While attempting to understand the big comment in
kvm_mmu_slot_remove_write_access() about TLB flushing, I discovered a
bug in the way the TDP MMU write-protects GFNs. I have not managed to
reproduce the bug as it requires a rather complex set up of live
migrating a VM that is using nested virtualization while the TDP MMU is
enabled.
Patch 1 fixes the bug and is CC'd to stable.
Patch 2-3 fix, document, and enforce invariants around MMU-writable
and Host-writable bits.
Patch 4 fixes up the aformentioned comment to be more readable.
Tested using the kvm-unit-tests and KVM selftests.
v2:
- Skip setting the SPTE when MMU-writable is already clear [Sean]
- Add patches for {MMU,Host}-writable invariants [Sean]
- Fix inaccuracies in kvm_mmu_slot_remove_write_access() comment [Sean]
v1: https://lore.kernel.org/kvm/20220112215801.3502286-1-dmatlack@xxxxxxxxxx/
David Matlack (4):
KVM: x86/mmu: Fix write-protection of PTs mapped by the TDP MMU
KVM: x86/mmu: Clear MMU-writable during changed_pte notifier
KVM: x86/mmu: Document and enforce MMU-writable and Host-writable
invariants
KVM: x86/mmu: Improve TLB flush comment in
kvm_mmu_slot_remove_write_access()
arch/x86/kvm/mmu/mmu.c | 31 ++++++++++++++++++++--------
arch/x86/kvm/mmu/spte.c | 1 +
arch/x86/kvm/mmu/spte.h | 42 ++++++++++++++++++++++++++++++++------
arch/x86/kvm/mmu/tdp_mmu.c | 6 +++---
4 files changed, 62 insertions(+), 18 deletions(-)
base-commit: fea31d1690945e6dd6c3e89ec5591490857bc3d4
--
2.34.1.703.g22d0c6ccf7-goog
