Commit:

  b94f5118

 ("kernel/watchdog: prevent false hardlockup on overloaded system")

tries to fix the following issue:

proc_write()
   set_sample_period()    <--- New sample period becoms visible
			  <----- Broken starts
   proc_watchdog_update()
     watchdog_enable_all_cpus()		watchdog_hrtimer_fn()
     update_watchdog_all_cpus()		   restart_timer(sample_period)
        watchdog_park_threads()

					thread->park()
					  disable_nmi()
			  <----- Broken ends

The reason why this is broken is that the update of the watchdog threshold
becomes immediately effective and visible for the hrtimer function which
uses that value to rearm the timer. But the NMI/perf side still uses the
old value up to the point where it is disabled. If the rate has been
lowered then the NMI can run fast enough to 'detect' a hard lockup because
the timer has not fired due to the longer period.

The patch 'fixed' this by adding a variable:

proc_write()
   set_sample_period()
					<----- Broken starts
   proc_watchdog_update()
     watchdog_enable_all_cpus()		watchdog_hrtimer_fn()
     update_watchdog_all_cpus()		   restart_timer(sample_period)
         watchdog_park_threads()
	  park_in_progress = 1
					<----- Broken ends
				        nmi_watchdog()
					  if (park_in_progress)
					     return;

The only effect of this variable was to make the window where the breakage
can hit small enough that it was not longer observable in testing. From a
correctness point of view it is a pointless bandaid which merily papers
over the root cause: the unsychronized update of the variable.

Looking deeper into the related code pathes unearthed similar problems in
the watchdog_start()/stop() functions.

 watchdog_start()
	perf_nmi_event_start()
	hrtimer_start()

 watchdog_stop()
	hrtimer_cancel()
	perf_nmi_event_stop()

In both cases the call order is wrong because if the tasks gets preempted
or the VM gets scheduled out long enough after the first call, then there is
a chance that the next NMI will see a stale hrtimer interrupt count and
trigger a false positive hard lockup splat.

Get rid of park_in_progress so the code can be gradually deobfuscated and
pruned from several layers of duct tape papering over the root cause,
which has been either ignored or not understood at all.

Once this is removed the underlying problem will be fixed by rewriting the
proc interface to do a proper synchronized update.

Address the start/stop() ordering problem as well by reverting the call
order, so this part is at least correct now.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Don Zickus <dzickus@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Siewior <bigeasy@linutronix.de>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Link: http://lkml.kernel.org/r/alpine.DEB.2.20.1709052038270.2393@nanos


Signed-off-by: Ingo Molnar <mingo@kernel.org>