Re: [PATCH] rcu/nocb: Add an option to ON/OFF an offloading from RT context

[Uladzislau Rezki]

> On Mon, May 09, 2022 at 01:17:00PM -0400, Joel Fernandes wrote:
> > On Sun, May 8, 2022 at 11:37 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > >
> > > On Sun, May 08, 2022 at 08:17:49PM -0400, Joel Fernandes wrote:
> > > > On Sun, May 8, 2022 at 5:32 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >  [...]
> > > > > > Also, I think it is wrong to assume that a certain kind of system will
> > > > > > always have a certain number of callbacks to process at a time. That
> > > > > > seems prone to poor design due to assumptions which may not always be
> > > > > > true.
> > > > >
> > > > > Who was assuming that?  Uladzislau was measuring rather than assuming,
> > > > > if that was what you were getting at.  Or if you are thinking about
> > > > > things like qhimark, your point is exactly why there is both a default
> > > > > (which has worked quite well for a very long time) and the ability to
> > > > > adjust based on the needs of your specific system.
> > > >
> > > > I was merely saying that based on measurements make assumptions, but
> > > > in the real world the assumption may not be true, then everything
> > > > falls apart. Instead I feel, callback threads should be RT only if 1.
> > > > As you mentioned, the time based thing. 2. If the CB list is long and
> > > > there's lot of processing. But instead, if it is made a CONFIG option,
> > > > then that forces a fixed behavior which may fall apart in the real
> > > > world. I think adding more CONFIGs and special cases is more complex
> > > > but that's my opinion.
> > >
> > > Again, exactly what problem are you trying to solve?
> > >
> > > From what I can see, Uladzislau's issue can be addressed by statically
> > > setting the rcuo kthreads to SCHED_OTHER at boot time.  The discussion
> > > is on exactly how RCU is to be informed of this, at kernel build time.
> > >
> > > > > > Can we not have 2 sets of RCU offload threads, one which operate at RT
> > > > > > and only process few callbacks at a time, while another which is the
> > > > > > lower priority CFS offload thread - executes whenever there is a lot
> > > > > > of CBs pending? Just a thought.
> > > > >
> > > > > How about if we start by solving the problems we know that we have?
> > > >
> > > > I don't know why you would say that, because we are talking about
> > > > solving the specific problem Vlad's patch addresses, not random
> > > > problems. Which is that, Android wants to run expedited GPs, but when
> > > > the callback list is large, the RT nocb thread can starve other
> > > > things. Did I misunderstand the patch? If so, sorry about that but
> > > > that's what my email was discussing. i.e. running of CBs in RT
> > > > threads. I suck at writing well as I clearly miscommunicated.
> > >
> > > OK.
> > >
> > > Why do you believe that this needs anything other than small adjustments
> > > the defaults of existing Kconfig options?  Or am I completely missing
> > > the point of your proposal?
> > >
> > > > > > Otherwise, I feel like we might be again proliferating CONFIG options
> > > > > > and increasing burden on the user to get it the CONFIG right.
> > > > >
> > > > > I bet that we will solve this without adding any new Kconfig options.
> > > > > And I bet that the burden is at worst on the device designer, not on
> > > > > the user.  Plus it is entirely possible that there might be a way to
> > > > > automatically configure things to handle what we know about today,
> > > > > again without adding Kconfig options.
> > > >
> > > > Yes, agreed.
> > >
> > > If I change my last sentence to read as follows, are we still in
> > > agreement?
> > >
> > >         Plus it is entirely possible that there might be a way to
> > >         automatically configure things to handle what we know about today,
> > >         again without adding Kconfig options and without changing runtime
> > >         code beyond that covered by Uladzislau's patch.
> > 
> > Yes, actually the automatic configuration of things is what I meant,
> > that's the "problem" I was referring to, where the system does the
> > right thing for a broader range of systems, without requiring the
> > users to find RCU issues and hand-tune them (that requires said users
> > to have tracing and debugging skills and get lucky finding a problem).
> > To be fair, I did not propose any solutions to such problems either,
> > it is just some ideas. I don't like knobs too much and I don't trust
> > users or system designers to get them right most of the time.
> > 
> > In that sense,  I don't think making rcuo threads run as RT or not
> > (which this patch does) is really fixing the problems. In one case,
> > you might have priority inversion, in another case you might cause
> > starvation. Probably what is needed is best of both worlds. That said,
> > I don't have better solutions right now than what I mentioned, which
> > is to assign priorities to the callbacks themselves and run them in
> > threads of different priorities.
> > 
> > For the record, I am not against the patch or anything like that (and
> > even if I was, I am not sure that it matters for merging :P)
> 
> Fair enough!
> 
> And for the record at this end, I would not be surprised if in 2032
> RCU offloaded callback invocation has sophisticated dynamic tuning of
> priorities and much else besides.  But one step at a time!  ;-)
> 
hh... It is hard to comment because i am a bit lost in this big conversation :)

What i have got so far. Joel does not like adding extra *_CONFIG
options, actually me too since it becomes more complicated thus
it requires more specific attention from users. I prefer to make
the code common but it is not possible sometimes to make it common,
because we have different kind of kernels and workloads.

>From the other hand the patch splits the BOOSTING logic into two peaces
because driving the grace periods kthreads in RT priority is not a big
issue because their run-times are short. Whereas running the "kthreads-callbacks"
in the RT context can be long so we end up in throttled situation for
other workloads.

I see that Paul would like to keep it for CONFIG_PREEMPT_RT, because it
was mainly designed for that kind of kernels. So we can align with Alison
patch and her decision, so i do not see any issues. So far RT folk seems
does not mind in having "callback-kthreads" as SCHED_FIFO :)

Do you agree with start from keeping it ON for CONFIG_PREEMPT_RT conf.
by default and OFF for other cases?

--
Uladzislau Rezki