Re: memory offline infinite loop after soft offline

[David Hildenbrand <david@redhat.com>]

On 18.10.19 13:00, David Hildenbrand wrote:
> On 18.10.19 10:55, Michal Hocko wrote:
>> On Fri 18-10-19 10:38:21, David Hildenbrand wrote:
>>> On 18.10.19 10:24, Michal Hocko wrote:
>>>> On Fri 18-10-19 10:13:36, David Hildenbrand wrote:
>>>> [...]
>>>>> However, if the compound page spans multiple pageblocks
>>>>
>>>> Although hugetlb pages spanning pageblocks are possible this shouldn't
>>>> matter in__test_page_isolated_in_pageblock because this function doesn't
>>>> really operate on pageblocks as the name suggests.  It is simply
>>>> traversing all valid RAM ranges (see walk_system_ram_range).
>>>
>>> As long as the hugepages don't span memory blocks/sections, you are right. I
>>> have no experience with gigantic pages in this regard.
>>
>> They can clearly span sections (1GB is larger than 128MB). Why do you
>> think it matters actually? walk_system_ram_range walks RAM ranges and no
>> allocation should span holes in RAM right?
>>
> 
> Let's explore what I was thinking. If we can agree that any compound
> page is always aligned to its size , then what I tell here is not
> applicable. I know it is true for gigantic pages.
> 
> Some extreme example to clarify
> 
> [ memory block 0 (128MB) ][ memory block 1 (128MB) ]
>                 [ compound page (128MB)  ]
> 
> If you would offline memory block 1, and you detect PG_offline on the

s/PG_offline/PG_hwpoison/ :)

> first page of that memory block (PageHWPoison(compound_head(page))), you
> would jump over the whole memory block (pfn += 1 <<
> compound_order(page)), leaving 64MB of the memory block unchecked.
> 
> Again, if any compound page has the alignment restrictions (PFN of head
> aligned to 1 << compound_order(page)), this is not possible.
> 
> 
> If it is, however, possible, the "clean" thing would be to only jump
> over the remaining part of the compound page, e.g., something like
> 
> pfn += (1 << compound_order(page)) - (page - compound_head(page)));
> 
> 
> 


-- 

Thanks,

David / dhildenb