Re: Why does the 2.6.37 scheduler hate RT tasks?

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Feb 14, 2011 at 10:55:49PM +0800, Yong Zhang wrote:
> On Tue, Feb 01, 2011 at 05:10:49PM -0800, Paul E. McKenney wrote:
> > Hello, Peter,
> > 
> > If I create an SCHED_FIFO-priority kthread in pre-2.6.37 kernels,
> > things work nicely.  If I try in in later kernels, __disable_runtime()
> > gets upset occasionally during CPU-hotplug calls when CPUs are inserted.
> > If I make the exact same kthreads run SCHED_NORMAL, no problem.
> > 
> > Unfortunately, I need an RT kthread so that I can do priority boosting.
> > 
> > My next thing to try is to demote my kthreads to SCHED_NORMAL at the
> > beginning of a CPU-hotplug event and promote them back to SCHED_FIFO at
> > the end of the CPU-hotplug event, which would leave the system vulnerable
> > to RCU priority inversion during CPU-hotplug events.  This would be
> > better than the current behavior, but I thought that I should check with
> > you to see if there is a better approach.
> 
> Just want to know the status of this issue :)
> 
> Does commit 06c3bc655697b19521901f9254eb0bbb2c67e7e8 cure it?

Ah, good to see that this got in!  I did extract it from the diagnostic
patch that Peter sent out, and it does seem to cure the issue.

							Thanx, Paul