Re: [PATCH] drm/xe: Unmap userptr in MMU invalidation notifier

[Matthew Brost <matthew.brost@intel.com>]

On Mon, Apr 29, 2024 at 07:55:22AM -0600, Zeng, Oak wrote:
> Hi Matt
> 
> > -----Original Message-----
> > From: Intel-xe <intel-xe-bounces@lists.freedesktop.org> On Behalf Of
> > Matthew Brost
> > Sent: Friday, April 26, 2024 7:33 PM
> > To: intel-xe@lists.freedesktop.org
> > Cc: Brost, Matthew <matthew.brost@intel.com>; Thomas Hellström
> > <thomas.hellstrom@linux.intel.com>; stable@vger.kernel.org
> > Subject: [PATCH] drm/xe: Unmap userptr in MMU invalidation notifier
> > 
> > To be secure, when a userptr is invalidated the pages should be dma
> > unmapped ensuring the device can no longer touch the invalidated pages.
> > 
> > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> > Fixes: 12f4b58a37f4 ("drm/xe: Use hmm_range_fault to populate user
> > pages")
> > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > Cc: stable@vger.kernel.org # 6.8
> > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_vm.c | 3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > index dfd31b346021..964a5b4d47d8 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -637,6 +637,9 @@ static bool vma_userptr_invalidate(struct
> > mmu_interval_notifier *mni,
> >  		XE_WARN_ON(err);
> >  	}
> > 
> > +	if (userptr->sg)
> > +		xe_hmm_userptr_free_sg(uvma);
> 
> Just some thoughts here. I think when we introduce system allocator, above should be made conditional. We should dma unmap userptr only for normal userptr but not for userptr created for system allocator (fault usrptr in the system allocator series). Because for system allocator the dma-unmapping would be part of the garbage collector and vma destroy process. Right?
> 

I don't think it should be conditional. In any case when a CPU address
is invalidated we need to ensure the dma mapping (IOMMU mapping) is
also invalid to ensure no path to the old (invalidate) pages exists.
This is an extra security that must be enforced. With removing the dma
mapping, in theory rouge accesses from the GPU could still access the
old pages.

Matt

> Oak 
> 
> > +
> >  	trace_xe_vma_userptr_invalidate_complete(vma);
> > 
> >  	return true;
> > --
> > 2.34.1
>