[PATCH v8] watchdog/hardlockup: simplify perf event probe and remove per-cpu dependency

[PATCH v8] watchdog/hardlockup: simplify perf event probe and remove per-cpu dependency

[Qiliang Yuan]

Simplify the hardlockup detector's probe path and remove its implicit
dependency on pinned per-cpu execution.

Refactor hardlockup_detector_event_create() to be stateless. Return
the created perf_event pointer to the caller instead of directly
modifying the per-cpu 'watchdog_ev' variable. This allows the probe
path to safely manage a temporary event without the risk of leaving
stale pointers should task migration occur.

Signed-off-by: Shouxin Sun <sunshx@chinatelecom.cn>
Signed-off-by: Junnan Zhang <zhangjn11@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <yuanql9@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <realwujing@gmail.com>
Cc: Douglas Anderson <dianders@chromium.org>
Cc: Song Liu <song@kernel.org>
Cc: Jinchao Wang <wangjinchao600@gmail.com>
---
v8:
- Add Signed-off-by with work email.
- v7 link: https://lore.kernel.org/all/20260128060833.1715622-1-realwujing@gmail.com/
v7:
- Update Signed-off-by to match author email.
- Remove redundant cpu_hotplug_disable() in probe path.
- v6 link: https://lore.kernel.org/all/20260128023136.1691973-1-realwujing@gmail.com/
v6:
- Change title to "simplify/cleanup" and remove "Fixes" tag since the issue
  is not reproducible on mainline.
- Rewrite commit message in imperative mood.
- Clarify that mainline is safe while this improves robustness.
- v5 link: https://lore.kernel.org/all/20260127022238.1182079-1-realwujing@gmail.com/
v5:
- Refine description: clarify that the retry path uses worker threads
  without PF_PERCPU_THREAD (though mainline is safe due to system_percpu_wq).
- v4 link: https://lore.kernel.org/all/20260124070814.806828-1-realwujing@gmail.com/
v4:
- Add cpu_hotplug_disable() in watchdog_hardlockup_probe() to stabilize
  the probe CPU.
- Update description to explain the relevance of 4.19 logs.
v3:
- Refactor hardlockup_detector_event_create() to be stateless by returning
  the event pointer instead of directly assigning to per-cpu variables.
- Restore PMU cycle fallback and unify the enable/probe paths.
v2:
- Add Cc: stable@vger.kernel.org.
v1:
- Avoid 'watchdog_ev' in probe path by manually creating and releasing a
  local perf event.

 kernel/watchdog_perf.c | 51 ++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 22 deletions(-)

diff --git a/kernel/watchdog_perf.c b/kernel/watchdog_perf.c
index d3ca70e3c256..4a798443033b 100644
--- a/kernel/watchdog_perf.c
+++ b/kernel/watchdog_perf.c
@@ -17,6 +17,7 @@
 #include <linux/atomic.h>
 #include <linux/module.h>
 #include <linux/sched/debug.h>
+#include <linux/cpu.h>
 
 #include <asm/irq_regs.h>
 #include <linux/perf_event.h>
@@ -118,18 +119,11 @@ static void watchdog_overflow_callback(struct perf_event *event,
 	watchdog_hardlockup_check(smp_processor_id(), regs);
 }
 
-static int hardlockup_detector_event_create(void)
+static struct perf_event *hardlockup_detector_event_create(unsigned int cpu)
 {
-	unsigned int cpu;
 	struct perf_event_attr *wd_attr;
 	struct perf_event *evt;
 
-	/*
-	 * Preemption is not disabled because memory will be allocated.
-	 * Ensure CPU-locality by calling this in per-CPU kthread.
-	 */
-	WARN_ON(!is_percpu_thread());
-	cpu = raw_smp_processor_id();
 	wd_attr = &wd_hw_attr;
 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 
@@ -143,14 +137,7 @@ static int hardlockup_detector_event_create(void)
 						       watchdog_overflow_callback, NULL);
 	}
 
-	if (IS_ERR(evt)) {
-		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
-			 PTR_ERR(evt));
-		return PTR_ERR(evt);
-	}
-	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
-	this_cpu_write(watchdog_ev, evt);
-	return 0;
+	return evt;
 }
 
