Re: [PATCH] kvm: don't clear level from irqfd

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Jun 27, 2012 at 08:41:57AM -0600, Alex Williamson wrote:
> On Wed, 2012-06-27 at 14:38 +0300, Michael S. Tsirkin wrote:
> > On Wed, Jun 27, 2012 at 01:22:54PM +0300, Gleb Natapov wrote:
> > > On Wed, Jun 27, 2012 at 01:20:24PM +0300, Michael S. Tsirkin wrote:
> > > > On Wed, Jun 27, 2012 at 01:03:46PM +0300, Gleb Natapov wrote:
> > > > > On Wed, Jun 27, 2012 at 12:13:43PM +0300, Michael S. Tsirkin wrote:
> > > > > > irqfd sets and clears the level immediately.
> > > > > > This used to be necessary for MSI interrupts but isn't anymore.
> > > > > > 
> > > > > > Note! We never officially supported irqfd for level
> > > > > > interrupts but it usually seemed to kind of work,
> > > > > > and this change will break it. There are no known
> > > > > > real users of that feature though - only old vfio prototypes.
> > > > > > Removing this line will make it easier to add
> > > > > > real support for level irqfds in the future.
> > > > > > 
> > > > > Did we officially supported irqfd for ioapic/pic edge interrupts?
> > > > > Because those will break to.
> > > > 
> > > > Hard to say - as Alex points out it was never documented.
> > > If we will start dropping everything that was never documented we will
> > > hardly have working VMM at the end :)
> > 
> > Generally true.
> > 
> > However I just realized that this interface does not work
> > reliably for ioapic/pic edge either. More or less for the same reason that
> > it does not work for edge: we set bit in IRR, try to inject,
> > then clear even if injection failed (e.g. because it is low
> > priority). Guest will never get an interrupt in this case.
> > 
> > So what happened e.g. for Alex's vfio prototype is he never
> > hit an interrupt window in light testing.
> > 
> > What we would need to do to fix it is:
> > > > > > -	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
> > > > > > -	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
> > > > > > +	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
> > > > > > +	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
> > 
> > Which will fix edge but break level same as this patch.
> > But is it worth fixing? Maybe declare it broken and be
> > done with it?
> 
> Ugh, so if I understand correctly, the case where kvm_set_irq() returns
> <0, we're broken.

Gleb says it's not broken. Maybe I misread the code.
Need more investigation.

>  VFIO has masked the physical device and has no error
> path to know that the interrupt inject was unsuccessful.  We almost need
> an error eventfd that bounces the interrupt back to userspace for the
> qemu interrupt model to handle, but that has it's own set of
> complications.  Suggestions?  Declaring the interface broken isn't very
> useful.  Thanks,
> 
> Alex