While the range notifier is executing, we have the write-side mmu interval
seqlock, and mmu_interval_read_retry() is always returning true,
which means that if amdgpu_cs_submit grabs the notifier lock during the
fence wait, it will retry anyway when checking the userptr validity
and block when retrying in mmu_interval_read_begin().
(See the extensive comments in mmu_interval_read_begin())
Hence we can release the notifier lock before the fence wait and avoid
a device-wide command submission block during invalidation.
Cc: Christian König <christian.koenig@xxxxxxx>
Signed-off-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
index 4b153daf283d..d3d340a6129c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
@@ -75,9 +75,10 @@ static bool amdgpu_mn_invalidate_gfx(struct mmu_interval_notifier *mni,
mmu_interval_set_seq(mni, cur_seq);
+ mutex_unlock(&adev->notifier_lock);
+
r = dma_resv_wait_timeout(bo->tbo.base.resv, true, false,
MAX_SCHEDULE_TIMEOUT);
- mutex_unlock(&adev->notifier_lock);
if (r <= 0)
DRM_ERROR("(%ld) failed to wait for user bo\n", r);
return true;
--
2.31.1
