Re: need_heavy_qs flag for PREEMPT=y kernels

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

On Sun, Aug 11, 2019 at 09:40:53PM -0400, Joel Fernandes wrote:
> On Sun, Aug 11, 2019 at 09:24:31PM -0400, Joel Fernandes wrote:
> > On Sun, Aug 11, 2019 at 04:30:24PM -0700, Paul E. McKenney wrote:
> [snip]
> > > Next question:  Why does rcu_implicit_dynticks_qs() check only for
> > > tick_nohz_full_cpu() and not also IS_ENABLED(CONFIG_PREEMPT)?  After
> > > all, a nohz_full CPU in a !CONFIG_PREEMPT kernel should be able to
> > > rely on cond_resched(), right?
> > > 
> > > Should this change?  Why or why not?
> > 
> > Let me think more about this :) I have an answer in mind but I will think a
> > bit more about it and responsd :)
> 
> It should not change. That's because (as somewhat mentioned in the comments),
> some code paths in the kernel check need_resched() before invoking
> cond_resched(). So even with PREEMPT=n having the help of cond_resched(), the
> cond_resched() may actually not even be invoked.  So in this case, the
> resched_cpu() from rcu_implicit_dynticks_qs() does the needful by setting the
> rescheduling flags on the CPU, so that cond_resched() on those CPUs actually
> get called. Is that a correct analysis?

Looks valid to me!  There might well be other scenarios as well, but
only one is required to justify the resched_cpu().

							Thanx, Paul