Re: [Xen-devel] [PATCH v2 09/11] x86/xen: use capabilities instead of fake cpuid values for xsave

[Andrew Cooper]

On 13/04/17 11:11, Juergen Gross wrote:
> @@ -281,22 +274,19 @@ static bool __init xen_check_mwait(void)
>  	return false;
>  #endif
>  }
> -static void __init xen_init_cpuid_mask(void)
> +
> +static bool __init xen_check_xsave(void)
>  {
> -	unsigned int ax, bx, cx, dx;
> -	unsigned int xsave_mask;
> +	unsigned int cx, xsave_mask;
>  
> -	ax = 1;
> -	cx = 0;
> -	cpuid(1, &ax, &bx, &cx, &dx);
> +	cx = cpuid_ecx(1);
>  
>  	xsave_mask =
>  		(1 << (X86_FEATURE_XSAVE % 32)) |
>  		(1 << (X86_FEATURE_OSXSAVE % 32));
>  
>  	/* Xen will set CR4.OSXSAVE if supported and not disabled by force */
> -	if ((cx & xsave_mask) != xsave_mask)
> -		cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
> +	return (cx & xsave_mask) == xsave_mask;

As you are taking the time to change this logic, could we see about
making it work in a way which doesn't require maintaining
non-architectural hackary in the hypervisor?

(For anyone reading this who isn't familiar with the history,
http://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=xen/arch/x86/cpuid.c;h=d359e090f33bb5793524186dc5f1c8bb620f17b3;hb=17cd6621688bce9972d9242611114fd7aba44c31#l726
ought to cover it)

In the presence of the bugs which caused this issue in the first place,
the best way to probe whether xsave is actually available is to try and
execute an xgetbv instruction, and use #UD as a signal that support
isn't actually available.

This also avoids Linux participating in the non-architectural behaviour
of forcing OXSAVE to be something other than a reflection of CR4.

~Andrew