Re: [RFC PATCH v4 0/7] KVM: pfncache: Add guest_memfd support to pfncache

[Sean Christopherson <seanjc@google.com>]

On Mon, Apr 20, 2026, Takahiro Itazuri wrote:
> [ based on 6.18 with [1] ]
> 
> This patch series adds guest_memfd support to gfn_to_pfn_cache (a.k.a.
> pfncache).  (This is still labelled RFC since its dependency [1] has not
> yet been merged.)
> 
> === Problem Statement ===
> 
> pfncache does not work with guest_memfd.  pfncaches resolve PFNs via
> hva_to_pfn(), which requires a userspace mapping and relies on GUP.
> This does not work for guest_memfd in the following two ways:
> 
>   * guest_memfd created without MMAP flag does not have a userspace
>     mapping due to the nature of private memory.

Sorry, but NAK.  At least until we have a *very* clear and generally agreed-upon
plan for how KVM is going to handle accessing guest memory NO_DIRECT_MAP (my brain
finally caught up to what this series wants to do).

Simply allowing gpcs to tap directly into guest_memfd doesn't allow KVM x86 to
use guest_memfd for non-CoCo VMs.  There might be a special class of VMs that can
squeak by, but any VM that finds itself in __kvm_{read,write}_guest_page() or in
FNAME(walk_addr_generic) needs to have guest memory mapped into host userspace.
Predicting what memory a "normal" VM will use for TDP or the source/destination
of emulated MMIO is simply not feasible.

At that point, the "guest_memfd created without MMAP flag" use case is meaningless,
because creating such a guest_memfd instance is *only* useful for backing PRIVATE
memory, which KVM obviously shouldn't be mapping.

This topic was discussed heavily at LPC 2024 (and/or LPC 2023?), and the consensus
was that mapping guest_memfd into host userspace is the least awful choice, the
only complete alternative being a *massive* rework of KVM x86 internals (other
architectures don't suffer as much due to limited emulation).

>   KVM: Rename invalidate_begin to invalidate_start for consistencyo

I'll grab this one for 7.2.

>   KVM: pfncache: Rename invalidate_start() helper
>   KVM: pfncache: Invalidate on gmem invalidation and memattr updates
>   KVM: selftests: Test pfncache with gmem-backed memory
>   KVM: selftests: Test pfncache invalidation for gmem-backed memory