[Qemu-devel] Re: [PATCH 06/41] virtio: Use DO_UPCAST instead of a cast

[Juan Quintela]

"Michael S. Tsirkin" <address@hidden> wrote:
> On Thu, Dec 03, 2009 at 12:56:57PM +0100, Juan Quintela wrote:
>> "Michael S. Tsirkin" <address@hidden> wrote:
>> > On Wed, Dec 02, 2009 at 08:03:22PM +0100, Juan Quintela wrote:
>> >> "Michael S. Tsirkin" <address@hidden> wrote:
>> >> 
>> >> > I don't understand.
>> >> > container_of is just more generic than DO_UPCAST.
>> >> > So why *ever* use DO_UPCAST? Let's get rid of it.
>> 
>> ....
>> 
>> >> See how you create a device of size struct_size, but then you access it
>> >> with vdev.  If vdev is _not_ the 1st element of the struct, you have got
>> >> corruption.
>> >
>> > A cleaner solution IMO would be to have callers allocate the memory
>> > and pass VirtIODevice * to virtio_common_init.
>> 
>> Been there, asked for that.  Basically qdev + passing initialized memory
>> = nono
>
> It does not have to be initialized.

sorry, already allocated memory.  basically -device foo
requires that qdev creates foo and then it call foo_init().

You can see in the mail archive that I wanted a qdev_create_here()
function, to do exactly what you wanted.  It has problems with getting
-device working realiablely.

>> >>  DO_UPCAST() prevent you for having that error.
>> >
>> >
>> > If we want to assert specific structure layout, this
>> > should be a compile-time check. There's no
>> > reason to do this check every time at a random place where
>> > DO_UPCAST is called.
>> 
>> See DO_UPCAST() definition :)
>> 
>> 
>> It is a compile time check.  It just do cpp magic to be sure that things
>> are right.  DO_UPCAST() == (cast *) with some typechecking.
>
> Yes, but it gives the error on the wrong place. So it is
> only good as long as you do not change the code.

No, it gives it in the right place.  Where you do a cast from

PCIDevice -> FOOState

or when you do a cast from VirtioDevice -> VirtIOFooDevice.

Really it is a "smarter" cast than just do (VirtioFooDevice *).

> In the example you give with virtio_init_common, there's no UPCAST
> to document the layout assumption *in the place where assumption is made*.
> On the other hand, DO_UPCAST is scattered all over the code
> in places which could work fine without any assumptions.
>
> Let's assume you want to change layout. You find all places
> with DO_UPCAST, fix them, and it will compile-but not work.
> Let's assume you change all code that makes layout assumptions
> to not make them: DO_UPCAST will still be hang around in other
> code.

Not requiring DO_UPCAST() mean changing how qdev work.  If you have
counted the Gerd qdev patches over the last months, you will see that it
is an almost "infinity" ammount of work.  I don't see why you want to
change that.  In this specific example, DO_UPCAST() is needed.  It is
required that VirtIODevice is the 1st field of any other structure.  I
really don't see why you want to change that.

Later, Juan.