[PATCHv2 0/2] *** Detect interrupt storm in softlockup ***

[PATCHv2 0/2] *** Detect interrupt storm in softlockup ***

[Bitao Hu]

I have implemented a low-overhead method for detecting interrupt storm
in softlockup. Please review it, all comments are welcome.

Changes from v1 to v2:

- From Douglas, optimize the memory of cpustats. With the maximum number
of CPUs, that's now this.
2 * 8192 * 4 + 1 * 8192 * 5 * 4 + 1 * 8192 = 237,568 bytes.

- From Liu Song, refactor the code format and add necessary comments. 

- From Douglas, use interrupt counts instead of interrupt time to
determine the cause of softlockup. 

- Remove the cmdline parameter added in PATCHv1.

Bitao Hu (2):
  watchdog/softlockup: low-overhead detection of interrupt storm
  watchdog/softlockup: report the most frequent interrupts

 kernel/watchdog.c | 222 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 222 insertions(+)

-- 
2.37.1 (Apple Git-137.1)

[PATCHv2 1/2] watchdog/softlockup: low-overhead detection of interrupt storm

[Bitao Hu]

The following softlockup is caused by interrupt storm, but it cannot be
identified from the call tree. Because the call tree is just a snapshot
and doesn't fully capture the behavior of the CPU during the soft lockup.
  watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
  ...
  Call trace:
    __do_softirq+0xa0/0x37c
    __irq_exit_rcu+0x108/0x140
    irq_exit+0x14/0x20
    __handle_domain_irq+0x84/0xe0
    gic_handle_irq+0x80/0x108
    el0_irq_naked+0x50/0x58

Therefore，I think it is necessary to report CPU utilization during the
softlockup_thresh period (report once every sample_period, for a total
of 5 reportings), like this:
  watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
  CPU#28 Utilization every 4s during lockup:
    #1: 0% system, 0% softirq, 100% hardirq, 0% idle
    #2: 0% system, 0% softirq, 100% hardirq, 0% idle
    #3: 0% system, 0% softirq, 100% hardirq, 0% idle
    #4: 0% system, 0% softirq, 100% hardirq, 0% idle
    #5: 0% system, 0% softirq, 100% hardirq, 0% idle
  ...

This would be helpful in determining whether an interrupt storm has
occurred or in identifying the cause of the softlockup. The criteria for
determination are as follows:
  a. If the hardirq utilization is high, then interrupt storm should be
  considered and the root cause cannot be determined from the call tree.
  b. If the softirq utilization is high, then we could analyze the call
  tree but it may cannot reflect the root cause.
  c. If the system utilization is high, then we could analyze the root
  cause from the call tree.

Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
---
 kernel/watchdog.c | 72 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 72 insertions(+)

diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 81a8862295d6..0efe9604c3c2 100644
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -23,6 +23,8 @@
 #include <linux/sched/debug.h>
 #include <linux/sched/isolation.h>
 #include <linux/stop_machine.h>
+#include <linux/kernel_stat.h>
+#include <linux/math64.h>
 
 #include <asm/irq_regs.h>
 #include <linux/kvm_para.h>
@@ -441,6 +443,73 @@ static int is_softlockup(unsigned long touch_ts,
 	return 0;
 }
 
+#ifdef CONFIG_IRQ_TIME_ACCOUNTING
+#define NUM_STATS_GROUPS	5
+#define STATS_SYSTEM		0
+#define STATS_SOFTIRQ		1
+#define STATS_HARDIRQ		2
+#define STATS_IDLE		3
+#define NUM_STATS_PER_GROUP	4
+static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
+static DEFINE_PER_CPU(u8, cpustat_utilization[NUM_STATS_GROUPS][NUM_STATS_PER_GROUP]);
+static DEFINE_PER_CPU(u8, cpustat_tail);
+static enum cpu_usage_stat idx_to_stat[NUM_STATS_PER_GROUP] = {
+	CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
+};
+
+static void update_cpustat(void)
+{
+	u8 i;
+	u16 *old = this_cpu_ptr(cpustat_old);
+	u8 (*utilization)[NUM_STATS_PER_GROUP] = this_cpu_ptr(cpustat_utilization);
+	u8 tail = this_cpu_read(cpustat_tail);
+	struct kernel_cpustat kcpustat;
+	u64 *cpustat = kcpustat.cpustat;
+	u16 sample_period_ms = sample_period >> 24LL; /* 2^24ns ~= 16.8ms */
+
+	kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
+	for (i = STATS_SYSTEM; i < NUM_STATS_PER_GROUP; i++) {
+		/*
+		 * We don't need nanosecond resolution. A granularity of 16ms is
+		 * sufficient for our precision, allowing us to use u16 to store
+		 * cpustats, which will roll over roughly every ~1000 seconds.
+		 * 2^24 ~= 16 * 10^6
+		 */
+		cpustat[idx_to_stat[i]] = lower_16_bits(cpustat[idx_to_stat[i]] >> 24LL);
+		utilization[tail][i] = 100 * (u16)(cpustat[idx_to_stat[i]] - old[i])
+					/ sample_period_ms;
+		old[i] = cpustat[idx_to_stat[i]];
+	}
+	this_cpu_write(cpustat_tail, (tail + 1) % NUM_STATS_GROUPS);
+}
+
+static void print_cpustat(void)
+{
+	u8 i, j;
+	u8 (*utilization)[NUM_STATS_PER_GROUP] = this_cpu_ptr(cpustat_utilization);
+	u8 tail = this_cpu_read(cpustat_tail);
+	u64 sample_period_second = sample_period;
+
+	do_div(sample_period_second, NSEC_PER_SEC);
+	/*
+	 * We do not want the "watchdog: " prefix on every line,
+	 * hence we use "printk" instead of "pr_crit".
+	 */
+	printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
+		smp_processor_id(), sample_period_second);
+	for (j = STATS_SYSTEM, i = tail; j < NUM_STATS_GROUPS;
+		j++, i = (i + 1) % NUM_STATS_GROUPS) {
+		printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t"
+			"%3u%% hardirq,\t%3u%% idle\n", j+1,
+			utilization[i][STATS_SYSTEM], utilization[i][STATS_SOFTIRQ],
+			utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
+	}
+}
+#else
+static inline void update_cpustat(void) { }
+static inline void print_cpustat(void) { }
+#endif
+
 /* watchdog detector functions */
 static DEFINE_PER_CPU(struct completion, softlockup_completion);
 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
@@ -504,6 +573,8 @@ static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 	 */
 	period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
 
+	update_cpustat();
+
 	/* Reset the interval when touched by known problematic code. */
 	if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
 		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
@@ -539,6 +610,7 @@ static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 		pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 			smp_processor_id(), duration,
 			current->comm, task_pid_nr(current));
+		print_cpustat();
 		print_modules();
 		print_irqtrace_events(current);
 		if (regs)
-- 
2.37.1 (Apple Git-137.1)

[PATCHv2 2/2] watchdog/softlockup: report the most frequent interrupts

[Bitao Hu]

When the watchdog determines that the current soft lockup is due
to an interrupt storm based on CPU utilization, reporting the
most frequent interrupts could be good enough for further
troubleshooting.

