[PATCH v3 04/25] KVM: x86/mmu: Add dedicated API to map guest_memfd pfn into TDP MMU

[Sean Christopherson]

On Wed, Nov 05, 2025, Yan Zhao wrote:
> On Wed, Nov 05, 2025 at 03:32:29PM +0800, Yan Zhao wrote:
> > On Tue, Nov 04, 2025 at 09:57:26AM -0800, Sean Christopherson wrote:
> > > On Thu, Oct 30, 2025, Yan Zhao wrote:
> > > > On Wed, Oct 22, 2025 at 12:53:53PM +0800, Yan Zhao wrote:
> > > > > On Thu, Oct 16, 2025 at 05:32:22PM -0700, Sean Christopherson wrote:
> > > > > > Link: https://lore.kernel.org/all/20250709232103.zwmufocd3l7sqk7y@amd.com
> > > > > 
> > > > > Hi Sean,                                                                         
> > > > > 
> > > > > Will you post [1] to fix the AB-BA deadlock issue for huge page in-place
> > > > > conversion as well?
> > > 
> > > If you (or anyone) has the bandwidth, please pick it up.  I won't have cycles to
> > > look at that for many weeks (potentially not even this calendar year).
> > Got it!
> > On the other hand, do you think we can address the warning as below?
> > The code is based on [2].
> Hmm, updated the diff.
> 
> diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
> index 7b4a4474d468..543e1eb9db65 100644
> --- a/virt/kvm/guest_memfd.c
> +++ b/virt/kvm/guest_memfd.c
> @@ -853,6 +853,9 @@ static int kvm_gmem_init_inode(struct inode *inode, loff_t size, u64 flags)
>         inode->i_size = size;
>         mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
>         mapping_set_inaccessible(inode->i_mapping);
> +       if (flags &GUEST_MEMFD_FLAG_MMAP)
> +               lockdep_set_subclass(&inode->i_mapping->invalidate_lock, 1);
> +
>         /* Unmovable mappings are supposed to be marked unevictable as well. */
>         WARN_ON_ONCE(!mapping_unevictable(inode->i_mapping));
> 
>  
> > As noted in [3], the only scenario can trigger the warning after [2] is when a
> > process creates a TDX VM with non-in-place-conversion guest_memfd and a normal
> > VM with in-place-conversion guest_memfd. The two invalidate_lock's don't contend
> > with each other theoretically.

Hmm, no, I think we need to hoist gup() call outside of filemap_invalidate_lock(),
because I don't think this is strictly limited to TDX VMs without in-place
conversion.  Even with in-place conversion, I think KVM should allow the source
page to be shared memory, at which point I believe this becomes a legimate AB-BA
issue.

In general, playing lockdep games with so many subsystems involved terrifies me.

> > [2] https://lore.kernel.org/all/cover.1760731772.git.ackerleytng@google.com/
> > [3] https://lore.kernel.org/all/aQMi%2Fn9DVyeaWsVH@yzhao56-desk.sh.intel.com/