drm/radeon: hold reference to fences in radeon_sa_bo_new (3.17 and older)

[Backport of upstream commit f6ff4f67, with
 an additional NULL pointer guard that is required for kernels 3.17 and older.

 To be precise, any kernel that does *not* have commit 954605ca "drm/radeon:
 use common fence implementation for fences, v4" requires this additional
 NULL pointer guard.]

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> (v1 + fix)
Tested-by: Lutz Euler <lutz.euler@freenet.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/gpu/drm/radeon/radeon_sa.c

@@ -349,8 +349,15 @@ 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) {
+			if (fences[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 && block) {