drm/radeon: add error handling to radeon_vm_unbind_locked

Waiting for a fence can fail for different reasons,
the most common is a deadlock.

Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 14 insertions, 3 deletions

diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
index 2b34c1a91421..ee11c5073726 100644
--- a/drivers/gpu/drm/radeon/radeon_gart.c
+++ b/drivers/gpu/drm/radeon/radeon_gart.c
@@ -316,10 +316,21 @@ static void radeon_vm_unbind_locked(struct radeon_device *rdev,
         }
 
         /* wait for vm use to end */
-        if (vm->fence) {
-                radeon_fence_wait(vm->fence, false);
-                radeon_fence_unref(&vm->fence);
+        while (vm->fence) {
+                int r;
+                r = radeon_fence_wait(vm->fence, false);
+                if (r)
+                        DRM_ERROR("error while waiting for fence: %d\n", r);
+                if (r == -EDEADLK) {
+                        mutex_unlock(&rdev->vm_manager.lock);
+                        r = radeon_gpu_reset(rdev);
+                        mutex_lock(&rdev->vm_manager.lock);
+                        if (!r)
+                                continue;
+                }
+                break;
         }
+        radeon_fence_unref(&vm->fence);
 
         /* hw unbind */
         rdev->vm_manager.funcs->unbind(rdev, vm);