On Mon, 2025-10-13 at 20:19 +0300, Ville Syrjälä wrote:
> On Mon, Oct 13, 2025 at 04:52:04PM +0300, Jani Nikula wrote:
> > On Thu, 18 Sep 2025, Jani Nikula <[email protected]> wrote:
> > > On Tue, 16 Sep 2025, Ville Syrjälä
> > > <[email protected]> wrote:
> > > > For now I'd be happy if someone just nukes that bogus page
> > > > alignemnt
> > > > of the stride on xe, allowing i915 and xe to use the same code
> > > > here.
> > > 
> > > I hope just [1] is enough for this.
> > > 
> > > [1]
> > > https://lore.kernel.org/r/7f4972104de8b179d5724ae83892ee294d3f3fd3.1758184771.git.jani.nik...@intel.com
> > 
> > So that regressed [2]. (Bisected internally, unfortunately not
> > reported
> > on the gitlab issue.) Any ideas, before I go on and resurrect this
> > patch
> > adding different strides for i915 and xe?
> 
> That bisect doesn't make any real sense to me unless there's an
> existing
> bug in the xe code where it fails to pin (and somehow the smaller
> stride
> alignment makes it fail) but it still reports success to the caller. 

Checked out drm-tip at 9c80ebfdd8460e69b35f5382f3e93a2a33a64e4f. Using
that hash + setup where seen earlier the issue is easily reproducing.
Reverted patches from the set which Jani is referring. After reverting
"[PATCH v2 01/10] drm/xe/fbdev: use the same 64-byte stride alignment
as i915" issue is not reproducing anymore.

I couldn't reproduce the issue using latest drm-tip but it is anyways
sporadic. It is seen in our CI testing with recent drm-tip as well.

BR,

Jouni Högander
> 
> Unfortunately that code in xe is completely illegible due the scoped
> guard mess. So it's darn near impossible to see with a visual
> inspection
> where it might silently fail. I think someone will need to sprinkle
> debugs all over that code to track what is happening to the pin
> count.
> 

Reply via email to