its mixing the lock
> ranking of vfsmnt_lock and dcache_lock, and only happens to get away
> with it (I think the race can only actually trigger if someone plays
> namespace manipulation games while automountd is trying to run an
> expire, which seems unlikely since only root can play such games, and
> presumably they're going to get autofs manage its own filesystems).
>
Hi Jeremy,
Do we have to use the reference counts?
I am thinking that I can rewrite the expire module without using them but
my understanding of the problem certainly not as clear as yours.
--
,-._|\ Ian Kent
/ \ Perth, Western Australia
*_.--._/ E-mail: [EMAIL PROTECTED]
v Web: http://themaw.net/
_______________________________________________
autofs mailing list
[EMAIL PROTECTED]
http://linux.kernel.org/mailman/listinfo/autofs
