Re: [virtio-dev] RFC: Doorbell suppression, packed-ring mode and hardware offload

[Jason Wang <jasowang@redhat.com>]

On 2019/2/15 äå12:23, Michael S. Tsirkin wrote:
> On Fri, Feb 15, 2019 at 11:59:55AM +0800, Jason Wang wrote:
> > On 2019/2/14 äå11:41, Michael S. Tsirkin wrote:
> > > > I think it's as simple as increase the avail idx by X? Since descriptor were
> > > > used in order, device can just read the next X-1 descriptors in this case.
> > > >
> > > > Thanks
> > > Right so a spec change would be needed, it's not transparent to guest.
> >
> > With the change, IN_ORDER + split_ring becomes something like submission
> > queue (descriptor ring) + completion queue (used ring). And used ring access
> > could be eliminated sometime, e.g for net for TX, we don't care about used
> > len.
> Oh yes but fundamentally this becomes very close to packed ring.


Looks not, it's a two rings vs one ring.


> So I'm
> not sure yet another option is justified by a small gain in PPS,


10% is done with less than 10 lines of code, I suspect maybe dpdk can 
see more.


> especially considering that it depends on in order and so doesn't
> support zero copy.


It looks to me in order is not the blocker, even if we allow out of 
order, we may still suffer from HOL.


> > What's more interesting is, this avail idx optimization is not required for
> > out of order completion which means it could be used for e.g block or SCSI
> > device.
> I don't get the last sentence. This only works for in order right?


Yes, you are right.

Thanks.