[RndTbl] Ryzen with Linux

[Trevor Cordes]

On 2019-01-21 Bill Reid wrote:
> I am running Ryzen 2nd generation with Ubuntu 18.10. The first CPU I
> had Ryzen 5 2400 started causing reboots after a couple of months.
> Waiting for replacement. Currently running Ryzen 3 2200.
> 
> Outside of the first CPU failing no other problems.
> 
> On 2019-01-20 6:59 p.m., Trevor Cordes wrote:
> > Is anyone here using Ryzen (1 or 2, or derivatives) with Linux?
> > (ECC irrelevant, fyi.)

Brad & my Ryzen build yielded interesting results.  Thanks for your
reply, and I hope to hear from anyone else with Ryzen + Linux.

We had it freeze within 5 hours of first booting.  Apparently there's a
nasty chip "bug".  If your reboots were more like freezes (or even if
not), you might have just been hitting this bug.  A CPU change will not
fix it.  It's a fundamental flaw.  Doesn't affect Windows, just Linux
and (most?) BSDs.  MS probably got tips from AMD on mitigation.

https://bugzilla.kernel.org/show_bug.cgi?id=196683

TLDR; easy fix is in the BIOS find the "CPU idle power" or similar and
change it to "typical" or "common" instead of AUTO.  Also add
idle=nomwait to kernel boot parms (may be superfluous after bios
tweak?).  As nearly everyone on the bz thread reports, these tweaks
instantly solve the problem.

More info: problem appears to be with C6 power state of CPU and when
the box goes too idle, too many cores go into too low a voltage and a)
some PS's get confused possibly, and b) the cpu, mobo and ps can't
react quick enough when a core suddenly wakes up from C6 to full-bore
mode.  The power starve crashes a core, or 6.

BIOS tweak tells the CPU to not go into the deepest states at all,
instead keeping a more even power draw.  However, lesser states should
still function so the tweak shouldn't cost more than a few watts (some
people claim 8W at idle) wastage (and none if you aren't idle very
much).

I'm pretty flabbergasted that this issue isn't further widespread, and
isn't really being addressed by AMD because it doesn't affect Windows.
It's 1.5+ years old and can affect all Zen-cored CPUs, 1st and 2nd
gen.  Kernel guys don't seem too excited about it, probably none of
them are using Ryzen yet.

On the bright side, ECC does indeed seem to work perfectly with the new
setup!  Yay.