Re: [ANNOUNCE] New utility: 'trace'

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Nov 17, 2010 at 02:36:16PM +0100, Peter Zijlstra wrote:
> On Wed, 2010-11-17 at 14:10 +0100, Ingo Molnar wrote:
> > * Peter Zijlstra <peterz@infradead.org> wrote:
> > 
> > > On Wed, 2010-11-17 at 13:53 +0100, Frederic Weisbecker wrote:
> > > > On Wed, Nov 17, 2010 at 12:35:50PM +0100, Peter Zijlstra wrote:
> > > > > On Wed, 2010-11-17 at 09:30 +0100, Ingo Molnar wrote:
> > > > > > > For example I'm currently working with dozens of trace_printk() and I would be 
> > > > > > > very happy to turn some of them off half of the time.
> > > > > > 
> > > > > > I guess we could try such a patch. If you send a prototype i'd be interested in 
> > > > > > testing it out.
> > > > > 
> > > > > I don't see the point, the kernel shouldn't contain any trace_printk()s
> > > > > to begin with..
> > > > 
> > > > 
> > > > It's oriented toward developers. Those who use dozens of tracepoints in
> > > > their tree because they are debugging something or developing a new feature,
> > > > they might to deactivate/reactivate some of these independant points.
> > > > 
> > > > This can also apply to dynamic_printk of course.
> > > > 
> > > > Well, the very first and main point is to standardize trace_printk into
> > > > a trace event so that it gets usable by perf tools. I have been asked many
> > > > times "how to use trace_printk() with perf?".
> > > 
> > > Thing is, since its these dev who add the trace_printk()s to begin with, I don't 
> > > see the point in splitting them out, if you didn't want them why did you add them 
> > > to begin with?!
> > 
> > That's a common workflow: lots of printks (trace_printk's) put all around the code - 
> > and sometimes one set of tracepoints is needed, one time another set.
> > 
> > _If_ we succeed in presenting them like Frederic suggested it, and if we make the 
> > turning on/off _simpler_ (no kernel modification) and faster (no kernel reboot) via 
> > the tooling, people like Frederic might start using it.
> > 
> > I dont think we should fight the workflow itself - it makes sense.
> > 
> > The only question is whether we can represent it all in a nicer fashion than 'modify 
> > the source code and reboot'. If we cannot then there's no point - but i'm not sure 
> > about it and Frederic seems to be convinced too that he can make such a switch 
> > on/off facility intuitive. We'll only see if we try it.
> > 
> > Also, i dont see any harm - do you?
> 
> Yes, trace_printk() is a pure debug interface, solely meant for the edit
> + reboot cycle.

So why prevent from making it even more handy?


> If you want anything more than that we've got tracepoints. The rule up
> until now has been to never merge a trace_printk() user.

Sure, that doesn't change the core idea of trace_prink(): none of them must
be merged. That new event interface would just make private uses of trace_printk()
more convenient.