Re: [Qemu-devel] [PATCH 0/4] Remove some qdev_get_machine() calls from CONFIG_USER_ONLY

[Like Xu]

On 2019/4/26 4:00, Eduardo Habkost wrote:
> This series moves some qdev code outside qdev.o, so it can be
> compiled only in CONFIG_SOFTMMU.
>
> The code being moved includes two qdev_get_machine() calls, so
> this will make it easier to move qdev_get_machine() to
> CONFIG_SOFTMMU later.
>
> After this series, there's one remaining qdev_get_machine() call
> that seems more difficult to remove:
>
>      static void device_set_realized(Object *obj, bool value, Error **errp)
>      {
>          /* [...] */
>          if (!obj->parent) {
>              gchar *name = g_strdup_printf("device[%d]", unattached_count++);
>
>              object_property_add_child(container_get(qdev_get_machine(),
>                                                      "/unattached"),
>                                        name, obj, &error_abort);
>              unattached_parent = true;
>              g_free(name);
>          }
>          /* [...] */
>      }

I may have an experimental patch to fix device_set_realized issue:

1. in qdev_get_machine():
replace
        dev = container_get(object_get_root(), "/machine");
with
        dev = object_resolve_path("/machine", NULL);

2. in device_set_realized():

Using
        Object *container = qdev_get_machine() ?
                qdev_get_machine() : object_get_root();
and pass it to
        object_property_add_child(
                container_get(container, "/unattached"),
                name, obj, &error_abort);

With this fix, we could say the qdev_get_machine() does
return the "/machine" object (or null) not a confused "/container".

We could continue to use qdev_get_machine() in system emulation mode,
getting rid of its surprising side effect as Markus said.

The return value of qdev_get_machine() in user-only mode
is the same object returned by object_get_root(),
so no semantic changes.


> This one is tricky because on system emulation mode it needs
> "/machine" to already exist, but in user-only mode it needs to
> implicitly create a "/machine" container.
>
> Eduardo Habkost (4):
>    machine: Move gpio code to hw/core/gpio.c
>    move qdev hotplug code to qdev-hotplug.c
>    qdev: Don't compile hotplug code in user-mode emulation
>    qdev-hotplug: Don't check type of qdev_get_machine()
>
>   hw/core/bus.c                |  11 --
>   hw/core/gpio.c               | 206 ++++++++++++++++++++++++++++++++
>   hw/core/qdev-hotplug-stubs.c |  44 +++++++
>   hw/core/qdev-hotplug.c       |  64 ++++++++++
>   hw/core/qdev.c               | 219 -----------------------------------
>   hw/core/Makefile.objs        |   5 +-
>   tests/Makefile.include       |   3 +-
>   7 files changed, 320 insertions(+), 232 deletions(-)
>   create mode 100644 hw/core/gpio.c
>   create mode 100644 hw/core/qdev-hotplug-stubs.c
>   create mode 100644 hw/core/qdev-hotplug.c