Discrete Device Assignment (GPU-Passtrough)

[Edvvde 7]

Nabend zusammen,

 

habe hier einen Hyper-V Server 2019 auf dem ich DDA für eine VM umsetzen wollte.

 

Der Server hat ein Supermicro H11Dsi-NT Board und 2 EPYC Prozessoren verbaut. Als bisherige PCI Geräte gibt es einen RAID-Controller sowie eine 2x10Gbits Netzwerkkarte.

Hinzugekommen ist nun eine RTX 4000 ADA SFF welche via DDA durchgereicht werden soll. SR-IOV ist im BIOS Enabled.

 

Das SurveyDDA Skript meldet für die RTX folgendes:

Traffic from this device may be redirected to other devices in the system.  Not assignable.

 

Kann damit vielleicht jemand was anfangen?

[testperson 1.706]

Hi,

 

dazu findest du bspw. Discrete Device Assignment -- Machines and devices - Microsoft Community Hub:

 

Zitat

There’s another sort of entry that might come up, one where the report says that the device’s traffic may be redirected to another device. Here’s an example:

Intel(R) Gigabit ET Dual Port Server Adapter
Traffic from this device may be redirected to other devices in the system. Not assignable.

Intel(R) Gigabit ET Dual Port Server Adapter #2
Traffic from this device may be redirected to other devices in the system. Not assignable.

What this is saying is that memory reads or writes from the device targeted at the VM’s memory (DMA) might end up being routed to other devices within the computer. This might be because there’s a PCI Express switch in the system, and there are multiple devices connected to the switch, and the switch doesn’t have the necessary mechanism to prevent DMA from one device from being targeted at the other devices. The PCI Express specifications optionally allow all of a device’s traffic to be forced all the way up to the I/O MMU in the system. This is called “Access Control Services” and Hyper-V looks for that and enables it to be sure that your VM can’t affect others within the same machine.

These messages also might show up because the device is “multi-function” where that means that a single chip has more than one thing within it that looks like a PCI device. In the example above, I have an Intel two-port gigabit Ethernet adapter. You could theoretically assign one of the Ethernet ports to a VM, and then that VM could take control of the other port by writing commands to it. Again, the PCI Express specifications allow a device designer to put in controls to stop this, via Access Control Services (ACS).

The funny thing is that the NIC above has neither the ACS control structure in it nor the ability to target one port from the other port. Unfortunately, the only way that I know this is that I happened to have discussed it with the man at Intel who led the team that designed that NIC. There’s no way to tell in software that one NIC port can’t target the other NIC port. The official way to make that distinction is to look for ACS in the device. To deal with this, we allow you to override the ACS check when dismounting a PCI Express device. (Dismount-VMHostAssignableDevice -force)

 

HTH

Jan