On Fri, Jul 10, 2020 at 06:20:17PM +0200, Stefano Garzarella wrote:
> On Fri, Jul 10, 2020 at 11:33:09AM -0400, Konrad Rzeszutek Wilk wrote:
> > .snip..
> > > Just to recap the proposal, the idea is to add some restrictions to the
> > > operations (sqe, register, fixed file) to safely allow untrusted applications
> > > or guests to use io_uring queues.
> > 
> > Hi!
> > 
> > This is neat and quite cool - but one thing that keeps nagging me is
> > what how much overhead does this cut from the existing setup when you use
> > virtio (with guests obviously)?
> 
> I need to do more tests, but the preliminary results that I reported on
> the original proposal [1] show an overhead of ~ 4.17 uS (with iodepth=1)
> when I'm using virtio ring processed in a dedicated iothread:
> 
>   - 73 kIOPS using virtio-blk + QEMU iothread + io_uring backend
>   - 104 kIOPS using io_uring passthrough
> 
> >                                 That is from a high level view the
> > beaty of io_uring being passed in the guest is you don't have the
> > virtio ring -> io_uring processing, right?
> 
> Right, and potentially we can share the io_uring queues directly to the
> guest userspace applications, cutting down the cost of Linux block
> layer in the guest.

Another factor is that the guest submits requests directly to the host
kernel sqpoll thread. When a virtqueue is used the sqpoll thread cannot
poll it directly so another host thread (QEMU) needs to poll the
virtqueue. The same applies for the completion code path.

Stefan