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

[Gabriel L. Somlo]

Hi,

I'm trying to get OS X to work as a QEMU guest, and one of the few
remaining "mysteries" I need to solve is that the OS X guest hangs
during boot, waiting for its boot disk to be available, unless the
following KVM patch is applied:


diff --git a/virt/kvm/ioapic.c b/virt/kvm/ioapic.c
index ce9ed99..1539d37 100644
--- a/virt/kvm/ioapic.c
+++ b/virt/kvm/ioapic.c
@@ -328,7 +328,6 @@ int kvm_ioapic_set_irq(struct kvm_ioapic *ioapic, int irq, 
int irq_source_id,
        irq_level = __kvm_irq_line_state(&ioapic->irq_states[irq],
                                         irq_source_id, level);
        entry = ioapic->redirtbl[irq];
-       irq_level ^= entry.fields.polarity;
        if (!irq_level) {
                ioapic->irr &= ~mask;
                ret = 1;
--


After digging around the KVM source for a bit, and printk-ing things
from Windows 7, Fedora 20, and OS X (10.9), I figured out the following:


1. Edge-triggered interrupts are invariably unaffected by the xor line
being removed by the patch. On all three guest types, edge-triggered
interrupts have polarity set to 0, so the xor is essentially a no-op,
and we can forget about it altogether.


2. Windows and Linux always configure all level-triggered interrupts
with polarity 0 (active-high, consistent with QEMU's ACPI/DSDT, in
particular q35-acpi-dsdt.dsl, which is what I'm using with -M q35).
As such, on Windows and Linux, the xor line in question is still a
no-op.


3. OS X (all versions I tried, at least since 10.5/Leopard) always
configures all level-triggered interrupts with polarity 1 (active-low),
regardless of what the QEMU DSDT says. As such, the xor line acts as
a negation of "irq_level", which at first glance sounds reasonable.

However: when KVM negates "irq_level" due to "polarity == 1", the OS X
guest hangs during boot.

OS X works fine when "polarity == 1" is ignored (with the xor line
commented out).

This may be another instance (similar to how OS X didn't use to check
with CPUID regarding monitor/mwait instruction availability) where
apple devs know that any of their supported hardware advertises
active-low in the DSDT, so no need to check, just hardcode that
assumption... :)


4. With s/ActiveHigh/ActiveLow/ in QEMU's q35-acpi-dsdt.dsl, Linux
actually switches to "polarity == 1" (active-low), and works fine
*with the xor line removed* !!!. With the xor line left intact (i.e.
without the above patch), the active-low fedora guest worked extremely
poorly, and printed out multiple error messages during boot:

        irq XX: nobody cared (try booting with the "irqpoll" option)
        ...
        Disabling IRQ #XX

for XX in [16, 18, 19, ...].


So, right now, I'm wondering about the following:


1. Regarding KVM and the polarity xor line in the patch above: Does
anyone have experience with any *other* guests which insist on setting
level-triggered interrupt polarity to 1/active-low ? Is that xor line
actually doing anything useful in practice, for any other guest, on
either QEMU or any other platform ?


2. Is there anything in QEMU (besides the ACPI DSDT .dsl files) which
has a hardcoded assumption re. "polarity == 0", or active-high, for
level-triggered interrupts? I tried to dig through hw/i386/kvm/ioapic.c
and a bunch of other files, but couldn't isolate anything that I could
"flip" to fix things in userspace.


Any ideas or suggestions about the appropriate way to move forward would
be much appreciated !!!


Thanks much,
--Gabriel