Re: [PATCHv4 10/13] node: Add memory caching attributes

From: Jonathan Cameron
Date: Tue Feb 12 2019 - 03:49:26 EST

On Mon, 11 Feb 2019 08:23:04 -0700
Keith Busch <keith.busch@xxxxxxxxx> wrote:

> On Sun, Feb 10, 2019 at 09:19:58AM -0800, Jonathan Cameron wrote:
> > On Sat, 9 Feb 2019 09:20:53 +0100
> > Brice Goglin <Brice.Goglin@xxxxxxxx> wrote:
> >
> > > Hello Keith
> > >
> > > Could we ever have a single side cache in front of two NUMA nodes ? I
> > > don't see a way to find that out in the current implementation. Would we
> > > have an "id" and/or "nodemap" bitmask in the sidecache structure ?
> >
> > This is certainly a possible thing for hardware to do.
> >
> > ACPI IIRC doesn't provide any means of representing that - your best
> > option is to represent it as two different entries, one for each of the
> > memory nodes. Interesting question of whether you would then claim
> > they were half as big each, or the full size. Of course, there are
> > other possible ways to get this info beyond HMAT, so perhaps the interface
> > should allow it to be exposed if available?
> HMAT doesn't do this, but I want this interface abstracted enough from
> HMAT to express whatever is necessary.
> The CPU cache is the closest existing exported attributes to this,
> and they provide "shared_cpu_list". To that end, I can export a
> "shared_node_list", though previous reviews strongly disliked multi-value
> sysfs entries. :(
> Would shared-node symlinks capture the need, and more acceptable?

My inclination is that it's better to follow an existing pattern than
invent a new one that breaks people's expectations.

However, don't feel that strongly about it as long as the interface
is functional and intuitive.

> > Also, don't know if it's just me, but calling these sidecaches is
> > downright confusing. In ACPI at least they are always
> > specifically referred to as Memory Side Caches.
> > I'd argue there should even by a hyphen Memory-Side Caches, the point
> > being that that they are on the memory side of the interconnected
> > rather than the processor side. Of course an implementation
> > choice might be to put them off to the side (as implied by sidecaches)
> > in some sense, but it's not the only one.
> >
> > </terminology rant> :)
> Now that you mention it, I agree "side" is ambiguous. Maybe call it
> "numa_cache" or "node_cache"?
I'm not sure any of the options work well. My inclination would be to
use the full name and keep the somewhat redundant memory there.

The other two feel like they could just as easily be coherent caches
at accelerators for example...


The fun of naming ;)