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

[Xiaoyao Li]

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:
> > > On Mon, Feb 24, 2020 at 02:04:46PM +0100, Vitaly Kuznetsov wrote:
> > > > Xiaoyao Li <xiaoyao.li@xxxxxxxxx> writes:
> > > >
> > > > > On 2/24/2020 6:16 PM, Vitaly Kuznetsov wrote:
> > > > > > Xiaoyao Li <xiaoyao.li@xxxxxxxxx> writes:
> > > > > Here variable basic_exit_reason is added for the cases where only basic
> > > > > exit reason number is needed.
> > > >
> > > > Can we do the other way around, i.e. introduce 'extended_exit_reason'
> > > > and use it where all 32 bits are needed? I'm fine with the change, just
> > > > trying to minimize the (unneeded) code churn.
> > >
> > > 100% agree.  Even better than adding a second field to vcpu_vmx would be
> > > to make it a union, though we'd probably want to call it something like
> > > full_exit_reason in that case.  That should give us compile-time checks on
> > > exit_reason, e.g. if we try to query one of the upper bits using a u16, e.g.
> >
> > 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.

> > Maybe we name it
> >
> > union {
> > 	u16 basic_exit_reason;
> > 	u32 exit_reason;
> > }
> >
> > as what SDM defines?
> >
> > > --- a/arch/x86/kvm/vmx/vmx.c
> > > +++ b/arch/x86/kvm/vmx/vmx.c
> > > @@ -5818,7 +5818,7 @@ static int vmx_handle_exit(struct kvm_vcpu *vcpu,
> > >          if (is_guest_mode(vcpu) && nested_vmx_exit_reflected(vcpu, exit_reason))
> > >                  return nested_vmx_reflect_vmexit(vcpu, exit_reason);
> > >
> > > -       if (exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY) {
> > > +       if (vmx->full_exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY) {
>
> If we do go the union route, this snippet of code is insufficient, the
> full/extended exit reason needs to be snapshotted early for use in the
> tracepoint and in fail_entry.hardware_entry_failure_reason.
>
> > >                  dump_vmcs();
> > >                  vcpu->run->exit_reason = KVM_EXIT_FAIL_ENTRY;
> > >                  vcpu->run->fail_entry.hardware_entry_failure_reason
> > > @@ -6620,11 +6620,12 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
> > >          vmx->nested.nested_run_pending = 0;
> > >          vmx->idt_vectoring_info = 0;
> > >
> > > -       vmx->exit_reason = vmx->fail ? 0xdead : vmcs_read32(VM_EXIT_REASON);
> > > -       if ((u16)vmx->exit_reason == EXIT_REASON_MCE_DURING_VMENTRY)
> > > +       vmx->full_exit_reason = vmx->fail ? 0xdead : vmcs_read32(VM_EXIT_REASON);
> > > +       if (vmx->exit_reason == EXIT_REASON_MCE_DURING_VMENTRY)
> > >                  kvm_machine_check();
> > >
> > > -       if (vmx->fail || (vmx->exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
> > > +       if (vmx->fail ||
> > > +           (vmx->full_exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
> > >                  return;
> > >
> > >          vmx->loaded_vmcs->launched = 1;
> > > diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
> > > index 7f42cf3dcd70..60c09640ea59 100644
> > > --- a/arch/x86/kvm/vmx/vmx.h
> > > +++ b/arch/x86/kvm/vmx/vmx.h
> > > @@ -260,7 +260,10 @@ struct vcpu_vmx {
> > >          int vpid;
> > >          bool emulation_required;
> > >
> > > -       u32 exit_reason;
> > > +       union {
> > > +               u16 exit_reason;
> > > +               u32 full_exit_reason;
> > > +       }
> > >
> > >          /* Posted interrupt descriptor */
> > >          struct pi_desc pi_desc;
> > >
> > >
> > >
> > >
> > >
> > > > --
> > > > Vitaly