Re: [Qemu-ppc] [Qemu-devel] [PATCH 2/2] vl: Print CPU help after we've registered the CPU accelerators

[Eduardo Habkost]

On Tue, Mar 07, 2017 at 10:02:26AM +0100, Thomas Huth wrote:
> On 03.03.2017 15:58, Eduardo Habkost wrote:
> [...]
> > I am not convinced that the output of "-cpu help" and
> > "-cpu help -machine accel=kvm" should look different. Do you have
> > an example of what exactly is wrong with the output currently?
> 
> The problem is that on POWER, we've got a "family" of CPUs with
> different sub-types, e.g. for POWER8:
> 
> $ qemu-system-ppc64 -cpu ? | grep POWER8
> PowerPC POWER8E_v2.1     PVR 004b0201
> PowerPC POWER8E          (alias for POWER8E_v2.1)
> PowerPC POWER8NVL_v1.0   PVR 004c0100
> PowerPC POWER8NVL        (alias for POWER8NVL_v1.0)
> PowerPC POWER8_v2.0      PVR 004d0200
> PowerPC POWER8           (alias for POWER8_v2.0)
> 
> Most of the users don't know about the current subtype that they are
> using, and just want to use "-cpu POWER8" - and for example we've also
> got an agreement with the libvirt folks that they can always use "-cpu
> POWER8" for any kind of POWER8 system, no matter whether the host is
> using a POWER8E or POWER8NVL chip.
> So the "POWER8" alias now gets updated internally in QEMU to the correct
> host CPU type ... but the output of "-cpu help" is then still wrong.
> I agree that it's kind of ugly to have different help texts depending on
> whether "accel=kvm" has been used or not, but that sounds still better
> to me than printing wrong information here.

I agree that incorrect information is even worse than showing
different help information depending on accel=kvm, but:

> Thinking about this again ... maybe it would be better if we'd rework
> the help text to print out something like this instead:
> 
> PowerPC POWER8           (alias for any POWER8 chip)
> 
> ... so that we simply get rid of the version/subtype information here
> completely?

Yes, making help output not depend on accel=kvm sounds better to
me.

This seems to be affected only by the alias table, so it can be
fixed even before we address the late-type_register() issue I was
discussing with David?

-- 
Eduardo