Re: [RFC, PATCH -rt] NMI-safe mb- and atomic-free RT RCU

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

On Thu, Jul 27, 2006 at 12:53:56PM -0700, Bill Huey wrote:
> On Thu, Jul 27, 2006 at 08:46:37AM -0700, Paul E. McKenney wrote:
> > On Thu, Jul 27, 2006 at 12:00:13PM +0100, Esben Nielsen wrote:
> > > No, RT tasks can still preempt the RCU read side lock. But SCHED_OTHER and 
> > > SCHED_BATCH can't. You can also the RCU read side boosting prioritiy 
> > > dynamic and let the system adjust it or just let the admin adjust it.
> > 
> > Fair enough -- I misread MAX_RT_PRIO as MAX_PRIO.
> > 
> > This approach I can get behind -- my thought has been to boost to
> > either MAX_RT_PRIO or MAX_RT_PRIO-1 when preempt_schedule() sees that
> > it is preempting an RCU read-side critical section.
> > 
> > So I agree with you on at least one point!  ;-)
> 
> This is the approach that I suggested to you, Paul, at OLS after your talk.

Yep.  I have been messing with it for some time, see the RCUboost*
patches in http://www.rdrop.com/users/paulmck/patches/ if you wish
to see a large number of subtly broken ways of approaching this.  ;-)

> Again, if you go about this path then you might think about extending the
> scheduler to have an additional parameter regarding a preemption threshold
> instead of doing this stuff indirectly with priority manipulations like
> the above. It was something that I was considering when I was doing my
> Linux kernel preemption project to fix the problems with RCU-ed read side
> thread migrating to another CPU.
>
> If folks go down this discussion track, it's going to open a can of
> scheduling worms with possiblities (threshold, priority, irq-thread
> priority, global run queue for SCHED_FIFO tasks) pushing into global run
> queue logic stuff. It's a bit spooky for the Linux kernel. Some of the
> thread migration pinning stuff with per CPU locks was rejected by Linus
> way back.

I have been dealing with worms enough for me without the preemption
threshold.  I agree that it might well be useful, but I am getting
my full quota of worms and spookiness with my current approach.  ;-)

> > A possible elaboration would be to keep a linked list of tasks preempted
> > in their RCU read-side critical sections so that they can be further
> > boosted to the highest possible priority (numerical value of zero,
> > not sure what the proper symbol is) if the grace period takes too many
> > jiffies to complete.  Another piece is priority boosting when blocking
> > on a mutex from within an RCU read-side critical section.
> 
> I'm not sure how folks feel about putting something like that in the
> scheduler path since it's such a specialized cases. Some of the scheduler
> folks might come out against this.

They might well.  And the resulting discussion might reveal a better
way.  Or it might well turn out that the simple approach of boosting
to an intermediate level without the list will suffice.

> > Doing it efficiently is the difficulty, particularly for tickless-idle
> > systems where CPUs need to be added and removed on a regular basis.
> > Also, what locking design would you use in order to avoid deadlock?
> > There is a hotplug mutex, but seems like you might need to acquire some
> > number of rq mutexes as well.
> 
> I'm not understanding what you exactly mean by tickless idle systems.

Some architectures power CPUs down (including stopping the scheduling
clock interrupt) when the CPU goes idle for a sufficient time.  The
CPU has sort of been hotplug-removed, but not quite.  In any case, once
the CPU is in that state, RCU needs to count it out of the grace-period
detection.

One difference between the tickless idle and the hot-unplug state is that
a CPU can (and does) enter and exit tickless idle state quite frequently.

> Are you talking about isolating a CPU for SCHED_FIFO and friends tasks
> only as in the CPU reservation stuff but with ticks off in many proprietary
> RTOSes ? Don't mean to be tangental here, I just need clarification.

No.  Although that was something I was messing with a couple of years back,
I am now leaving that to others.

							Thanx, Paul

> > Another approach I am looking at does not permit rcu_read_lock() in
> > NMI/SMI/hardirq, but is much simpler.  Its downside is that it cannot
> > serve as common code between CONFIG_PREEMPT_RT and CONFIG_PREEMPT.
> 
> bill
>