 /**
@@ -159,17 +146,26 @@ static int hardlockup_detector_event_create(void)
  */
 void watchdog_hardlockup_enable(unsigned int cpu)
 {
+	struct perf_event *evt;
+
 	WARN_ON_ONCE(cpu != smp_processor_id());
 
-	if (hardlockup_detector_event_create())
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
+		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
+			 PTR_ERR(evt));
 		return;
+	}
 
 	/* use original value for check */
 	if (!atomic_fetch_inc(&watchdog_cpus))
 		pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 
+	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
+	this_cpu_write(watchdog_ev, evt);
+
 	watchdog_init_timestamp();
-	perf_event_enable(this_cpu_read(watchdog_ev));
+	perf_event_enable(evt);
 }
 
 /**
@@ -263,19 +259,30 @@ bool __weak __init arch_perf_nmi_is_available(void)
  */
 int __init watchdog_hardlockup_probe(void)
 {
+	struct perf_event *evt;
+	unsigned int cpu;
 	int ret;
 
 	if (!arch_perf_nmi_is_available())
 		return -ENODEV;
 
-	ret = hardlockup_detector_event_create();
+	if (!hw_nmi_get_sample_period(watchdog_thresh))
+		return -EINVAL;
 
-	if (ret) {
+	/*
+	 * Test hardware PMU availability by creating a temporary perf event.
+	 * The event is released immediately.
+	 */
+	cpu = raw_smp_processor_id();
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
 		pr_info("Perf NMI watchdog permanently disabled\n");
+		ret = PTR_ERR(evt);
 	} else {
-		perf_event_release_kernel(this_cpu_read(watchdog_ev));
-		this_cpu_write(watchdog_ev, NULL);
+		perf_event_release_kernel(evt);
+		ret = 0;
 	}
+
 	return ret;
 }
 
-- 
2.51.0

[PATCH v8] watchdog/hardlockup: simplify perf event probe and remove per-cpu dependency

[Qiliang Yuan]

Simplify the hardlockup detector's probe path and remove its implicit
dependency on pinned per-cpu execution.

Refactor hardlockup_detector_event_create() to be stateless. Return
the created perf_event pointer to the caller instead of directly
modifying the per-cpu 'watchdog_ev' variable. This allows the probe
path to safely manage a temporary event without the risk of leaving
stale pointers should task migration occur.

Signed-off-by: Shouxin Sun <sunshx@chinatelecom.cn>
Signed-off-by: Junnan Zhang <zhangjn11@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <yuanql9@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <realwujing@gmail.com>
Cc: Douglas Anderson <dianders@chromium.org>
Cc: Song Liu <song@kernel.org>
Cc: Jinchao Wang <wangjinchao600@gmail.com>
---
v8:
- Add Signed-off-by with work email.
- v7 link: https://lore.kernel.org/all/20260128060833.1715622-1-realwujing@gmail.com/
v7:
- Update Signed-off-by to match author email.
- Remove redundant cpu_hotplug_disable() in probe path.
- v6 link: https://lore.kernel.org/all/20260128023136.1691973-1-realwujing@gmail.com/
v6:
- Change title to "simplify/cleanup" and remove "Fixes" tag since the issue
  is not reproducible on mainline.
- Rewrite commit message in imperative mood.
- Clarify that mainline is safe while this improves robustness.
- v5 link: https://lore.kernel.org/all/20260127022238.1182079-1-realwujing@gmail.com/
v5:
- Refine description: clarify that the retry path uses worker threads
  without PF_PERCPU_THREAD (though mainline is safe due to system_percpu_wq).
- v4 link: https://lore.kernel.org/all/20260124070814.806828-1-realwujing@gmail.com/
v4:
- Add cpu_hotplug_disable() in watchdog_hardlockup_probe() to stabilize
  the probe CPU.
- Update description to explain the relevance of 4.19 logs.
v3:
- Refactor hardlockup_detector_event_create() to be stateless by returning
  the event pointer instead of directly assigning to per-cpu variables.
- Restore PMU cycle fallback and unify the enable/probe paths.
v2:
- Add Cc: stable@vger.kernel.org.
v1:
- Avoid 'watchdog_ev' in probe path by manually creating and releasing a
  local perf event.

 kernel/watchdog_perf.c | 51 ++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 22 deletions(-)

diff --git a/kernel/watchdog_perf.c b/kernel/watchdog_perf.c
index d3ca70e3c256..4a798443033b 100644
--- a/kernel/watchdog_perf.c
+++ b/kernel/watchdog_perf.c
@@ -17,6 +17,7 @@
 #include <linux/atomic.h>
 #include <linux/module.h>
 #include <linux/sched/debug.h>
+#include <linux/cpu.h>
 
