drm/i915: Retry gtt fault when out of fence registers
authorVille Syrjälä <ville.syrjala@linux.intel.com>
Thu, 12 Oct 2023 13:28:01 +0000 (16:28 +0300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 25 Oct 2023 10:03:05 +0000 (12:03 +0200)
commitadd2eeda1d6e3ead1d61db70e37615e5ef6c5e75
treea71b1875d0315eeab35839c245e1a7512f6986b2
parentf691ec5a548257edb3aacd952e2a574e4e57b2c4
drm/i915: Retry gtt fault when out of fence registers

commit e339c6d628fe66c9b64bf31040a55770952aec57 upstream.

If we can't find a free fence register to handle a fault in the GMADR
range just return VM_FAULT_NOPAGE without populating the PTE so that
userspace will retry the access and trigger another fault. Eventually
we should find a free fence and the fault will get properly handled.

A further improvement idea might be to reserve a fence (or one per CPU?)
for the express purpose of handling faults without having to retry. But
that would require some additional work.

Looks like this may have gotten broken originally by
commit 39965b376601 ("drm/i915: don't trash the gtt when running out of fences")
as that changed the errno to -EDEADLK which wasn't handle by the gtt
fault code either. But later in commit 2feeb52859fc ("drm/i915/gt: Fix
-EDEADLK handling regression") I changed it again to -ENOBUFS as -EDEADLK
was now getting used for the ww mutex dance. So this fix only makes
sense after that last commit.

Cc: stable@vger.kernel.org
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/9479
Fixes: 2feeb52859fc ("drm/i915/gt: Fix -EDEADLK handling regression")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231012132801.16292-1-ville.syrjala@linux.intel.com
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
(cherry picked from commit 7f403caabe811b88ab0de3811ff3f4782c415761)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/gpu/drm/i915/gem/i915_gem_mman.c