Re: [PATCH] random32: Use rcuidle variant for tracepoint

[Peter Zijlstra]

On Fri, Aug 21, 2020 at 11:41:41AM -0400, Steven Rostedt wrote:
> On Fri, 21 Aug 2020 11:38:31 -0400
> Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> 
> > > > At some point we're going to have to introduce noinstr to idle as well.
> > > > But until that time this should indeed cure things.    
> > > 
> 
> What the above means, is that ideally we will get rid of all
> tracepoints and kasan checks from these RCU not watching locations. But

s/and kasan//

We only need to get rid of explicit tracepoints -- typically just move
them outside of the rcu_idle_enter/exit section.

> to do so, we need to move the RCU not watching as close as possible to
> where it doesn't need to be watching, and that is not as trivial of a
> task as one might think.

My recent patch series got a fair way towards that, but yes, there's
more work to be done still.

> Once we get to a minimal code path for RCU not
> to be watching, it will become "noinstr" and tracing and "debugging"
> will be disabled in these sections.

Right, noinstr is like notrace on steriods, not only does it disallow
tracing, it will also disable all the various *SAN/KCOV instrumentation.
It also has objtool based validation which ensures noinstr code doesn't
call out to regular code.