Re: [PATCH] configfs: Silence lockdep on mkdir(), rmdir() and configfs_depend_item()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2008-12-18 at 12:56 +0100, Peter Zijlstra wrote:
> On Thu, 2008-12-18 at 01:27 -0800, Joel Becker wrote:
> 
> > 	It's about the default_groups and how they build up and tear
> > down small bits of tree.
> > 	A simple creation of a config_item, a mkdir(2), is a normal VFS
> > lock set and doesn't make lockdep unhappy.  But if the new config_item
> > has a default_group or two, they need locking too.  Not so much on
> > mkdir(2), but on rmdir(2).
> 
> Hohumm,..
> 
> So the problem is that mkdir() doesn't just create a single entity but a
> whole tree:
> 
> configfs:/my_subsystem/$ mkdir foo
> 
> might result in:
> 
>  foo/
>  foo/A/
>  foo/B/
>  foo/B/C/
> 
> which on rmdir foo you'd have to tear down, but only if its that exact
> tree and not when say A has any user created directories.
> 
> VFS mkdir A/blah only synchronizes on A.i_mutex and checks S_DEAD to
> avoid races with rmdir A - which would lock first parent(A).i_mutex and
> then A.i_mutex before detaching A and marking it S_DEAD.
> 
> So what you're now doing is locking the full foo/ subtree in order to
> check there is no user content and block mkdir/creat from generating any
> - which is where the trouble comes from, right?
> 
> Like said on IRC, the whole populated thing made me think of
> mount/umount (steven whitehouse seems to have had a similar notion).
> 
> You basically want to synchronize any user mkdir/creat against foo
> instead of just the new parent so that rmdir foo can tell if there is
> any such content without having to lock the whole subtree.
> 
> That would mean them locking both foo and the new parent (when they're
> not one and the same). Trouble seems to be that vfs_mkdir() and such
> already have their new parent locked, which means you cannot go about
> locking foo anymore. But that would have resulted in a 3 deep
> lock-chain.
> 
> (and I don't see any filesystem hooks in user_path_parent() -- which is
> probably a good thing)
> 
> 
> Bugger..

In fact, both (configfs) mkdir and rmdir seem to synchronize on
su_mutex..

 mkdir B/C/bar

   C.i_mutex
     su_mutex

vs

 rmdir foo

   parent(foo).i_mutex
     foo.i_mutex
       su_mutex


once holding the rmdir su_mutex you can check foo's user-content, since
any mkdir will be blocked. All you have to do is then re-validate in
mkdir's su_mutex that !IS_DEADDIR(C).

Does that sound plausible, or am I missing something obvious.. ?