Re: [RFC PATCH 0/5] mm, hugetlb: allocation API and migration improvements

From: Mike Kravetz
Date: Thu Dec 21 2017 - 18:35:43 EST


On 12/20/2017 11:28 PM, Michal Hocko wrote:
> On Wed 20-12-17 14:43:03, Mike Kravetz wrote:
>> On 12/20/2017 01:53 AM, Michal Hocko wrote:
>>> On Wed 20-12-17 05:33:36, Naoya Horiguchi wrote:
>>>> I have one comment on the code path from mbind(2).
>>>> The callback passed to migrate_pages() in do_mbind() (i.e. new_page())
>>>> calls alloc_huge_page_noerr() which currently doesn't call SetPageHugeTemporary(),
>>>> so hugetlb migration fails when h->surplus_huge_page >= h->nr_overcommit_huge_pages.
>>>
>>> Yes, I am aware of that. I should have been more explicit in the
>>> changelog. Sorry about that and thanks for pointing it out explicitly.
>>> To be honest I wasn't really sure what to do about this. The code path
>>> is really complex and it made my head spin. I fail to see why we have to
>>> call alloc_huge_page and mess with reservations at all.
>>
>> Oops! I missed that in my review.
>>
>> Since alloc_huge_page was called with avoid_reserve == 1, it should not
>> do anything with reserve counts. One potential issue with the existing
>> code is cgroup accounting done by alloc_huge_page. When the new target
>> page is allocated, it is charged against the cgroup even though the original
>> page is still accounted for. If we are 'at the cgroup limit', the migration
>> may fail because of this.
>
> Yeah, the existing code seems just broken. I strongly suspect that the
> allocation API for hugetlb was so complicated that this was just a
> natural result of a confusion with some follow up changes on top.
>
>> I like your new code below as it explicitly takes reserve and cgroup
>> accounting out of the picture for migration. Let me think about it
>> for another day before providing a Reviewed-by.
>
> Thanks a lot!

You can add,

Reviewed-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>

I had some concerns about transferring huge page state during migration
not specific to this patch, so I did a bunch of testing.

--
Mike Kravetz