drm/intel: fix up set_tiling for untiled->tiled transition
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 18 Mar 2010 08:22:12 +0000 (09:22 +0100)
committerEric Anholt <eric@anholt.net>
Thu, 18 Mar 2010 23:48:02 +0000 (16:48 -0700)
Bug introduced in

commit 10ae9bd25acf394c8fa2f9d795dfa9cec4d19ed6
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Mon Feb 1 13:59:17 2010 +0100

    drm/i915: blow away userspace mappings before fence change

The problem is that when there's no fence reg assigned and the object
is mapped at a fenceable offset in the gtt, the userspace mappings won't
be torn down. Which happens on untiled->tiled transition quite often
on 4th gen and later because there fencing does not have any special
alignment constraints (as opposed to 2nd and 3rd gen on which I've tested
the original commit).

Bugzilla: http://bugs.freedesktop.org/show_bug.cgi?id=26993
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: Eric Anholt <eric@anholt.net> (fixes OpenArena)
Signed-off-by: Eric Anholt <eric@anholt.net>
drivers/gpu/drm/i915/i915_gem_tiling.c

index b5c55d88ff76f260f87a0dc64b140739dcc7ec4f..c01c878e51baf3ff206388c32312f685a73156f5 100644 (file)
@@ -325,9 +325,12 @@ i915_gem_set_tiling(struct drm_device *dev, void *data,
                 * need to ensure that any fence register is cleared.
                 */
                if (!i915_gem_object_fence_offset_ok(obj, args->tiling_mode))
-                   ret = i915_gem_object_unbind(obj);
+                       ret = i915_gem_object_unbind(obj);
+               else if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
+                       ret = i915_gem_object_put_fence_reg(obj);
                else
-                   ret = i915_gem_object_put_fence_reg(obj);
+                       i915_gem_release_mmap(obj);
+
                if (ret != 0) {
                        WARN(ret != -ERESTARTSYS,
                             "failed to reset object for tiling switch");