drm/radeon: hold reference to fences in radeon_sa_bo_new
authorNicolai Hähnle <nicolai.haehnle@amd.com>
Fri, 5 Feb 2016 19:35:53 +0000 (14:35 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 10 Feb 2016 19:07:44 +0000 (14:07 -0500)
An arbitrary amount of time can pass between spin_unlock and
radeon_fence_wait_any, so we need to ensure that nobody frees the
fences from under us.

Based on the analogous fix for amdgpu.

Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org
drivers/gpu/drm/radeon/radeon_sa.c

index c507896aca45a43a3fe271b79b84e7db61bfb85a..197b157b73d09b96f7ced03c6703b6696c210793 100644 (file)
@@ -349,8 +349,13 @@ int radeon_sa_bo_new(struct radeon_device *rdev,
                        /* see if we can skip over some allocations */
                } while (radeon_sa_bo_next_hole(sa_manager, fences, tries));
 
+               for (i = 0; i < RADEON_NUM_RINGS; ++i)
+                       radeon_fence_ref(fences[i]);
+
                spin_unlock(&sa_manager->wq.lock);
                r = radeon_fence_wait_any(rdev, fences, false);
+               for (i = 0; i < RADEON_NUM_RINGS; ++i)
+                       radeon_fence_unref(&fences[i]);
                spin_lock(&sa_manager->wq.lock);
                /* if we have nothing to wait for block */
                if (r == -ENOENT) {