Re: [Qemu-devel] [RFC] QOMification of AXI stream

["Andreas Färber" <afaerber@suse.de>]

Am 09.06.2012 03:53, schrieb Peter Crosthwaite:
> On Fri, Jun 8, 2012 at 7:45 PM, Andreas Färber <afaerber@suse.de> wrote:
>> Am 08.06.2012 06:23, schrieb Peter Crosthwaite:
>>> Each of the two core has three interfaces (+interrupt pins):
>>>
>>> 1: Sysbus attachment for device control
>>> 2: AXI stream TX link
>>> 3: AXI stream RX link
>> [...]
>>> struct XilinxDMAConnection {
>>>     void *dma;
>>>     void *client;
>>>
>>>     DMAPushFn to_dma;
>>>     DMAPushFn to_client;
>>> };
>>>
>>> So what im proposing is AXI stream is implemented as a unidirectional
>>> point to point bus. The xilinx ethernet system would consist of two of
>>> these buses one for tx, one for rx.
>> [...]
>>> A: Make AXI_STREAM_SLAVE an interface (not a sub-class of DEVICE). Its
>>> kind of annoying though if someone in the future whats the create a
>>> device thats only and axi stream slave, as they would have to
>>> explicitly inherit from DEVICE as well.
>>>
>>> or
>>>
>>> B: Have the slave attachment be a device within a device. Hard part is
>>> getting an accessor so machine models can retrieve the slave
>>> attachment and hook it up.
>>
>> If you dive into busses, note that Anthony has refactored QBus on
>> qom-next branch.
>>
> 
> How stable is this branch? It seems like I should use it as the
> development point. Is the merge immenent. If the merge is delayed, can
> I at least rely on the fundamental APIs define here (around links and
> stuff) no changing?

At this point we're pretty close to merging (hopefully next week) so I
would advise against basing new work on that branch. Just be prepared to
rebase onto the "qdev: Convert busses to QEMU Object Model" patch, i.e.
BusInfo gets replaced by TypeInfo and creation uses TYPE_FOO.

>> As Paul has already mentioned, the concept of tree-structured qdev
>> busses is deprecated by QOM in favor of link<> properties.
> 
> Ive had a brief look at the refactorings on qom-next, I notice that
> busses are now just children of the parent object TYPE_BUS.
> Essentially for point-to-point links this means that link itself has a
> QOM object. So for finer clarification, for new busses should or
> should I not have an object (whether it inheritTYPE_BUS or some other
> parent) for the link itself? Or should The master device interface
> directly with its slave? Im thinking the latter, no need for an object
> for a trivial point-to-point link.

No bus expert myself, deferring to Anthony and Paolo.

> Heres what i'm thinking now. each device will
> 
> Inherit from SYSBUS
> implement interface AXI_STREAM_SLAVE
> have a link property "axi_stream_connected_slave"
> 
> AXI_STREAM_SLAVE has a single function to push data down the link
> (what I believe you called DMAPushFn), but I will rename to
> axi_stream_push or the like as its not DMA specific.
> 
> Machine model then just sets axi_stream_connected_slave to each other.

Doesn't sound wrong so far under the premise of that simplistic
modelling approach. Not that I'm specifically advocating this approach.

>> That would of course limit the number of channels to one. Otherwise you
>> need a dedicated child<> object, of which a device can have multiple.
> 
> Im not too worried about that, but Peter and Paul have opened the
> discussion. Is the straight up interface on the sysbus device fine for
> what im trying to do - or should I have proxy objects for the sake of
> consistency?

I'm not aware of any use of interfaces in upstream nor of any proxy
object. In the end it'll be a compromise between fancy and quick... ;)

Regards,
Andreas

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg