Re: [PATCH] vfs namespace: Don't assume mount namespace has valid root

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

On Wed, Mar 14, 2012 at 11:03:43PM +0200, Sasha Levin wrote:
> > And that's all pointers to mnt_namespace that ever exist, aside of function
> > arguments and local variables. ?I'm not saying that I couldn't have possibly
> > fucked it up, but from rereading that code it doesn't look like we could
> > end up with dangling pointers to already freed instances...
> 
> I'm trying to find the exact chain of events leading it it at the
> moment, but it reproduces rather easily - so if you have any ideas on
> figuring it out I'm happy to try anything.

Which kernel, for starters?  I'd probably add dumping call chain +
return value in alloc_mnt_ns(), call chain + pointer being freed in
final put_mnt_ns() and failure exit of dup_mnt_ns(), address of
mnt_ns and value assigned to ->root on assignments to ->root in
create_mnt_ns() and dup_mnt_ns() and mnt_ns in dup_mnt_ns() if it
happens to have NULL ->root.

That should give you full history of allocation/freeing mnt_namespace
instances and of assignments to anyone's ->root.  Ought to make a sane
starting point...