Re: [Qemu-devel] [PATCH v6 05/11] vhost-user: add vhost_user_gpu_set_socket()

[Michael S. Tsirkin]

On Mon, Apr 29, 2019 at 09:12:43AM +0200, Gerd Hoffmann wrote:
>   Hi,
> 
> > > What questions for example?
> > 
> > This opens up different kind of possible replies, and error handling.
> > 
> > With current proposal and needs, the reply (or absence of reply) is
> > entirely driven by the request.
> > 
> > With your proposal, should all request have a reply?
> 
> Yes.
> 
> > which makes a lot
> > more code synchronous,
> 
> Why?  You don't have to wait for the reply before sending the next
> request.
> 
> Adding a request id to the messages might be useful, so it is possible
> to wait for a reply to a specific message without having to keeping
> track of all in-flight messages.
> 
> > and complicates both sides unnecessarily.
> 
> Having headers in the reply allows it to process them in generic code.
> There is a size header for the reply, so you can parse the stream
> without knowing what replay to expect.  You can use the status field to
> indicate the payload, simliar to virtio-gpu which has response code
> OK_NODATA, some OK_$whatpayload and some ERR_$failure codes.
> 
> You can dispatch based on the response/status code and run *fully*
> asynchronous without too much trouble.
> 
> > > > Can we leave that for future protocol extensions negotiated with
> > > > GET/SET_PROTOCOL_FEATURES ?
> > >
> > > I don't think negotiating such a basic protocol change is a good idea.
> > 
> > Well, then I would rather focus on improving protocol negociation,
> > rather than adding unnecessary protocol changes.
> > 
> > Given that GET/SET_PROTOCOL_FEATURES is the first messages being sent,
> > why couldn't it have flags indicating new protocol revision?
> 
> A properly structed reply allows a different approach in reply
> processing (see above).  But that only works if it is in the protocol
> right from the start.  As add-on feature it can't provide the benefits
> because the reply parser must be able to handle both protocol variants.
> 
> cheers,
>   Gerd

I think it can in theory - but if we know we want a feature we should
just add it as mandatory. More options does imply more overhead.

-- 
MST