Re: [PATCH] clocksource: exynos_mct: fix for sleeping in atomic ctx handling cpu hotplug notif.

[Marcin Jabrzyk <m.jabrzyk@samsung.com>]

On 30/01/15 15:21, Damian Eppel wrote:
> This is to fix an issue of sleeping in atomic context when processing
> hotplug notifications in Exynos MCT(Multi-Core Timer).
> The issue was reproducible on Exynos 3250 (Rinato board) and Exynos 5420
> (Arndale Octa board).
>
> Whilst testing cpu hotplug events on kernel configured with DEBUG_PREEMPT
> and DEBUG_ATOMIC_SLEEP we get following BUG message, caused by calling
> request_irq() and free_irq() in the context of hotplug notification
> (which is in this case atomic context).
>
> root@target:~# echo 0 > /sys/devices/system/cpu/cpu1/online
>
> [   25.157867] IRQ18 no longer affine to CPU1
> ...
> [   25.158445] CPU1: shutdown
>
> root@target:~# echo 1 > /sys/devices/system/cpu/cpu1/online
>
> [   40.785859] CPU1: Software reset
> [   40.786660] BUG: sleeping function called from invalid context at mm/slub.c:1241
> [   40.786668] in_atomic(): 1, irqs_disabled(): 128, pid: 0, name: swapper/1
> [   40.786678] Preemption disabled at:[<  (null)>]   (null)
> [   40.786681]
> [   40.786692] CPU: 1 PID: 0 Comm: swapper/1 Not tainted 3.19.0-rc4-00024-g7dca860 #36
> [   40.786698] Hardware name: SAMSUNG EXYNOS (Flattened Device Tree)
> [   40.786728] [<c0014a00>] (unwind_backtrace) from [<c0011980>] (show_stack+0x10/0x14)
> [   40.786747] [<c0011980>] (show_stack) from [<c0449ba0>] (dump_stack+0x70/0xbc)
> [   40.786767] [<c0449ba0>] (dump_stack) from [<c00c6124>] (kmem_cache_alloc+0xd8/0x170)
> [   40.786785] [<c00c6124>] (kmem_cache_alloc) from [<c005d6f8>] (request_threaded_irq+0x64/0x128)
> [   40.786804] [<c005d6f8>] (request_threaded_irq) from [<c0350b8c>] (exynos4_local_timer_setup+0xc0/0x13c)
> [   40.786820] [<c0350b8c>] (exynos4_local_timer_setup) from [<c0350ca8>] (exynos4_mct_cpu_notify+0x30/0xa8)
> [   40.786838] [<c0350ca8>] (exynos4_mct_cpu_notify) from [<c003b330>] (notifier_call_chain+0x44/0x84)
> [   40.786857] [<c003b330>] (notifier_call_chain) from [<c0022fd4>] (__cpu_notify+0x28/0x44)
> [   40.786873] [<c0022fd4>] (__cpu_notify) from [<c0013714>] (secondary_start_kernel+0xec/0x150)
> [   40.786886] [<c0013714>] (secondary_start_kernel) from [<40008764>] (0x40008764)
>
> Solution:
> Clockevent irqs cannot be requested/freed every time cpu is
> hot-plugged/unplugged as CPU_STARTING/CPU_DYING notifications
> that signals hotplug or unplug events are sent with both preemption
> and local interrupts disabled. Since request_irq() may sleep it is
> moved to the initialization stage and performed for all possible
> cpus prior putting them online. Then, in the case of hotplug event
> the irq asociated with the given cpu will simply be enabled.
> Similarly on cpu unplug event the interrupt is not freed but just
> disabled.
>
> Note that after successful request_irq() call for a clockevent device
> associated to given cpu the requested irq is disabled (via disable_irq()).
> That is to make things symmetric as we expect hotplug event as a next
> thing (which will enable irq again). This should not pose any problems
> because at the time of requesting irq the clockevent device is not
> fully initialized yet, therefore should not produce interrupts at that point.
>
> For disabling an irq at cpu unplug notification disable_irq_nosync() is
> chosen which is a non-blocking function. This again shouldn't be a problem as
> prior sending CPU_DYING notification interrupts are migrated away
