On Fri, Jul 12, 2019 at 06:01:05AM -0700, Paul E. McKenney wrote:
> On Thu, Jul 11, 2019 at 03:58:39PM -0400, Joel Fernandes wrote:
> > On Thu, Jul 11, 2019 at 12:48:18PM -0400, Joel Fernandes wrote:
> > > On Thu, Jul 11, 2019 at 08:02:15AM -0700, Paul E. McKenney wrote:
> > > > On Thu, Jul 11, 2019 at 09:08:49AM -0400, Joel Fernandes wrote:
> > > > > On Thu, Jul 11, 2019 at 05:30:52AM -0700, Paul E. McKenney wrote:
> > > > > > On Wed, Jul 10, 2019 at 10:20:25AM +0900, Byungchul Park wrote:
> > > > > > > On Tue, Jul 09, 2019 at 05:41:02AM -0700, Paul E. McKenney wrote:
> > > > > > > > > Hi Paul,
> > > > > > > > >
> > > > > > > > > IMHO, as much as we want to tune the time for fqs to be
> > > > > > > > > initiated, we
> > > > > > > > > can also want to tune the time for the help from scheduler to
> > > > > > > > > start.
> > > > > > > > > I thought only difference between them is a level of urgency.
> > > > > > > > > I might be
> > > > > > > > > wrong. It would be appreciated if you let me know if I miss
> > > > > > > > > something.
> > > > > > > >
> > > > > > > > Hello, Byungchul,
> > > > > > > >
> > > > > > > > I understand that one hypothetically might want to tune this at
> > > > > > > > runtime,
> > > > > > > > but have you had need to tune this at runtime on a real
> > > > > > > > production
> > > > > > > > workload? If so, what problem was happening that caused you to
> > > > > > > > want to
> > > > > > > > do this tuning?
> > > > > > >
> > > > > > > Not actually.
> > > > > > >
> > > > > > > > > And it's ok even if the patch is turned down based on your
> > > > > > > > > criteria. :)
> > > > > > > >
> > > > > > > > If there is a real need, something needs to be provided to meet
> > > > > > > > that
> > > > > > > > need. But in the absence of a real need, past experience has
> > > > > > > > shown
> > > > > > > > that speculative tuning knobs usually do more harm than good.
> > > > > > > > ;-)
> > > > > > >
> > > > > > > It makes sense, "A speculative tuning knobs do more harm than
> > > > > > > good".
> > > > > > >
> > > > > > > Then, it would be better to leave jiffies_till_{first,next}_fqs
> > > > > > > tunnable
> > > > > > > but jiffies_till_sched_qs until we need it.
> > > > > > >
> > > > > > > However,
> > > > > > >
> > > > > > > (1) In case that jiffies_till_sched_qs is tunnable:
> > > > > > >
> > > > > > > We might need all of jiffies_till_{first,next}_qs,
> > > > > > > jiffies_till_sched_qs and jiffies_to_sched_qs because
> > > > > > > jiffies_to_sched_qs can be affected by any of them. So we
> > > > > > > should be able to read each value at any time.
> > > > > > >
> > > > > > > (2) In case that jiffies_till_sched_qs is not tunnable:
> > > > > > >
> > > > > > > I think we don't have to keep the jiffies_till_sched_qs any
> > > > > > > longer since that's only for setting jiffies_to_sched_qs at
> > > > > > > *booting time*, which can be done with jiffies_to_sched_qs too.
> > > > > > > It's meaningless to keep all of tree variables.
> > > > > > >
> > > > > > > The simpler and less knobs that we really need we have, the
> > > > > > > better.
> > > > > > >
> > > > > > > what do you think about it?
> > > > > > >
> > > > > > > In the following patch, I (1) removed jiffies_till_sched_qs and
> > > > > > > then
> > > > > > > (2) renamed jiffies_*to*_sched_qs to jiffies_*till*_sched_qs
> > > > > > > because I
> > > > > > > think jiffies_till_sched_qs is a much better name for the
> > > > > > > purpose. I
> > > > > > > will resend it with a commit msg after knowing your opinion on it.
> > > > > >
> > > > > > I will give you a definite "maybe".
> > > > > >
> > > > > > Here are the two reasons for changing RCU's embarrassingly large
> > > > > > array
> > > > > > of tuning parameters:
> > > > > >
> > > > > > 1. They are causing a problem in production. This would represent
> > > > > > a
> > > > > > bug that clearly must be fixed. As you say, this change is not
> > > > > > in this category.
