On Mon, 2 Jul 2018 10:02:52 +0200
Daniel Vetter <daniel@xxxxxxxx> wrote:
> On Fri, Jun 29, 2018 at 01:17:17PM +0200, Boris Brezillon wrote:
> > In some cases CRTCs are active but are not able to generating events, at
> > least not at every frame at it's expected to.
> > This is typically the case when the CRTC is feeding a writeback connector
> > that has no job queued. In this situation the CRTC is usually stopped
> > until a new job is queued, and this can lead to timeouts when part of
> > the pipeline is updated but no new jobs are queued to the active
> > writeback connector.
> >
> > In order to solve that, we add a ->no_vblank flag to drm_crtc_state
> > and ask the CRTC drivers to set it to true when they know they're not
> > able to generate VBLANK events. The core drm_atomic_helper_fake_vblank()
> > helper can then be used to fake VBLANKs at commit time.
> >
> > Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxx>
> > ---
> > drivers/gpu/drm/drm_atomic_helper.c | 40 +++++++++++++++++++++++++++++++++++++
> > include/drm/drm_atomic_helper.h | 1 +
> > include/drm/drm_crtc.h | 15 ++++++++++++++
> > 3 files changed, 56 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
> > index 69063bcf2334..ca586993c2a2 100644
> > --- a/drivers/gpu/drm/drm_atomic_helper.c
> > +++ b/drivers/gpu/drm/drm_atomic_helper.c
> > @@ -2051,6 +2051,46 @@ void drm_atomic_helper_wait_for_dependencies(struct drm_atomic_state *old_state)
> > }
> > EXPORT_SYMBOL(drm_atomic_helper_wait_for_dependencies);
> >
> > +/**
> > + * drm_atomic_helper_fake_vblank - fake VBLANK events if needed
> > + * @old_state: atomic state object with old state structures
> > + *
> > + * This function walks all CRTCs and fake VBLANK events on those with
> > + * &drm_crtc_state.no_vblank set to true and &drm_crtc_state.event != NULL.
> > + * The primary use of this function is writeback connectors working in oneshot
> > + * mode and faking VBLANK events. In this case they only fake the VBLANK event
> > + * when a job is queued, and any change to the pipeline that does not touch the
> > + * connector is leading to timeouts when calling
> > + * drm_atomic_helper_wait_for_vblanks() or
> > + * drm_atomic_helper_wait_for_flip_done().
> > + *
> > + * This is part of the atomic helper support for nonblocking commits, see
> > + * drm_atomic_helper_setup_commit() for an overview.
> > + */
> > +void drm_atomic_helper_fake_vblank(struct drm_atomic_state *old_state)
> > +{
> > + struct drm_crtc_state *old_crtc_state, *new_crtc_state;
> > + struct drm_crtc *crtc;
> > + int i;
> > +
> > + for_each_oldnew_crtc_in_state(old_state, crtc, old_crtc_state,
> > + new_crtc_state, i) {
> > + unsigned long flags;
> > +
> > + if (!new_crtc_state->no_vblank && !old_crtc_state->no_vblank)
>
> Uh, this essentially makes it impossible to reset no_vblank.
I don't want ->no_vblank to be reset by the core. It's up to the CRTC
driver to clear/set it when something changes in the pipeline.
> For control
> flow state bits we only check the new state for it (see e.g. the various
> *_changed or plane_bits or whatever).
I tried with !new_crtc_state->no_vblank only, but then it does not
handle the case where the CRTC and connector are being disabled, and I
end up with a timeout.
