Re: [Qemu-devel] RFC: ioapic polarity vs. qemu os-x guest

[Michael S. Tsirkin]

On Sun, Feb 16, 2014 at 07:47:00AM -0700, Alex Williamson wrote:
> On Sun, 2014-02-16 at 13:41 +0200, Michael S. Tsirkin wrote:
> > On Fri, Feb 14, 2014 at 11:13:04PM +0100, Alexander Graf wrote:
> > > 
> > > On 14.02.2014, at 23:06, Gabriel L. Somlo <address@hidden> wrote:
> > > 
> > > > On Fri, Feb 14, 2014 at 10:21:09PM +0100, Alexander Graf wrote:
> > > >> 
> > > >> Can't you just turn the polarity around in the pci host adapter?
> > > > 
> > > > I tried this:
> > > > 
> > > > diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> > > > index 1221f32..0e86d21 100644
> > > > --- a/hw/pci/pci.c
> > > > +++ b/hw/pci/pci.c
> > > > @@ -118,13 +118,13 @@ static int pci_bar(PCIDevice *d, int reg)
> > > > 
> > > > static inline int pci_irq_state(PCIDevice *d, int irq_num)
> > > > {
> > > > -       return (d->irq_state >> irq_num) & 0x1;
> > > > +       return !(d->irq_state >> irq_num) & 0x1;
> > > > }
> > > > 
> > > > static inline void pci_set_irq_state(PCIDevice *d, int irq_num, int 
> > > > level)
> > > > {
> > > >         d->irq_state &= ~(0x1 << irq_num);
> > > > -       d->irq_state |= level << irq_num;
> > > > +       d->irq_state &= ~(level << irq_num);
> > > > }
> > > > 
> > > > static void pci_change_irq_level(PCIDevice *pci_dev, int irq_num, int 
> > > > change)
> > > > @@ -229,7 +229,7 @@ static void pcibus_reset(BusState *qbus)
> > > >     }
> > > > 
> > > >     for (i = 0; i < bus->nirq; i++) {
> > > > -        assert(bus->irq_count[i] == 0);
> > > > +        assert(bus->irq_count[i] != 0);
> > > >     }
> > > > }
> > > > 
> > > > ---
> > > > 
> > > > but now OS X freezes during boot right after
> > > > 
> > > >         [ PCI configuration begin ]
> > > >         [ PCI configuration end, bridges 1, devices 10 ]
> > > >         RTC: Only single RAM bank (128 bytes)
> > > > 
> > > > which all looks normal, except the process is supposed to continue on
> > > > from there and doesn't :)
> > > > 
> > > > On Linux, I get Fedora 20 live all the way up with no obvious/loud
> > > > complaints, but mouse and keyboard don't work at all...
> > > > 
> > > > I have to admit I'm a bit out of my depth here, though :)
> > > 
> > > Yeah, another thing we have to take into account is vhost-net which 
> > > generates IRQs directly through irqfd. I guess for those we'll have to 
> > > configure the polarity in the irq routing table?
> > > 
> > > 
> > > Alex
> > 
> > What will be affected is VFIO which uses IRQFD
> > for level interrupts with KVM_IRQFD_FLAG_RESAMPLE.
> > I suspect this will need a kernel change, maybe
> > a new flag for IRQFD: KVM_IRQFD_FLAG_ACTIVE_LOW,
> > since at the moment that does:
> > 
> > static void
> > irqfd_inject(struct work_struct *work)
> > {
> >         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
> >         struct kvm *kvm = irqfd->kvm;
> > 
> >         if (!irqfd->resampler) {
> >                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
> >                                 false);
> >                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
> >                                 false);
> >         } else
> >                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
> >                             irqfd->gsi, 1, false);
> > }
> 
> 
> 
> As you said in a previous message, devices just want assert & de-assert,
> 1 & 0, which is what we have here.  I would think that what asserted
> means only needs to be interpreted at the IOAPIC, so I'd hope we could
> get it right w/o an API change.


Well there is a bigger issue: any interrupt with
multiple sources is broken.

__kvm_irq_line_state does a logical OR of all sources,
before XOR with polarity.

This makes no sense if polarity is active low.


One is beginning to think the simplest fix
would be Gabriel's patch after all:
-      irq_level ^= entry.fields.polarity;


although it's ugly in that it perpetuates the
bug in more places instead of fixing it.


>  Thanks,
> 
> Alex