Re: [RFC PATCH v5 19/45] KVM: Allow owner of kvm_mmu_memory_cache to provide a custom page allocator

[Sean Christopherson]

On Tue, Feb 03, 2026, Rick P Edgecombe wrote:
> On Tue, 2026-02-03 at 12:12 -0800, Sean Christopherson wrote:
> > > E.g., I thought we might be able to use it to allocate a structure which has
> > > "pair of DPAMT pages" so it could be assigned to 'struct kvm_mmu_page'.  But
> > > it seems you abandoned this idea.  May I ask why?  Just want to understand
> > > the reasoning here.
> > 
> > Because that requires more complexity and there's no known use case, and I
> > don't see an obvious way for a use case to come along.  All of the
> > motiviations for a custom allocation scheme that I can think of apply only to
> > full pages, or fit nicely in a kmem_cache.
> > 
> > Specifically, the "cache" logic is already bifurcated between "kmem_cache' and
> > "page" usage.  Further splitting the "page" case doesn't require modifications
> > to the "kmem_cache" case, whereas providing a fully generic solution would
> > require additional changes, e.g. to handle this code:
> > 
> > 	page = (void *)__get_free_page(gfp_flags);
> > 	if (page && mc->init_value)
> > 		memset64(page, mc->init_value, PAGE_SIZE / sizeof(u64));
> > 
> > It certainly wouldn't be much complexity, but this code is already a bit
> > awkward, so I don't think it makes sense to add support for something that
> > will probably never be used.
> 
> The thing that the design needlessly works around is that we can rely on that
> there are only two DPAMT pages per 2MB range. We don't need the dynamic page
> count allocations.
> 
> This means we don't need to pass around the list of pages that lets arch/x86
> take as many pages as it needs. We can maybe just pass in a struct like Kai was
> suggesting to the get/put helpers. So I was in the process of trying to morph
> this series in that direction to get rid of the complexity resulting from the
> dynamic assumption. 
> 
> This was what I had done in response to v4 discussions, so now retrofitting it
> into this new ops scheme. Care to warn me off of this before I have something to
> show?

That's largely orthogonal to this change.  This change is about preparing the
DPAMT when S-EPT page is allocated versus being installed.  The fact that DPAMT
requires at most two pages versus a more dynamic maximum is irrelevant.

The caches aren't about dynamic sizes (though they play nicely with them), they're
about:

  (a) not having to deal with allocating under spinlock
  (b) not having to free memory that goes unused (for a single page fault)
  (c) batching allocations for performance reasons (with the caveat that I doubt
      anyone has measured the performance impact in many, many years).

None of those talking points change at all if KVM needs to provide 2 pages versus
N pages.  The max number of pages needed for page tables is pretty much the same
thing as DPAMT, just with a higher max (4/5 vs. 2).  In both cases, the allocated
pages may or may not be consumed for any given fault.

For the leaf pages (including the hugepage splitting cases), which don't utilize
KVM's kvm_mmu_memory_cache, I wouldn't expect the KVM details to change all that
much.  In fact, they shouldn't change at all, because tracking 2 pages versus N
pages in "struct tdx_pamt_cache" is a detail that is 100% buried in the TDX subsystem
(which was pretty much the entire goal of my design).

Though maybe I'm misunderstanding what you have in mind?