investigate kernel/qemu issue on compute nodes

[artificial-intelligence]

dmesg:

12345[3579432.962837] x86/split lock detection: #AC: CPU 0/KVM/3324070 took a split_lock trap at address: 0xfffff805742147bf

currently found links/information:

workaround from kernel 6.2. might be to set:

kernel.split_lock_mitigate=0

for older kernels this might mitigate this issue (everything here is currently untested!):

split_lock_detect=off

[artificial-intelligence]

investigating:

https://gitlab.com/qemu-project/qemu/-/issues/1069

[berendt]

1. Upgrade to linux-generic-hwe-22.04 (Linux Kernel 6.2.0)

[garloff]

12345[3579432.962837] x86/split lock detection: #AC: CPU 0/KVM/3324070 took a split_lock trap at address: 0xfffff805742147bf

So this is a warning from the kernel that it's doing something inefficient - switching the warning off hides it.
Real fix is to change the code to avoid locks that are split between cache lines.
So maybe the answer is to ignore this?

[berendt]

Yes, it was just an overlap with another problem. The problem itself was purely in the payload and independent of the compute nodes. I will document what Sven wrote. These are the parameters for ignoring it.

[artificial-intelligence]

as was pointed out this is more of an application issue, the mitigation from the kernel side is documented above. in the general case the application should be fixed.