Reto Buerki reported a failure in a nested VMM when running with HLT
interception disabled in L1. When putting L2 into HLT, KVM never actually
enters L2 and instead cancels the nested run and pretends that VM-Enter to
L2 completed and then exited on HLT (which KVM intercepted). Because KVM
never actually runs L2, KVM skips the pending MMU update for L2 and so
leaves a stale value in vmcs02.GUEST_CR3. If the next wake event for L2
triggers a nested VM-Exit, KVM will refresh vmcs12->guest_cr3 from
vmcs02.GUEST_CR3 and consume the stale value.
Fix the issue by unconditionally writing vmcs02.GUEST_CR3 during nested
VM-Enter instead of deferring the update to vmx_set_cr3(), and skip the
update of GUEST_CR3 in vmx_set_cr3() when running L2. I.e. make the
nested code fully responsible for vmcs02.GUEST_CR3.
I really wanted to go with a different fix of handling this as a one-off
case in the HLT flow (in nested_vmx_run()), and then following that up
with a cleanup of VMX's CR3 handling, e.g. to do proper dirty tracking
instead of having the nested code do manual VMREADs and VMWRITEs. I even
went so far as to hide vcpu->arch.cr3 (put CR3 in vcpu->arch.regs), but
things went south when I started working through the dirty tracking logic.
Because EPT can be enabled *without* unrestricted guest, enabling EPT
doesn't always mean GUEST_CR3 really is the guest CR3 (unlike SVM's NPT).
And because the unrestricted guest handling of GUEST_CR3 is dependent on
whether the guest has paging enabled, VMX can't even do a clean handoff
based on unrestricted guest. In a nutshell, dynamically handling the
transitions of GUEST_CR3 ownership in VMX is a nightmare, so fixing this
purely within the context of nested VMX turned out to be the cleanest fix.
Sean Christopherson (2):
KVM: nVMX: Always write vmcs02.GUEST_CR3 during nested VM-Enter
KVM: VMX: Skip GUEST_CR3 VMREAD+VMWRITE if the VMCS is up-to-date
arch/x86/kvm/vmx/nested.c | 8 ++++++++
arch/x86/kvm/vmx/vmx.c | 15 ++++++++++-----
2 files changed, 18 insertions(+), 5 deletions(-)
--
2.22.0
On 9/26/19 11:43 PM, Sean Christopherson wrote:
> Reto Buerki reported a failure in a nested VMM when running with HLT
> interception disabled in L1. When putting L2 into HLT, KVM never actually
> enters L2 and instead cancels the nested run and pretends that VM-Enter to
> L2 completed and then exited on HLT (which KVM intercepted). Because KVM
> never actually runs L2, KVM skips the pending MMU update for L2 and so
> leaves a stale value in vmcs02.GUEST_CR3. If the next wake event for L2
> triggers a nested VM-Exit, KVM will refresh vmcs12->guest_cr3 from
> vmcs02.GUEST_CR3 and consume the stale value.
>
> Fix the issue by unconditionally writing vmcs02.GUEST_CR3 during nested
> VM-Enter instead of deferring the update to vmx_set_cr3(), and skip the
> update of GUEST_CR3 in vmx_set_cr3() when running L2. I.e. make the
> nested code fully responsible for vmcs02.GUEST_CR3.
>
> I really wanted to go with a different fix of handling this as a one-off
> case in the HLT flow (in nested_vmx_run()), and then following that up
> with a cleanup of VMX's CR3 handling, e.g. to do proper dirty tracking
> instead of having the nested code do manual VMREADs and VMWRITEs. I even
> went so far as to hide vcpu->arch.cr3 (put CR3 in vcpu->arch.regs), but
> things went south when I started working through the dirty tracking logic.
>
> Because EPT can be enabled *without* unrestricted guest, enabling EPT
> doesn't always mean GUEST_CR3 really is the guest CR3 (unlike SVM's NPT).
> And because the unrestricted guest handling of GUEST_CR3 is dependent on
> whether the guest has paging enabled, VMX can't even do a clean handoff
> based on unrestricted guest. In a nutshell, dynamically handling the
> transitions of GUEST_CR3 ownership in VMX is a nightmare, so fixing this
> purely within the context of nested VMX turned out to be the cleanest fix.
>
> Sean Christopherson (2):
> KVM: nVMX: Always write vmcs02.GUEST_CR3 during nested VM-Enter
> KVM: VMX: Skip GUEST_CR3 VMREAD+VMWRITE if the VMCS is up-to-date
>
> arch/x86/kvm/vmx/nested.c | 8 ++++++++
> arch/x86/kvm/vmx/vmx.c | 15 ++++++++++-----
> 2 files changed, 18 insertions(+), 5 deletions(-)
Tested-by: Reto Buerki <[email protected]>
Thanks!
Sean Christopherson <[email protected]> writes:
> Reto Buerki reported a failure in a nested VMM when running with HLT
> interception disabled in L1. When putting L2 into HLT, KVM never actually
> enters L2 and instead cancels the nested run and pretends that VM-Enter to
> L2 completed and then exited on HLT (which KVM intercepted). Because KVM
> never actually runs L2, KVM skips the pending MMU update for L2 and so
> leaves a stale value in vmcs02.GUEST_CR3. If the next wake event for L2
> triggers a nested VM-Exit, KVM will refresh vmcs12->guest_cr3 from
> vmcs02.GUEST_CR3 and consume the stale value.
>
> Fix the issue by unconditionally writing vmcs02.GUEST_CR3 during nested
> VM-Enter instead of deferring the update to vmx_set_cr3(), and skip the
> update of GUEST_CR3 in vmx_set_cr3() when running L2. I.e. make the
> nested code fully responsible for vmcs02.GUEST_CR3.
>
> I really wanted to go with a different fix of handling this as a one-off
> case in the HLT flow (in nested_vmx_run()), and then following that up
> with a cleanup of VMX's CR3 handling, e.g. to do proper dirty tracking
> instead of having the nested code do manual VMREADs and VMWRITEs. I even
> went so far as to hide vcpu->arch.cr3 (put CR3 in vcpu->arch.regs), but
> things went south when I started working through the dirty tracking logic.
>
> Because EPT can be enabled *without* unrestricted guest, enabling EPT
> doesn't always mean GUEST_CR3 really is the guest CR3 (unlike SVM's NPT).
> And because the unrestricted guest handling of GUEST_CR3 is dependent on
> whether the guest has paging enabled, VMX can't even do a clean handoff
> based on unrestricted guest. In a nutshell, dynamically handling the
> transitions of GUEST_CR3 ownership in VMX is a nightmare, so fixing this
> purely within the context of nested VMX turned out to be the cleanest fix.
>
> Sean Christopherson (2):
> KVM: nVMX: Always write vmcs02.GUEST_CR3 during nested VM-Enter
> KVM: VMX: Skip GUEST_CR3 VMREAD+VMWRITE if the VMCS is up-to-date
>
Series:
Tested-by: Vitaly Kuznetsov <[email protected]>
--
Vitaly