Re: [PATCH v2] KVM: nVMX: Fix setting of CR0 and CR4 in guest mode

[Gleb Natapov <gleb@redhat.com>]

On Thu, Mar 07, 2013 at 09:53:49AM +0100, Jan Kiszka wrote:
> On 2013-03-07 09:43, Gleb Natapov wrote:
> > On Thu, Mar 07, 2013 at 09:12:19AM +0100, Jan Kiszka wrote:
> >> On 2013-03-07 08:51, Gleb Natapov wrote:
> >>> On Mon, Mar 04, 2013 at 08:40:29PM +0100, Jan Kiszka wrote:
> >>>> The logic for calculating the value with which we call kvm_set_cr0/4 was
> >>>> broken (will definitely be visible with nested unrestricted guest mode
> >>>> support). Also, we performed the check regarding CR0_ALWAYSON too early
> >>>> when in guest mode.
> >>>>
> >>>> What really needs to be done on both CR0 and CR4 is to mask out L1-owned
> >>>> bits and merge them in from GUEST_CR0/4. In contrast, arch.cr0/4 and
> >>>> arch.cr0/4_guest_owned_bits contain the mangled L0+L1 state and, thus,
> >>>> are not suited as input.
> >>>>
> >>>> For both CRs, we can then apply the check against VMXON_CRx_ALWAYSON and
> >>>> refuse the update if it fails. To be fully consistent, we implement this
> >>>> check now also for CR4.
> >>>>
> >>>> Finally, we have to set the shadow to the value L2 wanted to write
> >>>> originally.
> >>>>
> >>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
> >>>> ---
> >>>>
> >>>> Changes in v2:
> >>>>  - keep the non-misleading part of the comment in handle_set_cr0
> >>>>
> >>>>  arch/x86/kvm/vmx.c |   46 +++++++++++++++++++++++++++++++---------------
> >>>>  1 files changed, 31 insertions(+), 15 deletions(-)
> >>>>
> >>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> >>>> index 7cc566b..832b7b4 100644
> >>>> --- a/arch/x86/kvm/vmx.c
> >>>> +++ b/arch/x86/kvm/vmx.c
> >>>> @@ -4605,37 +4605,53 @@ vmx_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
> >>>>  /* called to set cr0 as appropriate for a mov-to-cr0 exit. */
> >>>>  static int handle_set_cr0(struct kvm_vcpu *vcpu, unsigned long val)
> >>>>  {
> >>>> -	if (to_vmx(vcpu)->nested.vmxon &&
> >>>> -	    ((val & VMXON_CR0_ALWAYSON) != VMXON_CR0_ALWAYSON))
> >>>> -		return 1;
> >>>> -
> >>>>  	if (is_guest_mode(vcpu)) {
> >>>> +		struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
> >>>> +		unsigned long orig_val = val;
> >>>> +
> >>>>  		/*
> >>>>  		 * We get here when L2 changed cr0 in a way that did not change
> >>>>  		 * any of L1's shadowed bits (see nested_vmx_exit_handled_cr),
> >>>> -		 * but did change L0 shadowed bits. This can currently happen
> >>>> -		 * with the TS bit: L0 may want to leave TS on (for lazy fpu
> >>>> -		 * loading) while pretending to allow the guest to change it.
> >>>> +		 * but did change L0 shadowed bits.
> >>>>  		 */
> >>>> -		if (kvm_set_cr0(vcpu, (val & vcpu->arch.cr0_guest_owned_bits) |
> >>>> -			 (vcpu->arch.cr0 & ~vcpu->arch.cr0_guest_owned_bits)))
> >>>> +		val = (val & ~vmcs12->cr0_guest_host_mask) |
> >>>> +			(vmcs_read64(GUEST_CR0) & vmcs12->cr0_guest_host_mask);
> >>> I think using GUEST_CR0 here is incorrect. It contains combination of bits
> >>> set by L2, L1 and L0 and here we need to get only L2/L1 mix which is in
> >>> vcpu->arch.cr0 (almost, but good enough for this case). Why vcpu->arch.cr0
> >>> contains right L2/L1 mix?
> >>
> >> L0/L1. E.g., kvm_set_cr0 unconditionally injects X86_CR0_ET and masks
> >> out reserved bits. But you are right, GUEST_CR0 isn't much better. And
> >> maybe that mangling in kvm_set_cr0 is a corner case we can ignore.
> >>
> > I think we can. ET is R/O and wired to 1, so it does not matter what
> > guest writes there it should be treated as 1. About reserved bits spec
> > says that software should write what it reads there and does not specify
> > what happens if software does not follow this.
> > 
> >>> Because it was set to vmcs12->guest_cr0 during
> >>> L2 #vmentry. While L2 is running three things may happen to CR0:
> >>>
> >>>  1. L2 writes to a bit that is not shadowed neither by L1 nor by L0. It
> >>>     will go strait to GUEST_CR0.
> >>>  2. L2 writes to a bit shadowed by L1. L1 #vmexit will be emulated. On the
> >>>     next #vmetry vcpu->arch.cr0 will be set to whatever value L1 calculated.
> >>>  3. L2 writes to a bit shadowed by L0, but not L1. This is the case we
> >>>     are handling here. And if we will do it right vcpu->arch.cr0 will be
> >>>     up-to-date at the end.
> >>>
> >>> The only case when, while this code running, vcpu->arch.cr0 has not
> >>> up-to-date value is if 1 happened, but since L2 guest overwriting cr0
> >>> here anyway it does not matter what it previously set in GUEST_CR0. The
> >>> correct bits are in a new cr0 value provided by val and accessible by
> >>> (val & ~vmcs12->cr0_guest_host_mask).
> >>
> >> I need to think about it again. Maybe vmcs12->guest_cr0 is best, but
> >> that's a shot from the hips now.
> >>
> > I do not think it is correct because case 3 does not update it. So if 3
> > happens twice without L1 #vmexit between then vmcs12->guest_cr0 will be
> > outdated.
> 
> Again, the only thing that matters here is L1's, not L0's view on the
> "real" CR0 value. So guest_cr0 is never outdated (/wrt
> cr0_guest_host_mask) as it will be updated by L1 in step 2. Even if
> arch.cr0 vs. guest_cr0 makes no difference in practice, the latter is
> more consistent, so I will go for it unless you can convince me it is wrong.
> 
Hmm, yes you are right that wrt cr0_guest_host_mask guest_cr0 should be
up-to-date. Please write a big comment about it. And what about moving
VMXON_CR0_ALWAYSON check into vmx_set_cr0()?

--
			Gleb.