Re: [PATCH v4 08/39] x86/resctrl: Generate default_ctrl instead of sharing it

[Reinette Chatre]

Hi James,

On 9/13/24 11:07 AM, James Morse wrote:
> Hi Reinette,
>
> On 14/08/2024 05:00, Reinette Chatre wrote:
> > On 8/2/24 10:28 AM, James Morse wrote:
> >
> > > +/**
> > > + * resctrl_get_default_ctrl() - Return the default control value for this
> > > + *                              resource.
> > > + * @r:        The resource whose default control type is queried.
> > > + */
> > > +static inline u32 resctrl_get_default_ctrl(struct rdt_resource *r)
> > > +{
> > > +    switch (r->schema_fmt) {
> > > +    case RESCTRL_SCHEMA_BITMAP:
> > > +        return BIT_MASK(r->cache.cbm_len) - 1;
> > > +    case RESCTRL_SCHEMA_PERCENTAGE:
> > > +        return 100u;
> > > +    case RESCTRL_SCHEMA_MBPS:
> > > +        return r->membw.max_bw;
> > > +    }
> > > +
> > > +    return WARN_ON_ONCE(1);
> > > +}
> > > +
>
> > I am concerned where this is headed. Since RESCTRL_SCHEMA_PERCENTAGE remains
> > in use when resctrl is mounted with mba_MBps the default cannot always
> > be 100u (it should be MBA_MAX_MBPS when software controller is active).
>
> I agree - and we can certainly tidy that up.
> But today when mba_sc is enable the bandwidth_gran and min-bandwidth files both report
> '10' (%?), which isn't particularly meaningful.

Indeed. The complication with mba_sc is that it is a user facing MB/s solution implemented
by percentages in hardware.

> I think these should both report '1'. There will be a minimum bandwidth, buts it not
> something that can be discovered by the mba_sc code.

I understand that it may simplify the interface but please take care that the implementation
relies on these values to determine which percentage to use to accomplish a user
requested Mb/s.

> This was an oversight because the mba_sc mode doesn't update default_ctrl or the format
> strings - it hijacks the parsing elsewhere. The default_ctrl isn't visible to user-space,
> the value used when reading the schema file comes from the mbps_val array, instead of
> ctrl_val.

Right.

> Some of this has been booted over the horizon - I'll add straightening out the mba_sc
> behaviour here to that list.

Ack. I think it will be simpler for this work to focus on splitting existing
implementation. If clearing these up are a priority we can also work on it before
the split, but not as part of this change that I believe was intended to not
include functional changes.

Reinette