[PATCH -tip/core/rcu] Fix rcu_migrate_callback() to allow for multiple waiters

[PATCH -tip/core/rcu] Fix rcu_migrate_callback() to allow for multiple waiters

[Paul E. McKenney]

The current interaction between rcu_barrier() and CPU hotplug can result
in multiple tasks waiting on rcu_migrate_wq: there can be one CPU-hotplug
notifier, and, because rcu_barrier() invokes wait_migrated_callbacks()
outside of rcu_barrier_mutex, an arbitrarily large number of tasks
executing in _rcu_barrier().  This situation could result in hangs,
because rcu_migrate_callback() would wake up but one task.

This patch addresses this problem by awakening all sleepers.

Located-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
---

 rcupdate.c |    9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/rcupdate.c b/kernel/rcupdate.c
index bd5d5c8..6c59e1b 100644
--- a/kernel/rcupdate.c
+++ b/kernel/rcupdate.c
@@ -211,10 +211,17 @@ void rcu_barrier_sched(void)
 }
 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
 
+/*
+ * Wake up anyone waiting on migration of RCU callbacks from a CPU
+ * going offline.  There can be only one CPU-hotplug notifier waiting,
+ * but there can be any number of rcu_barrier() invocations waiting,
+ * due to the fact that rcu_barrier() does its wait outside of the
+ * rcu_barrier_mutex.
+ */
 static void rcu_migrate_callback(struct rcu_head *notused)
 {
 	if (atomic_dec_and_test(&rcu_migrate_type_count))
-		wake_up(&rcu_migrate_wq);
+		wake_up_all(&rcu_migrate_wq);
 }
 
 extern int rcu_cpu_notify(struct notifier_block *self,

Re: [PATCH -tip/core/rcu] Fix rcu_migrate_callback() to allow for multiple waiters

[Lai Jiangshan]

Paul E. McKenney wrote:
> The current interaction between rcu_barrier() and CPU hotplug can result
> in multiple tasks waiting on rcu_migrate_wq: there can be one CPU-hotplug
> notifier, and, because rcu_barrier() invokes wait_migrated_callbacks()
> outside of rcu_barrier_mutex, an arbitrarily large number of tasks
> executing in _rcu_barrier().  This situation could result in hangs,
> because rcu_migrate_callback() would wake up but one task.
> 
> This patch addresses this problem by awakening all sleepers.
> 
> Located-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> ---
> 
>  rcupdate.c |    9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/rcupdate.c b/kernel/rcupdate.c
> index bd5d5c8..6c59e1b 100644
> --- a/kernel/rcupdate.c
> +++ b/kernel/rcupdate.c
> @@ -211,10 +211,17 @@ void rcu_barrier_sched(void)
>  }
>  EXPORT_SYMBOL_GPL(rcu_barrier_sched);
>  
> +/*
> + * Wake up anyone waiting on migration of RCU callbacks from a CPU
> + * going offline.  There can be only one CPU-hotplug notifier waiting,
> + * but there can be any number of rcu_barrier() invocations waiting,
> + * due to the fact that rcu_barrier() does its wait outside of the
> + * rcu_barrier_mutex.
> + */
>  static void rcu_migrate_callback(struct rcu_head *notused)
>  {
>  	if (atomic_dec_and_test(&rcu_migrate_type_count))
> -		wake_up(&rcu_migrate_wq);
> +		wake_up_all(&rcu_migrate_wq);
>  }
>  
>  extern int rcu_cpu_notify(struct notifier_block *self,
> 

Maybe wake_up_all() is good for readability. But wake_up()
is enough for the logic.

In wait_migrated_callbacks(), we use wait_event(), The current
thread is queued as a non-exclusive waiter. And wake_up()
wakes up all non-exclusive waiter. (We only have
non-exclusive waiters in the rcu_migrate_wq)

Re: [PATCH -tip/core/rcu] Fix rcu_migrate_callback() to allow for multiple waiters

[Paul E. McKenney]

On Thu, Aug 20, 2009 at 11:30:45AM +0800, Lai Jiangshan wrote:
> Paul E. McKenney wrote:
> > The current interaction between rcu_barrier() and CPU hotplug can result
> > in multiple tasks waiting on rcu_migrate_wq: there can be one CPU-hotplug
> > notifier, and, because rcu_barrier() invokes wait_migrated_callbacks()
> > outside of rcu_barrier_mutex, an arbitrarily large number of tasks
> > executing in _rcu_barrier().  This situation could result in hangs,
> > because rcu_migrate_callback() would wake up but one task.
> > 
> > This patch addresses this problem by awakening all sleepers.
> > 
> > Located-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > ---
> > 
> >  rcupdate.c |    9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/rcupdate.c b/kernel/rcupdate.c
> > index bd5d5c8..6c59e1b 100644
> > --- a/kernel/rcupdate.c
> > +++ b/kernel/rcupdate.c
> > @@ -211,10 +211,17 @@ void rcu_barrier_sched(void)
> >  }
> >  EXPORT_SYMBOL_GPL(rcu_barrier_sched);
> >  
> > +/*
> > + * Wake up anyone waiting on migration of RCU callbacks from a CPU
> > + * going offline.  There can be only one CPU-hotplug notifier waiting,
> > + * but there can be any number of rcu_barrier() invocations waiting,
> > + * due to the fact that rcu_barrier() does its wait outside of the
> > + * rcu_barrier_mutex.
> > + */
> >  static void rcu_migrate_callback(struct rcu_head *notused)
> >  {
> >  	if (atomic_dec_and_test(&rcu_migrate_type_count))
> > -		wake_up(&rcu_migrate_wq);
> > +		wake_up_all(&rcu_migrate_wq);
> >  }
> >  
> >  extern int rcu_cpu_notify(struct notifier_block *self,
> > 
> 
> Maybe wake_up_all() is good for readability. But wake_up()
> is enough for the logic.
> 
> In wait_migrated_callbacks(), we use wait_event(), The current
> thread is queued as a non-exclusive waiter. And wake_up()
> wakes up all non-exclusive waiter. (We only have
> non-exclusive waiters in the rcu_migrate_wq)

And we even discussed this very point on March 18th!!!

OK, never mind, this patch has no effect.

							Thanx, Paul