Re: [PATCH] rcu: Avoid unneeded function call in rcu_read_unlock()

["Paul E. McKenney" <paulmck@kernel.org>]

On Mon, Aug 30, 2021 at 11:36:51AM -0700, Andrii Nakryiko wrote:
> On Fri, Aug 27, 2021 at 11:34 AM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Thu, Aug 26, 2021 at 10:21:22PM -0400, Waiman Long wrote:
> > > Since commit aa40c138cc8f3 ("rcu: Report QS for outermost
> > > PREEMPT=n rcu_read_unlock() for strict GPs"). A real function call
> > > rcu_read_unlock_strict() is added to the inlined rcu_read_unlock().
> > > The rcu_read_unlock_strict() call is only needed if the performance
> > > sagging CONFIG_RCU_STRICT_GRACE_PERIOD option is set. This config
> > > option isn't set for most production kernels while the function call
> > > overhead remains.
> > >
> > > To provide a slight performance improvement, the
> > > CONFIG_RCU_STRICT_GRACE_PERIOD config check is moved from
> > > rcu_read_unlock_strict() to __rcu_read_unlock() so that the function
> > > call can be compiled out in most cases.
> > >
> > > Besides, the GPL exported rcu_read_unlock_strict() also impact the
> > > the compilation of non-GPL kernel modules as rcu_read_unlock() is a
> > > frequently used kernel API.
> > >
> > > Signed-off-by: Waiman Long <longman@redhat.com>
> >
> > Nice, and good eyes!!!
> >
> > I have queued this for v5.16, that is, not the upcoming merge window
> > but the one after that.
> >
> > I did my usual wordsmithing, so please check the following in case I
> > messed something up.  I intentionally omitted the EXPORT_SYMBOL_GPL()
> > discussion because:
> >
> > 1.      Kernels built with CONFIG_PREEMPT=y have the same issue
> >         with the __rcu_read_lock() and __rcu_read_unlock() functions.
> >
> > 2.      Many other RCU functions are EXPORT_SYMBOL_GPL() and have
> >         been for almost two decades.
> >
> > But if someone does use RCU readers within CONFIG_PREEMPT=n kernels from
> > a binary module, I will happily refer them to you for any RCU issues
> > that they encounter.  ;-)
> >
> > I am also CCing the BPF guys in case my interpretation of the code in
> > the BPF verifier is incorrect.
> >
> 
> LGTM from the BPF side, nothing really changed about when
> rcu_read_unlock_strict is an actual function vs no-op macro. It's also
> important to minimize the number of function calls in the context of
> recent LBR on-demand work done by Song, so this is a great
> improvement!

Thank you for looking this over!  May I add your Acked-by or similar?

                                                         Thanx, Paul

> > ------------------------------------------------------------------------
> >
> > commit 4a9f53b997b809c0256838e31c604aeeded2345a
> > Author: Waiman Long <longman@redhat.com>
> > Date:   Thu Aug 26 22:21:22 2021 -0400
> >
> >     rcu: Avoid unneeded function call in rcu_read_unlock()
> >
> >     Since commit aa40c138cc8f3 ("rcu: Report QS for outermost PREEMPT=n
> >     rcu_read_unlock() for strict GPs") the function rcu_read_unlock_strict()
> >     is invoked by the inlined rcu_read_unlock() function.  However,
> >     rcu_read_unlock_strict() is an empty function in production kernels,
> >     which are built with CONFIG_RCU_STRICT_GRACE_PERIOD=n.
> >
> >     There is a mention of rcu_read_unlock_strict() in the BPF verifier,
> >     but this is in a deny-list, meaning that BPF does not care whether
> >     rcu_read_unlock_strict() is ever called.
> >
> >     This commit therefore provides a slight performance improvement
> >     by hoisting the check of CONFIG_RCU_STRICT_GRACE_PERIOD from
> >     rcu_read_unlock_strict() into rcu_read_unlock(), thus avoiding the
> >     pointless call to an empty function.
> >
> >     Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com>
> >     Cc: Andrii Nakryiko <andrii@kernel.org>
> >     Signed-off-by: Waiman Long <longman@redhat.com>
> >     Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> >
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > index 434d12fe2d4f..5e0beb5c5659 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -71,7 +71,8 @@ static inline void __rcu_read_lock(void)
> >  static inline void __rcu_read_unlock(void)
> >  {
> >         preempt_enable();
> > -       rcu_read_unlock_strict();
> > +       if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
> > +               rcu_read_unlock_strict();
> >  }
> >
> >  static inline int rcu_preempt_depth(void)
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index 7a4876a3a882..0b55c647ab80 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -814,8 +814,7 @@ void rcu_read_unlock_strict(void)
> >  {
> >         struct rcu_data *rdp;
> >
> > -       if (!IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
> > -          irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
> > +       if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
> >                 return;
> >         rdp = this_cpu_ptr(&rcu_data);
> >         rcu_report_qs_rdp(rdp);