> from the cpu.
>
> Fixes: 7114cd749a12 ("clocksource: exynos_mct: use (request/free)_irq calls for local timer registration")
> Signed-off-by: Damian Eppel <d.eppel@samsung.com>
> Cc: <stable@vger.kernel.org>
> Reported-by: Krzysztof Kozlowski <k.kozlowski@samsung.com>
> ---
>   drivers/clocksource/exynos_mct.c | 45 ++++++++++++++++++++++++++++------------
>   1 file changed, 32 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/clocksource/exynos_mct.c b/drivers/clocksource/exynos_mct.c
> index 83564c9..9beca58 100644
> --- a/drivers/clocksource/exynos_mct.c
> +++ b/drivers/clocksource/exynos_mct.c
> @@ -466,15 +466,12 @@ static int exynos4_local_timer_setup(struct clock_event_device *evt)
>   	exynos4_mct_write(TICK_BASE_CNT, mevt->base + MCT_L_TCNTB_OFFSET);
>
>   	if (mct_int_type == MCT_INT_SPI) {
> -		evt->irq = mct_irqs[MCT_L0_IRQ + cpu];
> -		if (request_irq(evt->irq, exynos4_mct_tick_isr,
> -				IRQF_TIMER | IRQF_NOBALANCING,
> -				evt->name, mevt)) {
> -			pr_err("exynos-mct: cannot register IRQ %d\n",
> -				evt->irq);
> +
> +		if (evt->irq == -1)
>   			return -EIO;
> -		}
> -		irq_force_affinity(mct_irqs[MCT_L0_IRQ + cpu], cpumask_of(cpu));
> +
> +		irq_force_affinity(evt->irq, cpumask_of(cpu));
> +		enable_irq(evt->irq);
>   	} else {
>   		enable_percpu_irq(mct_irqs[MCT_L0_IRQ], 0);
>   	}
> @@ -487,10 +484,12 @@ static int exynos4_local_timer_setup(struct clock_event_device *evt)
>   static void exynos4_local_timer_stop(struct clock_event_device *evt)
>   {
>   	evt->set_mode(CLOCK_EVT_MODE_UNUSED, evt);
> -	if (mct_int_type == MCT_INT_SPI)
> -		free_irq(evt->irq, this_cpu_ptr(&percpu_mct_tick));
> -	else
> +	if (mct_int_type == MCT_INT_SPI) {
> +		if (evt->irq != -1)
> +			disable_irq_nosync(evt->irq);
> +	} else {
>   		disable_percpu_irq(mct_irqs[MCT_L0_IRQ]);
> +	}
>   }
>
>   static int exynos4_mct_cpu_notify(struct notifier_block *self,
> @@ -522,7 +521,7 @@ static struct notifier_block exynos4_mct_cpu_nb = {
>
>   static void __init exynos4_timer_resources(struct device_node *np, void __iomem *base)
>   {
> -	int err;
> +	int err, cpu;
>   	struct mct_clock_event_device *mevt = this_cpu_ptr(&percpu_mct_tick);
>   	struct clk *mct_clk, *tick_clk;
>
> @@ -549,7 +548,27 @@ static void __init exynos4_timer_resources(struct device_node *np, void __iomem
>   		WARN(err, "MCT: can't request IRQ %d (%d)\n",
>   		     mct_irqs[MCT_L0_IRQ], err);
>   	} else {
> -		irq_set_affinity(mct_irqs[MCT_L0_IRQ], cpumask_of(0));
> +		for_each_possible_cpu(cpu) {
> +			int mct_irq = mct_irqs[MCT_L0_IRQ + cpu];
> +			struct mct_clock_event_device *pcpu_mevt =
> +				per_cpu_ptr(&percpu_mct_tick, cpu);
> +
> +			pcpu_mevt->evt.irq = -1;
> +
> +			if (request_irq(mct_irq,
> +					exynos4_mct_tick_isr,
> +					IRQF_TIMER | IRQF_NOBALANCING,
> +					pcpu_mevt->name, pcpu_mevt)) {
> +				pr_err("exynos-mct: cannot register IRQ (cpu%d)\n",
> +									cpu);
> +
> +				continue;
> +			}
> +			pcpu_mevt->evt.irq = mct_irq;
> +			irq_force_affinity(mct_irq, cpumask_of(cpu));
> +
> +			disable_irq(mct_irq);
> +		}
>   	}
>
>   	err = register_cpu_notifier(&exynos4_mct_cpu_nb);
>

Tested-by: Marcin Jabrzyk <m.jabrzyk@samsung.com>

Test hardware: Rinato B2 (Exynos 3250) board

I can confirm that this fixes bug reported here [1] "BUG appearing when 
trying to allocate interrupt on Exynos MCT after CPU hotplug".


[1] 
http://lists.infradead.org/pipermail/linux-arm-kernel/2014-October/296500.html

Best regards,
Marcin Jabrzyk

Samsung R&D Institute Poland
Samsung Electronics