Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1671244 > unrolled thread

Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu offline

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-06-21 01:50 +0200
    Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline Jeffrey Hugo <jhugo@codeaurora.org> - 2017-06-21 16:40 +0200
      Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-06-21 22:50 +0200
        Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-06-23 05:40 +0200
          Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline Jeffrey Hugo <jhugo@codeaurora.org> - 2017-06-28 00:40 +0200
            Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-06-28 02:20 +0200
              Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline Jeffrey Hugo <jhugo@codeaurora.org> - 2017-06-29 18:30 +0200
                Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu  offline "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-06-30 02:20 +0200

#1671244 — Re: [BUG] Deadlock due due to interactions of block, RCU, and cpu offline

On Mon, Mar 27, 2017 at 11:17:11AM -0700, Paul E. McKenney wrote:
> On Mon, Mar 27, 2017 at 12:02:27PM -0600, Jeffrey Hugo wrote:
> > Hi Paul.
> > 
> > Thanks for the quick reply.
> > 
> > On 3/26/2017 5:28 PM, Paul E. McKenney wrote:
> > >On Sun, Mar 26, 2017 at 05:10:40PM -0600, Jeffrey Hugo wrote:
> > 
> > >>It is a race between this work running, and the cpu offline processing.
> > >
> > >One quick way to test this assumption is to build a kernel with Kconfig
> > >options CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y.  This will
> > >cause call_rcu_sched() to queue the work to a kthread, which can migrate
> > >to some other CPU.  If your analysis is correct, this should avoid
> > >the deadlock.  (Note that the deadlock should be fixed in any case,
> > >just a diagnostic assumption-check procedure.)
> > 
> > I enabled CONFIG_RCU_EXPERT=y, CONFIG_RCU_NOCB_CPU=y,
> > CONFIG_RCU_NOCB_CPU_ALL=y in my build.  I've only had time so far to
> > do one test run however the issue reproduced, but it took a fair bit
> > longer to do so.  An initial look at the data indicates that the
> > work is still not running.  An odd observation, the two threads are
> > no longer blocked on the same queue, but different ones.
> 
> I was afraid of that...
> 
> > Let me look at this more and see what is going on now.
> 
> Another thing to try would be to affinity the "rcuo" kthreads to
> some CPU that is never taken offline, just in case that kthread is
> sometimes somehow getting stuck during the CPU-hotplug operation.
> 
> > >>What is the opinion of the domain experts?
> > >
> > >I do hope that we can come up with a better fix.  No offense intended,
> > >as coming up with -any- fix in the CPU-hotplug domain is not to be
> > >denigrated, but this looks to be at vest quite fragile.
> > >
> > >							Thanx, Paul
> > >
> > 
> > None taken.  I'm not particularly attached to the current fix.  I
> > agree, it does appear to be quite fragile.
> > 
> > I'm still not sure what a better solution would be though.  Maybe
> > the RCU framework flushes the work somehow during cpu offline?  It
> > would need to ensure further work is not queued after that point,
> > which seems like it might be tricky to synchronize.  I don't know
> > enough about the working of RCU to even attempt to implement that.
> 
> There are some ways that RCU might be able to shrink the window during
> which the outgoing CPU's callbacks are in limbo, but they are not free
> of risk, so we really need to compleetly understand what is going on
> before making any possibly ill-conceived changes.  ;-)
> 
> > In any case, it seem like some more analysis is needed based on the
> > latest data.
> 
> Looking forward to hearing about you find!

Hearing nothing, I eventually took unilateral action (I am a citizen of
USA, after all!) and produced the lightly tested patch shown below.

Does it help?

							Thanx, Paul

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

commit faeb334286b762ba6b829ad62e3d39a97674bb56
Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Date:   Tue Jun 20 12:11:34 2017 -0700

    rcu: Migrate callbacks earlier in the CPU-offline timeline
    
    RCU callbacks must be migrated away from an outgoing CPU, and this is
    done near the end of the CPU-hotplug operation, after the outgoing CPU is
    long gone.  Unfortunately, this means that other CPU-hotplug callbacks
    can execute while the outgoing CPU's callbacks are still immobilized
    on the long-gone CPU's callback lists.  If any of these CPU-hotplug
    callbacks must wait, either directly or indirectly, for the invocation
    of any of the immobilized RCU callbacks, the system will hang.
    
    This commit avoids such hangs by migrating the callbacks away from the
    outgoing CPU immediately upon its departure.  This migration makes use
    of the cpuhp_report_idle_dead function that is called by the outgoing CPU
    just after it has entered the idle loop for the last time.  This function
    IPIs an online CPU in order to wake up the thread orchestrating the
    CPU-hotplug operation, and this commit makes this IPI handler migrate the
    immobilized RCU callbacks.  Thus, RCU is able to advance these callbacks
    and invoke them, which allows all the after-the-fact CPU-hotplug callbacks
    to wait on these RCU callbacks without risk of a hang.
    
    Reported-by: Jeffrey Hugo <jhugo@codeaurora.org>
    Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org
    Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
    Cc: Ingo Molnar <mingo@kernel.org>
    Cc: Anna-Maria Gleixner <anna-maria@linutronix.de>
    Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
    Cc: Richard Weinberger <richard@nod.at>

diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index ecca9d2e4e4a..96f1baf62ab8 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -109,6 +109,7 @@ void rcu_bh_qs(void);
 void rcu_check_callbacks(int user);
 void rcu_report_dead(unsigned int cpu);
 void rcu_cpu_starting(unsigned int cpu);
+void rcutree_migrate_callbacks(int cpu);
 
 #ifdef CONFIG_RCU_STALL_COMMON
 void rcu_sysrq_start(void);
diff --git a/kernel/cpu.c b/kernel/cpu.c
index 9ae6fbe5b5cf..2f86f2f22a72 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -59,6 +59,7 @@ struct cpuhp_cpu_state {
 	struct hlist_node	*node;
 	enum cpuhp_state	cb_state;
 	int			result;
+	int			cpu;
 	struct completion	done;
 #endif
 };
@@ -736,6 +737,7 @@ static void cpuhp_complete_idle_dead(void *arg)
 {
 	struct cpuhp_cpu_state *st = arg;
 
+	rcutree_migrate_callbacks(st->cpu);
 	complete(&st->done);
 }
 
@@ -1828,5 +1830,11 @@ void __init boot_cpu_init(void)
  */
 void __init boot_cpu_state_init(void)
 {
+	int cpu;
+
 	per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
+#ifdef CONFIG_SMP
+	for_each_possible_cpu(cpu)
+		per_cpu_ptr(&cpuhp_state, cpu)->cpu = cpu;
+#endif
 }
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 94ec7455fc46..e6d534946c7f 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2547,85 +2547,6 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 }
 
