Re: [PATCH v13 17/35] KVM: Add transparent hugepage support for dedicated guest memory

[Sean Christopherson]

On Wed, Nov 01, 2023, Xiaoyao Li wrote:
> On 10/31/2023 10:16 PM, Sean Christopherson wrote:
> > On Tue, Oct 31, 2023, Xiaoyao Li wrote:
> > > On 10/28/2023 2:21 AM, Sean Christopherson wrote:
> > > > Extended guest_memfd to allow backing guest memory with transparent
> > > > hugepages. Require userspace to opt-in via a flag even though there's no
> > > > known/anticipated use case for forcing small pages as THP is optional,
> > > > i.e. to avoid ending up in a situation where userspace is unaware that
> > > > KVM can't provide hugepages.
> > > 
> > > Personally, it seems not so "transparent" if requiring userspace to opt-in.
> > > 
> > > People need to 1) check if the kernel built with TRANSPARENT_HUGEPAGE
> > > support, or check is the sysfs of transparent hugepage exists; 2)get the
> > > maximum support hugepage size 3) ensure the size satisfies the alignment;
> > > before opt-in it.
> > > 
> > > Even simpler, userspace can blindly try to create guest memfd with
> > > transparent hugapage flag. If getting error, fallback to create without the
> > > transparent hugepage flag.
> > > 
> > > However, it doesn't look transparent to me.
> > 
> > The "transparent" part is referring to the underlying kernel mechanism, it's not
> > saying anything about the API.  The "transparent" part of THP is that the kernel
> > doesn't guarantee hugepages, i.e. whether or not hugepages are actually used is
> > (mostly) transparent to userspace.
> > 
> > Paolo also isn't the biggest fan[*], but there are also downsides to always
> > allowing hugepages, e.g. silent failure due to lack of THP or unaligned size,
> > and there's precedent in the form of MADV_HUGEPAGE.
> > 
> > [*] https://lore.kernel.org/all/84a908ae-04c7-51c7-c9a8-119e1933a189@xxxxxxxxxx
> 
> But it's different than MADV_HUGEPAGE, in a way. Per my understanding, the
> failure of MADV_HUGEPAGE is not fatal, user space can ignore it and
> continue.
>
> However, the failure of KVM_GUEST_MEMFD_ALLOW_HUGEPAGE is fatal, which leads
> to failure of guest memfd creation.

Failing KVM_CREATE_GUEST_MEMFD isn't truly fatal, it just requires different
action from userspace, i.e. instead of ignoring the error, userspace could redo
KVM_CREATE_GUEST_MEMFD with KVM_GUEST_MEMFD_ALLOW_HUGEPAGE=0.

We could make the behavior more like MADV_HUGEPAGE, e.g. theoretically we could
extend fadvise() with FADV_HUGEPAGE, or add a guest_memfd knob/ioctl() to let
userspace provide advice/hints after creating a guest_memfd.  But I suspect that
guest_memfd would be the only user of FADV_HUGEPAGE, and IMO a post-creation hint
is actually less desirable.

KVM_GUEST_MEMFD_ALLOW_HUGEPAGE will fail only if userspace didn't provide a
compatible size or the kernel doesn't support THP.  An incompatible size is likely
a userspace bug, and for most setups that want to utilize guest_memfd, lack of THP
support is likely a configuration bug.  I.e. many/most uses *want* failures due to
KVM_GUEST_MEMFD_ALLOW_HUGEPAGE to be fatal.

> For current implementation, I think maybe KVM_GUEST_MEMFD_DESIRE_HUGEPAGE
> fits better than KVM_GUEST_MEMFD_ALLOW_HUGEPAGE? or maybe *PREFER*?

Why?  Verbs like "prefer" and "desire" aren't a good fit IMO because they suggest
the flag is a hint, and hints are usually best effort only, i.e. are ignored if
there is a fundamental incompatibility.

"Allow" isn't perfect, e.g. I would much prefer a straight KVM_GUEST_MEMFD_USE_HUGEPAGES
or KVM_GUEST_MEMFD_HUGEPAGES flag, but I wanted the name to convey that KVM doesn't
(yet) guarantee hugepages.  I.e. KVM_GUEST_MEMFD_ALLOW_HUGEPAGE is stronger than
a hint, but weaker than a requirement.  And if/when KVM supports a dedicated memory
pool of some kind, then we can add KVM_GUEST_MEMFD_REQUIRE_HUGEPAGE.