Re: [PATCH] rcu: mark debug_lockdep_rcu_enabled() as pure

[Konstantin Khlebnikov]

On 20.05.2017 19:42, Paul E. McKenney wrote:
> On Fri, May 19, 2017 at 10:03:59AM +0300, Konstantin Khlebnikov wrote:
> > This allows to get rid of unneeded invocations.
> >
> > Function debug_lockdep_rcu_enabled() becomes really hot if several
> > debug options are enabled together with CONFIG_PROVE_RCU.
> >
> > Hottest path ends with:
> >    debug_lockdep_rcu_enabled
> >    is_ftrace_trampoline
> >    __kernel_text_address
> >
> > Here debug_lockdep_rcu_enabled() is called from condition
> > (debug_lockdep_rcu_enabled() && !__warned && (c)) inside macro
> > do_for_each_ftrace_op(), where "c" is false.
> >
> > With this patch "netperf -H localhost" shows boost from 2400 to 2500.
> >
> > Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
>
> Nice performance increase!
>
> The gcc documentation says that __attribute__((pure)) functions are
> supposed to have return values that depend only at the function's
> arguments and the values of global variables.  However, it also says:
>
> 	Interesting non-pure functions are functions with infinite loops
> 	or those depending on volatile memory or other system resource,
> 	that may change between two consecutive calls (such as feof in
> 	a multithreading environment).
>
> This is OK for current->lockdep_recursion because this variable is changed
> only by the current task (I think so, anyway).
>
> It is sort of OK for debug_locks.  This could be set to zero at any time
> by any other task, but if we have a race condition that very rarely causes
> two lockdep splats instead of just one, so what?  (But I am sure that
> some of the people on CC will correct me if I am wrong here.)
>
> It should be OK for rcu_scheduler_active because the transition from
> RCU_SCHEDULER_INACTIVE to RCU_SCHEDULER_INIT happens before the first
> context switch, and the various barrier() calls, implied as well as
> explicit, should keep things straight.
>
> But I don't totally trust my analysis.  Could you please get someone
> more gcc-savvy to review this and give their ack/review?  Given that,
> I will queue it.
>
> 							Thanx, Paul

Thank you for analisys.

__attribute__((pure)) allows compiler to deduplicate / eliminate calls.
This indeed might expand race windows when global switches like debug_locks
changes their state. But strict synchronization here isn't required.

> > ---
> >   include/linux/rcupdate.h |    2 +-
> >   kernel/rcu/update.c      |    2 +-
> >   2 files changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > index e1e5d002fdb9..9ecb3cb715bd 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -472,7 +472,7 @@ extern struct lockdep_map rcu_lock_map;
> >   extern struct lockdep_map rcu_bh_lock_map;
> >   extern struct lockdep_map rcu_sched_lock_map;
> >   extern struct lockdep_map rcu_callback_map;
> > -int debug_lockdep_rcu_enabled(void);
> > +int __pure debug_lockdep_rcu_enabled(void);
> >
> >   int rcu_read_lock_held(void);
> >   int rcu_read_lock_bh_held(void);
> > diff --git a/kernel/rcu/update.c b/kernel/rcu/update.c
> > index 273e869ca21d..a0c30abefdcd 100644
> > --- a/kernel/rcu/update.c
> > +++ b/kernel/rcu/update.c
> > @@ -292,7 +292,7 @@ struct lockdep_map rcu_callback_map =
> >   	STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
> >   EXPORT_SYMBOL_GPL(rcu_callback_map);
> >
> > -int notrace debug_lockdep_rcu_enabled(void)
> > +int __pure notrace debug_lockdep_rcu_enabled(void)
> >   {
> >   	return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
> >   	       current->lockdep_recursion == 0;