vcpu->arch.exception currently contains the vmcs02 IDT-vectored info
through the entire execution of the vmexit. This makes it harder
to keep that information safe when vcpu->arch.exception is reused for
an exception (such as a page fault) that happens while L0 handles a vmexit.
When this happens, there are two cases:
- the exception causes a vmexit to L1; in that case, the exception in the
IDT-vectored info is not reinjected; vcpu->arch.exception is
reused to build the VM-exit interruption info.
- the exception doesn't cause a vmexit to L1; in that case,
vcpu->arch.exception is changed to a double fault which is injected
normally into L2 via KVM_REQ_EVENT.
The fix is easy if the vmcs12 IDT-vectored info is prepared early, in
vmx_complete_interrupts (patches 1-2). I wanted to include this in 4.14,
but Radim is on vacation and also I'm not sure how this interacts with
Wanpeng's other refactoring of nested exceptions, so I'm sending it out
only as RFC.
I am applying only the first patch to kvm/queue. If I get a review, that
one can be put in 4.14.
Paolo
Paolo Bonzini (4):
KVM: nVMX: move vmentry tasks from prepare_vmcs12 to
enter_vmx_non_root_mode
KVM: nVMX: fill nested IDT-vectored event info on all L2->L0 exits
KVM: x86: pass struct kvm_queued_exception to kvm_multiple_exception
KVM: x86: allow overwriting L2 reinjected exception with L1 vmexit
arch/x86/include/asm/kvm_host.h | 2 +
arch/x86/kvm/svm.c | 79 +++++++-------
arch/x86/kvm/vmx.c | 224 +++++++++++++++++++++-------------------
arch/x86/kvm/x86.c | 86 +++++++++------
4 files changed, 217 insertions(+), 174 deletions(-)
--
1.8.3.1
