Re: [PATCH tip/core/rcu 1/9] rcu: Provide GP ordering in face of migrations and delays

[Peter Zijlstra]

On Thu, Oct 05, 2017 at 07:55:13AM -0700, Paul E. McKenney wrote:
> On Thu, Oct 05, 2017 at 11:41:14AM +0200, Peter Zijlstra wrote:
> > On Wed, Oct 04, 2017 at 02:29:27PM -0700, Paul E. McKenney wrote:
> > > Consider the following admittedly improbable sequence of events:
> > > 
> > > o	RCU is initially idle.
> > > 
> > > o	Task A on CPU 0 executes rcu_read_lock().
> > > 
> > > o	Task B on CPU 1 executes synchronize_rcu(), which must
> > > 	wait on Task A:
> > > 
> > > 	o	Task B registers the callback, which starts a new
> > > 		grace period, awakening the grace-period kthread
> > > 		on CPU 3, which immediately starts a new grace period.
> > > 
> > > 	o	Task B migrates to CPU 2, which provides a quiescent
> > > 		state for both CPUs 1 and 2.
> > > 
> > > 	o	Both CPUs 1 and 2 take scheduling-clock interrupts,
> > > 		and both invoke RCU_SOFTIRQ, both thus learning of the
> > > 		new grace period.
> > > 
> > > 	o	Task B is delayed, perhaps by vCPU preemption on CPU 2.
> > > 
> > > o	CPUs 2 and 3 pass through quiescent states, which are reported
> > > 	to core RCU.
> > > 
> > > o	Task B is resumed just long enough to be migrated to CPU 3,
> > > 	and then is once again delayed.
> > > 
> > > o	Task A executes rcu_read_unlock(), exiting its RCU read-side
> > > 	critical section.
> > > 
> > > o	CPU 0 passes through a quiescent sate, which is reported to
> > > 	core RCU.  Only CPU 1 continues to block the grace period.
> > > 
> > > o	CPU 1 passes through a quiescent state, which is reported to
> > > 	core RCU.  This ends the grace period, and CPU 1 therefore
> > > 	invokes its callbacks, one of which awakens Task B via
> > > 	complete().
> > > 
> > > o	Task B resumes (still on CPU 3) and starts executing
> > > 	wait_for_completion(), which sees that the completion has
> > > 	already completed, and thus does not block.  It returns from
> > > 	the synchronize_rcu() without any ordering against the
> > > 	end of Task A's RCU read-side critical section.
> > > 
> > > 	It can therefore mess up Task A's RCU read-side critical section,
> > > 	in theory, anyway.
> > 
> > I'm not sure I follow, at the very least the wait_for_completion() does
> > an ACQUIRE such that it observes the state prior to the RELEASE as done
> > by complete(), no?
> 
> Your point being that both wait_for_completion() and complete() acquire
> and release the same lock?  (Yes, I suspect that I was confusing this
> with wait_event() and wake_up(), just so you know.)

Well, fundamentally complete()/wait_for_completion() is a message-pass
and they include a RELEASE/ACQUIRE pair for causal reasons.

Per the implementation they use a spinlock, but any implementation needs
to provide at least that RELEASE/ACQUIRE pair.

> > And is not CPU0's QS reporting ordered against that complete()?
> 
> Mumble mumble mumble powerpc mumble mumble mumble...
> 
> OK, I will make this new memory barrier only execute for powerpc.
> 
> Or am I missing something else here?

So I'm not entirely clear on the required semantics here; why do we need
a full mb? I'm thinking CPU0's QS propagating through the tree and
arriving at the root node is a multi-copy-atomic / transitive thing and
all CPUs will agree the system QS has ended, right?

Whichever CPU establishes the system QS does complete() and the
wait_for_completion() then has the weak-transitive causal relation to
that, ensuring that -- in the above example -- CPU3 must be _after_
CPU0's rcu_read_unlock().