Below is an example of interrupt storm. The call tree does not
provide useful information, but we can analyze which interrupt
caused the soft lockup by comparing the counts of interrupts.

[ 2987.488075] watchdog: BUG: soft lockup - CPU#9 stuck for 23s! [kworker/9:1:214]
[ 2987.488607] CPU#9 Utilization every 4s during lockup:
[ 2987.488941]  #1:   0% system,          0% softirq,   100% hardirq,     0% idle
[ 2987.489357]  #2:   0% system,          0% softirq,   100% hardirq,     0% idle
[ 2987.489771]  #3:   0% system,          0% softirq,   100% hardirq,     0% idle
[ 2987.490186]  #4:   0% system,          0% softirq,   100% hardirq,     0% idle
[ 2987.490601]  #5:   0% system,          0% softirq,   100% hardirq,     0% idle
[ 2987.491034] CPU#9 Detect HardIRQ Time exceeds 50%. Most frequent HardIRQs:
[ 2987.491493]  #1: 330985      irq#7(IPI)
[ 2987.491743]  #2: 5000        irq#10(arch_timer)
[ 2987.492039]  #3: 9           irq#91(nvme0q2)
[ 2987.492318]  #4: 3           irq#118(virtio1-output.12)
...
[ 2987.492728] Call trace:
[ 2987.492729]  __do_softirq+0xa8/0x364

Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
---
 kernel/watchdog.c | 150 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 150 insertions(+)

diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 0efe9604c3c2..38fb18e17d71 100644
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -25,6 +25,9 @@
 #include <linux/stop_machine.h>
 #include <linux/kernel_stat.h>
 #include <linux/math64.h>
+#include <linux/irq.h>
+#include <linux/bitops.h>
+#include <linux/irqdesc.h>
 
 #include <asm/irq_regs.h>
 #include <linux/kvm_para.h>
@@ -431,11 +434,15 @@ void touch_softlockup_watchdog_sync(void)
 	__this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 }
 
+static void set_potential_softlockup(unsigned long now, unsigned long touch_ts);
+
 static int is_softlockup(unsigned long touch_ts,
 			 unsigned long period_ts,
 			 unsigned long now)
 {
 	if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
+		/* Softlockup may occur in the current period */
+		set_potential_softlockup(now, period_ts);
 		/* Warn about unreasonable delays. */
 		if (time_after(now, period_ts + get_softlockup_thresh()))
 			return now - touch_ts;
@@ -457,6 +464,8 @@ static enum cpu_usage_stat idx_to_stat[NUM_STATS_PER_GROUP] = {
 	CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
 };
 
+static void print_hardirq_counts(void);
+
 static void update_cpustat(void)
 {
 	u8 i;
@@ -504,10 +513,150 @@ static void print_cpustat(void)
 			utilization[i][STATS_SYSTEM], utilization[i][STATS_SOFTIRQ],
 			utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
 	}
+	print_hardirq_counts();
+}
+
+#define HARDIRQ_PERCENT_THRESH		50
+#define NUM_HARDIRQ_REPORT		5
+static DECLARE_BITMAP(softlockup_hardirq_cpus, CONFIG_NR_CPUS);
+static DEFINE_PER_CPU(u32 *, hardirq_counts);
+
+static void find_counts_top(u32 *irq_counts, int *irq, u32 perirq_counts, int perirq_id, int range)
+{
+	unsigned int i, j;
+
+	for (i = 0; i < range; i++) {
+		if (perirq_counts > irq_counts[i]) {
+			for (j = range - 1; j > i; j--) {
+				irq_counts[j] = irq_counts[j - 1];
+				irq[j] = irq[j - 1];
+			}
+			irq_counts[j] = perirq_counts;
+			irq[j] = perirq_id;
+			break;
+		}
+	}
+}
+
+/*
+ * If the proportion of time spent handling irq exceeds HARDIRQ_PERCENT_THRESH%
+ * during sample_period, then it is necessary to record the counts of each irq.
+ */
+static inline bool need_record_irq_counts(int type)
+{
+	int tail = this_cpu_read(cpustat_tail);
+	u8 utilization;
+
+	if (--tail == -1)
+		tail = 4;
+	utilization = this_cpu_read(cpustat_utilization[tail][type]);
+	return utilization > HARDIRQ_PERCENT_THRESH;
 }
+
+/*
+ * Mark softlockup as potentially caused by hardirq
+ */
+static void set_potential_softlockup_hardirq(void)
+{
+	u32 i;
+	u32 *counts = __this_cpu_read(hardirq_counts);
+	int cpu = smp_processor_id();
+	struct irq_desc *desc;
+
+	if (!need_record_irq_counts(STATS_HARDIRQ))
+		return;
+
+	if (!test_bit(cpu, softlockup_hardirq_cpus)) {
+		counts = kmalloc_array(nr_irqs, sizeof(u32), GFP_ATOMIC);
+		if (!counts)
+			return;
+		for_each_irq_desc(i, desc) {
+			if (!desc)
+				continue;
+			counts[i] = desc->kstat_irqs ?
+				*this_cpu_ptr(desc->kstat_irqs) : 0;
+		}
+		__this_cpu_write(hardirq_counts, counts);
+		set_bit(cpu, softlockup_hardirq_cpus);
+	}
+}
+
+static void clear_potential_softlockup_hardirq(void)
+{
+	u32 *counts = __this_cpu_read(hardirq_counts);
+	int cpu = smp_processor_id();
+
+	if (test_bit(cpu, softlockup_hardirq_cpus)) {
+		kfree(counts);
+		counts = NULL;
+		__this_cpu_write(hardirq_counts, counts);
+		clear_bit(cpu, softlockup_hardirq_cpus);
+	}
+}
+
+/*
+ * Mark that softlockup may occur
+ */
+static void set_potential_softlockup(unsigned long now, unsigned long period_ts)
+{
+	if (time_after_eq(now, period_ts + get_softlockup_thresh() / 5))
+		set_potential_softlockup_hardirq();
+}
+
+static void clear_potential_softlockup(void)
+{
+	clear_potential_softlockup_hardirq();
+}
+
+static void print_hardirq_counts(void)
+{
+	u32 i;
+	struct irq_desc *desc;
+	u32 counts_diff;
+	u32 *counts = __this_cpu_read(hardirq_counts);
+	int cpu = smp_processor_id();
+	u32 hardirq_counts_top[NUM_HARDIRQ_REPORT] = {0, 0, 0, 0, 0};
+	int hardirq_top[NUM_HARDIRQ_REPORT] = {-1, -1, -1, -1, -1};
+
+	if (test_bit(cpu, softlockup_hardirq_cpus)) {
+		/* Find the top NUM_HARDIRQ_REPORT most frequent interrupts */
+		for_each_irq_desc(i, desc) {
+			if (!desc)
+				continue;
+			counts_diff = desc->kstat_irqs ?
+				*this_cpu_ptr(desc->kstat_irqs) - counts[i] : 0;
+			find_counts_top(hardirq_counts_top, hardirq_top,
+					counts_diff, i, NUM_HARDIRQ_REPORT);
+		}
+		/*
+		 * We do not want the "watchdog: " prefix on every line,
+		 * hence we use "printk" instead of "pr_crit".
+		 */
+		printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n",
+			smp_processor_id(), HARDIRQ_PERCENT_THRESH);
+		for (i = 0; i < NUM_HARDIRQ_REPORT; i++) {
+			if (hardirq_top[i] == -1)
+				break;
+			desc = irq_to_desc(hardirq_top[i]);
+			if (desc && desc->action)
+				printk(KERN_CRIT "\t#%u: %-10u\tirq#%d(%s)\n",
+					i+1, hardirq_counts_top[i],
+					hardirq_top[i], desc->action->name);
+			else
+				printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n",
+					i+1, hardirq_counts_top[i],
+					hardirq_top[i]);
+		}
+		if (!need_record_irq_counts(STATS_HARDIRQ))
+			clear_potential_softlockup_hardirq();
+	}
+}
+
 #else
 static inline void update_cpustat(void) { }
 static inline void print_cpustat(void) { }
+static inline void set_potential_softlockup(unsigned long now, unsigned long period_ts) { }
+static inline void clear_potential_softlockup(void) { }
 #endif
 
