Message-ID: <20190804192454.GB83025@google.com>
Date:   Sun, 4 Aug 2019 15:24:54 -0400
From:   Joel Fernandes <joel@...lfernandes.org>
To:     "Paul E. McKenney" <paulmck@...ux.ibm.com>
Cc:     rcu@...r.kernel.org, linux-kernel@...r.kernel.org,
        mingo@...nel.org, jiangshanlai@...il.com, dipankar@...ibm.com,
        akpm@...ux-foundation.org, mathieu.desnoyers@...icios.com,
        josh@...htriplett.org, tglx@...utronix.de, peterz@...radead.org,
        rostedt@...dmis.org, dhowells@...hat.com, edumazet@...gle.com,
        fweisbec@...il.com, oleg@...hat.com
Subject: Re: [PATCH tip/core/rcu 03/11] rcu/nocb: Provide separate no-CBs
 grace-period kthreads

On Sat, Aug 03, 2019 at 12:46:11PM -0700, Paul E. McKenney wrote:
> On Sat, Aug 03, 2019 at 01:41:27PM -0400, Joel Fernandes wrote:
> > On Thu, Aug 01, 2019 at 03:50:20PM -0700, Paul E. McKenney wrote:
> > > Currently, there is one no-CBs rcuo kthread per CPU, and these kthreads
> > > are divided into groups.  The first rcuo kthread to come online in a
> > > given group is that group's leader, and the leader both waits for grace
> > > periods and invokes its CPU's callbacks.  The non-leader rcuo kthreads
> > > only invoke callbacks.
> > > 
> > > This works well in the real-time/embedded environments for which it was
> > > intended because such environments tend not to generate all that many
> > > callbacks.  However, given huge floods of callbacks, it is possible for
> > > the leader kthread to be stuck invoking callbacks while its followers
> > > wait helplessly while their callbacks pile up.  This is a good recipe
> > > for an OOM, and rcutorture's new callback-flood capability does generate
> > > such OOMs.
> > > 
> > > One strategy would be to wait until such OOMs start happening in
> > > production, but similar OOMs have in fact happened starting in 2018.
> > > It would therefore be wise to take a more proactive approach.
> > 
> > I haven't looked much into nocbs/nohz_full stuff (yet). In particular, I did
> > not even know that the rcuo threads do grace period life-cycle management and
> > waiting, I thought only the RCU GP threads did :-/. however, it seems this is
> > a completely separate grace-period management state machine outside of the
> > RCU GP thread right?
> 
> No, the rcuo kthreads interact with the main RCU GP kthread, initiating
> new grace periods when needed and being awakened as needed by the RCU
> GP kthread.

Ok, I see the interactions in rcu_nocb_wait_gp(). This what I was thinking
too is that there has to be these interactions with the main RCU GP kthread,
for anything to work :) Thanks for the explanation!

> > I was wondering for this patch, could we also just have the rcuo
> > leader handle both callback execution and waking other non-leader threads at
> > the same time? So like, execute few callbacks, then do the wake up of the
> > non-leaders to execute their callbacks, the get back to executing their own
> > callbacks, etc. That way we don't need a separate rcuog thread to wait for
> > grace period, would that not work?
> 
> I did look into that, but it was more complex and also didn't foster
> sharing of rcu_do_batch(), which used to only be for non-offloaded
> callbacks but now does both.  Besides which, invoking callbacks would
> degrade the rcuog kthread's response to new callbacks and the like.

Makes sense.

> > If you don't mind could you share with me a kvm.sh command (which has config,
> > boot parameters etc) that can produce the OOM without this patch? I'd
> > like to take a closer look at it.
> 
> Here you go:
> 
> tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 8 --duration 120 --configs TREE04
> 
> If you add "--memory 1G" in mainline, the OOMs go away.  Or at least
> decrease substantially in probability.

I could reproduce it, it look around 5-10 minutes for the OOM for 512MB
memory. Thanks.

> > Is there also a short answer for my the RCU GP thread cannot do the job of
> > these new rcuog threads?
> 
> First, the code is more complicated when you do it that way (and yes,
> I did actually write it out in pen on paper).  Second, if the CPU
> corresponding to the combined grace-period/callback kthread is doing the
> call_rcu() flooding, you are between a rock and a hard place.  On the
> one hand, you want that kthread to do nothing but invoke callbacks so
> as to have half a chance of keeping up, and on the other hand you need
> it to check state frequently so as to react in a timely fashion to a
> CPU corresponding to one of its callback kthreads starting a second
> callback flood.

