Re: [PATCH 1/2] kvm: vmx: Use basic exit reason to check if it's the specific VM EXIT

[Xiaoyao Li]

On 2/27/2020 7:59 AM, Sean Christopherson wrote:
> On Tue, Feb 25, 2020 at 02:41:20PM +0800, Xiaoyao Li wrote:
> > On 2/25/2020 2:13 PM, Sean Christopherson wrote:
> > > On Tue, Feb 25, 2020 at 08:13:15AM +0800, Xiaoyao Li wrote:
> > > > On 2/25/2020 12:17 AM, Sean Christopherson wrote:
> > > > I have thought about union, but it seems
> > > >
> > > > union {
> > > > 	u16 exit_reason;
> > > > 	u32 full_exit_reason;
> > > > }
> > > >
> > > > is not a good name. Since there are many codes in vmx.c and nested.c assume
> > > > that exit_reason stands for 32-bit EXIT REASON vmcs field as well as
> > > > evmcs->vm_exit_reason and vmcs12->vm_exit_reason. Do we really want to also
> > > > rename them to full_exit_reason?
> > >
> > > It's actually the opposite, almost all of the VMX code assumes exit_reason
> > > holds only the basic exit reason, i.e. a 16-bit value.  For example, SGX
> > > adds a modifier flag to denote a VM-Exit was from enclave mode, and that
> > > bit needs to be stripped from exit_reason, otherwise all the checks like
> > > "if (exit_reason == blah_blah_blah)" fail.
> > >
> > > Making exit_reason a 16-bit alias of the full/extended exit_reason neatly
> > > sidesteps that issue.  And it is an issue that has caused actual problems
> > > in the past, e.g. see commit beb8d93b3e42 ("KVM: VMX: Fix handling of #MC
> > > that occurs during VM-Entry").  Coincidentally, that commit also removes a
> > > local "u16 basic_exit_reason" :-).
> > >
> > > Except for one mistake, the pseudo-patch below is the entirety of required
> > > changes.  Most (all?) of the functions that take "u32 exit_reason" can (and
> > > should) continue to take a u32.
> > >
> > > As for the name, I strongly prefer keeping the exit_reason name for the
> > > basic exit reason.  The vast majority of VM-Exits do not have modifiers
> > > set, i.e. "basic exit reason" == vmcs.EXIT_REASON for nearly all normal
> > > usage.  This holds true in every form of communication, e.g. when discussing
> > > VM-Exit reasons, it's never qualified with "basic", it's simply the exit
> > > reason.  IMO the code is better off following the colloquial usage of "exit
> > > reason".  A simple comment above the union would suffice to clear up any
> > > confusion with respect to the SDM.
> >
> > Well, for this reason we can keep exit_reason for 16-bit usage, and define
> > full/extended_exit_reason for 32-bit cases. This makes less code churn.
> >
> > But after we choose to use exit_reason and full/extended_exit_reason, what
> > if someday new modifier flags are added and we want to enable some modifier
> > flags for nested case?
> > I guess we need to change existing exit_reason to full/extended_exit_reason
> > in nested.c/nested.h to keep the naming rule consistent.
>
> Ah, good point.  But, that's just another bug in my psuedo patch :-)
> It's literally one call site that needs to be updated.  E.g.
>
> 	if (is_guest_mode(vcpu) && nested_vmx_exit_reflected(vcpu, exit_reason))
> 		return nested_vmx_reflect_vmexit(vcpu, full_exit_reason);

shouldn't we also pass full_exit_reason to nested_vmx_exit_reflected()?

> Everywhere else KVM calls nested_vmx_reflect_vmexit() is (currently) done

I guess you wanted to say nested_vmx_vmexit() not 
nested_vmx_reflect_vmexit() here.

> with a hardcoded value (except handle_vmfunc(), but I actually want to
> change that one).