Re: [RFC PATCH v1 0/4] mm, hwpoison: improve handling workload related to hugetlb and memory_hotplug

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/5/12 0:22, David Hildenbrand wrote:
> On 11.05.22 18:10, HORIGUCHI NAOYA(堀口 直也) wrote:
>> On Wed, May 11, 2022 at 05:11:17PM +0200, David Hildenbrand wrote:
>>> On 09.05.22 12:53, Miaohe Lin wrote:
>>>> On 2022/5/9 17:58, Oscar Salvador wrote:
>>>>> On Mon, May 09, 2022 at 05:04:54PM +0800, Miaohe Lin wrote:
>>>>>>>> So that leaves us with either
>>>>>>>>
>>>>>>>> 1) Fail offlining -> no need to care about reonlining
>>>>>>
>>>>>> Maybe fail offlining will be a better alternative as we can get rid of many races
>>>>>> between memory failure and memory offline? But no strong opinion. :)
>>>>>
>>>>> If taking care of those races is not an herculean effort, I'd go with
>>>>> allowing offlining + disallow re-onlining.
>>>>> Mainly because memory RAS stuff.
>>>>
>>>> This dose make sense to me. Thanks. We can try to solve those races if
>>>> offlining + disallow re-onlining is applied. :)
>>>>
>>>>>
>>>>> Now, to the re-onlining thing, we'll have to come up with a way to check
>>>>> whether a section contains hwpoisoned pages, so we do not have to go
>>>>> and check every single page, as that will be really suboptimal.
>>>>
>>>> Yes, we need a stable and cheap way to do that.
>>>
>>> My simplistic approach would be a simple flag/indicator in the memory block devices
>>> that indicates that any page in the memory block was hwpoisoned. It's easy to
>>> check that during memory onlining and fail it.
>>>
>>> diff --git a/drivers/base/memory.c b/drivers/base/memory.c
>>> index 084d67fd55cc..3d0ef812e901 100644
>>> --- a/drivers/base/memory.c
>>> +++ b/drivers/base/memory.c
>>> @@ -183,6 +183,9 @@ static int memory_block_online(struct memory_block *mem)
>>>         struct zone *zone;
>>>         int ret;
>>>  
>>> +       if (mem->hwpoisoned)
>>> +               return -EHWPOISON;
>>> +
>>>         zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
>>>                                   start_pfn, nr_pages);
>>>  
>>
>> Thanks for the idea, a simple flag could work if we don't have to consider
>> unpoison.  If we need consider unpoison, we need remember the last hwpoison
>> page in the memory block, so mem->hwpoisoned should be the counter of
>> hwpoison pages.
> 
> Right, but unpoisoning+memory offlining+memory onlining is a yet more
> extreme use case we don't have to bother about I think.
> 
>>
>>>
>>>
>>> Once the problematic DIMM would actually get unplugged, the memory block devices
>>> would get removed as well. So when hotplugging a new DIMM in the same
>>> location, we could online that memory again.
>>
>> What about PG_hwpoison flags?  struct pages are also freed and reallocated
>> in the actual DIMM replacement?
> 
> Once memory is offline, the memmap is stale and is no longer
> trustworthy. It gets reinitialize during memory onlining -- so any
> previous PG_hwpoison is overridden at least there. In some setups, we
> even poison the whole memmap via page_init_poison() during memory offlining.
> 

I tend to agree with David. The memmap is unreliable after memory is offline. So preventing memory
re-online until a new DIMM replacement is a good idea.

Thanks!

> Apart from that, we should be freeing the memmap in all relevant cases
> when removing memory. I remember there are a couple of corner cases, but
> we don't really have to care about that.
>