On Mon, May 07, 2018 at 02:08:01PM -0700, Paul E. McKenney wrote: > On Mon, May 07, 2018 at 01:41:42PM -0700, Joel Fernandes wrote: > > From: "Joel Fernandes (Google)" <[email protected]> > > > > In recent tests with IRQ on/off tracepoints, a large performance > > overhead ~10% is noticed when running hackbench. This is root caused to > > calls to rcu_irq_enter_irqson and rcu_irq_exit_irqson from the > > tracepoint code. Following a long discussion on the list [1] about this, > > we concluded that srcu is a better alternative for use during rcu idle. > > Although it does involve extra barriers, its lighter than the sched-rcu > > version which has to do additional RCU calls to notify RCU idle about > > entry into RCU sections. > > > > In this patch, we change the underlying implementation of the > > trace_*_rcuidle API to use SRCU. This has shown to improve performance > > alot for the high frequency irq enable/disable tracepoints. [...] >> Signed-off-by: Joel Fernandes (Google) <[email protected]> > > --- > > include/linux/tracepoint.h | 46 +++++++++++++++++++++++++++++++------- > > kernel/tracepoint.c | 15 ++++++++++++- > > 2 files changed, 52 insertions(+), 9 deletions(-) > > > > diff --git a/include/linux/tracepoint.h b/include/linux/tracepoint.h > > index c94f466d57ef..f56f290cf8eb 100644 > > --- a/include/linux/tracepoint.h > > +++ b/include/linux/tracepoint.h > > @@ -15,6 +15,7 @@ > > */ > > > > #include <linux/smp.h> > > +#include <linux/srcu.h> > > #include <linux/errno.h> > > #include <linux/types.h> > > #include <linux/cpumask.h> > > @@ -33,6 +34,8 @@ struct trace_eval_map { > > > > #define TRACEPOINT_DEFAULT_PRIO 10 > > > > +extern struct srcu_struct tracepoint_srcu; > > + > > extern int > > tracepoint_probe_register(struct tracepoint *tp, void *probe, void *data); > > extern int > > @@ -77,6 +80,9 @@ int unregister_tracepoint_module_notifier(struct > > notifier_block *nb) > > */ > > static inline void tracepoint_synchronize_unregister(void) > > { > > +#ifdef CONFIG_TRACEPOINTS > > + synchronize_srcu(&tracepoint_srcu); > > +#endif > > synchronize_sched(); > > } > > > > @@ -129,18 +135,38 @@ extern void syscall_unregfunc(void); > > * as "(void *, void)". The DECLARE_TRACE_NOARGS() will pass in just > > * "void *data", where as the DECLARE_TRACE() will pass in "void *data, > > proto". > > */ > > -#define __DO_TRACE(tp, proto, args, cond, rcucheck) > > \ > > +#define __DO_TRACE(tp, proto, args, cond, rcuidle) \ > > do { \ > > struct tracepoint_func *it_func_ptr; \ > > void *it_func; \ > > void *__data; \ > > + int __maybe_unused idx = 0; \ > > \ > > if (!(cond)) \ > > return; \ > > - if (rcucheck) \ > > - rcu_irq_enter_irqson(); \ > > - rcu_read_lock_sched_notrace(); \ > > - it_func_ptr = rcu_dereference_sched((tp)->funcs); \ > > + \ > > + /* \ > > + * For rcuidle callers, use srcu since sched-rcu \ > > + * doesn't work from the idle path. \ > > + */ \ > > + if (rcuidle) { \ > > + if (in_nmi()) { \ > > + WARN_ON_ONCE(1); \ > > + return; /* no srcu from nmi */ \ > > + } \ > > + \ > > + idx = srcu_read_lock_notrace(&tracepoint_srcu); \ > > + it_func_ptr = \ > > + srcu_dereference_notrace((tp)->funcs, \ > > + &tracepoint_srcu); \ > > + /* To keep it consistent with !rcuidle path */ \ > > + preempt_disable_notrace(); \ > > + } else { \ > > + rcu_read_lock_sched_notrace(); \ > > + it_func_ptr = \ > > + rcu_dereference_sched((tp)->funcs); \ > > + } \ > > + \ > > if (it_func_ptr) { \ > > do { \ > > it_func = (it_func_ptr)->func; \ > > @@ -148,9 +174,13 @@ extern void syscall_unregfunc(void); > > ((void(*)(proto))(it_func))(args); \ > > } while ((++it_func_ptr)->func); \ > > } \ > > - rcu_read_unlock_sched_notrace(); \ > > - if (rcucheck) \ > > - rcu_irq_exit_irqson(); \ > > + \ > > + if (rcuidle) { \ > > Don't we also need an in_nmi() check here in order to avoid unbalanced > srcu_read_unlock_notrace() calls?
The in_nmi() in the lock path should take care of making sure its balanced. The diff the way its formatted appears confusing as Mathieu pointed. thanks, - Joel