 #include <asm/irq_regs.h>
 #include <linux/perf_event.h>
@@ -118,18 +119,11 @@ static void watchdog_overflow_callback(struct perf_event *event,
 	watchdog_hardlockup_check(smp_processor_id(), regs);
 }
 
-static int hardlockup_detector_event_create(void)
+static struct perf_event *hardlockup_detector_event_create(unsigned int cpu)
 {
-	unsigned int cpu;
 	struct perf_event_attr *wd_attr;
 	struct perf_event *evt;
 
-	/*
-	 * Preemption is not disabled because memory will be allocated.
-	 * Ensure CPU-locality by calling this in per-CPU kthread.
-	 */
-	WARN_ON(!is_percpu_thread());
-	cpu = raw_smp_processor_id();
 	wd_attr = &wd_hw_attr;
 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 
@@ -143,14 +137,7 @@ static int hardlockup_detector_event_create(void)
 						       watchdog_overflow_callback, NULL);
 	}
 
-	if (IS_ERR(evt)) {
-		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
-			 PTR_ERR(evt));
-		return PTR_ERR(evt);
-	}
-	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
-	this_cpu_write(watchdog_ev, evt);
-	return 0;
+	return evt;
 }
 
 /**
@@ -159,17 +146,26 @@ static int hardlockup_detector_event_create(void)
  */
 void watchdog_hardlockup_enable(unsigned int cpu)
 {
+	struct perf_event *evt;
+
 	WARN_ON_ONCE(cpu != smp_processor_id());
 
-	if (hardlockup_detector_event_create())
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
+		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
+			 PTR_ERR(evt));
 		return;
+	}
 
 	/* use original value for check */
 	if (!atomic_fetch_inc(&watchdog_cpus))
 		pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 
+	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
+	this_cpu_write(watchdog_ev, evt);
+
 	watchdog_init_timestamp();
-	perf_event_enable(this_cpu_read(watchdog_ev));
+	perf_event_enable(evt);
 }
 
 /**
@@ -263,19 +259,30 @@ bool __weak __init arch_perf_nmi_is_available(void)
  */
 int __init watchdog_hardlockup_probe(void)
 {
+	struct perf_event *evt;
+	unsigned int cpu;
 	int ret;
 
 	if (!arch_perf_nmi_is_available())
 		return -ENODEV;
 
-	ret = hardlockup_detector_event_create();
+	if (!hw_nmi_get_sample_period(watchdog_thresh))
+		return -EINVAL;
 
-	if (ret) {
+	/*
+	 * Test hardware PMU availability by creating a temporary perf event.
+	 * The event is released immediately.
+	 */
+	cpu = raw_smp_processor_id();
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
 		pr_info("Perf NMI watchdog permanently disabled\n");
+		ret = PTR_ERR(evt);
 	} else {
-		perf_event_release_kernel(this_cpu_read(watchdog_ev));
-		this_cpu_write(watchdog_ev, NULL);
+		perf_event_release_kernel(evt);
+		ret = 0;
 	}
+
 	return ret;
 }
 
-- 
2.51.0

[PATCH v8] watchdog/hardlockup: simplify perf event probe and remove per-cpu dependency

[Qiliang Yuan]

Simplify the hardlockup detector's probe path and remove its implicit
dependency on pinned per-cpu execution.

Refactor hardlockup_detector_event_create() to be stateless. Return
the created perf_event pointer to the caller instead of directly
modifying the per-cpu 'watchdog_ev' variable. This allows the probe
path to safely manage a temporary event without the risk of leaving
stale pointers should task migration occur.

