Re: [PATCH] EDAC/{i10nm,skx,skx_common}: Support multiple clumps

[Kyle Meyer]

On Thu, Dec 05, 2024 at 03:57:55PM -0800, Luck, Tony wrote:
> > +int skx_get_src_id(struct skx_dev *d, int off, u8 *id)
> > +{
> > +#ifdef CONFIG_NUMA
> > +	return skx_get_pkg_id(d, id);
> > +#else
> > +	u32 reg;
> > +
> > +	if (pci_read_config_dword(d->util_all, off, &reg)) {
> > +		skx_printk(KERN_ERR, "Failed to read src id\n");
> > +		return -ENODEV;
> > +	}
> > +
> > +	*id = GET_BITFIELD(reg, 12, 14);
> > +	return 0;
> > +#endif
> 
> Doh ... I alwasy forget about IS_ENABLED(). This can be written:
> 
> 
> int skx_get_src_id(struct skx_dev *d, int off, u8 *id)
> {
> 	u32 reg;
> 
> 	if (IS_ENABLED(CONFIG_NUMA))
> 		return skx_get_pkg_id(d, id);
> 
> 	if (pci_read_config_dword(d->util_all, off, &reg)) {
> 		skx_printk(KERN_ERR, "Failed to read src id\n");
> 		return -ENODEV;
> 	}
> 
> 	*id = GET_BITFIELD(reg, 12, 14);
> 	return 0;
> }

Looks good.

> 1) Does this work? I tried on a non-clumpy system that is NUMA.

Yes, I just tested this on a Sapphire Rapids system with multiple UPI domains.

> 2) Is it better (assuming #fidef factored off into a .h file)?

IMO, yes, but there's one subtle difference. EDAC will not load on systems
that have a single UPI domain when CONFIG_NUMA is enabled but numa=off, because
pcibus_to_node() in skx_get_pkg_id() will return NUMA_NO_NODE (-1). Is that a
case that we need to worry about?

Thanks,
Kyle Meyer