Re: [PATCH v4 2/5] docs/about/deprecated: Deprecate the qemu-system-i386 binary

[Daniel P. Berrangé <berrange@xxxxxxxxxx>]

On Mon, Mar 06, 2023 at 10:54:15AM +0100, Thomas Huth wrote:
> On 06/03/2023 10.27, Daniel P. Berrangé wrote:
> > On Mon, Mar 06, 2023 at 09:46:55AM +0100, Thomas Huth wrote:
> > > Aside from not supporting KVM on 32-bit hosts, the qemu-system-x86_64
> > > binary is a proper superset of the qemu-system-i386 binary. With the
> > > 32-bit host support being deprecated, it is now also possible to
> > > deprecate the qemu-system-i386 binary.
> > > 
> > > With regards to 32-bit KVM support in the x86 Linux kernel,
> > > the developers confirmed that they do not need a recent
> > > qemu-system-i386 binary here:
> > > 
> > >   https://lore.kernel.org/kvm/Y%2ffkTs5ajFy0hP1U@xxxxxxxxxx/
> > > 
> > > Reviewed-by: Daniel P. Berrangé <berrange@xxxxxxxxxx>
> > > Reviewed-by: Wilfred Mallawa <wilfred.mallawa@xxxxxxx>
> > > Signed-off-by: Thomas Huth <thuth@xxxxxxxxxx>
> > > ---
> > >   docs/about/deprecated.rst | 14 ++++++++++++++
> > >   1 file changed, 14 insertions(+)
> > > 
> > > diff --git a/docs/about/deprecated.rst b/docs/about/deprecated.rst
> > > index 1ca9dc33d6..c4fcc6b33c 100644
> > > --- a/docs/about/deprecated.rst
> > > +++ b/docs/about/deprecated.rst
> > > @@ -34,6 +34,20 @@ deprecating the build option and no longer defend it 
> > > in CI. The
> > >   ``--enable-gcov`` build option remains for analysis test case
> > >   coverage.
> > > +``qemu-system-i386`` binary (since 8.0)
> > > +'''''''''''''''''''''''''''''''''''''''
> > > +
> > > +The ``qemu-system-i386`` binary was mainly useful for running with KVM
> > > +on 32-bit x86 hosts, but most Linux distributions already removed their
> > > +support for 32-bit x86 kernels, so hardly anybody still needs this. The
> > > +``qemu-system-x86_64`` binary is a proper superset and can be used to
> > > +run 32-bit guests by selecting a 32-bit CPU model, including KVM support
> > > +on x86_64 hosts. Thus users are recommended to reconfigure their systems
> > > +to use the ``qemu-system-x86_64`` binary instead. If a 32-bit CPU guest
> > > +environment should be enforced, you can switch off the "long mode" CPU
> > > +flag, e.g. with ``-cpu max,lm=off``.
> > 
> > I had the idea to check this today and this is not quite sufficient,
> > because we have code that changes the family/model/stepping for
> > 'max' which is target dependent:
> > 
> > #ifdef TARGET_X86_64
> >      object_property_set_int(OBJECT(cpu), "family", 15, &error_abort);
> >      object_property_set_int(OBJECT(cpu), "model", 107, &error_abort);
> >      object_property_set_int(OBJECT(cpu), "stepping", 1, &error_abort);
> > #else
> >      object_property_set_int(OBJECT(cpu), "family", 6, &error_abort);
> >      object_property_set_int(OBJECT(cpu), "model", 6, &error_abort);
> >      object_property_set_int(OBJECT(cpu), "stepping", 3, &error_abort);
> > #endif
> > 
> > The former is a 64-bit AMD model and the latter is a 32-bit model.
> > 
> > Seems LLVM was sensitive to this distinction to some extent:
> > 
> >     https://gitlab.com/qemu-project/qemu/-/issues/191
> > 
> > A further difference is that qemy-system-i686 does not appear to enable
> > the 'syscall' flag, but I've not figured out where that difference is
> > coming from in the code.
> 
> Ugh, ok. I gave it a quick try with a patch like this:
> 
> diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> --- a/target/i386/cpu.c
> +++ b/target/i386/cpu.c
> @@ -4344,15 +4344,15 @@ static void max_x86_cpu_initfn(Object *obj)
>       */
>      object_property_set_str(OBJECT(cpu), "vendor", CPUID_VENDOR_AMD,
>                              &error_abort);
> -#ifdef TARGET_X86_64
> -    object_property_set_int(OBJECT(cpu), "family", 15, &error_abort);
> -    object_property_set_int(OBJECT(cpu), "model", 107, &error_abort);
> -    object_property_set_int(OBJECT(cpu), "stepping", 1, &error_abort);
> -#else
> -    object_property_set_int(OBJECT(cpu), "family", 6, &error_abort);
> -    object_property_set_int(OBJECT(cpu), "model", 6, &error_abort);
> -    object_property_set_int(OBJECT(cpu), "stepping", 3, &error_abort);
> -#endif
> +    if (object_property_get_bool(obj, "lm", &error_abort)) {
> +        object_property_set_int(obj, "family", 15, &error_abort);
> +        object_property_set_int(obj, "model", 107, &error_abort);
> +        object_property_set_int(obj, "stepping", 1, &error_abort);
> +    } else {
> +        object_property_set_int(obj, "family", 6, &error_abort);
> +        object_property_set_int(obj, "model", 6, &error_abort);
> +        object_property_set_int(obj, "stepping", 3, &error_abort);
> +    }
>      object_property_set_str(OBJECT(cpu), "model-id",
>                              "QEMU TCG CPU version " QEMU_HW_VERSION,
>                              &error_abort);
> 
> ... but it seems like the "lm" property is not initialized
> there yet, so this does not work... :-/
> 
> Giving that we have soft-freeze tomorrow, let's ignore this patch
> for now and revisit this topic during the 8.1 cycle. But I'll
> queue the other 4 patches to get some pressure out of our CI
> during the freeze time.

Yep, makes sense.

More generally the whole impl of the 'max' CPU feels somewhat
questionable even for qemu-system-i386. It exposes all features
that TCG supports. A large set of these features never existed
on *any* 32-bit silicon. Hands up who has seen 32-bit silicon
with AVX2 support ? From a correctness POV we should have
capped CPU features in some manner. Given the lack of interest
in 32-bit though, we've ignored the problem and it likely does
not affect apps anyway as they're not likely to be looking for
newish features.

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|