arm64 torture test hotplug failures (offlining causes -EBUSY)
Paul E. McKenney
paulmck at kernel.org
Tue Jan 17 20:00:58 PST 2023
On Wed, Jan 18, 2023 at 02:17:06AM +0000, Joel Fernandes wrote:
> Hi Paul,
>
> On Tue, Jan 17, 2023 at 8:42 PM Paul E. McKenney <paulmck at kernel.org> wrote:
> >
> > On Tue, Jan 17, 2023 at 08:02:24PM +0000, Joel Fernandes wrote:
> > > On Tue, Jan 17, 2023 at 4:54 AM Paul E. McKenney <paulmck at kernel.org> wrote:
> > > >
> > > > On Mon, Jan 16, 2023 at 11:36:57PM -0500, Joel Fernandes wrote:
> > > > > > On Jan 16, 2023, at 11:30 PM, Paul E. McKenney <paulmck at kernel.org> wrote:
> > > > > > On Tue, Jan 17, 2023 at 12:15:07AM +0000, Joel Fernandes wrote:
> > > > > >>> On Mon, Jan 16, 2023 at 05:38:00PM -0500, Joel Fernandes wrote:
> > > > > >>> Hi Zhouyi,
> > > > > >>>
> > > > > >>> On Mon, Jan 16, 2023 at 1:33 PM Zhouyi Zhou <zhouzhouyi at gmail.com> wrote:
> > > > > >>>>
> > > > > >>> [..]
> > > > > >>>> On Tue, Jan 17, 2023 at 1:27 AM Joel Fernandes <joel at joelfernandes.org> wrote:
> > > > > >>>>>
> > > > > >>>>> Hello,
> > > > > >>>>> I am seeing -EBUSY returned a lot during torture_onoff() when running
> > > > > >>>>> rcutorture on arm64. This causes hotplug failure 30% of the time. I am
> > > > > >>>>> also seeing this in 6.1-rc kernels. I believe see this only for CPU0.
> > > > > >>>>>
> > > > > >>>>> This causes warnings in torture tests:
> > > > > >>>>> [ 217.582290] rcu-torture:torture_onoff task: offline 0 failed: errno -16
> > > > > >>>>> [ 221.866362] rcu-torture:torture_onoff task: offline 0 failed: errno -16
> > > > > >>>>>
> > > > > >>>>> Full kernel log here:
> > > > > >>>>> http://box.joelfernandes.org:9080/job/rcutorture_stable_arm/job/linux-5.15.y/7/artifact/tools/testing/selftests/rcutorture/res/2023.01.15-14.51.11/TREE04/console.log
> > > > > >>>>>
> > > > > >>>>> Any ideas on why this is happening and only for CPU 0 (presumably the
> > > > > >>>>> boot CPU)? I'd personally need these warnings to go away for my tests
> > > > > >>>>> as this causes rcutorture's tests to not cleanly pass for me. It
> > > > > >>>>> appears remove_cpu() -> device_offline() is what returns the error.
> > > > > >>>>>
> > > > > >>>> I guess this probably because CPU 0 is the tick_do_timer_cpu in
> > > > > >>>> nohz_full mode, which prevent that cpu from
> > > > > >>>> going offline [1]. We have discussed this topic, but there is no
> > > > > >>>> agreement on how to solve it yet.
> > > > > >>>
> > > > > >>> But I am seeing the issue in TRACE02 config which is:
> > > > > >>> CONFIG_NO_HZ_IDLE=y
> > > > > >>> # CONFIG_NO_HZ_FULL is not set
> > > > > >>>
> > > > > >>> So that is not NO_HZ_FULL:
> > > > > >>> http://box.joelfernandes.org:9080/job/rcutorture_stable_arm/job/linux-5.15.y/7/artifact/tools/testing/selftests/rcutorture/res/2023.01.15-14.51.11/TRACE02/console.log.diags/
> > > > > >>> However, I can't seem to find the full kernel logs for that.
> > > > > >>>
> > > > > >>> Also, other than the TRACE02 fail, I only see the issue with configs
> > > > > >>> with CONFIG_NO_HZ_FULL=y
> > > > > >>>
> > > > > >>> Can you try TRACE02 specifically, and see if you can reproduce the
> > > > > >>> same issue on your setup? Meanwhile, I'll try to trace what is
> > > > > >>> returning the -EBUSY.
> > > > > >>
> > > > > >> How about something simple like the following? (untested)
> > > > > >>
> > > > > >> ---8<-----------------------
> > > > > >>
> > > > > >> diff --git a/kernel/torture.c b/kernel/torture.c
> > > > > >> index bc8fb361efc0..cd64110694c0 100644
> > > > > >> --- a/kernel/torture.c
> > > > > >> +++ b/kernel/torture.c
> > > > > >> @@ -220,6 +220,9 @@ bool torture_offline(int cpu, long *n_offl_attempts, long *n_offl_successes,
> > > > > >> // PCI probe frequently disables hotplug during boot.
> > > > > >> (*n_offl_attempts)--;
> > > > > >> s = " (-EBUSY forgiven during boot)";
> > > > > >> + } else if (tick_nohz_full_running && ret == -EBUSY) {
> > > > > >> + (*n_offl_attempts)--;
> > > > > >> + s = " (-EBUSY forgiven if nohz_full is running)";
> > > > > >
> > > > > > But this should be forgiven for the timekeeping CPU, not everyone,
> > > > > > correct?
