Re: [Qemu-devel] [PATCH v2 2/3] qobject: introduce QObjectCommon

[Marc-André Lureau]

Hi

On Wed, Mar 28, 2018 at 4:11 PM, Paolo Bonzini <address@hidden> wrote:
> On 28/03/2018 15:55, Eric Blake wrote:
>> On 03/28/2018 08:48 AM, Marc-André Lureau wrote:
>>
>>>>> +#define QOBJECT(x) \
>>>>> +   container_of(&(x)->base, QObject, base)
>>>>
>>>>
>>>> If I understand correctly, this still causes clang complaints when
>>>> called as
>>>> QOBJECT(NULL).  As long as we are touching this, should we improve this
>>>> macro to be friendly to NULL conversion?
>>>
>>> I don't see much need for allowing NULL (literally) to be passed to
>>> QOBJECT(). If it's a null pointer, as long as it has the right type,
>>> it should be fine, shouldn't it?
>>
>> Not with clang ubsan (okay, the failure is at runtime, not compile time):
>>
>> https://lists.gnu.org/archive/html/qemu-devel/2018-03/msg05143.html
>> https://lists.gnu.org/archive/html/qemu-devel/2018-03/msg05148.html
>>
>> Even when the offset is 0, the mere fact that you are computing an
>> offset relative to a NULL pointer is undefined behavior.
>
> Are there cases where we are passing NULL to qobject_{inc,dec}ref?  They
> currently have an "if" in they're body, but my opinion is they ought to
> crash and burn...

I find it convenient that unref() accepts NULL, just like free().

However, I agree than ref() should crash if given a NULL pointer.

While at it, I'd also prefer ref() to return a pointer to the same object...