Re: [PATCH 0/2] workqueue: fix a bug when numa mapping is changed

[Gu Zheng <guz.fnst@cn.fujitsu.com>]

Hi Kame-san,

On 03/26/2015 11:12 AM, Kamezawa Hiroyuki wrote:

> On 2015/03/26 11:17, Gu Zheng wrote:
>> Yasuaki Ishimatsu found that with node online/offline, cpu<->node
>> relationship is established. Because workqueue uses a info which was
>> established at boot time, but it may be changed by node hotpluging.
>>
>> Once pool->node points to a stale node, following allocation failure
>> happens.
>>    ==
>>       SLUB: Unable to allocate memory on node 2 (gfp=0x80d0)
>>        cache: kmalloc-192, object size: 192, buffer size: 192, default
>> order:
>>      1, min order: 0
>>        node 0: slabs: 6172, objs: 259224, free: 245741
>>        node 1: slabs: 3261, objs: 136962, free: 127656
>>    ==
>>
>> As the apicid <--> node relationship is persistent, so the root cause is the
> 　　　　 ^^^^^^^
>            pxm.
> 
>> cpu-id <-> lapicid mapping is not persistent (because the currently implementation
>> always choose the first free cpu id for the new added cpu), so if we can build
>> persistent cpu-id <-> lapicid relationship, this problem will be fixed.
>>
>> Please refer to https://lkml.org/lkml/2015/2/27/145 for the previous discussion.
>>
>> Gu Zheng (2):
>>    x86/cpu hotplug: make lapicid <-> cpuid mapping persistent
>>    workqueue: update per cpu workqueue's numa affinity when cpu
>>      preparing online
> 
> why patch(2/2) required ?

wq generates the numa affinity (pool->node) for all the possible cpu's
per cpu workqueue at init stage, that means the affinity of currently un-present
ones' may be incorrect, so we need to update the pool->node for the new added cpu
to the correct node when preparing online, otherwise it will try to create worker
on invalid node if node hotplug occurred.

Regards,
Gu

> 
> Thanks,
> -Kame
> 
>>
>>   arch/x86/kernel/apic/apic.c |   31 ++++++++++++++++++++++++++++++-
>>   kernel/workqueue.c          |    1 +
>>   2 files changed, 31 insertions(+), 1 deletions(-)
>>
> 
> 
> .
>