Re: [PATCH v3] target/s390x: filter deprecated properties based on model expansion type

[Markus Armbruster <armbru@redhat.com>]

Collin Walling <walling@linux.ibm.com> writes:

> On 7/24/24 3:56 AM, Markus Armbruster wrote:
>> Collin Walling <walling@linux.ibm.com> writes:
> Let me try to explain the purpose of @deprecated-props and see if it
> helps bring us closer to some semblance of a mutual understanding so we
> can work together on a concise documentation for this field.
>
> s390 has been announcing features as deprecated for some time now, which
> was fine as a way to let users know that they should tune their guests
> to no longer user these features.  Now that we are approaching the
> release of generations that will drop these deprecated features
> outright, we encounter an issue: if users have not been mindful with
> disabling these announced-deprecated-features, then their guests running
> on older models will not be able to migrate to machines running on newer
> hardware.
>
> To alleviate this, I've added the @deprecated-props array to the
> CpuModelInfo struct, and this field is populated by a
> query-cpu-model-expansion* return.  It is up the the user/management app
> to make use of this data.
>
> On the libvirt side (currently in development), I am able to easily
> retrieve the host-model with a full expansion, parse the
> @deprecated-props, and then cache them for later use (e.g. when
> reporting the host-model with these features disabled, or enabling a
> user to define their domain with deprecated-features disabled via a
> convenient XML attribute).
>
> tl;dr @deprecated-props is only reported via a
> query-cpu-model-expansion, and it is up to the user/management app to
> figure out what to do with them.

Got it.

Permit me a digression.  In QAPI/QMP, we do something similar: we expose
deprecation in introspection (query-qmp-schema), and what to do with the
information is up to the management application.  We provide one more
tool to it: policy for handling deprecated interfaces, set with -compat.
It permits "testing the future".  See qapi/compat.json for details.
Whether such a thing would be usful in your case I can't say.

>> On closer examination, more questions on CpuModelInfo emerge.  Uses:
>> 
>
> I will attempt to expand on each input @model (CpuModelInfo) as if they
> were documented in the file.
>
>> * query-cpu-model-comparison both arguments
>> 
>>   Documentation doesn't say how exactly the command uses the members of
>>   CpuModelInfo, i.e. @name, @props, @deprecated-props.  Can you tell me?
>> 
>
> Note: Compares ModelA and ModelB.
>
> Both @models must include @name.  @props is optional.  @deprecated-props
> is ignored.
>
> @name: the name of the CPU model definition to look up.  The definition
> will be compared against the generation, GA level, and a static set of
> properties of the opposing model.
>
> @props: a set of additional properties to include in the model's set of
> properties to be compared.
>
> @deprecated-props: ignored.  The user should consider these properties
> beforehand and decide if these properties should be disabled/omitted on
> the respective model.
>
>> * query-cpu-model-expansion argument @model and return value member
>>   @model.
>> 
>>   The other argument is the expansion type, on which the value of return
>>   value model.deprecated-props depends, I believe.  Fine.
>> 
>>   Documentation doesn't say how exactly the command uses the members of
>>   CpuModelInfo arguments, i.e. @name, @props, @deprecated-props.  Can
>>   you tell me?
>> 
>
> The @model must include @name.  @props is optional.  @deprecated-props
> is ignored.
>
> @name: the name of the CPU model definition to look up.  The definition
> is associated with a set of properties that will populate the return data.
>
> @props: a set of additional properties to include in the model's set of
> expanded properties.
>
> @deprecated-props: ignored.  The user should consider these properties
> beforehand and decide if these properties should be disabled/omitted on
> the model.

Return value member @model will have @name, may have @props and
@deprecated-props.

Absent @props is the same as {}.  Only x86 uses {}.

Absent @deprecated-props is the same as {}.  No target uses {}.  Can be
present only on S390.

Aside: returning the same thing in two different ways, like absent and
{}, is slightly more complex than necessary.  But let's ignore that
here.

>> * query-cpu-model-baseline both arguments and return value member
>>   @model.
>> 
>>   Same, except we don't have an expansion type here.  So same question,
>>   plus another one: how does return value model.deprecated-props behave?
>> 
>
> Note: Creates a baseline model based on ModelA and ModelB.
>
> The @models must include @name.  @props is optional.  @deprecated-props
> is ignored.
>
> @name: the name of the CPU model definition to look up.  The definition,
> GA level, and a static set of properties will be used to determine the
> maximum model between ModelA and ModelB.
>
> @props: a set of additional properties to include in the model's set of
> properties to be baselined.
>
> @deprecated-props: ignored.  The user should consider these properties
> beforehand and decide if these properties should be disabled/omitted on
> the respective model.

Return value member @model is just like in query-cpu-model-expansion.

Unlike query-cpu-model-expansion, we don't have an expansion type.  The
value of @deprecated-props depends on the expansion type.  Do we assume
a type?  Which one?

>> If you can't answer my questions, we need to find someone who can.
>> 
>
> Hopefully this provides clarity on how CpuModelInfo and its respective
> fields are used in each command.  @David should be able to fill in any
> missing areas / expand / offer corrections.
>
>> [...]

This helps, thanks!

Arguments that are silently ignored is bad interface design.

Observe: when CpuModelInfo is an argument, @deprecated-props is always
ignored.  When it's a return value, absent means {}, and it can be
present only for certain targets (currently S390).

The reason we end up with an argument we ignore is laziness: we use the
same type for both roles.  We can fix that easily:

    { 'struct': 'CpuModel',
      'data': { 'name': 'str',
                '*props': 'any' } }

    { 'struct': 'CpuModelInfo',
      'base': 'CpuModel',
      'data': { '*deprecated-props': ['str'] } }

Use CpuModel for arguments, CpuModelInfo for return values.

Since @deprecated-props is used only by some targets, I'd make it
conditional, i.e. 'if': 'TARGET_S390X'.

Thoughts?