 /*
- * Send the specified CPU's RCU callbacks to the orphanage.  The
- * specified CPU must be offline, and the caller must hold the
- * ->orphan_lock.
- */
-static void
-rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
-			  struct rcu_node *rnp, struct rcu_data *rdp)
-{
-	lockdep_assert_held(&rsp->orphan_lock);
-
-	/* No-CBs CPUs do not have orphanable callbacks. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
-		return;
-
-	/*
-	 * Orphan the callbacks.  First adjust the counts.  This is safe
-	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
-	 * cannot be running now.  Thus no memory barrier is required.
-	 */
-	rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
-	rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
-
-	/*
-	 * Next, move those callbacks still needing a grace period to
-	 * the orphanage, where some other CPU will pick them up.
-	 * Some of the callbacks might have gone partway through a grace
-	 * period, but that is too bad.  They get to start over because we
-	 * cannot assume that grace periods are synchronized across CPUs.
-	 */
-	rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
-
-	/*
-	 * Then move the ready-to-invoke callbacks to the orphanage,
-	 * where some other CPU will pick them up.  These will not be
-	 * required to pass though another grace period: They are done.
-	 */
-	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
-
-	/* Finally, disallow further callbacks on this CPU.  */
-	rcu_segcblist_disable(&rdp->cblist);
-}
-
-/*
- * Adopt the RCU callbacks from the specified rcu_state structure's
- * orphanage.  The caller must hold the ->orphan_lock.
- */
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
-{
-	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
-
-	lockdep_assert_held(&rsp->orphan_lock);
-
-	/* No-CBs CPUs are handled specially. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
-	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
-		return;
-
-	/* Do the accounting first. */
-	rdp->n_cbs_adopted += rsp->orphan_done.len;
-	if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
-		rcu_idle_count_callbacks_posted();
-	rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
-
-	/*
-	 * We do not need a memory barrier here because the only way we
-	 * can get here if there is an rcu_barrier() in flight is if
-	 * we are the task doing the rcu_barrier().
-	 */
-
-	/* First adopt the ready-to-invoke callbacks, then the done ones. */
-	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
-	WARN_ON_ONCE(rsp->orphan_done.head);
-	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
-	WARN_ON_ONCE(rsp->orphan_pend.head);
-	WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
-		     !rcu_segcblist_n_cbs(&rdp->cblist));
-}
-
-/*
  * Trace the fact that this CPU is going offline.
  */
 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
@@ -2695,7 +2616,6 @@ static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
  */
 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
-	unsigned long flags;
 	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
 
@@ -2704,18 +2624,6 @@ static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 
 	/* Adjust any no-longer-needed kthreads. */
 	rcu_boost_kthread_setaffinity(rnp, -1);
-
-	/* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
-	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
-	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
-	rcu_adopt_orphan_cbs(rsp, flags);
-	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
-
-	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
-		  !rcu_segcblist_empty(&rdp->cblist),
-		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
-		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
-		  rcu_segcblist_first_cb(&rdp->cblist));
 }
 
 /*
@@ -3927,6 +3835,116 @@ void rcu_report_dead(unsigned int cpu)
 	for_each_rcu_flavor(rsp)
 		rcu_cleanup_dying_idle_cpu(cpu, rsp);
 }
+
+/*
+ * Send the specified CPU's RCU callbacks to the orphanage.  The
+ * specified CPU must be offline, and the caller must hold the
+ * ->orphan_lock.
+ */
+static void
+rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
+			  struct rcu_node *rnp, struct rcu_data *rdp)
+{
+	lockdep_assert_held(&rsp->orphan_lock);
+
+	/* No-CBs CPUs do not have orphanable callbacks. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
+		return;
+
+	/*
+	 * Orphan the callbacks.  First adjust the counts.  This is safe
+	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
+	 * cannot be running now.  Thus no memory barrier is required.
+	 */
+	rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
+	rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
+
+	/*
+	 * Next, move those callbacks still needing a grace period to
+	 * the orphanage, where some other CPU will pick them up.
+	 * Some of the callbacks might have gone partway through a grace
+	 * period, but that is too bad.  They get to start over because we
+	 * cannot assume that grace periods are synchronized across CPUs.
+	 */
+	rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
+
+	/*
+	 * Then move the ready-to-invoke callbacks to the orphanage,
+	 * where some other CPU will pick them up.  These will not be
+	 * required to pass though another grace period: They are done.
+	 */
+	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
+
+	/* Finally, disallow further callbacks on this CPU.  */
+	rcu_segcblist_disable(&rdp->cblist);
+}
+
+/*
+ * Adopt the RCU callbacks from the specified rcu_state structure's
+ * orphanage.  The caller must hold the ->orphan_lock.
+ */
+static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
+{
+	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
+
+	lockdep_assert_held(&rsp->orphan_lock);
+
+	/* No-CBs CPUs are handled specially. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
+	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
+		return;
+
+	/* Do the accounting first. */
+	rdp->n_cbs_adopted += rsp->orphan_done.len;
+	if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
+		rcu_idle_count_callbacks_posted();
+	rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
+
+	/*
+	 * We do not need a memory barrier here because the only way we
+	 * can get here if there is an rcu_barrier() in flight is if
+	 * we are the task doing the rcu_barrier().
+	 */
+
+	/* First adopt the ready-to-invoke callbacks, then the done ones. */
+	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
+	WARN_ON_ONCE(rsp->orphan_done.head);
+	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
+	WARN_ON_ONCE(rsp->orphan_pend.head);
+	WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
+		     !rcu_segcblist_n_cbs(&rdp->cblist));
+}
+
+/* Orphan the dead CPU's callbacks, and then adopt them. */
+static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
+{
+	unsigned long flags;
+	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
+	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
+
+	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
+	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
+	rcu_adopt_orphan_cbs(rsp, flags);
+	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
+	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
+		  !rcu_segcblist_empty(&rdp->cblist),
+		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
+		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
+		  rcu_segcblist_first_cb(&rdp->cblist));
+}
+
+/*
+ * The outgoing CPU has just passed through the dying-idle state,
+ * and we are being invoked from the CPU that was IPIed to continue the
+ * offline operation.  We need to migrate the outgoing CPU's callbacks.
+ */
+void rcutree_migrate_callbacks(int cpu)
+{
+	struct rcu_state *rsp;
+
+	for_each_rcu_flavor(rsp)
+		rcu_migrate_callbacks(cpu, rsp);
+}
 #endif
 
