Skip to contentSkip to conversation

watchdog/hardlockup: set watchdog_hardlockup_warned to true as early as possible

PATCH In review Liu Song 3 participants 4 messages Aug 5, 2023 – Aug 7, 2023

lore
Share & tools
lore

Liu Song

author
#0 lore raw

PATCH 1/1 watchdog/hardlockup: set watchdog_hardlockup_warned to true as early as possible

Since we want to ensure only printing hardlockups once, it is necessary to set "watchdog_hardlockup_warned" to true as early as possible.

---
 kernel/watchdog.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/watchdog.c +2 -2
index 25d5627a6580..c4795f2d148c 100644
@@ -180,6 +180,8 @@ void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 		/* Only print hardlockups once. */
 		if (per_cpu(watchdog_hardlockup_warned, cpu))
 			return;
+		else
+			per_cpu(watchdog_hardlockup_warned, cpu) = true;
 
 		pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
 		print_modules();
@@ -206,8 +208,6 @@ void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 
 		if (hardlockup_panic)
 			nmi_panic(regs, "Hard LOCKUP");
-
-		per_cpu(watchdog_hardlockup_warned, cpu) = true;
 	} else {
 		per_cpu(watchdog_hardlockup_warned, cpu) = false;
 	}
Signature · 1 line
2.19.1.6.gb485710b
#1 lore raw

kernel/watchdog.c void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs) · line 184

Quoted 13 lines · On Sun, 6 Aug 2023 00:01:44 +0800 Liu Song <[email protected]> wrote:

Since we want to ensure only printing hardlockups once, it is necessary to set "watchdog_hardlockup_warned" to true as early as possible.

...

kernel/watchdog.c +2 -0
@@ -180,6 +180,8 @@ void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 		/* Only print hardlockups once. */
 		if (per_cpu(watchdog_hardlockup_warned, cpu))
 			return;
+		else
+			per_cpu(watchdog_hardlockup_warned, cpu) = true;

The "else" is unneeded.

kernel/watchdog.c void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs) · line 213

Quoted 11 lines
kernel/watchdog.c +0 -2
 		pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
 		print_modules();
@@ -206,8 +208,6 @@ void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 
 		if (hardlockup_panic)
 			nmi_panic(regs, "Hard LOCKUP");
-
-		per_cpu(watchdog_hardlockup_warned, cpu) = true;
 	} else {
 		per_cpu(watchdog_hardlockup_warned, cpu) = false;
 	}

When resending, please tell us some more about the effects of the change. Presumably there are circumstances in which excess output is produced? If so, describe these circumstances and the observed effects.

author
#2 lore raw
Quoted 4 lines · Andrew Morton

When resending, please tell us some more about the effects of the change. Presumably there are circumstances in which excess output is produced? If so, describe these circumstances and the observed effects.

Hi,

I haven't found duplicate warnings in the real environment.

However, considering that when system occurs hard lockup is basically abnormal, it

seems more reasonable to set "watchdog_hardlockup_warned" to ture, rather than

waiting for all kinds of information to be printed.

Thanks

#3 lore raw
Quoted 11 lines · Liu Song
Quoted 4 lines · Andrew Morton

When resending, please tell us some more about the effects of the change. Presumably there are circumstances in which excess output is produced? If so, describe these circumstances and the observed effects.

Hi,

I haven't found duplicate warnings in the real environment.

Quoted 7 lines

However, considering that when system occurs hard lockup is basically abnormal, it

seems more reasonable to set "watchdog_hardlockup_warned" to ture, rather than

waiting for all kinds of information to be printed.

I believe that this is not needed.

watchdog_hardlockup_check(cpu, regs) is called on a CPU periodically. There are two callers:

+ buddy detector checks the particular CPU when the solflockup's hrtimer callback is called. See watchdog_hardlockup_kick() in watchdog_timer_fn().

+ perf detector checks the particular CPU from a perf callback, see watchdog_overflow_callback().

Neither timer nor perf callbacks might be nested. They are naturally serialized on a given CPU. So, races are not possible in this case.

Best Regards,
Petr