Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <1599204397-17596-1-git-send-email-sumit.garg@linaro.org> <cd812613-03a6-30cd-74fd-25fdfcf28e42@arm.com>
In-Reply-To: <cd812613-03a6-30cd-74fd-25fdfcf28e42@arm.com>
From:   Sumit Garg <sumit.garg@linaro.org>
Date:   Thu, 1 Oct 2020 15:38:35 +0530
Message-ID: <CAFA6WYOmP-VPHxzCMF5Oy-jQB6nhycvAA3wO+BeYXB5icCO_TQ@mail.gmail.com>
Subject: Re: [PATCH v3] arm64: Enable perf events based hard lockup detector
To:     Alexandru Elisei <alexandru.elisei@arm.com>
Cc:     linux-arm-kernel <linux-arm-kernel@lists.infradead.org>,
        Catalin Marinas <catalin.marinas@arm.com>,
        Will Deacon <will@kernel.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Peter Zijlstra <peterz@infradead.org>, mingo@redhat.com,
        acme@kernel.org, alexander.shishkin@linux.intel.com,
        jolsa@redhat.com, namhyung@kernel.org,
        Thomas Gleixner <tglx@linutronix.de>,
        Stephen Boyd <swboyd@chromium.org>,
        julien.thierry.kdev@gmail.com,
        Douglas Anderson <dianders@chromium.org>,
        Daniel Thompson <daniel.thompson@linaro.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, 30 Sep 2020 at 18:33, Alexandru Elisei <alexandru.elisei@arm.com> wrote:
>
> Hi,
>
> Version 7 of the PMU NMI patches [1] has been picked up by Will, no major changes
> compared to v6.
>
> I would to try to review the PMU NMI bits, but I'm not familiar with how the
> watchdog functions. From my limited understanding, it uses an event that is reset
> periodically, and if it overflows, it triggers the watchdog, is that correct?

NMI watchdog basically uses perf event type: PERF_COUNT_HW_CPU_CYCLES
and registers watchdog_overflow_callback(). It doesn't reset perf
event but rather relies on the percpu watchdog_nmi_touch variable to
decide if the watchdog has actually expired or not when the perf event
is triggered. The defaut watchdog expiry timeout is 10 sec.

>
> [1] https://lkml.org/lkml/2020/9/24/458
>
> On 9/4/20 8:26 AM, Sumit Garg 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(). So we need to re-initialize lockup detection once
> > PMU has been initialized.
> >
> > Signed-off-by: Sumit Garg <sumit.garg@linaro.org>
> > ---
> >
> > Changes in v3:
> > - Rebased to latest pmu NMI patch-set [1].
> > - Addressed misc. comments from Stephen.
> >
> > [1] https://lkml.org/lkml/2020/8/19/671
> >
> > Changes since RFC:
> > - Rebased on top of Alex's WIP-pmu-nmi branch.
> > - Add comment for safe max. CPU frequency.
> > - Misc. cleanup.
> >
> >  arch/arm64/Kconfig             |  2 ++
> >  arch/arm64/kernel/perf_event.c | 41 +++++++++++++++++++++++++++++++++++++++--
> >  drivers/perf/arm_pmu.c         |  9 +++++++++
> >  include/linux/perf/arm_pmu.h   |  2 ++
> >  4 files changed, 52 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 6d23283..b5c2594 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -167,6 +167,8 @@ config ARM64
> >       select HAVE_NMI
> >       select HAVE_PATA_PLATFORM
> >       select HAVE_PERF_EVENTS
> > +     select HAVE_PERF_EVENTS_NMI if ARM64_PSEUDO_NMI
>
> This makes sense, as the PMU driver will use NMIs automatically if they are available.
>
> > +     select HAVE_HARDLOCKUP_DETECTOR_PERF if PERF_EVENTS && HAVE_PERF_EVENTS_NMI
> >       select HAVE_PERF_REGS
> >       select HAVE_PERF_USER_STACK_DUMP
> >       select HAVE_REGS_AND_STACK_ACCESS_API
> > diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> > index 5bf2835..2fb5b60 100644
> > --- a/arch/arm64/kernel/perf_event.c
> > +++ b/arch/arm64/kernel/perf_event.c
> > @@ -23,6 +23,8 @@
> >  #include <linux/platform_device.h>
> >  #include <linux/sched_clock.h>
> >  #include <linux/smp.h>
> > +#include <linux/nmi.h>
> > +#include <linux/cpufreq.h>
> >
> >  /* ARMv8 Cortex-A53 specific event types. */
> >  #define ARMV8_A53_PERFCTR_PREF_LINEFILL                              0xC2
> > @@ -1221,10 +1223,21 @@ static struct platform_driver armv8_pmu_driver = {
> >
> >  static int __init armv8_pmu_driver_init(void)
> >  {
> > +     int ret;
> > +
> >       if (acpi_disabled)
> > -             return platform_driver_register(&armv8_pmu_driver);
> > +             ret = platform_driver_register(&armv8_pmu_driver);
> >       else
> > -             return arm_pmu_acpi_probe(armv8_pmuv3_init);
> > +             ret = arm_pmu_acpi_probe(armv8_pmuv3_init);
>
> Shouldn't we return early here if the driver failed to bind instead of trying to
> initialize the lockup detector?

Makes sense, will only try to initialize the lockup detector if ret == 0.

>
> > +
> > +     /*
> > +      * Try to re-initialize lockup detector after PMU init in
> > +      * case PMU events are triggered via NMIs.
> > +      */
> > +     if (arm_pmu_irq_is_nmi())
> > +             lockup_detector_init();
> > +
> > +     return ret;
> >  }
> >  device_initcall(armv8_pmu_driver_init)
> >
> > @@ -1282,3 +1295,27 @@ void arch_perf_update_userpage(struct perf_event *event,
> >       userpg->cap_user_time_zero = 1;
> >       userpg->cap_user_time_short = 1;
> >  }
> > +
> > +#ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
> > +/*
> > + * 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;
> > +}
> > +#endif
> > diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c
> > index dd9d7f6..2cd0f40 100644
> > --- a/drivers/perf/arm_pmu.c
> > +++ b/drivers/perf/arm_pmu.c
> > @@ -718,6 +718,15 @@ static int armpmu_get_cpu_irq(struct arm_pmu *pmu, int cpu)
> >       return per_cpu(hw_events->irq, cpu);
> >  }
> >
> > +bool arm_pmu_irq_is_nmi(void)
> > +{
> > +     const struct pmu_irq_ops *irq_ops;
> > +
> > +     irq_ops = *this_cpu_ptr(&cpu_irq_ops);
> > +
> > +     return irq_ops == &pmunmi_ops || irq_ops == &percpu_pmunmi_ops;
> > +}
>
> In the latest iteration of the PMU NMI patches I introduced a static bool
> variable, has_nmi, which is used to print to dmesg if NMIs are in use. The
> function could be rewritten to return that variable.

Okay, will rebase onto the latest version.

-Sumit

>
> Thanks,
> Alex