 /*

[toc] | [next] | [standalone]

#1671716

On 6/20/2017 5:46 PM, Paul E. McKenney wrote:
> On Mon, Mar 27, 2017 at 11:17:11AM -0700, Paul E. McKenney wrote:
>> On Mon, Mar 27, 2017 at 12:02:27PM -0600, Jeffrey Hugo wrote:
>>> Hi Paul.
>>>
>>> Thanks for the quick reply.
>>>
>>> On 3/26/2017 5:28 PM, Paul E. McKenney wrote:
>>>> On Sun, Mar 26, 2017 at 05:10:40PM -0600, Jeffrey Hugo wrote:
>>>
>>>>> It is a race between this work running, and the cpu offline processing.
>>>>
>>>> One quick way to test this assumption is to build a kernel with Kconfig
>>>> options CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y.  This will
>>>> cause call_rcu_sched() to queue the work to a kthread, which can migrate
>>>> to some other CPU.  If your analysis is correct, this should avoid
>>>> the deadlock.  (Note that the deadlock should be fixed in any case,
>>>> just a diagnostic assumption-check procedure.)
>>>
>>> I enabled CONFIG_RCU_EXPERT=y, CONFIG_RCU_NOCB_CPU=y,
>>> CONFIG_RCU_NOCB_CPU_ALL=y in my build.  I've only had time so far to
>>> do one test run however the issue reproduced, but it took a fair bit
>>> longer to do so.  An initial look at the data indicates that the
>>> work is still not running.  An odd observation, the two threads are
>>> no longer blocked on the same queue, but different ones.
>>
>> I was afraid of that...
>>
>>> Let me look at this more and see what is going on now.
>>
>> Another thing to try would be to affinity the "rcuo" kthreads to
>> some CPU that is never taken offline, just in case that kthread is
>> sometimes somehow getting stuck during the CPU-hotplug operation.
>>
>>>>> What is the opinion of the domain experts?
>>>>
>>>> I do hope that we can come up with a better fix.  No offense intended,
>>>> as coming up with -any- fix in the CPU-hotplug domain is not to be
>>>> denigrated, but this looks to be at vest quite fragile.
>>>>
>>>> 							Thanx, Paul
>>>>
>>>
>>> None taken.  I'm not particularly attached to the current fix.  I
>>> agree, it does appear to be quite fragile.
>>>
>>> I'm still not sure what a better solution would be though.  Maybe
>>> the RCU framework flushes the work somehow during cpu offline?  It
>>> would need to ensure further work is not queued after that point,
>>> which seems like it might be tricky to synchronize.  I don't know
>>> enough about the working of RCU to even attempt to implement that.
>>
>> There are some ways that RCU might be able to shrink the window during
>> which the outgoing CPU's callbacks are in limbo, but they are not free
>> of risk, so we really need to compleetly understand what is going on
>> before making any possibly ill-conceived changes.  ;-)
>>
>>> In any case, it seem like some more analysis is needed based on the
>>> latest data.
>>
>> Looking forward to hearing about you find!
> 
> Hearing nothing, I eventually took unilateral action (I am a citizen of
> USA, after all!) and produced the lightly tested patch shown below.
> 
> Does it help?
> 
> 							Thanx, Paul
> 

Wow, has it been 3 months already?  I am extremely sorry, I've been 
preempted multiple times, and this has sat on my todo list where I keep 
thinking I need to find time to come back to it but apparently not doing 
enough to make that happen.

Thank you for not forgetting about this.  I promise I will somehow clear 
my schedule to test this next week.

Thank you again.

-- 
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm 
Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

[toc] | [prev] | [next] | [standalone]

#1672021

On Wed, Jun 21, 2017 at 08:39:45AM -0600, Jeffrey Hugo wrote:
> On 6/20/2017 5:46 PM, Paul E. McKenney wrote:
> >On Mon, Mar 27, 2017 at 11:17:11AM -0700, Paul E. McKenney wrote:
> >>On Mon, Mar 27, 2017 at 12:02:27PM -0600, Jeffrey Hugo wrote:
> >>>Hi Paul.
> >>>
> >>>Thanks for the quick reply.
> >>>
> >>>On 3/26/2017 5:28 PM, Paul E. McKenney wrote:
> >>>>On Sun, Mar 26, 2017 at 05:10:40PM -0600, Jeffrey Hugo wrote:
> >>>
> >>>>>It is a race between this work running, and the cpu offline processing.
> >>>>
> >>>>One quick way to test this assumption is to build a kernel with Kconfig
> >>>>options CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y.  This will
> >>>>cause call_rcu_sched() to queue the work to a kthread, which can migrate
> >>>>to some other CPU.  If your analysis is correct, this should avoid
> >>>>the deadlock.  (Note that the deadlock should be fixed in any case,
> >>>>just a diagnostic assumption-check procedure.)
> >>>
> >>>I enabled CONFIG_RCU_EXPERT=y, CONFIG_RCU_NOCB_CPU=y,
> >>>CONFIG_RCU_NOCB_CPU_ALL=y in my build.  I've only had time so far to
> >>>do one test run however the issue reproduced, but it took a fair bit
> >>>longer to do so.  An initial look at the data indicates that the
> >>>work is still not running.  An odd observation, the two threads are
> >>>no longer blocked on the same queue, but different ones.
> >>
> >>I was afraid of that...
> >>
> >>>Let me look at this more and see what is going on now.
> >>
> >>Another thing to try would be to affinity the "rcuo" kthreads to
> >>some CPU that is never taken offline, just in case that kthread is
> >>sometimes somehow getting stuck during the CPU-hotplug operation.
> >>
> >>>>>What is the opinion of the domain experts?
> >>>>
> >>>>I do hope that we can come up with a better fix.  No offense intended,
> >>>>as coming up with -any- fix in the CPU-hotplug domain is not to be
> >>>>denigrated, but this looks to be at vest quite fragile.
> >>>>
> >>>>							Thanx, Paul
> >>>>
> >>>
> >>>None taken.  I'm not particularly attached to the current fix.  I
> >>>agree, it does appear to be quite fragile.
> >>>
> >>>I'm still not sure what a better solution would be though.  Maybe
> >>>the RCU framework flushes the work somehow during cpu offline?  It
> >>>would need to ensure further work is not queued after that point,
> >>>which seems like it might be tricky to synchronize.  I don't know
> >>>enough about the working of RCU to even attempt to implement that.
> >>
> >>There are some ways that RCU might be able to shrink the window during
> >>which the outgoing CPU's callbacks are in limbo, but they are not free
> >>of risk, so we really need to compleetly understand what is going on
> >>before making any possibly ill-conceived changes.  ;-)
> >>
> >>>In any case, it seem like some more analysis is needed based on the
> >>>latest data.
> >>
> >>Looking forward to hearing about you find!
> >
> >Hearing nothing, I eventually took unilateral action (I am a citizen of
> >USA, after all!) and produced the lightly tested patch shown below.
> >
> >Does it help?
> >
> >							Thanx, Paul
> >
> 
> Wow, has it been 3 months already?  I am extremely sorry, I've been
> preempted multiple times, and this has sat on my todo list where I
> keep thinking I need to find time to come back to it but apparently
> not doing enough to make that happen.
> 
> Thank you for not forgetting about this.  I promise I will somehow
> clear my schedule to test this next week.

