When a GFN range is entirely private, it's unnecessary for
kvm_handle_hva_range() to invoke handlers for the GFN range, because
1) the gfn_range.attr_filter for the handler is KVM_FILTER_SHARED, which
   is for shared mappings only;
2) KVM has already zapped all shared mappings before setting the memory
   attribute to private.

This can avoid unnecessary zaps on private mappings for VMs of type
KVM_X86_SW_PROTECTED_VM, e.g., during auto numa balancing scans of VMAs.

Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
---
 virt/kvm/kvm_main.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index f769d1dccc21..e615ad405ce4 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -620,6 +620,17 @@ static __always_inline kvm_mn_ret_t kvm_handle_hva_range(struct kvm *kvm,
 			gfn_range.slot = slot;
 			gfn_range.lockless = range->lockless;
 
+#ifdef CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES
+			/*
+			 * If GFN range are all private, no need to invoke the
+			 * handler.
+			 */
+			if (kvm_range_has_memory_attributes(kvm, gfn_range.start,
+							    gfn_range.end, ~0,
+							    KVM_MEMORY_ATTRIBUTE_PRIVATE))
+				continue;
+#endif
+
 			if (!r.found_memslot) {
 				r.found_memslot = true;
 				if (!range->lockless) {
-- 
2.43.2

On Fri, Aug 22, 2025, Yan Zhao wrote:
> When a GFN range is entirely private, it's unnecessary for
> kvm_handle_hva_range() to invoke handlers for the GFN range, because
> 1) the gfn_range.attr_filter for the handler is KVM_FILTER_SHARED, which
>    is for shared mappings only;
> 2) KVM has already zapped all shared mappings before setting the memory
>    attribute to private.
> 
> This can avoid unnecessary zaps on private mappings for VMs of type
> KVM_X86_SW_PROTECTED_VM, e.g., during auto numa balancing scans of VMAs.

This feels like the wrong place to try and optimize spurious zaps.  x86 should
be skipping SPTEs that don't match.  For KVM_X86_SW_PROTECTED_VM, I don't think
we care about spurious zpas, because that's a testing-only type that doesn't have
line of sight to be being a "real" type.

For SNP, we might care?  But actually zapping private SPTEs would require
userspace to retain the shared mappings across a transition, _and_ be running
NUMA autobalancing in the first place.  If someone actually cares about optimizing
this scenario, KVM x86 could track private SPTEs via a software-available bit.

We also want to move away from KVM_MEMORY_ATTRIBUTE_PRIVATE and instead track
private vs. shared in the gmem instance.

So I'm inclined to skip this...

> Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
> ---
>  virt/kvm/kvm_main.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index f769d1dccc21..e615ad405ce4 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -620,6 +620,17 @@ static __always_inline kvm_mn_ret_t kvm_handle_hva_range(struct kvm *kvm,
>  			gfn_range.slot = slot;
>  			gfn_range.lockless = range->lockless;
>  
> +#ifdef CONFIG_KVM_GENERIC_MEMORY_ATTRIBUTES
> +			/*
> +			 * If GFN range are all private, no need to invoke the
> +			 * handler.
> +			 */
> +			if (kvm_range_has_memory_attributes(kvm, gfn_range.start,
> +							    gfn_range.end, ~0,
> +							    KVM_MEMORY_ATTRIBUTE_PRIVATE))
> +				continue;
> +#endif
> +
>  			if (!r.found_memslot) {
>  				r.found_memslot = true;
>  				if (!range->lockless) {
> -- 
> 2.43.2
>

On Mon, Aug 25, 2025 at 02:05:22PM -0700, Sean Christopherson wrote:
> On Fri, Aug 22, 2025, Yan Zhao wrote:
> > When a GFN range is entirely private, it's unnecessary for
> > kvm_handle_hva_range() to invoke handlers for the GFN range, because
> > 1) the gfn_range.attr_filter for the handler is KVM_FILTER_SHARED, which
> >    is for shared mappings only;
> > 2) KVM has already zapped all shared mappings before setting the memory
> >    attribute to private.
> > 
> > This can avoid unnecessary zaps on private mappings for VMs of type
> > KVM_X86_SW_PROTECTED_VM, e.g., during auto numa balancing scans of VMAs.
> 
> This feels like the wrong place to try and optimize spurious zaps.  x86 should
> be skipping SPTEs that don't match.  For KVM_X86_SW_PROTECTED_VM, I don't think
> we care about spurious zpas, because that's a testing-only type that doesn't have
> line of sight to be being a "real" type.
> 
> For SNP, we might care?  But actually zapping private SPTEs would require
> userspace to retain the shared mappings across a transition, _and_ be running
> NUMA autobalancing in the first place.  If someone actually cares about optimizing
Hmm, "running NUMA autobalancing" + "madvise(MADV_DONTNEED)" can still trigger
the spurious zaps.

task_numa_work  ==> found a VMA
  change_prot_numa
    change_protection
      change_pud_range ==> mmu_notifier_invalidate_range_start() if !pud_none()
 
Let me use munmap() in patch 3 to guard againt spurious zap then.

> this scenario, KVM x86 could track private SPTEs via a software-available bit.
> 
> We also want to move away from KVM_MEMORY_ATTRIBUTE_PRIVATE and instead track
> private vs. shared in the gmem instance.
>
> So I'm inclined to skip this...
Fair enough. Thank you for the detailed explanation!