Re: [PATCH 2/2] x86: Make the x86-64 stacktrace code safely callable from scheduler

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, May 12, 2011 at 11:28:10PM +0200, Ingo Molnar wrote:
> 
> * Frederic Weisbecker <fweisbec@gmail.com> wrote:
> 
> > Avoid potential scheduler recursion and deadlock from the
> > stacktrace code by avoiding rescheduling when we re-enable
> > preemption.
> > 
> > This robustifies some scheduler trace events like sched switch
> > when they are used to produce callchains in perf or ftrace.
> 
> > -	put_cpu();
> > +
> > +	/* We want stacktrace to be computable anywhere, even in the scheduler */
> > +	preempt_enable_no_resched();
> 
> So what happens if a callchain profiling happens to be interrupted by a hardirq 
> and the interrupt reschedules the current task? We'll miss the reschedule, 
> right?
> 
> preempt_enable_no_resched() is not a magic 'solve scheduler recursions' bullet 
> - it's to be used only if something else will guarantee the preemption check! 
> But nothing guarantees it here AFAICS.
> 
> A better fix would be to use local_irq_save()/restore().

Good point, but then lockdep itself might trigger a stacktrace from local_irq_save,
leading to a stacktrace recursion.

I can use raw_local_irq_disable(), or may be have a stacktrace recursion protection.
I fear the second solution could lead us to potentially lose useful information
if a stacktrace interrupts another one. Ok these are extreme cases...