Re: [PATCH v2 03/11] sched,idle,rcu: Push rcu_idle deeper into the idle path

[Joel Fernandes]

On Thu, Aug 27, 2020 at 09:47:48AM +0200, peterz@xxxxxxxxxxxxx wrote:
> On Wed, Aug 26, 2020 at 09:24:19PM -0400, Joel Fernandes wrote:
> > On Wed, Aug 26, 2020 at 09:18:26PM -0400, Joel Fernandes wrote:
> > > On Fri, Aug 21, 2020 at 10:47:41AM +0200, Peter Zijlstra wrote:
> > > > Lots of things take locks, due to a wee bug, rcu_lockdep didn't notice
> > > > that the locking tracepoints were using RCU.
> > > > 
> > > > Push rcu_idle_{enter,exit}() as deep as possible into the idle paths,
> > > > this also resolves a lot of _rcuidle()/RCU_NONIDLE() usage.
> > > > 
> > > > Specifically, sched_clock_idle_wakeup_event() will use ktime which
> > > > will use seqlocks which will tickle lockdep, and
> > > > stop_critical_timings() uses lock.
> > > 
> > > I was wondering if those tracepoints should just use _rcuidle variant of the
> > > trace call. But that's a terrible idea considering that would add unwanted
> > > overhead I think.
> > > 
> > > Reviewed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > 
> > BTW, if tracepoint is converted to use RCU-trace flavor, then these kinds of
> > issues go away, no? That RCU flavor is always watching.
> 
> All trace_*_rcuidle() and RCU_NONIDLE() usage is a bug IMO.
> 
> Ideally RCU-trace goes away too.

I was thinking that unless the rcu_idle_enter/exit calls coincide with points
in the kernel where instrumentation is not allowed, there is always a chance
somebody wants to use tracepoints after rcu_idle_enter or before exit. In
this case, trace_*_rcuidle() is unavoidable unless you can move the
rcu_idle_enter/exit calls deeper as you are doing.

Maybe objtool can help with that?

The other solution is RCU-trace if you can't push the rcu_idle_enter/exit
calls any deeper and still want to get rid of trace_*_rcuidle thingies. Me
and Mathieu were talking at LPC about tracepoint conversion to RCU-trace and
we can work on it if that's the right direction.

thanks,

 - Joel