Re: [RFC, PATCH 08/12] KVM: x86/tdp_mmu: Add phys_prepare() and phys_cleanup() to kvm_x86_ops

[Huang, Kai]

On Wed, 2025-05-14 at 09:43 +0300, kirill.shutemov@xxxxxxxxxxxxxxx wrote:
> On Wed, May 14, 2025 at 12:00:17AM +0000, Huang, Kai wrote:
> > On Mon, 2025-05-12 at 12:55 +0300, Kirill A. Shutemov wrote:
> > > On Fri, May 09, 2025 at 09:25:58AM +0800, Yan Zhao wrote:
> > > > On Thu, May 08, 2025 at 04:23:56PM +0300, Kirill A. Shutemov wrote:
> > > > > On Tue, May 06, 2025 at 07:55:17PM +0800, Yan Zhao wrote:
> > > > > > On Fri, May 02, 2025 at 04:08:24PM +0300, Kirill A. Shutemov wrote:
> > > > > > > The functions kvm_x86_ops::link_external_spt() and
> > > > > > > kvm_x86_ops::set_external_spte() are used to assign new memory to a VM.
> > > > > > > When using TDX with Dynamic PAMT enabled, the assigned memory must be
> > > > > > > covered by PAMT.
> > > > > > > 
> > > > > > > The new function kvm_x86_ops::phys_prepare() is called before
> > > > > > > link_external_spt() and set_external_spte() to ensure that the memory is
> > > > > > > ready to be assigned to the virtual machine. In the case of TDX, it
> > > > > > > makes sure that the memory is covered by PAMT.
> > > > > > > 
> > > > > > > kvm_x86_ops::phys_prepare() is called in a context where struct kvm_vcpu
> > > > > > > is available, allowing the implementation to allocate memory from a
> > > > > > > per-VCPU pool.
> > > > > > > 
> > > > > > Why not invoke phys_prepare() and phys_cleanup() in set_external_spte_present()?
> > > > > > Or in tdx_sept_set_private_spte()/tdx_sept_link_private_spt()?
> > > > > 
> > > > > Because the memory pool we allocated from is per-vcpu and we lost access
> > > > > to vcpu by then. And not all callers provide vcpu.
> > > > Maybe we can get vcpu via kvm_get_running_vcpu(), as in [1].
> > > > Then for callers not providing vcpu (where vcpu is NULL), we can use per-KVM
> > > > cache? 
> > > 
> > > Hm. I was not aware of kvm_get_running_vcpu(). Will play with it, thanks.
> > 
> > I am not sure why per-vcpu cache matters.
> > 
> > For non-leaf SEPT pages, AFAICT the "vcpu->arch.mmu_external_spt_cache" is just
> > an empty cache, and eventually __get_free_page() is used to allocate in:
> >                                                                                             
> >   sp->external_spt = 
> > 	kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_external_spt_cache);
> > 
> > So why not we actually create a kmem_cache for it with an actual 'ctor', and we
> > can call tdx_alloc_page() in that.  This makes sure when the "external_spt" is
> > allocated, the underneath PAMT entry is there.
> 
> This would make hard to debug PAMT memory leaks. external_spt pages in the
> pool will have PAMT memory tied to them, so we will have non-zero PAMT
> memory usage with zero TDs running.

Why is that?  AFAICT all 'external_spt' pages are freed when TD is gone.

> 
> > For the last level guest memory page, similar to SEV, we can hook the
> > kvm_arch_gmem_prepare() to call tdx_alloc_page() to make PAMT entry ready.
> 
> I don't think kvm_arch_gmem_prepare() is right place to allocate PAMT
> memory. THPs are dynamic and page order can change due to split or
> collapse between the time the page is allocated and gets mapped into EPT.
> I am not sure if SEV code is correct in this regard.

Yeah, agreed.  Not sure how does SEV-SNP handles large page split/merge either.