Re: [PATCH 5/5 v2] rcu: Fix dyntick-idle tracing

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 7 Apr 2017 07:40:11 -0700
"Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:

> On Fri, Apr 07, 2017 at 10:01:11AM -0400, Steven Rostedt wrote:
> > From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
> > 
> > The tracing subsystem started using rcu_irq_entry() and rcu_irq_exit()
> > (with my blessing) to allow the current _rcuidle alternative tracepoint
> > name to be dispensed with while still maintaining good performance.
> > Unfortunately, this causes RCU's dyntick-idle entry code's tracing to
> > appear to RCU like an interrupt that occurs where RCU is not designed
> > to handle interrupts.
> > 
> > This commit fixes this problem by moving the zeroing of ->dynticks_nesting
> > after the offending trace_rcu_dyntick() statement, which narrows the
> > window of vulnerability to a pair of adjacent statements that are now
> > marked with comments to that effect.
> > 
> > Link: http://lkml.kernel.org/r/20170405193928.GM1600@linux.vnet.ibm.com
> > 
> > Reported-by: Steven Rostedt <rostedt@goodmis.org>
> > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
> > ---
> >  kernel/rcu/tree.c | 48 +++++++++++++++++++++++-------------------------
> >  1 file changed, 23 insertions(+), 25 deletions(-)
> > 
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 50fee7689e71..8b4d273331e4 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -57,6 +57,7 @@
> >  #include <linux/random.h>
> >  #include <linux/trace_events.h>
> >  #include <linux/suspend.h>
> > +#include <linux/ftrace.h>
> > 
> >  #include "tree.h"
> >  #include "rcu.h"
> > @@ -771,25 +772,24 @@ cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
> >  }
> > 
> >  /*
> > - * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
> > + * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
> >   *
> > - * If the new value of the ->dynticks_nesting counter now is zero,
> > - * we really have entered idle, and must do the appropriate accounting.
> > - * The caller must have disabled interrupts.
> > + * Enter idle, doing appropriate accounting.  The caller must have
> > + * disabled interrupts.
> >   */
> > -static void rcu_eqs_enter_common(long long oldval, bool user)
> > +static void rcu_eqs_enter_common(bool user)
> >  {
> >  	struct rcu_state *rsp;
> >  	struct rcu_data *rdp;
> > -	RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
> > +	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
> > 
> > -	trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
> > +	trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
> >  	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
> >  	    !user && !is_idle_task(current)) {
> >  		struct task_struct *idle __maybe_unused =
> >  			idle_task(smp_processor_id());
> > 
> > -		trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
> > +		trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
> >  		rcu_ftrace_dump(DUMP_ORIG);
> >  		WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
> >  			  current->pid, current->comm,
> > @@ -800,7 +800,10 @@ static void rcu_eqs_enter_common(long long oldval, bool user)
> >  		do_nocb_deferred_wakeup(rdp);
> >  	}
> >  	rcu_prepare_for_idle();
> > -	rcu_dynticks_eqs_enter();
> > +	stack_tracer_disable();
> > +	rdtp->dynticks_nesting = 0; /* Breaks tracing momentarily. */
> > +	rcu_dynticks_eqs_enter(); /* After this, tracing works again. */
> > +	stack_tracer_enable();  
> 
> Hmmm...  There is not supposed to be any tracing in this interval,

Why not? function tracing happens without an issue. But then again,
function tracing doesn't depend on RCU.

> and interrupts are disabled.  Wouldn't it be better to have something
> that made tracing illegal during this interval?

I don't see an issue here. Function tracing is fine. There should be no
trace_events() as those are static events and shouldn't dynamically
appear in this interval.

The problem I hit is that stack tracing uses function tracing to check
the stack of all functions. It doesn't need RCU either, unless it hits
a new "max stack", which it then calls save_stack_trace(), which does a
lot, and it does perform an rcu_read_lock(), which is what broke.

I'm fine with tracing, as that works. What doesn't work is tracing a
new max stack.

> 
> Yeah, I am a bit concerned about idle-entry latency...
> 

Which should now be fine because of the inlined this_cpu_inc/dec()
which is very efficient and made for fast paths like this.

-- Steve