 /* watchdog detector functions */
@@ -525,6 +674,7 @@ static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 static int softlockup_fn(void *data)
 {
 	update_touch_ts();
+	clear_potential_softlockup();
 	complete(this_cpu_ptr(&softlockup_completion));
 
 	return 0;
-- 
2.37.1 (Apple Git-137.1)

Re: [PATCHv2 1/2] watchdog/softlockup: low-overhead detection of interrupt storm

[Liu Song]

在 2024/1/30 15:47, Bitao Hu 写道:
> The following softlockup is caused by interrupt storm, but it cannot be
> identified from the call tree. Because the call tree is just a snapshot
> and doesn't fully capture the behavior of the CPU during the soft lockup.
>    watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>    ...
>    Call trace:
>      __do_softirq+0xa0/0x37c
>      __irq_exit_rcu+0x108/0x140
>      irq_exit+0x14/0x20
>      __handle_domain_irq+0x84/0xe0
>      gic_handle_irq+0x80/0x108
>      el0_irq_naked+0x50/0x58
>
> Therefore，I think it is necessary to report CPU utilization during the
> softlockup_thresh period (report once every sample_period, for a total
> of 5 reportings), like this:
>    watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>    CPU#28 Utilization every 4s during lockup:
>      #1: 0% system, 0% softirq, 100% hardirq, 0% idle
>      #2: 0% system, 0% softirq, 100% hardirq, 0% idle
>      #3: 0% system, 0% softirq, 100% hardirq, 0% idle
>      #4: 0% system, 0% softirq, 100% hardirq, 0% idle
>      #5: 0% system, 0% softirq, 100% hardirq, 0% idle
>    ...
>
> This would be helpful in determining whether an interrupt storm has
> occurred or in identifying the cause of the softlockup. The criteria for
> determination are as follows:
>    a. If the hardirq utilization is high, then interrupt storm should be
>    considered and the root cause cannot be determined from the call tree.
>    b. If the softirq utilization is high, then we could analyze the call
>    tree but it may cannot reflect the root cause.
>    c. If the system utilization is high, then we could analyze the root
>    cause from the call tree.
>
> Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
> ---
>   kernel/watchdog.c | 72 +++++++++++++++++++++++++++++++++++++++++++++++
>   1 file changed, 72 insertions(+)
>
> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
> index 81a8862295d6..0efe9604c3c2 100644
> --- a/kernel/watchdog.c
> +++ b/kernel/watchdog.c
> @@ -23,6 +23,8 @@
>   #include <linux/sched/debug.h>
>   #include <linux/sched/isolation.h>
>   #include <linux/stop_machine.h>
> +#include <linux/kernel_stat.h>
> +#include <linux/math64.h>
>   
>   #include <asm/irq_regs.h>
>   #include <linux/kvm_para.h>
> @@ -441,6 +443,73 @@ static int is_softlockup(unsigned long touch_ts,
>   	return 0;
>   }
>   
> +#ifdef CONFIG_IRQ_TIME_ACCOUNTING
> +#define NUM_STATS_GROUPS	5
> +#define STATS_SYSTEM		0
> +#define STATS_SOFTIRQ		1
> +#define STATS_HARDIRQ		2
> +#define STATS_IDLE		3
> +#define NUM_STATS_PER_GROUP	4
This is a set of related numbers; wouldn't it be better to use an enum?
> +static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
> +static DEFINE_PER_CPU(u8, cpustat_utilization[NUM_STATS_GROUPS][NUM_STATS_PER_GROUP]);
> +static DEFINE_PER_CPU(u8, cpustat_tail);
> +static enum cpu_usage_stat idx_to_stat[NUM_STATS_PER_GROUP] = {
> +	CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
> +};
To be honest, I'm not particularly fond of the name 'idx_to_stat' as the 
concept of an
index is already implied by the nature of an array, so adding 'idx' is 
redundant.
I suggest shortening the name.

> +
> +static void update_cpustat(void)
> +{
> +	u8 i;
> +	u16 *old = this_cpu_ptr(cpustat_old);
> +	u8 (*utilization)[NUM_STATS_PER_GROUP] = this_cpu_ptr(cpustat_utilization);
> +	u8 tail = this_cpu_read(cpustat_tail);
> +	struct kernel_cpustat kcpustat;
> +	u64 *cpustat = kcpustat.cpustat;
> +	u16 sample_period_ms = sample_period >> 24LL; /* 2^24ns ~= 16.8ms */

There are two instances where right shift operations are used; it is 
suggested to employ a helper macro for a more comfortable look.


