4:26 p.m.
The Grace SOC introduces Extended GPU Memory (EGM) [3], a feature that enables
GPUs to efficiently access system memory within and across nodes. This patch
series adds support for virtualizing EGM (vEGM) in libvirt, allowing VMs to
utilize dedicated EGM memory regions through ACPI.
RFC Status
==========
This patch series is submitted as an RFC to gather feedback from the libvirt
community on the overall approach and implementation details. While kernel EGM
driver support and QEMU acpi-egm-memory device support are not yet upstream,
reference implementations are available [1][2] to enable testing and validation
of the libvirt integration.
Any community feedback is appreciated.
Background and Use Cases
=========================
EGM allows host memory to be partitioned into two regions:
1. Standard memory for Host OS usage
2. EGM region assigned to VMs as their system memory
This technology enables various high-performance computing scenarios [3]:
- Large memory pools for AI/ML workloads
- High-performance computing applications
- Memory extension for systems with limited main memory
- GPU-accelerated workloads requiring large addressable memory
Implementation Overview
=======================
This 8-patch series adds a new device type 'acpi-egm-memory' with the following
structure:
1. Schema definition - Add XML schema definition for the new ACPI EGM memory device
2. XML parsing - Implement XML parsing and internal data structures
3. Capability detection - Add QEMU capability detection for EGM support
4. Validation - Add validation logic for EGM device configuration
5. Command generation - Implement QEMU command line generation
6. Resource management - Setup required cgroup and namespace configurations
7. Documentation - Add comprehensive documentation
8. Testing - Add qemuxmlconftest for ACPI EGM memory device
XML Configuration
=================
Example usage in domain XML:
<devices>
<hostdev mode="subsystem" type="pci" managed="yes">
<alias name="ua-hostdev0"/>
<source>
<address domain="0x0000" bus="0x01" slot="0x00"
function="0x0"/>
</source>
</hostdev>
<acpiEgmMemory>
<alias name="egm0"/>
<pciDev>ua-hostdev0</pciDev>
<numaNode>0</numaNode>
</acpiEgmMemory>
</devices>
This configuration results in appropriate QEMU command line options:
-object memory-backend-file,id=m0,mem-path=/dev/egm0,size=32G,share=on,prealloc=on
-object acpi-egm-memory,id=egm0,pci-dev=ua-hostdev0,node=0
Implementation Notes
====================
The device validation includes checking that referenced PCI devices exist,
NUMA nodes are valid, and device paths are accessible with proper permissions.
Memory backing is configured automatically to use the EGM device path, and
cgroups/namespaces are set up to allow safe access.
Testing
=======
I've tested XML parsing, validation, and qemu command line generation. The
qemuxmlconftest passes for XML handling, though the command generation test
currently fails since QEMU doesn't have acpi-egm-memory support yet, but
the generated args look correct for when it does.
Requirements
============
This feature requires:
- NVIDIA ARM64 Grace platform with EGM support
- Host kernel with EGM driver support [1]
- QEMU with ACPI EGM device support [2]
[1] https://github.com/ianm-nv/NV-Kernels/tree/6.8_ghvirt_egm_may2025
[2] https://github.com/ianm-nv/qemu/tree/6.8_ghvirt_egm_may2025
[3]
https://developer.nvidia.com/blog/nvidia-grace-hopper-superchip-architect...
Ian May (8):
conf: Add schema definition for ACPI EGM memory device
conf: Add definitions and XML parsing for ACPI EGM memory device
qemu: Add capability detection for ACPI EGM memory device
qemu: Add validation for ACPI EGM memory device configuration
qemu: Add command line generation for ACPI EGM memory device
qemu: Add cgroup and namespace setup for ACPI EGM memory device
docs: Document ACPI EGM memory device
tests: Add qemuxmlconftest for ACPI EGM memory device
docs/formatdomain.rst | 80 ++++++++++++++
src/ch/ch_domain.c | 1 +
src/conf/domain_conf.c | 102 ++++++++++++++++++
src/conf/domain_conf.h | 11 ++
src/conf/domain_postparse.c | 8 ++
src/conf/domain_validate.c | 23 ++++
src/conf/schemas/domaincommon.rng | 19 ++++
src/conf/virconftypes.h | 2 +
src/libxl/libxl_driver.c | 6 ++
src/lxc/lxc_driver.c | 6 ++
src/qemu/qemu_capabilities.c | 2 +
src/qemu/qemu_capabilities.h | 1 +
src/qemu/qemu_cgroup.c | 21 ++++
src/qemu/qemu_command.c | 37 +++++++
src/qemu/qemu_domain.c | 2 +
src/qemu/qemu_domain_address.c | 2 +
src/qemu/qemu_driver.c | 3 +
src/qemu/qemu_hotplug.c | 5 +
src/qemu/qemu_namespace.c | 21 ++++
src/qemu/qemu_postparse.c | 1 +
src/qemu/qemu_validate.c | 99 +++++++++++++++++
src/test/test_driver.c | 4 +
tests/meson.build | 1 +
.../caps_10.0.0_aarch64.xml | 1 +
tests/qemuegmmock.c | 67 ++++++++++++
.../acpi-egm-memory.aarch64-latest.args | 1 +
.../acpi-egm-memory.aarch64-latest.xml | 56 ++++++++++
tests/qemuxmlconfdata/acpi-egm-memory.xml | 27 +++++
tests/qemuxmlconftest.c | 5 +-
29 files changed, 613 insertions(+), 1 deletion(-)
create mode 100644 tests/qemuegmmock.c
create mode 100644 tests/qemuxmlconfdata/acpi-egm-memory.aarch64-latest.args
create mode 100644 tests/qemuxmlconfdata/acpi-egm-memory.aarch64-latest.xml
create mode 100644 tests/qemuxmlconfdata/acpi-egm-memory.xml
--
2.43.0