Re: [PATCH] workqueue: update numa affinity when node hotplug

[Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On 2015/03/03 1:28, Tejun Heo wrote:
> Hello,
>
> On Mon, Mar 02, 2015 at 05:41:05PM +0900, Kamezawa Hiroyuki wrote:
>> Let me start from explaining current behavior.
>>
>> - cpu-id is determined when a new processor(lapicid/x2apicid) is founded.
>>    cpu-id<->nodeid relationship is _not_ recorded.
>
> Is this something from the firmware side or is it just that we aren't
> maintaining the association right now?
>

I think it's not just maintained.

>> - node-id is determined when a new pxm(firmware info) is founded.
>>    pxm<->nodeid relationship is recorded.
>>
>> By this, there are 2 cases of cpu<->nodeid change.
>>
>> Case A) In x86, cpus on memory-less nodes are all tied to existing nodes(round robin).
>> At memory-hotadd happens and a new node comes, cpus are moved to a newly added node
>> based on pxm.
>
> Ah, okay, so the firmware doesn't provide proximity information at all
> for memory-less nodes so we end up putting all of them somewhere
> random and when memory is added to one of the memory-less nodes, the
> mapping information changes?
>

With memory-less node, proximity domain for processors are given but ignored.
When memory(node) hotplug happens, the information revisited and cpuid<->nodeid
relationship is updated.

> Am I understanding it correctly?  If so, it's super weird tho.  Why
> wouldn't there be proximity information for a memless node?  Not
> having memory doesn't mean it's at the same distance from all existing
> nodes.
>
Firmware gives pxm for memory-less node but it's ignored.
I'm not sure why the current implemetaion is.

>> Case B) Adding a node after removing another node, if pxm of them were different from
>> each other, cpu<->node relatiionship changes.
>
> I don't get this either.  Does proximity relationship actually change?
> Or is it that we're assigning different IDs to the same thing?Isn't
> proximity pretty much hardwired to how the system is architected to
> begin with?
>

relationship between proximity domain and lapic id doesn't change.
relationship between lapic-id and cpu-id changes.

pxm <-> memory address  : no change
pxm <-> lapicid         : no change
pxm <-> node id         : no change
lapicid <-> cpu id      : change.

>> I personally thinks proper fix is building persistent cpu-id <-> lapicid relationship as
>> pxm does rather than creating band-aid.
>
> Oh if this is possible, I agree that's the right direction too.
>

Implementation is a bit complicated now :(.

Thanks,
-Kame

> Thanks.
>