Re: [RFC PATCH 3/7] x86: Grant AMX permission for guest

[Paolo Bonzini]

On 1/7/22 10:31, Yang Zhong wrote:
> +static void x86_xsave_req_perm(void)
> +{
> +    unsigned long bitmask;
> +
> +    long rc = syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_GUEST_PERM,
> +                      XSTATE_XTILE_DATA_BIT);
> +    if (rc) {
> +        /*
> +         * The older kernel version(<5.15) can't support
> +         * ARCH_REQ_XCOMP_GUEST_PERM and directly return.
> +         */
> +        return;
> +    }
> +
> +    rc = syscall(SYS_arch_prctl, ARCH_GET_XCOMP_GUEST_PERM, &bitmask);
> +    if (rc) {
> +        error_report("prctl(ARCH_GET_XCOMP_GUEST_PERM) error: %ld", rc);
> +    } else if (!(bitmask & XFEATURE_XTILE_MASK)) {
> +        error_report("prctl(ARCH_REQ_XCOMP_GUEST_PERM) failure "
> +                     "and bitmask=0x%lx", bitmask);
> +        exit(EXIT_FAILURE);
> +    }
> +}
> +
>   void x86_cpus_init(X86MachineState *x86ms, int default_cpu_version)
>   {
>       int i;
> @@ -124,6 +150,8 @@ void x86_cpus_init(X86MachineState *x86ms, int 
> default_cpu_version)
>       MachineState *ms = MACHINE(x86ms);
>       MachineClass *mc = MACHINE_GET_CLASS(x86ms);
>   
> +    /* Request AMX pemission for guest */
> +    x86_xsave_req_perm();
>       x86_cpu_set_default_version(default_cpu_version);
>   

This should be done before creating a CPU with support for state 
component 18.  It happens in kvm_init_vcpu, with the following call stack:

        kvm_init_vcpu
        kvm_vcpu_thread_fn
        kvm_start_vcpu_thread
        qemu_init_vcpu
        x86_cpu_realizefn

The issue however is that this has to be done before 
KVM_GET_SUPPORTED_CPUID and KVM_CHECK_EXTENSION(KVM_CAP_XSAVE2).

For the former, you can assume that anything returned by 
ARCH_GET_XCOMP_GUEST_PERM will be returned by KVM_GET_SUPPORTED_CPUID in 
CPUID[0xD].EDX:EAX, so you can:

- add it to kvm_arch_get_supported_cpuid