Re: linux-next-20110923: warning kernel/rcutree.c:1833

[Paul E. McKenney]

On Fri, Sep 30, 2011 at 12:24:38PM -0700, Paul E. McKenney wrote:
> On Fri, Sep 30, 2011 at 08:29:46AM -0700, Paul E. McKenney wrote:
> > On Fri, Sep 30, 2011 at 03:11:09PM +0200, Frederic Weisbecker wrote:
> > > On Thu, Sep 29, 2011 at 10:12:05AM -0700, Paul E. McKenney wrote:
> > > > On Thu, Sep 29, 2011 at 02:30:44PM +0200, Frederic Weisbecker wrote:
> > > > > I was thinking about the fact that idle is a caller of rcu_enter_nohz().
> > > > > And there may be more callers of it in the future. So I thought it may
> > > > > be better to keep rcu_enter_nohz() idle-agnostic.
> > > > > 
> > > > > But it's fine, there are other ways to call rcu_idle_enter()/rcu_idle_exit()
> > > > > from the right places other than from rcu_enter/exit_nohz().
> > > > > We have tick_check_idle() on irq entry and tick_nohz_irq_exit(), both are called
> > > > > on the first interrupt level in idle.
> > > > > 
> > > > > So I can change that easily for the nohz cpusets.
> > > > 
> > > > Heh!  From what I can see, we were both wrong!
> > > > 
> > > > My thought at this point is to make it so that rcu_enter_nohz() and
> > > > rcu_exit_nohz() are renamed to rcu_enter_idle() and rcu_exit_idle()
> > > > respectively.  I drop the per-CPU variable and the added functions
> > > > from one of my patches.  These functions, along with rcu_irq_enter(),
> > > > rcu_irq_exit(), rcu_nmi_enter(), and rcu_nmi_exit(), are moved out from
> > > > under CONFIG_NO_HZ.  This allows these functions to track idle state
> > > > regardless of the setting of CONFIG_NO_HZ.  It also separates the state
> > > > of the scheduling-clock tick from RCU's view of CPU idleness, which
> > > > simplifies things.
> > > > 
> > > > I will put something together along these lines.
> > > 
> > > Should I wait for your updated patch before rebasing?
> > 
> > Gah!!!  I knew I was forgetting something!  I will get that out.
> > 
> > > > > > > > The problem I have with this is that it is rcu_enter_nohz() that tracks
> > > > > > > > the irq nesting required to correctly decide whether or not we are going
> > > > > > > > to really go to idle state.  Furthermore, there are cases where we
> > > > > > > > do enter idle but do not enter nohz, and that has to be handled correctly
> > > > > > > > as well.
> > > > > > > > 
> > > > > > > > Now, it is quite possible that I am suffering a senior moment and just
> > > > > > > > failing to see how to structure this in the design where rcu_idle_enter()
> > > > > > > > invokes rcu_enter_nohz(), but regardless, I am failing to see how to
> > > > > > > > structure this so that it works correctly.
> > > > > > > > 
> > > > > > > > Please feel free to enlighten me!
> > > > > > > 
> > > > > > > Ah I realize that you want to call rcu_idle_exit() when we enter
> > > > > > > the first level interrupt and rcu_idle_enter() when we exit it
> > > > > > > to return to idle loop.
> > > > > > > 
> > > > > > > But we use that check:
> > > > > > > 
> > > > > > > 	if (user ||
> > > > > > > 	    (rcu_is_cpu_idle() &&
> > > > > > >  	     !in_softirq() &&
> > > > > > >  	     hardirq_count() <= (1 << HARDIRQ_SHIFT)))
> > > > > > >  		rcu_sched_qs(cpu);
> > > > > > > 
> > > > > > > So we ensure that by the time we call rcu_check_callbacks(), we are not nesting
> > > > > > > in another interrupt.
> > > > > > 
> > > > > > But I would like to enable checks for entering/exiting idle while
> > > > > > within an RCU read-side critical section. The idea is to move
> > > > > > the checks from their currently somewhat problematic location in
> > > > > > rcu_needs_cpu_quick_check() to somewhere more sensible.  My current
> > > > > > thought is to move them rcu_enter_nohz() and rcu_exit_nohz() near the
> > > > > > calls to rcu_idle_enter() and rcu_idle_exit(), respectively.
> > > > > 
> > > > > So, checking if we are calling rcu_idle_enter() while in an RCU
> > > > > read side critical section?
> > > > > 
> > > > > But we already have checks that RCU read side API are not called in
> > > > > extended quiescent state.
> > > > 
> > > > Both checks are good.  The existing checks catch this kind of error:
> > > > 
> > > > 1.	CPU 0 goes idle, entering an RCU extended quiescent state.
> > > > 2.	CPU 0 illegally enters an RCU read-side critical section.
> > > > 
> > > > The new check catches this kind of error:
> > > > 
> > > > 1.	CPU 0 enters an RCU read-side critical section.
> > > > 2.	CPU 0 goes idle, entering an RCU extended quiescent state,
> > > > 	but illegally so because it is still in an RCU read-side
> > > > 	critical section.
> > > 
> > > Right.
> > > 
> > > > 
> > > > > > This would mean that they operated only in NO_HZ kernels with lockdep
> > > > > > enabled, but I am good with that because to do otherwise would require
> > > > > > adding nesting-level counters to the non-NO_HZ case, which I would like
> > > > > > to avoid, expecially for TINY_RCU.
> > > > 
> > > > And my reworking of RCU's NO_HZ code to instead be idle code removes
> > > > the NO_HZ-only restriction.  Getting rid of the additional per-CPU
> > > > variable reduces the TINY_RCU overhead to acceptable levels.
> > > > 
> > > > > There can be a secondary check in rcu_read_lock_held() and friends to
> > > > > ensures that rcu_is_idle_cpu(). In the non-NO_HZ case it's useful to
> > > > > find similar issues.
> > > > > 
> > > > > In fact we could remove the check for rcu_extended_qs() in read side
> > > > > APIs and check instead rcu_is_idle_cpu(). That would work in any
> > > > > config and not only NO_HZ.
> > > > > 
> > > > > But I hope we can actually keep the check for RCU extended quiescent
> > > > > state so that when rcu_enter_nohz() is called from other places than
> > > > > idle, we are ready for it.
> > > > > 
> > > > > I believe it's fine to have both checks in PROVE_RCU.
> > > > 
> > > > Agreed, I have not yet revisited rcu_extended_qs(), but some change
> > > > might be useful.
> > > 
> > > Yep.
> > > 
> > > > > > OK, my current plans are to start forward-porting to -rc8, and I would
> > > > > > like to have this pair of delta patches or something like them pulled
> > > > > > into your stack.
> > > > > 
> > > > > Sure I can take your patches (I'm going to merge the delta into the first).
> > > > > But if you want a rebase against -rc8, it's going to be easier if you
> > > > > do that rebase on the branch you want me to work on. Then I work on top
> > > > > of it.
> > > > > 
> > > > > For example we can take your rcu/dynticks, rewind to
> > > > > "rcu: Make synchronize_sched_expedited() better at work sharing"
> > > > > 771c326f20029a9f30b9a58237c9a5d5ddc1763d, rebase on top of -rc8
> > > > > and I rebase my patches (yours included) on top of it and I repost.
> > > > > 
> > > > > Right?
> > > > 
> > > > Yep!  Your earlier three patches look to need some extended-quiescent-state
> > > > rework as well:
> > > > 
> > > > b5566f3d: Detect illegal rcu dereference in extended quiescent state
> > > > ee05e5a4: Inform the user about dynticks-idle mode on PROVE_RCU warning
> > > > fa5d22cf: Warn when rcu_read_lock() is used in extended quiescent state
> > > > 
> > > > So I will leave these out and let you rebase them.
> > > 
> > > Fine. Just need to know if they need an update against a patch from you
> > > that is to come or something.
> > 
> > I am on it, apologies for the delay!
> 
> And here is a first cut, probably totally broken, but a start.
> 
> With this change, I am wondering about tick_nohz_stop_sched_tick()'s
> invocation of rcu_idle_enter() -- this now needs to be called regardless
> of whether or not tick_nohz_stop_sched_tick() actually stops the tick.
> Except that if tick_nohz_stop_sched_tick() is invoked with inidle==0,
> it looks like we should -not- call rcu_idle_enter().
> 
> I eventually just left the rcu_idle_enter() calls in their current
> places due to paranoia about messing up and ending up with unbalanced
> rcu_idle_enter() and rcu_idle_exit() calls.  Any thoughts on how to
> make this work better?

