On Wed, Jun 14, 2017 at 05:32:12PM +0200, Paolo Bonzini wrote:
>
>
> On 14/06/2017 17:28, Roman Kagan wrote:
> > On Wed, Jun 14, 2017 at 05:08:02PM +0200, Paolo Bonzini wrote:
> >>
> >>
> >> On 06/06/2017 20:19, Roman Kagan wrote:
> >>> + sint_route->msg_status = ret;
> >>> + /* notify the msg originator of the progress made; if the slot was
> >>> busy we
> >>> + * set msg_pending flag in it so it will be the guest who will do
> >>> EOM and
> >>> + * trigger the notification from KVM via sint_ack_notifier */
> >>> + if (ret != -EAGAIN) {
> >>> + qemu_bh_schedule(sint_route->msg_bh);
> >>> + }
> >>
> >> It may be faster to use aio_bh_schedule_oneshot, depending on the number
> >> of devices.
> >
> > Messages aren't used on fast paths, they are only exchanged at device
> > setup/teardown. So I cared more about readability than speed here.
>
> Then you really want to use aio_bh_schedule_oneshot, because bottom
> halves incur a (small) cost even when you don't use them: each iteration
> of the event loop visits the list of bottom halves.
I didn't realize that (yes that's easy to see in the code but the API
didn't suggest I needed to ;).
> Persistent bottom halves, thus, are only a good idea if they are
> expected to trigger very often on a busy VM. If this bottom half is
> only triggered at setup/teardown, it shouldn't use them.
Thanks for pointing that out; we'll have to adjust our vmbus code too as
we used a persistent bottom half for message interactions there too.
Roman.
