Re: thread/core siblings info for guests

[Avi Kivity <avi@redhat.com>]

Nitin A Kamble wrote:
> Hi Avi,
>  Attached is a patch to handle the "host mode" as you mentioned earlier.
>
> It works this way:
>    Qemu's -cpu option is extended to take a new paramenter: -cpu _host_.
> With that parameter the cpuid bits are collected from the host cpu. And
> few bits of cleared to get it working with the kvm.
>    I tested the patch with Windows & Linux guests, and I see the host
> topology is propagating into the guest. With that windows server is not
> imposing the incorrect licensing restriction to guest, as it was doing
> earlier. With the patch I could bring up the windows server guest with
> 16 cpus on kvm.
>   I also fixed implementation of indexed cpuid leaves (0x4 & 0xb) in
> qemu. And also some code cleanup are there for consistent terminology to
> have more readable code, 
>
>   Please check-in the patch in kvm-usermode.git tree. There are no
> changes for the kernel side or kvm.git tree for this.
>
> Signed-Off-By: Nitin A Kamble <nitin.a.kamble@intel.com>
>
> Next things:
>  * Default mode: I am not planning to do anything for this. I see the
> current situation is good enough for the default mode.
>
>  * Managed mode: For this, I am thinking few things. 
>     1. Either an option can be added to qemu to take (almost) all of 
>        it's cpuid bits from a file.
>     2. Or use command line options to modify particular bits of guest  
>        cpuid.
>     3. Or do the both.
>
> What do you think about the managed mode implementation? you probably
> have better idea regarding how and what kind of tools will be managing
> it.
>
>   

> +static int fill_x86_defs_for_host(CPUX86State *env, x86_def_t * def)
> +{
> +    uint32 eax, ebx, ecx, edx;
> +
> +    env->cpu_host_cpu = 1;
> +
> +    host_cpuid(0, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_level = eax;
> +    env->cpuid_vendor1 = ebx;
> +    env->cpuid_vendor2 = ecx;
> +    env->cpuid_vendor3 = edx;
> +
> +    host_cpuid(1, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_version = eax;
> +    env->cpuid_features = edx & 0xdfffffff; /* disable Thermal Monitro */
> +    env->cpuid_ext_features = ecx;
> +
> +    host_cpuid(0x80000000, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_xlevel = eax;
> +
> +    host_cpuid(0x80000001, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_ext3_features = ecx;
> +    env->cpuid_ext2_features = edx & 0xf7ffffff; /* disable rdtscp */
> +
> +    host_cpuid(0x80000002, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_model[0] = eax;
> +    env->cpuid_model[1] = ebx;
> +    env->cpuid_model[2] = ecx;
> +    env->cpuid_model[3] = edx;
> +
> +    host_cpuid(0x80000003, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_model[4] = eax;
> +    env->cpuid_model[5] = ebx;
> +    env->cpuid_model[6] = ecx;
> +    env->cpuid_model[7] = edx;
> +
> +    host_cpuid(0x80000004, 0, &eax, &ebx, &ecx, &edx);
> +    env->cpuid_model[8] = eax;
> +    env->cpuid_model[9] = ebx;
> +    env->cpuid_model[10] = ecx;
> +    env->cpuid_model[11] = edx;
> +
> +    return 0;
> +}
> +

There is already a kernel function for this (KVM_GET_SUPPORTED_CPUID, 
IIRC) that does this.  It has the advantage of knowing which bits the 
kernel supports, so there's no need to keep the kernel and userspace at 
the same version.

-- 
error compiling committee.c: too many arguments to function