Message-ID: <20181019012645.GC89903@joelaf.mtv.corp.google.com>
Date:   Thu, 18 Oct 2018 18:26:45 -0700
From:   Joel Fernandes <joel@...lfernandes.org>
To:     "Paul E. McKenney" <paulmck@...ux.ibm.com>
Cc:     Nikolay Borisov <nborisov@...e.com>, linux-kernel@...r.kernel.org,
        Jonathan Corbet <corbet@....net>,
        Josh Triplett <josh@...htriplett.org>,
        Lai Jiangshan <jiangshanlai@...il.com>,
        linux-doc@...r.kernel.org,
        Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
        Steven Rostedt <rostedt@...dmis.org>
Subject: Re: [PATCH RFC] doc: rcu: remove obsolete (non-)requirement about
 disabling preemption

On Thu, Oct 18, 2018 at 05:19:32PM -0700, Paul E. McKenney wrote:
> On Thu, Oct 18, 2018 at 05:03:50PM -0700, Joel Fernandes wrote:
> > On Thu, Oct 18, 2018 at 07:46:37AM -0700, Paul E. McKenney wrote:
> > [..]
> > > > > > > > > ------------------------------------------------------------------------
> > > > > > > > > 
> > > > > > > > > commit 07921e8720907f58f82b142f2027fc56d5abdbfd
> > > > > > > > > Author: Paul E. McKenney <paulmck@...ux.ibm.com>
> > > > > > > > > Date:   Tue Oct 16 04:12:58 2018 -0700
> > > > > > > > > 
> > > > > > > > >     rcu: Speed up expedited GPs when interrupting RCU reader
> > > > > > > > >     
> > > > > > > > >     In PREEMPT kernels, an expedited grace period might send an IPI to a
> > > > > > > > >     CPU that is executing an RCU read-side critical section.  In that case,
> > > > > > > > >     it would be nice if the rcu_read_unlock() directly interacted with the
> > > > > > > > >     RCU core code to immediately report the quiescent state.  And this does
> > > > > > > > >     happen in the case where the reader has been preempted.  But it would
> > > > > > > > >     also be a nice performance optimization if immediate reporting also
> > > > > > > > >     happened in the preemption-free case.
> > > > > > > > >     
> > > > > > > > >     This commit therefore adds an ->exp_hint field to the task_struct structure's
> > > > > > > > >     ->rcu_read_unlock_special field.  The IPI handler sets this hint when
> > > > > > > > >     it has interrupted an RCU read-side critical section, and this causes
> > > > > > > > >     the outermost rcu_read_unlock() call to invoke rcu_read_unlock_special(),
> > > > > > > > >     which, if preemption is enabled, reports the quiescent state immediately.
> > > > > > > > >     If preemption is disabled, then the report is required to be deferred
> > > > > > > > >     until preemption (or bottom halves or interrupts or whatever) is re-enabled.
> > > > > > > > >     
> > > > > > > > >     Because this is a hint, it does nothing for more complicated cases.  For
> > > > > > > > >     example, if the IPI interrupts an RCU reader, but interrupts are disabled
> > > > > > > > >     across the rcu_read_unlock(), but another rcu_read_lock() is executed
> > > > > > > > >     before interrupts are re-enabled, the hint will already have been cleared.
> > > > > > > > >     If you do crazy things like this, reporting will be deferred until some
> > > > > > > > >     later RCU_SOFTIRQ handler, context switch, cond_resched(), or similar.
> > > > > > > > >     
> > > > > > > > >     Reported-by: Joel Fernandes <joel@...lfernandes.org>
> > > > > > > > >     Signed-off-by: Paul E. McKenney <paulmck@...ux.ibm.com>
> > > > > > > > > 
> > > > > > > > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > > > > > > > index 004ca21f7e80..64ce751b5fe9 100644
> > > > > > > > > --- a/include/linux/sched.h
> > > > > > > > > +++ b/include/linux/sched.h
> > > > > > > > > @@ -571,8 +571,10 @@ union rcu_special {
> > > > > > > > >  	struct {
> > > > > > > > >  		u8			blocked;
> > > > > > > > >  		u8			need_qs;
> > > > > > > > > +		u8			exp_hint; /* Hint for performance. */
> > > > > > > > > +		u8			pad; /* No garbage from compiler! */
> > > > > > > > >  	} b; /* Bits. */
> > > > > > > > > -	u16 s; /* Set of bits. */
> > > > > > > > > +	u32 s; /* Set of bits. */
> > > > > > > > >  };
> > > > > > > > >  
> > > > > > > > >  enum perf_event_task_context {
> > > > > > > > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > > > > > > > > index e669ccf3751b..928fe5893a57 100644
> > > > > > > > > --- a/kernel/rcu/tree_exp.h
> > > > > > > > > +++ b/kernel/rcu/tree_exp.h
> > > > > > > > > @@ -692,8 +692,10 @@ static void sync_rcu_exp_handler(void *unused)
> > > > > > > > >  	 */
> > > > > > > > >  	if (t->rcu_read_lock_nesting > 0) {
> > > > > > > > >  		raw_spin_lock_irqsave_rcu_node(rnp, flags);
> > > > > > > > > -		if (rnp->expmask & rdp->grpmask)
> > > > > > > > > +		if (rnp->expmask & rdp->grpmask) {
> > > > > > > > >  			rdp->deferred_qs = true;