> +
> +	kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
> +	for (i = STATS_SYSTEM; i < NUM_STATS_PER_GROUP; i++) {
> +		/*
> +		 * We don't need nanosecond resolution. A granularity of 16ms is
> +		 * sufficient for our precision, allowing us to use u16 to store
> +		 * cpustats, which will roll over roughly every ~1000 seconds.
> +		 * 2^24 ~= 16 * 10^6
> +		 */
> +		cpustat[idx_to_stat[i]] = lower_16_bits(cpustat[idx_to_stat[i]] >> 24LL);
> +		utilization[tail][i] = 100 * (u16)(cpustat[idx_to_stat[i]] - old[i])
> +					/ sample_period_ms;
> +		old[i] = cpustat[idx_to_stat[i]];
> +	}
> +	this_cpu_write(cpustat_tail, (tail + 1) % NUM_STATS_GROUPS);
> +}
> +
> +static void print_cpustat(void)
> +{
> +	u8 i, j;
> +	u8 (*utilization)[NUM_STATS_PER_GROUP] = this_cpu_ptr(cpustat_utilization);
> +	u8 tail = this_cpu_read(cpustat_tail);
> +	u64 sample_period_second = sample_period;
> +
> +	do_div(sample_period_second, NSEC_PER_SEC);
> +	/*
> +	 * We do not want the "watchdog: " prefix on every line,
> +	 * hence we use "printk" instead of "pr_crit".
> +	 */
> +	printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
> +		smp_processor_id(), sample_period_second);
> +	for (j = STATS_SYSTEM, i = tail; j < NUM_STATS_GROUPS;
> +		j++, i = (i + 1) % NUM_STATS_GROUPS) {
> +		printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t"
> +			"%3u%% hardirq,\t%3u%% idle\n", j+1,
> +			utilization[i][STATS_SYSTEM], utilization[i][STATS_SOFTIRQ],
> +			utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
> +	}
> +}
> +#else
> +static inline void update_cpustat(void) { }
> +static inline void print_cpustat(void) { }
> +#endif
> +
>   /* watchdog detector functions */
>   static DEFINE_PER_CPU(struct completion, softlockup_completion);
>   static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
> @@ -504,6 +573,8 @@ static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
>   	 */
>   	period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
>   
> +	update_cpustat();
> +
>   	/* Reset the interval when touched by known problematic code. */
>   	if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
>   		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
> @@ -539,6 +610,7 @@ static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
>   		pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
>   			smp_processor_id(), duration,
>   			current->comm, task_pid_nr(current));
> +		print_cpustat();
>   		print_modules();
>   		print_irqtrace_events(current);
>   		if (regs)

Re: [PATCHv2 2/2] watchdog/softlockup: report the most frequent interrupts

[Liu Song]

