Re: [PATCH 2/2] drm/i915: Make sure we have enough memory bandwidth on ICL

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Wed, Mar 20, 2019 at 11:46:35PM +0200, Ville Syrjala wrote:
> +	/*
> +	 * Try to muzzle SAGV to prevent it from
> +	 * messing up the memory controller readout.
> +	 */
> +	intel_disable_sagv(dev_priv);
> +
> +	/*
> +	 * Magic sleep to avoid observing very high DDR clock.
> +	 * Not sure what's going on but on a system with DDR4-3200
> +	 * clock of 4800 MT/s is often observed here. A short
> +	 * sleep manages to hide that.. Is that actually
> +	 * the "min latency" SAGV point?. Maybe the SA clocks
> +	 * things way up when there is no memory traffic?
> +	 * But polling the register never seems to show this
> +	 * except during i915 unload/load. Sleeping before the
> +	 * SAGV disable usually returns 2133 MT/s.
> +	 *
> +	 * FIXME what is going on?
> +	 */
> +	msleep(5);

Argh. Looks like this isn't working on the ci machines. We get

<7>[   12.419386] [drm:i915_driver_load [i915]] SAGV 0 DCLK=64 tRP=15 tRDPRE=8 tRAS=35 tRCD=15 tRC=50
<7>[   12.419417] [drm:i915_driver_load [i915]] SAGV 1 DCLK=64 tRP=15 tRDPRE=8 tRAS=35 tRCD=15 tRC=50
<7>[   12.419447] [drm:i915_driver_load [i915]] SAGV 2 DCLK=64 tRP=15 tRDPRE=8 tRAS=35 tRCD=15 tRC=50

Which would indicate 2133 MT/s even though the machines have
3200 MT/s memory (at least according to DMI).

-- 
Ville Syrjälä
Intel
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx