On Tue, Jul 03, 2018 at 07:36:16PM +0530, Vaidyanathan Srinivasan wrote: > * Gautham R Shenoy <e...@linux.vnet.ibm.com> [2018-07-03 10:54:16]: > > > From: "Gautham R. Shenoy" <e...@linux.vnet.ibm.com> > > > > In the situations where snooze is the only cpuidle state due to > > firmware not exposing any platform idle states, the idle CPUs will > > remain in snooze for a long time with interrupts disabled causing the > > Hard-lockup detector to complain. > > snooze_loop() will spin in SMT low priority with interrupt enabled. We > have local_irq_enable() before we get into the snooze loop. > Since this is a polling state, we should wakeup without an interrupt > and hence we set TIF_POLLING_NRFLAG as well. >
You are right. We have a local_irq_enable() inside the snooze_loop. > > > watchdog: CPU 51 detected hard LOCKUP on other CPUs 59 > > watchdog: CPU 51 TB:535296107736, last SMP heartbeat TB:527472229239 > > (15281ms ago) > > watchdog: CPU 59 Hard LOCKUP > > watchdog: CPU 59 TB:535296252849, last heartbeat TB:526554725466 (17073ms > > ago) > > hmm.. not sure why watchdog will complain, maybe something more is > going on. Will look into this Vaidy. > > > Fix this by adding CPUIDLE_FLAG_POLLING flag to the state, so that the > > cpuidle governor will do the right thing, such as not stopping the > > tick if it is going to put the idle cpu to snooze. > > > > Reported-by: Akshay Adiga <akshay.ad...@linux.vnet.ibm.com> > > Cc: Nicholas Piggin <npig...@gmail.com> > > Signed-off-by: Gautham R. Shenoy <e...@linux.vnet.ibm.com> > > --- > > drivers/cpuidle/cpuidle-powernv.c | 1 + > > 1 file changed, 1 insertion(+) > > > > diff --git a/drivers/cpuidle/cpuidle-powernv.c > > b/drivers/cpuidle/cpuidle-powernv.c > > index d29e4f0..b73041b 100644 > > --- a/drivers/cpuidle/cpuidle-powernv.c > > +++ b/drivers/cpuidle/cpuidle-powernv.c > > @@ -156,6 +156,7 @@ static int stop_loop(struct cpuidle_device *dev, > > { /* Snooze */ > > .name = "snooze", > > .desc = "snooze", > > + .flags = CPUIDLE_FLAG_POLLING, > > .exit_latency = 0, > > .target_residency = 0, > > .enter = snooze_loop }, > > Adding the CPUIDLE_FLAG_POLLING is good and enables more optimization. > But the reason that we spin with interrupt disabled does not seem > right. Fair point. > > --Vaidy >