Re: [PATCH 5/7] drm/xe/nvlp: Implement Wa_14026539277

[Matt Roper <matthew.d.roper@intel.com>]

On Thu, Mar 05, 2026 at 12:23:52PM -0300, Gustavo Sousa wrote:
> Gustavo Sousa <gustavo.sousa@intel.com> writes:
> 
> > Implement the KMD part of Wa_14026539277, which applies to NVL-P A0.
> > The KMD implementation is just one component of the workaround, which
> > also depends on Pcode to implement its part in order to be complete.
> >
> > Signed-off-by: Gustavo Sousa <gustavo.sousa@intel.com>
> > ---
> >  drivers/gpu/drm/xe/regs/xe_gt_regs.h |  4 ++++
> >  drivers/gpu/drm/xe/xe_gt.c           | 27 +++++++++++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_wa_oob.rules   |  2 ++
> >  3 files changed, 33 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/xe/regs/xe_gt_regs.h b/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> > index 66ddad767ad4..a83cafbe03fd 100644
> > --- a/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> > +++ b/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> > @@ -452,6 +452,10 @@
> >  
> >  #define XEHPC_L3CLOS_MASK(i)			XE_REG_MCR(0xb194 + (i) * 8)
> >  
> > +#define L2COMPUTESIDECTRL			XE_REG_MCR(0xb1c0)
> > +#define   CECTRL				REG_GENMASK(2, 1)
> > +#define   CECTRL_CENODATA_ALWAYS		REG_FIELD_PREP(CECTRL, 0x0)
> > +
> >  #define XE2_GLOBAL_INVAL			XE_REG(0xb404)
> >  
> >  #define XE2LPM_L3SQCREG2			XE_REG_MCR(0xb604)
> > diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
> > index b455af1e6072..3c8692f9b8cf 100644
> > --- a/drivers/gpu/drm/xe/xe_gt.c
> > +++ b/drivers/gpu/drm/xe/xe_gt.c
> > @@ -450,6 +450,25 @@ int xe_gt_record_default_lrcs(struct xe_gt *gt)
> >  	return err;
> >  }
> >  
> > +static void xe_gt_wa_14026539277(struct xe_gt *gt)
> > +{
> > +	u32 val;
> > +
> > +	if (!XE_GT_WA(gt, 14026539277))
> > +		return;
> 
> It appears SRIOV VF wouldn't have access to the register below, right?
> 
> Adding Piotr here, since we talked about it offline recently. Does
> implementing this only on PF side suffice?
> 
> If so, should we return early here if on a VF or should we rather add a
> FUNC(xe_rtp_match_not_sriov_vf) in the rules for Wa_14026539277?
> 
> One thing that I realize with the latter approach is that 14026539277
> would not be listed in the workarounds debugfs file.  Do we really want
> that? I think this question applies to all cases where we use
> FUNC(xe_rtp_match_not_sriov_vf), not only this one.

Right, registers like this only have a single global copy in hardware,
and whatever the PF sets in there applies to everything on the system
(PFs and VFs).  So adding FUNC(xe_rtp_match_not_sriov_vf) sounds correct
to me since this workaround only applies to PF drivers and isn't
applicable to VFs.


Matt

> 
> --
> Gustavo Sousa
> 
> > +
> > +	/*
> > +	 * L2COMPUTESIDECTRL has a specific offset for media and the GSI offset
> > +	 * does not apply.
> > +	 */
> > +	xe_gt_assert(gt, xe_gt_is_main_type(gt));
> > +
> > +	val = xe_gt_mcr_unicast_read_any(gt, L2COMPUTESIDECTRL);
> > +	val &= ~CECTRL;
> > +	val |= CECTRL_CENODATA_ALWAYS;
> > +	xe_gt_mcr_multicast_write(gt, L2COMPUTESIDECTRL, val);
> > +}
> > +
> >  int xe_gt_init_early(struct xe_gt *gt)
> >  {
> >  	int err;
> > @@ -575,6 +594,14 @@ static int gt_init_with_gt_forcewake(struct xe_gt *gt)
> >  	 */
> >  	gt->info.gmdid = xe_mmio_read32(&gt->mmio, GMD_ID);
> >  
> > +	/*
> > +	 * Wa_14026539277 can't be implemented as a regular GT workaround (i.e.
> > +	 * as an entry in gt_was[]) because we would get the hardware already in
> > +	 * a bad state by the time it would be applied.  Hence, we implement it
> > +	 * as an OOB workaround and apply it early to prevent that.
> > +	 */
> > +	xe_gt_wa_14026539277(gt);
> > +
> >  	return 0;
> >  }
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_wa_oob.rules b/drivers/gpu/drm/xe/xe_wa_oob.rules
> > index 80b54b195f20..7f42436df6ab 100644
> > --- a/drivers/gpu/drm/xe/xe_wa_oob.rules
> > +++ b/drivers/gpu/drm/xe/xe_wa_oob.rules
> > @@ -58,3 +58,5 @@
> >  
> >  14025883347	MEDIA_VERSION_RANGE(1301, 3503)
> >  		GRAPHICS_VERSION_RANGE(2004, 3005)
> > +
> > +14026539277	PLATFORM(NOVALAKE_P), PLATFORM_STEP(A0, B0), GRAPHICS_VERSION(3510)
> >
> > -- 
> > 2.52.0

-- 
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation