On Tue, Jun 20, 2017 at 09:45:23AM -0700, Paul E. McKenney wrote: > On Sun, Jun 18, 2017 at 06:40:00AM -0400, Tejun Heo wrote: > > Hello, > > > > On Sat, Jun 17, 2017 at 10:31:05AM -0700, Paul E. McKenney wrote: > > > On Sat, Jun 17, 2017 at 07:53:14AM -0400, Tejun Heo wrote: > > > > Hello, > > > > > > > > On Fri, Jun 16, 2017 at 10:36:58AM -0700, Paul E. McKenney wrote: > > > > > And no test failures from yesterday evening. So it looks like we get > > > > > somewhere on the order of one failure per 138 hours of TREE07 > > > > > rcutorture > > > > > runtime with your printk() in the mix. > > > > > > > > > > Was the above output from your printk() output of any help? > > > > > > > > Yeah, if my suspicion is correct, it'd require new kworker creation > > > > racing against CPU offline, which would explain why it's so difficult > > > > to repro. Can you please see whether the following patch resolves the > > > > issue? > > > > > > That could explain why only Steve Rostedt and I saw the issue. As far > > > as I know, we are the only ones who regularly run CPU-hotplug stress > > > tests. ;-) > > > > I was a bit confused. It has to be racing against either new kworker > > being created on the wrong CPU or rescuer trying to migrate to the > > CPU, and it looks like we're mostly seeing the rescuer condition, but, > > yeah, this would only get triggered rarely. Another contributing > > factor could be the vmstat work putting on a workqueue w/ rescuer > > recently. It runs quite often, so probably has increased the chance > > of hitting the right condition. > > Sounds like too much fun! ;-) > > But more constructively... If I understand correctly, it is now possible > to take a CPU partially offline and put it back online again. This should > allow much more intense testing of this sort of interaction. > > And no, I haven't yet tried this with RCU because I would probably need > to do some mix of just-RCU online/offline and full-up online-offline. > Plus RCU requires pretty much a full online/offline cycle to fully > exercise it. :-/ > > > > I have a weekend-long run going, but will give this a shot overnight on > > > Monday, Pacific Time. Thank you for putting it together, looking forward > > > to seeing what it does! > > > > Thanks a lot for the testing and patience. Sorry that it took so > > long. I'm not completely sure the patch is correct. It might have to > > be more specifc about which type of migration or require further > > synchronization around migration, but hopefully it'll at least be able > > to show that this was the cause of the problem. > > And last night's tests had no failures. Which might actually mean > something, will get more info when I run without your patch this > evening. ;-)
And it didn't fail without the patch, either. 45 hours of test vs. 60 hours with the patch. This one is not going to be easy to prove either way. I will try again this evening without the patch and see what that gets us. Thanx, Paul