在 2024/1/30 15:47, Bitao Hu 写道:
> When the watchdog determines that the current soft lockup is due
> to an interrupt storm based on CPU utilization, reporting the
> most frequent interrupts could be good enough for further
> troubleshooting.
>
> Below is an example of interrupt storm. The call tree does not
> provide useful information, but we can analyze which interrupt
> caused the soft lockup by comparing the counts of interrupts.
>
> [ 2987.488075] watchdog: BUG: soft lockup - CPU#9 stuck for 23s! [kworker/9:1:214]
> [ 2987.488607] CPU#9 Utilization every 4s during lockup:
> [ 2987.488941]  #1:   0% system,          0% softirq,   100% hardirq,     0% idle
> [ 2987.489357]  #2:   0% system,          0% softirq,   100% hardirq,     0% idle
> [ 2987.489771]  #3:   0% system,          0% softirq,   100% hardirq,     0% idle
> [ 2987.490186]  #4:   0% system,          0% softirq,   100% hardirq,     0% idle
> [ 2987.490601]  #5:   0% system,          0% softirq,   100% hardirq,     0% idle
> [ 2987.491034] CPU#9 Detect HardIRQ Time exceeds 50%. Most frequent HardIRQs:
> [ 2987.491493]  #1: 330985      irq#7(IPI)
> [ 2987.491743]  #2: 5000        irq#10(arch_timer)
> [ 2987.492039]  #3: 9           irq#91(nvme0q2)
> [ 2987.492318]  #4: 3           irq#118(virtio1-output.12)
> ...
> [ 2987.492728] Call trace:
> [ 2987.492729]  __do_softirq+0xa8/0x364
>
> Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
> ---
>   kernel/watchdog.c | 150 ++++++++++++++++++++++++++++++++++++++++++++++
>   1 file changed, 150 insertions(+)
>
> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
> index 0efe9604c3c2..38fb18e17d71 100644
> --- a/kernel/watchdog.c
> +++ b/kernel/watchdog.c
> @@ -25,6 +25,9 @@
>   #include <linux/stop_machine.h>
>   #include <linux/kernel_stat.h>
>   #include <linux/math64.h>
> +#include <linux/irq.h>
> +#include <linux/bitops.h>
> +#include <linux/irqdesc.h>
>   
>   #include <asm/irq_regs.h>
>   #include <linux/kvm_para.h>
> @@ -431,11 +434,15 @@ void touch_softlockup_watchdog_sync(void)
>   	__this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
>   }
>   
> +static void set_potential_softlockup(unsigned long now, unsigned long touch_ts);
> +
>   static int is_softlockup(unsigned long touch_ts,
>   			 unsigned long period_ts,
>   			 unsigned long now)
>   {
>   	if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
> +		/* Softlockup may occur in the current period */
> +		set_potential_softlockup(now, period_ts);
>   		/* Warn about unreasonable delays. */
>   		if (time_after(now, period_ts + get_softlockup_thresh()))
>   			return now - touch_ts;
> @@ -457,6 +464,8 @@ static enum cpu_usage_stat idx_to_stat[NUM_STATS_PER_GROUP] = {
>   	CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
>   };
>   
> +static void print_hardirq_counts(void);
> +
>   static void update_cpustat(void)
>   {
>   	u8 i;
> @@ -504,10 +513,150 @@ static void print_cpustat(void)
>   			utilization[i][STATS_SYSTEM], utilization[i][STATS_SOFTIRQ],
>   			utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
>   	}
> +	print_hardirq_counts();
> +}
> +
> +#define HARDIRQ_PERCENT_THRESH		50
> +#define NUM_HARDIRQ_REPORT		5
> +static DECLARE_BITMAP(softlockup_hardirq_cpus, CONFIG_NR_CPUS);
> +static DEFINE_PER_CPU(u32 *, hardirq_counts);
> +
> +static void find_counts_top(u32 *irq_counts, int *irq, u32 perirq_counts, int perirq_id, int range)
> +{
> +	unsigned int i, j;
> +
> +	for (i = 0; i < range; i++) {
> +		if (perirq_counts > irq_counts[i]) {
> +			for (j = range - 1; j > i; j--) {
> +				irq_counts[j] = irq_counts[j - 1];
> +				irq[j] = irq[j - 1];
> +			}
> +			irq_counts[j] = perirq_counts;
> +			irq[j] = perirq_id;
> +			break;
> +		}
> +	}
> +}
> +
> +/*
> + * If the proportion of time spent handling irq exceeds HARDIRQ_PERCENT_THRESH%
> + * during sample_period, then it is necessary to record the counts of each irq.
> + */
> +static inline bool need_record_irq_counts(int type)
> +{
> +	int tail = this_cpu_read(cpustat_tail);
> +	u8 utilization;
> +
> +	if (--tail == -1)
> +		tail = 4;
> +	utilization = this_cpu_read(cpustat_utilization[tail][type]);
> +	return utilization > HARDIRQ_PERCENT_THRESH;
>   }
> +
> +/*
> + * Mark softlockup as potentially caused by hardirq
> + */
> +static void set_potential_softlockup_hardirq(void)
> +{
> +	u32 i;
> +	u32 *counts = __this_cpu_read(hardirq_counts);
> +	int cpu = smp_processor_id();
> +	struct irq_desc *desc;
> +
> +	if (!need_record_irq_counts(STATS_HARDIRQ))
> +		return;
> +
> +	if (!test_bit(cpu, softlockup_hardirq_cpus)) {
> +		counts = kmalloc_array(nr_irqs, sizeof(u32), GFP_ATOMIC);
> +		if (!counts)
> +			return;
> +		for_each_irq_desc(i, desc) {
> +			if (!desc)
> +				continue;
> +			counts[i] = desc->kstat_irqs ?
> +				*this_cpu_ptr(desc->kstat_irqs) : 0;
> +		}
> +		__this_cpu_write(hardirq_counts, counts);
> +		set_bit(cpu, softlockup_hardirq_cpus);
> +	}
> +}
> +
> +static void clear_potential_softlockup_hardirq(void)
> +{
> +	u32 *counts = __this_cpu_read(hardirq_counts);
> +	int cpu = smp_processor_id();
> +
> +	if (test_bit(cpu, softlockup_hardirq_cpus)) {
> +		kfree(counts);
> +		counts = NULL;
> +		__this_cpu_write(hardirq_counts, counts);
> +		clear_bit(cpu, softlockup_hardirq_cpus);
> +	}
> +}
> +
> +/*
> + * Mark that softlockup may occur
> + */
> +static void set_potential_softlockup(unsigned long now, unsigned long period_ts)
> +{
> +	if (time_after_eq(now, period_ts + get_softlockup_thresh() / 5))
> +		set_potential_softlockup_hardirq();
> +}
> +
> +static void clear_potential_softlockup(void)
> +{
> +	clear_potential_softlockup_hardirq();
> +}
> +
> +static void print_hardirq_counts(void)
> +{
> +	u32 i;
> +	struct irq_desc *desc;
> +	u32 counts_diff;
> +	u32 *counts = __this_cpu_read(hardirq_counts);
> +	int cpu = smp_processor_id();
> +	u32 hardirq_counts_top[NUM_HARDIRQ_REPORT] = {0, 0, 0, 0, 0};
> +	int hardirq_top[NUM_HARDIRQ_REPORT] = {-1, -1, -1, -1, -1};
"hardirq_counts_top" and "hardirq_top" seem like two members of a struct,
working together to record the most suspicious irq, so wouldn't using a 
struct make it clearer?
> +
> +	if (test_bit(cpu, softlockup_hardirq_cpus)) {
> +		/* Find the top NUM_HARDIRQ_REPORT most frequent interrupts */
> +		for_each_irq_desc(i, desc) {
> +			if (!desc)
> +				continue;
> +			counts_diff = desc->kstat_irqs ?
> +				*this_cpu_ptr(desc->kstat_irqs) - counts[i] : 0;
> +			find_counts_top(hardirq_counts_top, hardirq_top,
> +					counts_diff, i, NUM_HARDIRQ_REPORT);
> +		}
> +		/*
> +		 * We do not want the "watchdog: " prefix on every line,
> +		 * hence we use "printk" instead of "pr_crit".
> +		 */
> +		printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n",
> +			smp_processor_id(), HARDIRQ_PERCENT_THRESH);
> +		for (i = 0; i < NUM_HARDIRQ_REPORT; i++) {
> +			if (hardirq_top[i] == -1)
> +				break;
> +			desc = irq_to_desc(hardirq_top[i]);
> +			if (desc && desc->action)
> +				printk(KERN_CRIT "\t#%u: %-10u\tirq#%d(%s)\n",
> +					i+1, hardirq_counts_top[i],
> +					hardirq_top[i], desc->action->name);
> +			else
> +				printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n",
> +					i+1, hardirq_counts_top[i],
> +					hardirq_top[i]);
> +		}
> +		if (!need_record_irq_counts(STATS_HARDIRQ))
> +			clear_potential_softlockup_hardirq();
> +	}
> +}
> +
>   #else
>   static inline void update_cpustat(void) { }
>   static inline void print_cpustat(void) { }
> +static inline void set_potential_softlockup(unsigned long now, unsigned long period_ts) { }
> +static inline void clear_potential_softlockup(void) { }
>   #endif
>   
>   /* watchdog detector functions */
> @@ -525,6 +674,7 @@ static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
>   static int softlockup_fn(void *data)
>   {
>   	update_touch_ts();
> +	clear_potential_softlockup();
>   	complete(this_cpu_ptr(&softlockup_completion));
>   
>   	return 0;

Re: [PATCHv2 1/2] watchdog/softlockup: low-overhead detection of interrupt storm

[kernel test robot]

Hi Bitao,

kernel test robot noticed the following build warnings:

[auto build test WARNING on akpm-mm/mm-everything]
[also build test WARNING on linus/master v6.8-rc2 next-20240131]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Bitao-Hu/watchdog-softlockup-low-overhead-detection-of-interrupt-storm/20240130-155029
base:   https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git mm-everything
patch link:    https://lore.kernel.org/r/20240130074744.45759-2-yaoma%40linux.alibaba.com
patch subject: [PATCHv2 1/2] watchdog/softlockup: low-overhead detection of interrupt storm
config: x86_64-randconfig-122-20240130 (https://download.01.org/0day-ci/archive/20240131/202401311206.nS3XM5Li-lkp@intel.com/config)
compiler: clang version 17.0.6 (https://github.com/llvm/llvm-project 6009708b4367171ccdbf4b5905cb6a803753fe18)
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20240131/202401311206.nS3XM5Li-lkp@intel.com/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202401311206.nS3XM5Li-lkp@intel.com/

