On Thu, 2025-06-05 at 16:01 +0300, kirill.shutemov@xxxxxxxxxxxxxxx wrote:
> On Fri, May 23, 2025 at 03:00:56PM +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.
> >
> > I looked closer to this and while it is good idea, but ctor in kmem_cache
> > cannot fail which makes this approach not viable.
> >
> > I guess we can a constructor directly into struct kvm_mmu_memory_cache.
> > Let me play with this.
>
> I failed to make it work.
>
> We need to have destructor paired with the constructor that would do
> PAMT-aware freeing. And redirect all free paths to it. It requires
> substantial rework. I don't think it worth the effort.
>
> Will do manual PAMT management for SPT in TDX code.
Thanks for the effort.
Maybe something below?
diff --git a/arch/x86/kvm/mmu/mmu_internal.h
b/arch/x86/kvm/mmu/mmu_internal.h
index db8f33e4de62..48732270bff0 100644
--- a/arch/x86/kvm/mmu/mmu_internal.h
+++ b/arch/x86/kvm/mmu/mmu_internal.h
@@ -164,8 +164,10 @@ static inline bool is_mirror_sp(const struct
kvm_mmu_page *sp)
return sp->role.is_mirror;
}
-static inline void kvm_mmu_alloc_external_spt(struct kvm_vcpu *vcpu, struct
kvm_mmu_page *sp)
+static inline int kvm_mmu_alloc_external_spt(struct kvm_vcpu *vcpu, struct
kvm_mmu_page *sp)
{
+ int r;
+
/*
* external_spt is allocated for TDX module to hold private EPT
mappings,
* TDX module will initialize the page by itself.
@@ -173,6 +175,12 @@ static inline void kvm_mmu_alloc_external_spt(struct
kvm_vcpu *vcpu, struct kvm_
* KVM only interacts with sp->spt for private EPT operations.
*/
sp->external_spt = kvm_mmu_memory_cache_alloc(&vcpu-
>arch.mmu_external_spt_cache);
+
+ r = tdx_pamt_get(virt_to_page(sp->external_spt));
+ if (r)
+ free_page((unsigned long)sp->external_spt);
+
+ return r;
}
static inline gfn_t kvm_gfn_root_bits(const struct kvm *kvm, const struct
kvm_mmu_page *root)
diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
index 7f3d7229b2c1..2d3a716d9195 100644
--- a/arch/x86/kvm/mmu/tdp_mmu.c
+++ b/arch/x86/kvm/mmu/tdp_mmu.c
@@ -55,7 +55,10 @@ void kvm_mmu_uninit_tdp_mmu(struct kvm *kvm)
static void tdp_mmu_free_sp(struct kvm_mmu_page *sp)
{
- free_page((unsigned long)sp->external_spt);
+ if (sp->external_spt) {
+ free_page((unsigned long)sp->external_spt);
+ tdx_pamt_put(virt_to_page(sp->external_spt));
+ }
free_page((unsigned long)sp->spt);
kmem_cache_free(mmu_page_header_cache, sp);
}
@@ -1277,8 +1280,13 @@ int kvm_tdp_mmu_map(struct kvm_vcpu *vcpu, struct
kvm_page_fault *fault)
*/
sp = tdp_mmu_alloc_sp(vcpu);
tdp_mmu_init_child_sp(sp, &iter);
- if (is_mirror_sp(sp))
- kvm_mmu_alloc_external_spt(vcpu, sp);
+ if (is_mirror_sp(sp)) {
+ r = kvm_mmu_alloc_external_spt(vcpu, sp);
+ if (r) {
+ tdp_mmu_free_sp(sp);
+ goto retry;
+ }
+ }
sp->nx_huge_page_disallowed = fault->huge_page_disallowed;
