Re: [PATCH 51/55] qdev: Make qdev_realize() support bus-less devices

[Markus Armbruster <armbru@redhat.com>]

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 25/05/20 08:38, Markus Armbruster wrote:
>> Paolo Bonzini <pbonzini@redhat.com> writes:
>> 
>>> On 20/05/20 17:02, Markus Armbruster wrote:
>>>>>>
>>>>>> qdev_realize_and_unref() remains restricted, because its reference
>>>>>> counting would become rather confusing for bus-less devices.
>>>>> I think it would be fine, you would just rely on the reference held by
>>>>> the QOM parent (via the child property).
>>>> I took one look at the contract I wrote for it, and balked :)
>>>>
>>>> qdev_realize()'s contract before this patch:
>>>>
>>>>     /*
>>>>      * Realize @dev.
>>>>      * @dev must not be plugged into a bus.
>>>>      * Plug @dev into @bus.  This takes a reference to @dev.
>>>>      * If @dev has no QOM parent, make one up, taking another reference.
>>>>      * On success, return true.
>>>>      * On failure, store an error through @errp and return false.
>>>>      */
>>>>     bool qdev_realize(DeviceState *dev, BusState *bus, Error **errp)
>>>>
>>>> Simple enough.
>>>>
>>>> This patch merely adds "If @bus, " before "plug".  Still simple enough.
>>>>
>>>> qdev_realize_and_unref()'s contract:
>>>>
>>>>     /*
>>>>      * Realize @dev and drop a reference.
>>>>      * This is like qdev_realize(), except it steals a reference rather
>>>>      * than take one to plug @dev into @bus.  On failure, it drops that
>>>>      * reference instead.  @bus must not be null.  Intended use:
>>>>      *     dev = qdev_new();
>>>>      *     [...]
>>>>      *     qdev_realize_and_unref(dev, bus, errp);
>>>>      * Now @dev can go away without further ado.
>>>>      */
>>>>     bool qdev_realize_and_unref(DeviceState *dev, BusState *bus, Error **errp)
>>>>
>>>> If @bus is null, who gets to hold the stolen reference?
>>>>
>>>> You seem to suggest the QOM parent.  What if @dev already has a parent?
>>>
>>> The caller would still hold the stolen reference, and it would be
>>> dropped.
>> 
>> I read this sentence three times, and still don't get it.  Is the
>> reference held or is it dropped?
>
> To call qdev_realize_and_unref, you need to have your own reference,
> which you probably got from qdev_new.
>
> The function might add one via object_property_add_child or it might
> not; it might add one via qdev_set_parent_bus or it might not.  But in
> any case, when it returns you won't have a reference anymore.
>
> One possibility is to think of it in terms of stealing the reference and
> passing it to the bus.  However, as in the lifetime phases that I posted
> earlier, once you realize a device you are no longer in charge of its
> lifetime.  Instead, the unparent callback will take care of unrealizing
> the device and dropping all outstanding long-living references.
>
> So...
>
>>> Or alternatively, ignore all the stolen references stuff, and merely see
>>> qdev_realize_and_unref as a shortcut for qdev_realize+object_unref,
>>> because it's a common idiom.
>> 
>> Even common idioms need to make sense :)
>
> ... that's why the common idiom makes sense.
>
>> The contract must specify exactly what happens to the reference count,
>> case by case.
>
> For both qdev_realize and qdev_realize_and_unref, on return the caller
> need not care about keeping alive the device in the long-term.
>
> For qdev_realize_and_unref, the caller must _also_ have a "private"
> reference to the object, which will be dropped on return.
>
> For qdev_realize, the caller _can_ have a private reference that it has
> to later drop at a convenient time, but it could also ensure that the
> device has a long-term reference via object->parent instead.

I need a contract.  The difficulty of writing a clear contract, caused
by a case that doesn't actually occur, is what made me limit null bus to
qdev_realize().  I admittedly didn't try hard.  Next try:

    /*
     * Realize @dev and drop a reference.
     * This is like qdev_realize(), except the caller must hold a
     * (private) reference, which is dropped on return regardless of
     * success or failure.  Intended use:
     *     dev = qdev_new();
     *     [...]
     *     qdev_realize_and_unref(dev, bus, errp);
     * Now @dev can go away without further ado.
     */

> Perhaps this tells us that the /machine/unattached automation actually
> shouldn't be moved to qdev_realize, but rather to
> qdev_realize_and_unref, and qdev_realize could assert that there is a
> parent already at the time of the call.  However it is probably too
> early to make a decision on that.

The common pairings are qdev_new() with qdev_realize_and_unref(), and
object_initialize_child() with qdev_realize().  Your idea obviously
works for these.  Whether there are other uses where it might not work,
I can't say offhand.