Re: [PATCH RFC] tracepoint: Introduce tracepoint callbacks executing with preempt on

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 27 Apr 2018 11:42:15 -0400 (EDT)
Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:

> ----- On Apr 27, 2018, at 10:47 AM, rostedt rostedt@goodmis.org wrote:
> 
> > On Fri, 27 Apr 2018 10:26:29 -0400 (EDT)
> > Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> >   
> >> The general approach and the implementation look fine, except for
> >> one small detail: I would be tempted to explicitly disable preemption
> >> around the call to the tracepoint callback for the rcuidle variant,
> >> unless we plan to audit every tracer right away to remove any assumption
> >> that preemption is disabled in the callback implementation.  
> > 
> > I'm thinking that we do that audit. There shouldn't be many instances
> > of it. I like the idea that a tracepoint callback gets called with
> > preemption enabled.  
> 
> I see that ftrace explicitly disables preemption in its ring buffer
> code. FWIW, this is redundant when called from sched-rcu tracepoints
> and from kprobes which adds unnecessary performance overhead.

Sure, but that code is called from other locations that do not have
preemption disabled. Calling preempt_disable() is far from the biggest
overhead of that code path.

> 
> LTTng expects preemption to be disabled when invoked. I can adapt on my
> side as needed, but would prefer not to have redundant preemption disabling
> for probes hooking on sched-rcu tracepoints (which is the common case).

Why not? Really, preempt_disable is simply a per cpu counter, with only
need of adding compiler barriers.

> 
> Do perf callbacks expect preemption to be disabled ?

I'll have to look, but wouldn't be hard to change.

-- Steve