Re: [PATCH 3/5] workqueue: fixup existing pool->node

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

On 12/13/2014 01:25 AM, Tejun Heo wrote:
> On Fri, Dec 12, 2014 at 06:19:53PM +0800, Lai Jiangshan wrote:
>> Yasuaki Ishimatsu hit a bug when the numa mapping between CPU and node
>> is changed.  And the previous path fixup wq_numa_possible_cpumask.
>> (See more information form the changelog of that patch)
>>
>> After wq_numa_possible_cpumask was updated, the new pool->node will be
>> correct, but the existing pools (and workers) are still running, some of
>> them are running with the wrong pool->node, or even worse, with pool->node
>> which is quitted node, they create_worker() on wrong pool->node.
>> These create_worker() may create workers on wrong node or failed without
>> any progress (when with pool->node which is quitted node).
>>
>> So we need to update the pool->node when the numa mapping is changed.
>>
>> We simply re-calc the pool->node when the numa mapping changed. It reuses
>> the code from get_unbound_pool() for unbound pool.
> 
> I don't get this patch.  If a node is gone, all its cpus would be gone
> and the pool should be discarded.  If a new node comes online with
> different mappings, new sets of pools should serve them instead of
> recycling the old ones.  Wouldn't it make a lot more sense to make
> sure we don't reuse the pools w/ old mappings for new pwqs?
> 


The pwqs of the old node's cpumask do be discarded. But the pools of the old
node's cpumask maybe recycle. For example, a new workqueue's affinity is set to
the old node's cpumask before the pool is dead. Any old pool can long live.


> Thanks.
>