Re: [PATCH 07/17] drm/i915/ddi: Simplify waiting for a port to idle via DDI_BUF_CTL

[Imre Deak <imre.deak@intel.com>]

On Wed, Feb 05, 2025 at 02:35:18PM +0200, Jani Nikula wrote:
> On Wed, 29 Jan 2025, Imre Deak <imre.deak@intel.com> wrote:
> > When waiting for a port to idle, there is no point in distinguishing the
> > platform specific timeouts, instead of just using the maximum timeout.
> 
> Why?
> 
> All of this sounds kind of reasonable, but we'll need a better rationale
> than "there is no point".

The rational is that there is no point in the complexity of specifying
an exact timeout and for that the suitable wait API. The sequence in
particular is not performance critical at all either and due to
scheduling it's not guaranteed anyhow how long the wait will last at the
given timescale. In the usual case where the wait succeeds the actual
time waited does not change with the increased timeout.

> > Simplify things accordingly, describing the Bspec platform specific
> > timeouts in code comments.
> >
> > Signed-off-by: Imre Deak <imre.deak@intel.com>
> > ---
> >  drivers/gpu/drm/i915/display/intel_ddi.c | 78 +++++++++++-------------
> >  1 file changed, 36 insertions(+), 42 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/display/intel_ddi.c b/drivers/gpu/drm/i915/display/intel_ddi.c
> > index 24c56d2aa5f31..d040558b5d029 100644
> > --- a/drivers/gpu/drm/i915/display/intel_ddi.c
> > +++ b/drivers/gpu/drm/i915/display/intel_ddi.c
> > @@ -177,69 +177,63 @@ static void hsw_prepare_hdmi_ddi_buffers(struct intel_encoder *encoder,
> >  		       trans->entries[level].hsw.trans2);
> >  }
> >  
> > -static void mtl_wait_ddi_buf_idle(struct drm_i915_private *i915, enum port port)
> > +static i915_reg_t intel_ddi_buf_status_reg(struct intel_display *display, enum port port)
> >  {
> > -	int ret;
> > +	struct drm_i915_private *i915 = to_i915(display->drm);
> 
> Please don't add new i915 uses, display will work just fine here.
> 
> >  
> > -	/* FIXME: find out why Bspec's 100us timeout is too short */
> > -	ret = wait_for_us((intel_de_read(i915, XELPDP_PORT_BUF_CTL1(i915, port)) &
> > -			   XELPDP_PORT_BUF_PHY_IDLE), 10000);
> > -	if (ret)
> > -		drm_err(&i915->drm, "Timeout waiting for DDI BUF %c to get idle\n",
> > -			port_name(port));
> > +	if (DISPLAY_VER(display) >= 14)
> > +		return XELPDP_PORT_BUF_CTL1(i915, port);
> > +	else
> > +		return DDI_BUF_CTL(port);
> >  }
> >  
> >  void intel_wait_ddi_buf_idle(struct drm_i915_private *dev_priv,
> >  			     enum port port)
> >  {
> > -	if (IS_BROXTON(dev_priv)) {
> > +	struct intel_display *display = &dev_priv->display;
> > +
> > +	/*
> > +	 * Bspec's platform specific timeouts:
> > +	 * MTL+   : 100 us
> > +	 * BXT    : fixed 16 us
> > +	 * HSW-ADL: 8 us
> > +	 *
> > +	 * FIXME: MTL requires 10 ms based on tests, find out why 100 us is too short
> > +	 */
> 
> Seems a bit odd to me to list all the platform specific timeouts, and
> then largely ignore them without explanation!

It's documented so after any future platform requirement changes
(dropping support for a platform, adding a new platform with a new
timeout) can be applied to the timeout used below.

> Also, 10 ms is several orders of magnitude longer than it should need to
> be on all platforms.

I described above why this doesn't matter (in the usual case the wait
duration will not change).

