Re: [Qemu-devel] [PATCH 0/4] Tweaks around virtio-blk start/stop

[Cornelia Huck <cornelia.huck@de.ibm.com>]

On Wed, 16 Mar 2016 13:49:10 +0100
Paolo Bonzini <pbonzini@redhat.com> wrote:

> On 16/03/2016 13:42, Cornelia Huck wrote:
> > On Wed, 16 Mar 2016 13:32:59 +0100
> > Paolo Bonzini <pbonzini@redhat.com> wrote:
> > 
> >> On 16/03/2016 13:22, Cornelia Huck wrote:
> >>>>> Yeah, it doesn't help that the functions are underdocumented (as in the
> >>>>> "assign" parameter above).
> >>> My understanding is:
> >>>
> >>> - 'assign': set up a new notifier (true) or disable it (false)
> >>> - 'set_handler': use our handler (true) or have it handled elsewhere
> >>>   (false)
> >>
> >> Right.  So if we're setting up a new notifier in
> >> virtio_queue_aio_set_host_notifier_handler, virtio_pci_stop_ioeventfd should
> 
> ... not call virtio_queue_host_notifier_read.

This needs to be cascaded into
virtio_queue_set_host_notifier_fd_handler(), I guess.

> 
> >>>>> I don't think the ->set_host_notifiers() api really allows for this.
> >>>>
> >>>> I think it does, assign is the last argument to k->set_host_notifier().
> >>>
> >>> This depends on whether we want 'assign' to clean up any old notifiers
> >>> before setting up new ones. I think we want different behaviour for
> >>> dataplane and vhost.
> >>
> >> I think dataplane and vhost are the same.
> >>
> >> The question is whether ioeventfd=off,vhost=on or
> >> ioeventfd=off,dataplane=on are valid combinations; I think they aren't.
> > 
> > We should disallow that even temporary, then.
> > 
> > (This would imply that we need to drop the _stop_ioeventfd() call in
> > ->set_host_notifier(), correct?)
> 
> No, it would not.  ioeventfd=off,vhost=on would mean: "when vhost is
> off, use vCPU thread notification".

*confused*

Is ioeventfd=off supposed to mean "don't talk to the kernel, do
everything in qemu"?

> 
> When turning on vhost you'd still stop ioeventfd (i.e. stop processing
> the virtqueue in QEMU's main iothread), but you don't need to do
> anything to the event notifier.  vhost will pick it up and work on the
> virtqueue if necessary.  Likewise for dataplane.

So "disassociate the handler and switch over to the new one"?

> 
> >> If they aren't, it should be okay to remove the
> >> virtio_queue_host_notifier_read call in
> >> virtio_queue_set_host_notifier_fd_handler and
> >> virtio_queue_aio_set_host_notifier_handler.  That's because a handler
> >> for the notifier will always be set _somewhere_.  It could be the usual
> >> ioeventfd handler, the vhost handler or the dataplane handler, but one
> >> will be there.
> > 
> > It should; but we probably need to do a final read when we stop the
> > ioeventfd.
> 
> I was thinking of handing the final read directly to the next guy who
> polls the event notifier instead.  So, when called from vhost or
> dataplane, virtio_pci_stop_ioeventfd would use
> assign=true/set_handler=false ("a new notifier is going to be set up by
> the caller").

OK, then we'd need to pass a new parameter for this.

> 
> The host notifier API unfortunately is full of indirections.  I'm not
> sure how many of them are actually necessary.

Oh yes, it's very hard to follow, especially with not-very-well defined
parameters.