Re: drm: xe: Kernel-submitted job timed out

On Tue, 18 Aug 2026 at 23:55, Linus Torvalds
<[email protected]> wrote:
>
> It turns out that now I can generate those drm job timedout issues at will.
>
> [  +1.222075] xe 0000:4b:00.0: [drm] Tile0: GT0: Engine reset:

Ok. I've spent all day today on chasing this down in between merge
window work, since it finally *was* repeatable.

And the fix ends up being basically a one-liner, even if the debug
session to get there was not.

The Xe driver takes the memory address used for the memory stolen for
CCS, and rounds it up to the nearest 128kB area.

And that is very VERY wrong, because it means that the xe driver will
then *use* that part of memory that wasn't at a 128kB boundary. And
the HW engine will too and write to it.

And when that memory happens ot be used for GPU page tables, very bad
things happen.

I bet this also explains some occasional random screen corruption I've
seen - when the memory isn't used for something as important as a page
table, it "only" corrupts random bitmap memory and the like.

I committed the fix as commit 818bebeb63dd ("drm/xe: Don't hand out
the flat CCS storage as usable VRAM").

Adding some people from commit 37173392741c ("drm/xe/vram: fix ccs
offset calculation") to the cc, because that's where the problem came
from. Two years ago. I'm not sure why it became so repeatable for me
now, but some user space behavior change clearly triggered it now
pretty much every single boot.

                 Linus