No worries, and I am very much looking forward to seeing the results of
your testing.

							Thanx, Paul

[toc] | [prev] | [next] | [standalone]

#1673197

On Wed, Jun 21, 2017 at 09:18:53AM -0700, Paul E. McKenney wrote:
> On Wed, Jun 21, 2017 at 08:39:45AM -0600, Jeffrey Hugo wrote:
> > On 6/20/2017 5:46 PM, Paul E. McKenney wrote:
> > >On Mon, Mar 27, 2017 at 11:17:11AM -0700, Paul E. McKenney wrote:
> > >>On Mon, Mar 27, 2017 at 12:02:27PM -0600, Jeffrey Hugo wrote:
> > >>>Hi Paul.
> > >>>
> > >>>Thanks for the quick reply.
> > >>>
> > >>>On 3/26/2017 5:28 PM, Paul E. McKenney wrote:
> > >>>>On Sun, Mar 26, 2017 at 05:10:40PM -0600, Jeffrey Hugo wrote:
> > >>>
> > >>>>>It is a race between this work running, and the cpu offline processing.
> > >>>>
> > >>>>One quick way to test this assumption is to build a kernel with Kconfig
> > >>>>options CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y.  This will
> > >>>>cause call_rcu_sched() to queue the work to a kthread, which can migrate
> > >>>>to some other CPU.  If your analysis is correct, this should avoid
> > >>>>the deadlock.  (Note that the deadlock should be fixed in any case,
> > >>>>just a diagnostic assumption-check procedure.)
> > >>>
> > >>>I enabled CONFIG_RCU_EXPERT=y, CONFIG_RCU_NOCB_CPU=y,
> > >>>CONFIG_RCU_NOCB_CPU_ALL=y in my build.  I've only had time so far to
> > >>>do one test run however the issue reproduced, but it took a fair bit
> > >>>longer to do so.  An initial look at the data indicates that the
> > >>>work is still not running.  An odd observation, the two threads are
> > >>>no longer blocked on the same queue, but different ones.
> > >>
> > >>I was afraid of that...
> > >>
> > >>>Let me look at this more and see what is going on now.
> > >>
> > >>Another thing to try would be to affinity the "rcuo" kthreads to
> > >>some CPU that is never taken offline, just in case that kthread is
> > >>sometimes somehow getting stuck during the CPU-hotplug operation.
> > >>
> > >>>>>What is the opinion of the domain experts?
> > >>>>
> > >>>>I do hope that we can come up with a better fix.  No offense intended,
> > >>>>as coming up with -any- fix in the CPU-hotplug domain is not to be
> > >>>>denigrated, but this looks to be at vest quite fragile.
> > >>>>
> > >>>>							Thanx, Paul
> > >>>>
> > >>>
> > >>>None taken.  I'm not particularly attached to the current fix.  I
> > >>>agree, it does appear to be quite fragile.
> > >>>
> > >>>I'm still not sure what a better solution would be though.  Maybe
> > >>>the RCU framework flushes the work somehow during cpu offline?  It
> > >>>would need to ensure further work is not queued after that point,
> > >>>which seems like it might be tricky to synchronize.  I don't know
> > >>>enough about the working of RCU to even attempt to implement that.
> > >>
> > >>There are some ways that RCU might be able to shrink the window during
> > >>which the outgoing CPU's callbacks are in limbo, but they are not free
> > >>of risk, so we really need to compleetly understand what is going on
> > >>before making any possibly ill-conceived changes.  ;-)
> > >>
> > >>>In any case, it seem like some more analysis is needed based on the
> > >>>latest data.
> > >>
> > >>Looking forward to hearing about you find!
> > >
> > >Hearing nothing, I eventually took unilateral action (I am a citizen of
> > >USA, after all!) and produced the lightly tested patch shown below.
> > >
> > >Does it help?
> > >
> > >							Thanx, Paul
> > >
> > 
> > Wow, has it been 3 months already?  I am extremely sorry, I've been
> > preempted multiple times, and this has sat on my todo list where I
> > keep thinking I need to find time to come back to it but apparently
> > not doing enough to make that happen.
> > 
> > Thank you for not forgetting about this.  I promise I will somehow
> > clear my schedule to test this next week.
> 
> No worries, and I am very much looking forward to seeing the results of
> your testing.

And please see below for an updated patch based on LKML review and
more intensive testing.

							Thanx, Paul

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

commit 4384c26f62d90e9685a50d65dcd352dbe96ae220
Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Date:   Tue Jun 20 12:11:34 2017 -0700

    rcu: Migrate callbacks earlier in the CPU-offline timeline
    
    RCU callbacks must be migrated away from an outgoing CPU, and this is
    done near the end of the CPU-hotplug operation, after the outgoing CPU is
    long gone.  Unfortunately, this means that other CPU-hotplug callbacks
    can execute while the outgoing CPU's callbacks are still immobilized
    on the long-gone CPU's callback lists.  If any of these CPU-hotplug
    callbacks must wait, either directly or indirectly, for the invocation
    of any of the immobilized RCU callbacks, the system will hang.
    
    This commit avoids such hangs by migrating the callbacks away from the
    outgoing CPU immediately upon its departure, shortly after the return
    from __cpu_die() in takedown_cpu().  Thus, RCU is able to advance these
    callbacks and invoke them, which allows all the after-the-fact CPU-hotplug
    callbacks to wait on these RCU callbacks without risk of a hang.
    
    While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage()
    and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including
    dead code on the one hand and to avoid define-without-use warnings on the
    other hand.
    
