Red Hat Bugzilla – Bug 1309320
qemu-kvm feature request to hide hypervisor from guest
Last modified: 2016-02-25 08:56:18 EST
Description of problem:
It would be nice if qemu-kvm in RHEL 7.x has feature to
hide hypervisor from guest via command line option kvm=off
(-cpu host,kvm=off). This is essential for GPU passthrough
with current NVIDIA cards and latest drivers. See page 23
RHEL 7.2 libvirt already supports this via
QEMU-KVM-EV is new enough to have this support as well.
So for GPU passthrough one has to go with qemu-kvm-ev
only because it supports kvm=off instead of qemu-kvm.
Enabling feature in domain XML
Check hypervisor in guest.
Steps to Reproduce:
1. enable feature in domain XML (see above)
2. virt-what in guest should return empty string
3. dmesg in guest should contain info about bare matal (instead of "on KVM")
[ 0.000000] Booting paravirtualized kernel on bare hardware
No way to hide KVM in qemu-kvm.
This is not a supported configuration by either Red Hat Enterprise Linux or Nvidia. Closing bug as WONTFIX.
"Red Hat Enterprise Linux 7 supports PCI device assignment of NVIDIA K-Series Quadro (model 2000 series or higher), GRID, and Tesla as non-VGA graphics devices."
That means you support NVIDIA GPU passthrough, but you don't care the GPU doesn't work?
Those supported models have no requirement to hide the hypervisor, the NVIDIA driver works with KVM exposed.
I have two different Tesla cards
1. NVIDIA Corporation GF110GL [Tesla M2090]
2. NVIDIA Corporation GK110BGL [Tesla K40c]
and first doesn't work without hiding KVM for sure.
Second one is currently running Debian 8 guest with old distribution driver 340.96-1, which isn't broken yet. I'll test some newer driver tomorrow.
Are these Teslas unsupported models?
(In reply to Vlastimil Holer from comment #9)
> I have two different Tesla cards
> 1. NVIDIA Corporation GF110GL [Tesla M2090]
> 2. NVIDIA Corporation GK110BGL [Tesla K40c]
> and first doesn't work without hiding KVM for sure.
> Second one is currently running Debian 8 guest with old distribution driver
> 340.96-1, which isn't broken yet. I'll test some newer driver tomorrow.
> Are these Teslas unsupported models?
According to our data from NVIDIA, only the second is supported, we have no information on M-series Teslas.
OK. If it's on your agreement with NVIDIA, I understand. On the other hand I don't see anything wrong against backporting the KVM hiding feature (e.g. for QA environments).
I'll check the second card first, but I have strong feeling, that it didn't work as well.
Also if there is a limited set of supported GPUs, it should be written somewhere.
Documentation and knowledgebase is missing these things
Anyway, thank you Alex for your comments.
(In reply to Alex Williamson from comment #10)
> According to our data from NVIDIA, only the second is supported, we have no
> information on M-series Teslas.
You are right, the second Tesla K40c works without hiding KVM.