hw/isa/ich9: Add comment on imperfect emulation of PIC vs. I/O APIC routing

As noted in the comment, the PCI INTx lines are supposed to be routed
to *both* the PIC and the I/O APIC. It's just that we don't cope with
the concept of an IRQ being asserted to two *different* pins on the
two irqchips.

So we have this hack of routing to I/O APIC only if the PIRQ routing to
the PIC is disabled. Which seems to work well enough, even when I try
hard to break it with kexec. But should be explicitly documented and
understood.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <112a09643b8191c4eae7d92fa247a861ab90a9ee.camel@infradead.org>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

hw/isa/lpc_ich9.c

@@ -304,6 +304,21 @@ static PCIINTxRoute ich9_route_intx_pin_to_irq(void *opaque, int pirq_pin)
             route.irq = -1;
         }
     } else {
+        /*
+         * Strictly speaking, this is wrong. The PIRQ should be routed
+         * to *both* the I/O APIC and the PIC, on different pins. The
+         * I/O APIC has a fixed mapping to IRQ16-23, while the PIC is
+         * routed according to the PIRQx_ROUT configuration. But QEMU
+         * doesn't (yet) cope with the concept of pin numbers differing
+         * between PIC and I/O APIC, and neither does the in-kernel KVM
+         * irqchip support. So we route to the I/O APIC *only* if the
+         * routing to the PIC is disabled in the PIRQx_ROUT settings.
+         *
+         * This seems to work even if we boot a Linux guest with 'noapic'
+         * to make it use the legacy PIC, and then kexec directly into a
+         * new kernel which uses the I/O APIC. The new kernel explicitly
+         * disables the PIRQ routing even though it doesn't need to care.
+         */
         route.irq = ich9_pirq_to_gsi(pirq_pin);
     }