Re: [PATCH 1/2] tracing/function-graph-tracer: signal interrupts entry/exit points on outpout

[Ingo Molnar <mingo@elte.hu>]

* Frédéric Weisbecker <fweisbec@gmail.com> wrote:

> 2008/12/8 Ingo Molnar <mingo@elte.hu>:
> >
> > * Frederic Weisbecker <fweisbec@gmail.com> wrote:
> >
> >> +/* Used during output to find the interrupts entry/exit points */
> >> +const char *ftrace_graph_irq_entries[] = {
> >> +     "smp_call_function_single_interrupt",
> >> +     "xen_call_function_single_interrupt",
> >> +     "wrapper_smp_local_timer_interrupt",
> >> +     "smp_irq_move_cleanup_interrupt",
> >> +     "smp_call_function_interrupt",
> >> +     "xen_call_function_interrupt",
> >> +     "smp_apic_timer_interrupt",
> >> +     "uv_bau_message_interrupt",
> >> +     "mce_threshold_interrupt",
> >> +     "smp_spurious_interrupt",
> >> +     "smp_thermal_interrupt",
> >> +     "smp_error_interrupt",
> >> +     "do_IRQ",
> >> +     NULL
> >
> > hm, couldnt we move these symbols to a separate section, and then only
> > check for [section.start ... section.end] instead of this ugly and slow
> > array?
> >
> > Missing a few annotations initially is no big deal - we wont have
> > pretty-print. do_IRQ() and smp_apic_timer_interrupt is what matters most
> > in practice. "__irqentry" section annotation or so, which puts them into
> > .text.irqentry or so - and then irqentry_start/end are extracted via
> > appropriate glue in the arch/x86/kernel/vmlinux* linker script.
> 
> 
> I found it a bit ugly too while I wrote it :-( I like this idea of a 
> section, I will just have to verify if it is between the start and the 
> end of it to check if its an irq entry. But I think that even if the 
> others than do_IRQ and smp_apic_timer_interrupt are more rare, they 
> should be annotated for this new section (and it seems there are new 
> coming interrupts like smp_perf_counter_interrupt() in perf 
> counter....)... Perhaps someone would profile them...
> 
> Hm?

not annotating an IRQ entry is not a big problem in practice: people will 
notice them in traces that they are not annotated, will send a feature 
request, we annotate them. If they are not noticed, it means they are 
rare in one way or another, and dont matter in practice. So this is a 
self-maintaining concept.

	Ingo