Re: [PATCH v1 3/3] x86/MCE/AMD: Handle reassigned bit definitions for CS SMCA

[Yazen Ghannam <yazen.ghannam@amd.com>]

On Tue, Jan 17, 2023 at 10:23:09AM +0100, Ingo Molnar wrote:
> 
> * Avadhut Naik <avadnaik@amd.com> wrote:
> 
> > @@ -178,6 +178,8 @@ static const struct smca_hwid smca_hwid_mcatypes[] = {
> >  	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0)	},
> >  	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1)	},
> >  	{ SMCA_CS_V2,	 HWID_MCATYPE(0x2E, 0x2)	},
> > +	/* Software defined SMCA bank type to handle erratum 1384*/
> > +	{ SMCA_CS_V2_QUIRK, HWID_MCATYPE(0x0, 0x1)  },
> >  
> >  	/* Unified Memory Controller MCA type */
> >  	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0)	},
> > @@ -259,6 +261,17 @@ static inline void fixup_hwid(unsigned int *hwid_mcatype)
> >  
> >  	if (c->x86 == 0x19) {
> >  		switch (c->x86_model) {
> > +		/*
> > +		 * Per Genoa's revision guide, erratum 1384, some SMCA Extended
> > +		 * Error Codes and SMCA Control bits are incorrect for SMCA CS
> > +		 * bank type.
> > +		 */
> > +		case 0x10 ... 0x1F:
> > +		case 0x60 ... 0x7B:
> > +		case 0xA0 ... 0xAF:
> > +			if (*hwid_mcatype == HWID_MCATYPE(0x2E, 0x2))
> > +				*hwid_mcatype = HWID_MCATYPE(0x0, 0x1);
> 
> Why are we open-coding these types?
> 
> Why not use smca_hwid_mcatypes[SMCA_CS_V2], etc.?
> 
> > +			if (*hwid_mcatype == HWID_MCATYPE(0x2E, 0x2))
> > +				*hwid_mcatype = HWID_MCATYPE(0x0, 0x1);
> 
> Ditto.
>

This code runs before matching on a struct with the enums. It seems simplest
to fixup the hardware-provided value before setting things up rather than
changing things later.

Current code flow:
1) Read HWID/McaType values from hardware.
2) Loop through known types and match on the HWID_MCATYPE() tuple.
3) Continue setup based on matched type and its Linux enum.

New code flow:
1) Read HWID/McaType values from hardware.
--> Fixup HWID/McaType values due to any known hardware quirks.
2) Loop through known types and match on the HWID_MCATYPE() tuple.
3) Continue setup based on matched type and its Linux enum.

What do you think?

Also, a further update (maybe you're alluding to?) is get rid of the struct
smca_hwid and just define an enum with "NAME = HWID_MCATYPE(XXX, YYY)".

The struct smca_hwid had another field that was removed, so it seems
unnecessary at the moment.

Thanks,
Yazen