[QEMU,1.1,v2] Expose CPUID leaf 7 only for -cpu host

Changes v1 -> v2:
  - Use kvm_arch_get_supported_cpuid() instead of host_cpuid() on
    cpu_x86_fill_host().

  We should use GET_SUPPORTED_CPUID for all bits on "-cpu host"
  eventually, but I am not changing all the other CPUID leaves because
  we may not be able to test such an intrusive change in time for 1.1.

Description of the bug:

Since QEMU 0.15, the CPUID information on CPUID[EAX=7,ECX=0] is being
returned unfiltered to the guest, directly from the GET_SUPPORTED_CPUID
return value.

The problem is that this makes the resulting CPU feature flags
unpredictable and dependent on the host CPU and kernel version. This
breaks live-migration badly if migrating from a host CPU that supports
some features on that CPUID leaf (running a recent kenrel) to a kernel
or host CPU that doesn't support it.

Migration also is incorrect (the virtual CPU changes under the guest's
feet) if you migrate in the opposite direction (from an old CPU/kernel
to a new CPU/kernel), but with less serious consequences (guests
normally query CPUID information only once on boot).

Fortunately, the bug affects only users using cpudefs with level >= 7.

The right behavior should be to explicitly enable those features on
[cpudef] config sections or on the "-cpu" command-line arguments. Right
now there is no predefined CPU model on QEMU that has those features:
the latest Intel model we have is Sandy Bridge.

I would like to get this fixed on 1.1, so I am submitting this patch,
that enables those features only if "-cpu host" is being used (as we
don't have any pre-defined CPU model that actually have those features).
After 1.1 is released, we can make those features properly configurable
on [cpudef] and -cpu configuration.

One problem is: with this patch, users with the following setup:
- Running QEMU 1.0;
- Using a cpudef having level >= 7;
- Running a kernel that supports the features on CPUID leaf 7; and
- Running on a CPU that supports some features on CPUID leaf 7
won't be able to live-migrate to QEMU 1.1. But for these users
live-migration is already broken (they can't live-migrate to hosts with
older CPUs or older kernels, already), I don't see how to avoid this
problem.

Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
---
 target-i386/cpu.c |   22 +++++++++++++++-------
 target-i386/cpu.h |    2 ++
 2 files changed, 17 insertions(+), 7 deletions(-)

On 05/17/2012 10:26 AM, Eduardo Habkost wrote:

> The problem is that this makes the resulting CPU feature flags
> unpredictable and dependent on the host CPU and kernel version. This
> breaks live-migration badly if migrating from a host CPU that supports
> some features on that CPUID leaf (running a recent kenrel) to a kernel

s/kenrel/kernel/

On Thu, May 17, 2012 at 10:56:36AM -0600, Eric Blake wrote:
> On 05/17/2012 10:26 AM, Eduardo Habkost wrote:
> 
> > The problem is that this makes the resulting CPU feature flags
> > unpredictable and dependent on the host CPU and kernel version. This
> > breaks live-migration badly if migrating from a host CPU that supports
> > some features on that CPUID leaf (running a recent kenrel) to a kernel
> 
> s/kenrel/kernel/

Thanks for catching it. I guess this can be edited before applying the
patch, to avoid having to resubmit it.

On Thu, May 17, 2012 at 01:26:55PM -0300, Eduardo Habkost wrote:
[...]
> @@ -521,6 +523,12 @@ static int cpu_x86_fill_host(x86_def_t *x86_cpu_def)
>      x86_cpu_def->ext_features = ecx;
>      x86_cpu_def->features = edx;
>  
> +    if (x86_cpu_def->level >= 7) {
> +        x86_cpu_def->cpuid_7_0_ebx_features = kvm_arch_get_supported_cpuid(kvm_state, 0x7, 0, R_EBX);

This needs a check for kvm_enabled(). The code for "-cpu host" checks
for kvm_enabled() before calling cpu_x86_fill_host() already, but "-cpu
enforce" uses cpu_x86_fill_host() and doesn't check for kvm_enabled()
(that doesn't make much sense, but I won't try to fix 2 bugs in the same
patch). I will send v3.

Thanks to Gleb Natapov for catching it.