Hi Paul,
Thanks for your inputs.
On Wed, 16 Dec 2020 at 21:33, Paul E. McKenney <[email protected]> wrote:
>
> On Wed, Dec 16, 2020 at 03:40:04PM +0530, Naresh Kamboju wrote:
> > Linux Kernel Functional Testing (LKFT) started running rcu-torture tests on
> > qemu_arm64, qemu_arm qemu_x86_64 and qemu_i386 from our CI build systems.
> >
> > The following warning(s) noticed on qemu_i386 while running rcu-torture test
> > on Linux mainline and Linux -next master branch. Since we do not have baseline
> > results i can not comment this as regression but when compared with
> > stable-rc 5.4 kernel this warning is new on mainline and next.
>
> The rcutorture testing "stutters", that is, it periodically intentionally
> drops the test load down to zero for a few seconds. The expectation is
> that with no load, rcutorture will have no trouble finishing any needed
> grace periods within that zero-load period. If at the end of the stutter
> period, RCU work remains undone, then this warning is emitted.
>
> This warning can be a false positive in the following situations:
>
> 1. The system on which you are running rcutorture is under
> additional heavy load.
The DUT is running the test - rcutorture - only.
> 2. You are running multiple guest OSes, each of which is running
> rcutorture, and vCPUs from each of the guest OSes ends up
> sharing a core with a vCPU from one of the other guests. This
> can cause the zero-load period to not be so unloaded.
>
> 3. You built rcutorture into your kernel, so that rcutorture starts
> immediately at boot time (CONFIG_RCU_TORTURE_TEST=y). If your
> boot takes long enough, rcutorture can massively overload the
> single boot CPU, which can in turn result in this warning.
The test was built as a module.
CONFIG_RCU_TORTURE_TEST=m
>
> If you are in situation #1, I suggest disabling stuttering using the
> rcutorture.stutter=0 kernel boot parameter.
>
> If you are in situation #2, I suggest binding the guest-OS vCPUs
> to avoid them sharing cores with each other.
>
> If you are in situation #3, I have patches that I expect to submit
> upstream in the v5.12 merge window that can help. Hey, they work for me!
> If you would like to test them before then, please let me know.
>
> If something else is going on, please let me know what it is so that
> I can fix it one way or another.
We were running on qemu_i386 today. I have tested on real hardware
and the reported problem has been reproduced.
> This warning has been present for quite some time, but I continually
> make rcutorture more aggressive, and this could well be part of the
> fallout of additional rcutorture aggression.
>
> And either way, thank you for trying out rcutorture!
We are happy to test :)
- Naresh
On Thu, Dec 17, 2020 at 04:18:52PM +0530, Naresh Kamboju wrote:
> Hi Paul,
>
> Thanks for your inputs.
>
> On Wed, 16 Dec 2020 at 21:33, Paul E. McKenney <[email protected]> wrote:
> >
> > On Wed, Dec 16, 2020 at 03:40:04PM +0530, Naresh Kamboju wrote:
> > > Linux Kernel Functional Testing (LKFT) started running rcu-torture tests on
> > > qemu_arm64, qemu_arm qemu_x86_64 and qemu_i386 from our CI build systems.
> > >
> > > The following warning(s) noticed on qemu_i386 while running rcu-torture test
> > > on Linux mainline and Linux -next master branch. Since we do not have baseline
> > > results i can not comment this as regression but when compared with
> > > stable-rc 5.4 kernel this warning is new on mainline and next.
> >
> > The rcutorture testing "stutters", that is, it periodically intentionally
> > drops the test load down to zero for a few seconds. The expectation is
> > that with no load, rcutorture will have no trouble finishing any needed
> > grace periods within that zero-load period. If at the end of the stutter
> > period, RCU work remains undone, then this warning is emitted.
> >
> > This warning can be a false positive in the following situations:
> >
> > 1. The system on which you are running rcutorture is under
> > additional heavy load.
>
> The DUT is running the test - rcutorture - only.
>
> > 2. You are running multiple guest OSes, each of which is running
> > rcutorture, and vCPUs from each of the guest OSes ends up
> > sharing a core with a vCPU from one of the other guests. This
> > can cause the zero-load period to not be so unloaded.
> >
> > 3. You built rcutorture into your kernel, so that rcutorture starts
> > immediately at boot time (CONFIG_RCU_TORTURE_TEST=y). If your
> > boot takes long enough, rcutorture can massively overload the
> > single boot CPU, which can in turn result in this warning.
>
> The test was built as a module.
> CONFIG_RCU_TORTURE_TEST=m
>
> >
> > If you are in situation #1, I suggest disabling stuttering using the
> > rcutorture.stutter=0 kernel boot parameter.
> >
> > If you are in situation #2, I suggest binding the guest-OS vCPUs
> > to avoid them sharing cores with each other.
> >
> > If you are in situation #3, I have patches that I expect to submit
> > upstream in the v5.12 merge window that can help. Hey, they work for me!
> > If you would like to test them before then, please let me know.
> >
> > If something else is going on, please let me know what it is so that
> > I can fix it one way or another.
>
> We were running on qemu_i386 today. I have tested on real hardware
> and the reported problem has been reproduced.
>
> > This warning has been present for quite some time, but I continually
> > make rcutorture more aggressive, and this could well be part of the
> > fallout of additional rcutorture aggression.
> >
> > And either way, thank you for trying out rcutorture!
>
> We are happy to test :)
Is this reproducible? If so, could you please try bisection?
Thanx, Paul
> Is this reproducible? If so, could you please try bisection?
Yes. This issue is reproducible.
I will run bisection and get back to you (might take some time).
- Naresh