sparse warnings: (new ones prefixed by >>)
>> kernel/watchdog.c:464:50: sparse: sparse: incorrect type in initializer (different modifiers) @@     expected unsigned char [usertype] ( *utilization )[4] @@     got unsigned char [noderef] [usertype] ( * )[4] @@
   kernel/watchdog.c:464:50: sparse:     expected unsigned char [usertype] ( *utilization )[4]
   kernel/watchdog.c:464:50: sparse:     got unsigned char [noderef] [usertype] ( * )[4]
   kernel/watchdog.c:489:50: sparse: sparse: incorrect type in initializer (different modifiers) @@     expected unsigned char [usertype] ( *utilization )[4] @@     got unsigned char [noderef] [usertype] ( * )[4] @@
   kernel/watchdog.c:489:50: sparse:     expected unsigned char [usertype] ( *utilization )[4]
   kernel/watchdog.c:489:50: sparse:     got unsigned char [noderef] [usertype] ( * )[4]

vim +464 kernel/watchdog.c

   459	
   460	static void update_cpustat(void)
   461	{
   462		u8 i;
   463		u16 *old = this_cpu_ptr(cpustat_old);
 > 464		u8 (*utilization)[NUM_STATS_PER_GROUP] = this_cpu_ptr(cpustat_utilization);
   465		u8 tail = this_cpu_read(cpustat_tail);
   466		struct kernel_cpustat kcpustat;
   467		u64 *cpustat = kcpustat.cpustat;
   468		u16 sample_period_ms = sample_period >> 24LL; /* 2^24ns ~= 16.8ms */
   469	
   470		kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
   471		for (i = STATS_SYSTEM; i < NUM_STATS_PER_GROUP; i++) {
   472			/*
   473			 * We don't need nanosecond resolution. A granularity of 16ms is
   474			 * sufficient for our precision, allowing us to use u16 to store
   475			 * cpustats, which will roll over roughly every ~1000 seconds.
   476			 * 2^24 ~= 16 * 10^6
   477			 */
   478			cpustat[idx_to_stat[i]] = lower_16_bits(cpustat[idx_to_stat[i]] >> 24LL);
   479			utilization[tail][i] = 100 * (u16)(cpustat[idx_to_stat[i]] - old[i])
   480						/ sample_period_ms;
   481			old[i] = cpustat[idx_to_stat[i]];
   482		}
   483		this_cpu_write(cpustat_tail, (tail + 1) % NUM_STATS_GROUPS);
   484	}
   485	

-- 
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki

Re: [PATCHv2 1/2] watchdog/softlockup: low-overhead detection of interrupt storm

[Bitao Hu]

On 2024/1/31 09:19, Liu Song wrote:
> 
> 在 2024/1/30 15:47, Bitao Hu 写道:
>> The following softlockup is caused by interrupt storm, but it cannot be
>> identified from the call tree. Because the call tree is just a snapshot
>> and doesn't fully capture the behavior of the CPU during the soft lockup.
>>    watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>>    ...
>>    Call trace:
>>      __do_softirq+0xa0/0x37c
>>      __irq_exit_rcu+0x108/0x140
>>      irq_exit+0x14/0x20
>>      __handle_domain_irq+0x84/0xe0
>>      gic_handle_irq+0x80/0x108
>>      el0_irq_naked+0x50/0x58
>>
>> Therefore，I think it is necessary to report CPU utilization during the
>> softlockup_thresh period (report once every sample_period, for a total
>> of 5 reportings), like this:
>>    watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>>    CPU#28 Utilization every 4s during lockup:
>>      #1: 0% system, 0% softirq, 100% hardirq, 0% idle
>>      #2: 0% system, 0% softirq, 100% hardirq, 0% idle
>>      #3: 0% system, 0% softirq, 100% hardirq, 0% idle
>>      #4: 0% system, 0% softirq, 100% hardirq, 0% idle
>>      #5: 0% system, 0% softirq, 100% hardirq, 0% idle
>>    ...
>>
>> This would be helpful in determining whether an interrupt storm has
>> occurred or in identifying the cause of the softlockup. The criteria for
>> determination are as follows:
>>    a. If the hardirq utilization is high, then interrupt storm should be
>>    considered and the root cause cannot be determined from the call tree.
>>    b. If the softirq utilization is high, then we could analyze the call
>>    tree but it may cannot reflect the root cause.
>>    c. If the system utilization is high, then we could analyze the root
>>    cause from the call tree.
>>
>> Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
>> ---
>>   kernel/watchdog.c | 72 +++++++++++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 72 insertions(+)
>>
>> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
>> index 81a8862295d6..0efe9604c3c2 100644
>> --- a/kernel/watchdog.c
>> +++ b/kernel/watchdog.c
>> @@ -23,6 +23,8 @@
>>   #include <linux/sched/debug.h>
>>   #include <linux/sched/isolation.h>
>>   #include <linux/stop_machine.h>
>> +#include <linux/kernel_stat.h>
>> +#include <linux/math64.h>
>>   #include <asm/irq_regs.h>
>>   #include <linux/kvm_para.h>
>> @@ -441,6 +443,73 @@ static int is_softlockup(unsigned long touch_ts,
>>       return 0;
>>   }
>> +#ifdef CONFIG_IRQ_TIME_ACCOUNTING
>> +#define NUM_STATS_GROUPS    5
>> +#define STATS_SYSTEM        0
>> +#define STATS_SOFTIRQ        1
>> +#define STATS_HARDIRQ        2
>> +#define STATS_IDLE        3
>> +#define NUM_STATS_PER_GROUP    4
> This is a set of related numbers; wouldn't it be better to use an enum?
Agree.
>> +static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
>> +static DEFINE_PER_CPU(u8, 
>> cpustat_utilization[NUM_STATS_GROUPS][NUM_STATS_PER_GROUP]);
>> +static DEFINE_PER_CPU(u8, cpustat_tail);
>> +static enum cpu_usage_stat idx_to_stat[NUM_STATS_PER_GROUP] = {
>> +    CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
>> +};
> To be honest, I'm not particularly fond of the name 'idx_to_stat' as the 
> concept of an
> index is already implied by the nature of an array, so adding 'idx' is 
> redundant.
> I suggest shortening the name.
OK, the name 'stats' is clear enough here.
> 
>> +
>> +static void update_cpustat(void)
>> +{
>> +    u8 i;
>> +    u16 *old = this_cpu_ptr(cpustat_old);
>> +    u8 (*utilization)[NUM_STATS_PER_GROUP] = 
>> this_cpu_ptr(cpustat_utilization);
>> +    u8 tail = this_cpu_read(cpustat_tail);
>> +    struct kernel_cpustat kcpustat;
>> +    u64 *cpustat = kcpustat.cpustat;
>> +    u16 sample_period_ms = sample_period >> 24LL; /* 2^24ns ~= 16.8ms */
> 
> There are two instances where right shift operations are used; it is 
> suggested to employ a helper macro for a more comfortable look.
OK.
> 
> 
>> +
>> +    kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
>> +    for (i = STATS_SYSTEM; i < NUM_STATS_PER_GROUP; i++) {
>> +        /*
>> +         * We don't need nanosecond resolution. A granularity of 16ms is
>> +         * sufficient for our precision, allowing us to use u16 to store
>> +         * cpustats, which will roll over roughly every ~1000 seconds.
>> +         * 2^24 ~= 16 * 10^6
>> +         */
>> +        cpustat[idx_to_stat[i]] = 
>> lower_16_bits(cpustat[idx_to_stat[i]] >> 24LL);
>> +        utilization[tail][i] = 100 * (u16)(cpustat[idx_to_stat[i]] - 
>> old[i])
>> +                    / sample_period_ms;
>> +        old[i] = cpustat[idx_to_stat[i]];
>> +    }
>> +    this_cpu_write(cpustat_tail, (tail + 1) % NUM_STATS_GROUPS);
>> +}
>> +
>> +static void print_cpustat(void)
>> +{
>> +    u8 i, j;
>> +    u8 (*utilization)[NUM_STATS_PER_GROUP] = 
>> this_cpu_ptr(cpustat_utilization);
>> +    u8 tail = this_cpu_read(cpustat_tail);
>> +    u64 sample_period_second = sample_period;
>> +
>> +    do_div(sample_period_second, NSEC_PER_SEC);
>> +    /*
>> +     * We do not want the "watchdog: " prefix on every line,
>> +     * hence we use "printk" instead of "pr_crit".
>> +     */
>> +    printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
>> +        smp_processor_id(), sample_period_second);
>> +    for (j = STATS_SYSTEM, i = tail; j < NUM_STATS_GROUPS;
>> +        j++, i = (i + 1) % NUM_STATS_GROUPS) {
>> +        printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t"
>> +            "%3u%% hardirq,\t%3u%% idle\n", j+1,
>> +            utilization[i][STATS_SYSTEM], utilization[i][STATS_SOFTIRQ],
>> +            utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
>> +    }
>> +}
>> +#else
>> +static inline void update_cpustat(void) { }
>> +static inline void print_cpustat(void) { }
>> +#endif
>> +
>>   /* watchdog detector functions */
>>   static DEFINE_PER_CPU(struct completion, softlockup_completion);
>>   static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
>> @@ -504,6 +573,8 @@ static enum hrtimer_restart 
>> watchdog_timer_fn(struct hrtimer *hrtimer)
>>        */
>>       period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
>> +    update_cpustat();
>> +
>>       /* Reset the interval when touched by known problematic code. */
>>       if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
>>           if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
>> @@ -539,6 +610,7 @@ static enum hrtimer_restart 
>> watchdog_timer_fn(struct hrtimer *hrtimer)
>>           pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
>>               smp_processor_id(), duration,
>>               current->comm, task_pid_nr(current));
>> +        print_cpustat();
>>           print_modules();
>>           print_irqtrace_events(current);
>>           if (regs)

