How to use vfio-pci in vmware VM with PCI passthrough?

How to use vfio-pci in vmware VM with PCI passthrough?

[Oleksandr Nahnybida]

[-- Attachment #1: Type: text/plain, Size: 2546 bytes --]

Hi all,

We are working on migrating our DPDK application from igb_uio to vfio-pci.
Our target environment is a VMware ESXi host running on an AMD Epyc server
with NICs configured for PCI Passthrough to a guest VM with Debian Bookworm
(Kernel 6.1.0-39-amd64)

We've encountered a couple of issues.

Problem 1:

Initially, attempting to use vfio-pci failed with an error code of -22, and
the /sys/class/iommu/ directory was empty. We discovered the "expose IOMMU
to guest OS" option in VMware and enabled it.

This led to a new error:
"The virtual machine cannot be powered on because IOMMU virtualization is
not compatible with PCI passthru on AMD platforms"

We found a workaround by adding amd.iommu.supportsPcip = "TRUE" to the VM's
configuration. The VM now boots, and the IOMMU is visible in the guest.

However, when we run our DPDK application, it hangs after printing "EAL:
VFIO support initialized", and shortly after, the guest kernel panics with
a soft lockup error, making the system eventually unresponsive.
BUG: soft lockup - CPU#34 stuck for 75s! [kcompactd0:529]

Problem 2:

Separately, we've noticed that our IOMMU groups are not ideal. Many groups
contain not only the NICs we need to bind, but also other devices like PCI
bridges.

IOMMU Group 7:
        0000:00:17.0 - PCI bridge: VMware PCI Express Root Port
        0000:00:17.1
        0000:00:17.2
        0000:00:17.3
        0000:00:17.4
        0000:00:17.5
        0000:00:17.6
        0000:00:17.7
        0000:13:00.0 - nic
        0000:13:00.1 - nic
        0000:14:00.0 - nic
        0000:14:00.1 - nic

Questions:

   1.  Is enabling guest IOMMU virtualization in VMware with the
   amd.iommu.supportsPcip workaround, the correct approach here?
   2. I see that vfio-pci can be run in unsafe mode, but is there any
   benefit to using it over igb_uio in this case?
      1. In my understanding, the hypervisor is using actual hardware IOMMU
      to implement PCI pass-through anyway, so what's the point of having it
      inside of a guest VM again?
      2. Also, usually enable_unsafe_noiommu_mode is not compiled in, so we
      need to recompile vfio separately, and since we already compile igb_uio
      anyway, vfio won't be any better in terms of deployment.
   3. There also seems to be an option of using SR-IOV instead of
   passthrough, but we haven't explored this option yet. The question here is,
   do you still need to "expose iommu" to be able to bind VF to vfio? And
   what's the correct workflow here in general?

Best Regarads,
Oleksandr

[-- Attachment #2: Type: text/html, Size: 2897 bytes --]