Re: [Qemu-devel] [PATCH] qmp: object-add: Validate class before creating object

[Markus Armbruster]

Eduardo Habkost <address@hidden> writes:

> Currently it is very easy to crash QEMU by issuing an object-add command
> using an abstract class or a class that doesn't support
> TYPE_USER_CREATABLE as parameter.
>
> Example: with the following QMP command:
>
>     (QEMU) object-add qom-type=cpu id=foo
>
> QEMU aborts at:
>
>     ERROR:qom/object.c:335:object_initialize_with_type: assertion failed: 
> (type->abstract == false)
>
> This patch moves the check for TYPE_USER_CREATABLE before object_new(),
> and adds a check to prevent the code from trying to instantiate abstract
> classes.
>
> Signed-off-by: Eduardo Habkost <address@hidden>

Related device_add fixes:

061e84f qdev-monitor: Avoid device_add crashing on non-device driver name
2fa4e56 qdev-monitor: Fix crash when device_add is called with abstract driver
7ea5e78 qdev: Do not let the user try to device_add when it cannot work

> ---
>  qmp.c | 21 ++++++++++++++-------
>  1 file changed, 14 insertions(+), 7 deletions(-)
>
> diff --git a/qmp.c b/qmp.c
> index 87a28f7..e7ecbb7 100644
> --- a/qmp.c
> +++ b/qmp.c
> @@ -540,14 +540,27 @@ void object_add(const char *type, const char *id, const 
> QDict *qdict,
>                  Visitor *v, Error **errp)
>  {
>      Object *obj;
> +    ObjectClass *klass;
>      const QDictEntry *e;
>      Error *local_err = NULL;
>  
> -    if (!object_class_by_name(type)) {
> +    klass = object_class_by_name(type);
> +    if (!klass) {
>          error_setg(errp, "invalid class name");
>          return;
>      }
>  
> +    if (object_class_is_abstract(klass)) {
> +        error_setg(errp, "object type '%s' is abstract", type);
> +        return;
> +    }
> +
> +    if (!object_class_dynamic_cast(klass, TYPE_USER_CREATABLE)) {
> +        error_setg(errp, "object type '%s' isn't supported by object-add",
> +                   type);
> +        return;
> +    }
> +
>      obj = object_new(type);
>      if (qdict) {
>          for (e = qdict_first(qdict); e; e = qdict_next(qdict, e)) {

For comparison, this is qdev_device_add():

    oc = object_class_by_name(driver);
    if (!oc) {
        const char *typename = find_typename_by_alias(driver);

        if (typename) {
            driver = typename;
            oc = object_class_by_name(driver);
        }
    }

    if (!object_class_dynamic_cast(oc, TYPE_DEVICE)) {
        qerror_report(ERROR_CLASS_GENERIC_ERROR,
                      "'%s' is not a valid device model name", driver);
        return NULL;
    }

    if (object_class_is_abstract(oc)) {
        qerror_report(QERR_INVALID_PARAMETER_VALUE, "driver",
                      "non-abstract device type");
        return NULL;
    }

    dc = DEVICE_CLASS(oc);
    if (dc->cannot_instantiate_with_device_add_yet) {
        qerror_report(QERR_INVALID_PARAMETER_VALUE, "driver",
                      "pluggable device type");
        return NULL;
    }

Got the "subtype of" and the "not abstract" check in the opposite order,
and the additional cannot_instantiate_with_device_add_yet check.

Does the different order matter?

> @@ -558,12 +571,6 @@ void object_add(const char *type, const char *id, const 
> QDict *qdict,
>          }
>      }
>  
> -    if (!object_dynamic_cast(obj, TYPE_USER_CREATABLE)) {
> -        error_setg(&local_err, "object type '%s' isn't supported by 
> object-add",
> -                   type);
> -        goto out;
> -    }
> -
>      user_creatable_complete(obj, &local_err);
>      if (local_err) {
>          goto out;