[PATCH v2 hmm 11/11] mm/hmm: Remove confusing comment and logic from hmm_release
Jason Gunthorpe
jgg at mellanox.com
Mon Jun 10 16:02:56 UTC 2019
On Fri, Jun 07, 2019 at 02:37:07PM -0700, Ralph Campbell wrote:
>
> On 6/6/19 11:44 AM, Jason Gunthorpe wrote:
> > From: Jason Gunthorpe <jgg at mellanox.com>
> >
> > hmm_release() is called exactly once per hmm. ops->release() cannot
> > accidentally trigger any action that would recurse back onto
> > hmm->mirrors_sem.
> >
> > This fixes a use after-free race of the form:
> >
> > CPU0 CPU1
> > hmm_release()
> > up_write(&hmm->mirrors_sem);
> > hmm_mirror_unregister(mirror)
> > down_write(&hmm->mirrors_sem);
> > up_write(&hmm->mirrors_sem);
> > kfree(mirror)
> > mirror->ops->release(mirror)
> >
> > The only user we have today for ops->release is an empty function, so this
> > is unambiguously safe.
> >
> > As a consequence of plugging this race drivers are not allowed to
> > register/unregister mirrors from within a release op.
> >
> > Signed-off-by: Jason Gunthorpe <jgg at mellanox.com>
>
> I agree with the analysis above but I'm not sure that release() will
> always be an empty function. It might be more efficient to write back
> all data migrated to a device "in one pass" instead of relying
> on unmap_vmas() calling hmm_start_range_invalidate() per VMA.
I think we have to focus on the *current* kernel - and we have two
users of release, nouveau_svm.c is empty and amdgpu_mn.c does
schedule_work() - so I believe we should go ahead with this simple
solution to the actual race today that both of those will suffer from.
If we find a need for a more complex version then it can be debated
and justified with proper context...
Ok?
Jason
More information about the amd-gfx
mailing list