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

From: Marc Zyngier
Date: Mon Aug 03 2020 - 09:24:44 EST

On 2020-08-03 12:48, Valentin Schneider wrote:
On 03/08/20 12:38, Vladimir Oltean wrote:
On Mon, Aug 03, 2020 at 10:51:32AM +0100, Robin Murphy wrote:

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.



Thanks Robin. I've applied Marc's "[PATCH v2 00/17] arm/arm64: Turning
IPIs into normal interrupts" series and the LS1028A I'm debugging hangs
in absolutely the same way.

I'm not too surprised by that, wrt accounting this mostly changes where the
stores go to and barely shuffles when they happen (slightly earlier on the
IPI handling path).

Indeed. This series is just a "let's make things be the way they should be",
and isn't really fixing any bug. It actually may introduce a couple...

FWIW I've had 'stress-ng --hrtimers 1' running on my Juno and eMAG for ~15
minutes and haven't had a splat yet.

I've started a couple of VMs with that workload too. 400K irq/s on an
8 vcpu guest, nothing to report so far.

But removing IRQ_TIME_ACCOUNTING from defconfig is a resounding "no way!".
It looks like we have a bug, and it should be squashed, not glanced over.

Jazz is not dead. It just smells funny...