> Introducing rcug grace-period-only kthreads means that you get the best of
> both worlds.  Plus last year's flavor consolidation decreased the number
> of rcuo kthreads from either 2N or 3N to N, so increasing it to only
> (N + sqrt(N)) should be just fine.  Though I would expect that there
> will be at least some screaming and shouting.  ;-)

Ok got it. Yes, fewer newer threads now even with nocb improvements :)

thanks,

 - Joel


> 
> 							Thanx, Paul
> 
> > thanks a lot,
> > 
> >  - Joel
> > 
> > 
> > > This commit therefore features per-CPU rcuo kthreads that do nothing
> > > but invoke callbacks.  Instead of having one of these kthreads act as
> > > leader, each group has a separate rcog kthread that handles grace periods
> > > for its group.  Because these rcuog kthreads do not invoke callbacks,
> > > callback floods on one CPU no longer block callbacks from reaching the
> > > rcuc callback-invocation kthreads on other CPUs.
> > > 
> > > This change does introduce additional kthreads, however:
> > > 
> > > 1.	The number of additional kthreads is about the square root of
> > > 	the number of CPUs, so that a 4096-CPU system would have only
> > > 	about 64 additional kthreads.  Note that recent changes
> > > 	decreased the number of rcuo kthreads by a factor of two
> > > 	(CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so
> > > 	this still represents a significant improvement on most systems.
> > > 
> > > 2.	The leading "rcuo" of the rcuog kthreads should allow existing
> > > 	scripting to affinity these additional kthreads as needed, the
> > > 	same as for the rcuop and rcuos kthreads.  (There are no longer
> > > 	any rcuob kthreads.)
> > > 
> > > 3.	A state-machine approach was considered and rejected.  Although
> > > 	this would allow the rcuo kthreads to continue their dual
> > > 	leader/follower roles, it complicates callback invocation
> > > 	and makes it more difficult to consolidate rcuo callback
> > > 	invocation with existing softirq callback invocation.
> > > 
> > > The introduction of rcuog kthreads should thus be acceptable.
> > > 
> > > Signed-off-by: Paul E. McKenney <paulmck@...ux.ibm.com>
> > > ---
> > >  kernel/rcu/tree.h        |   6 +-
> > >  kernel/rcu/tree_plugin.h | 115 +++++++++++++++++++--------------------
> > >  2 files changed, 61 insertions(+), 60 deletions(-)
> > > 
> > > diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
> > > index 32b3348d3a4d..dc3c53cb9608 100644
> > > --- a/kernel/rcu/tree.h
> > > +++ b/kernel/rcu/tree.h
> > > @@ -200,8 +200,8 @@ struct rcu_data {
> > >  	atomic_long_t nocb_q_count_lazy; /*  invocation (all stages). */
> > >  	struct rcu_head *nocb_cb_head;	/* CBs ready to invoke. */
> > >  	struct rcu_head **nocb_cb_tail;
> > > -	struct swait_queue_head nocb_wq; /* For nocb kthreads to sleep on. */
> > > -	struct task_struct *nocb_cb_kthread;
> > > +	struct swait_queue_head nocb_cb_wq; /* For nocb kthreads to sleep on. */
> > > +	struct task_struct *nocb_gp_kthread;
> > >  	raw_spinlock_t nocb_lock;	/* Guard following pair of fields. */
> > >  	int nocb_defer_wakeup;		/* Defer wakeup of nocb_kthread. */
> > >  	struct timer_list nocb_timer;	/* Enforce finite deferral. */
> > > @@ -211,6 +211,8 @@ struct rcu_data {
> > >  					/* CBs waiting for GP. */
> > >  	struct rcu_head **nocb_gp_tail;
> > >  	bool nocb_gp_sleep;		/* Is the nocb GP thread asleep? */
> > > +	struct swait_queue_head nocb_gp_wq; /* For nocb kthreads to sleep on. */
> > > +	struct task_struct *nocb_cb_kthread;
> > >  	struct rcu_data *nocb_next_cb_rdp;
> > >  					/* Next rcu_data in wakeup chain. */
> > >  
> > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > index 5a72700c3a32..c3b6493313ab 100644
> > > --- a/kernel/rcu/tree_plugin.h
> > > +++ b/kernel/rcu/tree_plugin.h
> > > @@ -1531,7 +1531,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,
> > >  	struct rcu_data *rdp_leader = rdp->nocb_gp_rdp;
> > >  
> > >  	lockdep_assert_held(&rdp->nocb_lock);
> > > -	if (!READ_ONCE(rdp_leader->nocb_cb_kthread)) {
> > > +	if (!READ_ONCE(rdp_leader->nocb_gp_kthread)) {
> > >  		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
> > >  		return;
> > >  	}
> > > @@ -1541,7 +1541,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,
> > >  		del_timer(&rdp->nocb_timer);
> > >  		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
> > >  		smp_mb(); /* ->nocb_gp_sleep before swake_up_one(). */