> > > > > >
> > > > > > 2. The change simplifies either RCU's code or the process of tuning
> > > > > > RCU, but without degrading RCU's ability to run everywhere and
> > > > > > without removing debugging tools.
> > > > > >
> > > > > > The change below clearly simplifies things by removing a few lines
> > > > > > of
> > > > > > code, and it does not change RCU's default self-configuration. But
> > > > > > are
> > > > > > we sure about the debugging aspect? (Please keep in mind that many
> > > > > > more
> > > > > > sites are willing to change boot parameters than are willing to
> > > > > > patch
> > > > > > their kernels.)
> > > > > >
> > > > > > What do you think?
> > > > >
> > > > > Just to add that independent of whether the runtime tunable make
> > > > > sense or
> > > > > not, may be it is still worth correcting the 0444 to be 0644 to be a
> > > > > separate
> > > > > patch?
> > > >
> > > > You lost me on this one. Doesn't changing from 0444 to 0644 make it be
> > > > a runtime tunable?
> > >
> > > I was going by our earlier discussion that the parameter is still
> > > writable at
> > > boot time. You mentioned something like the following:
> > > ---
> > > In Byungchul's defense, the current module_param() permissions are
> > > 0444, which really is read-only. Although I do agree that they can
> > > be written at boot, one could use this same line of reasoning to argue
> > > that const variables can be written at compile time (or, for on-stack
> > > const variables, at function-invocation time). But we still call them
> > > "const".
> > > ---
> > >
> > > Sorry if I got confused. You are right that we could leave it as
> > > read-only.
> > >
> > > > > > Finally, I urge you to join with Joel Fernandes and go through these
> > > > > > grace-period-duration tuning parameters. Once you guys get your
> > > > > > heads
> > > > > > completely around all of them and how they interact across the
> > > > > > different
> > > > > > possible RCU configurations, I bet that the two of you will have
> > > > > > excellent
> > > > > > ideas for improvement.
> > > > >
> > > > > Yes, I am quite happy to join forces. Byungchul, let me know what
> > > > > about this
> > > > > or other things you had in mind. I have some other RCU topics too I
> > > > > am trying
> > > > > to get my head around and planning to work on more patches.
> > > > >
> > > > > Paul, in case you had any other specific tunables or experiments in
> > > > > mind, let
> > > > > me know. I am quite happy to try out new experiments and learn
> > > > > something
> > > > > based on tuning something.
> > > >
> > > > These would be the tunables controlling how quickly RCU takes its
> > > > various actions to encourage the current grace period to end quickly.
> > > > I would be happy to give you the exact list if you wish, but most of
> > > > them have appeared in this thread.
> > > >
> > > > The experiments should be designed to work out whether the current
> > > > default settings have configurations where they act badly. This might
> > > > also come up with advice for people attempting hand-tuning, or proposed
> > > > parameter-checking code to avoid bad combinations.
> > > >
> > > > For one example, setting the RCU CPU stall timeout too low will
> > > > definitely
> > > > cause some unwanted splats. (Yes, one could argue that other things in
> > > > the kernel should change to allow this value to decrease, but things
> > > > like latency tracer and friends are probably more useful and important.)
> > >
> > > Ok, thank you for the hints.
> >
> > Hmm, speaking of grace period durations, it seems to me the maximum grace
> > period ever is recorded in rcu_state.gp_max. However it is not read from
> > anywhere.
> >
> > Any idea why it was added but not used?
>
> If I remember correclty, it used to be used in debugfs prints. It is
> useful for working out how low you can decrease rcutorture.stall_cpu to
> without getting RCU CPU stall warnings. A rather infrequent need,
> given that the mainline default has been adjusted only once.
Got it.
> > I am interested in dumping this value just for fun, and seeing what I get.
> >
> > I wonder also it is useful to dump it in rcutorture/rcuperf to find any
> > issues, or even expose it in sys/proc fs to see what worst case grace
> > periods
> > look like.
>
> That might be worthwhile.
Thanks. I am thinking also it will help see whether something else like GP
thread priority or other configuration could be causing long GP times that is
otherwise not possible to see without a counter.
I will work out with Byungchul and come up with something.
Cheers,
- Joel
