[PATCH v14 2/6] Documentation, dt, numa: dt bindings for NUMA.
gpkulkarni at gmail.com
Mon Mar 7 21:01:33 PST 2016
On Tue, Mar 8, 2016 at 1:17 AM, David Daney <ddaney at caviumnetworks.com> wrote:
> On 03/07/2016 11:22 AM, Robert Richter wrote:
>> On 03.03.16 15:55:35, David Daney wrote:
>>> From: Ganapatrao Kulkarni <gkulkarni at caviumnetworks.com>
>>> Add DT bindings for numa mapping of memory, CPUs and IOs.
>>> Reviewed-by: Robert Richter <rrichter at cavium.com>
>>> Signed-off-by: Ganapatrao Kulkarni <gkulkarni at caviumnetworks.com>
>>> Signed-off-by: David Daney <david.daney at cavium.com>
>>> Documentation/devicetree/bindings/numa.txt | 272
>>> 1 file changed, 272 insertions(+)
>>> create mode 100644 Documentation/devicetree/bindings/numa.txt
>>> diff --git a/Documentation/devicetree/bindings/numa.txt
>>> new file mode 100644
>>> index 0000000..ec5ed7c
>>> --- /dev/null
>>> +++ b/Documentation/devicetree/bindings/numa.txt
>>> +3 - distance-map
>>> +The device tree node distance-map describes the relative
>>> +distance (memory latency) between all numa nodes.
>>> +- compatible : Should at least contain "numa-distance-map-v1".
>>> +- 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.
>>> + The distances are equal in either direction.
>>> + 2. The distance from a node to self (local distance) is
>>> + with value 10 and all internode distance should be represented
>>> + a value greater than 10.
>>> + 3. distance-matrix should have entries in lexicographical
>>> + order of nodes.
>>> + 4. There must be only one device node distance-map which must
>>> reside in the root node.
>> There is no note that this one is optional, but is it right? The
>> default is 10 for local and 20 for remote connections.
> Do we need to explicitly state that it is optional? Many node types are
> optional, and their binding specifications don't really talk about their
> being optional.
> If the node is present then it has the meaning specified.
> If the node is *not* present, then the special meaning described in the
> bindings document does not apply.
> In the case of NUMA, this means that all memory is equally distant (i.e. it
> is *Uniform*), and we are not talking about a *Non* *Uniform* Memory
> Architecture (NUMA) system.
>> If so, then ...
>> static int __init of_numa_parse_distance_map(void)
>> int ret = -EINVAL;
>> struct device_node *np = of_find_node_by_path("/distance-map");
>> if (!np)
>> return ret;
>> must return 0 instead of -EINVAL here.
> No, I don't think doing that would be correct.
> If there is no "distance-map", then of_numa_init() returns the error code.
> This causes the code in arch/arm64/kernel/numa.c to fall back to the
> non-NUMA "dummy_numa" case.
IMO, return 0 will allow 2 node system to have distance-map optional.
by default node distance is set to 10 for local node and for remote node is 20
and this will suffice the need of 2 node system.
by returning EINVAL, we are forcing 2 node system (and even for
systems which has equal remote distances)
to define distance-map.
> By adding your Reviewed-by: Robert Richter <rrichter at cavium.com> tag to
> patch 5/6, where we select between "real" and "dummy_numa", I had assumed
> that you agreed with this approach.
> David Daney
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
More information about the linux-arm-kernel