[Intel-gfx] [PATCH] drm/i915: Clear bogus DMC BIOS/debug power well requests

[Imre Deak]

On Wed, Dec 05, 2018 at 10:20:23PM +0200, Ville Syrjala wrote:
> From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> 
> The DMC firmware is confused and forces on the BIOS and debug
> power well requests for PW1 and MISC IO on some platforms. On
> BXT I measured this to waste about 10mW in the freeze system
> suspend state with the SoC in s0. I didn't get conclusive
> numbers for s0ix on account of the power consumption being
> much more noisy than in s0.
> 
> This is pretty much undoing part of commit 42d9366d41a9
> ("drm/i915/gen9+: Don't remove secondary power well requests")
> where we stopped sanitizing the DMCs bogus request bits.
> 
> Cc: Imre Deak <imre.deak at intel.com>
> Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>

Thanks for the effort for clarifying this.

The justification in 42d9366d41a9 for not removing the PW#1 and MISC_IO driver
request bits is not too solid, I admit. Bspec 4230 has this to say about them
(for SKL, but same in practice for the rest):

"""
4. Disable Power Well 1 (PG1) and Misc IO Power

    Clear PWR_WELL_CTL Power Well 1 Request to 0b. Do not clear Misc IO Power Request.
    Wait for 10us. Do not poll for the power well to disable. Other clients may be keeping it enabled.
"""

With MISC_IO we would clearly do the opposite in this patch wrt. the spec.

With PW#1 it's unclear if the spec means only the driver's request bit or all
of them, but "other clients may be keeping it enabled" suggests to me it means
only the driver's request bit. OTOH removing only the driver's request bit
alone doesn't make sense due to DMC forcing on the BIOS (and debug) request
bits anyway.

I opened a BSpec update request to clarify this, I suggest waiting for an
update there before going ahead with this change.

> ---
>  drivers/gpu/drm/i915/intel_runtime_pm.c | 35 +++++++++++++++++++++++--
>  1 file changed, 33 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_runtime_pm.c b/drivers/gpu/drm/i915/intel_runtime_pm.c
> index 4350a5270423..6e349181dd1c 100644
> --- a/drivers/gpu/drm/i915/intel_runtime_pm.c
> +++ b/drivers/gpu/drm/i915/intel_runtime_pm.c
> @@ -336,10 +336,17 @@ static void hsw_wait_for_power_well_disable(struct drm_i915_private *dev_priv,
>  	 * Bspec doesn't require waiting for PWs to get disabled, but still do
>  	 * this for paranoia. The known cases where a PW will be forced on:
>  	 * - a KVMR request on any power well via the KVMR request register
> -	 * - a DMC request on PW1 and MISC_IO power wells via the BIOS and
> -	 *   DEBUG request registers
> +	 * - a debug request on any power well via the DEBUG request register
>  	 * Skip the wait in case any of the request bits are set and print a
>  	 * diagnostic message.
> +	 *
> +	 * Note that DMC firmware will also force on the PW1 BIOS request
> +	 * on SKL-CNL, MISC_IO BIOS request on SKL-GLK (although MISC_IO
> +	 * does not even exits on BXT/GLK so the bit doesn't stick),
> +	 * and the PW1/MISC_IO debug request on BXT. We simply clear
> +	 * those spurious requests in hsw_power_well_disable() to make
> +	 * sure they don't waste power. Starting from ICL the DMC firmware
> +	 * has been fixed to only force on the PW1 driver request bit.
>  	 */
>  	wait_for((disabled = !(I915_READ(regs->driver) &
>  			       HSW_PWR_WELL_CTL_STATE(pw_idx))) ||
> @@ -347,6 +354,11 @@ static void hsw_wait_for_power_well_disable(struct drm_i915_private *dev_priv,
>  	if (disabled)
>  		return;
>  
> +	WARN(reqs & 3,
> +	     "%s left on (bios:%d driver:%d kvmr:%d debug:%d)\n",
> +	     power_well->desc->name,
> +	     !!(reqs & 1), !!(reqs & 2), !!(reqs & 4), !!(reqs & 8));
> +
>  	DRM_DEBUG_KMS("%s forced on (bios:%d driver:%d kvmr:%d debug:%d)\n",
>  		      power_well->desc->name,
>  		      !!(reqs & 1), !!(reqs & 2), !!(reqs & 4), !!(reqs & 8));
> @@ -409,6 +421,7 @@ static void hsw_power_well_disable(struct drm_i915_private *dev_priv,
>  				   struct i915_power_well *power_well)
>  {
>  	const struct i915_power_well_regs *regs = power_well->desc->hsw.regs;
> +	enum i915_power_well_id id = power_well->desc->id;
>  	int pw_idx = power_well->desc->hsw.idx;
>  	u32 val;
>  
> @@ -417,6 +430,24 @@ static void hsw_power_well_disable(struct drm_i915_private *dev_priv,
>  
>  	val = I915_READ(regs->driver);
>  	I915_WRITE(regs->driver, val & ~HSW_PWR_WELL_CTL_REQ(pw_idx));
> +	/*
> +	 * On SKL-CNL DMC firmware forces on the BIOS request.
> +	 * This wastes a bit of power so clear it.
> +	 */
> +	if (INTEL_GEN(dev_priv) < 11 &&
> +	    (id == SKL_DISP_PW_1 || id == SKL_DISP_PW_MISC_IO)) {
> +		val = I915_READ(regs->bios);
> +		I915_WRITE(regs->bios, val & ~HSW_PWR_WELL_CTL_REQ(pw_idx));
> +	}
> +	/*
> +	 * On BXT DMC firmware forces on the debug request.
> +	 * This wastes a bit of power so clear it.
> +	 */
> +	if (IS_BROXTON(dev_priv) &&
> +	    (id == SKL_DISP_PW_1 || id == SKL_DISP_PW_MISC_IO)) {

No MISC_IO on BXT, so is this just for symmetry with the above if-block?

> +		val = I915_READ(regs->debug);
> +		I915_WRITE(regs->debug, val & ~HSW_PWR_WELL_CTL_REQ(pw_idx));
> +	}
>  	hsw_wait_for_power_well_disable(dev_priv, power_well);
>  }
>  
> -- 
> 2.18.1
>