> If the HW watchdog is running, the start function should not be called
> on open. It would be called after the watchdog was stopped and is then
> started again. With that, opening the watchdog the first time would not
> call this function, and started_by_fw would remain true. Closing it would
> then stopp the watchdog and call pm_runtime_put(). The next open() would
> hit the above case, and not call pm_runtime_get_sync(). pm would then be
> out of sync.
> 
> What am I missing ?

Obviously, I was missing something. I didn't see that the watchdog core
calls ->ping instead of ->start when WDOG_HW_RUNNING is set. And once
I knew about it, I even found it in the documentation:

+  Note: when you register the watchdog timer device with this bit set,
+  then opening /dev/watchdog will skip the start operation but send a keepalive
+  request instead.

This makes all my code to keep the refcnt balanced very moot.

Sorry for the noise. I will fix up things and retest.