Re: [PATCH 1/2] perf, x86-mm: Add exit-fault tracing

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Wed, Oct 26, 2016 at 10:51:16AM +0200, Peter Zijlstra escreveu:
> On Wed, Oct 26, 2016 at 01:51:59AM +0200, Alexis Berlemont wrote:
> 
> -ENOCHANGELOG

Yeah, please add one explaining why we want this, how we can use it, for
example, in 'perf trace', to measure how long each page fault took,
examples of it in use, etc.
 
> > Signed-off-by: Alexis Berlemont <alexis.berlemont@gmail.com>
> > ---
> 
> > diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
> > index 9f72ca3..e31e8ef 100644
> > --- a/arch/x86/mm/fault.c
> > +++ b/arch/x86/mm/fault.c
> > @@ -1488,6 +1488,7 @@ trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
> >  	prev_state = exception_enter();
> >  	trace_page_fault_entries(address, regs, error_code);
> >  	__do_page_fault(regs, error_code, address);
> > +	trace_page_fault_exit(address);
> 
> Aside from my general hatred of tracepoint, it bugs me that its not
> symmetric like the irq vector ones. But I'll leave that to x86 people.

What is the simmetry problem, you think we should have:

  exceptions:page_fault_kernel_exit
  exceptions:page_fault_user_exit

To be counterparts of:

[root@jouet ~]# perf list exceptions:*

List of pre-defined events (to be used in -e):

  exceptions:page_fault_kernel                       [Tracepoint event]
  exceptions:page_fault_user                         [Tracepoint event]

[root@jouet ~]# 

? 

For perf usage yeah, the good thing would be to have just
exceptions:page_fault_entry and exceptions:page_fault_exit, and then use
perf_event_attr fields to filter what kind was desired, with a
header.misc stating where it took place, just like with other events,
but by now we have two for entry, so two for exit?

- Arnaldo

> >  	exception_exit(prev_state);
> >  }
> >  NOKPROBE_SYMBOL(trace_do_page_fault);
> > -- 
> > 2.10.1
> >