Re: [PATCH v2] EXP rcu: Move expedited grace period (GP) work to RT kthread_worker

[Paul E. McKenney]

On Fri, Apr 08, 2022 at 06:21:03AM -0400, Joel Fernandes wrote:
> On Fri, Apr 8, 2022 at 12:57 AM Kalesh Singh <kaleshsingh@xxxxxxxxxx> wrote:
> [..]
> > +config RCU_EXP_KTHREAD
> > +       bool "Perform RCU expedited work in a real-time kthread"
> > +       depends on RCU_BOOST && RCU_EXPERT
> 
> Doesn't this disable the fix if a system is not RCU_EXPERT ?  Please
> see the definition of RCU_EXPERT:
> "This option needs to be enabled if you wish to make expert-level
> adjustments to RCU configuration."
> 
> I don't think a bug fix counts as an expert-level adjustment.

Good catch!

The constraints that this Kconfig option must meet are as follows:

1.	It must not expose innocent bystanders to additional questions
	in response to "make oldconfig" and friends.  So, if you build
	a kernel, then apply this patch, then run "make oldconfig",
	this last must run to completion without waiting for user input.

	This point is not trivial, and everyone who has submitted a
	mainline pull request violating this rule can attest.  ;-)

2.	Enabing RCU_EXPERT should be the exception rather than the
	rule.  (This rule is all too often honored in the breach.)

3.	If a given Kconfig option is to be selected, it must not depend
	on RCU_EXPERT.  (Otherwise, the build system complains.)

So the question is "Does Android already enable RCU_EXPERT?"  If so,
having this depend on RCU_EXPERT, though not great, is OK given that
Android is the only expected user.  If not, another option is to make
this option depend on ANDROID or some similar Kconfig option that Linus
does not enable in his kernel builds.

> > +       default !PREEMPT_RT && NR_CPUS <= 32
> 
> What is the benefit of turning it off on PREEMPT_RT, even if
> PREEMPT_RT does not use expedited GPs much post-boot? I would think in
> the future if PREEMPT_RT ever uses expedited GPs, they would want this
> feature even more. I'd rather be future-proof now as I don't see any
> advantages of disabling it on !PREEMPT_RT (And a drawback that the fix
> won't apply to those systems). Also will keep the config simple.

Current PREEMPT_RT workloads do not care about synchronize_rcu_expedited()
latencies, so much so that they boot with rcupdate.rcu_normal=y.  Plus
the kthread_create_worker() function does not yet have that many users.
So it makes sense to fence in the effects of this change.

Of course, if you are instead saying that ChromeOS also wants low-latency
synchronize_rcu_expedited(), let's figure out the best way to get this
change to ChromeOS users as well as to Android users.  ;-)

							Thanx, Paul

> > +       help
> > +         Use this option to further reduce the latencies of expedited
> > +         grace periods at the expense of being more disruptive.
> > +
> > +         Accept the default if unsure.
> > +
> >  config RCU_NOCB_CPU
> >         bool "Offload RCU callback processing from boot-selected CPUs"
> >         depends on TREE_RCU
> [...]
> 
> Thanks,
> 
>  - Joel