Re: [PATCH] fix race in mark_mounts_for_expiry()

[Jamie Lokier <jamie@shareable.org>]

Miklos Szeredi wrote:
> > Some archs already have an atomic_dec_if_positive() (see for instance
> > the PPC). It won't take much work to convert that to an
> > atomic_inc_if_positive().
> > 
> > For those arches that don't have that sort of thing, then writing a
> > generic atomic_inc_if_positive() using cmpxchg() will often be possible,
> > but there are exceptions (for instance the original 386 does not have a
> > cmpxchg, so there you will have to use something else).
> 
> The problem with introducing architecture specific code, is that it's
> just asking for new bugs.
> 
> If it's something used all over the kernel, than obviously it's OK,
> but for the sake of just one caller it's a bit crazy I think.

I agree.

And I think you're just adding to the case for removing mnt_namespace
entirely.  We'd still keep CLONE_NS, and users currently using
namespaces (in the normal ways) would see no difference.

mnt_namespace has these visible effects:

    - Prevents some tasks from mounting/umounting in a "foreign"
      namespace, even when they are granted access to the directory
      tree of the foreign namespace.

      It's not clear if the restriction is a useful security tool.

    - Causes every mount in a mount tree to be detached (independently),
      when last task associated with a namespace is destroyed.

And this invisible effect:

    - More concurrency than a global mount lock would have.

Is that all?  Are any of these effects important enough to keep?

-- Jamie