Message-ID: <5541A542.4050508@linux.vnet.ibm.com>
Date:	Thu, 30 Apr 2015 09:15:06 +0530
From:	Preeti U Murthy <preeti@...ux.vnet.ibm.com>
To:	"Rafael J. Wysocki" <rjw@...ysocki.net>,
	Sudeep Holla <sudeep.holla@....com>,
	Peter Zijlstra <peterz@...radead.org>
CC:	Linus Walleij <linus.walleij@...aro.org>,
	"Rafael J. Wysocki" <rafael@...nel.org>,
	Daniel Lezcano <daniel.lezcano@...aro.org>,
	Linux PM list <linux-pm@...r.kernel.org>,
	Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...nel.org>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	ACPI Devel Maling List <linux-acpi@...r.kernel.org>
Subject: Re: [PATCH 16/20] sched/idle: Use explicit broadcast oneshot control
 function

On 04/29/2015 06:34 AM, Rafael J. Wysocki wrote:
> On Wednesday, April 29, 2015 02:50:22 AM Rafael J. Wysocki wrote:
>> On Tuesday, April 28, 2015 02:58:37 PM Sudeep Holla wrote:
>>>
>>> On 28/04/15 15:14, Rafael J. Wysocki wrote:
>>>> On Tuesday, April 28, 2015 03:37:44 PM Rafael J. Wysocki wrote:
>>>>> On Tuesday, April 28, 2015 03:31:54 PM Rafael J. Wysocki wrote:
>>>>>> On Tuesday, April 28, 2015 02:37:10 PM Linus Walleij wrote:
>>>>>>> On Tue, Apr 28, 2015 at 2:19 PM, Rafael J. Wysocki <rafael@...nel.org> wrote:
>>>>>>>> Sudeep:
>>>>>>>>> At-least I observed issue only when I am using hardware broadcast timer.
>>>>>>>>> It doesn't hang when I am using hrtimer as broadcast timer in which case
>>>>>>>>> one of the cpu will be not enter deeper idle states that lose timer.
>>>>>>>>> I will rerun on v4.1-rc1 and post the complete log.
>>>>>>>>
>>>>>>>> So the bug here is that cpuidle_enter() enables interrupts, so the
>>>>>>>> assumption about them being not enabled made by
>>>>>>>> tick_broadcast_oneshot_control() is actually not valid.
>>>>>>>>
>>>>>>>> It looks like we need to acquire the clockevents_lock at least in this
>>>>>>>> particular case.  Let me see where to put it and I'll send a patch for
>>>>>>>> testing.
>>>>>>>
>>>>>>> Aha that looks very much like it. Put me on the patch and I'll
>>>>>>> take it for a spin.
>>>>>>
>>>>>> OK, so something like the below for starters (the _irqsave variant is used to
>>>>>> avoid adding one more WARN_ON(irqs_disabled()) in there).
>>>>>>
>>>>>> I haven't tested it, but then I can't reproduce the original issue in the
>>>>>> first place.
>>>>>
>>>>> Of course, the whole "broadcast" thing could be done from cpuidle_enter()
>>>>> in the first place, but then we could not avoid the problem with the cpuidle
>>>>> *callback* enabling interrupts possibly in there anyway (not to mention the
>>>>> "coupled" stuff).
>>>>
>>>> That said, if the given state is marked with CPUIDLE_FLAG_TIMER_STOP, I really
>>>> wouldn't expect it to re-enable interrupts on exit and the "coupled" thing
>>>> seems to be fundamentally at odds with that flag either.
>>>>
>>>> So it should be possible to move the "broadcast" logic into the cpuidle layer,
>>>> which I'm going to try to do.
>>>>
>>>
>>> Makes sense.
>>>
>>>> Please test the patch I've sent, though, as it should bring the code back to
>>>> where it was before the clockevents_notify() removal and it'd be good to verify
>>>> that.
>>>>
>>>
>>> I tested your patch and it works now. Anyways I am continuing to run
>>> stress tests on my board. I will report if I find any issues.
>>
>> Great, thanks!
>>
>> Below is the patch I came up with in the meantime.
>>
>> This moves the "switch to broadcast" timer logic into
>> cpuidle_enter_state() which allows tick_broadcast_exit() to be
>> called directly with interrupts disabled (as required), but
>> it also adds a fallback branch reflecting the 4.0 and earlier
>> behavior for idle states that enable interrupts on exit
>> from their ->enter callbacks.
>>
>> I'm not aware of any valid cases when CPUIDLE_FLAG_TIMER_STOP can be
