Re: [PATCH] x86emul: recognize CLDEMOTE

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Tue, Jan 25, 2022 at 03:22:25PM +0100, Jan Beulich wrote:
> We claim to support the insn, but so far the emulator has been handling
> it as a NOP.

While not ideal, the SDM mentions that "The CLDEMOTE instruction may
be ignored by hardware in certain cases and is not a guarantee.".

> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> ---
> While handling x86emul_cldemote separately in hvmemul_cache_op() means
> to carry some redundant code, folding it with CLFLUSH{,OPT} / CLWB
> didn't seem very attractive either.
> 
> --- a/xen/arch/x86/arch.mk
> +++ b/xen/arch/x86/arch.mk
> @@ -23,6 +23,7 @@ $(call as-option-add,CFLAGS,CC,"xsaveopt
>  $(call as-option-add,CFLAGS,CC,"rdseed %eax",-DHAVE_AS_RDSEED)
>  $(call as-option-add,CFLAGS,CC,"clac",-DHAVE_AS_CLAC_STAC)
>  $(call as-option-add,CFLAGS,CC,"clwb (%rax)",-DHAVE_AS_CLWB)
> +$(call as-option-add,CFLAGS,CC,"cldemote (%rax)",-DHAVE_AS_CLDEMOTE)
>  $(call as-option-add,CFLAGS,CC,".equ \"x\"$$(comma)1",-DHAVE_AS_QUOTED_SYM)
>  $(call as-option-add,CFLAGS,CC,"invpcid 
> (%rax)$$(comma)%rax",-DHAVE_AS_INVPCID)
>  $(call as-option-add,CFLAGS,CC,"movdiri 
> %rax$$(comma)(%rax)",-DHAVE_AS_MOVDIR)
> --- a/xen/arch/x86/hvm/emulate.c
> +++ b/xen/arch/x86/hvm/emulate.c
> @@ -2351,6 +2351,28 @@ static int hvmemul_cache_op(
>           * to be sensibly used is in (virtualization unaware) firmware.
>           */
>          break;
> +
> +    case x86emul_cldemote:
> +        ASSERT(!is_x86_system_segment(seg));
> +
> +        if ( !boot_cpu_has(X86_FEATURE_CLDEMOTE) ||
> +             hvmemul_virtual_to_linear(seg, offset, 0, NULL, hvm_access_none,
> +                                       hvmemul_ctxt, &addr) != X86EMUL_OKAY )
> +            break;
> +
> +        if ( hvmemul_ctxt->seg_reg[x86_seg_ss].dpl == 3 )
> +            pfec |= PFEC_user_mode;
> +
> +        mapping = hvmemul_map_linear_addr(addr, 0, pfec, hvmemul_ctxt);

I think the emulator should map the address using the same cache
attributes as the guest, or else the result might be different than
intended?

Thanks, Roger.