On 03/04/2015 04:53 PM, Kamezawa Hiroyuki wrote:
> On 2015/03/04 17:03, Xishi Qiu wrote:
>> On 2015/3/4 11:56, Gu Zheng wrote:
>>
>>> Hi Xishi,
>>> On 03/04/2015 10:52 AM, Xishi Qiu wrote:
>>>
>>>> On 2015/3/4 10:22, Xishi Qiu wrote:
>>>>
>>>>> On 2015/3/3 18:20, Gu Zheng wrote:
>>>>>
>>>>>> Hi Xishi,
>>>>>> On 03/03/2015 11:30 AM, Xishi Qiu wrote:
>>>>>>
>>>>>>> When hot-remove a numa node, we will clear pgdat,
>>>>>>> but is memset 0 safe in try_offline_node()?
>>>>>>
>>>>>> It is not safe here. In fact, this is a temporary solution here.
>>>>>> As you know, pgdat is accessed lock-less now, so protection
>>>>>> mechanism (RCU？) is needed to make it completely safe here,
>>>>>> but it seems a bit over-kill.
>>>>>>
>>>>
>>>> Hi Gu,
>>>>
>>>> Can we just remove "memset(pgdat, 0, sizeof(*pgdat));" ?
>>>> I find this will be fine in the stress test except the warning
>>>> when hot-add memory.
>>>
>>> As you see, it will trigger the warning in free_area_init_node().
>>> Could you try the following patch? It will reset the pgdat before reuse it.
>>>
>>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>>> index 1778628..0717649 100644
>>> --- a/mm/memory_hotplug.c
>>> +++ b/mm/memory_hotplug.c
>>> @@ -1092,6 +1092,9 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
>>> return NULL;
>>>
>>> arch_refresh_nodedata(nid, pgdat);
>>> + } else {
>>> + /* Reset the pgdat to reuse */
>>> + memset(pgdat, 0, sizeof(*pgdat));
>>> }
>>
>> Hi Gu,
>>
>> If schedule last a long time, next_zone may be still access the pgdat here,
>> so it is not safe enough, right?
Hi Xishi,
IMO, the scheduled time is rather short if compares with the time gap
between hot remove and hot re-add a node, so we can say it is safe here.
>>
>
> How about just reseting pgdat->nr_zones and pgdat->classzone_idx to be 0 rather than
> memset() ?
>
> It seems breaking pointer information in pgdat is not a choice.
> Just proper "values" should be reset.
Anyway, sounds reasonable.
Best regards,
Gu
>
> Thanks,
> -Kame
>
>
>
> .
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href