>> set for such states, but people may try to add stuff like that in the
>> future, so it's better to catch that (hence the WARN_ON_ONCE) and do
>> our best to handle it gracefully anyway, IMO.
>>
>> The "if (entered_state == -EBUSY)" check is conservative.  It may
>> be better to do "if (entered_state < 0)" and fall back to the default
>> on all errors, but that's not what we do today (I guess the concern
>> would be "what if the state ->enter returns an error after entering
>> and exiting the idle state, in which case we may miss a wakeup event
>> if we fall back to the default").
> 
> Actually, if my understanding of things is correct (the local clock event
> device cannot go away from under code executed with interrupts disabled
> on the local CPU), the simplified one below should be sufficient.
> 
> ---
>  drivers/cpuidle/cpuidle.c |   16 ++++++++++++++++
>  kernel/sched/idle.c       |   16 ++--------------
>  2 files changed, 18 insertions(+), 14 deletions(-)
> 
> Index: linux-pm/kernel/sched/idle.c
> ===================================================================
> --- linux-pm.orig/kernel/sched/idle.c
> +++ linux-pm/kernel/sched/idle.c
> @@ -81,7 +81,6 @@ static void cpuidle_idle_call(void)
>  	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
>  	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
>  	int next_state, entered_state;
> -	unsigned int broadcast;
>  	bool reflect;
> 
>  	/*
> @@ -150,17 +149,6 @@ static void cpuidle_idle_call(void)
>  		goto exit_idle;
>  	}
> 
> -	broadcast = drv->states[next_state].flags & CPUIDLE_FLAG_TIMER_STOP;
> -
> -	/*
> -	 * Tell the time framework to switch to a broadcast timer
> -	 * because our local timer will be shutdown. If a local timer
> -	 * is used from another cpu as a broadcast timer, this call may
> -	 * fail if it is not available
> -	 */
> -	if (broadcast && tick_broadcast_enter())
> -		goto use_default;
> -
>  	/* Take note of the planned idle state. */
>  	idle_set_state(this_rq(), &drv->states[next_state]);
> 
> @@ -174,8 +162,8 @@ static void cpuidle_idle_call(void)
>  	/* The cpu is no longer idle or about to enter idle. */
>  	idle_set_state(this_rq(), NULL);
> 
> -	if (broadcast)
> -		tick_broadcast_exit();
> +	if (entered_state == -EBUSY)
> +		goto use_default;
> 
>  	/*
>  	 * Give the governor an opportunity to reflect on the outcome
> Index: linux-pm/drivers/cpuidle/cpuidle.c
> ===================================================================
> --- linux-pm.orig/drivers/cpuidle/cpuidle.c
> +++ linux-pm/drivers/cpuidle/cpuidle.c
> @@ -158,9 +158,18 @@ int cpuidle_enter_state(struct cpuidle_d
>  	int entered_state;
> 
>  	struct cpuidle_state *target_state = &drv->states[index];
> +	bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
>  	ktime_t time_start, time_end;
>  	s64 diff;
> 
> +	/*
> +	 * Tell the time framework to switch to a broadcast timer because our
> +	 * local timer will be shut down.  If a local timer is used from another
> +	 * CPU as a broadcast timer, this call may fail if it is not available.
> +	 */
> +	if (broadcast && tick_broadcast_enter())
> +		return -EBUSY;
> +
>  	trace_cpu_idle_rcuidle(index, dev->cpu);
>  	time_start = ktime_get();
> 
> @@ -169,6 +178,13 @@ int cpuidle_enter_state(struct cpuidle_d
>  	time_end = ktime_get();
>  	trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
> 
> +	if (broadcast) {
> +		if (WARN_ON_ONCE(!irqs_disabled()))
> +			local_irq_disable();
> +
> +		tick_broadcast_exit();
> +	}
> +
>  	if (!cpuidle_state_is_coupled(dev, drv, entered_state))
>  		local_irq_enable();
> 
> 

Looks good.

Reviewed-by: Preeti U Murthy <preeti@...ux.vnet.ibm.com>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pm" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists