Re: Linux 5.19 __NR_move_pages failed for hugepage

From: Miaohe Lin
Date: Fri Aug 12 2022 - 02:41:00 EST


On 2022/8/12 11:04, Wang, Haiyue wrote:
>> -----Original Message-----
>> From: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>> Sent: Friday, August 12, 2022 09:59
>> To: Wang, Haiyue <haiyue.wang@xxxxxxxxx>
>> Cc: akpm@xxxxxxxxxxxxxxxxxxxx; Linux-MM <linux-mm@xxxxxxxxx>; linux-kernel <linux-
>> kernel@xxxxxxxxxxxxxxx>; Naoya Horiguchi <naoya.horiguchi@xxxxxxxxx>; David Hildenbrand
>> <david@xxxxxxxxxx>
>> Subject: Re: Linux 5.19 __NR_move_pages failed for hugepage
>>
>> On 2022/8/11 16:01, Wang, Haiyue wrote:
>>> Hi Miaohe,
>>>
>>>
>>
>> Hi Haiyue,
>>
>> Many thanks for your report and debug.
>>
>>>
>>> When I call "*syscall (__NR_move_pages, 0, n_pages, ptr, 0, status, 0)*" to get the huge page node
>>>
>>> information, it is failed with '-2' returned in 'status' array.
>>>
>>>
>>>
>>> After some debug, I found that "*follow_huge_pud*" will return NULL if '*FOLL_GET*' is set.
>>>
>>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e66f17ff71772b209eed39de
>> 35aaa99ba819c93d
>> <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e66f17ff71772b209eed39d
>> e35aaa99ba819c93d>
>>>
>>>
>>>
>>> This will make your patch doesn't work for huge page.
>>>
>>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=4cd614841c06338a087769ee
>> 3cfa96718784d1f5
>> <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=4cd614841c06338a087769e
>> e3cfa96718784d1f5>
>>>
>>
>> Supporting of '*FOLL_GET*' in follow_huge_pud is introduced via the below commit:
>>
>> https://lore.kernel.org/all/20220714042420.1847125-9-
>> naoya.horiguchi@xxxxxxxxx/T/#mb3c83df087fba454b7b4ea32227fb8775ca70081
>>
>> But that's still not perfect yet. For s390 version of follow_huge_pud, FOLL_GET is still not supported.
>> And pgd level
>> hugepage doesn't support FOLL_GET now.
>>
>>>
>>>
>>> Not sure you know this issue or not, just share my debug information.
>>
>> I'm not sure whether it's better to revert my above "problematic" patch first then add it back when
>> all hugetlb pages support FOLL_GET.
>> Or we could just live with it? Any thoughts?
>>
>
> TBH, the issue is more complicated than I think. :-(
>
> Looks like only '[PATCH v7 5/8] mm, hwpoison: set PG_hwpoison for busy hugetlb pages' will be
> backported to 5.19 ? Only this patch has "Fixes:" tag. If so, it will break 5.19.

If you want to mitigate the problem of __NR_move_pages failing for hugepage, "[PATCH v7 2/8] mm/hugetlb:
make pud_huge() and follow_huge_pud() aware of non-present pud entry" could be backported to 5.19.

>
> I just run VPP 'https://fd.io/' to find the error message about huge page allocation
> after I switched from 5.18 to 5.19.

Do you mean the reported problem is found by VPP? Anyway, you can send a patch to fix the problem if you like. :)
I will try fixing it if requested of course (but I'm not sure how to fix it yet).

Thanks,
Miaohe Lin