Re: Guest migration between different Ryzen CPU generations

[Sean Christopherson <seanjc@google.com>]

On Thu, Jun 02, 2022, mike tancsa wrote:
> On 6/2/2022 8:42 AM, Igor Mammedov wrote:
> > On Tue, 31 May 2022 13:00:07 -0400
> > mike tancsa <mike@sentex.net> wrote:
> > 
> > > Hello,
> > > 
> > >       I have been using kvm since the Ubuntu 18 and 20.x LTS series of
> > > kernels and distributions without any issues on a whole range of Guests
> > > up until now. Recently, we spun up an Ubuntu LTS 22 hypervisor to add to
> > > the mix and eventually upgrade to. Hardware is a series of Ryzen 7 CPUs
> > > (3700x).  Migrations back and forth without issue for Ubuntu 20.x
> > > kernels.  The first Ubuntu 22 machine was on identical hardware and all
> > > was good with that too. The second Ubuntu 22 based machine was spun up
> > > with a newer gen Ryzen, a 5800x.  On the initial kernel version that
> > > came with that release back in April, migrations worked as expected
> > > between hardware as well as different kernel versions and qemu / KVM
> > > versions that come default with the distribution. Not sure if migrations
> > > between kernel and KVM versions "accidentally" worked all these years,
> > > but they did.  However, we ran into an issue with the kernel
> > > 5.15.0-33-generic (possibly with 5.15.0-30 as well) thats part of
> > > Ubuntu.  Migrations no longer worked to older generation CPUs.  I could
> > > send a guest TO the box and all was fine, but upon sending the guest to
> > > another hypervisor, the sender would see it as successfully migrated,
> > > but the VM would typically just hang, with 100% CPU utilization, or
> > > sometimes crash.  I tried a 5.18 kernel from May 22nd and again the
> > > behavior is different. If I specify the CPU as EPYC or EPYC-IBPB, I can
> > > migrate back and forth.
> > perhaps you are hitting issue fixed by:
> > https://lore.kernel.org/lkml/CAJ6HWG66HZ7raAa+YK0UOGLF+4O3JnzbZ+a-0j8GNixOhLk9dA@mail.gmail.com/T/
> > 
> Thanks for the response. I am not sure.

I suspect Igor is right.  PKRU/PKU, the offending XSAVE feature in that bug, is
in the "new in 5800" list below, and that bug fix went into v5.17, i.e. should
also be fixed in v5.18.

Unfortunately, there's no Fixes: provided and I'm having a hell of a time trying
to figure out when the bug was actually introduced.  The v5.15 code base is quite
different due to a rather massive FPU rework in v5.16.  That fix definitely would
not apply cleanly, but it doesn't mean that the underlying root cause is different,
e.g. the buggy code could easily have been lurking for multiple kernel versions
before the rework in v5.16.

> That patch is from Feb. Would the bug have been introduced sometime in May to
> the 5.15 kernel than Ubuntu 22 would have tracked ?

Dates don't necessarily mean a whole lot when it comes to stable kernels, e.g.
it's not uncommon for a change to be backported to a stable kernel weeks/months
after it initially landed in the upstream tree.

Is moving to v5.17 or later an option for you?  If not, what was the "original"
Ubuntu 22 kernel version that worked?  Ideally, assuming it's the same FPU/PKU bug,
the fix would be backported to v5.15, but that's likely going to be quite difficult,
especially without knowing exactly which commit introduced the bug.

> Looking at the CPU flags diff between the 5800 and the 3700,
> 
> diff -u 3700x 5800x
> --- 3700x       2022-06-02 14:57:00.331309878 +0000
> +++ 5800x       2022-06-02 14:56:52.403340136 +0000
> @@ -77,6 +77,7 @@
>  hw_pstate
>  ssbd
>  mba
> +ibrs
>  ibpb
>  stibp
>  vmmcall
> @@ -85,6 +86,8 @@
>  avx2
>  smep
>  bmi2
> +erms
> +invpcid
>  cqm
>  rdt_a
>  rdseed
> @@ -122,13 +125,15 @@
>  vgif
>  v_spec_ctrl
>  umip
> +pku
> +ospke
> +vaes
> +vpclmulqdq
>  rdpid
>  overflow_recov
>  succor
>  smca
> -sme
> -sev
> -sev_es
> +fsrm
>  bugs
>  sysret_ss_attrs
>  spectre_v1