> > > -		swake_up_one(&rdp_leader->nocb_wq);
> > > +		swake_up_one(&rdp_leader->nocb_gp_wq);
> > >  	} else {
> > >  		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
> > >  	}
> > > @@ -1646,7 +1646,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
> > >  	smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
> > >  
> > >  	/* If we are not being polled and there is a kthread, awaken it ... */
> > > -	t = READ_ONCE(rdp->nocb_cb_kthread);
> > > +	t = READ_ONCE(rdp->nocb_gp_kthread);
> > >  	if (rcu_nocb_poll || !t) {
> > >  		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
> > >  				    TPS("WakeNotPoll"));
> > > @@ -1786,7 +1786,7 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)
> > >   * No-CBs GP kthreads come here to wait for additional callbacks to show up.
> > >   * This function does not return until callbacks appear.
> > >   */
> > > -static void nocb_leader_wait(struct rcu_data *my_rdp)
> > > +static void nocb_gp_wait(struct rcu_data *my_rdp)
> > >  {
> > >  	bool firsttime = true;
> > >  	unsigned long flags;
> > > @@ -1794,12 +1794,10 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)
> > >  	struct rcu_data *rdp;
> > >  	struct rcu_head **tail;
> > >  
> > > -wait_again:
> > > -
> > >  	/* Wait for callbacks to appear. */
> > >  	if (!rcu_nocb_poll) {
> > >  		trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu, TPS("Sleep"));
> > > -		swait_event_interruptible_exclusive(my_rdp->nocb_wq,
> > > +		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
> > >  				!READ_ONCE(my_rdp->nocb_gp_sleep));
> > >  		raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
> > >  		my_rdp->nocb_gp_sleep = true;
> > > @@ -1838,7 +1836,7 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)
> > >  			trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu,
> > >  					    TPS("WokeEmpty"));
> > >  		}
> > > -		goto wait_again;
> > > +		return;
> > >  	}
> > >  
> > >  	/* Wait for one grace period. */
> > > @@ -1862,34 +1860,47 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)
> > >  		rdp->nocb_cb_tail = rdp->nocb_gp_tail;
> > >  		*tail = rdp->nocb_gp_head;
> > >  		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
> > > -		if (rdp != my_rdp && tail == &rdp->nocb_cb_head) {
> > > +		if (tail == &rdp->nocb_cb_head) {
> > >  			/* List was empty, so wake up the kthread.  */
> > > -			swake_up_one(&rdp->nocb_wq);
> > > +			swake_up_one(&rdp->nocb_cb_wq);
> > >  		}
> > >  	}
> > > +}
> > >  
> > > -	/* If we (the GP kthreads) don't have CBs, go wait some more. */
> > > -	if (!my_rdp->nocb_cb_head)
> > > -		goto wait_again;
> > > +/*
> > > + * No-CBs grace-period-wait kthread.  There is one of these per group
> > > + * of CPUs, but only once at least one CPU in that group has come online
> > > + * at least once since boot.  This kthread checks for newly posted
> > > + * callbacks from any of the CPUs it is responsible for, waits for a
> > > + * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
> > > + * that then have callback-invocation work to do.
> > > + */
> > > +static int rcu_nocb_gp_kthread(void *arg)
> > > +{
> > > +	struct rcu_data *rdp = arg;
> > > +
> > > +	for (;;)
> > > +		nocb_gp_wait(rdp);
> > > +	return 0;
> > >  }
> > >  
> > >  /*
> > >   * No-CBs CB kthreads come here to wait for additional callbacks to show up.
> > > - * This function does not return until callbacks appear.
> > > + * This function returns true ("keep waiting") until callbacks appear and
> > > + * then false ("stop waiting") when callbacks finally do appear.
> > >   */
> > > -static void nocb_follower_wait(struct rcu_data *rdp)
> > > +static bool nocb_follower_wait(struct rcu_data *rdp)
> > >  {
> > > -	for (;;) {
> > > -		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FollowerSleep"));
> > > -		swait_event_interruptible_exclusive(rdp->nocb_wq,
> > > -					 READ_ONCE(rdp->nocb_cb_head));
> > > -		if (smp_load_acquire(&rdp->nocb_cb_head)) {
> > > -			/* ^^^ Ensure CB invocation follows _head test. */
> > > -			return;
> > > -		}
> > > -		WARN_ON(signal_pending(current));
> > > -		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
> > > +	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FollowerSleep"));
> > > +	swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
> > > +				 READ_ONCE(rdp->nocb_cb_head));
> > > +	if (smp_load_acquire(&rdp->nocb_cb_head)) { /* VVV */