    Reported-by: Jeffrey Hugo <jhugo@codeaurora.org>
    Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org
    Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
    Cc: Ingo Molnar <mingo@kernel.org>
    Cc: Anna-Maria Gleixner <anna-maria@linutronix.de>
    Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
    Cc: Richard Weinberger <richard@nod.at>

diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index ecca9d2e4e4a..96f1baf62ab8 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -109,6 +109,7 @@ void rcu_bh_qs(void);
 void rcu_check_callbacks(int user);
 void rcu_report_dead(unsigned int cpu);
 void rcu_cpu_starting(unsigned int cpu);
+void rcutree_migrate_callbacks(int cpu);
 
 #ifdef CONFIG_RCU_STALL_COMMON
 void rcu_sysrq_start(void);
diff --git a/kernel/cpu.c b/kernel/cpu.c
index 9ae6fbe5b5cf..f5f3db2403fa 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -729,6 +729,7 @@ static int takedown_cpu(unsigned int cpu)
 	__cpu_die(cpu);
 
 	tick_cleanup_dead_cpu(cpu);
+	rcutree_migrate_callbacks(cpu);
 	return 0;
 }
 
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 94ec7455fc46..e6d534946c7f 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2547,85 +2547,6 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 }
 
 /*
- * Send the specified CPU's RCU callbacks to the orphanage.  The
- * specified CPU must be offline, and the caller must hold the
- * ->orphan_lock.
- */
-static void
-rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
-			  struct rcu_node *rnp, struct rcu_data *rdp)
-{
-	lockdep_assert_held(&rsp->orphan_lock);
-
-	/* No-CBs CPUs do not have orphanable callbacks. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
-		return;
-
-	/*
-	 * Orphan the callbacks.  First adjust the counts.  This is safe
-	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
-	 * cannot be running now.  Thus no memory barrier is required.
-	 */
-	rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
-	rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
-
-	/*
-	 * Next, move those callbacks still needing a grace period to
-	 * the orphanage, where some other CPU will pick them up.
-	 * Some of the callbacks might have gone partway through a grace
-	 * period, but that is too bad.  They get to start over because we
-	 * cannot assume that grace periods are synchronized across CPUs.
-	 */
-	rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
-
-	/*
-	 * Then move the ready-to-invoke callbacks to the orphanage,
-	 * where some other CPU will pick them up.  These will not be
-	 * required to pass though another grace period: They are done.
-	 */
-	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
-
-	/* Finally, disallow further callbacks on this CPU.  */
-	rcu_segcblist_disable(&rdp->cblist);
-}
-
-/*
- * Adopt the RCU callbacks from the specified rcu_state structure's
- * orphanage.  The caller must hold the ->orphan_lock.
- */
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
-{
-	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
-
-	lockdep_assert_held(&rsp->orphan_lock);
-
-	/* No-CBs CPUs are handled specially. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
-	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
-		return;
-
-	/* Do the accounting first. */
-	rdp->n_cbs_adopted += rsp->orphan_done.len;
-	if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
-		rcu_idle_count_callbacks_posted();
-	rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
-
-	/*
-	 * We do not need a memory barrier here because the only way we
-	 * can get here if there is an rcu_barrier() in flight is if
-	 * we are the task doing the rcu_barrier().
-	 */
-
-	/* First adopt the ready-to-invoke callbacks, then the done ones. */
-	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
-	WARN_ON_ONCE(rsp->orphan_done.head);
-	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
-	WARN_ON_ONCE(rsp->orphan_pend.head);
-	WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
-		     !rcu_segcblist_n_cbs(&rdp->cblist));
-}
-
-/*
  * Trace the fact that this CPU is going offline.
  */
 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
@@ -2695,7 +2616,6 @@ static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
  */
 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
-	unsigned long flags;
 	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
 
@@ -2704,18 +2624,6 @@ static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 
 	/* Adjust any no-longer-needed kthreads. */
 	rcu_boost_kthread_setaffinity(rnp, -1);
-
-	/* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
-	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
-	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
-	rcu_adopt_orphan_cbs(rsp, flags);
-	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
-
-	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
-		  !rcu_segcblist_empty(&rdp->cblist),
-		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
-		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
-		  rcu_segcblist_first_cb(&rdp->cblist));
 }
 
 /*
@@ -3927,6 +3835,116 @@ void rcu_report_dead(unsigned int cpu)
 	for_each_rcu_flavor(rsp)
 		rcu_cleanup_dying_idle_cpu(cpu, rsp);
 }
+
+/*
+ * Send the specified CPU's RCU callbacks to the orphanage.  The
+ * specified CPU must be offline, and the caller must hold the
+ * ->orphan_lock.
+ */
+static void
+rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
+			  struct rcu_node *rnp, struct rcu_data *rdp)
+{
+	lockdep_assert_held(&rsp->orphan_lock);
+
+	/* No-CBs CPUs do not have orphanable callbacks. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
+		return;
+
+	/*
+	 * Orphan the callbacks.  First adjust the counts.  This is safe
+	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
+	 * cannot be running now.  Thus no memory barrier is required.
+	 */
+	rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
+	rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
+
+	/*
+	 * Next, move those callbacks still needing a grace period to
+	 * the orphanage, where some other CPU will pick them up.
+	 * Some of the callbacks might have gone partway through a grace
+	 * period, but that is too bad.  They get to start over because we
+	 * cannot assume that grace periods are synchronized across CPUs.
+	 */
+	rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
+
+	/*
+	 * Then move the ready-to-invoke callbacks to the orphanage,
+	 * where some other CPU will pick them up.  These will not be
+	 * required to pass though another grace period: They are done.
+	 */
+	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
+
+	/* Finally, disallow further callbacks on this CPU.  */
+	rcu_segcblist_disable(&rdp->cblist);
+}
+
+/*
+ * Adopt the RCU callbacks from the specified rcu_state structure's
+ * orphanage.  The caller must hold the ->orphan_lock.
+ */
+static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
+{
+	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
+
+	lockdep_assert_held(&rsp->orphan_lock);
+
+	/* No-CBs CPUs are handled specially. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
+	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
+		return;
+
+	/* Do the accounting first. */
+	rdp->n_cbs_adopted += rsp->orphan_done.len;
+	if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
+		rcu_idle_count_callbacks_posted();
+	rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
+
+	/*
+	 * We do not need a memory barrier here because the only way we
+	 * can get here if there is an rcu_barrier() in flight is if
+	 * we are the task doing the rcu_barrier().
+	 */
+
+	/* First adopt the ready-to-invoke callbacks, then the done ones. */
+	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
+	WARN_ON_ONCE(rsp->orphan_done.head);
+	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
+	WARN_ON_ONCE(rsp->orphan_pend.head);
+	WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
+		     !rcu_segcblist_n_cbs(&rdp->cblist));
+}
+
+/* Orphan the dead CPU's callbacks, and then adopt them. */
+static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
+{
+	unsigned long flags;
+	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
+	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
+
+	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
+	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
+	rcu_adopt_orphan_cbs(rsp, flags);
+	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
+	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
+		  !rcu_segcblist_empty(&rdp->cblist),
+		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
+		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
+		  rcu_segcblist_first_cb(&rdp->cblist));
+}
+
+/*
+ * The outgoing CPU has just passed through the dying-idle state,
+ * and we are being invoked from the CPU that was IPIed to continue the
+ * offline operation.  We need to migrate the outgoing CPU's callbacks.
+ */
+void rcutree_migrate_callbacks(int cpu)
+{
+	struct rcu_state *rsp;
+
+	for_each_rcu_flavor(rsp)
+		rcu_migrate_callbacks(cpu, rsp);
+}
 #endif
 
