drm/atomic: Update docs around locking and commit sequencing

Both locking and especially sequencing of nonblocking commits have
evolved a lot. The details are all there, but I noticed that the big
picture and connections have fallen behind a bit. Apply polish.

Motivated by some review discussions with Thierry.

v2: Review from Thierry

Reviewed-by: Thierry Reding <treding@nvidia.com>
Cc: Thierry Reding <treding@nvidia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191204100011.859468-1-daniel.vetter@ffwll.ch

Documentation/gpu/drm-kms.rst

@@ -260,7 +260,8 @@ Taken all together there's two consequences for the atomic design:
   drm_connector_state <drm_connector_state>` for connectors. These are the only
   objects with userspace-visible and settable state. For internal state drivers
   can subclass these structures through embeddeding, or add entirely new state
-  structures for their globally shared hardware functions.
+  structures for their globally shared hardware functions, see :c:type:`struct
+  drm_private_state<drm_private_state>`.
 
 - An atomic update is assembled and validated as an entirely free-standing pile
   of structures within the :c:type:`drm_atomic_state <drm_atomic_state>`
@@ -269,6 +270,14 @@ Taken all together there's two consequences for the atomic design:
   to the driver and modeset objects. This way rolling back an update boils down
   to releasing memory and unreferencing objects like framebuffers.
 
+Locking of atomic state structures is internally using :c:type:`struct
+drm_modeset_lock <drm_modeset_lock>`. As a general rule the locking shouldn't be
+exposed to drivers, instead the right locks should be automatically acquired by
+any function that duplicates or peeks into a state, like e.g.
+:c:func:`drm_atomic_get_crtc_state()`.  Locking only protects the software data
+structure, ordering of committing state changes to hardware is sequenced using
+:c:type:`struct drm_crtc_commit <drm_crtc_commit>`.
+
 Read on in this chapter, and also in :ref:`drm_atomic_helper` for more detailed
 coverage of specific topics.
 

drivers/gpu/drm/drm_atomic.c

@@ -688,10 +688,12 @@ static void drm_atomic_plane_print_state(struct drm_printer *p,
  * associated state struct &drm_private_state.
  *
  * Similar to userspace-exposed objects, private state structures can be
- * acquired by calling drm_atomic_get_private_obj_state(). Since this function
+ * acquired by calling drm_atomic_get_private_obj_state(). This also takes care
- * does not take care of locking, drivers should wrap it for each type of
+ * of locking, hence drivers should not have a need to call drm_modeset_lock()
- * private state object they have with the required call to drm_modeset_lock()
+ * directly. Sequence of the actual hardware state commit is not handled,
- * for the corresponding &drm_modeset_lock.
+ * drivers might need to keep track of struct drm_crtc_commit within subclassed
+ * structure of &drm_private_state as necessary, e.g. similar to
+ * &drm_plane_state.commit. See also &drm_atomic_state.fake_commit.
  *
  * All private state structures contained in a &drm_atomic_state update can be
  * iterated using for_each_oldnew_private_obj_in_state(),