> > > +		/* ^^^ Ensure CB invocation follows _head test. */
> > > +		return false;
> > >  	}
> > > +	WARN_ON(signal_pending(current));
> > > +	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
> > > +	return true;
> > >  }
> > >  
> > >  /*
> > > @@ -1899,7 +1910,7 @@ static void nocb_follower_wait(struct rcu_data *rdp)
> > >   * have to do quite so many wakeups (as in they only need to wake the
> > >   * no-CBs GP kthreads, not the CB kthreads).
> > >   */
> > > -static int rcu_nocb_kthread(void *arg)
> > > +static int rcu_nocb_cb_kthread(void *arg)
> > >  {
> > >  	int c, cl;
> > >  	unsigned long flags;
> > > @@ -1911,10 +1922,8 @@ static int rcu_nocb_kthread(void *arg)
> > >  	/* Each pass through this loop invokes one batch of callbacks */
> > >  	for (;;) {
> > >  		/* Wait for callbacks. */
> > > -		if (rdp->nocb_gp_rdp == rdp)
> > > -			nocb_leader_wait(rdp);
> > > -		else
> > > -			nocb_follower_wait(rdp);
> > > +		while (nocb_follower_wait(rdp))
> > > +			continue;
> > >  
> > >  		/* Pull the ready-to-invoke callbacks onto local list. */
> > >  		raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
> > > @@ -2048,7 +2057,8 @@ void __init rcu_init_nohz(void)
> > >  static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
> > >  {
> > >  	rdp->nocb_tail = &rdp->nocb_head;
> > > -	init_swait_queue_head(&rdp->nocb_wq);
> > > +	init_swait_queue_head(&rdp->nocb_cb_wq);
> > > +	init_swait_queue_head(&rdp->nocb_gp_wq);
> > >  	rdp->nocb_cb_tail = &rdp->nocb_cb_head;
> > >  	raw_spin_lock_init(&rdp->nocb_lock);
> > >  	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
> > > @@ -2056,50 +2066,39 @@ static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
> > >  
> > >  /*
> > >   * If the specified CPU is a no-CBs CPU that does not already have its
> > > - * rcuo kthread, spawn it.  If the CPUs are brought online out of order,
> > > - * this can require re-organizing the GP-CB relationships.
> > > + * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
> > > + * for this CPU's group has not yet been created, spawn it as well.
> > >   */
> > >  static void rcu_spawn_one_nocb_kthread(int cpu)
> > >  {
> > > -	struct rcu_data *rdp;
> > > -	struct rcu_data *rdp_last;
> > > -	struct rcu_data *rdp_old_leader;
> > > -	struct rcu_data *rdp_spawn = per_cpu_ptr(&rcu_data, cpu);
> > > +	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
> > > +	struct rcu_data *rdp_gp;
> > >  	struct task_struct *t;
> > >  
> > >  	/*
> > >  	 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
> > >  	 * then nothing to do.
> > >  	 */
> > > -	if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_cb_kthread)
> > > +	if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
> > >  		return;
> > >  
> > >  	/* If we didn't spawn the GP kthread first, reorganize! */
> > > -	rdp_old_leader = rdp_spawn->nocb_gp_rdp;
> > > -	if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_cb_kthread) {
> > > -		rdp_last = NULL;
> > > -		rdp = rdp_old_leader;
> > > -		do {
> > > -			rdp->nocb_gp_rdp = rdp_spawn;
> > > -			if (rdp_last && rdp != rdp_spawn)
> > > -				rdp_last->nocb_next_cb_rdp = rdp;
> > > -			if (rdp == rdp_spawn) {
> > > -				rdp = rdp->nocb_next_cb_rdp;
> > > -			} else {
> > > -				rdp_last = rdp;
> > > -				rdp = rdp->nocb_next_cb_rdp;
> > > -				rdp_last->nocb_next_cb_rdp = NULL;
> > > -			}
> > > -		} while (rdp);
> > > -		rdp_spawn->nocb_next_cb_rdp = rdp_old_leader;
> > > +	rdp_gp = rdp->nocb_gp_rdp;
> > > +	if (!rdp_gp->nocb_gp_kthread) {
> > > +		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
> > > +				"rcuog/%d", rdp_gp->cpu);
> > > +		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
> > > +			return;
> > > +		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
> > >  	}
> > >  
> > >  	/* Spawn the kthread for this CPU. */
> > > -	t = kthread_run(rcu_nocb_kthread, rdp_spawn,
> > > +	t = kthread_run(rcu_nocb_cb_kthread, rdp,
> > >  			"rcuo%c/%d", rcu_state.abbr, cpu);
> > > -	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo kthread, OOM is now expected behavior\n", __func__))
> > > +	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
> > >  		return;
> > > -	WRITE_ONCE(rdp_spawn->nocb_cb_kthread, t);
> > > +	WRITE_ONCE(rdp->nocb_cb_kthread, t);
> > > +	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
> > >  }
> > >  
> > >  /*
> > > -- 
> > > 2.17.1
> > > 
> > 

Powered by blists - more mailing lists