watchdog/hardlockup/perf: Use atomics to track in-use cpu counter
authorDon Zickus <dzickus@redhat.com>
Wed, 1 Nov 2017 18:11:27 +0000 (14:11 -0400)
committerThomas Gleixner <tglx@linutronix.de>
Wed, 1 Nov 2017 20:18:40 +0000 (21:18 +0100)
Guenter reported:
  There is still a problem. When running
    echo 6 > /proc/sys/kernel/watchdog_thresh
    echo 5 > /proc/sys/kernel/watchdog_thresh
  repeatedly, the message

   NMI watchdog: Enabled. Permanently consumes one hw-PMU counter.

  stops after a while (after ~10-30 iterations, with fluctuations).
  Maybe watchdog_cpus needs to be atomic ?

That's correct as this again is affected by the asynchronous nature of the
smpboot thread unpark mechanism.

CPU 0 CPU1 CPU2
write(watchdog_thresh, 6)
  stop()
    park()
  update()
  start()
    unpark()
thread->unpark()
  cnt++;
write(watchdog_thresh, 5) thread->unpark()
  stop()
    park() thread->park()
   cnt--;   cnt++;
  update()
  start()
    unpark()

That's not a functional problem, it just affects the informational message.

Convert watchdog_cpus to atomic_t to prevent the problem

Reported-and-tested-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Don Zickus <dzickus@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20171101181126.j727fqjmdthjz4xk@redhat.com
kernel/watchdog_hld.c

index a7f137c1933aa207175f8643bc19f503b09b6ea4..a84b205fac9a0be08c6e96570857260efe5b3f41 100644 (file)
@@ -12,6 +12,7 @@
 #define pr_fmt(fmt) "NMI watchdog: " fmt
 
 #include <linux/nmi.h>
 #define pr_fmt(fmt) "NMI watchdog: " fmt
 
 #include <linux/nmi.h>
+#include <linux/atomic.h>
 #include <linux/module.h>
 #include <linux/sched/debug.h>
 
 #include <linux/module.h>
 #include <linux/sched/debug.h>
 
@@ -25,7 +26,7 @@ static DEFINE_PER_CPU(struct perf_event *, dead_event);
 static struct cpumask dead_events_mask;
 
 static unsigned long hardlockup_allcpu_dumped;
 static struct cpumask dead_events_mask;
 
 static unsigned long hardlockup_allcpu_dumped;
-static unsigned int watchdog_cpus;
+static atomic_t watchdog_cpus = ATOMIC_INIT(0);
 
 void arch_touch_nmi_watchdog(void)
 {
 
 void arch_touch_nmi_watchdog(void)
 {
@@ -189,7 +190,8 @@ void hardlockup_detector_perf_enable(void)
        if (hardlockup_detector_event_create())
                return;
 
        if (hardlockup_detector_event_create())
                return;
 
-       if (!watchdog_cpus++)
+       /* use original value for check */
+       if (!atomic_fetch_inc(&watchdog_cpus))
                pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 
        perf_event_enable(this_cpu_read(watchdog_ev));
                pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
 
        perf_event_enable(this_cpu_read(watchdog_ev));
@@ -207,7 +209,7 @@ void hardlockup_detector_perf_disable(void)
                this_cpu_write(watchdog_ev, NULL);
                this_cpu_write(dead_event, event);
                cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
                this_cpu_write(watchdog_ev, NULL);
                this_cpu_write(dead_event, event);
                cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
-               watchdog_cpus--;
+               atomic_dec(&watchdog_cpus);
        }
 }
 
        }
 }