[PATCH] kernel/watchdog: Prevent false hardlockup on overloaded system

[Don Zickus <dzickus@redhat.com>]

On an overloaded system, it is possible that a change in the watchdog threshold
can be delayed long enough to trigger a false positive.

This can easily be achieved by having a cpu spinning indefinitely on a task,
while another cpu updates watchdog threshold.

What happens is while trying to park the watchdog threads, the hrtimers on the
other cpus trigger and reprogram themselves with the new slower watchdog
threshold.  Meanwhile, the nmi watchdog is still programmed with the old faster
threshold.

Because the one cpu is blocked, it prevents the thread parking on the other
cpus from completing, which is needed to shutdown the nmi watchdog and
reprogram it correctly.  As a result, a false positive from the nmi watchdog is
reported.

Fix this by setting a park_in_progress flag to block all lockups
until the parking is complete.

Fix provided by Ulrich Obergfell.

Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Don Zickus <dzickus@redhat.com>
---
 include/linux/nmi.h   | 1 +
 kernel/watchdog.c     | 9 +++++++++
 kernel/watchdog_hld.c | 3 +++
 3 files changed, 13 insertions(+)

diff --git a/include/linux/nmi.h b/include/linux/nmi.h
index 0ea0a38..67e3392 100644
--- a/include/linux/nmi.h
+++ b/include/linux/nmi.h
@@ -111,6 +111,7 @@ static inline bool trigger_single_cpu_backtrace(int cpu)
 extern unsigned long watchdog_enabled;
 extern DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
 extern unsigned long *watchdog_cpumask_bits;
+extern atomic_t park_in_progress;
 extern int sysctl_softlockup_all_cpu_backtrace;
 extern int sysctl_hardlockup_all_cpu_backtrace;
 struct ctl_table;
diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 4ea7752..fea6ebd 100644
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -52,6 +52,8 @@
 #define for_each_watchdog_cpu(cpu) \
 	for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
 
+atomic_t park_in_progress = ATOMIC_INIT(0);
+
 /*
  * The 'watchdog_running' variable is set to 1 when the watchdog threads
  * are registered/started and is set to 0 when the watchdog threads are
@@ -246,6 +248,9 @@ static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 	int duration;
 	int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 
+	if (atomic_read(&park_in_progress) != 0)
+		return HRTIMER_NORESTART;
+
 	/* kick the hardlockup detector */
 	watchdog_interrupt_count();
 
@@ -453,12 +458,16 @@ static int watchdog_park_threads(void)
 {
 	int cpu, ret = 0;
 
+	atomic_set(&park_in_progress, 1);
+
 	for_each_watchdog_cpu(cpu) {
 		ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
 		if (ret)
 			break;
 	}
 
+	atomic_set(&park_in_progress, 0);
+
 	return ret;
 }
 
diff --git a/kernel/watchdog_hld.c b/kernel/watchdog_hld.c
index cd690fb..06d7c09 100644
--- a/kernel/watchdog_hld.c
+++ b/kernel/watchdog_hld.c
@@ -95,6 +95,9 @@ static void watchdog_overflow_callback(struct perf_event *event,
 	/* Ensure the watchdog never gets throttled */
 	event->hw.interrupts = 0;
 
+	if (atomic_read(&park_in_progress) != 0)
+		return;
+
 	if (__this_cpu_read(watchdog_nmi_touch) == true) {
 		__this_cpu_write(watchdog_nmi_touch, false);
 		return;
-- 
1.8.4