Re: [RFC PATCH] arm64: defconfig: Disable fine-grained task level IRQ time accounting

From: Robin Murphy
Date: Mon Aug 03 2020 - 05:51:39 EST

On 2020-08-03 09:16, Vladimir Oltean wrote:
On Mon, Aug 03, 2020 at 10:04:01AM +0200, Kurt Kanzenbach wrote:
On Thu Jul 30 2020, Vladimir Oltean wrote:
On Thu, Jul 30, 2020 at 09:23:44AM +0200, Kurt Kanzenbach wrote:
On Wed Jul 29 2020, Vladimir Oltean wrote:
For more context, here is my original report of the issue:

Just like you, I could not reproduce the RCU stalls and system hang on a
5.6-rt kernel, just on mainline and derivatives, using the plain

The issue is not specific to Layerscape or i.MX8, but rather I was able
to see the same behavior on Marvell Armada 37xx as well as Qualcomm

So, while of course I agree that disabling IRQ time accounting for arm64
isn't a real solution, it isn't by far an exaggerated proposal either.
Nonetheless, the patch is just a RFC and should be treated as such. We
are at a loss when it comes to debugging this any further and we would
appreciate some pointers.

Yeah, sure. I'll try to reproduce this issue first. So it triggers with:

* arm64
* mainline, not -rt kernel
* opened serial console
* irq accounting enabled

Anything else?


Thanks for giving a helping hand, Kurt. The defconfig should be enough.
In the interest of full disclosure, the only arm64 device on which we
didn't reproduce this was the 16-core LX2160A. But we did reproduce on
that with maxcpus=1 though. And also on msm8976 with all 8 cores booted.
Just mentioning this in case you're testing on a 16-core system, you
might want to reduce the number a bit.

OK. I've reproduced it on a Marvell Armada SoC with v5.6 mainline. See
splats below. Running with irq time accounting enabled, kills the
machine immediately. However, I'm not getting the possible deadlock
warnings in 8250 as you did. So that might be unrelated.

Yes, the console lockdep warnings are unrelated. They are discussed

Unfortunately I have no idea what to debug here.


So, this means we could submit a formal version of this patch? :)

Defconfig is absolutely not the appropriate way to work around bugs - it's merely a starting point for users and distros to set up their own kernel, and if they can still enable this option and render their system unusable then patching some other config that they aren't using is pointless. To usefully mitigate a problem you'd need to make sure the offending option cannot be selected at all (i.e. prohibit HAVE_IRQ_TASK_ACCOUNTING as well).

Having glanced across another thread that mentions IRQ accounting recently[1], I wonder if the underlying bug here might have something do to with the stuff that Marc's trying to clean up.