On Thu, Feb 13, 2020 at 09:27:16AM +0100, Peter Zijlstra wrote:

On Wed, Feb 12, 2020 at 06:20:05PM -0500, Joel Fernandes wrote:

quoted

On Wed, Feb 12, 2020 at 10:01:42PM +0100, Peter Zijlstra wrote:

quoted

quoted

+#define trace_rcu_enter()					\
+({								\
+	unsigned long state = 0;				\
+	if (!rcu_is_watching())	{				\
+		if (in_nmi()) {					\
+			state = __TR_NMI;			\
+			rcu_nmi_enter();			\
+		} else {					\
+			state = __TR_IRQ;			\
+			rcu_irq_enter_irqsave();		\

I think this can be simplified. You don't need to rely on in_nmi() here. I
believe for NMI's, you can just call rcu_irq_enter_irqsave() and that should
be sufficient to get RCU watching. Paul can correct me if I'm wrong, but I am
pretty sure that would work.

In fact, I think a better naming for rcu_irq_enter_irqsave() pair could be
(in the first patch):

rcu_ensure_watching_begin();
rcu_ensure_watching_end();

So I hadn't looked deeply into rcu_irq_enter(), it seems to call
rcu_nmi_enter_common(), but with @irq=true.

What exactly is the purpose of that @irq argument, and how much will it
hurt to lie there? Will it come apart if we have @irq != !in_nmi()
for example?

There is a comment in there that says ->dynticks_nmi_nesting ought to be
odd only if we're in NMI. The only place that seems to care is
rcu_nmi_exit_common(), and that does indeed do something different for
IRQs vs NMIs.

So I don't think we can blindly unify this. But perhaps Paul sees a way?

The reason for the irq argument is to avoid invoking
rcu_prepare_for_idle() and rcu_dynticks_task_enter() from NMI context
from rcu_nmi_exit_common().  Similarly, we need to avoid invoking
rcu_dynticks_task_exit() and rcu_cleanup_after_idle() from NMI context
from rcu_nmi_enter_common().

It might well be that I could make these functions be NMI-safe, but
rcu_prepare_for_idle() in particular would be a bit ugly at best.
So, before looking into that, I have a question.  Given these proposed
changes, will rcu_nmi_exit_common() and rcu_nmi_enter_common() be able
to just use in_nmi()?

							Thanx, Paul