Re: [PATCH 3/4] kvm: Extend irqfd to support level interrupts

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 2012-06-25 at 02:02 +0300, Michael S. Tsirkin wrote:
> On Sun, Jun 24, 2012 at 03:59:27PM -0600, Alex Williamson wrote:
> > On Sun, 2012-06-24 at 18:49 +0300, Michael S. Tsirkin wrote:
> > > On Sun, Jun 24, 2012 at 09:18:38AM -0600, Alex Williamson wrote:
> > > > > > @@ -242,7 +299,8 @@ kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
> > > > > >  
> > > > > >  	ret = 0;
> > > > > >  	list_for_each_entry(tmp, &kvm->irqfds.items, list) {
> > > > > > -		if (irqfd->eventfd != tmp->eventfd)
> > > > > > +		if (irqfd->eventfd != tmp->eventfd &&
> > > > > > +		    irqfd->eventfd != tmp->eoi_eventfd)
> > > > > >  			continue;
> > > > > 
> > > > > So we allow duplicate irqfd with differing eoifd (or edge-triggered and
> > > > > level-triggered irqfd on the same context).
> > > > > 
> > > > > (why the check in the first place? just so we can have a reliable
> > > > > deassign or is it avoiding a deeper problem?)
> > > > 
> > > > I really wasn't sure to what extent we wanted to prevent duplicates.  My
> > > > guess was that we don't want to have an irqfd trigger more than one
> > > > thing.  That seems to be what the current code does.  I don't see any
> > > > problems with multiple irqfds triggering the same eventfd though.  I
> > > > only added a test that a new irqfd can't be triggered by an existing
> > > > eoi_eventfd as that could make a nasty loop.
> > > 
> > > How would that make a loop? You can have the same thing
> > > with e.g. ioeventfd - why isn't it a problem there?
> > 
> > eoi_eventfd1 -> irqfd2 [eoi] eoi_eventfd2 -> irqfd1 [eoi] eoi_eventfd1 ->...
> 
> Sorry I don't understand.
> What does this [eoi] mean? How is eoi eventfd different from ioeventfd?

[eoi] is simply the guest doing an EOI write.  There's some interaction
required from the guest which could rate limit the loop.  I'm not really
following your comparison to an ioeventfd.  If the question is can you
create loops with an ioeventfd, I imagine the answer is probably so.  I
certainly don't plan on adding code to test every fd in use by a vm and
validate it can't do crazy things, so if even this minor sanity test
could prevent useful things, I'll happily remove it.  Thanks,

Alex