Re: [PATCH tip/core/rcu 10/20] rcu: Add functions to test for trivial grace periods

From: Paul E. McKenney
Date: Mon Jan 16 2017 - 22:05:47 EST


On Mon, Jan 16, 2017 at 12:01:29AM -0800, Josh Triplett wrote:
> On Sat, Jan 14, 2017 at 01:13:11AM -0800, Paul E. McKenney wrote:
> > Under some circumstances, RCU grace periods are zero cost. For
> > RCU-preempt, this is the case during boot, and for RCU-bh and RCU-sched,
> > this is the case if there is only one CPU. This means that RCU users
> > might wish to dispense with grace-period-avoidance strategies when
> > grace periods are zero cost, so this commit adds rcu_trivial_gp(),
> > rcu_bh_trivial_gp(), and rcu_sched_trivial_gp() to test for these
> > conditions. Because the conditions leading to zero-cost grace periods
> > can change at any time (for example, when a second CPU is onlined), these
> > functions should be used as performance hints, and must not be relied
> > on for correctness. For example, even if rcu_trivial_gp() returns true,
> > you are required to invoke synchronize_rcu().
> >
> > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
>
> Do you have anything planned that uses these functions?
>
> Exposing this information, rather than just letting callers call
> synchronize_rcu() and sometimes get "free" grace periods, seems
> potentially error-prone.
>
> If you keep these functions, please expand the comments above them to
> explicitly include the explanation in the commit message, and
> specifically that you must always call the corresponding synchronize
> function even if these functions return true.

Hmmm...

Someone asked for this, I failed to add their Reported-by, and
I do not remember who it was. I have tagged this commit with
trivial_gp.2017.01.17a and am dropping it from my commits for the 4.11
merge window. If someone speaks up for it, I can always add it in later.

The use case was something about doing batching if !rcu_trivial_gp(),
but not bothering if rcu_trivial_gp(), which means that grace periods
are free anyway. So performance, no risk of incorrectness. If used
correctly, anyway...

Thanx, Paul

> > include/linux/rcupdate.h | 8 ++++++++
> > include/linux/rcutiny.h | 10 ++++++++++
> > include/linux/rcutree.h | 2 ++
> > kernel/rcu/tree.c | 24 ++++++++++++++++++++++++
> > kernel/rcu/tree_plugin.h | 11 +++++++++++
> > 5 files changed, 55 insertions(+)
> >
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > index 01f71e1d2e94..a6222478b87d 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -293,6 +293,7 @@ void __rcu_read_lock(void);
> > void __rcu_read_unlock(void);
> > void rcu_read_unlock_special(struct task_struct *t);
> > void synchronize_rcu(void);
> > +bool rcu_trivial_gp(void);
> >
> > /*
> > * Defined as a macro as it is a very low level header included from
> > @@ -448,6 +449,13 @@ bool __rcu_is_watching(void);
> > #define RCU_SCHEDULER_INIT 1
> > #define RCU_SCHEDULER_RUNNING 2
> >
> > +#ifndef CONFIG_PREEMPT_RCU
> > +static inline bool rcu_trivial_gp(void)
> > +{
> > + return rcu_sched_trivial_gp();
> > +}
> > +#endif /* #ifndef CONFIG_PREEMPT_RCU */
> > +
> > /*
> > * init_rcu_head_on_stack()/destroy_rcu_head_on_stack() are needed for dynamic
> > * initialization and destruction of rcu_head on the stack. rcu_head structures
> > diff --git a/include/linux/rcutiny.h b/include/linux/rcutiny.h
> > index ac81e4063b40..a77dafe79813 100644
> > --- a/include/linux/rcutiny.h
> > +++ b/include/linux/rcutiny.h
> > @@ -82,6 +82,16 @@ static inline void synchronize_sched_expedited(void)
> > synchronize_sched();
> > }
> >
> > +static inline bool rcu_sched_trivial_gp(void)
> > +{
> > + return true;
> > +}
> > +
> > +static inline bool rcu_bh_trivial_gp(void)
> > +{
> > + return true;
> > +}
> > +
> > static inline void kfree_call_rcu(struct rcu_head *head,
> > rcu_callback_t func)
> > {
> > diff --git a/include/linux/rcutree.h b/include/linux/rcutree.h
> > index 63a4e4cf40a5..fcd61cb08851 100644
> > --- a/include/linux/rcutree.h
> > +++ b/include/linux/rcutree.h
> > @@ -47,6 +47,8 @@ static inline void rcu_virt_note_context_switch(int cpu)
> > void synchronize_rcu_bh(void);
> > void synchronize_sched_expedited(void);
> > void synchronize_rcu_expedited(void);
> > +bool rcu_sched_trivial_gp(void);
> > +bool rcu_bh_trivial_gp(void);
> >
> > void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func);
> >
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index d7b63b88434b..ed5a17aca281 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -3293,6 +3293,18 @@ void synchronize_sched(void)
> > EXPORT_SYMBOL_GPL(synchronize_sched);
> >
> > /**
> > + * rcu_sched_trivial_gp - Are RCU-sched grace periods trivially zero cost?
> > + *
> > + * Returns true if RCU-sched grace periods are currently zero cost, which
> > + * they are if there is only one CPU. Note that unless you take steps to
> > + * prevent it, the number of CPUs might change at any time.
> > + */
> > +bool rcu_sched_trivial_gp(void)
> > +{
> > + return rcu_blocking_is_gp();
> > +}
> > +
> > +/**
> > * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
> > *
> > * Control will return to the caller some time after a full rcu_bh grace
> > @@ -3320,6 +3332,18 @@ void synchronize_rcu_bh(void)
> > EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
> >
> > /**
> > + * rcu_bh_trivial_gp - Are RCU-bh grace periods trivially zero cost?
> > + *
> > + * Returns true if RCU-bh grace periods are currently zero cost, which
> > + * they are if there is only one CPU. Note that unless you take steps to
> > + * prevent it, the number of CPUs might change at any time.
> > + */
> > +bool rcu_bh_trivial_gp(void)
> > +{
> > + return rcu_blocking_is_gp();
> > +}
> > +
> > +/**
> > * get_state_synchronize_rcu - Snapshot current RCU state
> > *
> > * Returns a cookie that is used by a later call to cond_synchronize_rcu()
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index 56583e764ebf..e92d67a3fad2 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -680,6 +680,17 @@ void synchronize_rcu(void)
> > EXPORT_SYMBOL_GPL(synchronize_rcu);
> >
> > /**
> > + * rcu_trivial_gp - Are RCU grace periods trivially zero cost?
> > + *
> > + * Returns true if RCU grace periods are currently zero cost, which
> > + * they are during boot.
> > + */
> > +bool rcu_trivial_gp(void)
> > +{
> > + return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
> > +}
> > +
> > +/**
> > * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
> > *
> > * Note that this primitive does not necessarily wait for an RCU grace period
> > --
> > 2.5.2
> >
>