Re: [PATCH 02/10] perf, trace: Use per-tracepoint-per-cpu hlist to track events

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, May 21, 2010 at 12:02:05PM +0200, Peter Zijlstra wrote:
> On Fri, 2010-05-21 at 11:40 +0200, Frederic Weisbecker wrote:
> > On Fri, May 21, 2010 at 11:02:03AM +0200, Peter Zijlstra wrote:
> 
> > > Also, avoid conditionals on the fast path by ordering with probe unregister
> > > so that we should never get on the callback path without the data being there.
> > > 
> > 									\
> > > +	head = per_cpu_ptr(event_call->perf_events, smp_processor_id());\
> 
> > Should be rcu_dereference_sched ?
> 
> No, I removed all that rcu stuff and synchronized against the probe
> unregister.
> 
> I assumed that after probe unregister a tracepoint callback doesn't
> happen, which then guarantees we should never get !head.



I'm not sure about this. The tracepoints are called under rcu_read_lock(),
but there is not synchronize_rcu() after we unregister a tracepoint, which
means you can have a pending preempted one somewhere.

There is a call_rcu that removes the callbacks, but that only protect
the callback themselves.



> 
> > > +	for_each_possible_cpu(cpu)
> > > +		INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
> > > +
> > > +	tp_event->perf_events = list;
> > 
> > 
> > 
> > I suspect this must be rcu_assign_pointer.
> 
> Same thing as above, I do this before probe register, so I see no need
> for RCU.
> 
> > > +	list = per_cpu_ptr(list, smp_processor_id());
> > > +	hlist_add_head_rcu(&p_event->hlist_entry, list);
> > 
> > 
> > 
> > Ah and may be small comment, because using the hlist api here
> > may puzzle more people than just me ;)
> 
> What exactly is the puzzlement about?



The fact we use the hlist API not for hlist purpose but for a list.




> > > +	if (--tp_event->perf_refcount > 0)
> > > +		return;
> > > +
> > > +	tp_event->perf_event_disable(tp_event);
> > 
> > 
> > 
> > Don't we need a rcu_synchronize_sched() here?
> 
> Doesn't probe unregister synchronize things against its own callback?



May be I missed it but it doesn't seem so.



> > > +	raw_data = per_cpu_ptr(perf_trace_buf[*rctxp], smp_processor_id());
> > 
> > 
> > 
> > Needs rcu_dereference_sched too. And this could be __this_cpu_var()
> 
> Ahh! so that is what its called.


:)


 
> > > +	preempt_disable_notrace();
> > 
> > 
> > Why is this needed. We have the recursion context protection already.
> 
> Because:
> 
> @@ -4094,7 +4087,7 @@ end:
>  
>  int perf_swevent_get_recursion_context(void)
>  {
> -       struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
> +       struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
>         int rctx;
>  
>         if (in_nmi())
> 



Right.

Thanks.