Hello Marc,
This patch is breaking the POWER9/POWER10 XIVE driver (these are not
old PPC systems :) on machines sharing the same LSI HW IRQ. For instance,
a linux KVM guest with a virtio-rng and a virtio-balloon device. In that
case, Linux creates two distinct IRQ mappings which can lead to some
unexpected behavior.
Either the irq domain translates, or it doesn't. If the driver creates
a nomap domain, and yet expects some sort of translation to happen,
then the driver is fundamentally broken. And even without that: how do
you end-up with a single HW interrupt having two mappings?
A fix to go forward would be to change the XIVE IRQ domain to use a
'Tree' domain for reverse mapping and not the 'No Map' domain mapping.
I will keep you updated for XIVE.
I bet there is a bit more to it. From what you are saying above,
something rather ungodly is happening in the XIVE code.
It's making progress.
This change in irq_find_mapping() is what 'breaks' XIVE :
+ if (irq_domain_is_nomap(domain)) {
+ if (hwirq < domain->revmap_size) {
+ data = irq_domain_get_irq_data(domain, hwirq);
+ if (data && data->hwirq == hwirq)
+ return hwirq;
+ }
+
+ return 0;
With the introduction of IRQ_DOMAIN_FLAG_NO_MAP, the revmap_tree lookup
is skipped and the previously mapped IRQ is not found. XIVE was relying
on a side effect of irq_domain_set_mapping() which is not true anymore.
I guess the easiest fix for 5.14 and 5.15 (in which was introduced MSI
domains) is to change the XIVE IRQ domain to a domain tree. Since the HW
can handle 1MB interrupts, this looks like a better choice for the driver.
Thanks,
C.
