lkml.org 
[lkml]   [2015]   [Apr]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/2] workqueue: fix a bug when numa mapping is changed
On 2015/04/02 10:36, Gu Zheng wrote:
> Hi Kame, TJ,
>
> On 04/01/2015 04:30 PM, Kamezawa Hiroyuki wrote:
>
>> On 2015/04/01 12:02, Tejun Heo wrote:
>>> On Wed, Apr 01, 2015 at 11:55:11AM +0900, Kamezawa Hiroyuki wrote:
>>>> Now, hot-added cpus will have the lowest free cpu id.
>>>>
>>>> Because of this, in most of systems which has only cpu-hot-add, cpu-ids are always
>>>> contiguous even after cpu hot add.
>>>> In enterprise, this would be considered as imcompatibility.
>>>>
>>>> determining cpuid <-> lapicid at boot will make cpuids sparse. That may corrupt
>>>> exisiting script or configuration/resource management software.
>>>
>>> Ugh... so, cpu number allocation on hot-add is part of userland
>>> interface that we're locked into?
>>
>> We checked most of RHEL7 packages and didn't find a problem yet.
>> But, for examle, we know some performance test team's test program assumed contiguous
>> cpuids and it failed. It was an easy case because we can ask them to fix the application
>> but I guess there will be some amount of customers that cpuids are contiguous.
>>
>>> Tying hotplug and id allocation
>>> order together usually isn't a good idea. What if the cpu up fails
>>> while running the notifiers? The ID is already allocated and the next
>>> cpu being brought up will be after a hole anyway. Is this even
>>> actually gonna affect userland?
>>>
>>
>> Maybe. It's not fail-safe but....
>>
>> In general, all kernel engineers (and skilled userland engineers) knows that
>> cpuids cannot be always contiguous and cpuids/nodeids should be checked before
>> running programs. I think most of engineers should be aware of that but many
>> users have their own assumption :(
>>
>> Basically, I don't have strong objections, you're right technically.
>>
>> In summary...
>> - users should not assume cpuids are contiguous.
>> - all possible ids should be fixed at boot time.
>> - For uses, some clarification document should be somewhere in Documenatation.
>
> Fine to me.
>
>>
>> So, Gu-san
>> 1) determine all possible ids at boot.
>> 2) clarify cpuid/nodeid can have hole because of 1) in Documenation.
>> 3) It would be good if other guys give us ack.
>
> Also fine.
> But before this going, could you please reconsider determining the ids when firstly
> present (the implementation on this patchset)?
> Though it is not the perfect one in some words, but we can ignore the doubts that
> mentioned above as the cpu/node hotplug is not frequent behaviours, and there seems
> not anything harmful to us if we go this way.
>

Is it so heavy work ? Hmm. My requests are

Implement your patches as
- Please don't change current behavior at boot.
- Remember all possible apicids and give them future cpuids if not assigned.
as step 1.

Please fix dynamic pxm<->node detection in step2.

In future, memory-less node handling in x86 should be revisited.

Thanks,
-Kame









\
 
 \ /
  Last update: 2015-04-02 05:21    [W:0.050 / U:0.276 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site