Re: [PATCH] KVM: VMX: Micro-optimize SPEC_CTRL handling in __vmx_vcpu_run()

[Uros Bizjak]

On Tue, Aug 19, 2025 at 5:00 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Thu, Aug 07, 2025, Uros Bizjak wrote:
> > Use memory operand in CMP instruction to avoid usage of a
> > temporary register. Use %eax register to hold VMX_spec_ctrl
> > and use it directly in the follow-up WRMSR.
> >
> > The new code saves a few bytes by removing two MOV insns, from:
> >
> >   2d: 48 8b 7c 24 10          mov    0x10(%rsp),%rdi
> >   32: 8b bf 48 18 00 00       mov    0x1848(%rdi),%edi
> >   38: 65 8b 35 00 00 00 00    mov    %gs:0x0(%rip),%esi
> >   3f: 39 fe                   cmp    %edi,%esi
> >   41: 74 0b                   je     4e <...>
> >   43: b9 48 00 00 00          mov    $0x48,%ecx
> >   48: 31 d2                   xor    %edx,%edx
> >   4a: 89 f8                   mov    %edi,%eax
> >   4c: 0f 30                   wrmsr
> >
> > to:
> >
> >   2d: 48 8b 7c 24 10          mov    0x10(%rsp),%rdi
> >   32: 8b 87 48 18 00 00       mov    0x1848(%rdi),%eax
> >   38: 65 3b 05 00 00 00 00    cmp    %gs:0x0(%rip),%eax
> >   3f: 74 09                   je     4a <...>
> >   41: b9 48 00 00 00          mov    $0x48,%ecx
> >   46: 31 d2                   xor    %edx,%edx
> >   48: 0f 30                   wrmsr
> >
> > No functional change intended.
> >
> > Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> > Cc: Sean Christopherson <seanjc@xxxxxxxxxx>
> > Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > Cc: Borislav Petkov <bp@xxxxxxxxx>
> > Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> > Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> > ---
> >  arch/x86/kvm/vmx/vmenter.S | 6 ++----
> >  1 file changed, 2 insertions(+), 4 deletions(-)
> >
> > diff --git a/arch/x86/kvm/vmx/vmenter.S b/arch/x86/kvm/vmx/vmenter.S
> > index 0a6cf5bff2aa..c65de5de92ab 100644
> > --- a/arch/x86/kvm/vmx/vmenter.S
> > +++ b/arch/x86/kvm/vmx/vmenter.S
> > @@ -118,13 +118,11 @@ SYM_FUNC_START(__vmx_vcpu_run)
> >        * and vmentry.
> >        */
> >       mov 2*WORD_SIZE(%_ASM_SP), %_ASM_DI
> > -     movl VMX_spec_ctrl(%_ASM_DI), %edi
> > -     movl PER_CPU_VAR(x86_spec_ctrl_current), %esi
> > -     cmp %edi, %esi
> > +     movl VMX_spec_ctrl(%_ASM_DI), %eax
> > +     cmp PER_CPU_VAR(x86_spec_ctrl_current), %eax
>
> Huh.  There's a pre-existing bug lurking here, and in the SVM code.  SPEC_CTRL
> is an MSR, i.e. a 64-bit value, but the assembly code assumes bits 63:32 are always
> zero.

But MSBs are zero, MSR is defined in arch/x86/include/msr-index.h as:

#define MSR_IA32_SPEC_CTRL 0x00000048 /* Speculation Control */

and "movl $..., %eax" zero-extends the value to full 64-bit width.

FWIW, MSR_IA32_SPEC_CTR is handled in the same way in arch/x86/entry/entry.S:

movl $MSR_IA32_PRED_CMD, %ecx

So, the insn is OK when MSR fits 32-bits. The insn is a bit smaller, too:

       movl $0x01, %ecx
       movq $0x01, %rcx

assembles to:

  0:   b9 01 00 00 00          mov    $0x1,%ecx
  5:   48 c7 c1 01 00 00 00    mov    $0x1,%rcx

Rest assured that the assembler checks the immediate when 32-bit
registers are involved, e.g.:

mov.s: Assembler messages:
mov.s:1: Warning: 0xaaaaaaaaaa shortened to 0xaaaaaaaa

Uros.