 /*

[toc] | [prev] | [next] | [standalone]

#1676223

On 6/22/2017 9:34 PM, Paul E. McKenney wrote:
> On Wed, Jun 21, 2017 at 09:18:53AM -0700, Paul E. McKenney wrote:
>> No worries, and I am very much looking forward to seeing the results of
>> your testing.
> 
> And please see below for an updated patch based on LKML review and
> more intensive testing.
> 

I spent some time on this today.  It didn't go as I expected.  I 
validated the issue is reproducible as before on 4.11 and 4.12 rcs 1 
through 4.  However, the version of stress-ng that I was using ran into 
constant errors starting with rc5, making it nearly impossible to make 
progress toward reproduction.  Upgrading stress-ng to tip fixes the 
issue, however, I've still been unable to repro the issue.

Its my unfounded suspicion that something went in between rc4 and rc5 
which changed the timing, and didn't actually fix the issue.  I will run 
the test overnight for 5 hours to try to repro.

The patch you sent appears to be based on linux-next, and appears to 
have a number of dependencies which prevent it from cleanly applying on 
anything current that I'm able to repro on at this time.  Do you want to 
provide a rebased version of the patch which applies to say 4.11?  I 
could easily test that and report back.

-- 
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm 
Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

[toc] | [prev] | [next] | [standalone]

#1676283

On Tue, Jun 27, 2017 at 04:32:09PM -0600, Jeffrey Hugo wrote:
> On 6/22/2017 9:34 PM, Paul E. McKenney wrote:
> >On Wed, Jun 21, 2017 at 09:18:53AM -0700, Paul E. McKenney wrote:
> >>No worries, and I am very much looking forward to seeing the results of
> >>your testing.
> >
> >And please see below for an updated patch based on LKML review and
> >more intensive testing.
> >
> 
> I spent some time on this today.  It didn't go as I expected.  I
> validated the issue is reproducible as before on 4.11 and 4.12 rcs 1
> through 4.  However, the version of stress-ng that I was using ran
> into constant errors starting with rc5, making it nearly impossible
> to make progress toward reproduction.  Upgrading stress-ng to tip
> fixes the issue, however, I've still been unable to repro the issue.
> 
> Its my unfounded suspicion that something went in between rc4 and
> rc5 which changed the timing, and didn't actually fix the issue.  I
> will run the test overnight for 5 hours to try to repro.
> 
> The patch you sent appears to be based on linux-next, and appears to
> have a number of dependencies which prevent it from cleanly applying
> on anything current that I'm able to repro on at this time.  Do you
> want to provide a rebased version of the patch which applies to say
> 4.11?  I could easily test that and report back.

Here is a very lightly tested backport to v4.11.

							Thanx, Paul

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

commit 5557f5d8435dcfdc81ff090aceb57734448d3051
Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Date:   Tue Jun 20 12:11:34 2017 -0700

    rcu: Migrate callbacks earlier in the CPU-offline timeline
    
    RCU callbacks must be migrated away from an outgoing CPU, and this is
    done near the end of the CPU-hotplug operation, after the outgoing CPU is
    long gone.  Unfortunately, this means that other CPU-hotplug callbacks
    can execute while the outgoing CPU's callbacks are still immobilized
    on the long-gone CPU's callback lists.  If any of these CPU-hotplug
    callbacks must wait, either directly or indirectly, for the invocation
    of any of the immobilized RCU callbacks, the system will hang.
    
    This commit avoids such hangs by migrating the callbacks away from the
    outgoing CPU immediately upon its departure, shortly after the return
    from __cpu_die() in takedown_cpu().  Thus, RCU is able to advance these
    callbacks and invoke them, which allows all the after-the-fact CPU-hotplug
    callbacks to wait on these RCU callbacks without risk of a hang.
    
    While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage()
    and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including
    dead code on the one hand and to avoid define-without-use warnings on the
    other hand.
    
    Reported-by: Jeffrey Hugo <jhugo@codeaurora.org>
    Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org
    Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
    Cc: Ingo Molnar <mingo@kernel.org>
    Cc: Anna-Maria Gleixner <anna-maria@linutronix.de>
    Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
    Cc: Richard Weinberger <richard@nod.at>

diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index de88b33c0974..183d69438776 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -295,6 +295,7 @@ void rcu_bh_qs(void);
 void rcu_check_callbacks(int user);
 void rcu_report_dead(unsigned int cpu);
 void rcu_cpu_starting(unsigned int cpu);
+void rcutree_migrate_callbacks(int cpu);
 
 #ifndef CONFIG_TINY_RCU
 void rcu_end_inkernel_boot(void);
diff --git a/kernel/cpu.c b/kernel/cpu.c
index 37b223e4fc05..21be6ab54ea2 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -729,6 +729,7 @@ static int takedown_cpu(unsigned int cpu)
 	__cpu_die(cpu);
 
 	tick_cleanup_dead_cpu(cpu);
+	rcutree_migrate_callbacks(cpu);
 	return 0;
 }
 
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 50fee7689e71..63206f81574a 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2636,114 +2636,6 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 }
 
