In mid April of 2018, AMD seemed to have acknowledged the mwait
"problem", and said BIOS changes would be pushed out "soon". In the
meantime, the idle=nomwait kernel command line would affect a
reasonable temporary solution.
If you have the linux-cpupower package installed, the command
will tell you about the idle subsystem.
In the acknowledgement, it was mentioned that this was only affecting
kernels newer than 4.13 (I am running a 4.17 or 4.18 kernel). I think
most people running stable system, are probably running 4.9.
If you go to play with that zenstates.py program I mentioned in the
previous message, the instruction for that say that root is the user who
runs it, and the msr kernel module must be loaded for it to work.
Will this solve the problem? These EPYC/Threadripper people are
talking about reboots. What I am seeing is freezes (CPU halts and
never starts again). Is this the same problem? I don't know. At the
moment I am running with idle-nomwait, and I'll see if it freezes
again. If it doesn't freeze for a while, I guess I will start dropping
the number of boinc jobs down, so that it spends more time in the idle