Re: Bisected post-3.9 regression: Resume takes 5 times as much time as with v3.9

["Bjørn Mork" <bjorn@mork.no>]

Borislav Petkov <bp@alien8.de> writes:

> On Sun, May 12, 2013 at 06:13:34PM +0200, Bjørn Mork wrote:
>> diff --git a/kernel/rcutree.c b/kernel/rcutree.c
>> index 6934238..2dcbf84 100644
>> --- a/kernel/rcutree.c
>> +++ b/kernel/rcutree.c
>> @@ -3103,9 +3103,11 @@ static int rcu_pm_notify(struct notifier_block *self,
>>  {
>>  	switch (action) {
>>  	case PM_HIBERNATION_PREPARE:
>> +	case PM_SUSPEND_PREPARE:
>>  		rcu_expedited = 1;
>>  		break;
>> -	case PM_POST_RESTORE:
>> +	case PM_POST_HIBERNATION:
>> +	case PM_POST_SUSPEND:
>>  		rcu_expedited = 0;
>>  		break;
>>  	default:
>
> If I'm reading Documentation/power/notifiers.txt correctly, we only need
> PM_HIBERNATION_PREPARE when we go to sleep (whatever hibernation method
> we use) and PM_POST_HIBERNATION when we restore.

Well, that's not the way I read it. And testing also supports that.

Adding the above to your patch makes restore from suspend use < 3
seconds again.  Using only PM_HIBERNATION_PREPARE had no effect on the
restore from suspend time, still measured at around 15 seconds on my
laptop.

Note that I did not test hibernation at all.  But based on your patch I
assume the issue is exactly the same.

> I don't think it matters for expediting RCU grace periods whether
> we had an error during resume or not and I'd basically want to set
> rcu_expedited to 0 unconditionally when resuming..

Yes, and I believe that's what PM_POST_HIBERNATION will achieve if I
read the docs correctly.  But I guess this should be sanity checked via
Rafael or someone else knowing how these notifiers work.


Bjørn