Re: [PATCHv2 2/2] watchdog/softlockup: report the most frequent interrupts

[Bitao Hu]

On 2024/1/31 09:50, Liu Song wrote:
> 
> 在 2024/1/30 15:47, Bitao Hu 写道:
>> When the watchdog determines that the current soft lockup is due
>> to an interrupt storm based on CPU utilization, reporting the
>> most frequent interrupts could be good enough for further
>> troubleshooting.
>>
>> Below is an example of interrupt storm. The call tree does not
>> provide useful information, but we can analyze which interrupt
>> caused the soft lockup by comparing the counts of interrupts.
>>
>> [ 2987.488075] watchdog: BUG: soft lockup - CPU#9 stuck for 23s! 
>> [kworker/9:1:214]
>> [ 2987.488607] CPU#9 Utilization every 4s during lockup:
>> [ 2987.488941]  #1:   0% system,          0% softirq,   100% 
>> hardirq,     0% idle
>> [ 2987.489357]  #2:   0% system,          0% softirq,   100% 
>> hardirq,     0% idle
>> [ 2987.489771]  #3:   0% system,          0% softirq,   100% 
>> hardirq,     0% idle
>> [ 2987.490186]  #4:   0% system,          0% softirq,   100% 
>> hardirq,     0% idle
>> [ 2987.490601]  #5:   0% system,          0% softirq,   100% 
>> hardirq,     0% idle
>> [ 2987.491034] CPU#9 Detect HardIRQ Time exceeds 50%. Most frequent 
>> HardIRQs:
>> [ 2987.491493]  #1: 330985      irq#7(IPI)
>> [ 2987.491743]  #2: 5000        irq#10(arch_timer)
>> [ 2987.492039]  #3: 9           irq#91(nvme0q2)
>> [ 2987.492318]  #4: 3           irq#118(virtio1-output.12)
>> ...
>> [ 2987.492728] Call trace:
>> [ 2987.492729]  __do_softirq+0xa8/0x364
>>
>> Signed-off-by: Bitao Hu <yaoma@linux.alibaba.com>
>> ---
>>   kernel/watchdog.c | 150 ++++++++++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 150 insertions(+)
>>
>> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
>> index 0efe9604c3c2..38fb18e17d71 100644
>> --- a/kernel/watchdog.c
>> +++ b/kernel/watchdog.c
>> @@ -25,6 +25,9 @@
>>   #include <linux/stop_machine.h>
>>   #include <linux/kernel_stat.h>
>>   #include <linux/math64.h>
>> +#include <linux/irq.h>
>> +#include <linux/bitops.h>
>> +#include <linux/irqdesc.h>
>>   #include <asm/irq_regs.h>
>>   #include <linux/kvm_para.h>
>> @@ -431,11 +434,15 @@ void touch_softlockup_watchdog_sync(void)
>>       __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
>>   }
>> +static void set_potential_softlockup(unsigned long now, unsigned long 
>> touch_ts);
>> +
>>   static int is_softlockup(unsigned long touch_ts,
>>                unsigned long period_ts,
>>                unsigned long now)
>>   {
>>       if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && 
>> watchdog_thresh) {
>> +        /* Softlockup may occur in the current period */
>> +        set_potential_softlockup(now, period_ts);
>>           /* Warn about unreasonable delays. */
>>           if (time_after(now, period_ts + get_softlockup_thresh()))
>>               return now - touch_ts;
>> @@ -457,6 +464,8 @@ static enum cpu_usage_stat 
>> idx_to_stat[NUM_STATS_PER_GROUP] = {
>>       CPUTIME_SYSTEM, CPUTIME_SOFTIRQ, CPUTIME_IRQ, CPUTIME_IDLE
>>   };
>> +static void print_hardirq_counts(void);
>> +
>>   static void update_cpustat(void)
>>   {
>>       u8 i;
>> @@ -504,10 +513,150 @@ static void print_cpustat(void)
>>               utilization[i][STATS_SYSTEM], 
>> utilization[i][STATS_SOFTIRQ],
>>               utilization[i][STATS_HARDIRQ], utilization[i][STATS_IDLE]);
>>       }
>> +    print_hardirq_counts();
>> +}
>> +
>> +#define HARDIRQ_PERCENT_THRESH        50
>> +#define NUM_HARDIRQ_REPORT        5
>> +static DECLARE_BITMAP(softlockup_hardirq_cpus, CONFIG_NR_CPUS);
>> +static DEFINE_PER_CPU(u32 *, hardirq_counts);
>> +
>> +static void find_counts_top(u32 *irq_counts, int *irq, u32 
>> perirq_counts, int perirq_id, int range)
>> +{
>> +    unsigned int i, j;
>> +
>> +    for (i = 0; i < range; i++) {
>> +        if (perirq_counts > irq_counts[i]) {
>> +            for (j = range - 1; j > i; j--) {
>> +                irq_counts[j] = irq_counts[j - 1];
>> +                irq[j] = irq[j - 1];
>> +            }
>> +            irq_counts[j] = perirq_counts;
>> +            irq[j] = perirq_id;
>> +            break;
>> +        }
>> +    }
>> +}
>> +
>> +/*
>> + * If the proportion of time spent handling irq exceeds 
>> HARDIRQ_PERCENT_THRESH%
>> + * during sample_period, then it is necessary to record the counts of 
>> each irq.
>> + */
>> +static inline bool need_record_irq_counts(int type)
>> +{
>> +    int tail = this_cpu_read(cpustat_tail);
>> +    u8 utilization;
>> +
>> +    if (--tail == -1)
