Re: [Qemu-devel] [RFC PATCH v0 2/9] cpu: Store CPU typename in MachineState

[Eduardo Habkost]

On Fri, Dec 18, 2015 at 11:46:05AM +0100, Igor Mammedov wrote:
> On Thu, 17 Dec 2015 16:09:23 -0200
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Wed, Dec 16, 2015 at 11:26:20PM +0100, Igor Mammedov wrote:
> > > On Wed, 16 Dec 2015 17:39:02 -0200
> > > Eduardo Habkost <address@hidden> wrote:
> > > 
> > > > On Wed, Dec 16, 2015 at 05:54:25PM +0100, Igor Mammedov wrote:
> > > > > On Tue, 15 Dec 2015 14:08:09 +0530
> > > > > Bharata B Rao <address@hidden> wrote:
> > > > > 
> > > > > > On Mon, Dec 14, 2015 at 03:29:49PM -0200, Eduardo Habkost wrote:
> > > > > > > On Thu, Dec 10, 2015 at 11:45:37AM +0530, Bharata B Rao wrote:
> > > > > > > > Storing CPU typename in MachineState lets us to create CPU
> > > > > > > > threads for all architectures in uniform manner from
> > > > > > > > arch-neutral code.
> > > > > > > > 
> > > > > > > > TODO: Touching only i386 and spapr targets for now
> > > > > > > > 
> > > > > > > > Signed-off-by: Bharata B Rao <address@hidden>
> > > > > > > 
> > > > > > > Suggestions:
> > > > > > > 
> > > > > > > * Name the field "cpu_base_type" to indicate it is the base CPU
> > > > > > >   class name, not the actual CPU class name used when creating
> > > > > > >   CPUs.
> > > > > > > * Put it in MachineClass, as it may be useful for code that
> > > > > > >   runs before machine->init(), in the future.
> > > > > > 
> > > > > > Ok.
> > > > > > 
> > > > > > > * Maybe make it a CPUClass* field instead of a string?
> > > > > > 
> > > > > > In the current use case, this base cpu type string is being passed
> > > > > > to cpu_generic_init(const char *typename, const char *cpu_model)
> > > > > > to create boot time CPUs with given typename and cpu_mode. So for
> > > > > > now the string makes sense for use case.
> > > > > > 
> > > > > > Making it CPUClass* would necessiate more changes to
> > > > > > cpu_generic_init().
> > > > > how about actually leaving it as "cpu_type" and putting in it
> > > > > actual cpu type that could be used with device_add().
> > > > > 
> > > > > that would get rid of keeping and passing around intermediate
> > > > > cpu_model.
> > > > 
> > > > Makes sense. We only need to save both typename and cpu_model
> > > > today because cpu_generic_init() currently encapsulates three
> > > > steps: CPU class lookup + CPU creation + CPU feature parsing. But
> > > > we shouldn't need to redo CPU class lookup every time.
> > > BTW: Eduardo do you know if QEMU could somehow provide a list of
> > > supported CPU types (i.e. not cpumodels) to libvirt?
> > 
> > Not sure I understand the question. Could you clarify what you
> > mean by "supported CPU types", and what's the problem it would
> > solve?
> device_add TYPE, takes only type name so I'd like to kep it that way
> and make sure that libvirt/user can list cpu types that hi would
> be able to use with device_add/-device.
> 
> for PC they are generated from cpu_model with help of x86_cpu_type_name()

What about adding a "qom-type" field to query-cpu-definitions?

> 
> > > 
> > > > 
> > > > We could just split cpu_model once, and save the resulting
> > > > CPUClass* + featurestr, instead of saving the full cpu_model
> > > > string and parsing it again every time.
> > > isn't featurestr as x86/sparc specific?
> > > 
> > > Could we have field in  x86_cpu_class/sparc_cpu_class for it and set it
> > > when cpu_model is parsed?
> > > That way generic cpu_model parser would handle only cpu names and
> > > target specific overrides would handle both.
> > 
> > I always assumed we want to have a generic CPU model + featurestr
> > mechanism that could be reused by multiple architectures.
> 
> I've thought the opposite way, that we wanted to faze out featurestr
> in favor of generic option parsing of generic device, i.e.
>  -device TYPE,option=X,...
> but we would have to keep compatibility with old CLI
> that supplies cpu definition via -cpu cpu_model,featurestr
> so cpu_model translated into "cpu_type" field make sense for every
> target but featurestr is x86/sparc specific and I'd prefer to
> keep it that way and do not introduce it to other targets.

I see, and it may make sense long term. But do you really think
we will be able to deprecate "-cpu" and "-smp" soon?

We already have CPUClass::parse_features, and cpu_generic_init()
already makes the model/featurestr split. Do you propose we
remove that generic code and move it back to x86/sparc?

Also, are you sure no other architectures support options in
"-cpu"? cpu_common_parse_features() already supports setting QOM
properties, and it is already available on every architecture
that calls CPUClass::parse_features or uses cpu_generic_init(). I
see that both ARM and PPC have properties available in their CPU
classes, and call parse_features() too.

-- 
Eduardo