> > > > > > > > > +			WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, true);
> > > > > > > > > +		}
> > > > > > > > >  		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > > > > > > >  	}
> > > > > > > > >  
> > > > > > > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > > > > > > index 8b48bb7c224c..d6286eb6e77e 100644
> > > > > > > > > --- a/kernel/rcu/tree_plugin.h
> > > > > > > > > +++ b/kernel/rcu/tree_plugin.h
> > > > > > > > > @@ -643,8 +643,9 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > > > > > > > >  	local_irq_save(flags);
> > > > > > > > >  	irqs_were_disabled = irqs_disabled_flags(flags);
> > > > > > > > >  	if ((preempt_bh_were_disabled || irqs_were_disabled) &&
> > > > > > > > > -	    t->rcu_read_unlock_special.b.blocked) {
> > > > > > > > > +	    t->rcu_read_unlock_special.s) {
> > > > > > > > >  		/* Need to defer quiescent state until everything is enabled. */
> > > > > > > > > +		WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > > > > > > >  		raise_softirq_irqoff(RCU_SOFTIRQ);
> > > > > > > > 
> > > > > > > > Still going through this patch, but it seems to me like the fact that
> > > > > > > > rcu_read_unlock_special is called means someone has requested for a grace
> > > > > > > > period. Then in that case, does it not make sense to raise the softirq
> > > > > > > > for processing anyway?
> > > > > > > 
> > > > > > > Not necessarily.  Another reason that rcu_read_unlock_special() might
> > > > > > > be called is if the RCU read-side critical section had been preempted,
> > > > > > > in which case there might not even be a grace period in progress.
> > > > > > 
> > > > > > Yes true, it was at the back of my head ;) It needs to remove itself from the
> > > > > > blocked lists on the unlock. And ofcourse the preemption case is alsoo
> > > > > > clearly mentioned in this function's comments. (slaps self).
> > > > > 
> > > > > Sometimes rcutorture reminds me of interesting RCU corner cases...  ;-)
> > > > > 
> > > > > > > In addition, if interrupts, bottom halves, and preemption are all enabled,
> > > > > > > the code in rcu_preempt_deferred_qs_irqrestore() doesn't need to bother
> > > > > > > raising softirq, as it can instead just immediately report the quiescent
> > > > > > > state.
> > > > > > 
> > > > > > Makes sense. I will go through these code paths more today. Thank you for the
> > > > > > explanations!
> > > > > > 
> > > > > > I think something like need_exp_qs instead of 'exp_hint' may be more
> > > > > > descriptive?
> > > > > 
> > > > > Well, it is only a hint due to the fact that it is not preserved across
> > > > > complex sequences of overlapping RCU read-side critical sections of
> > > > > different types.  So if you have the following sequence:
> > > > > 
> > > > > 	rcu_read_lock();
> > > > > 	/* Someone does synchronize_rcu_expedited(), which sets ->exp_hint. */
> > > > > 	preempt_disable();
> > > > > 	rcu_read_unlock(); /* Clears ->exp_hint. */
> > > > > 	preempt_enable(); /* But ->exp_hint is already cleared. */
> > > > > 
> > > > > This is OK because there will be some later event that passes the quiescent
> > > > > state to the RCU core.  This will slow down the expedited grace period,
> > > > > but this case should be uncommon.  If it does turn out to be common, then
> > > > > some more complex scheme can be put in place.
> > > > > 
> > > > > Hmmm...  This patch does need some help, doesn't it?  How about the following
> > > > > to be folded into the original?
> > > > > 
> > > > > commit d8d996385055d4708121fa253e04b4272119f5e2
> > > > > Author: Paul E. McKenney <paulmck@...ux.ibm.com>
> > > > > Date:   Wed Oct 17 13:32:25 2018 -0700
> > > > > 
> > > > >     fixup! rcu: Speed up expedited GPs when interrupting RCU reader
> > > > >     
> > > > >     Signed-off-by: Paul E. McKenney <paulmck@...ux.ibm.com>
> > > > > 
> > > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > > index d6286eb6e77e..117aeb582fdc 100644
> > > > > --- a/kernel/rcu/tree_plugin.h
> > > > > +++ b/kernel/rcu/tree_plugin.h
> > > > > @@ -650,6 +650,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
> > > > >  		local_irq_restore(flags);
> > > > >  		return;
> > > > >  	}
> > > > > +	WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > > >  	rcu_preempt_deferred_qs_irqrestore(t, flags);
> > > > >  }
> > > > >  
> > > > 
> > > > Sure, I believe so. I was also thinking out load about if we can avoid
> > > > raising of the softirq for some cases in rcu_read_unlock_special:
> > > > 
> > > > For example, in rcu_read_unlock_special()
> > > > 
> > > > static void rcu_read_unlock_special(struct task_struct *t)
> > > > {
> > > > [...]
