Re: [patch 3/6] vfs: mountinfo stable peer group id

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Mar 19, 2008 at 05:41:15PM +0100, Miklos Szeredi wrote:
> > > From: Miklos Szeredi <mszeredi@suse.cz>
> > Um?  Do you ever need to take it outside of vfsmount_lock?
> > 
> 
> Tried to think this through:
> 
> It's always called with namespace_sem, which is enough, no need for a
> new lock.  The bigger problem, is that it _is_ called with
> vfsmount_lock in one case, which is bad, since the allocation may
> sleep.

It is called with vfsmount_lock in *all* cases.  You've missed one
in umount_tree(), BTW; you won't block in that case, though.

> That is in do_change_type().  But do we really need to hold
> vfsmount_lock in that case?

Not the issue.

>  I think not, the propagation tree has no
> relevance outside namespace_sem, so that one should be sufficient.

Callers manipulate more than propagation tree.  Note that e.g.
umount_tree() changes all sorts of data structures, including ones
that are traversed without namespace_sem.

I _really_ don't like the idea of different locking rules for caller
of a function depending on the value of argument of that function.
They are complicated enough as it is.

Argh...  OK, I'll try to put something together tonight, after I get some
sleep - 31 hours of uptime _suck_ ;-/  BTW, on top of everything else,
the current variant plays interesting games with CL_PROPAGATION behaviour
and I really don't like the look of what it's doing there.  Later...