>
> > + continue;
> > +
> > + spin_lock_irqsave(&old_state->dev->event_lock, flags);
> > + if (new_crtc_state->event) {
> > + drm_crtc_send_vblank_event(crtc,
> > + new_crtc_state->event);
> > + new_crtc_state->event = NULL;
> > + }
> > + spin_unlock_irqrestore(&old_state->dev->event_lock, flags);
> > + }
> > +}
> > +EXPORT_SYMBOL(drm_atomic_helper_fake_vblank);
> > +
> > /**
> > * drm_atomic_helper_commit_hw_done - setup possible nonblocking commit
> > * @old_state: atomic state object with old state structures
> > diff --git a/include/drm/drm_atomic_helper.h b/include/drm/drm_atomic_helper.h
> > index 26aaba58d6ce..99e2a5297c69 100644
> > --- a/include/drm/drm_atomic_helper.h
> > +++ b/include/drm/drm_atomic_helper.h
> > @@ -100,6 +100,7 @@ int __must_check drm_atomic_helper_swap_state(struct drm_atomic_state *state,
> > int drm_atomic_helper_setup_commit(struct drm_atomic_state *state,
> > bool nonblock);
> > void drm_atomic_helper_wait_for_dependencies(struct drm_atomic_state *state);
> > +void drm_atomic_helper_fake_vblank(struct drm_atomic_state *state);
> > void drm_atomic_helper_commit_hw_done(struct drm_atomic_state *state);
> > void drm_atomic_helper_commit_cleanup_done(struct drm_atomic_state *state);
> >
> > diff --git a/include/drm/drm_crtc.h b/include/drm/drm_crtc.h
> > index 23eddbccab10..7435dc66c08a 100644
> > --- a/include/drm/drm_crtc.h
> > +++ b/include/drm/drm_crtc.h
> > @@ -87,6 +87,20 @@ struct drm_plane_helper_funcs;
> > * @zpos_changed: zpos values of planes on this crtc have been updated
> > * @color_mgmt_changed: color management properties have changed (degamma or
> > * gamma LUT or CSC matrix)
> > + * @no_vblank: reflects the ability of a CRTC to send VBLANK events. This state
> > + * usually depends on the pipeline configuration, and the main usuage is
> > + * CRTCs feeding a writeback connector operating in oneshot mode. In this
> > + * case the VBLANK event is only generated when a job is queued to the
> > + * writeback connector, and we want the core to fake VBLANK events when
> > + * this part of the pipeline hasn't changed but others had or when the
> > + * CRTC and connectors are disabled.
>
> s/are disabled/are getting disabled/
>
> You'll never get a request for an event when it's already off.
That's true.
>
> > + * __drm_atomic_helper_crtc_duplicate_state() will the value from the
> > + * current state, the CRTC driver is then responsible for updating this
> > + * field when needed.
>
> Not parsing the above ... probably missing a "... will not reset ..."
Exactly. Will add the missing words.
>
> > + * Note that, even when no_blank is set to true, the CRTC driver can still
> > + * steal the &drm_crtc_state.event object and send the event on its own.
> > + * That's usually what happens when a job is queued to the writeback
> > + * connector.
>
> The last sentence is confusing imo. Just drop it?
Yes, I know, but it's also important to state that the ->no_blank +
event == NULL is a valid combination, and just means that the driver
decided to generate the event (that happens when a new WB job is
queued).
>
> Please use the inline comment style for struct members, and then also
> polish the formatting a bit (e.g. paragraph breaks, which are only
> possible with the inline style).
I considered that, but the other fields were already documented in the
single block above the struct, so I thought keeping things consistent
was better. Should I just add this field doc inline and keep the other
ones where they are, or should I add a patch moving all docs inline?
>
> With the nits addressed:
>
> Reviewed-by: Daniel Vetter <daniel.vetter@xxxxxxxx>
>
> > * @plane_mask: bitmask of (1 << drm_plane_index(plane)) of attached planes
> > * @connector_mask: bitmask of (1 << drm_connector_index(connector)) of attached connectors
> > * @encoder_mask: bitmask of (1 << drm_encoder_index(encoder)) of attached encoders
> > @@ -118,6 +132,7 @@ struct drm_crtc_state {
> > bool connectors_changed : 1;
> > bool zpos_changed : 1;
> > bool color_mgmt_changed : 1;
> > + bool no_vblank : 1;
> >
> > /* attached planes bitmask:
> > * WARNING: transitional helpers do not maintain plane_mask so
> > --
> > 2.14.1
> >
>
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
