Re: [PATCH 0/1] x86: Eliminate bp argument from the stack tracing routines

[Soeren Sandmann <sandmann@daimi.au.dk>]

Frederic Weisbecker <fweisbec@gmail.com> writes:

> > (FWIW, this
> > 
> > diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
> > index 461a85d..d977d26 100644
> > --- a/arch/x86/kernel/cpu/perf_event.c
> > +++ b/arch/x86/kernel/cpu/perf_event.c
> > @@ -1653,7 +1653,7 @@ static const struct stacktrace_ops backtrace_ops = {
> >  	.warning_symbol		= backtrace_warning_symbol,
> >  	.stack			= backtrace_stack,
> >  	.address		= backtrace_address,
> > -	.walk_stack		= print_context_stack_bp,
> > +	.walk_stack		= print_context_stack,
> >  };
> > 
> > makes it produce correct kernel callchains. And yes, I did compile the
> > kernel with CONFIG_FRAME_POINTER).
> 
> 
> 
> What do you see is broken in 64 bits perf callchains? Can you please provide
> me more details so that I can fix the issue?

Apparently, the problem only happens when using the hrtimer based
events. I don't think there is a way to make perf use those from the
command line, but if you apply this:

--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -294,6 +294,12 @@ static void create_counter(int counter, int cpu)
                attr->enable_on_exec = 1;
        }
 
+       if (attr->config == PERF_COUNT_HW_CPU_CYCLES && attr->type ==
PERF_TYPE_HARDWARE)
+       {
+           attr->type = PERF_TYPE_SOFTWARE;
+           attr->config = PERF_COUNT_SW_CPU_CLOCK;
+       }
+       
        for (thread_index = 0; thread_index < thread_num; thread_index++) {
 try_again:
                fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,

you should be able to see the problem.  You can also try sysprof; it
always uses the software counters.


Soren