Mixing IRQF_NO_SUSPEND and !IRQF_NO_SUSPEND on the same IRQ line is highly
discouraged, but in some cases (IRQ shared by a timer and other devices)
you don't have any other choice.
Since some devices sharing the IRQ line might tag it as a wakeup source,
you might end up with your handler that requested IRQF_NO_SUSPEND not
being called in suspended state, or invalid system wakeup (the system is
woken up without any wakeup source actually requesting it).

To deal with such unlikely situations, you'll have to:
1/ prevent any automatic wakeup when at least one of the IRQ users
   registered with IRQF_NO_SUSPEND
2/ let IRQ users decide if/when they should wake the system up

This patch is taking care of 1.

Signed-off-by: Boris Brezillon <redacted>
---
 kernel/irq/pm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/irq/pm.c b/kernel/irq/pm.c
index 3ca5325..1743162 100644
--- a/kernel/irq/pm.c
+++ b/kernel/irq/pm.c

@@ -16,7 +16,8 @@
 
 bool irq_pm_check_wakeup(struct irq_desc *desc)
 {
-	if (irqd_is_wakeup_armed(&desc->irq_data)) {
+	if (irqd_is_wakeup_armed(&desc->irq_data) &&
+	    !desc->no_suspend_depth) {
 		irqd_clear(&desc->irq_data, IRQD_WAKEUP_ARMED);
 		desc->istate |= IRQS_SUSPENDED | IRQS_PENDING;
 		desc->depth++;

-- 
1.9.1