Re: [PATCH v3 03/70] RAMBlock/guest_memfd: Enable KVM_GUEST_MEMFD_ALLOW_HUGEPAGE

[Claudio Fontana]

On 11/30/23 18:51, Daniel P. Berrangé wrote:
> On Thu, Nov 30, 2023 at 05:54:26PM +0100, David Hildenbrand wrote:
>> On 30.11.23 17:01, Sean Christopherson wrote:
>>> On Thu, Nov 30, 2023, David Hildenbrand wrote:
>>>> On 30.11.23 08:32, Xiaoyao Li wrote:
>>>>> On 11/20/2023 5:26 PM, David Hildenbrand wrote:
>>>>>>
>>>>>>>> ... did you shamelessly copy that from hw/virtio/virtio-mem.c ? ;)
>>>>>>>
>>>>>>> Get caught.
>>>>>>>
>>>>>>>> This should be factored out into a common helper.
>>>>>>>
>>>>>>> Sure, will do it in next version.
>>>>>>
>>>>>> Factor it out in a separate patch. Then, this patch is get small that
>>>>>> you can just squash it into #2.
>>>>>>
>>>>>> And my comment regarding "flags = 0" to patch #2 does no longer apply :)
>>>>>>
>>>>>
>>>>> I see.
>>>>>
>>>>> But it depends on if KVM_GUEST_MEMFD_ALLOW_HUGEPAGE will appear together
>>>>> with initial guest memfd in linux (hopefully 6.8)
>>>>> https://lore.kernel.org/all/CABgObfa=DH7FySBviF63OS9sVog_wt-AqYgtUAGKqnY5Bizivw@mail.gmail.com/
>>>>>
>>>>
>>>> Doesn't seem to be in -next if I am looking at the right tree:
>>>>
>>>> https://git.kernel.org/pub/scm/virt/kvm/kvm.git/log/?h=next
>>>
>>> Yeah, we punted on adding hugepage support for the initial guest_memfd 
>>> merge so
>>> as not to rush in kludgy uABI.  The internal KVM code isn't problematic, we 
>>> just
>>> haven't figured out exactly what the ABI should look like, e.g. should 
>>> hugepages
>>> be dependent on THP being enabled, and if not, how does userspace discover 
>>> the
>>> supported hugepage sizes?
>>
>> Are we talking about THP or hugetlb? They are two different things, and
>> "KVM_GUEST_MEMFD_ALLOW_HUGEPAGE" doesn't make it clearer what we are talking
>> about.
>>
>> This patch here "get_thp_size()" indicates that we care about THP, not
>> hugetlb.
>>
>>
>> THP lives in:
>>      /sys/kernel/mm/transparent_hugepage/
>> and hugetlb in:
>>      /sys/kernel/mm/hugepages/
>>
>> THP for shmem+anon currently really only supports PMD-sized THP, that size
>> can be observed via:
>>      /sys/kernel/mm/transparent_hugepage/hpage_pmd_size
>>
>> hugetlb sizes can be detected simply by looking at the folders inside
>> /sys/kernel/mm/hugepages/. "tools/testing/selftests/mm/vm_util.c" in the
>> kernel has a function "detect_hugetlb_page_sizes()" that uses that interface
>> to detect the sizes.
>>
>>
>> But likely we want THP support here. Because for hugetlb, one would actually
>> have to instruct the kernel which size to use, like we do for memfd with
>> hugetlb.
> 
> Would we not want both ultimately ?
> 
> THP is good because it increases performance vs non-HP out of the box
> without the user or mgmt app having to make any decisions.
> 
> It does not give you deterministic performance though, because it has
> to opportunistically assign huge pages basd on what is available and
> that may differ each time a VM is launched.  Explicit admin/mgmt app
> controlled huge page usage gives determinism, at the cost of increased
> mgmt overhead.
> 
> Both are valid use cases depending on the tradeoff a deployment and/or
> mgmt app wants to make.

Absolutely, it really depends on the definition of "performance" for the 
specific goal the user is trying to achieve.
There are very prominent use cases where THP is a big no-no due to the latency 
introduced.

C