Re: [Qemu-devel] [PATCH 07/23] target-i386: convert cpuid features into properties

[Eduardo Habkost]

On Thu, Oct 04, 2012 at 03:50:34PM +0200, Igor Mammedov wrote:
> On Thu, 4 Oct 2012 10:33:30 -0300
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Thu, Oct 04, 2012 at 03:25:59PM +0200, Igor Mammedov wrote:
> > > On Thu, 4 Oct 2012 10:10:27 -0300
> > > Eduardo Habkost <address@hidden> wrote:
> > > 
> > > > On Thu, Oct 04, 2012 at 03:01:19PM +0200, Igor Mammedov wrote:
> > > > > On Thu, 4 Oct 2012 09:43:41 -0300
> > > > > Eduardo Habkost <address@hidden> wrote:
> > > > > 
> > > > > > On Thu, Oct 04, 2012 at 08:53:22AM +0200, Igor Mammedov wrote:
> > > > > > > On Wed, 3 Oct 2012 13:54:34 -0300
> > > > > > > Eduardo Habkost <address@hidden> wrote:
> > > > > > > 
> > > > > > > > On Wed, Oct 03, 2012 at 06:24:11PM +0200, Igor Mammedov wrote:
> > > > > > > > > On Wed, 03 Oct 2012 17:20:46 +0200
> > > > > > > > > Paolo Bonzini <address@hidden> wrote:
> > > > > > > > > 
> > > > > > > > > > Il 03/10/2012 17:03, Eduardo Habkost ha scritto:
> > > > > > > > > > > On Tue, Oct 02, 2012 at 05:38:45PM -0300, Eduardo Habkost
> > > > > > > > > > > wrote:
> > > > > > > > > > >> (Now replying on the right thread, to keep the
> > > > > > > > > > >> discussion in the right place. I don't know how I ended
> > > > > > > > > > >> up replying to a pre-historic version of the patch,
> > > > > > > > > > >> sorry.)
> > > > > > > > > > >>
> > > > > > > > > > >> On Tue, Oct 02, 2012 at 05:36:59PM +0200, Igor Mammedov
> > > > > > > > > > >> wrote: [...]
> > > > > > > > > > >>> @@ -1938,6 +2043,12 @@ static void x86_cpu_initfn(Object
> > > > > > > > > > >>> *obj) object_property_add(obj, "tsc-frequency", "int",
> > > > > > > > > > >>>                          x86_cpuid_get_tsc_freq,
> > > > > > > > > > >>>                          x86_cpuid_set_tsc_freq, NULL,
> > > > > > > > > > >>> NULL, NULL);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj, feature_name);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj,
> > > > > > > > > > >>> ext_feature_name);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj,
> > > > > > > > > > >>> ext2_feature_name);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj,
> > > > > > > > > > >>> ext3_feature_name);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj,
> > > > > > > > > > >>> kvm_feature_name);
> > > > > > > > > > >>> +    x86_register_cpuid_properties(obj,
> > > > > > > > > > >>> svm_feature_name);
> > > > > > > > > > >>
> > > > > > > > > > >> Stupid question about qdev:
> > > > > > > > > > >>
> > > > > > > > > > >> - qdev_prop_set_globals() is called from device_initfn()
> > > > > > > > > > >> - device_initfn() is called before the child class
> > > > > > > > > > >> instance_init() function (x86_cpu_initfn())
> > > > > > > > > > >> - So, qdev_prop_set_globals() gets called before the CPU
> > > > > > > > > > >> class properties are registered.
> > > > > > > > > > >>
> > > > > > > > > > >> So this would defeat the whole point of all the work
> > > > > > > > > > >> we're doing, that is to allow compatibility bits to be
> > > > > > > > > > >> set as machine-type global properties. But I don't know
> > > > > > > > > > >> what's the right solution here.
> > > > > > > > > > >>
> > > > > > > > > > >> Should the qdev_prop_set_globals() call be moved to
> > > > > > > > > > >> qdev_init() instead? Should the CPU properties be
> > > > > > > > > > >> registered somewhere else?
> > > > > > > > > > 
> > > > > > > > > > Properties should be registered (for all objects, not just
> > > > > > > > > > CPUs) in the instance_init function.  This is device_initfn.
> > > > > > > > > > 
> > > > > > > > > > I would add an instance_postinit function that is called at
> > > > > > > > > > the end of object_initialize_with_type, that is after
> > > > > > > > > > instance_init, and in the opposite order (i.e. from the
> > > > > > > > > > leaf to the root).
> > > > > > > > > 
> > > > > > > > > You've meant something like that?
> > > > > > > > > 
> > > > > > > > 
> > > > > > > > That's almost exactly the same code I wrote here. :-)
> > > > > > > > 
> > > > > > > > The only difference is that I added post_init to the struct
> > > > > > > > Object documentation comments, and added a unit test. The unit
> > > > > > > > test required the qdev-core/qdev split, so we could compile it
> > > > > > > > without bringing too many dependencies. I will submit it soon.
> > > > > > > > 
> > > > > > > After irc discussion, Anthony suggested to use static properties
> > > > > > > instead of dynamic ones that we use now. 
> > > > > > > 
> > > > > > > But  qdev_prop_set_globals() in device_initfn() is still causes
> > > > > > > problems even with static properties.
> > > > > > > 
> > > > > > > For x86 CPU classes we were going dynamically generate CPU
> > > > > > > classes and store pointer to appropriate cpudef from
> > > > > > > builtin_x86_defs in class field for each CPU class and then init
> > > > > > > default feature words values from this field int x86_cpu_initfn().
> > > > > > > 
> > > > > > > However with qdev_prop_set_globals() in device_initfn() that is
> > > > > > > called before x86_cpu_initfn() it won't work because defaults in
> > > > > > > x86_cpu_initfn() will overwrite whatever was set by
> > > > > > > qdev_prop_set_globals().
> > > > > > 
> > > > > > We can set the default values on class_init, instead. The class_init
> > > > > > function for each CPU model can get the x86_def_t struct as the data
> > > > > > pointer.
> > > > > > 
> > > > > > I still think that the interface to build the DeviceClass.props
> > > > > > array on class_init is really painful to use, but it's still doable.
> > > > > 
> > > > > You mean dynamic building of DeviceClass.props arrays for each CPU
> > > > > sub-class?
> > > > 
> > > > That's the only solution I see if we want to make all the CPU properties
> > > > static, yes.
> > > 
> > > Well I could generate compile time arrays for every built-in cpu model
> > > and we can remove then x86_def_t struct & builtins altogether.
> > > Only 'host' would be left for dynamic generation then.
> > 
> > You mean duplicating the property list in the code? Then the
> > feature-name -> CPUID-bit mapping information would be duplicated on all
> > those arrays, and adding support to a new CPU feature would require
> > adding entries to all the arrays.
> Not to all arrays, but only to ones which cpumodel-s support specific feature.
> 
> Here is possible ups&downs this:
>  + full introspection, including default cpu features values.
>  -/+ it would be possible to represent cpumodel more faithfully, i.e. include
>    only features that specific cpu supports. (so no AVX=on in 486 model), not
>    sure if it is plus but it would more like real hw.

