[PATCH v2 1/2] drm/i915/gem: Allow EXEC_CAPTURE on recoverable contexts on DG1
Joonas Lahtinen
joonas.lahtinen at linux.intel.com
Thu May 22 10:58:07 UTC 2025
Quoting Ville Syrjälä (2025-05-22 12:35:05)
> On Thu, May 22, 2025 at 09:25:04AM +0300, Joonas Lahtinen wrote:
> > (+ Tvrkto)
> >
> > Quoting Ville Syrjala (2025-04-11 17:43:12)
<SNIP>
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
> > > @@ -2013,7 +2013,7 @@ static int eb_capture_stage(struct i915_execbuffer *eb)
> > > continue;
> > >
> > > if (i915_gem_context_is_recoverable(eb->gem_context) &&
> > > - (IS_DGFX(eb->i915) || GRAPHICS_VER_FULL(eb->i915) > IP_VER(12, 0)))
> > > + GRAPHICS_VER_FULL(eb->i915) > IP_VER(12, 10))
> >
> > The IS_DGFX check was there because the error capture is expected to be
> > broken on anything with VRAM.
>
> I don't care. It's a regression that prevents current userspace
> from working.
(Spoiler: The userspace fix seems to be accepted and released.)
It's always bit murky when a platform stays under force_probe for
extended period of time, but it was never considered to be finalized from
memory management perspective at the time of adding this check.
Now you are just unblocking codepaths that are simply not expected to work,
and as it's in rather fragile part of the device resets so that's bit of
a no-go.
So if you really would prefer to drop this check for DG1, options would be to
implement the page copying for VRAM (probably bit much work) or alternatively
we could just ignore the flag for DG1 specifically.
> > If we have already submitted an userspace fix to remove that flag, that
> > would be the right way to go.
>
> There has a been an open pull request for that for who knows how long
> without any action.
>
> >
> > So reverting this for now.
>
> *You* make sure a userspace fix actually gets released then.
Per GIT history[1] it should already be part of media-driver release 25.2.3
which was released last week. Or am I looking at the wrong fix?
Regards, Joonas
[1] https://github.com/intel/media-driver/commit/93c07d9b4b96a78bab21f6acd4eb863f4313ea4a
More information about the Intel-gfx
mailing list