Nvidia Quadro P400 not working on Ubuntu server via GPU/PCIe passthrough

olivierlambert

Okay so at least "it's good news": it's not a regression So I assume it's required to hide the hypervisor underneath.

TheFrisianClause

@olivierlambert Yes I believe so, do you need more information on this?

olivierlambert

Not yet, I need to talk with Xen devs when I can on how to achieve that.

TheFrisianClause

@olivierlambert Alright, thanks for the effort

warriorcookie

@olivierlambert just wanted to clarify a couple things:
-nvidia drivers support passthrough recently on windows guests only. Linux guests are still unsupported.
-while most quadro cards are supported for passthrough including P4000, the P400 is excluded from support. It's considered consumer grade (GP107 basically a striped down gtx1050)
-might be a given but for those that stumble on this thread, both the video device and audio device need to be passed through separately as they show as separate devices under lspci.

I'm confident this is an issue with the nvidia driver disabling when it see's the CPUID hypervisor present bit.

On my hardware (SM x9 with e5-2667v2) proxmox and esxi both worked flawlessly with near bare metal performance but you're able to change the CPUID hypervisor present bit.

On xcp-ng 8.2 with windows guest it works with no issues.
Once I install Linux as the guest then smi-nvidia produces "no devices were found" even though it shows with lspci.

I'm trying to migrate from esxi. I'm jumping in on this thread as this is a major issue for me that would prevent me switching. Unfortunate as this project ticks all the other boxes that vmware failed to....

olivierlambert

Okay so it's "just" a matter of hiding the CPUID then Thanks a lot for the detailed feedback @warriorcookie !

Let me push again on that feature request on my side, I'll see what we can do (and how much time/money to get a solution)

warriorcookie

Parralel work into what I think could be the solution, but stalled as of now: https://xcp-ng.org/forum/topic/4643/nested-virtualization-of-windows-hyper-v-on-xcp-ng/39

TheFrisianClause

@warriorcookie Well this is a nice and detailed post, never knew about the CPUID though so I think that could be the problem here.

olivierlambert

It is, but it doesn't answer how much effort is needed to "solve it". ATM, there's no way to change it in Xen. So the great question is the modification scope required

warriorcookie

@olivierlambert said in Nvidia Quadro P400 not working on Ubuntu server via GPU/PCIe passthrough:

It is, but it doesn't answer how much effort is needed to "solve it". ATM, there's no way to change it in Xen. So the great question is the modification scope required

My Pa always said "don't bring me a dead cat without a shovel..."

Sorry, I'm all cats and no shovel today!

olivierlambert

Maybe a better approach would be to modify Nvidia drivers to change or remove the check. After all, it should be only a kind of grep on the word Xen

And obviously, this would be also against the EULA.

TheFrisianClause

Apparently this seem to be an bigger issue than thought or am I wrong thinking that?

warriorcookie

@olivierlambert said in Nvidia Quadro P400 not working on Ubuntu server via GPU/PCIe passthrough:

Maybe a better approach would be to modify Nvidia drivers to change or remove the check. After all, it should be only a kind of grep on the word Xen

And obviously, this would be also against the EULA.

I've found this but have not had time to play: https://github.com/DualCoder/vgpu_unlock

olivierlambert

So the most advanced discussion within Xen about it, is here: https://lists.archive.carbon60.com/xen/devel/439501

Nothing happened since sadly, but it might be a start. Also, hiding Xen will probably have a lot of consequences regarding performances (no PV drivers and so on)

TheFrisianClause

@olivierlambert hmmm seems something, but why can this so easily be accomplished with Proxmox and even ESXi (Which I thought would be a pain considering the age of the P400). So if I for example would buy an nvidia K4000 I would be able to succesfully passthrough the GPU to Plex? (I don't know if you have done that)

olivierlambert

Because… they are totally different projects? ESXi got thousand engineers working on it. KVM model is different and less isolated than Xen and ESXi, so you can make important modifications in it more easily.

TheFrisianClause

@olivierlambert Yeah that they are different projects is obvious, but my question was more like 'what makes it so hard to do this in Xen?'

olivierlambert

True Type-1 hypervisor (like ESXi, unlike KVM) makes it more isolated but harder to do things in general
It's as hard in ESXi, but resources on the hypervisor are 2 or 3 order of magnitude higher than for the Xen project.

Obviously, we are working hard here at Vates to get more people directly involved in the Xen project. But it takes time and a vast amount of money to reach our target Anyway, I'll try to see what I can do with our resources. The main issue for me now on this feature: it's mainly for non-pro usage, so no company will finance that.

warriorcookie

@olivierlambert said in Nvidia Quadro P400 not working on Ubuntu server via GPU/PCIe passthrough:

True Type-1 hypervisor (like ESXi, unlike KVM) makes it more isolated but harder to do things in general

It's as hard in ESXi, but resources on the hypervisor are 2 or 3 order of magnitude higher than for the Xen project.

Obviously, we are working hard here at Vates to get more people directly involved in the Xen project. But it takes time and a vast amount of money to reach our target Anyway, I'll try to see what I can do with our resources. The main issue for me now on this feature: it's mainly for non-pro usage, so no company will finance that.

I certainly appreciate the challange, and I wish I had something to offer to help development wise.
Perhaps a more "pro" use case could be from the standpoint of nested VM with the likes of HyperV?

olivierlambert

I think I never heard of any company/industry requiring the possibility to hide the hypervisor for now (doesn't mean it doesn't exists, though).

It's all a matter of priorities and limited resources sadly… That's why taking the problem from the other angle (ie drivers that doesn't check it) might be a correct solution.