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 <[email protected]>
---
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
+ 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);
+
+ /*
+ * 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;
+}
+
/*
* PMU hardware loses all context when a CPU goes offline.
* When a CPU is hotplugged back in, since some hardware registers are
diff --git a/include/linux/perf/arm_pmu.h b/include/linux/perf/arm_pmu.h
index 5b616dd..5765069 100644
--- a/include/linux/perf/arm_pmu.h
+++ b/include/linux/perf/arm_pmu.h
@@ -160,6 +160,8 @@ int arm_pmu_acpi_probe(armpmu_init_fn init_fn);
static inline int arm_pmu_acpi_probe(armpmu_init_fn init_fn) { return 0; }
#endif
+bool arm_pmu_irq_is_nmi(void);
+
/* Internal functions only for core arm_pmu code */
struct arm_pmu *armpmu_alloc(void);
struct arm_pmu *armpmu_alloc_atomic(void);
--
2.7.4
On Fri, Sep 04, 2020 at 12:56:37PM +0530, 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 <[email protected]>
> ---
>
> 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(-)
It would be great to have Alexandru's ack on this, since it builds directly
on top of his code.
Will
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?
[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 <[email protected]>
> ---
>
> 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?
> +
> + /*
> + * 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.
Thanks,
Alex
On Wed, 30 Sep 2020 at 18:33, Alexandru Elisei <[email protected]> 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 <[email protected]>
> > ---
> >
> > 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