Re: [PATCH 2/3] trace: fix default boot up tracer

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 2 Feb 2009 23:12:37 -0500 (EST) Steven Rostedt <rostedt@goodmis.org> wrote:

> 
> On Mon, 2 Feb 2009, Andrew Morton wrote:
> 
> > On Mon, 02 Feb 2009 21:38:32 -0500 Steven Rostedt <rostedt@goodmis.org> wrote:
> > 
> > > @@ -510,8 +520,25 @@ int register_tracer(struct tracer *type)
> > >   out:
> > >  	tracing_selftest_running = false;
> > >  	mutex_unlock(&trace_types_lock);
> > > -	lock_kernel();
> > >  
> > > +	if (!ret && default_bootup_tracer) {
> > > +		if (!strncmp(default_bootup_tracer, type->name,
> > > +			     BOOTUP_TRACER_SIZE)) {
> > > +			printk(KERN_INFO "Starting tracer '%s'\n",
> > > +			       type->name);
> > > +			/* Do we want this tracer to start on bootup? */
> > > +			tracing_set_tracer(type->name);
> > > +			default_bootup_tracer = NULL;
> > > +			/* disable other selftests, since this will break it. */
> > > +			tracing_selftest_disabled = 1;
> > > +#ifdef CONFIG_FTRACE_STARTUP_TEST
> > > +			printk(KERN_INFO "Disabling FTRACE selftests due"
> > > +			       " to running tracer '%s'\n", type->name);
> > > +#endif
> > > +		}
> > > +	}
> > > +
> > > +	lock_kernel();
> > >  	return ret;
> > >  }
> > 
> > The fun and games which register_tracer() plays with lock_kernel() tell
> > us that this function is only called at bootup time and hence could be
> > __init.  A quick whizz through callers confirms that.
> > 
> > And if register_tracer() is also only callable at bootup, one suspects
> > that unregister_tracer() isn't useful.  And lo, it has no callers.
> > 
> > This leads one to further surmise that trace_types_lock a) could be
> > __initdata and b) could be removed (the list is only altered when we're
> > running single-threaded).  This appears to be the case.
> 
> The lock_kernel addition was added when the BKL became a spinlock again. 
> The selftests needed to be able to sleep, and this caused issues.

Sleeping inside lock_kernel() is quite OK.  Confused.

What is the call path to this function?  Does it all happen under
ftrace_init()?  If not, do we risk breaking start_kernel()'s
thou-shalt-not-enable-interrupts-early rule which powerpc (at least)
imposes?

> The register_tracer was initial written to be pluggable at any time. 
> Perhaps in the future to allow modules. But this does not seem to have 
> panned out.
> 
> Since we have the lock_kernel there anyway, if we ever need to handle 
> modules, that will need a different interface anyway. I guess I can nuke 
> the unregister tracer.

And add some __init/__initdatas?

> As for the trace_types_lock mutex, that is still needed. Not only did it 
> guard against the tracer registry, it also guards the switching of tracers.
> 
>  # echo function > /debug/tracing/current_tracer
> 

OK, it's a dual-use lock, as the comment indicates.