On Tue, Feb 18, 2020 at 12:17:28PM -0800, Paul E. McKenney wrote:

On Tue, Feb 18, 2020 at 08:58:31PM +0100, Peter Zijlstra wrote:

quoted

On Thu, Feb 13, 2020 at 03:44:44PM -0500, Joel Fernandes wrote:

quoted

quoted

quoted

That _should_ already be the case today. That is, if we end up in a
tracer and in_nmi() is unreliable we're already screwed anyway.

quoted

I removed the static from rcu_nmi_enter()/exit() as it is called from
outside, that makes it build now. Updated below is Paul's diff. I also added
NOKPROBE_SYMBOL() to rcu_nmi_exit() to match rcu_nmi_enter() since it seemed
asymmetric.

quoted

+__always_inline void rcu_nmi_exit(void)
 {
 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 

@@ -651,25 +653,15 @@ static __always_inline void rcu_nmi_exit_common(bool irq)
 	trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
 	WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
 
-	if (irq)
+	if (!in_nmi())
 		rcu_prepare_for_idle();
 
 	rcu_dynticks_eqs_enter();
 
-	if (irq)
+	if (!in_nmi())
 		rcu_dynticks_task_enter();
 }

Boris and me have been going over the #MC code (and finding loads of
'interesting' code) and ran into ist_enter(), whish has the following
code:

                /*
                 * We might have interrupted pretty much anything.  In
                 * fact, if we're a machine check, we can even interrupt
                 * NMI processing.  We don't want in_nmi() to return true,
                 * but we need to notify RCU.
                 */
                rcu_nmi_enter();


Which, to me, sounds all sorts of broken. The IST (be it #DB or #MC) can
happen while we're holding all sorts of locks. This must be an NMI-like
context.

Ouch!  Looks like I need to hold off on getting rid of the "irq"
parameters if in_nmi() isn't going to be accurate.

I'm currently trying to twist my brain around all this, because I
suspect it's all completely broken one way or another.

But yes, we definitely need to fix this before your patch goes in.