Re: [PATCH v3 03/38] x86/resctrl: Add a schema format enum and use this for fflags

[Reinette Chatre]

Hi James,

On 7/1/24 11:17 AM, James Morse wrote:
> Hi Reinette,
>
> On 28/06/2024 17:43, Reinette Chatre wrote:
> > On 6/14/24 7:59 AM, James Morse wrote:
> > > resctrl has three types of control, these emerge from the way the
> > > architecture initialises a number of properties in struct rdt_resource.
> > >
> > > A group of these properties need to be set the same on all architectures,
> > > it would be better to specify the format the schema entry should use, and
> > > allow resctrl to generate all the other properties it needs. This avoids
> > > architectures having divergant behaviour here.
> >
> > divergant -> divergent ?
> >
> > > Add a schema format enum, and as a first use, replace the fflags member
> > > of struct rdt_resource.
> > >
> > > The MBA schema has a different format between AMD and Intel systems.
> > > The schema_fmt property is changed by __rdt_get_mem_config_amd() to
> > > enable the MBPS format.
>
> > > diff --git a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> > > b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> > > index e3edc41882dc..b12307d465bc 100644
> > > --- a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> > > +++ b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> > > @@ -2162,6 +2162,19 @@ static int rdtgroup_mkdir_info_resdir(void *priv, char *name,
> > >        return ret;
> > >    }
> > >    +static u32 fflags_from_resource(struct rdt_resource *r)
> > > +{
> > > +    switch (r->schema_fmt) {
> > > +    case RESCTRL_SCHEMA_BITMAP:
> > > +        return RFTYPE_RES_CACHE;
> > > +    case RESCTRL_SCHEMA_PERCENTAGE:
> > > +    case RESCTRL_SCHEMA_MBPS:
> > > +        return RFTYPE_RES_MB;
> > > +    }
> > > +
> > > +    return WARN_ON_ONCE(1);
> > > +}
> > > +
> >
> > The fflags returned specifies which files will be associated with the resource
> > in the "info" directory. Basing this on a property of the schema does not look
> > right to me. I understand that many of the info files relate to, for example,
> > information related to the bitmap used by the cache,
>
> Do we agree that some of them are?
>
> One reason for doing this is it decouples the parsing and management of bitmaps from "this
> is the L3 cache", which will make it much easier to support bitmaps on some other kind of
> resource.

The way I see it is that it changes the meaning of the RFTYPE_RES_CACHE flag from "this is a
file related to the cache resource" to "this is a file containing a bitmap property".
It prevents us from easily adding a file related to the cache resource, which
the info directory is intended to contain.

> Ultimately I'd like to expose these to user-space, so that user-space can work out how to
> configure resources it doesn't recognise. Today '100' could be a percentage, a bitmap, or
> a value in MB/s. Today some knowledge of the control type is needed to work this out.
>
>
> > but that is not the same for
> > info files related to the MBA resource (all info files related to MBA resource
> > are not about the schema property format).
>
> Hmmm, because the files min_bandwidth and bandwidth_gran both have bandwidth in their name?
>
> I agree 'delay_linear' and 'thread_throttle_mode' are a bit strange.

Right. This is not a clean association.

> > I do not think the type of values of a schema should dictate which files
> > appear in the info directory.
>
> Longer term I think this will be a problem. We probably only have 3 types of control:
> percentage, bitmap and MB/s... but if each resource on each architecture adds files here
> the list will quickly grow. User-space won't be able to work out how to configure a
> resource type it hadn't seen before.

That is fair. This makes the type of control a property of the resource as is done in this
series. Perhaps this can be exposed to user space via the info directory?

Possibly the files related to control can have new flags that that reflect the control type
instead of the resource. For example, "bit_usage" currently has
"RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE" and that could be (for lack of better
term) "RFTYPE_CTRL_INFO | RFTYPE_CTRL_BITMAP" to disconnect the control type from the
resource. Doing so may then map nicely to the fflags_from_resource() in this patch that
connects the schema format to the _control_ type flag. As we have found there is not
a clear mapping between the control type and the resource type so I expect RFTYPE_RES_CACHE
and RFTYPE_RES_MB to remain and be associated with files that contain information
specific to that resource. This enables future additions of files containing cache specific
(non-bitmap) properties to still be added (with RFTYPE_RES_CACHE flag) without impacting
everything that uses a bitmap.

What do you think?

> This may not be the time - but I think eventually resctrl shouldn't have to care about
> what resources the architecture is presenting.
> For these files, we may need to duplicate 'min_bandwidth' as 'min_percentage'. MBA would
> have both, but any new controls using percentage wouldn't expose them.
>
>
> > Doesn't MPAM support percentage for cache resources
> > and bitmaps for memory resources?
>
> It can have fixed-point-fractions and bitmaps for both caches and memory. Unfortunately
> everything in MPAM is optional - the driver converts whatever it finds for memory
> bandwidth to a percentage as that is what resctrl and user-space expect.
> I can't do the same for cache controls as bitmaps implicitly isolate portions, something
> that can't be done with the fractional control. So far everyone has built the bitmaps
> because its the easiest implementation - but I have had requests to support the cache
> fixed-point-fraction. Doing it as a percentage is least invasive to resctrl...
>
>
> > Can the fflags rather depend on the resource type itself, by using the rid?
>
> Sure.

Reinette