irq: mips: avoid nested irq_enter()
authorMark Rutland <mark.rutland@arm.com>
Wed, 20 Oct 2021 16:25:22 +0000 (17:25 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 18 Nov 2021 18:16:40 +0000 (19:16 +0100)
[ Upstream commit c65b52d02f6c1a06ddb20cba175ad49eccd6410d ]

As bcm6345_l1_irq_handle() is a chained irqchip handler, it will be
invoked within the context of the root irqchip handler, which must have
entered IRQ context already.

When bcm6345_l1_irq_handle() calls arch/mips's do_IRQ() , this will nest
another call to irq_enter(), and the resulting nested increment to
`rcu_data.dynticks_nmi_nesting` will cause rcu_is_cpu_rrupt_from_idle()
to fail to identify wakeups from idle, resulting in failure to preempt,
and RCU stalls.

Chained irqchip handlers must invoke IRQ handlers by way of thee core
irqchip code, i.e. generic_handle_irq() or generic_handle_domain_irq()
and should not call do_IRQ(), which is intended only for root irqchip
handlers.

Fix bcm6345_l1_irq_handle() by calling generic_handle_irq() directly.

Fixes: c7c42ec2baa1de7a ("irqchips/bmips: Add bcm6345-l1 interrupt controller")
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Marc Zyngier <maz@kernel.org>
Acked-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/irqchip/irq-bcm6345-l1.c

index e348378..1bd0621 100644 (file)
@@ -140,7 +140,7 @@ static void bcm6345_l1_irq_handle(struct irq_desc *desc)
                for_each_set_bit(hwirq, &pending, IRQS_PER_WORD) {
                        irq = irq_linear_revmap(intc->domain, base + hwirq);
                        if (irq)
-                               do_IRQ(irq);
+                               generic_handle_irq(irq);
                        else
                                spurious_interrupt();
                }