Hello,
As is probably known, there are quite a few issues with reliable PCI passthrough.
We've been able to get PCI passthrough working well, with one caveat: if GPU 1 is assigned to VM A, and is then re-assigned to VM B, it sometimes does not come up properly. The error we get within the VM is that the GPU has "fallen off the bus".
We have tried many things, but the only way to reliably resolve this is to run a pre-start hookscript that removes the PCI device and re-discovers it.
We use a resource mapping to simplify the VM configs, and to determine which device has to be removed and re-discovered, we use the "qm showcmd <vmid>" command.
However, we found out this does not work reliably. the showcmd command actually always appears to show the first PCI address of the resource mapping. And if >1 GPU is present it would show the first addresses of the GPU resource mapping.
This appears to be caused by this change: https://www.mail-archive.com/pve-devel@lists.proxmox.com/msg27318.html
This appears to be an intended change, but I don't know if my usecase was considered when the change was made.
Is there a different way to determine the PCI address(es) that will be used when starting a VM using a resource mapping? Or is this actually an unintended side-effect?
Alternatively we could remove the resource mapping of course, but if possible we would like to avoid doing that (for a simpler VM configuration)
Thank you
Niels
As is probably known, there are quite a few issues with reliable PCI passthrough.
We've been able to get PCI passthrough working well, with one caveat: if GPU 1 is assigned to VM A, and is then re-assigned to VM B, it sometimes does not come up properly. The error we get within the VM is that the GPU has "fallen off the bus".
We have tried many things, but the only way to reliably resolve this is to run a pre-start hookscript that removes the PCI device and re-discovers it.
We use a resource mapping to simplify the VM configs, and to determine which device has to be removed and re-discovered, we use the "qm showcmd <vmid>" command.
However, we found out this does not work reliably. the showcmd command actually always appears to show the first PCI address of the resource mapping. And if >1 GPU is present it would show the first addresses of the GPU resource mapping.
This appears to be caused by this change: https://www.mail-archive.com/pve-devel@lists.proxmox.com/msg27318.html
This appears to be an intended change, but I don't know if my usecase was considered when the change was made.
Is there a different way to determine the PCI address(es) that will be used when starting a VM using a resource mapping? Or is this actually an unintended side-effect?
Alternatively we could remove the resource mapping of course, but if possible we would like to avoid doing that (for a simpler VM configuration)
Thank you
Niels
