Severe system freeze with NFS on Proxmox 9 running kernel 6.14.8-2-pve when mounting NFS shares

[gch3000]

I've been watching this thread for a while - glad people are making progress.

I'm running an OpenMediaVault VM using normal disk passthrough of a 4TB NVME and a couple of old sata hard drives. I'd seen that NFS client activity from an LXC to this VM's NFS server can lock things up. NFS traffic from an machine external to this VM seems solid. My less-that-ideal workaround was to have the LXC use sshfs to connect to OMV. If I can figure out how to move to PCI passthrough of the NVME without breaking things I probably will. I'm curious what people think about doing PCI passthrough of the SATA controller for the two hard drives. I currently don't NFS share these drives (they are just used for backup of the NVME drive). This would make the OMV instance look more like a physical machine (it would see SMART, control spindown of all the storage drives etc), but is there some downside I don't see?

 

[das1996]

Some more data points.

b550/3700x
NFS exports defined in a truenas vm
NFS client was the pve host itself
"cat /dev/random > filename.ext" in the exported folder to geneate nfs load
Monitoring tn pool status with "zpool iostat 1" in the tn vm

Installed prox 8.2 from scratch and retested.. same nfs lockups/disconnects with the virtual disk as well as 2.5" external hd via usb 3.0, kernel 6.8.4-2. I couldn't test with the hba as that was unavailable.

Upgrading 8.2 to the latest 8.4 with latest 6.8.12 (?) kernel was successful. No crashes writing to either the virtual disk or the usb drive. Wrote over 20GB using either method.

Upgraded to the latest pve 9.1 with 6.17.4 kernel. Lockup/disconnect with both external usb and virtual disk again. Box is still on 9.1. I so added the hba back in and retested. As expected it worked with the hba in passthrough. Retested virtual disk again too (with hba installed and passed through). This time it worked fine!@# Test a few more times, writing to the hba disk was 100% reliable. Writing to the virt disk with the hba passed through wasn't as reliable. It didn't crash/lock every time, but did once or twice.

With the hba removed from pass through, writing to the virt disk locked/disconnected each time.

Virtual disks (and phsyical disks passed through?) appears to be buggy when used in a vm based nas over nfs. So if one wants to use pve 9.1, stick to using an hba would be my advice. I'm done testing.

Given the above, I don't perceive I should encounter issues on the production box upgrading to 9.1 as it has the sata controllers (equiv to the hba) in pass through and there are no virtual disks or physical disks in passthrough. However, at this point, I think I will update to the latest 8.4 and leave it there until pve 9.2 or 9.3 rolls out. It's a single node server doing modest tasks. Nic pinning would be useful as it is headless and sometimes adding pci cards does change pci enumeration.