From: Andrew Cooper [mailto:andrew.cooper3@xxxxxxxxxx]
Sent: Friday, October 27, 2017 5:55 PM
To: Hao, Xudong <xudong.hao@xxxxxxxxx>; xen-devel@xxxxxxxxxxxxx
Cc: Lars Kurth <lars.kurth@xxxxxxxxxx>; Julien Grall <julien.grall@xxxxxxx>
Subject: Re: [Xen-devel] Xen 4.10.0 RC1 test result
On 27/10/17 09:28, Hao, Xudong wrote:
We performed Xen 4.10 RC1 testing on Intel Xeon Skylake, Broadwell server, Intel Atom Denverton platforms, verified many functional features, which include new features Local MCE, L2 CAT and UMIP on Xen 4.10. We’d like to share the result
out.
Most of features passed to testing on Xen 4.10 RC1, VT-d, RAS and nested has some bugs.
VT-d:
[BUG] win2008 guest cannot get ip through sriov
https://www.mail-archive.com/xen-devel@xxxxxxxxxxxxx/msg127433.html
RAS:
[BUG] xen-mceinj tool testing cause dom0 crash
https://www.mail-archive.com/xen-devel@xxxxxxxxxxxxx/msg108671.html
Nested:
Nested status is better than Xen 4.9.0, KVM on Xen, HyperV on Xen works, while Xen on Xen, VMware on Xen fail.
https://wiki.xenproject.org/wiki/Nested_Virtualization_in_Xen
Do you have any further details on your HyperV scenarios, in particular versions of HyperV and the hardware involved, and guests booted under HyperV?
XenServers current nested-virt testing status shows a rather bleaker picture.
More modern version of Windows Server fail to initialise the HyperV role, because Xen doesn't advertise Virtual NMI support to L1. (One version, Server 2012 R2 I believe, indicates the same, but with a BSOD instead). Older versions still do actually boot
successfully.
When booting windows guests under nested HyperV, old versions appear to be stable with a single one-vcpu guest, but unstable with multiple vcpus or multiple single-vcpu guests. The instability here is a VMEntry failure trying to inject an NMI, and occurs because
HyperV and Xen disagree on whether to use Virtual NMI, resulting in HyperV thinking virtual NMI is disabled, but it is actually enabled in hardware.
When booting windows guests under more modern nested HyperV, the guest is crashing because of a pagefault when trying to access the APIC page. We haven't tracked down the cause of this, but I expect it is something to do with emulating instruction while in
nested vcpu context.
Thanks,
~Andrew