 /*
- * Send the specified CPU's RCU callbacks to the orphanage.  The
- * specified CPU must be offline, and the caller must hold the
- * ->orphan_lock.
- */
-static void
-rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
-			  struct rcu_node *rnp, struct rcu_data *rdp)
-{
-	/* No-CBs CPUs do not have orphanable callbacks. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
-		return;
-
-	/*
-	 * Orphan the callbacks.  First adjust the counts.  This is safe
-	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
-	 * cannot be running now.  Thus no memory barrier is required.
-	 */
-	if (rdp->nxtlist != NULL) {
-		rsp->qlen_lazy += rdp->qlen_lazy;
-		rsp->qlen += rdp->qlen;
-		rdp->n_cbs_orphaned += rdp->qlen;
-		rdp->qlen_lazy = 0;
-		WRITE_ONCE(rdp->qlen, 0);
-	}
-
-	/*
-	 * Next, move those callbacks still needing a grace period to
-	 * the orphanage, where some other CPU will pick them up.
-	 * Some of the callbacks might have gone partway through a grace
-	 * period, but that is too bad.  They get to start over because we
-	 * cannot assume that grace periods are synchronized across CPUs.
-	 * We don't bother updating the ->nxttail[] array yet, instead
-	 * we just reset the whole thing later on.
-	 */
-	if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
-		*rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
-		rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
-		*rdp->nxttail[RCU_DONE_TAIL] = NULL;
-	}
-
-	/*
-	 * Then move the ready-to-invoke callbacks to the orphanage,
-	 * where some other CPU will pick them up.  These will not be
-	 * required to pass though another grace period: They are done.
-	 */
-	if (rdp->nxtlist != NULL) {
-		*rsp->orphan_donetail = rdp->nxtlist;
-		rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
-	}
-
-	/*
-	 * Finally, initialize the rcu_data structure's list to empty and
-	 * disallow further callbacks on this CPU.
-	 */
-	init_callback_list(rdp);
-	rdp->nxttail[RCU_NEXT_TAIL] = NULL;
-}
-
-/*
- * Adopt the RCU callbacks from the specified rcu_state structure's
- * orphanage.  The caller must hold the ->orphan_lock.
- */
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
-{
-	int i;
-	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
-
-	/* No-CBs CPUs are handled specially. */
-	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
-	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
-		return;
-
-	/* Do the accounting first. */
-	rdp->qlen_lazy += rsp->qlen_lazy;
-	rdp->qlen += rsp->qlen;
-	rdp->n_cbs_adopted += rsp->qlen;
-	if (rsp->qlen_lazy != rsp->qlen)
-		rcu_idle_count_callbacks_posted();
-	rsp->qlen_lazy = 0;
-	rsp->qlen = 0;
-
-	/*
-	 * We do not need a memory barrier here because the only way we
-	 * can get here if there is an rcu_barrier() in flight is if
-	 * we are the task doing the rcu_barrier().
-	 */
-
-	/* First adopt the ready-to-invoke callbacks. */
-	if (rsp->orphan_donelist != NULL) {
-		*rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
-		*rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
-		for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
-			if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
-				rdp->nxttail[i] = rsp->orphan_donetail;
-		rsp->orphan_donelist = NULL;
-		rsp->orphan_donetail = &rsp->orphan_donelist;
-	}
-
-	/* And then adopt the callbacks that still need a grace period. */
-	if (rsp->orphan_nxtlist != NULL) {
-		*rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
-		rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
-		rsp->orphan_nxtlist = NULL;
-		rsp->orphan_nxttail = &rsp->orphan_nxtlist;
-	}
-}
-
-/*
  * Trace the fact that this CPU is going offline.
  */
 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
@@ -2805,14 +2697,12 @@ static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
 
 /*
  * The CPU has been completely removed, and some other CPU is reporting
- * this fact from process context.  Do the remainder of the cleanup,
- * including orphaning the outgoing CPU's RCU callbacks, and also
- * adopting them.  There can only be one CPU hotplug operation at a time,
- * so no other CPU can be attempting to update rcu_cpu_kthread_task.
+ * this fact from process context.  Do the remainder of the cleanup.
+ * There can only be one CPU hotplug operation at a time, so no need for
+ * explicit locking.
  */
 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
-	unsigned long flags;
 	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
 
@@ -2821,16 +2711,6 @@ static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 
 	/* Adjust any no-longer-needed kthreads. */
 	rcu_boost_kthread_setaffinity(rnp, -1);
-
-	/* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
-	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
-	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
-	rcu_adopt_orphan_cbs(rsp, flags);
-	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
-
-	WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
-		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
-		  cpu, rdp->qlen, rdp->nxtlist);
 }
 
 /*
@@ -4011,6 +3891,140 @@ void rcu_report_dead(unsigned int cpu)
 	for_each_rcu_flavor(rsp)
 		rcu_cleanup_dying_idle_cpu(cpu, rsp);
 }
+
+/*
+ * Send the specified CPU's RCU callbacks to the orphanage.  The
+ * specified CPU must be offline, and the caller must hold the
+ * ->orphan_lock.
+ */
+static void
+rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
+			  struct rcu_node *rnp, struct rcu_data *rdp)
+{
+	/* No-CBs CPUs do not have orphanable callbacks. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
+		return;
+
+	/*
+	 * Orphan the callbacks.  First adjust the counts.  This is safe
+	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
+	 * cannot be running now.  Thus no memory barrier is required.
+	 */
+	if (rdp->nxtlist != NULL) {
+		rsp->qlen_lazy += rdp->qlen_lazy;
+		rsp->qlen += rdp->qlen;
+		rdp->n_cbs_orphaned += rdp->qlen;
+		rdp->qlen_lazy = 0;
+		WRITE_ONCE(rdp->qlen, 0);
+	}
+
+	/*
+	 * Next, move those callbacks still needing a grace period to
+	 * the orphanage, where some other CPU will pick them up.
+	 * Some of the callbacks might have gone partway through a grace
+	 * period, but that is too bad.  They get to start over because we
+	 * cannot assume that grace periods are synchronized across CPUs.
+	 * We don't bother updating the ->nxttail[] array yet, instead
+	 * we just reset the whole thing later on.
+	 */
+	if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
+		*rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
+		rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
+		*rdp->nxttail[RCU_DONE_TAIL] = NULL;
+	}
+
+	/*
+	 * Then move the ready-to-invoke callbacks to the orphanage,
+	 * where some other CPU will pick them up.  These will not be
+	 * required to pass though another grace period: They are done.
+	 */
+	if (rdp->nxtlist != NULL) {
+		*rsp->orphan_donetail = rdp->nxtlist;
+		rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
+	}
+
+	/*
+	 * Finally, initialize the rcu_data structure's list to empty and
+	 * disallow further callbacks on this CPU.
+	 */
+	init_callback_list(rdp);
+	rdp->nxttail[RCU_NEXT_TAIL] = NULL;
+}
+
+/*
+ * Adopt the RCU callbacks from the specified rcu_state structure's
+ * orphanage.  The caller must hold the ->orphan_lock.
+ */
+static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
+{
+	int i;
+	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
+
+	/* No-CBs CPUs are handled specially. */
+	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
+	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
+		return;
+
+	/* Do the accounting first. */
+	rdp->qlen_lazy += rsp->qlen_lazy;
+	rdp->qlen += rsp->qlen;
+	rdp->n_cbs_adopted += rsp->qlen;
+	if (rsp->qlen_lazy != rsp->qlen)
+		rcu_idle_count_callbacks_posted();
+	rsp->qlen_lazy = 0;
+	rsp->qlen = 0;
+
+	/*
+	 * We do not need a memory barrier here because the only way we
+	 * can get here if there is an rcu_barrier() in flight is if
+	 * we are the task doing the rcu_barrier().
+	 */
+
+	/* First adopt the ready-to-invoke callbacks. */
+	if (rsp->orphan_donelist != NULL) {
+		*rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
+		*rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
+		for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
+			if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
+				rdp->nxttail[i] = rsp->orphan_donetail;
+		rsp->orphan_donelist = NULL;
+		rsp->orphan_donetail = &rsp->orphan_donelist;
+	}
+
+	/* And then adopt the callbacks that still need a grace period. */
+	if (rsp->orphan_nxtlist != NULL) {
+		*rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
+		rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
+		rsp->orphan_nxtlist = NULL;
+		rsp->orphan_nxttail = &rsp->orphan_nxtlist;
+	}
+}
+
+/* Orphan the dead CPU's callbacks, and then adopt them. */
+static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
+{
+	unsigned long flags;
+	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
+	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
+
+	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
+	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
+	rcu_adopt_orphan_cbs(rsp, flags);
+	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
+}
+
+/*
+ * The outgoing CPU has just passed through the dying-idle state,
+ * and we are being invoked from the CPU that was IPIed to continue the
+ * offline operation.  We need to migrate the outgoing CPU's callbacks.
+ */
+void rcutree_migrate_callbacks(int cpu)
+{
+	struct rcu_state *rsp;
+
+	for_each_rcu_flavor(rsp)
+		rcu_migrate_callbacks(cpu, rsp);
+}
 #endif
 
