On Thu, Dec 01, 2016 at 06:30:35AM +0100, Peter Zijlstra wrote:
> On Wed, Nov 30, 2016 at 11:40:19AM -0800, Paul E. McKenney wrote:
> > > See commit:
> > >
> > > 4a81e8328d37 ("rcu: Reduce overhead of cond_resched() checks for RCU")
> > >
> > > Someone actually wrote down what the problem was.
> > Don't worry, it won't happen again. ;-)
> > OK, so the regressions were in the "open1" test of Anton Blanchard's
> > "will it scale" suite, and were due to faster (and thus more) grace
> > periods rather than path length.
> > I could likely counter the grace-period speedup by regulating the rate
> > at which the grace-period machinery pays attention to the rcu_qs_ctr
> > per-CPU variable. Actually, this looks pretty straightforward (famous
> > last words). But see patch below, which is untested and probably
> > completely bogus.
> Possible I suppose. Didn't look too hard at it.
> > > > > Also, I seem to have missed, why are we going through this again?
> > > >
> > > > Well, the point I've brought that up is because having basically two
> > > > APIs for cond_resched is more than confusing. Basically all longer in
> > > > kernel loops do cond_resched() but it seems that this will not help the
> > > > silence RCU lockup detector in rare cases where nothing really wants to
> > > > schedule. I am really not sure whether we want to sprinkle
> > > > cond_resched_rcu_qs at random places just to silence RCU detector...
> > >
> > > Right.. now, this is obviously all PREEMPT=n code, which therefore also
> > > implies this is rcu-sched.
> > >
> > > Paul, now doesn't rcu-sched, when the grace-period has been long in
> > > coming, try and force it? And doesn't that forcing include prodding CPUs
> > > with resched_cpu() ?
> > It does in the v4.8.4 kernel that Boris is running. It still does in my
> > -rcu tree, but only after an RCU CPU stall (something about people not
> > liking IPIs). I may need to do a resched_cpu() halfway to stall-warning
> > time or some such.
> Sure, we all dislike IPIs, but I'm thinking this half-way point is
> sensible, no point in issuing user visible annoyance if indeed we can
> prod things back to life, no?
> Only if we utterly fail to make it respond should we bug the user with
> our failure..
I will put together a patch later today.
My intent is to hold off on the "upgrade cond_resched()" patch, one
step at a time. Longer term, I do very much like the idea of having
cond_resched() do both scheduling and RCU quiescent states, assuming
that this avoids performance pitfalls.
> > > I'm thinking not, because if it did, that would make cond_resched()
> > > actually schedule, which would then call into rcu_note_context_switch()
> > > which would then make RCU progress, no?
> > Sounds plausible, but from what I can see some of the loops pointed
> > out by Boris's stall-warning messages don't have cond_resched().
> > There was another workload that apparently worked better when moved from
> > cond_resched() to cond_resched_rcu_qs(), but I don't know what kernel
> > version was running.
> Egads.. cursed if you do, cursed if you dont eh..
Almost like this was real life! ;-)