Signed-off-by: Shouxin Sun <sunshx@chinatelecom.cn>
Signed-off-by: Junnan Zhang <zhangjn11@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <yuanql9@chinatelecom.cn>
Signed-off-by: Qiliang Yuan <realwujing@gmail.com>
Cc: Douglas Anderson <dianders@chromium.org>
Cc: Song Liu <song@kernel.org>
Cc: Jinchao Wang <wangjinchao600@gmail.com>
---
v8:
- Add Signed-off-by with work email.
- v7 link: https://lore.kernel.org/all/20260128060833.1715622-1-realwujing@gmail.com/
v7:
- Update Signed-off-by to match author email.
- Remove redundant cpu_hotplug_disable() in probe path.
- v6 link: https://lore.kernel.org/all/20260128023136.1691973-1-realwujing@gmail.com/
v6:
- Change title to "simplify/cleanup" and remove "Fixes" tag since the issue
  is not reproducible on mainline.
- Rewrite commit message in imperative mood.
- Clarify that mainline is safe while this improves robustness.
- v5 link: https://lore.kernel.org/all/20260127022238.1182079-1-realwujing@gmail.com/
v5:
- Refine description: clarify that the retry path uses worker threads
  without PF_PERCPU_THREAD (though mainline is safe due to system_percpu_wq).
- v4 link: https://lore.kernel.org/all/20260124070814.806828-1-realwujing@gmail.com/
v4:
- Add cpu_hotplug_disable() in watchdog_hardlockup_probe() to stabilize
  the probe CPU.
- Update description to explain the relevance of 4.19 logs.
v3:
- Refactor hardlockup_detector_event_create() to be stateless by returning
  the event pointer instead of directly assigning to per-cpu variables.
- Restore PMU cycle fallback and unify the enable/probe paths.
v2:
- Add Cc: stable@vger.kernel.org.
v1:
- Avoid 'watchdog_ev' in probe path by manually creating and releasing a
  local perf event.

 kernel/watchdog_perf.c | 51 ++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 22 deletions(-)

diff --git a/kernel/watchdog_perf.c b/kernel/watchdog_perf.c
index d3ca70e3c256..4a798443033b 100644
--- a/kernel/watchdog_perf.c
+++ b/kernel/watchdog_perf.c
@@ -17,6 +17,7 @@
 #include <linux/atomic.h>
 #include <linux/module.h>
 #include <linux/sched/debug.h>
+#include <linux/cpu.h>
 
 #include <asm/irq_regs.h>
 #include <linux/perf_event.h>
@@ -118,18 +119,11 @@ static void watchdog_overflow_callback(struct perf_event *event,
 	watchdog_hardlockup_check(smp_processor_id(), regs);
 }
 
-static int hardlockup_detector_event_create(void)
+static struct perf_event *hardlockup_detector_event_create(unsigned int cpu)
 {
-	unsigned int cpu;
 	struct perf_event_attr *wd_attr;
 	struct perf_event *evt;
 
-	/*
-	 * Preemption is not disabled because memory will be allocated.
-	 * Ensure CPU-locality by calling this in per-CPU kthread.
-	 */
-	WARN_ON(!is_percpu_thread());
-	cpu = raw_smp_processor_id();
 	wd_attr = &wd_hw_attr;
 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
 
@@ -143,14 +137,7 @@ static int hardlockup_detector_event_create(void)
 						       watchdog_overflow_callback, NULL);
 	}
 
-	if (IS_ERR(evt)) {
-		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
-			 PTR_ERR(evt));
-		return PTR_ERR(evt);
-	}
-	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
-	this_cpu_write(watchdog_ev, evt);
-	return 0;
+	return evt;
 }
 
 /**
@@ -159,17 +146,26 @@ static int hardlockup_detector_event_create(void)
  */
 void watchdog_hardlockup_enable(unsigned int cpu)
 {
+	struct perf_event *evt;
+
 	WARN_ON_ONCE(cpu != smp_processor_id());
 
-	if (hardlockup_detector_event_create())
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
+		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
+			 PTR_ERR(evt));
 		return;
+	}
 
 	/* use original value for check */
 	if (!atomic_fetch_inc(&watchdog_cpus))
 		pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 
+	WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
+	this_cpu_write(watchdog_ev, evt);
+
 	watchdog_init_timestamp();
-	perf_event_enable(this_cpu_read(watchdog_ev));
+	perf_event_enable(evt);
 }
 
 /**
@@ -263,19 +259,30 @@ bool __weak __init arch_perf_nmi_is_available(void)
  */
 int __init watchdog_hardlockup_probe(void)
 {
+	struct perf_event *evt;
+	unsigned int cpu;
 	int ret;
 
 	if (!arch_perf_nmi_is_available())
 		return -ENODEV;
 
-	ret = hardlockup_detector_event_create();
+	if (!hw_nmi_get_sample_period(watchdog_thresh))
+		return -EINVAL;
 
