Re: [PATCH 2/5] workqueue: update wq_numa_possible_cpumask

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 12/13/2014 01:18 AM, Tejun Heo wrote:
> On Fri, Dec 12, 2014 at 06:19:52PM +0800, Lai Jiangshan wrote:
> ...
>> +static void wq_update_numa_mapping(int cpu)
>> +{
>> +	int node, orig_node = NUMA_NO_NODE, new_node = cpu_to_node(cpu);
>> +
>> +	lockdep_assert_held(&wq_pool_mutex);
>> +
>> +	if (!wq_numa_enabled)
>> +		return;
>> +
>> +	/* the node of onlining CPU is not NUMA_NO_NODE */
>> +	if (WARN_ON(new_node == NUMA_NO_NODE))
>> +		return;
>> +
>> +	/* test whether the NUMA node mapping is changed. */
>> +	if (cpumask_test_cpu(cpu, wq_numa_possible_cpumask[new_node]))
>> +		return;
>> +
>> +	/* find the origin node */
>> +	for_each_node(node) {
>> +		if (cpumask_test_cpu(cpu, wq_numa_possible_cpumask[node])) {
>> +			orig_node = node;
>> +			break;
>> +		}
>> +	}
>> +
>> +	/* there may be multi mappings changed, re-initial. */
>> +	cpumask_clear(wq_numa_possible_cpumask[new_node]);
>> +	if (orig_node != NUMA_NO_NODE)
>> +		cpumask_clear(wq_numa_possible_cpumask[orig_node]);
>> +	for_each_possible_cpu(cpu) {
>> +		node = cpu_to_node(node);
>> +		if (node == new_node)
>> +			cpumask_set_cpu(cpu, wq_numa_possible_cpumask[new_node]);
>> +		else if (orig_node != NUMA_NO_NODE && node == orig_node)
>> +			cpumask_set_cpu(cpu, wq_numa_possible_cpumask[orig_node]);
>> +	}
>> +}
> 
> Let's please move this to NUMA code and properly update it on actual
> mapping changes.
> 

Hi, TJ

I didn't get your means.  What did you mean "NUMA code"?  Which one did you mean?

1) "NUMA code" = system's NUMA memory hotplug code, AKA, keep the numa mapping stable

   I think this is the better idea.  This idea came to my mind immediately at the time
   I received the bug report.  And after some discussions, I was told that it is too HARD
   to keep the numa mapping stable across multiple physical system-board/node online/offline.

   This idea makes the assumption "the numa mapping is stable after system booted" as
   a restriction of the NUMA.  And it will favor all the code outside of the numa code,
   otherwise (we deny the assumption like this patchset) all the code which use
   "cpu_to_node()" and cache the return value will have to be fixed up like this patchset.

   Hi, hotplug-team, any idea to keep the numa mapping stable?

2) "NUMA code" = workqueue's NUMA code
   I think I already did it, the code I added was right below the code of
   wq_update_unbound_numa().  Or I missed something?

Thanks,
Lai