Re: [PATCH 00/11] rcu: Make rcu_core() safe in PREEMPT_RT with NOCB + a few other fixes v2

From: Frederic Weisbecker
Date: Thu Oct 14 2021 - 06:44:04 EST


On Wed, Oct 13, 2021 at 09:27:33AM -0700, Paul E. McKenney wrote:
> On Wed, Oct 13, 2021 at 01:43:35PM +0200, Frederic Weisbecker wrote:
> > On Tue, Oct 12, 2021 at 08:28:32PM -0700, Paul E. McKenney wrote:
> > > On Tue, Oct 12, 2021 at 05:32:15PM -0700, Paul E. McKenney wrote:
> > > > On Mon, Oct 11, 2021 at 04:51:29PM +0200, Frederic Weisbecker wrote:
> > > > > Hi,
> > > > >
> > > > > No code change in this v2, only changelogs:
> > > > >
> > > > > * Add tags from Valentin and Sebastian
> > > > >
> > > > > * Remove last reference to SEGCBLIST_SOFTIRQ_ONLY (thanks Valentin)
> > > > >
> > > > > * Rewrite changelog for "rcu/nocb: Check a stable offloaded state to manipulate qlen_last_fqs_check"
> > > > > after off-list debates with Paul.
> > > > >
> > > > > * Remove the scenario with softirq interrupting rcuc on
> > > > > "rcu/nocb: Limit number of softirq callbacks only on softirq" as it's
> > > > > probably not possible (thanks Valentin).
> > > > >
> > > > > * Remove the scenario with task spent scheduling out accounted on tlimit
> > > > > as it's not possible (thanks Valentin)
> > > > > (see "rcu: Apply callbacks processing time limit only on softirq")
> > > > >
> > > > > * Fixed changelog of
> > > > > "rcu/nocb: Don't invoke local rcu core on callback overload from nocb kthread"
> > > > > (thanks Sebastian).
> > > > >
> > > > > git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
> > > > > rcu/rt-v2
> > > > >
> > > > > HEAD: 2c9349986d5f70a555195139665841cd98e9aba4
> > > > >
> > > > > Thanks,
> > > > > Frederic
> > > >
> > > > Nice!
> > > >
> > > > I queued these for further review and testing. I reworked the commit log
> > > > of 6/11 to give my idea of the reason, though I freely admit that this
> > > > reason is not as compelling as it no doubt seemed when I wrote that code.
> > >
> > > But in initial tests TREE04.5, TREE04.6, and TREE04.9 all hit the
> > > WARN_ON(1) in rcu_torture_barrier(), which indicates rcu_barrier()
> > > breakage. My best (but not so good) guess is a five-hour MTBF on a
> > > dual-socket system.
> > >
> > > I started an automated "git bisect" with each step running 100 hours
> > > of TREE04, but I would be surprised if anything useful comes of it.
> > > Pleased, mind you, but surprised.
> >
> > Ok I can reproduce.
> >
> > I'm launching a bisect from my side as well.
>
> Mine converged on 2a4200944750 ("rcu/nocb: Prepare state machine for
> a new step"). The surprise is that I was running "git bisect run"
> on a script wrappering kvm-remote.sh, which means that it managed to
> repeatedly request 10 systems, download to them, run the test, collect
> the results, and finally return the systems.
>
> Huh. I should probably refactor my local script to avoid the pointless
> repeated request/return work.
>
> But which commit did your bisect find? ;-)

So my bisection got confused with two different issues: one with an
oom and one with rcu_barrier() being unhappy.

I'm re-running it but I'll investigate both.

>
> Anyway, I am keeping the first commit 4b246eab4750 ("rcu/nocb: Make
> local rcu_nocb_lock_irqsave() safe against concurrent deoffloading"),
> but dropping the others for the time being.

Fair enough!

Thanks.



> Thanx, Paul