 static int rcu_pm_notify(struct notifier_block *self,

[toc] | [prev] | [next] | [standalone]

#1677922

On 6/27/2017 6:11 PM, Paul E. McKenney wrote:
> On Tue, Jun 27, 2017 at 04:32:09PM -0600, Jeffrey Hugo wrote:
>> On 6/22/2017 9:34 PM, Paul E. McKenney wrote:
>>> On Wed, Jun 21, 2017 at 09:18:53AM -0700, Paul E. McKenney wrote:
>>>> No worries, and I am very much looking forward to seeing the results of
>>>> your testing.
>>>
>>> And please see below for an updated patch based on LKML review and
>>> more intensive testing.
>>>
>>
>> I spent some time on this today.  It didn't go as I expected.  I
>> validated the issue is reproducible as before on 4.11 and 4.12 rcs 1
>> through 4.  However, the version of stress-ng that I was using ran
>> into constant errors starting with rc5, making it nearly impossible
>> to make progress toward reproduction.  Upgrading stress-ng to tip
>> fixes the issue, however, I've still been unable to repro the issue.
>>
>> Its my unfounded suspicion that something went in between rc4 and
>> rc5 which changed the timing, and didn't actually fix the issue.  I
>> will run the test overnight for 5 hours to try to repro.
>>
>> The patch you sent appears to be based on linux-next, and appears to
>> have a number of dependencies which prevent it from cleanly applying
>> on anything current that I'm able to repro on at this time.  Do you
>> want to provide a rebased version of the patch which applies to say
>> 4.11?  I could easily test that and report back.
> 
> Here is a very lightly tested backport to v4.11.
> 

Works for me. Always reproduced the lockup within 2 minutes on stock 
4.11.  With the change applied, I was able to test for 2 hours in the 
same conditions, and 4 hours with the full system and not encounter an 
issue.

Feel free to add:
Tested-by: Jeffrey Hugo <jhugo@codeaurora.org>

I'm going to go back to 4.12-rc5 and see if I can get either repro the 
issue, or identify what changed.  Hopefully I can get to linux-next and 
double check the original version of the change as well.

-- 
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm 
Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

[toc] | [prev] | [next] | [standalone]

#1678330

On Thu, Jun 29, 2017 at 10:29:12AM -0600, Jeffrey Hugo wrote:
> On 6/27/2017 6:11 PM, Paul E. McKenney wrote:
> >On Tue, Jun 27, 2017 at 04:32:09PM -0600, Jeffrey Hugo wrote:
> >>On 6/22/2017 9:34 PM, Paul E. McKenney wrote:
> >>>On Wed, Jun 21, 2017 at 09:18:53AM -0700, Paul E. McKenney wrote:
> >>>>No worries, and I am very much looking forward to seeing the results of
> >>>>your testing.
> >>>
> >>>And please see below for an updated patch based on LKML review and
> >>>more intensive testing.
> >>>
> >>
> >>I spent some time on this today.  It didn't go as I expected.  I
> >>validated the issue is reproducible as before on 4.11 and 4.12 rcs 1
> >>through 4.  However, the version of stress-ng that I was using ran
> >>into constant errors starting with rc5, making it nearly impossible
> >>to make progress toward reproduction.  Upgrading stress-ng to tip
> >>fixes the issue, however, I've still been unable to repro the issue.
> >>
> >>Its my unfounded suspicion that something went in between rc4 and
> >>rc5 which changed the timing, and didn't actually fix the issue.  I
> >>will run the test overnight for 5 hours to try to repro.
> >>
> >>The patch you sent appears to be based on linux-next, and appears to
> >>have a number of dependencies which prevent it from cleanly applying
> >>on anything current that I'm able to repro on at this time.  Do you
> >>want to provide a rebased version of the patch which applies to say
> >>4.11?  I could easily test that and report back.
> >
> >Here is a very lightly tested backport to v4.11.
> >
> 
> Works for me. Always reproduced the lockup within 2 minutes on stock
> 4.11.  With the change applied, I was able to test for 2 hours in
> the same conditions, and 4 hours with the full system and not
> encounter an issue.
> 
> Feel free to add:
> Tested-by: Jeffrey Hugo <jhugo@codeaurora.org>

Applied, thank you!

> I'm going to go back to 4.12-rc5 and see if I can get either repro
> the issue, or identify what changed.  Hopefully I can get to
> linux-next and double check the original version of the change as
> well.

Looking forward to hearing what you find!

							Thanx, Paul

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web