It looks like a downside, as it would break configurations that could
have worked previously.

I mean: a 486 with AVX may not exist, but that doesn't mean people don't
have existing (and working) VMs in their systems with those settings,
and they would break on a QEMU upgrade.


>  - a lot of lines of code , but it could be dealt with extra macros, so
>    resulting arrays could look like built-in now (if we add feature there we
>    anyway should replicate it other relevant builtins).

The main problem I see is that the macros that define the property have
to define its default value, too.

For example: Westmere supports the tsc-deadline flag, but it
default=false, and SandyBride would support the same flag, but with
default=true. That means that somewhere in the code we will have a line
saying:
 DEFINE_PROP_BIT("tsc-deadline", CPUX86State, ext_features, 24, false)
and somewhere else, we will have a line saying:
 DEFINE_PROP_BIT("tsc-deadline", CPUX86State, ext_features, 24, true)

I don't see how we could let different classes have different defaults
without duplicating the list of properties, even with help of macros.

>  + maybe generating only "host"'s DeviceClass.props could be simplier.

True. But we can still make the "host" class simpler, even if we don't
do any of the above. (It could use a different instance init function,
for example).


Anyway, I don't see why so much effort to generate those property lists
at compile time. The list is set on DeviceClass.props at runtime (at
class_init) so we can generate the property list at runtime inside
class_init, already.


> 
> And then for builtin models I could make a series static CPU classes that
> would use this arrays.
>  
> 
> > 
> > > 
> > > > 
> > > > I'm still not convinced we really need to do that, though. Maybe we can
> > > > make static only the ones we really need to be able to implement
> > > > machine-type-compatibility global properties?
> > > > 
> > > > Machine-type compatibility global properties were the initial reason for
> > > > the static-properties requirement. We don't really need to allow _all_
> > > > CPU features to be controlled by global properties, only the ones we
> > > > need for machine-type compatibility.
> > > > 
> > > 
> > 
> 

-- 
Eduardo