[PATCHv2 6/8] arm_pmu: explicitly enable/disable SPIs at hotplug

[Mark Rutland]

On Mon, Feb 26, 2018 at 03:22:53PM +0000, Will Deacon wrote:
> On Mon, Feb 26, 2018 at 04:16:19PM +0100, Geert Uytterhoeven wrote:
> > On Mon, Feb 5, 2018 at 5:42 PM, Mark Rutland <mark.rutland at arm.com> wrote:
> > > To support ACPI systems, we need to request IRQs before CPUs are
> > > hotplugged, and thus we need to request IRQs before we know their
> > > associated PMU.
> > >
> > > This is problematic if a PMU IRQ is pending out of reset, as it may be
> > > taken before we know the PMU, and thus the IRQ handler won't be able to
> > > handle it, leaving it screaming.
> > >
> > > To avoid such problems, lets request all IRQs in a disabled state, and
> > > explicitly enable/disable them at hotplug time, when we're sure the PMU
> > > has been probed.
> > >
> > > Signed-off-by: Mark Rutland <mark.rutland at arm.com>
> > 
> > This is now commit 6de3f79112cc26bf in v4.16-rc3, and causes a BUG during
> > CPU offlining (e.g. during system suspend, or during boot with
> > CONFIG_ARM_PSCI_CHECKER=y).
> > 
> > With CONFIG_ARM_PSCI_CHECKER=y:
> > 
> > psci_checker: PSCI checker started using 6 CPUs
> > psci_checker: Starting hotplug tests
> > psci_checker: Trying to turn off and on again all CPUs
> > BUG: sleeping function called from invalid context at kernel/irq/manage.c:112
> > in_atomic(): 1, irqs_disabled(): 128, pid: 15, name: migration/1
> > no locks held by migration/1/15.
> > irq event stamp: 192
> > hardirqs last  enabled at (191): [<00000000803c2507>]
> > _raw_spin_unlock_irq+0x2c/0x4c
> > hardirqs last disabled at (192): [<000000007f57ad28>] multi_cpu_stop+0x9c/0x140
> > softirqs last  enabled at (0): [<0000000004ee1b58>]
> > copy_process.isra.77.part.78+0x43c/0x1504
> > softirqs last disabled at (0): [<          (null)>]           (null)
> > CPU: 1 PID: 15 Comm: migration/1 Not tainted 4.16.0-rc3-salvator-x #1651
> > Hardware name: Renesas Salvator-X board based on r8a7796 (DT)
> > Call trace:
> >  dump_backtrace+0x0/0x140
> >  show_stack+0x14/0x1c
> >  dump_stack+0xb4/0xf0
> >  ___might_sleep+0x1fc/0x218
> >  __might_sleep+0x70/0x80
> >  synchronize_irq+0x40/0xa8
> >  disable_irq+0x20/0x2c
> 
> Given that these things are CPU-affine, I reckon this should be
> disable_irq_nosync. Mark?

Given IRQs are disabled, this should be fine, yes.

FWIW, if you spin this as a patch:

Acked-by: Mark Rutland <mark.rutland at arm.com>

Mark.

> 
> Will
> 
> --->8
> 
> diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c
> index 0c2ed11c0603..f63db346c219 100644
> --- a/drivers/perf/arm_pmu.c
> +++ b/drivers/perf/arm_pmu.c
> @@ -638,7 +638,7 @@ static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
>  		if (irq_is_percpu_devid(irq))
>  			disable_percpu_irq(irq);
>  		else
> -			disable_irq(irq);
> +			disable_irq_nosync(irq);
>  	}
>  
>  	per_cpu(cpu_armpmu, cpu) = NULL;