RFC Protect VMM from speculative attacks using x86 PKRU
Get series with b4
b4 am [email protected] change-id 20260729-feature-pkey-dev-c8a026db85b4Versions newest first
- v1 RFC Latest6 patches 5 repliesIn review
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/6 | x86: Introduce basic PKRU hardware wrappersMAINTAINERS util/ | +97 | 0 replies | |
| 2/6 | kvm: Add guest memory Pkey initializationinclude/ system/ util/ +1 more | +26 | 2 replies | |
| 3/6 | physmem: Tag guest RAMBlocks with Protection Keyinclude/ include/ system/ +2 more | +37 | 0 replies | |
| 4/6 | kvm: Lock guest RAMBlocks via PKRU during host userspace executionaccel/ include/ system/ +1 more | +58 | 0 replies | |
| 5/6 | x86: Add xstate parsing and PKRU offset detectionutil/ | +53 | 0 replies | |
| 6/6 | kvm: Implement SIGSEGV sentinel for Pkey recoveryutil/ | +155 | 2 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [RFC PATCH 0/6] Protect VMM from speculative attacks using x86 PKRU
Hi,
This RFC patch series introduces a hardware-assisted memory protection mechanism in QEMU using x86 Protection Keys for Userspace (PKRU / MPK) to complete the Address Space Isolation ecosystem by protecting the VMM from guest-to-host speculative execution attacks.
In a guest-to-host speculative side-channel attack against QEMU, a malicious guest mistrains the VMM branch predictor to execute a speculative gadget in host mode. This gadget speculatively reads a host secret and uses that secret as an index to touch guest RAMBlocks, leaving an observable microarchitectural footprint in the CPU cache. Existing mitigations like issuing IBPB on every VM-exit incur high overhead, while guest_memfd cannot protect memory that must stay mapped.
To break this covert channel, we protect guest RAMBlocks using a dedicated pkey (pkey_mprotect).
- During normal QEMU host userspace execution (I/O handling, timers, event loops), we lock this pkey by setting PKEY_DISABLE_ACCESS in the PKRU register via WRPKRU. Because CPU pipelines natively enforce PKRU restrictions during out-of-order execution, any speculative gadget in host mode attempting to touch guest RAMBlocks is blocked by the MMU at the hardware level, leaving zero footprint in the cache.
- Before entering KVM (KVM_RUN), we temporarily unlock the pkey, and immediately re-lock it upon returning to QEMU userspace.
- To gracefully handle legitimate host accesses to guest RAMBlocks when locked, a SEGV_PKUERR signal handler unlocks the pkey in the interrupted thread's xstate and issues an IBPB (PR_SPEC_INDIRECT_BRANCH) to prevent speculative leakage before resuming execution.
Comments and feedback on this approach are very welcome!
To: [email protected]
Cc: Paolo Bonzini <[email protected]>
Cc: Zhao Liu <[email protected]>
Cc: Richard Henderson <[email protected]>
Cc: Philippe Mathieu-Daudé <[email protected]>
Cc: Peter Xu <[email protected]>
Cc: [email protected]
Cc: James Houghton <[email protected]>
Cc: Mingwei Zhang <[email protected]>
Cc: Dave Hansen <[email protected]>
Cc: Brendan Jackman <[email protected]>
Cc: Reiji Watanabe <[email protected]>
Cc: Jacky Li <[email protected]>
---
Jacky Li (6):
[RFC PATCH 1/6] x86: Introduce basic PKRU hardware wrappers
[RFC PATCH 2/6] kvm: Add guest memory Pkey initialization
[RFC PATCH 3/6] physmem: Tag guest RAMBlocks with Protection Key [RFC PATCH 4/6] kvm: Lock guest RAMBlocks via PKRU during host userspace execution [RFC PATCH 5/6] x86: Add xstate parsing and PKRU offset detection [RFC PATCH 6/6] kvm: Implement SIGSEGV sentinel for Pkey recovery
MAINTAINERS | 1 + accel/kvm/kvm-all.c | 7 +- include/exec/cpu-common.h | 4 + include/qemu/mmap-alloc.h | 5 + system/physmem.c | 22 ++- system/vl.c | 1 + util/meson.build | 2 +- util/mmap-alloc.c | 3 +- util/pkey.c | 386 ++++++++++++++++++++++++++++++++++++++++++++++ 9 files changed, 424 insertions(+), 7 deletions(-)
---
base-commit: fa19879df1658f96ac07365fca8835b7decd6995
change-id: 20260729-feature-pkey-dev-c8a026db85b4
Best regards,
Signature · 1 line
Jacky Li <[email protected]>