KVM/QEMU live migration fails with Ubuntu wily kernel 4.2.0-30.35

[Tim Gardner]

On 03/01/2016 12:49 AM, Thomas Lamprecht wrote:
> Hi all!
>
> First of all, most of the information I'm posting here is available at:
> https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1540532
> But as it seems this "bug report" won't get much attention anymore I'm
> posting it here again, in the hope that I do not step on any feet and
> that it gets resolved
>
> The Ubuntu-4.2.0-30.35 introduced a regression regarding live migration.
> When migrating live the VM will start on the target host but is
> completely frozen.
> This affects all types of VMs, I after receiving a bug report with a
> WIndows VM I tested it with various distros, like: Alpine Linux, Linux
> Mint, Ubuntu Desktop 15.10, Windows 7.
> I always get the same result, the VM is completely frozen and in no way
> responsive or reachable. In some cases the VM was not frozen but rather
> reset on the target.
>
> A git bisect and a bit of guessing gave me the culprit, its
> 3f11933efc9ef55ecb2ac7e6d626e8d05a99a4b1 (KVM: x86: expose MSR_TSC_AUX
> to userspace) from the ubuntu repo, or the respective kernel upstream
> patch9dbe6cf941a6fe82933aef565e4095fb10f65023 .
>
> Reverting this fixes the issue (with no introduced regression, as far as
> I've seen).
>
> But as I tested the master and master-next branches from the Ubuntu
> Xenial kernel repos and I knew that this patch was there and live
> migration was working I reverse bisect (i.e. switch good and bad) the
> 4.4 to find the fix to this issue and backported it back to the 4.2 Wily
> kernel and tested it successfully.
>
> The fix is commit 81b1b9ca6d5ca5f3ce91c0095402def657cf5db3 (KVM: VMX:
> Fix host initiated access to guest MSR_TSC_AUX) upstream and the
> backport patch is attached to the Ubuntu launchpad bug link i posted at
> the beginning.
> https://launchpadlibrarian.net/243097677/0001-KVM-VMX-Fix-host-initiated-access-to-guest-MSR_TSC_A.patch
>
> The backport was rather trivial, I only had to replace the new
> guest_cpuid_has_rdtscp(vcpu) with the old check on the vmx struct.
>
> If any action by my side should/must be taken to land this in ubuntu
> kernel repo, e.g. sending the patch to this list, feel free to notify me.
>
> best regards,
> Thomas
>
>

Kamal - this looks like a stable update regression.

-- 
Tim Gardner tim.gardner at canonical.com