Re: [PATCH 17/41] rcu: Restart tick if we enqueue a callback in anohz/cpuset CPU

[Frederic Weisbecker]

On Wed, May 23, 2012 at 09:15:14AM -0700, Paul E. McKenney wrote:
> On Wed, May 23, 2012 at 04:03:36PM +0200, Frederic Weisbecker wrote:
> > On Tue, May 22, 2012 at 10:30:47AM -0700, Paul E. McKenney wrote:
> > > On Tue, May 22, 2012 at 10:27:14AM -0700, Paul E. McKenney wrote:
> > > > On Tue, May 01, 2012 at 01:54:51AM +0200, Frederic Weisbecker wrote:
> > > > > If we enqueue an rcu callback, we need the CPU tick to stay
> > > > > alive until we take care of those by completing the appropriate
> > > > > grace period.
> > > > > 
> > > > > Thus, when we call_rcu(), send a self IPI that checks rcu_needs_cpu()
> > > > > so that we restore a periodic tick behaviour that can take care of
> > > > > everything.
> > > > 
> > > > Ouch, I hadn't considered RCU callbacks being posted from within an
> > > > extended quiescent state.  I guess I need to make __call_rcu() either
> > > > complain about this or handle it correctly...  It would -usually- be
> > > > harmless, but there is getting to be quite a bit of active machinery
> > > > in the various idle loops, so just assuming that it cannot happen is
> > > > probably getting to be an obsolete assumption.
> > > 
> > > Adaptive ticks does restart the tick upon entering the kernel, correct?
> > 
> > No, it keeps the tick down. The tick is restarted only if it's needed:
> > when more than one task are on the runqueue, a posix cpu timer is running,
> > a CPU needs the current one to report a quiescent state, etc...
> 
> Ah, I didn't realize that you didn't restart the tick upon entry to the
> kernel.  So this is why you need the IPI -- because there is no tick, if
> the system call runs for a long time, RCU is not guaranteed to make any
> progress on that CPU.
> 
> In the common case, this will not be a problem because system calls
> normally spend a short amount of time in the kernel, so normally RCU's
> dyntick-idle detection will handle this case.  The exception to this
> rule is when there is a long CPU-bound code path in the kernel, where
> "long" means many milliseconds.  In this exception case, this CPU needs
> to be interrupted or whatever is needed to force the CPU to progress
> through RCU.

Exactly!
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/