Re: [PATCH RFC] rcu: Make rcu_barrier() less disruptive

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> On Thu, Mar 15, 2012 at 11:21:59AM -0700, Paul E. McKenney wrote:
> > On Thu, Mar 15, 2012 at 01:45:27PM -0400, Mathieu Desnoyers wrote:
> > > * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > > > The rcu_barrier() primitive interrupts each and every CPU, registering
> > > > a callback on every CPU.  Once all of these callbacks have been invoked,
> > > > rcu_barrier() knows that every callback that was registered before
> > > > the call to rcu_barrier() has also been invoked.
> > > > 
> > > > However, there is no point in registering a callback on a CPU that
> > > > currently has no callbacks, most especially if that CPU is in a
> > > > deep idle state.  This commit therefore makes rcu_barrier() avoid
> > > > interrupting CPUs that have no callbacks.  Doing this requires reworking
> > > > the handling of orphaned callbacks, otherwise callbacks could slip through
> > > > rcu_barrier()'s net by being orphaned from a CPU that rcu_barrier() had
> > > > not yet interrupted to a CPU that rcu_barrier() had already interrupted.
> > > > This reworking was needed anyway to take a first step towards weaning
> > > > RCU from the CPU_DYING notifier's use of stop_cpu().
> > > 
> > > Quoting Documentation/RCU/rcubarrier.txt:
> > > 
> > > "We instead need the rcu_barrier() primitive. This primitive is similar
> > > to synchronize_rcu(), but instead of waiting solely for a grace
> > > period to elapse, it also waits for all outstanding RCU callbacks to
> > > complete. Pseudo-code using rcu_barrier() is as follows:"
> > > 
> > > The patch you propose seems like a good approach to make rcu_barrier
> > > less disruptive, but everyone need to be aware that rcu_barrier() would
> > > quit having the side-effect of doing the equivalent of
> > > "synchronize_rcu()" from now on: within this new approach, in the case
> > > where there are no pending callbacks, rcu_barrier() could, AFAIU, return
> > > without waiting for the current grace period to complete.
> > > 
> > > Any use of rcu_barrier() that would assume that a synchronize_rcu() is
> > > implicit with the rcu_barrier() execution would be a bug anyway, but
> > > those might only show up after this patch is applied. I would therefore
> > > recommend to audit all rcu_barrier() users to ensure none is expecting
> > > rcu_barrier to act as a synchronize_rcu before pushing this change.
> > 
> > Good catch!
> > 
> > I am going to chicken out and explicitly wait for a grace period if there
> > were no callbacks.  Having rcu_barrier() very rarely be a quick no-op does
> > sound like a standing invitation for subtle non-reproducible bugs.  ;-)
> 
> I take it back...
> 
> After adopting callbacks (rcu_adopt_orphan_cbs()), _rcu_barrier()
> unconditionally posts a callback on the current CPU and waits for it.
> So _rcu_barrier() actually does always wait for a grace period.

Ah ok, that should handle it then.

> 
> Yes, I could be more dainty and make rcu_adopt_orphan_cbs() return an
> indication of whether there were any callbacks, and then post the callback
> only if either there were some callbacks adopted or if there were no calls
> to smp_call_function_single().  But that adds complexity for almost no
> benefit -- and no one can accuse _rcu_barrier() of being a fastpath!  ;-)
> 
> Or am I missing something here?

Nope, I think it all makes sense.

Thanks,

Mathieu

> 
> 							Thanx, Paul
> 

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com