RFC Add SEV-SNP restricted injection hypervisor support

Hi all,

This v5 of the Restricted Injection hypervisor support patches.

The previous version:

https://lore.kernel.org/all/[email protected]

This is based on d4b7fb647204f0c81dfeae2d1a708e4d858e0c94 (kvm/next).

This series addresses v4 upstream review comments from Sashiko, Tom, and Joerg.

The main changes in v5 include:

1. Guest-Driven IRR-to-ISR Transition: Instead of moving the vector from IRR to ISR directly upon reading (matching legacy interrupt injection), KVM now waits for the guest acknowledgment before moving the interrupt vector from IRR to ISR.

2. No-EOI Handling: Added handling for the 'no_eoi_required' case.

3. HLT Intercept and Architectural Alignment: Added a new patch to properly intercept HLT when Restricted Injection is active. This addresses the guest's loss of control over HLT/interrupt interactions by immediate guest re-entry on pending events.

4. Sleep-in-Atomic and Scheduling Fixes: Resolved critical issues caused by kvm_vcpu_map() calling paths in the HLT fast-path (sleeping with local IRQs disabled) and during normal vCPU block polling (incorrect TASK_INTERRUPTIBLE/TASK_RUNNING state transitions causing livelocks).

5. Refined Cancel Injection: When canceling injection, KVM now only cancels the #HV event while preserving the pending interrupt information in the doorbell page.

6. Priority-Based Vector Overwriting: KVM now overwrites the interrupt vector in the doorbell page when a higher-priority vector appears, avoiding sending an unnecessary duplicate #HV.

7. Documentation and Review Feedback: Improved several comments for clarity and implemented code fixes for valid issues pointed out by Sashiko during review.

This patch set remains an RFC. There is an alternative approach that I will present at KVM Forum 2026 in the talk: "COCONUT-SVSM, KVM Planes, and Secure Interrupt Injection - Upstream". I plan to update this series further after gathering feedback from the KVM Forum community.

Thanks,
Melody

Changelog:
---------
v4

This is v4 of the Restricted Injection hypervisor support patches.

The previous version was submitted here:
https://lore.kernel.org/all/[email protected]

The new series are rebased on commit 6b802031877a995456c528095c41d1948546bf45 (kvm/next).

Restricted Injection (RI) is not a prerequisite for Planes support. Even though the SVSM requires RI, RI support can be submitted independently. Since the number of Planes patches is large, submit RI as its own series.

v3

This is v3 of the Restricted Injection hypervisor support patches.

The previous, v2 version was submitted here:
https://lore.kernel.org/r/[email protected]

Since then, the series experienced the following changes:

1. Remove unused field in struct hvdb_events.

2. Add comments explaining sev_snp_blocked() with no_further_signal and vector. Explain that only a single interrupt vector is presented by the hypervisor in the doorbell page. Only when the interrupt is acked, the next interrupt is presented.

3. Add warning in sev_snp_cancel_injection() when vector is not #HV_VECTOR.

4. Update hvdb_map() failure handling. When it fails, no interrupt will be injected.

5. Remove sev_snp_queue_exception() in svm_update_soft_interrupt_rip(), since the soft interrupt is only BP_VECTOR and OF_VECTOR, so it will always return false, then sev_snp_queue_exception() will be executed in svm_inject_exception() always.

6. Add new #HV IPI feature in XAPIC and X2APIC mode, test each mode with three IPI types: broadcast, self-IPI, and allbutself.

v2

Hi all,

This is a v2 of the restricted injection hypervisor support patches.

The previous version was submitted here:
https://lore.kernel.org/r/[email protected]

Since the previous submission, one issue reported by the kernel test robot was fixed.

All comments and review feedback are appreciated.

Thanks.

v1

Operating systems may not handle unexpected interrupt or exception sequences. A malicious hypervisor can inject random interrupt or exception sequences, putting guest drivers or guest OS kernels into an unexpected state, which could lead to security issues.

To address this concern, SEV-SNP restricts the injection of interrupts and exceptions to those only allowed by the guest. Restricted Injection disables all hypervisor-based interrupt queuing and event injection for all vectors, allowing only a single vector, #HV (28), which is reserved for SNP guest use but is never generated by hardware. #HV is only permitted to be injected into VMSAs that execute with Restricted Injection.

Guests operating with Restricted Injection are expected to communicate with the hypervisor about events via a software-managed para-virtualization interface. This interface can utilize #HV injection as a doorbell to inform the guest that new events have occurred. This patch set implements Restricted Injection on the KVM side directly into VMPL0.

Overview:

The GHCB 2.0 specification[1] defines #HV doorbell page and the #HV doorbell page NAE event allows for an SEV-SNP guest to register a doorbell page for use with the hypervisor injection exception (#HV). When Restricted Injection is active, only #HV exceptions can be injected into the guest, and the hypervisor follows the GHCB #HV doorbell communication to inject the exception or interrupt. Restricted Injection can be enabled by setting the bit in vmsa_features.

The patchset is rebased on the kvm/next (commit 15e1c3d65975524c5c792fcd59f7d89f00402261).

Testing:

The patchset has been tested with the sev-snp guest, ovmf and qemu supporting restricted injection.

Four test sets:
1.ls -lr /
2.apt update
3.fio
4.perf

Melody Wang (8):
x86/sev: Define the #HV doorbell page structure
KVM: SVM: Add support for the SEV-SNP #HV doorbell page NAE event
KVM: SVM: Inject #HV when Restricted Injection is active
KVM: SVM: Inject NMIs when Restricted Injection is active
KVM: SVM: Inject MCEs when Restricted Injection is active
KVM: SVM: Handle HLT intercept when Restricted Injection is enabled.
KVM: SVM: Add support for the SEV-SNP #HV IPI NAE event
KVM: SVM: Enable Restricted Injection for an SEV-SNP guest

 arch/x86/include/asm/cpufeatures.h |   1 +
 arch/x86/include/asm/kvm-x86-ops.h |   3 +
 arch/x86/include/asm/kvm_host.h    |   4 +-
 arch/x86/include/asm/sev-common.h  |   1 +
 arch/x86/include/asm/svm.h         |  35 +++
 arch/x86/include/uapi/asm/svm.h    |   6 +
 arch/x86/kvm/i8259.c               |  45 +++-
 arch/x86/kvm/irq.c                 |  51 +++-
 arch/x86/kvm/irq.h                 |   3 +
 arch/x86/kvm/lapic.c               |  46 +++-
 arch/x86/kvm/lapic.h               |   3 +
 arch/x86/kvm/svm/sev.c             | 403 ++++++++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.c             |  96 ++++++-
 arch/x86/kvm/svm/svm.h             |  35 +++
 arch/x86/kvm/vmx/main.c            |  10 +
 arch/x86/kvm/vmx/vmx.c             |   5 +
 arch/x86/kvm/vmx/x86_ops.h         |   1 +
 arch/x86/kvm/x86.c                 |   9 +
 18 files changed, 735 insertions(+), 22 deletions(-)

2.43.0