drm/atomic: better doc for implicit vs explicit fencing

Note that a pile of drivers don't seem to take implicit fencing into
account, or at least don't call drm_atoimc_set_fence_for_plane().
Cc'ing relevant people, or at least some. Some drivers also look like
they don't disable implicit fencing (e.g. amdgpu) because the explicit
fences and implicit fences are handled by entirely independent code
paths.

I also wonder whether we shouldn't just make the recommended helpers
the default ones, since a lot of drivers don't bother to handle the
implicit fences at all it seems. The helpers won't blow up even for
non-GEM drivers or GEM drivers which don't fill out the gem bo
pointers in struct drm_framebuffer.

v2: Comments from Eric.

Cc: Gerd Hoffmann <kraxel@redhat.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Sinclair Yeh <syeh@vmware.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Gustavo Padovan <gustavo@padovan.org>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Sean Paul <seanpaul@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180405154449.23038-7-daniel.vetter@ffwll.ch

drivers/gpu/drm/drm_atomic.c

@@ -1496,6 +1496,14 @@ EXPORT_SYMBOL(drm_atomic_set_fb_for_plane);
  * Otherwise, if &drm_plane_state.fence is not set this function we just set it
  * with the received implicit fence. In both cases this function consumes a
  * reference for @fence.
+ *
+ * This way explicit fencing can be used to overrule implicit fencing, which is
+ * important to make explicit fencing use-cases work: One example is using one
+ * buffer for 2 screens with different refresh rates. Implicit fencing will
+ * clamp rendering to the refresh rate of the slower screen, whereas explicit
+ * fence allows 2 independent render and display loops on a single buffer. If a
+ * driver allows obeys both implicit and explicit fences for plane updates, then
+ * it will break all the benefits of explicit fencing.
  */
 void
 drm_atomic_set_fence_for_plane(struct drm_plane_state *plane_state,

include/drm/drm_modeset_helper_vtables.h

@@ -1004,11 +1004,14 @@ struct drm_plane_helper_funcs {
 	 * This function must not block for outstanding rendering, since it is
 	 * called in the context of the atomic IOCTL even for async commits to
 	 * be able to return any errors to userspace. Instead the recommended
-	 * way is to fill out the fence member of the passed-in
+	 * way is to fill out the &drm_plane_state.fence of the passed-in
 	 * &drm_plane_state. If the driver doesn't support native fences then
 	 * equivalent functionality should be implemented through private
 	 * members in the plane structure.
 	 *
+	 * Drivers which always have their buffers pinned should use
+	 * drm_gem_fb_prepare_fb() for this hook.
+	 *
 	 * The helpers will call @cleanup_fb with matching arguments for every
 	 * successful call to this hook.
 	 *

include/drm/drm_plane.h

@@ -80,8 +80,15 @@ struct drm_plane_state {
 	/**
 	 * @fence:
 	 *
-	 * Optional fence to wait for before scanning out @fb. Do not write this
-	 * directly, use drm_atomic_set_fence_for_plane()
+	 * Optional fence to wait for before scanning out @fb. The core atomic
+	 * code will set this when userspace is using explicit fencing. Do not
+	 * write this directly for a driver's implicit fence, use
+	 * drm_atomic_set_fence_for_plane() to ensure that an explicit fence is
+	 * preserved.
+	 *
+	 * Drivers should store any implicit fence in this from their
+	 * &drm_plane_helper.prepare_fb callback. See drm_gem_fb_prepare_fb()
+	 * and drm_gem_fb_simple_display_pipe_prepare_fb() for suitable helpers.
 	 */
 	struct dma_fence *fence;