Re: [PATCH 1/2] tracing,mm - add kernel pagefault tracepoint for x86 & x86_64

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Nov 10, 2010 at 04:17:50PM +0100, Jiri Olsa wrote:
> On Wed, Nov 10, 2010 at 04:00:25PM +0100, Frederic Weisbecker wrote:
> > On Wed, Nov 10, 2010 at 02:52:44PM +0100, Ingo Molnar wrote:
> > > 
> > > * Jiri Olsa <jolsa@redhat.com> wrote:
> > > 
> > > > On Wed, Nov 10, 2010 at 08:29:54AM -0500, Christoph Hellwig wrote:
> > > > > On Wed, Nov 10, 2010 at 12:56:11PM +0100, Jiri Olsa wrote:
> > > > > > +	TP_printk("task=%lx, address=%lx, regs=%lx",
> > > > > > +		(unsigned long)__entry->task, (unsigned long)__entry->address,
> > > > > > +			__entry->regs)
> > > > > 
> > > > > How exactly do you use the information in this trace point?  Especially
> > > > > the undecoded pt_regs doesn't seem very useful to me at all.
> > > > 
> > > > agreed, the registers pointer are not very useful in the trace file output,
> > > > and could be taken away.. just wanted to be complete I guess
> > > > 
> > > > but I believe they are useful when you register the mm_kernel_pagefault
> > > > tracepoint and process the information by yourself
> > > 
> > > That would be expressed in a better and more generic fashion via adding 
> > > PERF_SAMPLE_REGS to perf_event_sample_format, and add a ptregs dump in 
> > > kernel/perf_event.c, perf_output_sample(). That way any tracepoint can request a 
> > > (user-space)ptregs state snapshot, not just the pagefault ones.
> > > 
> > > Thanks,
> > > 
> > > 	Ingo
> > 
> > 
> > We are going to have that with the dwarf based callchain patchset. I'm cooking
> > this.
> > 
> > Thanks.
> > 
> 
> I guess I can take the regs out then.. would that patch be acceptable
> afterwards..?


Please take out the regs yeah.

Thanks.