> > > > 	if ((preempt_bh_were_disabled || irqs_were_disabled) &&
> > > > 	    t->rcu_read_unlock_special.s) {
> > > > 		/* Need to defer quiescent state until everything is enabled. */
> > > > 		raise_softirq_irqoff(RCU_SOFTIRQ);
> > > > 		local_irq_restore(flags);
> > > > 		return;
> > > > 	}
> > > > 	rcu_preempt_deferred_qs_irqrestore(t, flags);
> > > > }
> > > > 
> > > > Instead of raising the softirq, for the case where irqs are enabled, but
> > > > preemption is disabled, can we not just do:
> > > > 
> > > > 		set_tsk_need_resched(current);
> > > > 		set_preempt_need_resched();
> > > > 
> > > > and return? Not sure the benefits of doing that are, but it seems nice to
> > > > avoid raising the softirq if possible, for benefit of real-time workloads.
> > > 
> > > This approach would work very well in the case when preemption or bottom
> > > halves were disabled, but would not handle the case where interrupts were
> > > enabled during the RCU read-side critical section, an expedited grace
> > > period started (thus setting ->exp_hint), interrupts where then disabled,
> > > and finally rcu_read_unlock() was invoked.  Re-enabling interrupts would
> > > not cause either softirq or the scheduler to do anything, so the end of
> > > the expedited grace period might be delayed for some time, for example,
> > > until the next scheduling-clock interrupt.
> > > 
> > > But please see below.
> > > 
> > > > Also it seems like there is a chance the softirq might run before the
> > > > preemption is reenabled anyway right?
> > > 
> > > Not unless the rcu_read_unlock() is invoked from within a softirq
> > > handler on the one hand or within an interrupt handler that interrupted
> > > a preempt-disable region of code.  Otherwise, because interrupts are
> > > disabled, the raise_softirq() will wake up ksoftirqd, which cannot run
> > > until both preemption and bottom halves are enabled.
> > > 
> > > > Also one last thing, in your patch - do we really need to test for
> > > > "t->rcu_read_unlock_special.s" in rcu_read_unlock_special()? AFAICT,
> > > > rcu_read_unlock_special would only be called if t->rcu_read_unlock_special.s
> > > > is set in the first place so we can drop the test for that.
> > > 
> > > Good point!
> > > 
> > > How about the following?
> > > 
> > > 							Thanx, Paul
> > > 
> > > ------------------------------------------------------------------------
> > > 
> > > static void rcu_read_unlock_special(struct task_struct *t)
> > > {
> > > 	unsigned long flags;
> > > 	bool preempt_bh_were_disabled =
> > > 			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
> > > 	bool irqs_were_disabled;
> > > 
> > > 	/* NMI handlers cannot block and cannot safely manipulate state. */
> > > 	if (in_nmi())
> > > 		return;
> > > 
> > > 	local_irq_save(flags);
> > > 	irqs_were_disabled = irqs_disabled_flags(flags);
> > > 	if (preempt_bh_were_disabled || irqs_were_disabled) {
> > > 		WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
> > > 		/* Need to defer quiescent state until everything is enabled. */
> > > 		if (irqs_were_disabled) {
> > > 			raise_softirq_irqoff(RCU_SOFTIRQ);
> > > 		} else {
> > > 			set_tsk_need_resched(current);
> > > 			set_preempt_need_resched();
> > > 		}
> > 
> > Looks good to me, thanks! Maybe some code comments would be nice as well.
> > 
> > Shouldn't we also set_tsk_need_resched for the irqs_were_disabled case, so
> > that say if we are in an IRQ disabled region (local_irq_disable), then
> > ksoftirqd would run as possible once IRQs are renabled?
> 
> Last I checked, local_irq_restore() didn't check for reschedules, instead
> relying on IPIs and scheduling-clock interrupts to do its dirty work.
> Has that changed?

Yes, local_irq_restore is light weight, and does not check for reschedules.

I was thinking of case where ksoftirqd is woken up, but does not run unless
we set the NEED_RESCHED flag. But that should get set anyway since probably
ksoftirqd is of high enough priority than the currently running task..

Roughly speaking the scenario could be something like:

rcu_read_lock();
                 <-- IPI comes in for the expedited GP, sets exp_hint
local_irq_disable();
// do a bunch of stuff
rcu_read_unlock();   <-- This calls the rcu_read_unlock_special which raises
                         the soft irq, and wakesup softirqd.
local_irq_enable();

// Now ksoftirqd is ready to run but we don't switch into the
// scheduler for sometime because tif_need_resched() returns false and
// any cond_resched calls do nothing. So we potentially spend lots of
// time before the next scheduling event.

You think this should not be an issue?

thanks,

 - Joel

Powered by blists - more mailing lists