Re: [RFC PATCH] KVM: TDX: Decouple TDX init mem region from kvm_gmem_populate()

[Yan Zhao]

On Mon, Jul 28, 2025 at 05:45:35PM -0700, Vishal Annapurve wrote:
> On Mon, Jul 28, 2025 at 2:49 AM Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote:
> >
> > On Fri, Jul 18, 2025 at 08:57:10AM -0700, Vishal Annapurve wrote:
> > > On Fri, Jul 18, 2025 at 2:15 AM Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, Jul 15, 2025 at 09:10:42AM +0800, Yan Zhao wrote:
> > > > > On Mon, Jul 14, 2025 at 08:46:59AM -0700, Sean Christopherson wrote:
> > > > > > > >         folio = __kvm_gmem_get_pfn(file, slot, index, &pfn, &is_prepared, &max_order);
> > > > > > > If max_order > 0 is returned, the next invocation of __kvm_gmem_populate() for
> > > > > > > GFN+1 will return is_prepared == true.
> > > > > >
> > > > > > I don't see any reason to try and make the current code truly work with hugepages.
> > > > > > Unless I've misundertood where we stand, the correctness of hugepage support is
> > > > > Hmm. I thought your stand was to address the AB-BA lock issue which will be
> > > > > introduced by huge pages, so you moved the get_user_pages() from vendor code to
> > > > > the common code in guest_memfd :)
> > > > >
> > > > > > going to depend heavily on the implementation for preparedness.  I.e. trying to
> > > > > > make this all work with per-folio granulartiy just isn't possible, no?
> > > > > Ah. I understand now. You mean the right implementation of __kvm_gmem_get_pfn()
> > > > > should return is_prepared at 4KB granularity rather than per-folio granularity.
> > > > >
> > > > > So, huge pages still has dependency on the implementation for preparedness.
> > > > Looks with [3], is_prepared will not be checked in kvm_gmem_populate().
> > > >
> > > > > Will you post code [1][2] to fix non-hugepages first? Or can I pull them to use
> > > > > as prerequisites for TDX huge page v2?
> > > > So, maybe I can use [1][2][3] as the base.
> > > >
> > > > > [1] https://lore.kernel.org/all/aG_pLUlHdYIZ2luh@xxxxxxxxxx/
> > > > > [2] https://lore.kernel.org/all/aHEwT4X0RcfZzHlt@xxxxxxxxxx/
> >
> > From the PUCK, looks Sean said he'll post [1][2] for 6.18 and Michael will post
> > [3] soon.
> >
> > hi, Sean, is this understanding correct?
> >
> > > IMO, unless there is any objection to [1], it's un-necessary to
> > > maintain kvm_gmem_populate for any arch (even for SNP). All the
> > > initial memory population logic needs is the stable pfn for a given
> > > gfn, which ideally should be available using the standard mechanisms
> > > such as EPT/NPT page table walk within a read KVM mmu lock (This patch
> > > already demonstrates it to be working).
> > >
> > > It will be hard to clean-up this logic once we have all the
> > > architectures using this path.
> > >
> > > [1] https://lore.kernel.org/lkml/CAGtprH8+x5Z=tPz=NcrQM6Dor2AYBu3jiZdo+Lg4NqAk0pUJ3w@xxxxxxxxxxxxxx/
> > IIUC, the suggestion in the link is to abandon kvm_gmem_populate().
> > For TDX, it means adopting the approach in this RFC patch, right?
> 
> Yes, IMO this RFC is following the right approach as posted.
Ira has been investigating this for a while, see if he has any comment.