Re: [PATCH rcu 3/3] rcu: Allow expedited RCU grace periods on incoming CPUs

[Paul E. McKenney]

On Sat, Feb 12, 2022 at 12:14:20AM +0530, Mukesh Ojha wrote:
> 
> On 2/10/2022 3:36 AM, Paul E. McKenney wrote:
> > On Wed, Feb 09, 2022 at 11:53:33PM +0530, Mukesh Ojha wrote:
> > > On 2/5/2022 4:25 AM, Paul E. McKenney wrote:
> > > > Although it is usually safe to invoke synchronize_rcu_expedited() from a
> > > > preemption-enabled CPU-hotplug notifier, if it is invoked from a notifier
> > > > between CPUHP_AP_RCUTREE_ONLINE and CPUHP_AP_ACTIVE, its attempts to
> > > > invoke a workqueue handler will hang due to RCU waiting on a CPU that
> > > > the scheduler is not paying attention to.  This commit therefore expands
> > > > use of the existing workqueue-independent synchronize_rcu_expedited()
> > > > from early boot to also include CPUs that are being hotplugged.
> > > > 
> > > > Link: https://lore.kernel.org/lkml/7359f994-8aaf-3cea-f5cf-c0d3929689d6@xxxxxxxxxxx/
> > > > Reported-by: Mukesh Ojha <quic_mojha@xxxxxxxxxxx>
> > > > Cc: Tejun Heo <tj@xxxxxxxxxx>
> > > > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > > > ---
> > > >    kernel/rcu/tree_exp.h | 14 ++++++++++----
> > > >    1 file changed, 10 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > > > index 60197ea24ceb9..1a45667402260 100644
> > > > --- a/kernel/rcu/tree_exp.h
> > > > +++ b/kernel/rcu/tree_exp.h
> > > > @@ -816,7 +816,7 @@ static int rcu_print_task_exp_stall(struct rcu_node *rnp)
> > > >     */
> > > >    void synchronize_rcu_expedited(void)
> > > >    {
> > > > -	bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
> > > > +	bool no_wq;
> > > >    	struct rcu_exp_work rew;
> > > >    	struct rcu_node *rnp;
> > > >    	unsigned long s;
> > > > @@ -841,9 +841,15 @@ void synchronize_rcu_expedited(void)
> > > >    	if (exp_funnel_lock(s))
> > > >    		return;  /* Someone else did our work for us. */
> > > > +	/* Don't use workqueue during boot or from an incoming CPU. */
> > > > +	preempt_disable();
> > > > +	no_wq = rcu_scheduler_active == RCU_SCHEDULER_INIT ||
> > > > +		!cpumask_test_cpu(smp_processor_id(), cpu_active_mask);
> > > > +	preempt_enable();
> > > > +
> > > >    	/* Ensure that load happens before action based on it. */
> > > > -	if (unlikely(boottime)) {
> > > > -		/* Direct call during scheduler init and early_initcalls(). */
> > > > +	if (unlikely(no_wq)) {
> > > > +		/* Direct call for scheduler init, early_initcall()s, and incoming CPUs. */
> > > >    		rcu_exp_sel_wait_wake(s);
> > > >    	} else {
> > > >    		/* Marshall arguments & schedule the expedited grace period. */
> > > > @@ -861,7 +867,7 @@ void synchronize_rcu_expedited(void)
> > > >    	/* Let the next expedited grace period start. */
> > > >    	mutex_unlock(&rcu_state.exp_mutex);
> > > > -	if (likely(!boottime))
> > > > +	if (likely(!no_wq))
> > > >    		destroy_work_on_stack(&rew.rew_work);
> > > >    }
> > > >    EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
> > > Can we reach a condition after this change where no_wq = true and during
> > > rcu_stall report where exp_task = 0 list and exp_mask contain only this cpu
> > > ?
> > Hello, Mukesh, and thank you for looking this over!
> > 
> > At first glance, I do not believe that this can happen because the
> > expedited grace-period machinery avoids waiting on the current CPU.
> > (See sync_rcu_exp_select_node_cpus(), both the raw_smp_processor_id()
> > early in the function and the get_cpu() later in the function.)
> > 
> > But please let me know if I am missing something here.
> > 
> > But suppose that we could in fact reach this condition.  What bad thing
> > would happen?  Other than a resched_cpu() having been invoked several
> > times on a not-yet-online CPU, of course.  ;-)
> 
> 
> I thought more about this, what if  synchronize_rcu_expedited thread got
> schedule out and run on some other cpu
> and we clear out cpu on which it ran next from exp_mask.
> 
> Queuing the work on same cpu ensures that it will always be right cpu to
> clear out.
> Do you think this can happen ?

Indeed it might.

But if it did, the scheduler would invoke RCU's hook, which is named
rcu_note_context_switch(), and do so on the pre-switch CPU.  There are
two implementations for this function, one for CONFIG_PREEMPT=y
and another for CONFIG_PREEMPT=n.  Both look to me like they invoke
rcu_report_exp_rdp() when needed, one directly and the other via the
CONFIG_PREEMPT=n variant of rcu_qs().

Am I missing something?

						Thanx, Paul