[PATCH v3 5/5] arm64: Enable perf events based hard lockup detector

[Petr Mladek]

On Tue 2022-04-05 20:53:04, Lecopzer Chen wrote:
>  
> > On Thu 2022-03-24 22:14:05, Lecopzer Chen wrote:
> > > With the recent feature added to enable perf events to use pseudo NMIs
> > > as interrupts on platforms which support GICv3 or later, its now been
> > > possible to enable hard lockup detector (or NMI watchdog) on arm64
> > > platforms. So enable corresponding support.
> > > 
> > > One thing to note here is that normally lockup detector is initialized
> > > just after the early initcalls but PMU on arm64 comes up much later as
> > > device_initcall(). To cope with that, overriding watchdog_nmi_probe() to
> > > let the watchdog framework know PMU not ready, and inform the framework
> > > to re-initialize lockup detection once PMU has been initialized.
> > > 
> > > [1]: http://lore.kernel.org/linux-arm-kernel/1610712101-14929-1-git-send-email-sumit.garg@linaro.org
> > > 
> > > --- /dev/null
> > > +++ b/arch/arm64/kernel/watchdog_hld.c
> > > @@ -0,0 +1,37 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +#include <linux/nmi.h>
> > > +#include <linux/cpufreq.h>
> > > +#include <linux/perf/arm_pmu.h>
> > > +
> > > +/*
> > > + * Safe maximum CPU frequency in case a particular platform doesn't implement
> > > + * cpufreq driver. Although, architecture doesn't put any restrictions on
> > > + * maximum frequency but 5 GHz seems to be safe maximum given the available
> > > + * Arm CPUs in the market which are clocked much less than 5 GHz. On the other
> > > + * hand, we can't make it much higher as it would lead to a large hard-lockup
> > > + * detection timeout on parts which are running slower (eg. 1GHz on
> > > + * Developerbox) and doesn't possess a cpufreq driver.
> > > + */
> > > +#define SAFE_MAX_CPU_FREQ	5000000000UL // 5 GHz
> > > +u64 hw_nmi_get_sample_period(int watchdog_thresh)
> > > +{
> > > +	unsigned int cpu = smp_processor_id();
> > > +	unsigned long max_cpu_freq;
> > > +
> > > +	max_cpu_freq = cpufreq_get_hw_max_freq(cpu) * 1000UL;
> > > +	if (!max_cpu_freq)
> > > +		max_cpu_freq = SAFE_MAX_CPU_FREQ;
> > > +
> > > +	return (u64)max_cpu_freq * watchdog_thresh;
> > > +}
> > 
> > This change is not mentioned in the commit message.
> > Please, put it into a separate patch.
> 
> 
> Actully, This cames from
> [1]: http://lore.kernel.org/linux-arm-kernel/1610712101-14929-1-git-send-email-sumit.garg@linaro.org
> And I didn't touch the commit message from the origin patch.
> But of course, I could imporve it with proper description if
> anyone thinks it's not good enough.

I see.

> Would you mean put this function hw_nmi_get_sample_period() in patch
> 6th?
> In the view of "arm64 uses delayed init with all the functionality it need to set up",
> IMO, this make sense for me to put into a single patch.

Or you could split it in two patches and add
hw_nmi_get_sample_period() in the earlier patch.


> But if you still think this should put into a separate patch, I'll do it:)

It is always better to split the changes whenever possible. It makes
the review easier. And it also helps to find the real culprit of
a regression using bisection.


> > > +int __init watchdog_nmi_probe(void)
> > > +{
> > > +	if (!allow_lockup_detector_init_retry)
> > > +		return -EBUSY;
> > 
> > How do you know that you should return -EBUSY
> > when retry in not enabled?
> > 
> > I guess that it is an optimization to make it fast
> > during the first call. But the logic is far from
> > obvious.
> > 
> 
> Yes, you can see this as an optimization, because arm64 PMU is not ready
> during lockup_detector_init(), so the watchdog_nmi_probe() must fail.
>
> Thus we only want to do watchdog_nmi_probe() in delayed init,
> so if not in the state (allow_lockup_detector_init_retry=true), just tell
> 
> if it's unclear

Yes, it is far from obvious.

> maybe a brief comment can be add like this:
> 
> +	/* arm64 is only able to initialize lockup detecor during delayed init */
> +	if (!allow_lockup_detector_init_retry)
> +		return -EBUSY;

No, please, remove this optimization. It just makes problems:

   + it requires a comment here because the logic is far from obvious.

   + it is the reason why we need another variable to avoid the race in
     lockup_detector_check(), see the discussion about the 4th patch.


> > > +
> > > +	if (!arm_pmu_irq_is_nmi())
> > > +		return -ENODEV;
> > > +
> > > +	return hardlockup_detector_perf_init();
> > > +}
> > 
> For arm_pmu_irq_is_nmi() checking, we do need it, becasue arm64 needs
> explictly turns on Pseudo-NMI to support base function for NMI.
>
> hardlockup_detector_perf_init() will success even if we haven't had
> Pseudo-NMI turns on, however, the pmu interrupts will act like a
> normal interrupt instead of NMI and the hardlockup detector would be broken.

I see. Please, explain this in a comment. It is another thing
that is far from obvious.

Best Regards,
Petr