Re: [PATCH RFC tip/core/rcu 0/5] Expedited grace periods encouraging normal ones

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

On Wed, Jul 01, 2015 at 08:43:54AM -0700, Josh Triplett wrote:
> On Tue, Jun 30, 2015 at 08:37:01PM -0700, Paul E. McKenney wrote:
> > On Tue, Jun 30, 2015 at 05:42:14PM -0700, Josh Triplett wrote:
> > > On Tue, Jun 30, 2015 at 05:15:58PM -0700, Paul E. McKenney wrote:
> > > > On Tue, Jun 30, 2015 at 04:46:33PM -0700, josh@joshtriplett.org wrote:
> > > > > On Tue, Jun 30, 2015 at 03:12:24PM -0700, Paul E. McKenney wrote:
> > > > > > On Tue, Jun 30, 2015 at 03:00:15PM -0700, josh@joshtriplett.org wrote:
> > > > > > > On Tue, Jun 30, 2015 at 02:48:05PM -0700, Paul E. McKenney wrote:
> > > > > > > > Hello!
> > > > > > > > 
> > > > > > > > This series contains some highly experimental patches that allow normal
> > > > > > > > grace periods to take advantage of the work done by concurrent expedited
> > > > > > > > grace periods.  This can reduce the overhead incurred by normal grace
> > > > > > > > periods by eliminating the need for force-quiescent-state scans that
> > > > > > > > would otherwise have happened after the expedited grace period completed.
> > > > > > > > It is not clear whether this is a useful tradeoff.  Nevertheless, this
> > > > > > > > series contains the following patches:
> > > > > > > 
> > > > > > > While it makes sense to avoid unnecessarily delaying a normal grace
> > > > > > > period if the expedited machinery has provided the necessary delay, I'm
> > > > > > > also *deeply* concerned that this will create a new class of
> > > > > > > nondeterministic performance issues.  Something that uses RCU may
> > > > > > > perform badly due to grace period latency, but then suddenly start
> > > > > > > performing well because an unrelated task starts hammering expedited
> > > > > > > grace periods.  This seems particularly likely during boot, for
> > > > > > > instance, where RCU grace periods can be a significant component of boot
> > > > > > > time (when you're trying to boot to userspace in small fractions of a
> > > > > > > second).
> > > > > > 
> > > > > > I will take that as another vote against.  And for a reason that I had
> > > > > > not yet come up with, so good show!  ;-)
> > > > > 
> > > > > Consider it a fairly weak concern against.  Increasing performance seems
> > > > > like a good thing in general; I just don't relish the future "feels less
> > > > > responsive" bug reports that take a long time to track down and turn out
> > > > > to be "this completely unrelated driver was loaded and started using
> > > > > expedited grace periods".
> > > > 
> > > > From what I can see, this one needs a good reason to go in, as opposed
> > > > to a good reason to stay out.
> > > > 
> > > > > Then again, perhaps the more relevant concern would be why drivers use
> > > > > expedited grace periods in the first place.
> > > > 
> > > > Networking uses expedited grace periods when RTNL is held to reduce
> > > > contention on that lock.
> > > 
> > > Wait, what?  Why is anything using traditional (non-S) RCU while *any*
> > > lock is held?
> > 
> > In their defense, it is a sleeplock that is never taken except when
> > rearranging networking configuration.  Sometimes they need a grace period
> > under the lock.  So synchronize_net() checks to see if RTNL is held, and
> > does a synchronize_rcu_expedited() if so and a synchronize_rcu() if not.
> > 
> > But maybe I am misunderstanding your question?
> 
> No, you understood my question.  It seems wrong that they would need a
> grace period *under* the lock, rather than the usual case of making a
> change under the lock, dropping the lock, and *then* synchronizing.

On that I must defer to the networking folks.

							Thanx, Paul