Re: Re: Re: Re: [PATCH v2 1/2] virtio: abstract virtqueue related methods

[zhenwei pi]

On 5/18/23 18:09, Michael S. Tsirkin wrote:
> On Thu, May 18, 2023 at 08:47:22AM +0800, zhenwei pi wrote:
> > On 5/17/23 18:39, Michael S. Tsirkin wrote:
> > > On Wed, May 17, 2023 at 04:35:55PM +0800, zhenwei pi wrote:
> > > > On 5/17/23 15:46, Christoph Hellwig wrote:
> > > > > On Wed, May 17, 2023 at 03:43:03PM +0800, zhenwei pi wrote:
> > > > > > I have a plan to introduce 'Virtio Over Fabrics'(TCP&RDMA) as Virtio
> > > > > > transport, as mentioned in cover letter of this series:
> > > > > > 3 weeks ago, I posted a proposal 'Virtio Over Fabrics':
> > > > > > https://lists.oasis-open.org/archives/virtio-comment/202304/msg00442.html
> > > > >
> > > > > Just don't do it.  Please define your own protocols over RDMA or TCP
> > > > > for exactly the operations you need (for many they will already exist)
> > > > > instead of piggyg backing on virtio and making everyone else pay the
> > > > > price.
> > > >
> > > > Hi
> > > >
> > > > 1, `virtqueue_add_inbuf` in current version:
> > > > static inline int virtqueue_add_inbuf(struct virtqueue *vq,
> > > >                                         struct scatterlist *sg,
> > > >                                         unsigned int num,
> > > >                                         void *data,
> > > >                                         gfp_t gfp)
> > > > {
> > > >           if (likely(!vq->abstract))
> > > >                   return vring_virtqueue_add_sgs(vq, &sg, num, 0, 1, data,
> > > > NULL, gfp);
> > > >
> > > >           return vq->add_sgs(vq, &sg, num, 0, 1, data, NULL, gfp);
> > > > }
> > > >
> > > > And disassemble 'virtinput_queue_evtbuf':
> > > > static void virtinput_queue_evtbuf(struct virtio_input *vi,
> > > >                                      struct virtio_input_event *evtbuf)
> > > > {
> > > >           struct scatterlist sg[1];
> > > >
> > > >           sg_init_one(sg, evtbuf, sizeof(*evtbuf));
> > > >           virtqueue_add_inbuf(vi->evt, sg, 1, evtbuf, GFP_ATOMIC);
> > > > }
> > > >
> > > > I notice that two instructions are newly added for vring like:
> > > >    24d:   80 78 35 00             cmpb   $0x0,0x35(%rax)
> > > >    251:   75 3f                   jne    292
> > > >
> > > > Is it an expensive price...
> > >
> > > Can we somehow only override the kick method?
> > > Then take the ring and send it over ...
> >
> > Could you please take a look at this code?
> > https://github.com/pizhenwei/linux/blob/virtio-of-github/drivers/virtio/virtio_fabrics.c#LL861C13-L861C23
>
> what am I looking at here?
>
> Looks like at least vof_handle_vq duplicates some code from vringh.
> But besides that yes, that's the idea.

OK, I'd drop this series.

Cc Jason & Stefan.


--
zhenwei pi