> 
> > +	if (display->platform.broxton) {
> >  		udelay(16);
> >  		return;
> >  	}
> >  
> > -	if (wait_for_us((intel_de_read(dev_priv, DDI_BUF_CTL(port)) &
> > -			 DDI_BUF_IS_IDLE), 8))
> > -		drm_err(&dev_priv->drm, "Timeout waiting for DDI BUF %c to get idle\n",
> > +	static_assert(DDI_BUF_IS_IDLE == XELPDP_PORT_BUF_PHY_IDLE);
> > +	if (intel_de_wait_for_set(display, intel_ddi_buf_status_reg(display, port),
> > +				  DDI_BUF_IS_IDLE, 10))
> > +		drm_err(display->drm, "Timeout waiting for DDI BUF %c to get idle\n",
> >  			port_name(port));
> >  }
> >  
> >  static void intel_wait_ddi_buf_active(struct intel_encoder *encoder)
> >  {
> > -	struct drm_i915_private *dev_priv = to_i915(encoder->base.dev);
> > +	struct intel_display *display = to_intel_display(encoder);
> >  	enum port port = encoder->port;
> > -	int timeout_us;
> > -	int ret;
> >  
> > -	/* Wait > 518 usecs for DDI_BUF_CTL to be non idle */
> > -	if (DISPLAY_VER(dev_priv) < 10) {
> > +	/*
> > +	 * Bspec's platform specific timeouts:
> > +	 * MTL+             : 10000 us
> > +	 * DG2              : 1200 us
> > +	 * TGL-ADL combo PHY: 1000 us
> > +	 * TGL-ADL TypeC PHY: 3000 us
> > +	 * HSW-ICL          : fixed 518 us
> > +	 */
> > +	if (DISPLAY_VER(display) < 10) {
> >  		usleep_range(518, 1000);
> >  		return;
> >  	}
> >  
> > -	if (DISPLAY_VER(dev_priv) >= 14) {
> > -		timeout_us = 10000;
> > -	} else if (IS_DG2(dev_priv)) {
> > -		timeout_us = 1200;
> > -	} else if (DISPLAY_VER(dev_priv) >= 12) {
> > -		if (intel_encoder_is_tc(encoder))
> > -			timeout_us = 3000;
> > -		else
> > -			timeout_us = 1000;
> > -	} else {
> > -		timeout_us = 500;
> > -	}
> > -
> > -	if (DISPLAY_VER(dev_priv) >= 14)
> > -		ret = _wait_for(!(intel_de_read(dev_priv,
> > -						XELPDP_PORT_BUF_CTL1(dev_priv, port)) &
> > -				  XELPDP_PORT_BUF_PHY_IDLE),
> > -				timeout_us, 10, 10);
> > -	else
> > -		ret = _wait_for(!(intel_de_read(dev_priv, DDI_BUF_CTL(port)) & DDI_BUF_IS_IDLE),
> > -				timeout_us, 10, 10);
> > -
> > -	if (ret)
> > -		drm_err(&dev_priv->drm, "Timeout waiting for DDI BUF %c to get active\n",
> > +	static_assert(DDI_BUF_IS_IDLE == XELPDP_PORT_BUF_PHY_IDLE);
> > +	if (intel_de_wait_for_clear(display, intel_ddi_buf_status_reg(display, port),
> > +				    DDI_BUF_IS_IDLE, 10))
> > +		drm_err(display->drm, "Timeout waiting for DDI BUF %c to get active\n",
> >  			port_name(port));
> >  }
> >  
> > @@ -3067,7 +3061,7 @@ static void mtl_disable_ddi_buf(struct intel_encoder *encoder,
> >  	intel_de_rmw(dev_priv, DDI_BUF_CTL(port), DDI_BUF_CTL_ENABLE, 0);
> >  
> >  	/* 3.c Poll for PORT_BUF_CTL Idle Status == 1, timeout after 100us */
> 
> Comment is now stale. (Which is why we should never add comments like
> that.)

Right, I remove all these later in the patchset.

> 
> > -	mtl_wait_ddi_buf_idle(dev_priv, port);
> > +	intel_wait_ddi_buf_idle(dev_priv, port);
> >  
> >  	/* 3.d Disable D2D Link */
> >  	mtl_ddi_disable_d2d_link(encoder);
> 
> -- 
> Jani Nikula, Intel