Re: cpu stopper threads and load balancing leads to deadlock

From: Peter Zijlstra
Date: Thu May 03 2018 - 10:52:34 EST


On Thu, May 03, 2018 at 07:39:41AM -0700, Paul E. McKenney wrote:
> Huh.
>
> No, RCU_NONIDLE() only works for idle, not for offline.

Oh bummer..

> Maybe... Let me take a look. There must be some way to mark a
> specific lock acquisition and release as being lockdep-invisible...

But I suspect it flags a real issue. Shutting it up isn't the issue as
such. There could be a concurrent sched domain rebuild while we do this
setup, and if we free that domain while we're looking at it, bad things
happen.

... I wonder what brought this failure to pass now. This code has been
like this a for a long time.