Re: [Qemu-devel] [PATCH v3 5/6] target-i386: Don't enable nested VMX by default

[Eduardo Habkost]

On Wed, Oct 29, 2014 at 06:40:48PM +0100, Andreas Färber wrote:
> Am 03.10.2014 um 21:39 schrieb Eduardo Habkost:
> > TCG doesn't support VMX, and nested VMX is not enabled by default on the
> > KVM kernel module.
> > 
> > So, there's no reason to have VMX enabled by default on the core2duo and
> > coreduo CPU models, today. Even the newer Intel CPU model definitions
> > don't have it enabled.
> > 
> > In this case, we need machine-type compat code, as people may be running
> > the older machine-types on hosts that had VMX nesting enabled.
> > 
> > Signed-off-by: Eduardo Habkost <address@hidden>
> > ---
> >  hw/i386/pc_piix.c | 2 ++
> >  hw/i386/pc_q35.c  | 2 ++
> >  target-i386/cpu.c | 8 ++++----
> >  3 files changed, 8 insertions(+), 4 deletions(-)
> [...]
> > diff --git a/target-i386/cpu.c b/target-i386/cpu.c
> > index 1e9fff9..c336003 100644
> > --- a/target-i386/cpu.c
> > +++ b/target-i386/cpu.c
> > @@ -720,10 +720,10 @@ static X86CPUDefinition builtin_x86_defs[] = {
> >              CPUID_MTRR | CPUID_CLFLUSH | CPUID_MCA |
> >              CPUID_PSE36 | CPUID_VME | CPUID_ACPI | CPUID_SS,
> >          /* Missing: CPUID_EXT_DTES64, CPUID_EXT_DSCPL, CPUID_EXT_EST,
> > -         * CPUID_EXT_TM2, CPUID_EXT_XTPR, CPUID_EXT_PDCM */
> > +         * CPUID_EXT_TM2, CPUID_EXT_XTPR, CPUID_EXT_PDCM, CPUID_EXT_VMX */
> >          .features[FEAT_1_ECX] =
> >              CPUID_EXT_SSE3 | CPUID_EXT_MONITOR | CPUID_EXT_SSSE3 |
> > -            CPUID_EXT_VMX | CPUID_EXT_CX16,
> > +            CPUID_EXT_CX16,
> >          .features[FEAT_8000_0001_EDX] =
> >              CPUID_EXT2_LM | CPUID_EXT2_SYSCALL | CPUID_EXT2_NX,
> >          .features[FEAT_8000_0001_ECX] =
> [snip]
> 
> Here I'm less certain what the best approach is. As you point out,
> there's an inconsistency that I agree should be fixed. I wonder however
> whether an approach similar to 3/6 for KVM only would be better? I.e.,
> have VMX as a sometimes-KVM-supported feature be listed in the model and
> filter it out for accel=kvm so that -cpu enforce works, but let
> accel=tcg fail with features not implemented.

I don't mind either way, this is up for TCG maintainers. But I don't
understand what would be the benefit in keeping it enabled by default
for TCG.

Fortunately the answer for KVM and TCG are completely independent from
each other, now. We just need to follow this depending on the defaults
we choose:

  | Default on |
  |  TCG | KVM | Implementation
  +------+-----+-------------------------------------------------
A |  NO  | NO  | Unset it on CPU model table (builtin_x86_defs)
B |  NO  | YES | Unset on table, set on kvm_default_features
C |  YES | NO  | Set on table, set on kvm_default_unset_features
B |  YES | YES | Set it on CPU model table
  +------+-----+-------------------------------------------------

This patch implements row A.

If you tell me you really want VMX to be enabled by default on TCG mode,
then we can implement row C like we already did for CPUID_ACPI,
CPUID_EXT_MONITOR, and CPUID_EXT3_SVM.

But, are you really sure you do?

Considering accel=tcg behavior only (ignore KVM by now): why is this
different from all the features listed on patch 1/6? VMX was never
supported by TCG, so (like all the features from 1/6) it is not useful
to have it enabled by default when accel=tcg.

-- 
Eduardo