Re: [RFC PATCH 0/3] Add hugetlb MADV_DONTNEED support

[David Hildenbrand <david@redhat.com>]

On 27.01.22 18:52, Axel Rasmussen wrote:
> On Thu, Jan 27, 2022 at 3:57 AM David Hildenbrand <david@redhat.com> wrote:
>>
>> On 13.01.22 19:03, Mike Kravetz wrote:
>>> Userfaultfd selftests for hugetlb does not perform UFFD_EVENT_REMAP
>>> testing.  However, mremap support was recently added in commit
>>> 550a7d60bd5e ("mm, hugepages: add mremap() support for hugepage backed
>>> vma").  While attempting to enable mremap support in the test, it was
>>> discovered that the mremap test indirectly depends on MADV_DONTNEED.
>>>
>>> hugetlb does not support MADV_DONTNEED.  However, the only thing
>>> preventing support is a check in can_madv_lru_vma().  Simply removing
>>> the check will enable support.
>>>
>>> This is sent as a RFC because there is no existing use case calling
>>> for hugetlb MADV_DONTNEED support except possibly the userfaultfd test.
>>> However, adding support makes sense as it is fairly trivial and brings
>>> hugetlb functionality more in line with 'normal' memory.
>>>
>>
>> Just a note:
>>
>> QEMU doesn't use huge anonymous memory directly (MAP_ANON | MAP_HUGE...)
>> but instead always goes either via hugetlbfs or via memfd.
>>
>> For MAP_PRIVATE hugetlb mappings, fallocate(FALLOC_FL_PUNCH_HOLE) seems
>> to get the job done (IOW: also discards private anon pages). See the
>> comments in the QEMU code below. I remember that that is somewhat
>> inconsistent. For ordinary MAP_PRIVATE mapped files I remember that we
>> always need fallocate(FALLOC_FL_PUNCH_HOLE) + madvise(QEMU_MADV_DONTNEED)
>> to make sure
>>
>> a) All file pages are removed
>> b) All private anon pages are removed
>>
>> IIRC hugetlbfs really is different in that regard, but maybe other fs
>> behave similarly.
>>
>> That's why QEMU was able to live for now without MADV_DONTNEED support
>> for hugetlbfs and most probably won't ever need it.
> 
> Agreed, all of the production use cases I'm aware of use hugetlbfs,
> not MAP_HUGE...
> 
> But, I would say this is convenient for testing purposes. It's
> slightly more convenient to not have to mount hugetlbfs / perform the
> associated setup for tests.

Creating a memfd is not too hard, but yes, not a single-liner. Maybe the
uffd test should go via memfds for hugetlb instead. But maybe that
limits the mremap functionality? No expert.

> 
> Perhaps that's only a small motivation for enabling this, but then
> again Mike's patch to do so is likewise very small. :)

... and apparently buggy :P

-- 
Thanks,

David / dhildenb