Re: perf: fuzzer BUG: KASAN: stack-out-of-bounds in __unwind_start

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Dec 01, 2016 at 05:41:07PM +0100, Peter Zijlstra wrote:
> On Thu, Dec 01, 2016 at 04:33:16AM -0800, Paul E. McKenney wrote:
> > On Thu, Dec 01, 2016 at 06:52:35AM +0100, Peter Zijlstra wrote:
> > > On Wed, Nov 30, 2016 at 01:13:03PM -0600, Josh Poimboeuf wrote:
> > > > This question was probably intended for other folks, but I should point
> > > > out that idle tasks *do* invoke the scheduler.  cpu_idle_loop() calls
> > > > schedule_preempt_disabled().
> > > 
> > > Right, but that doesn't matter I think. The below will simply not call
> > > rcu_note_voluntary_context_switch() from the idle task, which would be
> > > fine I think.
> > > 
> > > > > So is the following a sensible approach, or should I look elsewhere?
> > > > > 
> > > > > 	#define cond_resched_rcu_qs() \
> > > > > 	do { \
> > > > > 		if (!is_idle_task(current) && !cond_resched()) \
> > > > > 			rcu_note_voluntary_context_switch(current); \
> > > 
> > > You should reverse your conditions though:
> > > 
> > > 		if (!cond_resched() && !is_idle_task(current))
> > > 			rcu_note_voluntary_context_switch(current);
> > > 
> > > That way we'll still do cond_resched() and you only gate the RCU call.
> > 
> > This makes it illegal at early boot.
> 
> Humm, how early are we talking?

The case I saw was during start_kernel(), IIRC.  But again, it turned
out that the patch putting cond_resched_rcu_qs() in that early was
(1) broken and (2) unnecessary.

							Thanx, Paul