[Intel-gfx] [PATCH] drm/i915/dg1: Disable D3 for gfx's VSP

[Gupta, Anshuman]

> -----Original Message-----
> From: Vivi, Rodrigo <rodrigo.vivi at intel.com>
> Sent: Wednesday, July 20, 2022 7:33 PM
> To: Gupta, Anshuman <anshuman.gupta at intel.com>
> Cc: intel-gfx at lists.freedesktop.org; Summers, Stuart
> <stuart.summers at intel.com>
> Subject: Re: [PATCH] drm/i915/dg1: Disable D3 for gfx's VSP
> 
> On Wed, Jul 20, 2022 at 06:51:37PM +0530, Anshuman Gupta wrote:
> > When DG1 gfx pci endpoint's parent upstream bridge (VSP) enters to D3,
> > any direct memory read from lmem returns with 0xff.
> > DG1 requires i915 driven runtime resume in order to read properly from
> > lmem, DG1 is unable to wakeup from local memory reads.
> > This may breaks using mmap() with lmem object on DG1 as it relies on
> > mercy of i915 driven runtime resume.
> 
> why our disable at root_port not working to block this case?
> If the root power is not at D3Cold our power shouldn't be cut and our memory
> should be there.
This is not d3cold issue, there was no ACPI real_power_state means d3cold was not supported.  
the issue here is reading from memory is not transitioning the d3 -> d0 on DG1.
As far as I understand , reading form mmap memory will not cause i915 kernel wakeup.
It will be a device wakeup (D3-> D0) without any i915 runtime resume, 
If we force dpms_on before accessing mmap region.
It can read the valid data from lmem.
Thanks,
Anshuman Gupta.
> 
> >
> > Let's not break using mmap() for lmem object on DG1.
> > Using pci_d3cold_disable(i915) accordingly to disable D3 for upstream
> > bridge.
> >
> > Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/6331
> > Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > Signed-off-by: Anshuman Gupta <anshuman.gupta at intel.com>
> > ---
> >  drivers/gpu/drm/i915/intel_runtime_pm.c | 8 ++++++++
> >  1 file changed, 8 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/i915/intel_runtime_pm.c
> > b/drivers/gpu/drm/i915/intel_runtime_pm.c
> > index 6ed5786bcd29..671009fd351a 100644
> > --- a/drivers/gpu/drm/i915/intel_runtime_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_runtime_pm.c
> > @@ -559,6 +559,7 @@ void intel_runtime_pm_enable(struct
> intel_runtime_pm *rpm)
> >  	struct drm_i915_private *i915 = container_of(rpm,
> >  						     struct drm_i915_private,
> >  						     runtime_pm);
> > +	struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
> >  	struct device *kdev = rpm->kdev;
> >
> >  	/*
> > @@ -574,6 +575,9 @@ void intel_runtime_pm_enable(struct
> intel_runtime_pm *rpm)
> >  	pm_runtime_set_autosuspend_delay(kdev, 10000); /* 10s */
> >  	pm_runtime_mark_last_busy(kdev);
> >
> > +	if (IS_DG1(i915))
> > +		pci_d3cold_disable(pdev);
> > +
> >  	/*
> >  	 * Take a permanent reference to disable the RPM functionality and
> drop
> >  	 * it only when unloading the driver. Use the low level get/put
> > helpers, @@ -607,12 +611,16 @@ void intel_runtime_pm_disable(struct
> intel_runtime_pm *rpm)
> >  	struct drm_i915_private *i915 = container_of(rpm,
> >  						     struct drm_i915_private,
> >  						     runtime_pm);
> > +	struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
> >  	struct device *kdev = rpm->kdev;
> >
> >  	/* Transfer rpm ownership back to core */
> >  	drm_WARN(&i915->drm, pm_runtime_get_sync(kdev) < 0,
> >  		 "Failed to pass rpm ownership back to core\n");
> >
> > +	if (IS_DG1(i915))
> > +		pci_d3cold_enable(pdev);
> > +
> >  	pm_runtime_dont_use_autosuspend(kdev);
> >
> >  	if (!rpm->available)
> > --
> > 2.26.2
> >