Re: [PATCH v2 00/44] Make qdev static property API usable by any QOM type

[Eduardo Habkost <ehabkost@redhat.com>]

On Mon, Nov 09, 2020 at 06:33:04PM +0100, Paolo Bonzini wrote:
> On 09/11/20 18:16, Eduardo Habkost wrote:
> > On Mon, Nov 09, 2020 at 05:34:01PM +0100, Paolo Bonzini wrote:
> > > On 09/11/20 16:21, Eduardo Habkost wrote:
> > > > Nothing prevents us from describing those properties inside the
> > > > same property array.
> > > 
> > > Do you mean adding PropertyInfos for them?  Adding a once-only PropertyInfo
> > > is worse than writing a custom getter/setter pair, because:
> > > 
> > > - without (DEFINE_)PROP_* you lose the type safety.
> > > 
> > > - with (DEFINE_)PROP_* you have much more boilerplate to write
> > 
> > I mean extending the API to let custom setters and getters appear
> > on the Property array, not using the existing API.
> 
> That seems like conflicting goals.  The field property API is based on
> getters and setters hidden in PropertyInfo.  The "other" property API is
> based on getters and setters in plain sight in the declaration of the
> property.

There's nothing that prevents a
  void object_class_add_properties(oc, Property *props);
function from supporting both.

> 
> > > > > > I think having different ways for different things (class vs. object) is
> > > > > > better than having different ways for the same things (class in qdev vs.
> > > > > > class in non-qdev).
> > > > > 
> > > > > Right, but qdev's DEFINE_PROP_STRING would be easy to change to something
> > > > > like
> > > > > 
> > > > > - DEFINE_PROP_STRING("name", ...),
> > > > > + device_class_add_field_property(dc, "name", PROP_STRING(...));
> > > > 
> > > > I'm not worried about this direction of conversion (which is
> > > > easy).  I'm worried about the function call => QAPI schema
> > > > conversion.  Function calls are too flexible and requires parsing
> > > > and executing C code.
> > > 
> > > Converting DEFINE_PROP_STRING to a schema also requires parsing C code,
> > > since you can have handwritten Property literals (especially for custom
> > > PropertyInfo).  Converting DEFINE_PROP_STRING it also requires matching the
> > > array against calls to object_class_add_field_properties (which could be
> > > hidden behind helpers such as device_class_set_props).  (Plus matching
> > > class_init functions against TypeInfo).
> > 
> > Parsing an array containing a handful of macros (a tiny subset of
> > C) isn't even comparable to parsing and executing C code where
> > object*_property_add*() calls can be buried deep in many levels
> > of C function calls (which may or may not be conditional).
> 
> Finding the array would also require finding calls buried deep in C code,
> wouldn't they?

Yes, but I don't expect this to happen if the API doesn't
encourage that.

> 
> > (Also, I don't think we should allow handwritten Property literals.)
> 
> How would you do custom setters and getters then---without separate
> PropertyInfos, without Property literals, and without an exploding number of
> macros?

Property with struct field:

  /* We call this DEFINE_PROP_UINT32 today.  We can keep the
   * existing name just to reduce churn.
   */
  DEFINE_PROP_UINT32_FIELD("myproperty", MyState, my_field)


Prop with struct field but custom setter:

  DEFINE_PROP_UINT32_FIELD("myproperty", MyState, my_field,
                           .custom_setter = my_custom_setter)

Prop with no struct field, and custom setter/getter:

  DEFINE_PROP("myproperty", prop_type_uint32,
              .custom_getter = my_getter,
              .custom_setter = my_setter)


Definitions for above:

#define DEFINE_PROP(_name, _typeinfo, ...) \
    { .name = _name,
      .info = &_typeinfo,
      __VA_ARGS__
    }

#define DEFINE_FIELD_PROP(name, typeinfo, type, state, field, ...) \
    DEFINE_PROP(name, typeinfo,
                .offset = offsetof(state, field) +
                          type_check(typeof_field(state, field), type),
                __VA_ARGS__)

#define DEFINE_PROP_UINT32_FIELD(name, state, field, ...) \
    DEFINE_FIELD_PROP(name, prop_type_uint32, uint32_t, state, field, __VA_ARGS__)


Alternative DEFINE_FIELD_PROP definition if we implement some
macro magic to declare the expected type for each typeinfo
variable:

/* Will make ACTUAL_C_TYPE(prop_type_uint32) expand to uint32_t */
DECLARE_QOM_TYPE(prop_type_uint32, uint32_t)
/* Will make ACTUAL_C_TYPE(prop_type_uint64) expand to uint64_t) 
DECLARE_QOM_TYPE(prop_type_uint64, uint64_t)

#define DEFINE_FIELD_PROP(name, typeinfo, state, field, ...) \
    DEFINE_PROP(name, typeinfo,
                .offset = offsetof(state, field) +
                          type_check(typeof_field(state, field),
                                     ACTUAL_C_TYPE(typeinfo)),
                __VA_ARGS__)


> 
> > > So, you don't save any parsing by using arrays.  (In fact I would probably
> > > skip the parsing, and use your suggestion of *executing* C code: write the
> > > QAPI schema generator in C, link into QEMU and run it just once to generate
> > > the QOM schema).
> > 
> > If we do that with the existing code, we can't be sure the
> > generated schema doesn't depend on configure flags or run time
> > checks inside class_init.
> 
> We can use grep or Coccinelle or manual code review to identify problematic
> cases.

We can, but I believe it is better and simpler to have an API
that enforces (or at least encourages) this.

> 
> > Even locating the cases where this is
> > happening is being a challenge because the API is too flexible.
> > 
> > However, if we require the property list to be always evaluated
> > at compile time, we can be sure that this method will be
> > reliable.
> > 
> > > QOM has been using function calls for many years, are there any cases of
> > > misuse of that flexibility that you have in mind?  I can only think of two
> > > *uses*, in fact.  One is eepro100_register_types is the only case I can
> > > remember where types are registered dynamically.  The other is S390 CPU
> > > features.  [...]
> > 
> > The list of tricky dynamic properties is large and I don't think
> > we even found all cases yet.  John documented many of them here:
> > 
> > https://gitlab.com/jsnow/qemu/-/blob/cli_audit/docs/cli_audit.md
> > 
> > Look, for example, for the sections named "Features" for CPU
> > options.
> 
> Yes, I'm only considering object_class_property calls.  Those are the ones
> that I claim aren't being misused enough for this to be a problem.
> 

instance-level properties are where most of the complexity was
introduced because the class API didn't exist yet.  I don't think
we should ignore them, or we risk having the same issues when
converting them to class properties.


> Making instance-level properties appear in the schema is a completely
> different kind of conversion, because there is plenty of manual work (and
> unsolved problems for e.g. subobject property aliases).

I'd like us to convert instance-level properties to an API that
is easy to use and where the same problems won't happen again.

> 
> > You are also ignoring the complexity of the code path that leads
> > to the object*_property_add*() calls, which is the main problem
> > on most cases.
> 
> I would like an example of the complexity of those code paths.  I don't see
> much complexity, as long as the object exists at all, and I don't see how it
> would be simpler to find the code paths that lead to
> object_class_add_field_properties.

Possibly the most complex case is x86_cpu_register_bit_prop().
The qdev_property_add_static() calls at arm_cpu_post_init() are
tricky too.

If object*_property_add*() is hidden behind a function call or a
`if` statement, it's already too much complexity to me.  I don't
want us to need a second audit like the one John made when we
decide to represent QOM class properties in a QAPI schema.

-- 
Eduardo