On Mon, Mar 04, 2013 at 08:23:52PM +0100, Jan Kiszka wrote:
> On 2013-03-04 19:39, Gleb Natapov wrote:
> > On Mon, Mar 04, 2013 at 07:08:08PM +0100, Jan Kiszka wrote:
> >> On 2013-03-04 18:56, Gleb Natapov wrote:
> >>> On Mon, Mar 04, 2013 at 03:25:47PM +0100, Jan Kiszka wrote:
> >>>> On 2013-03-04 15:15, Gleb Natapov wrote:
> >>>>> On Mon, Mar 04, 2013 at 03:09:51PM +0100, Jan Kiszka wrote:
> >>>>>> On 2013-03-04 14:22, Gleb Natapov wrote:
> >>>>>>> On Thu, Feb 28, 2013 at 10:44:47AM +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 <[email protected]>
> >>>>>>>> ---
> >>>>>>>>
> >>>>>>>> Found while making unrestricted guest mode working. Not sure what
> >>>>>>>> impact
> >>>>>>>> the bugs had on current feature level, if any.
> >>>>>>>>
> >>>>>>>> For interested folks, I've pushed my nEPT environment here:
> >>>>>>>>
> >>>>>>>> git://git.kiszka.org/linux-kvm.git nept-hacking
> >>>>>>>>
> >>>>>>>> arch/x86/kvm/vmx.c | 49
> >>>>>>>> ++++++++++++++++++++++++++++++-------------------
> >>>>>>>> 1 files changed, 30 insertions(+), 19 deletions(-)
> >>>>>>>>
> >>>>>>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> >>>>>>>> index 7cc566b..d1dac08 100644
> >>>>>>>> --- a/arch/x86/kvm/vmx.c
> >>>>>>>> +++ b/arch/x86/kvm/vmx.c
> >>>>>>>> @@ -4605,37 +4605,48 @@ 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)) {
> >>>>>>>> - /*
> >>>>>>>> - * 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.
> >>>>>>>> - */
> >>>>>>> Can't say I understand this patch yet, but it looks like the comment
> >>>>>>> is
> >>>>>>> still valid. Why have you removed it?
> >>>>>>
> >>>>>> L0 allows L1 or L2 at most to own TS, the rest is host-owned. I think
> >>>>>> the comment was always misleading.
> >>>>>>
> >>>>> I do not see how it is misleading. For everything but TS we will not get
> >>>>> here (if L1 is kvm). For TS we will get here if L1 allows L2 to change
> >>>>> it, but L0 does not.
> >>>>
> >>>> For everything *but guest-owned* we will get here, thus for most CR0
> >>>> accesses (bit-wise, not regarding frequency).
> >>>>
> >>> I do not see how. If bit is trapped by L1 we will not get here. We will
> >>> do vmexit to L1 instead. nested_vmx_exit_handled_cr() check this
> >>> condition.
> >>> I am not arguing about you code (didn't grok it yet), but the comment
> >>> still make sense to me.
> >>
> >> "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." That I can sign. But the rest about TS is just
> >> misleading as we trap _every_ change in L0 - except for TS under certain
> >> conditions. The old code was tested against TS only, that's what the
> >> comment witness.
> >>
> > TS is just an example of how we can get here with KVM on KVM. Obviously
> > other hypervisors may have different configuration. L2 may allow full
> > guest access to CR0 and then each CR0 write by L2 will be handled here.
> > Under what other condition "we trap _every_ change in L0 - except for
> > TS" here?
>
> On FPU activation:
>
> cr0_guest_owned_bits = X86_CR0_TS;
>
> And on FPU deactivation:
>
> cr0_guest_owned_bits = 0;
>
That's exactly TS case that comment explains. Note that
CR0_GUEST_HOST_MASK = ~cr0_guest_owned_bits.
> >
> >> If you prefer, I'll leave part one in.
> >>
> > Please do so. Without the comment it is not obvious why exit condition
> > is not checked here. Still do not see why you object to TS part.
>
> It describes a corner case in a way that suggests this is the only
> reason why we get here.
>
For KVM on KVM it is.
--
Gleb.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
