drm/amd/display: Skip fast cursor updates for fb changes
authorNicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Fri, 14 Dec 2018 17:26:58 +0000 (12:26 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 18 Dec 2018 22:39:26 +0000 (17:39 -0500)
[Why]
The behavior of drm_atomic_helper_cleanup_planes differs depending on
whether the commit was asynchronous or not. When it's called from
amdgpu_dm_atomic_commit_tail during a typical atomic commit the
plane state has been swapped so it calls cleanup_fb on the old plane
state.

However, in the asynchronous commit codepath the call to
drm_atomic_helper_commit also calls dm_plane_helper_cleanup_fb after
atomic_async_update has been called. Since the plane state is updated
in place and has not been swapped the cleanup_fb call affects the new
plane state.

This results in a use after free for the given sequence:

- Fast update, fb1 pin/ref, fb1 unpin/unref
- Fast update, fb2 pin/ref, fb2 unpin/unref
- Slow update, fb1 pin/ref, fb2 unpin/unref
- Fast update, fb2 pin/ref -> use after free. bug

[How]
Disallow framebuffer changes in the fast path. Since this includes
a NULL framebuffer, this means that only framebuffers that have
been previously pin+ref at least once will be used, preventing a
use after free.

This has a significant throughput reduction for cursor updates where
the framebuffer changes. For most desktop usage this isn't a problem,
but it does introduce performance regressions for two specific IGT
tests:

- cursor-vs-flip-toggle
- cursor-vs-flip-varying-size

Fixes: 2cc751931afc ("drm/amd/display: Add fast path for cursor plane updates")
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c

index c13856a..753c6c2 100644 (file)
@@ -3628,10 +3628,20 @@ static int dm_plane_atomic_check(struct drm_plane *plane,
 static int dm_plane_atomic_async_check(struct drm_plane *plane,
                                       struct drm_plane_state *new_plane_state)
 {
+       struct drm_plane_state *old_plane_state =
+               drm_atomic_get_old_plane_state(new_plane_state->state, plane);
+
        /* Only support async updates on cursor planes. */
        if (plane->type != DRM_PLANE_TYPE_CURSOR)
                return -EINVAL;
 
+       /*
+        * DRM calls prepare_fb and cleanup_fb on new_plane_state for
+        * async commits so don't allow fb changes.
+        */
+       if (old_plane_state->fb != new_plane_state->fb)
+               return -EINVAL;
+
        return 0;
 }