Re: [Qemu-devel] [PATCH RFC 2/2] qmp: New command qom-new

[Igor Mammedov]

On 05/24/2012 03:04 PM, Anthony Liguori wrote:
> On 05/24/2012 06:43 AM, Markus Armbruster wrote:
> > To create objects via QMP.
> >
> > Test case:
> >
> > $ upstream-qemu --enable-kvm -S -m 384 -vnc :0 -monitor stdio -chardev 
> > socket,id=qmp,path=test-qmp,server=on,wait=off -mon mode=control,chardev=qmp
> >
> > Conversation on the qmp socket:
> > {"QMP": {"version": {"qemu": {"micro": 93, "minor": 0, "major": 1}, "package": ""}, 
> > "capabilities": []}}
> > { "execute": "qmp_capabilities" }
> > {"return": {}}
> > {"execute":"qom-new","arguments":{"parent":"/xxx", "prop-name":"test-qmp-new", 
> > "type-name":"xxx"}}
> > {"error": {"class": "DeviceNotFound", "desc": "Device '/xxx' not found", "data": 
> > {"device": "/xxx"}}}
> > {"execute":"qom-new","arguments":{"parent":"/machine", "prop-name":"test-qmp-new", 
> > "type-name":"xxx"}}
> > {"error": {"class": "InvalidParameterValue", "desc": "Parameter 'type-name' expects a type name", "data": 
> > {"name": "type-name", "expected": "a type
> > name"}}}
> > {"execute":"qom-new","arguments":{"parent":"/machine", "prop-name":"test-qmp-new", 
> > "type-name":"container"}}
> > {"return": "/machine/test-qmp-new"}
> > {"execute":"qom-list","arguments":{"path":"/machine"}}
> > {"return": [{"name": "test-qmp-new", "type": "child<container>"}, {"name": "i440fx", "type": 
> > "child<i440FX-pcihost>"}, {"name": "unattached",
> > "type": "child<container>"}, {"name": "peripheral", "type": "child<container>"}, {"name": 
> > "peripheral-anon", "type": "child<container>"}]}
> >
> > Note: qdev objects (subtype of TYPE_DEVICE) created with qom-new lack
> > additional magic performed by qdev_try_create(), and almost certainly
> > won't work.
> >
> > Signed-off-by: Markus Armbruster<address@hidden>
> > ---
> > qapi-schema.json | 22 ++++++++++++++++++++++
> > qmp-commands.hx | 5 +++++
> > qmp.c | 27 +++++++++++++++++++++++++++
> > 3 files changed, 54 insertions(+), 0 deletions(-)
> >
> > diff --git a/qapi-schema.json b/qapi-schema.json
> > index 2ca7195..ab9e68b 100644
> > --- a/qapi-schema.json
> > +++ b/qapi-schema.json
> > @@ -1699,6 +1699,28 @@
> > 'returns': [ 'ObjectTypeInfo' ] }
> >
> > ##
> > +# qom-new:
> > +#
> > +# Create a new object
> > +#
> > +# @parent: the parent's path within the object model. See @qom-get
> > +# for a description of paths.
> > +#
> > +# @prop-name: the name of the property to add to the parent.
> > +#
> > +# @type: the new object's type name
> > +#
> > +# Returns: The new object's canonical absolute path
> > +#
> > +# Since: 1.2
> > +#
> > +# Notes: This command is experimental and may change syntax in future releases.
> > +##
> > +{ 'command': 'qom-new',
> > + 'data': { 'parent': 'str', 'prop-name': 'str', 'type-name': 'str' },
> > + 'returns': 'str' }
> > +
> > +##
> > # @migrate
> > #
> > # Migrates the current running guest to another Virtual Machine.
> > diff --git a/qmp-commands.hx b/qmp-commands.hx
> > index db980fa..53adda2 100644
> > --- a/qmp-commands.hx
> > +++ b/qmp-commands.hx
> > @@ -2142,3 +2142,8 @@ EQMP
> > .args_type = "implements:s?,abstract:b?",
> > .mhandler.cmd_new = qmp_marshal_input_qom_list_types,
> > },
> > + {
> > + .name = "qom-new",
> > + .args_type = "parent:s,prop-name:s,type-name:s",
> > + .mhandler.cmd_new = qmp_marshal_input_qom_new,
> > + },
> > diff --git a/qmp.c b/qmp.c
> > index fee9fb2..cad5610 100644
> > --- a/qmp.c
> > +++ b/qmp.c
> > @@ -417,3 +417,30 @@ ObjectTypeInfoList *qmp_qom_list_types(bool has_implements,
> >
> > return ret;
> > }
> > +
> > +char *qmp_qom_new(const char *parent, const char *prop_name,
> > + const char *type_name, Error **errp)
> > +{
> > + Object *p, *obj;
> > + Type type;
> > +
> > + // TODO anything fancy with containger_get() needed?
>
> I'm not sure how I feel about this. I never intended for a user to be able to 
> create objects that were arbitrary children of other objects.
>
> In some ways, I think this is almost too powerful of an interface to expose to 
> users. I like things like device_add() better that only creates objects
> of TYPE_DEVICE that are always in /peripherial.
>
> For block, we'd have a similar interface that always created objects of 
> TYPE_BLOCK_DRIVER and put them in /block.

Will we have a special cases for every incompatible device types that is going 
to be hot-plugged via device_add monitor command?

For CPUs my thoughts were moving in opposite direction, like:
 - make possible to create and initialize CPU as a regular QOM object
 - hack qdev_device_add() to allow not only TYPE_DEVICE to be created there

There are patches out there that make cpu a child of /machine at board level.
But for hot-added objects parent could be specified as a property
or knowledge about parent hard-coded inside of object itself or
hard-coded in device_add().
Which one of them likely to be adopted?

> > + p = object_resolve_path(parent, NULL);
> > + if (!p) {
> > + error_set(errp, QERR_DEVICE_NOT_FOUND, parent);
> > + return NULL;
> > + }
> > +
> > + type = type_get_by_name(type_name);
> > + if (!type) {
> > + error_set(errp, QERR_INVALID_PARAMETER_VALUE,
> > + "type-name", "a type name");
> > + return NULL;
> > + }
> > + obj = object_new_with_type(type);
> > +
> > + // TODO bombs if p is an interface object; can this happen?
>
> All interface types have .abstract set and there's an assert() to validate that 
> an object isn't abstract when creating. So it should be
> object_new_with_type() that's asserting, not object_property_add_child.
>
> We'll probably want to enforce that qom-new isn't given an abstract type by 
> introducing a type_is_abstract() or something like that.
>
> Regards,
>
> Anthony Liguori
>
> > + object_property_add_child(p, prop_name, obj, NULL);
> > +
> > + return object_get_canonical_path(obj);
> > +}

--
-----
 Igor