Re: [PATCH 1/4] KVM: nVMX: Always make an attempt to map eVMCS after migration

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Maxim Levitsky <mlevitsk@redhat.com> writes:

> On Wed, 2021-05-05 at 10:39 +0200, Vitaly Kuznetsov wrote:
>> Maxim Levitsky <mlevitsk@redhat.com> writes:
>> 
>> > On Mon, 2021-05-03 at 17:08 +0200, Vitaly Kuznetsov wrote:
>> > > When enlightened VMCS is in use and nested state is migrated with
>> > > vmx_get_nested_state()/vmx_set_nested_state() KVM can't map evmcs
>> > > page right away: evmcs gpa is not 'struct kvm_vmx_nested_state_hdr'
>> > > and we can't read it from VP assist page because userspace may decide
>> > > to restore HV_X64_MSR_VP_ASSIST_PAGE after restoring nested state
>> > > (and QEMU, for example, does exactly that). To make sure eVMCS is
>> > > mapped /vmx_set_nested_state() raises KVM_REQ_GET_NESTED_STATE_PAGES
>> > > request.
>> > > 
>> > > Commit f2c7ef3ba955 ("KVM: nSVM: cancel KVM_REQ_GET_NESTED_STATE_PAGES
>> > > on nested vmexit") added KVM_REQ_GET_NESTED_STATE_PAGES clearing to
>> > > nested_vmx_vmexit() to make sure MSR permission bitmap is not switched
>> > > when an immediate exit from L2 to L1 happens right after migration (caused
>> > > by a pending event, for example). Unfortunately, in the exact same
>> > > situation we still need to have eVMCS mapped so
>> > > nested_sync_vmcs12_to_shadow() reflects changes in VMCS12 to eVMCS.
>> > > 
>> > > As a band-aid, restore nested_get_evmcs_page() when clearing
>> > > KVM_REQ_GET_NESTED_STATE_PAGES in nested_vmx_vmexit(). The 'fix' is far
>> > > from being ideal as we can't easily propagate possible failures and even if
>> > > we could, this is most likely already too late to do so. The whole
>> > > 'KVM_REQ_GET_NESTED_STATE_PAGES' idea for mapping eVMCS after migration
>> > > seems to be fragile as we diverge too much from the 'native' path when
>> > > vmptr loading happens on vmx_set_nested_state().
>> > > 
>> > > Fixes: f2c7ef3ba955 ("KVM: nSVM: cancel KVM_REQ_GET_NESTED_STATE_PAGES on nested vmexit")
>> > > Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
>> > > ---
>> > >  arch/x86/kvm/vmx/nested.c | 29 +++++++++++++++++++----------
>> > >  1 file changed, 19 insertions(+), 10 deletions(-)
>> > > 
>> > > diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
>> > > index 1e069aac7410..2febb1dd68e8 100644
>> > > --- a/arch/x86/kvm/vmx/nested.c
>> > > +++ b/arch/x86/kvm/vmx/nested.c
>> > > @@ -3098,15 +3098,8 @@ static bool nested_get_evmcs_page(struct kvm_vcpu *vcpu)
>> > >  			nested_vmx_handle_enlightened_vmptrld(vcpu, false);
>> > >  
>> > >  		if (evmptrld_status == EVMPTRLD_VMFAIL ||
>> > > -		    evmptrld_status == EVMPTRLD_ERROR) {
>> > > -			pr_debug_ratelimited("%s: enlightened vmptrld failed\n",
>> > > -					     __func__);
>> > > -			vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
>> > > -			vcpu->run->internal.suberror =
>> > > -				KVM_INTERNAL_ERROR_EMULATION;
>> > > -			vcpu->run->internal.ndata = 0;
>> > > +		    evmptrld_status == EVMPTRLD_ERROR)
>> > >  			return false;
>> > > -		}
>> > >  	}
>> > >  
>> > >  	return true;
>> > > @@ -3194,8 +3187,16 @@ static bool nested_get_vmcs12_pages(struct kvm_vcpu *vcpu)
>> > >  
>> > >  static bool vmx_get_nested_state_pages(struct kvm_vcpu *vcpu)
>> > >  {
>> > > -	if (!nested_get_evmcs_page(vcpu))
>> > > +	if (!nested_get_evmcs_page(vcpu)) {
>> > > +		pr_debug_ratelimited("%s: enlightened vmptrld failed\n",
>> > > +				     __func__);
>> > > +		vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
>> > > +		vcpu->run->internal.suberror =
>> > > +			KVM_INTERNAL_ERROR_EMULATION;
>> > > +		vcpu->run->internal.ndata = 0;
>> > > +
>> > >  		return false;
>> > > +	}
>> > 
>> > Hi!
>> > 
>> > Any reason to move the debug prints out of nested_get_evmcs_page?
>> > 
>> 
>> Debug print could've probably stayed or could've been dropped
>> completely -- I don't really believe it's going to help
>> anyone. Debugging such issues without instrumentation/tracing seems to
>> be hard-to-impossible...
>> 
>> > >  
>> > >  	if (is_guest_mode(vcpu) && !nested_get_vmcs12_pages(vcpu))
>> > >  		return false;
>> > > @@ -4422,7 +4423,15 @@ void nested_vmx_vmexit(struct kvm_vcpu *vcpu, u32 vm_exit_reason,
>> > >  	/* trying to cancel vmlaunch/vmresume is a bug */
>> > >  	WARN_ON_ONCE(vmx->nested.nested_run_pending);
>> > >  
>> > > -	kvm_clear_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
>> > > +	if (kvm_check_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu)) {
>> > > +		/*
>> > > +		 * KVM_REQ_GET_NESTED_STATE_PAGES is also used to map
>> > > +		 * Enlightened VMCS after migration and we still need to
>> > > +		 * do that when something is forcing L2->L1 exit prior to
>> > > +		 * the first L2 run.
>> > > +		 */
>> > > +		(void)nested_get_evmcs_page(vcpu);
>> > > +	}
>> > Yes this is a band-aid, but it has to be done I agree.
>> > 
>> 
>> To restore the status quo, yes.
>> 
>> > >  
>> > >  	/* Service the TLB flush request for L2 before switching to L1. */
>> > >  	if (kvm_check_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu))
>> > 
>> > 
>> > 
>> > I also tested this and it survives a bit better (used to crash instantly
>> > after a single migration cycle, but the guest still crashes after around ~20 iterations of my 
>> > regular nested migration test).
>> > 
>> > Blues screen shows that stop code is HYPERVISOR ERROR and nothing else.
>> > 
>> > I tested both this patch alone and all 4 patches.
>> > 
>> > Without evmcs, the same VM with same host kernel and qemu survived an overnight
>> > test and passed about 1800 migration iterations.
>> > (my synthetic migration test doesn't yet work on Intel, I need to investigate why)
>> > 
>> 
>> It would be great to compare on Intel to be 100% sure the issue is eVMCS
>> related, Hyper-V may be behaving quite differently on AMD.
> Hi!
>
> I tested this on my Intel machine with and without eVMCS, without changing
> any other parameters, running the same VM from a snapshot.
>
> As I said without eVMCS the test survived overnight stress of ~1800 migrations.
> With eVMCs, it fails pretty much on first try. 
> With those patches, it fails after about 20 iterations.
>

Ah, sorry, misunderstood your 'synthetic migration test doesn't yet work
on Intel' :-) 

-- 
Vitaly