Skip to content

Switch to QEMU auto SMMU property detection mode #43

Description

@lyarwood

Summary

Switch from hardcoded SMMU properties to the QEMU auto SMMU property detection mode, so that the SMMU configuration is automatically inferred from the host and we reduce static configuration drift.

Background

A QEMU feature to auto-detect SMMU properties from the host is under upstream review:
https://lore.kernel.org/all/20260422204335.23116-1-nathanc@nvidia.com/

This is planned for the QEMU 26.03 (Voyager) release. Once available, kubevirt-aie can switch to auto mode instead of hardcoding the following values:

  • OAS hardcoded to 48
  • Ril set to off
  • Accel set to on

The current hardcoded values are set in:
https://github.com/kubevirt/kubevirt-aie/blob/release-1.8-aie-nv/pkg/virt-launcher/virtwrap/converter/pci-placement.go#L371

Acceptance Criteria

  • When auto SMMU mode is enabled, OAS, Ril, and Accel are no longer hardcoded
  • SMMU properties are correctly inferred from the host SMMUv3 configuration
  • GPU passthrough continues to work on Grace Blackwell platforms (GB200, GB300) with auto-detected SMMU properties
  • One SMMUv3 device per PCIe expander bus is still correctly configured

Dependencies

  • Upstream QEMU auto SMMU property series must land (targeted for 26.03 Voyager release)
  • Eric Auger (eauger@redhat.com) will notify when the feature is available

Jira

https://redhat.atlassian.net/browse/CNV-86646

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions