Re: [PATCH v4 4/8] KVM: VMX: Handle MMIO Stale Data in VM-Enter assembly via ALTERNATIVES_2

[Pawan Gupta <pawan.kumar.gupta@linux.intel.com>]

On Mon, Nov 03, 2025 at 10:17:49AM +0100, Peter Zijlstra wrote:
> On Fri, Oct 31, 2025 at 04:55:24PM -0700, Pawan Gupta wrote:
> > On Thu, Oct 30, 2025 at 05:30:36PM -0700, Sean Christopherson wrote:
> > ...
> > > diff --git a/arch/x86/kvm/vmx/vmenter.S b/arch/x86/kvm/vmx/vmenter.S
> > > index 1f99a98a16a2..61a809790a58 100644
> > > --- a/arch/x86/kvm/vmx/vmenter.S
> > > +++ b/arch/x86/kvm/vmx/vmenter.S
> > > @@ -71,6 +71,7 @@
> > >   * @regs:	unsigned long * (to guest registers)
> > >   * @flags:	VMX_RUN_VMRESUME:	use VMRESUME instead of VMLAUNCH
> > >   *		VMX_RUN_SAVE_SPEC_CTRL: save guest SPEC_CTRL into vmx->spec_ctrl
> > > + *		VMX_RUN_CLEAR_CPU_BUFFERS_FOR_MMIO: vCPU can access host MMIO
> > >   *
> > >   * Returns:
> > >   *	0 on VM-Exit, 1 on VM-Fail
> > > @@ -137,6 +138,12 @@ SYM_FUNC_START(__vmx_vcpu_run)
> > >  	/* Load @regs to RAX. */
> > >  	mov (%_ASM_SP), %_ASM_AX
> > >  
> > > +	/* Stash "clear for MMIO" in EFLAGS.ZF (used below). */
> > > +	ALTERNATIVE_2 "",								\
> > > +		      __stringify(test $VMX_RUN_CLEAR_CPU_BUFFERS_FOR_MMIO, %ebx), 	\
> > > +		      X86_FEATURE_CLEAR_CPU_BUF_MMIO,					\
> > > +		      "", X86_FEATURE_CLEAR_CPU_BUF_VM
> > > +
> > >  	/* Check if vmlaunch or vmresume is needed */
> > >  	bt   $VMX_RUN_VMRESUME_SHIFT, %ebx
> > >  
> > > @@ -161,7 +168,12 @@ SYM_FUNC_START(__vmx_vcpu_run)
> > >  	mov VCPU_RAX(%_ASM_AX), %_ASM_AX
> > >  
> > >  	/* Clobbers EFLAGS.ZF */
> > > -	VM_CLEAR_CPU_BUFFERS
> > > +	ALTERNATIVE_2 "",							\
> > > +		      __stringify(jz .Lskip_clear_cpu_buffers;			\
> > > +				  CLEAR_CPU_BUFFERS_SEQ;			\
> > > +				  .Lskip_clear_cpu_buffers:),			\
> > > +		      X86_FEATURE_CLEAR_CPU_BUF_MMIO,				\
> > > +		      __CLEAR_CPU_BUFFERS, X86_FEATURE_CLEAR_CPU_BUF_VM
> > 
> > Another way to write this could be:
> > 
> > 	ALTERNATIVE_2 "jmp .Lskip_clear_cpu_buffers",					\
> > 		      "jz  .Lskip_clear_cpu_buffers", X86_FEATURE_CLEAR_CPU_BUF_MMIO,	\
> > 		      "",			      X86_FEATURE_CLEAR_CPU_BUF_VM
> > 
> > 	CLEAR_CPU_BUFFERS_SEQ
> > .Lskip_clear_cpu_buffers:
> > 
> > With this jmp;verw; would show up in the disassembly on unaffected CPUs, I
> > don't know how big a problem is that. OTOH, I find this easier to understand.
> 
> Generating larger code just to keep disassembly 'simple' seems wrong.
> Also, see this:
> 
>   https://lkml.kernel.org/r/194ad779-f41f-46a5-9973-e886f483b60a@oracle.com

Ok thanks, that settles it.