Move NMI VM-Exit handling into vmx_vcpu_enter_exit() to fix a (mostly
benign?) bug where NMIs can be unblocked prior to servicing the NMI that
triggered the VM-Exit, e.g. if instrumentation triggers a fault and thus
an IRET. I deliberately didn't tag any of these for stable@ as the odds
of me screwing something up or of a backport going sideways seems higher
than out-of-order NMIs causing major problems.
The bulk of this series is just getting various helpers/paths ready for
noinstr usage.
I kept the use of a direct call to a dedicated entry point for NMIs
(doubled down really). AFAICT, there are no issues with the direct call
in the current code, and I don't know enough about FRED to know if using
INT $2 would be better or worse, i.e. less churn seemed like the way to
go. And if reverting to INT $2 in the future is desirable, splitting NMI
and IRQ handling makes it quite easy to do so as all the relevant code
that needs to be ripped out is isolated.
Sean Christopherson (7):
KVM: x86: Make vmx_get_exit_qual() and vmx_get_intr_info()
noinstr-friendly
KVM: VMX: Allow VM-Fail path of VMREAD helper to be instrumented
KVM: VMX: Always inline eVMCS read/write helpers
KVM: VMX: Always inline to_vmx() and to_kvm_vmx()
x86/entry: KVM: Use dedicated VMX NMI entry for 32-bit kernels too
KVM: VMX: Provide separate subroutines for invoking NMI vs. IRQ
handlers
KVM: VMX: Handle NMI VM-Exits in noinstr region
arch/x86/include/asm/idtentry.h | 16 +++-----
arch/x86/kernel/nmi.c | 8 ++--
arch/x86/kvm/kvm_cache_regs.h | 12 ++++++
arch/x86/kvm/vmx/hyperv.h | 20 ++++-----
arch/x86/kvm/vmx/vmcs.h | 4 +-
arch/x86/kvm/vmx/vmenter.S | 72 ++++++++++++++++++---------------
arch/x86/kvm/vmx/vmx.c | 55 +++++++++++++------------
arch/x86/kvm/vmx/vmx.h | 18 ++++-----
arch/x86/kvm/vmx/vmx_ops.h | 2 +
arch/x86/kvm/x86.h | 6 +--
10 files changed, 117 insertions(+), 96 deletions(-)
base-commit: 208f1c64e255fe3a29083880818e010ebdf585c6
--
2.39.0.rc1.256.g54fd8350bd-goog
