NVidia GPU / Mapped device 'Parameter verification failed (400)'

[dooferorg]

I keep getting the error on only *some* of my nodes when trying to edit the PCIe resource setting on a VM that 'Parameter verification failed (400)'.


I have resource mapping set up across all nodes that have GPUs (Tesla P4s), they are all the same GPU and since they're on the same slot on the same model computer the PCIe id happens to be identical.





I deleted all the mappings and redid them, still the same error. The really weird thing is too I can migrate the VM between hosts and select the 'force' to just tell it to get on with it despite the warning that the mapping doesn't work. The VM will start up just fine.

So it seems like it's a bug in Proxmox that it's just not working well in 'Edit: PCI Device' dialog to *actually* read the mappings that are there. I don't know what regex it's talking about (is that stored anywhere because then I could see if there was an error in it). As it stands the error message is useless.

So, what's the best way to sort this out so that I can use the mapping dialog once again?

This is in PVE 8.4.1

 

[dcsapak]

it's probably due to some nodes not on the most current version?
the api to query the models recently gained the ability to use a mapping in addition to give a pciid directly, so my guess is that when you get the error those nodes are not on that version yet

you can verify that by comparing the output
of

pveversion -v

of all your nodes

 

[dooferorg]

Ah, yes, you're right. The nodes in question are on 8.3 still, rather than the 8.4.1 that the node I am using for the interface was using.

It'd be nice if that was highlighted somewhere about different PVE versions not matching. You really have to dig for it. It's also not apparent from the error message that that is the problem. I'd expect that if the UI is trying to do an operation on an older PVE node that it would realize that and not do something that breaks. At least, that'd be my expectation as a user.

As a feature request it would be nice at the 'Datacenter' level to have an option to 'normalize the PVE versions' and pick a version that then can update all the nodes so they're on the same one, if incompatibilities are going to be introduced between releases.