> > > > > >
> > > > > > Yes, I know that CPU-hotplug operations can fail, but in my testing
> > > > > > they almost never do. This means that a new failure might well be a
> > > > > > real bug somewhere that needs attention.
> > > > >
> > > > > Sure. We may need to expose some API to reveal that.
> > > > >
> > > > > It appeared though that Thomas in the other thread related to patch
> > > > > from Zhouyi, was suggesting that rcutorture tolerate hotplug failure
> > > > > though, because they are not abnormal, right?
> > > >
> > > > Based on my rcutorture testing experience on x86, they are not at all
> > > > normal. The only time I have seen rcutorture CPU-hotplug failures has
> > > > been due to some bug that needed fixing.
> > >
> > > I see, ok. I need to debug what is returning -EBUSY for !NO_HZ_FULL on
> > > arm64, I will report back once I do.
> > >
> > > Meanwhile, Marc I am wondering if you are able to reproduce the issue
> > > on your side on TRACE02 config, like I am?
> > >
> > > Here is the TRACE02 config fragment:
> > > http://box.joelfernandes.org:9080/job/rcutorture_stable_arm/job/linux-5.15.y/7/artifact/tools/testing/selftests/rcutorture/res/2023.01.15-14.51.11/TRACE02/ConfigFragment/*view*/
> > >
> > > Here are instructions on how to run it (torture test parameters etc)
> > > if you are loading the module yourself:
> > > http://box.joelfernandes.org:9080/job/rcutorture_stable_arm/job/linux-5.15.y/7/artifact/tools/testing/selftests/rcutorture/res/2023.01.15-14.51.11/TRACE02/bare-metal/*view*/
> >
> > I am assuming that this is directed to someone having easy access
> > to ARM hardware.
> >
> > > > Is there a plan to make CPU hotplug failures more frequent?
> > >
> > > I am not aware of such a plan but I was going by "There are quite some
> > > reasons why a CPU-hotplug or a hot-unplug operation can fail, which is
> > > not a fatal problem, really." in [1].
> > >
> > > What about an rcutorture to skip hotplug for a certain cpu id,
> > > rcutorture.skip_hotplug_cpus="0". Can be a last resort. But we/I
> > > should debug this issue more before getting to that.
> >
> > Yes, in fact there already are some checks along those lines, for example,
> > the torture_offline() function's check of cpu_is_hotpluggable(). So for
> > example, as I understand it, a CONFIG_NO_HZ_FULL=y system should mark
> > the housekeeping CPU as !cpu_is_hotpluggable().
>
> I don't think CONFIG_NO_HZ_FULL does any such marking (at least I am
> not seeing it). Even on x86, if you enable
> CONFIG_BOOTPARAM_HOTPLUG_CPU0=y , and CONFIG_NO_HZ_FULL=y, and run
> rcutorture with boot args:
>
> nohz_full=0-3 rcutorture.onoff_interval=100 rcutorture.onoff_holdoff=2
> rcutorture.shutdown_secs=30
>
> You will see this in the kernel logs:
> [ 2.816022] rcu-torture:torture_onoff task: offline 0 failed: errno -16
> [ 2.975913] rcu-torture:torture_onoff task: offline 0 failed: errno -16
>
> So RCU torture test clearly thought the CPUs were hot-pluggable, when
> they was chance for them to return -EBUSY (due to housekeeping and
> what not). So this issue seems to be architecture independent, in that
> sense.
>
> So the 2 ways forward I see are:
> - Make the torture test aware of which CPUs are 'house keeping'
> - Make it possible to turn off CPU0 hotplugging on ARM64 by default
> (via CONFIG or boot option).
>
> Another option could be, forgive -EBUSY on CPU0 for
> CONFIG_NO_HZ_FULL=y. Is it possible to assign a non-0 CPU id as a
> housekeeping CPU?
I would be happier to forgive failure to offline housekeeping CPUs than
blanket forgiveness of CPU 0. Especially given that I recently got
burned by a non-zero boot cpu. ;-)
But wouldn't it be even better for cpu_is_hotpluggable() to know the
NO_HZ_FULL rules of the road?
> Adding Frederic to CC as well as we are talking about
> housekeeping/isolation stuff.
But as you say, perhaps Frederic has a better idea.
> > And topology_init() sets this based on platform_can_hotplug_cpu(cpu).
> > And this function sets CPU 0 as !cpu_is_hotpluggable() unless the
> > architecture specifies a .cpu_can_disable() function.
>
> Ah, that is 32-bit ARM code only. This issue is on 64-bit ARM (arch/arm64/).
Apologies! I will look more carefully at the pathnames next time!
But maybe arm64 needs something similar?
Thanx, Paul
> Thanks,
>
> - Joel
>
>
> > So architectures that don't want specific CPUs to be hotpluggable
> > can and should so specify.
> >
> > Thanx, Paul
> >
> > > Thanks,
> > >
> > > - Joel
> > > [1] https://lkml.org/lkml/2022/11/27/182
More information about the linux-arm-kernel
mailing list