Re: [PATCH 01/10] KVM: nVMX: Move reflection check into nested_vmx_reflect_vmexit()

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Sean Christopherson <sean.j.christopherson@intel.com> writes:

> On Mon, Mar 16, 2020 at 10:33:27PM -0700, Sean Christopherson wrote:
>> On Fri, Mar 13, 2020 at 01:12:33PM +0100, Vitaly Kuznetsov wrote:
>> > Sean Christopherson <sean.j.christopherson@intel.com> writes:
>> > 
>> > > -static inline int nested_vmx_reflect_vmexit(struct kvm_vcpu *vcpu,
>> > > -					    u32 exit_reason)
>> > > +static inline bool nested_vmx_reflect_vmexit(struct kvm_vcpu *vcpu,
>> > > +					     u32 exit_reason)
>> > >  {
>> > > -	u32 exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
>> > > +	u32 exit_intr_info;
>> > > +
>> > > +	if (!nested_vmx_exit_reflected(vcpu, exit_reason))
>> > > +		return false;
>> > 
>> > (unrelated to your patch)
>> > 
>> > It's probably just me but 'nested_vmx_exit_reflected()' name always
>> > makes me thinkg 'the vmexit WAS [already] reflected' and not 'the vmexit
>> > NEEDS to be reflected'. 'nested_vmx_exit_needs_reflecting()' maybe?
>> 
>> Not just you.  It'd be nice if the name some how reflected (ha) that the
>> logic is mostly based on whether or not L1 expects the exit, with a few
>> exceptions.  E.g. something like
>> 
>> 	if (!l1_expects_vmexit(...) && !is_system_vmexit(...))
>> 		return false;
>
> Doh, the system VM-Exit logic is backwards, it should be
>
> 	if (!l1_expects_vmexit(...) || is_system_vmexit(...))
> 		return false;
>> 
>> The downside of that is the logic is split, which is probably a net loss?
>

Yea,

(just thinking out loud below)

the problem with the split is that we'll have to handle the same exit
reason twice, e.g. EXIT_REASON_EXCEPTION_NMI (is_nmi() check goes to
is_system_vmexit() and vmcs12->exception_bitmap check goes to
l1_expects_vmexit()). Also, we have two 'special' cases: vmx->fail and
nested_run_pending. While the former belongs to to l1_expects_vmexit(),
the later doesn't belong to either (but we can move it to
nested_vmx_reflect_vmexit() I believe).

On the other hand, I'm a great fan of splitting checkers ('pure'
functions) from actors (functions with 'side-effects') and
nested_vmx_exit_reflected() while looking like a checker does a lot of
'acting': nested_mark_vmcs12_pages_dirty(), trace printk.

-- 
Vitaly