Re: [PATCH 1/6] KVM: nVMX: delay loading of PDPTRs to KVM_REQ_GET_NESTED_STATE_PAGES

[Maxim Levitsky <mlevitsk@redhat.com>]

On Fri, 2021-04-02 at 17:27 +0000, Sean Christopherson wrote:
> On Thu, Apr 01, 2021, Maxim Levitsky wrote:
> > Similar to the rest of guest page accesses after migration,
> > this should be delayed to KVM_REQ_GET_NESTED_STATE_PAGES
> > request.
> 
> FWIW, I still object to this approach, and this patch has a plethora of issues.
> 
> I'm not against deferring various state loading to KVM_RUN, but wholesale moving
> all of GUEST_CR3 processing without in-depth consideration of all the side
> effects is a really bad idea.
It could be, I won't argue about this.

> 
> > Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
> > ---
> >  arch/x86/kvm/vmx/nested.c | 14 +++++++++-----
> >  1 file changed, 9 insertions(+), 5 deletions(-)
> > 
> > diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> > index fd334e4aa6db..b44f1f6b68db 100644
> > --- a/arch/x86/kvm/vmx/nested.c
> > +++ b/arch/x86/kvm/vmx/nested.c
> > @@ -2564,11 +2564,6 @@ static int prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
> >  		return -EINVAL;
> >  	}
> >  
> > -	/* Shadow page tables on either EPT or shadow page tables. */
> > -	if (nested_vmx_load_cr3(vcpu, vmcs12->guest_cr3, nested_cpu_has_ept(vmcs12),
> > -				entry_failure_code))
> > -		return -EINVAL;
> > -
> >  	/*
> >  	 * Immediately write vmcs02.GUEST_CR3.  It will be propagated to vmcs12
> >  	 * on nested VM-Exit, which can occur without actually running L2 and
> > @@ -3109,11 +3104,16 @@ static bool nested_get_evmcs_page(struct kvm_vcpu *vcpu)
> >  static bool nested_get_vmcs12_pages(struct kvm_vcpu *vcpu)
> >  {
> >  	struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
> > +	enum vm_entry_failure_code entry_failure_code;
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> >  	struct kvm_host_map *map;
> >  	struct page *page;
> >  	u64 hpa;
> >  
> > +	if (nested_vmx_load_cr3(vcpu, vmcs12->guest_cr3, nested_cpu_has_ept(vmcs12),
> > +				&entry_failure_code))
> 
> This results in KVM_RUN returning 0 without filling vcpu->run->exit_reason.
> Speaking from experience, debugging those types of issues is beyond painful.
> 
> It also means CR3 is double loaded in the from_vmentry case.
> 
> And it will cause KVM to incorrectly return NVMX_VMENTRY_KVM_INTERNAL_ERROR
> if a consistency check fails when nested_get_vmcs12_pages() is called on
> from_vmentry.  E.g. run unit tests with this and it will silently disappear.

I do remember now that you said something about this, but I wasn't able
to find it in my email. Sorry about this.
I agree with you.

I think that a question I should ask is why do we really need to 
delay accessing guest memory after a migration.

So far I mostly just assumed that we need to do so, thinking that qemu
updates the memslots or something, or maybe because guest memory
isn't fully migrated and relies on post-copy to finish it.

Also I am not against leaving CR3 processing in here and doing only PDPTR load
in KVM_RUN (and only when *SREG2 API is not used).

> 
> diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c
> index bbb006a..b8ccc69 100644
> --- a/x86/vmx_tests.c
> +++ b/x86/vmx_tests.c
> @@ -8172,6 +8172,16 @@ static void test_guest_segment_base_addr_fields(void)
>         vmcs_write(GUEST_AR_ES, ar_saved);
>  }
> 
> +static void test_guest_cr3(void)
> +{
> +       u64 cr3_saved = vmcs_read(GUEST_CR3);
> +
> +       vmcs_write(GUEST_CR3, -1ull);
> +       test_guest_state("Bad CR3 fails VM-Enter", true, -1ull, "GUEST_CR3");
> +
> +       vmcs_write(GUEST_DR7, cr3_saved);
> +}
> +
Could you send this test to kvm unit tests?

>  /*
>   * Check that the virtual CPU checks the VMX Guest State Area as
>   * documented in the Intel SDM.
> @@ -8181,6 +8191,8 @@ static void vmx_guest_state_area_test(void)
>         vmx_set_test_stage(1);
>         test_set_guest(guest_state_test_main);
> 
> +       test_guest_cr3();
> +
>         /*
>          * The IA32_SYSENTER_ESP field and the IA32_SYSENTER_EIP field
>          * must each contain a canonical address.
> 
> 
> > +		return false;
> > +
> >  	if (nested_cpu_has2(vmcs12, SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES)) {
> >  		/*
> >  		 * Translate L1 physical address to host physical
> > @@ -3357,6 +3357,10 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu,
> >  	}
> >  
> >  	if (from_vmentry) {
> > +		if (nested_vmx_load_cr3(vcpu, vmcs12->guest_cr3,
> > +		    nested_cpu_has_ept(vmcs12), &entry_failure_code))
> 
> This alignment is messed up; it looks like two separate function calls.
Sorry about this, I see it now.
> 
> > +			goto vmentry_fail_vmexit_guest_mode;
> > +
> >  		failed_index = nested_vmx_load_msr(vcpu,
> >  						   vmcs12->vm_entry_msr_load_addr,
> >  						   vmcs12->vm_entry_msr_load_count);
> > -- 
> > 2.26.2
> > 


Best regards,
	Maxim Levitsky