[PATCH v12 05/12] KVM: guest_memfd: Enforce NUMA mempolicy using shared policy

[Sean Christopherson]

On Fri, Oct 10, 2025, Ackerley Tng wrote:
> Sean Christopherson <seanjc at google.com> writes:
> 
> > On Fri, Oct 10, 2025, Shivank Garg wrote:
> >> >> @@ -112,6 +114,19 @@ static int kvm_gmem_prepare_folio(struct kvm *kvm, struct kvm_memory_slot *slot,
> >> >>  	return r;
> >> >>  }
> >> >>  
> >> >> +static struct mempolicy *kvm_gmem_get_folio_policy(struct gmem_inode *gi,
> >> >> +						   pgoff_t index)
> >> > 
> >> > How about kvm_gmem_get_index_policy() instead, since the policy is keyed
> >> > by index?
> >
> > But isn't the policy tied to the folio?  I assume/hope that something will split
> > folios if they have different policies for their indices when a folio contains
> > more than one page.  In other words, how will this work when hugepage support
> > comes along?
> >
> > So yeah, I agree that the lookup is keyed on the index, but conceptually aren't
> > we getting the policy for the folio?  The index is a means to an end.
> >
> 
> I think the policy is tied to the index.
> 
> When we mmap(), there may not be a folio at this index yet, so any folio
> that gets allocated for this index then is taken from the right NUMA
> node based on the policy.
> 
> If the folio is later truncated, the folio just goes back to the NUMA
> node, but the memory policy remains for the next folio to be allocated
> at this index.

Right.  Though thinking about this more, there's no reason to have "index" in
the name, kvm_gmem_get_policy() is sufficient.  E.g. we don't have "index" in
the name for things like kvm_get_vcpu().

Luckily, it's all made moot by Shivank's fixup :-)

> >> >> +{
> >> >> +#ifdef CONFIG_NUMA
> >> >> +	struct mempolicy *mpol;
> >> >> +
> >> >> +	mpol = mpol_shared_policy_lookup(&gi->policy, index);
> >> >> +	return mpol ? mpol : get_task_policy(current);
> >> > 
> >> > Should we be returning NULL if no shared policy was defined?
> >> > 
> >> > By returning NULL, __filemap_get_folio_mpol() can handle the case where
> >> > cpuset_do_page_mem_spread().
> >> > 
> >> > If we always return current's task policy, what if the user wants to use
> >> > cpuset_do_page_mem_spread()?
> >> > 
> >> 
> >> I initially followed shmem's approach here.
> >> I agree that returning NULL maintains consistency with the current default
> >> behavior of cpuset_do_page_mem_spread(), regardless of CONFIG_NUMA.
> >> 
> >> I'm curious what could be the practical implications of cpuset_do_page_mem_spread()
> >> v/s get_task_policy() as the fallback?
> >
> > Userspace could enable page spreading on the task that triggers guest_memfd
> > allocation.  I can't conjure up a reason to do that, but I've been surprised
> > more than once by KVM setups.
> >
> >> Which is more appropriate for guest_memfd when no policy is explicitly set
> >> via mbind()?
> >
> > I don't think we need to answer that question?  Userspace _has_ set a policy,
> > just through cpuset, not via mbind().  So while I can't imagine there's a sane
> > use case for cpuset_do_page_mem_spread() with guest_memfd, I also don't see a
> > reason why KVM should effectively disallow it.
> >
> > And unless I'm missing something, allocation will eventually fallback to
> > get_task_policy() (in alloc_frozen_pages_noprof()), so by explicitly getting the
> > task policy in guest_memfd, KVM is doing _more_ work than necessary _and_ is
> > unnecessarily restricting usersepace.
> >
> > Add in that returning NULL would align this code with the ->get_policy hook (and
> > could be shared again, I assume), and my vote is definitely to return NULL and
> > not get in the way.
> 
> ... although if we are going to return NULL then we can directly use
> mpol_shared_policy_lookup(), so the first discussion is moot.

Ha!  Great minds think alike, right!!?!

> Though looking slightly into the future, shareability (aka memory
> attributes or shared/private state within guest_memfd inodes) are also
> keyed by index, and is a property of the index and not the folio (since
> shared/private state is defined even before folios are allocated for a
> given index.

Yeah, which further reinforces that having "index" in the function name is
superfluous (and potentially confusing), e.g. IMO the proposed helpers:

  kvm_gmem_get_attributes()
  kvm_gmem_is_private_mem()
  kvm_gmem_is_shared_mem()

are far better than e.g.:

  kvm_gmem_get_index_attributes()
  kvm_gmem_is_index_private_mem()
  kvm_gmem_is_index_shared_mem()