2013/5/21 Srivatsa S. Bhat <srivatsa.b...@linux.vnet.ibm.com>: > On 05/20/2013 09:31 PM, Frederic Weisbecker wrote: >> When the watchdog code is boot-disabled by the user, for example >> through the 'nmi_watchdog=0' boot option, the setup() callback of >> the watchdog kthread requests to park the task, and that until the >> user later re-enables the watchdog through sysctl or procfs. >> >> However the parking request is not well handled when done from >> the setup() callback. After ->setup() is called, the generic smpboot >> kthread loop directly tries to call the thread function or wait >> for some event if ->thread_should_run() is false. >> >> In the case of the watchdog kthread, ->thread_should_run() returns >> false and the kthread goes to sleep and wait for the watchdog timer >> to wake it up. But the timer is not enabled since the user requested >> to disable the watchdog. We want the kthread to park instead of waiting >> for events that can't happen. >> >> As a result, later unpark requests after sysctl write through >> 'sysctl -w kernel.watchdog=1' won't wake up/unpark the task as >> expected, since it's not parked anyway, leaving the value modified >> without triggering any action. >> >> We could workaround some solution in the watchdog code like forcing >> one pass to the thread function and immediately return to park. >> >> But supporting parking requests from ->setup() or ->unpark() > > unpark() can already do a proper park, because immediately after > coming out of the parked state, the 'continue' statement helps > re-evaluate the stop/park condition.
Right. > > So this fix is only for the ->setup() case. > >> callbacks look like proper way to implement cancellation in >> general. So let's fix it that way. >> > > Sounds good to me. > > Reviewed-by: Srivatsa S. Bhat <srivatsa.b...@linux.vnet.ibm.com> > > But I wonder what nmi_watchdog=0 should actually mean, semantically.. > The current code works as if the user asked us not to run the watchdog > threads, but it could as well be interpreted as if the user does not > want to run *any* watchdog-related *code*. In that case, ideally we > should *unregister* the watchdog threads, instead of just parking them. > And when the user enables them again via sysctl/procfs, we should > register the watchdog threads with the smpboot infrastructure. > > I'm not saying that the current semantics is wrong, but if we really > implement it the other way I proposed above, then we won't have to deal > with weird issues like ->setup() code wanting to park, and watchdog > threads unparking and parking themselves on every CPU hotplug operation, > despite the fact that the user specified nmi_watchdog=0 on the kernel > command line. That's an interesting point. It seems that using smpboot register/unregister operations for sysctl control would be more appropriate and less complicated. I'm going to give it a try. Thanks! -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
