Re: [PATCH 2/4] mm: memory_hotplug: check hwpoisoned page firstly in do_migrate_range()

[Kefeng Wang <wangkefeng.wang@huawei.com>]

On 2024/8/2 4:14, David Hildenbrand wrote:
> On 25.07.24 03:16, Kefeng Wang wrote:
>> The commit b15c87263a69 ("hwpoison, memory_hotplug: allow hwpoisoned
>> pages to be offlined") don't handle the hugetlb pages, the dead loop
>> still occur if offline a hwpoison hugetlb, luckly, after the commit
>> e591ef7d96d6 ("mm,hwpoison,hugetlb,memory_hotplug: hotremove memory
>> section with hwpoisoned hugepage"), the HPageMigratable of hugetlb
>> page will be clear, and the hwpoison hugetlb page will be skipped in
>> scan_movable_pages(), so the deed loop issue is fixed.
>>
>> However if the HPageMigratable() check passed(without reference and
>> lock), the hugetlb page may be hwpoisoned, it won't cause issue since
>> the hwpoisoned page will be handled correctly in the next movable
>> pages scan loop, and it will be isolated in do_migrate_range() and
>> but fails to migrated. In order to avoid the unnecessary isolation and
>> unify all hwpoisoned page handling, let's unconditionally check hwpoison
>> firstly, and if it is a hwpoisoned hugetlb page, try to unmap it as
>> the catch all safety net like normal page does.
>>
>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
>> ---
>>   mm/memory_hotplug.c | 27 ++++++++++++++++-----------
>>   1 file changed, 16 insertions(+), 11 deletions(-)
>>
>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>> index 66267c26ca1b..ccaf4c480aed 100644
>> --- a/mm/memory_hotplug.c
>> +++ b/mm/memory_hotplug.c
>> @@ -1788,28 +1788,33 @@ static void do_migrate_range(unsigned long 
>> start_pfn, unsigned long end_pfn)
>>           folio = page_folio(page);
>>           head = &folio->page;
>> -        if (PageHuge(page)) {
>> -            pfn = page_to_pfn(head) + compound_nr(head) - 1;
>> -            isolate_hugetlb(folio, &source);
>> -            continue;
>> -        } else if (PageTransHuge(page))
>> -            pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
>> -
>>           /*
>>            * HWPoison pages have elevated reference counts so the 
>> migration would
>>            * fail on them. It also doesn't make any sense to migrate 
>> them in the
>>            * first place. Still try to unmap such a page in case it is 
>> still mapped
>> -         * (e.g. current hwpoison implementation doesn't unmap KSM 
>> pages but keep
>> -         * the unmap as the catch all safety net).
>> +         * (keep the unmap as the catch all safety net).
>>            */
>> -        if (PageHWPoison(page)) {
>> +        if (unlikely(PageHWPoison(page))) {
>> +            folio = page_folio(page);
>> +
>>               if (WARN_ON(folio_test_lru(folio)))
>>                   folio_isolate_lru(folio);
>> +
>>               if (folio_mapped(folio))
>> -                try_to_unmap(folio, TTU_IGNORE_MLOCK);
>> +                unmap_posioned_folio(folio, TTU_IGNORE_MLOCK);
>> +
>> +            if (folio_test_large(folio))
>> +                pfn = folio_pfn(folio) + folio_nr_pages(folio) - 1;
>>               continue;
>>           }
>> +        if (PageHuge(page)) {
>> +            pfn = page_to_pfn(head) + compound_nr(head) - 1;
>> +            isolate_hugetlb(folio, &source);
>> +            continue;
>> +        } else if (PageTransHuge(page))
>> +            pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
> 
> If we can use a folio in the PageHWPoison() case, can we use one here as 
> well? I know that it's all unreliable when not holding a folio 
> reference, and we have to be a bit careful.

Using a folio here is part of patch4, I want to unify hugetlb/thp(or 
large folio) with "pfn = folio_pfn(folio) + folio_nr_pages(folio) - 1" 
when large folio after get a ref.

> 
> It feels like using folios here would mostly be fine, because things 
> like PageHuge() already use folios internally.
> 
> And using it in the PageHWPoison() but not here looks a bit odd.

We will convert to use folio in the following patch.

> 
> The important part is that we don't segfault if we'd overshoot our target.
>