[PATCH v7 2/4] Documentation, dt, arm64/arm: dt bindings for numa.

[Mark Rutland]

Hi,

On Fri, Dec 11, 2015 at 08:11:07PM +0530, Ganapatrao Kulkarni wrote:
> On Fri, Dec 11, 2015 at 7:23 PM, Mark Rutland <mark.rutland at arm.com> wrote:
> > Hi,
> >
> > On Tue, Nov 17, 2015 at 10:50:41PM +0530, Ganapatrao Kulkarni wrote:
> >> DT bindings for numa mapping of memory, cores and IOs.
> >>
> >> Reviewed-by: Robert Richter <rrichter at cavium.com>
> >> Signed-off-by: Ganapatrao Kulkarni <gkulkarni at caviumnetworks.com>
> >
> > Overall this looks good to me. However, I have a couple of concerns.
> thanks.

[...]

> >> +==============================================================================
> >> +2 - numa-node-id
> >> +==============================================================================
> >> +The device node property numa-node-id describes numa domains within a
> >> +machine. This property can be used in device nodes like cpu, memory, bus and
> >> +devices to map to respective numa nodes.
> >> +
> >> +numa-node-id property is a 32-bit integer which defines numa node id to which
> >> +this device node has numa domain association.
> >
> > I'd prefer if the above two paragraphs were replaced with:
> >
> >         For the purpose of identification, each NUMA node is associated
> >         with a unique token known as a node id. For the purpose of this
> >         binding a node id is a 32-bit integer.
> >
> >         A device node is associated with a NUMA node by the presence of
> >         a numa-node-id property which contains the node id of the
> >         device.
> ok, will do.

[...]

> >> +==============================================================================
> >> +3 - distance-map
> >> +==============================================================================
> >> +
> >> +The device tree node distance-map describes the relative
> >> +distance (memory latency) between all numa nodes.
> >
> > Is this not a combined approximation for latency and bandwidth?
> AFAIK, it is to represent inter-node memory access latency.
> >
> >> +- compatible : Should at least contain "numa,distance-map-v1".
> >
> > Please use "numa-distance-map-v1", as "numa" is not a vendor.
> ok
> >
> >> +- distance-matrix
> >> +  This property defines a matrix to describe the relative distances
> >> +  between all numa nodes.
> >> +  It is represented as a list of node pairs and their relative distance.
> >> +
> >> +  Note:
> >> +     1. Each entry represents distance from first node to second node.
> >> +     2. If both directions between 2 nodes have the same distance, only
> >> +            one entry is required.
> >
> > I still don't understand what direction means in this context. Are there
> > systems (of any architecture) which don't have symmetric distances?
> > Which accesses does this apply differently to?
> >
> > Given that, I think that it might be best to explicitly call out
> > distances as being equal, and leave any directionality for a later
> > revision of the binding when we have some semantics for directionality.
> agreed, given that there is no know system to substantiate dual direction,
> let us not explicit about direction.

Regarding your comment in [1], I was expecting a respin of this series
with the above comments addressed. I will not provide an ack until I've
seen that.

Additional concerns below also apply.

> >> +     2. distance-matrix shold have entries in lexicographical ascending order of nodes.
> >> +     3. There must be only one Device node distance-map and must reside in the root node.
> >> +
> >> +Example:
> >> +     4 nodes connected in mesh/ring topology as below,
> >> +
> >> +             0_______20______1
> >> +             |               |
> >> +             |               |
> >> +           20|               |20
> >> +             |               |
> >> +             |               |
> >> +             |_______________|
> >> +             3       20      2
> >> +
> >> +     if relative distance for each hop is 20,
> >> +     then inter node distance would be for this topology will be,
> >> +           0 -> 1 = 20
> >> +           1 -> 2 = 20
> >> +           2 -> 3 = 20
> >> +           3 -> 0 = 20
> >> +           0 -> 2 = 40
> >> +           1 -> 3 = 40
> >
> > How is this scaled relative to a local access?
> this is based on representing local distance with 10 and
> all inter-node latency being represented as multiple of 10.
> 
> >
> > Do we assume that a local access has value 1, e.g. each hop takes 20x a
> > local access in this example?
> The local distance is represented as 10, this is fixed and same as in ACPI.
> Inter-node distance can be any number greater than 10.
> this information can be added here to make it clear.

This seems rather arbitrary.

Why can we not define the local distance in the DT? I appreciate that
the value is hard-coded for ACPI, but we don't have to copy that
limitation.

I'm not sure if asymmetric local distances matter.

Thanks,
Mark.

[1] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-December/394634.html