>> +        tail = 4;
>> +    utilization = this_cpu_read(cpustat_utilization[tail][type]);
>> +    return utilization > HARDIRQ_PERCENT_THRESH;
>>   }
>> +
>> +/*
>> + * Mark softlockup as potentially caused by hardirq
>> + */
>> +static void set_potential_softlockup_hardirq(void)
>> +{
>> +    u32 i;
>> +    u32 *counts = __this_cpu_read(hardirq_counts);
>> +    int cpu = smp_processor_id();
>> +    struct irq_desc *desc;
>> +
>> +    if (!need_record_irq_counts(STATS_HARDIRQ))
>> +        return;
>> +
>> +    if (!test_bit(cpu, softlockup_hardirq_cpus)) {
>> +        counts = kmalloc_array(nr_irqs, sizeof(u32), GFP_ATOMIC);
>> +        if (!counts)
>> +            return;
>> +        for_each_irq_desc(i, desc) {
>> +            if (!desc)
>> +                continue;
>> +            counts[i] = desc->kstat_irqs ?
>> +                *this_cpu_ptr(desc->kstat_irqs) : 0;
>> +        }
>> +        __this_cpu_write(hardirq_counts, counts);
>> +        set_bit(cpu, softlockup_hardirq_cpus);
>> +    }
>> +}
>> +
>> +static void clear_potential_softlockup_hardirq(void)
>> +{
>> +    u32 *counts = __this_cpu_read(hardirq_counts);
>> +    int cpu = smp_processor_id();
>> +
>> +    if (test_bit(cpu, softlockup_hardirq_cpus)) {
>> +        kfree(counts);
>> +        counts = NULL;
>> +        __this_cpu_write(hardirq_counts, counts);
>> +        clear_bit(cpu, softlockup_hardirq_cpus);
>> +    }
>> +}
>> +
>> +/*
>> + * Mark that softlockup may occur
>> + */
>> +static void set_potential_softlockup(unsigned long now, unsigned long 
>> period_ts)
>> +{
>> +    if (time_after_eq(now, period_ts + get_softlockup_thresh() / 5))
>> +        set_potential_softlockup_hardirq();
>> +}
>> +
>> +static void clear_potential_softlockup(void)
>> +{
>> +    clear_potential_softlockup_hardirq();
>> +}
>> +
>> +static void print_hardirq_counts(void)
>> +{
>> +    u32 i;
>> +    struct irq_desc *desc;
>> +    u32 counts_diff;
>> +    u32 *counts = __this_cpu_read(hardirq_counts);
>> +    int cpu = smp_processor_id();
>> +    u32 hardirq_counts_top[NUM_HARDIRQ_REPORT] = {0, 0, 0, 0, 0};
>> +    int hardirq_top[NUM_HARDIRQ_REPORT] = {-1, -1, -1, -1, -1};
> "hardirq_counts_top" and "hardirq_top" seem like two members of a struct,
> working together to record the most suspicious irq, so wouldn't using a 
> struct make it clearer?
OK, I will define a struct.
>> +
>> +    if (test_bit(cpu, softlockup_hardirq_cpus)) {
>> +        /* Find the top NUM_HARDIRQ_REPORT most frequent interrupts */
>> +        for_each_irq_desc(i, desc) {
>> +            if (!desc)
>> +                continue;
>> +            counts_diff = desc->kstat_irqs ?
>> +                *this_cpu_ptr(desc->kstat_irqs) - counts[i] : 0;
>> +            find_counts_top(hardirq_counts_top, hardirq_top,
>> +                    counts_diff, i, NUM_HARDIRQ_REPORT);
>> +        }
>> +        /*
>> +         * We do not want the "watchdog: " prefix on every line,
>> +         * hence we use "printk" instead of "pr_crit".
>> +         */
>> +        printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. 
>> Most frequent HardIRQs:\n",
>> +            smp_processor_id(), HARDIRQ_PERCENT_THRESH);
>> +        for (i = 0; i < NUM_HARDIRQ_REPORT; i++) {
>> +            if (hardirq_top[i] == -1)
>> +                break;
>> +            desc = irq_to_desc(hardirq_top[i]);
>> +            if (desc && desc->action)
>> +                printk(KERN_CRIT "\t#%u: %-10u\tirq#%d(%s)\n",
>> +                    i+1, hardirq_counts_top[i],
>> +                    hardirq_top[i], desc->action->name);
>> +            else
>> +                printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n",
>> +                    i+1, hardirq_counts_top[i],
>> +                    hardirq_top[i]);
>> +        }
>> +        if (!need_record_irq_counts(STATS_HARDIRQ))
>> +            clear_potential_softlockup_hardirq();
>> +    }
>> +}
>> +
>>   #else
>>   static inline void update_cpustat(void) { }
>>   static inline void print_cpustat(void) { }
>> +static inline void set_potential_softlockup(unsigned long now, 
>> unsigned long period_ts) { }
>> +static inline void clear_potential_softlockup(void) { }
>>   #endif
>>   /* watchdog detector functions */
>> @@ -525,6 +674,7 @@ static DEFINE_PER_CPU(struct cpu_stop_work, 
>> softlockup_stop_work);
>>   static int softlockup_fn(void *data)
>>   {
>>       update_touch_ts();
>> +    clear_potential_softlockup();
>>       complete(this_cpu_ptr(&softlockup_completion));
>>       return 0;