[Intel-gfx] [PATCH] drm/i915/gem: Atomically invalidate userptr on mmu-notifier
Cavitt, Jonathan
jonathan.cavitt at intel.com
Mon Dec 18 18:33:44 UTC 2023
-----Original Message-----
From: Andi Shyti <andi.shyti at linux.intel.com>
Sent: Monday, December 18, 2023 8:06 AM
To: Cavitt, Jonathan <jonathan.cavitt at intel.com>
Cc: intel-gfx at lists.freedesktop.org; Gupta, saurabhg <saurabhg.gupta at intel.com>; chris.p.wilson at linux.intel.com
Subject: Re: [Intel-gfx] [PATCH] drm/i915/gem: Atomically invalidate userptr on mmu-notifier
>
> Hi Jonathan,
>
> On Tue, Nov 28, 2023 at 08:25:05AM -0800, Jonathan Cavitt wrote:
> > Never block for outstanding work on userptr object upon receipt of a
> > mmu-notifier. The reason we originally did so was to immediately unbind
> > the userptr and unpin its pages, but since that has been dropped in
> > commit b4b9731b02c3c ("drm/i915: Simplify userptr locking"), we never
> > return the pages to the system i.e. never drop our page->mapcount and so
> > do not allow the page and CPU PTE to be revoked. Based on this history,
> > we know we are safe to drop the wait entirely.
> >
> > Upon return from mmu-notifier, we will still have the userptr pages
> > pinned preventing the following PTE operation (such as try_to_unmap)
> > adjusting the vm_area_struct, so it is safe to keep the pages around for
> > as long as we still have i/o pending.
> >
> > We do not have any means currently to asynchronously revalidate the
> > userptr pages, that is always prior to next use.
> >
> > Signed-off-by: Chris Wilson <chris.p.wilson at linux.intel.com>
> > Signed-off-by: Jonathan Cavitt <jonathan.cavitt at intel.com>
>
> Reviewed-by: Andi Shyti <andi.shyti at linux.intel.com>
Thank you for the review! I don't think I have permission to push this upstream,
though, so you or someone else will have to complete the push.
-Jonathan Cavitt
>
> Thanks,
> Andi
>
More information about the Intel-gfx
mailing list