Re: [f2fs-dev] [PATCH kvm-next V11 6/7] KVM: guest_memfd: Enforce NUMA mempolicy using shared policy

[Sean Christopherson <seanjc@google.com>]

On Wed, Oct 15, 2025, Gregory Price wrote:
> On Fri, Sep 26, 2025 at 12:36:27PM -0700, Sean Christopherson via Linux-f2fs-devel wrote:
> > > 
> > > static struct mempolicy *kvm_gmem_get_policy(struct vm_area_struct *vma,
> > > 					     unsigned long addr, pgoff_t *pgoff)
> > > {
> > > 	*pgoff = vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT);
> > > 
> > > 	return __kvm_gmem_get_policy(GMEM_I(file_inode(vma->vm_file)), *pgoff);
> > 
> > Argh!!!!!  This breaks the selftest because do_get_mempolicy() very specifically
> > falls back to the default_policy, NOT to the current task's policy.  That is
> > *exactly* the type of subtle detail that needs to be commented, because there's
> > no way some random KVM developer is going to know that returning NULL here is
> > important with respect to get_mempolicy() ABI.
> > 
> 
> Do_get_mempolicy was designed to be accessed by the syscall, not as an
> in-kernel ABI.

Ya, by "get_mempolicy() ABI" I meant the uABI for the get_mempolicy syscall.

> get_task_policy also returns the default policy if there's nothing
> there, because that's what applies.
> 
> I have dangerous questions:

Not dangerous at all, I find them very helpful!

> why is __kvm_gmem_get_policy using
> 	mpol_shared_policy_lookup()
> instead of
> 	get_vma_policy()

With the disclaimer that I haven't followed the gory details of this series super
closely, my understanding is...

Because the VMA is a means to an end, and we want the policy to persist even if
the VMA goes away.

With guest_memfd, KVM effectively inverts the standard MMU model.  Instead of mm/
being the primary MMU and KVM being a secondary MMU, guest_memfd is the primary
MMU and any VMAs are secondary (mostly; it's probably more like 1a and 1b).  This
allows KVM to map guest_memfd memory into a guest without a VMA, or with more
permissions than are granted to host userspace, e.g. guest_memfd memory could be
writable by the guest, but read-only for userspace.

But we still want to support things like mbind() so that userspace can ensure
guest_memfd allocations align with the vNUMA topology presented to the guest,
or are bound to the NUMA node where the VM will run.  We considered adding equivalent
file-based syscalls, e.g. fbind(), but IIRC the consensus was that doing so was
unnecessary (and potentially messy?) since we were planning on eventually adding
mmap() support to guest_memfd anyways.

> get_vma_policy does this all for you

I assume that doesn't work if the intent is for new VMAs to pick up the existing
policy from guest_memfd?  And more importantly, guest_memfd needs to hook
->set_policy so that changes through e.g. mbind() persist beyond the lifetime of
the VMA.

> struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
>                                  unsigned long addr, int order, pgoff_t *ilx)
> {
>         struct mempolicy *pol;
> 
>         pol = __get_vma_policy(vma, addr, ilx);
>         if (!pol)
>                 pol = get_task_policy(current);
>         if (pol->mode == MPOL_INTERLEAVE ||
>             pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
>                 *ilx += vma->vm_pgoff >> order;
>                 *ilx += (addr - vma->vm_start) >> (PAGE_SHIFT + order);
>         }
>         return pol;
> }
> 
> Of course you still have the same issue: get_task_policy will return the
> default, because that's what applies.
> 
> do_get_mempolicy just seems like the completely incorrect interface to
> be using here.