drivers/gpu/drm/drm_atomic_helper.c

@@ -1832,17 +1832,21 @@ EXPORT_SYMBOL(drm_atomic_helper_commit);
 /**
  * DOC: implementing nonblocking commit
  *
- * Nonblocking atomic commits have to be implemented in the following sequence:
+ * Nonblocking atomic commits should use struct &drm_crtc_commit to sequence
+ * different operations against each another. Locks, especially struct
+ * &drm_modeset_lock, should not be held in worker threads or any other
+ * asynchronous context used to commit the hardware state.
  *
- * 1. Run drm_atomic_helper_prepare_planes() first. This is the only function
+ * drm_atomic_helper_commit() implements the recommended sequence for
- * which commit needs to call which can fail, so we want to run it first and
+ * nonblocking commits, using drm_atomic_helper_setup_commit() internally:
+ *
+ * 1. Run drm_atomic_helper_prepare_planes(). Since this can fail and we
+ * need to propagate out of memory/VRAM errors to userspace, it must be called
  * synchronously.
  *
  * 2. Synchronize with any outstanding nonblocking commit worker threads which
- * might be affected the new state update. This can be done by either cancelling
+ * might be affected by the new state update. This is handled by
- * or flushing the work items, depending upon whether the driver can deal with
+ * drm_atomic_helper_setup_commit().
- * cancelled updates. Note that it is important to ensure that the framebuffer
- * cleanup is still done when cancelling.
  *
  * Asynchronous workers need to have sufficient parallelism to be able to run
  * different atomic commits on different CRTCs in parallel. The simplest way to
@@ -1853,21 +1857,29 @@ EXPORT_SYMBOL(drm_atomic_helper_commit);
  * must be done as one global operation, and enabling or disabling a CRTC can
  * take a long time. But even that is not required.
  *
+ * IMPORTANT: A &drm_atomic_state update for multiple CRTCs is sequenced
+ * against all CRTCs therein. Therefore for atomic state updates which only flip
+ * planes the driver must not get the struct &drm_crtc_state of unrelated CRTCs
+ * in its atomic check code: This would prevent committing of atomic updates to
+ * multiple CRTCs in parallel. In general, adding additional state structures
+ * should be avoided as much as possible, because this reduces parallelism in
+ * (nonblocking) commits, both due to locking and due to commit sequencing
+ * requirements.
+ *
  * 3. The software state is updated synchronously with
  * drm_atomic_helper_swap_state(). Doing this under the protection of all modeset
- * locks means concurrent callers never see inconsistent state. And doing this
+ * locks means concurrent callers never see inconsistent state. Note that commit
- * while it's guaranteed that no relevant nonblocking worker runs means that
+ * workers do not hold any locks; their access is only coordinated through
- * nonblocking workers do not need grab any locks. Actually they must not grab
+ * ordering. If workers would access state only through the pointers in the
- * locks, for otherwise the work flushing will deadlock.
+ * free-standing state objects (currently not the case for any driver) then even
+ * multiple pending commits could be in-flight at the same time.
  *
  * 4. Schedule a work item to do all subsequent steps, using the split-out
  * commit helpers: a) pre-plane commit b) plane commit c) post-plane commit and
  * then cleaning up the framebuffers after the old framebuffer is no longer
- * being displayed.
+ * being displayed. The scheduled work should synchronize against other workers
- *
+ * using the &drm_crtc_commit infrastructure as needed. See
- * The above scheme is implemented in the atomic helper libraries in
+ * drm_atomic_helper_setup_commit() for more details.
- * drm_atomic_helper_commit() using a bunch of helper functions. See
- * drm_atomic_helper_setup_commit() for a starting point.
  */
 
 static int stall_checks(struct drm_crtc *crtc, bool nonblock)
@@ -2096,7 +2108,7 @@ EXPORT_SYMBOL(drm_atomic_helper_setup_commit);
  *
  * This function waits for all preceeding commits that touch the same CRTC as
  * @old_state to both be committed to the hardware (as signalled by
- * drm_atomic_helper_commit_hw_done) and executed by the hardware (as signalled
+ * drm_atomic_helper_commit_hw_done()) and executed by the hardware (as signalled
  * by calling drm_crtc_send_vblank_event() on the &drm_crtc_state.event).
  *
  * This is part of the atomic helper support for nonblocking commits, see

include/drm/drm_atomic.h

@@ -60,8 +60,8 @@
  * 	wait for flip_done		<----
  * 	clean up atomic state
  *
- * The important bit to know is that cleanup_done is the terminal event, but the
+ * The important bit to know is that &cleanup_done is the terminal event, but the
- * ordering between flip_done and hw_done is entirely up to the specific driver
+ * ordering between &flip_done and &hw_done is entirely up to the specific driver
  * and modeset state change.
  *
  * For an implementation of how to use this look at
@@ -92,6 +92,9 @@ struct drm_crtc_commit {
 	 * commit is sent to userspace, or when an out-fence is singalled. Note
 	 * that for most hardware, in most cases this happens after @hw_done is
 	 * signalled.
+	 *
+	 * Completion of this stage is signalled implicitly by calling
+	 * drm_crtc_send_vblank_event() on &drm_crtc_state.event.
 	 */
 	struct completion flip_done;
 
@@ -107,6 +110,9 @@ struct drm_crtc_commit {
 	 * Note that this does not need to include separately reference-counted
 	 * resources like backing storage buffer pinning, or runtime pm
 	 * management.
+	 *
+	 * Drivers should call drm_atomic_helper_commit_hw_done() to signal
+	 * completion of this stage.
 	 */
 	struct completion hw_done;
 
@@ -118,6 +124,9 @@ struct drm_crtc_commit {
 	 * a vblank wait completed it might be a bit later. This completion is
 	 * useful to throttle updates and avoid hardware updates getting ahead
 	 * of the buffer cleanup too much.
+	 *
+	 * Drivers should call drm_atomic_helper_commit_cleanup_done() to signal
+	 * completion of this stage.
 	 */
 	struct completion cleanup_done;