Re: [PATCH] tracing/perf: Move rcu_irq_enter/exit_irqson() to perf trace point hook

[Peter Zijlstra]

On Tue, Feb 11, 2020 at 04:05:32PM +0100, Peter Zijlstra wrote:

> So we haz:
> 
> | #define nmi_enter()						\
> | 	do {							\
> | 		arch_nmi_enter();				\
> 
> arm64 only, lets ignore for now
> 
> | 		printk_nmi_enter();				\
> 
> notrace
> 
> | 		lockdep_off();					\
> 
> notrace
> 
> | 		ftrace_nmi_enter();				\
> 
> !notrace !!!
> 
> | 		BUG_ON(in_nmi());				\
> | 		preempt_count_add(NMI_OFFSET + HARDIRQ_OFFSET);\
> 
> lets make this __preempt_count_add() ASAP !

preempt_count_add() first frobs the actual preempt_count and then does
the trace, so that might just work. But it does need a notrace
annotation, I'm thinking, because calling into the function tracer
_before_ we do the preempt_count increment is irrecoverable crap.

> | 		rcu_nmi_enter();				\
> 
> are you _really_ sure you want to go trace that ?!?
> 
> | 		trace_hardirq_enter();				\
> | 	} while (0)
> 
>