Paul Kocialkowski <paul.kocialkowski@bootlin.com> writes:

> Hi,
>
> On Wed, 2019-01-23 at 10:34 -0800, Eric Anholt wrote:
>> Paul Kocialkowski <paul.kocialkowski@bootlin.com> writes:
>> 
>> > During an atomic commit, the HVS is configured with a display list
>> > for the channel matching the associated CRTC. The Pixel Valve (CRTC)
>> > and encoder are also configured for the new setup at that time.
>> > While the Pixel Valve and encoder are reconfigured synchronously, the
>> > HVS is only reconfigured after the display list address (DISPLIST) has
>> > been updated to the current display list address (DISPLACTX), which is
>> > the responsibility of the hardware.
>> > 
>> > The time frame during which the HVS is still running on its previous
>> > configuration but the CRTC and encoder have been reconfigured already
>> > can lead to a number of synchronization issues. They will eventually
>> > cause errors reported on the FIFOs, such as underruns.
>> > 
>> > With underrun detection enabled (from Boris Brezillon's series), this
>> > leads to unreliable underrun detection with random false positives.
>> > 
>> > To ensure a coherent state, wait for each enabled channel of the HVS
>> > to synchronize its current display list address. This fixes the issue
>> > of random underrun reporting on commits.
>> > 
>> > Signed-off-by: Paul Kocialkowski <paul.kocialkowski@bootlin.com>
>> > ---
>> >  drivers/gpu/drm/vc4/vc4_drv.h  |  1 +
>> >  drivers/gpu/drm/vc4/vc4_hvs.c  | 17 +++++++++++++++++
>> >  drivers/gpu/drm/vc4/vc4_kms.c  |  2 ++
>> >  drivers/gpu/drm/vc4/vc4_regs.h |  2 ++
>> >  4 files changed, 22 insertions(+)
>> > 
>> > diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
>> > index c24b078f0593..955f157f5ad0 100644
>> > --- a/drivers/gpu/drm/vc4/vc4_drv.h
>> > +++ b/drivers/gpu/drm/vc4/vc4_drv.h
>> > @@ -772,6 +772,7 @@ void vc4_irq_reset(struct drm_device *dev);
>> >  extern struct platform_driver vc4_hvs_driver;
>> >  void vc4_hvs_dump_state(struct drm_device *dev);
>> >  int vc4_hvs_debugfs_regs(struct seq_file *m, void *unused);
>> > +void vc4_hvs_sync_dlist(struct drm_device *dev);
>> >  
>> >  /* vc4_kms.c */
>> >  int vc4_kms_load(struct drm_device *dev);
>> > diff --git a/drivers/gpu/drm/vc4/vc4_hvs.c b/drivers/gpu/drm/vc4/vc4_hvs.c
>> > index 5d8c749c9749..1ba60b8e0c2d 100644
>> > --- a/drivers/gpu/drm/vc4/vc4_hvs.c
>> > +++ b/drivers/gpu/drm/vc4/vc4_hvs.c
>> > @@ -166,6 +166,23 @@ static int vc4_hvs_upload_linear_kernel(struct vc4_hvs *hvs,
>> >  	return 0;
>> >  }
>> >  
>> > +void vc4_hvs_sync_dlist(struct drm_device *dev)
>> > +{
>> > +	struct vc4_dev *vc4 = to_vc4_dev(dev);
>> > +	unsigned int i;
>> > +	int ret;
>> > +
>> > +	for (i = 0; i < SCALER_CHANNELS_COUNT; i++) {
>> > +		if (!(HVS_READ(SCALER_DISPCTRLX(i)) & SCALER_DISPCTRLX_ENABLE))
>> > +			continue;
>> > +
>> > +		ret = wait_for(HVS_READ(SCALER_DISPLACTX(i)) ==
>> > +			       HVS_READ(SCALER_DISPLISTX(i)), 1000);
>> > +		WARN(ret, "Timeout waiting for channel %d display list sync\n",
>> > +		     i);
>> > +	}
>> > +}
>> > +
>> >  static int vc4_hvs_bind(struct device *dev, struct device *master, void *data)
>> >  {
>> >  	struct platform_device *pdev = to_platform_device(dev);
>> > diff --git a/drivers/gpu/drm/vc4/vc4_kms.c b/drivers/gpu/drm/vc4/vc4_kms.c
>> > index 0490edb192a1..2d66a2b57a91 100644
>> > --- a/drivers/gpu/drm/vc4/vc4_kms.c
>> > +++ b/drivers/gpu/drm/vc4/vc4_kms.c
>> > @@ -155,6 +155,8 @@ vc4_atomic_complete_commit(struct drm_atomic_state *state)
>> >  
>> >  	drm_atomic_helper_commit_hw_done(state);
>> >  
>> > +	vc4_hvs_sync_dlist(dev);
>> > +
>> >  	drm_atomic_helper_wait_for_flip_done(dev, state);
>> 
>> wait_for_flip_done should already be waiting for DISPLACTX to match our
>> crtc's display list, though, right (see vc4_crtc_handle_page_flip())?
>> This seems like a no-op to me.
>
> Right, at this stage it's pretty much a no-op. It does make a
> difference when bringing-in vc4_hvs_unmask_underrun, because this
> vc4_hvs_sync_dlist call comes before it.
>
> When the display lists are not yet synchronized (and differ), an
> underrun occurs so it will be reported although it's not related to the
> new display list configuration. Waiting for the display lists to sync
> before unmasking the interrupt prevents that.
>
> I think waiting for the page flip to unmask the underrun interrupt
> would be too late because the HVS should have finished processing the
> new dlist by then (but maybe I'm mistaken about that), so we would miss
> the underrun indication.

For load tracking I'm mainly concerned about making sure that we don't
have long-standing underruns -- if the first frame after a modeset is a
little wonky, you probably won't notice it.

Getting a clean modeset sequence would be really nice to have,
just... don't block on that.