Hi, the MEI softirq does not use an irq domain, so the NUMA node cannot be inferred from the node of the controller, and the map function does not automatically clear the IRQ_NOREQUEST flag.
On x86, the IRQ_NOREQUEST flag is not set by platform code on allocation, but that appears to be an oversight; other platforms set it on allocation, and expect either irq_domain_associate_locked() or the device setup code to clear it. The flags are probably worth backporting (without that patch, the GSC is unavailable because request_irq() fails), the NUMA patch is probably not important enough (but doesn't hurt either). Simon Simon Richter (2): drm/xe: allow request_irq on GSC interrupt drm/xe: Use same NUMA node for MEI interrupt drivers/gpu/drm/xe/xe_heci_gsc.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) -- 2.47.3
