Skip to content

hung_task debugging: Add tracepoint to report the hang

Oleg Nesterov lkml 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest1 patch 2 repliesIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/1hung_task debugging: Add tracepoint to report the hanginclude/trace/events/sched.h kernel/hung_task.c+23 -02 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.

Read the v1 thread

Cover letter [PATCH 0/1] hung_task debugging: Add tracepoint to report the hang

Hi,

We have a feature request, the customer needs something more hookable than just printk's from check_hung_task() to implement the user-space watchdog which can potentially resolve the problems which caused the hang.

The patch simply adds a tracepoint into check_hung_task(), do you think we can (ab)use tp this way?

And I am just curious, perhaps another patch (below) makes sense too?

Oleg.

------------------------------------------------------------------------------
[PATCH] hung_task debugging: Do not report the same task twice

If a task hangs, check_hung_task() will blame it again and again until it wakes up or sysctl_hung_task_warnings becomes zero. IMO, this just adds the unnecessary noise and if another task hangs after sysctl_hung_task_timeout_secs * sysctl_hung_task_warnings it won't be reported.

With this patch check_hung_task() simply sets the most significant bit in ->last_switch_count to mark this task as "already reported". This bit is cleared if we notice the change in nvcsw/nivcsw, so we do not skip this task if it hangs again later.

Note that we also ignore the MSB in switch_count, we need this to avoid the false-positive "already reported". This means that ->last_switch_count is not necessarily equal to ->nvcsw + ->nivcsw but we do not care, we have enough bits to notice the change.

And this allows to remove the special switch_count == 0 case, just we need to initialize ->last_switch_count = LONG_MIN.

---
 kernel/fork.c      |    2 +-
 kernel/hung_task.c |   19 +++++++++----------
 2 files changed, 10 insertions(+), 11 deletions(-)
kernel/fork.c +1 -1
index 8531609..aea397b 100644
@@ -866,7 +866,7 @@ static int copy_mm(unsigned long clone_flags, struct task_struct *tsk)
 	tsk->min_flt = tsk->maj_flt = 0;
 	tsk->nvcsw = tsk->nivcsw = 0;
 #ifdef CONFIG_DETECT_HUNG_TASK
-	tsk->last_switch_count = tsk->nvcsw + tsk->nivcsw;
+	tsk->last_switch_count = LONG_MIN; /* see check_hung_task() */
 #endif
 
 	tsk->mm = NULL;
kernel/hung_task.c +9 -10
index 3952ab1..0f6233c 100644
@@ -72,7 +72,7 @@ static struct notifier_block panic_block = {
 
 static void check_hung_task(struct task_struct *t, unsigned long timeout)
 {
-	unsigned long switch_count = t->nvcsw + t->nivcsw;
+	unsigned long switch_count;
 
 	/*
 	 * Ensure the task is not frozen.
@@ -81,19 +81,18 @@ static void check_hung_task(struct task_struct *t, unsigned long timeout)
 	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
 	    return;
 
-	/*
-	 * When a freshly created task is scheduled once, changes its state to
-	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
-	 * musn't be checked.
-	 */
-	if (unlikely(!switch_count))
-		return;
-
-	if (switch_count != t->last_switch_count) {
+	/* Ignore MSB, see below. LONG_MAX = ~LONG_MIN. */
+	switch_count = (t->nvcsw + t->nivcsw) & LONG_MAX;
+	if (switch_count != (LONG_MAX & t->last_switch_count)) {
 		t->last_switch_count = switch_count;
 		return;
 	}
 
+	/* We use MSB to mark this task as already reported. */
+	if (t->last_switch_count & LONG_MIN)
+		return;
+	t->last_switch_count |= LONG_MIN;
+
 	trace_sched_process_hang(t);
 
 	if (!sysctl_hung_task_warnings)
Signature · 1 line
1.5.5.1