[PATCH] KVM: x86/mmu: Prevent hugepage recovery on non-guest_memfd private GFNs
From: Ackerley Tng via B4 Relay
Date: Fri Nov 20 2026 - 16:07:42 EST
From: Ackerley Tng <ackerleytng@xxxxxxxxxx>
When determining the maximum mapping level outside of a page fault
(such as when disabling dirty logging), KVM queries the VM memory
attributes to determine whether a GFN is private. Userspace can mark a
GFN as private in the VM attributes even when the covering memslot is
not backed by guest_memfd.
Previously, this caused the mapping level calculation to fall into
kvm_gmem_max_mapping_level(), triggering a kernel warning.
Here's 1 way to trigger the warning on a TDX VM, with VM memory attributes
enabled on KVM.
1. Create a non-guest_memfd memslot covering a 1G-aligned base GFN with
dirty logging enabled.
2. In that memslot, fault a shared page outside the first 2M chunk so a
non-leaf PTE at the 1G level is installed.
3. Mark the first 2M chunk (at the 1G-aligned base GFN) private. Conversion
zaps that 2M chunk, while the shared mapping outside it keeps the 1G
non-leaf PTE present.
4. Disable dirty logging on the memslot.
Hugepage recovery iterates over the present 1G non-leaf PTE, sees that its
base GFN is private and the 2M range is not mixed, and calls into
kvm_gmem_max_mapping_level() without a fault context, hitting the WARN.
Return 4K as the maximum mapping level when a GFN is marked private in
a non-fault path for a memslot without guest_memfd. Fault paths already
disallow mapping private GFNs in non-guest_memfd memslots, and
returning 4K prevents hugepage recovery on such GFNs.
Fixes: a3522ac71f5e ("KVM: x86/mmu: Enforce guest_memfd's max order when recovering hugepages")
Signed-off-by: Ackerley Tng <ackerleytng@xxxxxxxxxx>
---
arch/x86/kvm/mmu/mmu.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index f89e9f8a6cc2f..af40fbe9001a7 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -3489,6 +3489,14 @@ int kvm_mmu_max_mapping_level(struct kvm *kvm, struct kvm_page_fault *fault,
} else {
max_level = PG_LEVEL_NUM;
is_private = kvm_is_private_gfn(kvm, gfn);
+
+ /*
+ * is_private (set in VM attributes) should never have been used
+ * with a non-guest_memfd memslot. Return PG_LEVEL_4K to prevent
+ * hugepage recovery.
+ */
+ if (is_private && !kvm_slot_has_gmem(slot))
+ return PG_LEVEL_4K;
}
max_level = min(max_level, max_huge_page_level);
---
base-commit: d0f871bfb831705d9f55479f5b2b40f636ea7b16
change-id: 20261120-gmem-fix-gmem-max-mapping-level-accbb71c3352
Best regards,
--
Ackerley Tng <ackerleytng@xxxxxxxxxx>