Well, rcutorture didn't like this one much.  Turns out that I messed
up the count balances on the NO_HZ=n case, and perhaps more besides.
I am now trying the following patch on top of my previous one.

							Thanx, Paul

------------------------------------------------------------------------

diff --git a/include/linux/tick.h b/include/linux/tick.h
index 35d2ffc..ca40838 100644
--- a/include/linux/tick.h
+++ b/include/linux/tick.h
@@ -129,7 +129,8 @@ extern u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time);
 # else
 static inline void tick_nohz_stop_sched_tick(int inidle)
 {
-	rcu_idle_enter();
+	if (inidle)
+		rcu_idle_enter();
 }
 static inline void tick_nohz_restart_sched_tick(void)
 {
diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index d61b908..4692907 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -405,7 +405,6 @@ void tick_nohz_stop_sched_tick(int inidle)
 			ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
 			ts->tick_stopped = 1;
 			ts->idle_jiffies = last_jiffies;
-			rcu_idle_enter();
 		}
 
 		ts->idle_sleeps++;
@@ -444,6 +443,8 @@ out:
 	ts->last_jiffies = last_jiffies;
 	ts->sleep_length = ktime_sub(dev->next_event, now);
 end:
+	if (inidle)
+		rcu_idle_enter();
 	local_irq_restore(flags);
 }
 
@@ -500,6 +501,7 @@ void tick_nohz_restart_sched_tick(void)
 	ktime_t now;
 
 	local_irq_disable();
+	rcu_idle_exit();
 	if (ts->idle_active || (ts->inidle && ts->tick_stopped))
 		now = ktime_get();
 
@@ -514,8 +516,6 @@ void tick_nohz_restart_sched_tick(void)
 
 	ts->inidle = 0;
 
-	rcu_idle_exit();
-
 	/* Update jiffies first */
 	select_nohz_load_balancer(0);
 	tick_do_update_jiffies64(now);
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/