-	if (ret) {
+	/*
+	 * Test hardware PMU availability by creating a temporary perf event.
+	 * The event is released immediately.
+	 */
+	cpu = raw_smp_processor_id();
+	evt = hardlockup_detector_event_create(cpu);
+	if (IS_ERR(evt)) {
 		pr_info("Perf NMI watchdog permanently disabled\n");
+		ret = PTR_ERR(evt);
 	} else {
-		perf_event_release_kernel(this_cpu_read(watchdog_ev));
-		this_cpu_write(watchdog_ev, NULL);
+		perf_event_release_kernel(evt);
+		ret = 0;
 	}
+
 	return ret;
 }
 
-- 
2.51.0

Re: [PATCH v8] watchdog/hardlockup: simplify perf event probe and remove per-cpu dependency

[Doug Anderson]

Hi,

On Wed, Jan 28, 2026 at 5:29 AM Qiliang Yuan <realwujing@gmail.com> wrote:
>
> Simplify the hardlockup detector's probe path and remove its implicit
> dependency on pinned per-cpu execution.
>
> Refactor hardlockup_detector_event_create() to be stateless. Return
> the created perf_event pointer to the caller instead of directly
> modifying the per-cpu 'watchdog_ev' variable. This allows the probe
> path to safely manage a temporary event without the risk of leaving
> stale pointers should task migration occur.
>
> Signed-off-by: Shouxin Sun <sunshx@chinatelecom.cn>
> Signed-off-by: Junnan Zhang <zhangjn11@chinatelecom.cn>
> Signed-off-by: Qiliang Yuan <yuanql9@chinatelecom.cn>
> Signed-off-by: Qiliang Yuan <realwujing@gmail.com>
> Cc: Douglas Anderson <dianders@chromium.org>
> Cc: Song Liu <song@kernel.org>
> Cc: Jinchao Wang <wangjinchao600@gmail.com>
> ---
> v8:
> - Add Signed-off-by with work email.
> - v7 link: https://lore.kernel.org/all/20260128060833.1715622-1-realwujing@gmail.com/
> v7:
> - Update Signed-off-by to match author email.
> - Remove redundant cpu_hotplug_disable() in probe path.
> - v6 link: https://lore.kernel.org/all/20260128023136.1691973-1-realwujing@gmail.com/
> v6:
> - Change title to "simplify/cleanup" and remove "Fixes" tag since the issue
>   is not reproducible on mainline.
> - Rewrite commit message in imperative mood.
> - Clarify that mainline is safe while this improves robustness.
> - v5 link: https://lore.kernel.org/all/20260127022238.1182079-1-realwujing@gmail.com/
> v5:
> - Refine description: clarify that the retry path uses worker threads
>   without PF_PERCPU_THREAD (though mainline is safe due to system_percpu_wq).
> - v4 link: https://lore.kernel.org/all/20260124070814.806828-1-realwujing@gmail.com/
> v4:
> - Add cpu_hotplug_disable() in watchdog_hardlockup_probe() to stabilize
>   the probe CPU.
> - Update description to explain the relevance of 4.19 logs.
> v3:
> - Refactor hardlockup_detector_event_create() to be stateless by returning
>   the event pointer instead of directly assigning to per-cpu variables.
> - Restore PMU cycle fallback and unify the enable/probe paths.
> v2:
> - Add Cc: stable@vger.kernel.org.
> v1:
> - Avoid 'watchdog_ev' in probe path by manually creating and releasing a
>   local perf event.
>
>  kernel/watchdog_perf.c | 51 ++++++++++++++++++++++++------------------
>  1 file changed, 29 insertions(+), 22 deletions(-)
>
> diff --git a/kernel/watchdog_perf.c b/kernel/watchdog_perf.c
> index d3ca70e3c256..4a798443033b 100644
> --- a/kernel/watchdog_perf.c
> +++ b/kernel/watchdog_perf.c
> @@ -17,6 +17,7 @@
>  #include <linux/atomic.h>
>  #include <linux/module.h>
>  #include <linux/sched/debug.h>
> +#include <linux/cpu.h>

IIRC, the above include was added for the call to
cpu_hotplug_disable(), which has now been removed. The #include can be
removed now, right?

In any case:

Reviewed-by: Douglas Anderson <dianders@chromium.org>