Skip to content

runtime: Add support for physical VMBus devices#447

Draft
Camelron wants to merge 27 commits into
msft-previewfrom
cameronbaird/msft-preview/physical-vmbus-2
Draft

runtime: Add support for physical VMBus devices#447
Camelron wants to merge 27 commits into
msft-previewfrom
cameronbaird/msft-preview/physical-vmbus-2

Conversation

@Camelron

Copy link
Copy Markdown

This change adds support for NICs of type VMBus in the physical_endpoint paths. VMBus devices cannot be bound to the vfio-pci driver however, so add handling in physical_endpoint such that PCI devices are passed through via VFIO as normal, but physical NICs of bus type vmbus are handled with a network pair.

The change in our fork includes a cherry-pick of upstream change to keep things aligned with upstream: kata-containers@6436922

Merge Checklist
  • Followed patch format from upstream recommendation: https://github.com/kata-containers/community/blob/main/CONTRIBUTING.md#patch-format
  • Included a single commit in a given PR - at least unless there are related commits and each makes sense as a change on its own.
  • Merged using "create a merge commit" rather than "squash and merge" (or similar)
  • genPolicy only: Builds on Windows
  • genPolicy only: Updated sample YAMLs' policy annotations, if applicable
Summary
Associated issues
Links to CVEs
Test Methodology

sprt and others added 27 commits March 18, 2026 13:32
Bug: https://microsoft.visualstudio.com/OS/_workitems/edit/43668151

Rationale: This is a temporary solution for optimizing memory usage for
the current mechanism of requesting resources through pod Limit
annotations:
- if no Limits are specified and hence WorkloadMemMB is 0, set a default
  value 'StaticWorkloadDefaultMem' to allocate a default amount of
  memory for use for containers in the sandbox in addition to the base
  memory
- if Limits are specified, the base memory and the sum of Limits are
  allocated. The end user needs to be aware of the minimum memory
  requirements for their pods, otherwise the pod will be stuck in the
  ContainerCreating state

Testing: Manual testing, creating pods with Limits and without limits,
and with two containers where each container has a limit, tested with
integration in a SPEC file where the config variables were set via
environment variables via the make command

Adapted by @mfrw from 3.1.0 to apply to 3.2.0

Signed-off-by: Muhammad Falak R Wani <mwani@microsoft.com>
Signed-off-by: Manuel Huber <mahuber@microsoft.com>

runtime: Remove unused VMM options for mem alloc

- We only ever tested these fork changes with CLH+MSHV
- Remove these options as we don't use QEMU/FC

Signed-off-by: Manuel Huber <mahuber@microsoft.com>
This branch starts introducing additional scripting to build, deploy
and evaluate the components used in AKS' Pod Sandboxing and
Confidential Containers preview features. This includes the capability
to build the IGVM file and its reference measurement file for remote
attestation.

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

tools: Improve igvm-builder and node-builder/azure-linux scripting

- Support for Mariner 3 builds using OS_VERSION variable
- Improvements to IGVM build process and flow as described in README
- Adoption of using only cloud-hypervisor-cvm on CBL-Mariner

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

tools: Add package-tools-install functionality

- Add script to install kata-containers(-cc)-tools bits
- Minor improvements in README.md
- Minor fix in package_install
- Remove echo outputs in package_build

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

tools: Enable setting IGVM SVN

- Allow setting SVN parameter for IGVM build scripting

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: introduce BUILD_TYPE variable

This lets developers build and deploy Kata in debug mode without having to make
manual edits to the build scripts.

With BUILD_TYPE=debug (default is release):

 * The agent is built in debug mode.
 * The agent is built with a permissive policy (using allow-all.rego).
 * The shim debug config file is used, ie. we create the symlink
   configuration-clh-snp-debug.toml <- configuration-clh-snp.toml.

For example, building and deploying Kata-CC in debug mode is now as simple as:

   make BUILD_TYPE=debug all-confpods deploy-confpods

Also do note that make still lets you override the other variables even after
setting BUILD_TYPE. For example, you can use the production shim config with
BUILD_TYPE=debug:

   make BUILD_TYPE=debug SHIM_USE_DEBUG_CONFIG=no all-confpods deploy-confpods

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>

node-builder: introduce SHIM_REDEPLOY_CONFIG

See README: when SHIM_REDEPLOY_CONFIG=no, the shim configuration is NOT
redeployed, so that potential config changes made directly on the host
during development aren't lost.

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>

node-builder: Use img for Pod Sandboxing

Switch from UVM initrd to image format

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Adapt README instructions

- Sanitize containerd config snippet
- Set podOverhead for Kata runtime class

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

tools: Adapt AGENT_POLICY_FILE path

- Adapt path in uvm_build.sh script to comply
  with the usptream changes we pulled in

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Use Azure Linux 3 as default path

- update recipe and node-builder scripting
- change default value on rootfs-builder

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Deploy-only for AzL3 VMs

- split deployment sections in node-builder README.md
- install jq, curl dependencies within IGVM script
- add path parameter to UVM install script

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Minor updates to README.md

- no longer install make package, is part of meta package
- remove superfluous popd
- add note on permissive policy for ConfPods UVM builds

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Updates to README.md

- with the latest 3.2.0.azl4 package on PMC, can remove OS_VERSION parameter
  and use the make deploy calls instead of copying files by hand for variant
  I (now aligned with Variant II)
- with the latest changes on msft-main, set the podOverhead to 600Mi

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: Fix SHIM_USE_DEBUG_CONFIG behavior

Using a symlink would create a cycle after calling this script again when
copying the final configuration at line 74 so we just use cp instead.

Also, I moved this block to the end of the file to properly override the final
config file.

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>

node-builder: Build and install debug configuration for pod sandboxing

For ease of debugging, install a configuration-clh-debug.toml for pod
sandboxing as we do in Conf pods.

Signed-off-by: Cameron Baird <cameronbaird@microsoft.com>

runtime: remove clh-snp config file usage in makefile

Not needed to build vanilla kata

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>

package_tools_install.sh: include nsdax.gpl.c

Include nsdax.gpl.c

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>

node-builder: fix typo in string comparison

This also fixes a shellcheck error and lets us require the
shellcheck-required job:

In ./tools/osbuilder/node-builder/azure-linux/uvm_build.sh line 34:
        if [ -z "${UVM_KERNEL_HEADER_DIR}}" ]; then
                                         ^-- SC2157 (error): Argument to -z is always false due to literal strings.

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>

docs: node-builder: fix static check error

This fixes the below static check error to follow up on the infra fix from
kata-containers#11646:

2025-07-31T19:32:45.0031829Z time="2025-07-31T19:32:44.990004665Z" level=fatal msg="found 2 parse errors:\nfile=\"tools/osbuilder/node-builder/azure-linux/README.md\": duplicate heading: \"Set up environment\" (heading: {Name:Set up environment MDName:Set up environment LinkName:set-up-environment Level:2})\nfile=\"tools/osbuilder/node-builder/azure-linux/README.md\": duplicate heading: \"Install build dependencies\" (heading: {Name:Install build dependencies MDName:Install build dependencies LinkName:install-build-dependencies Level:2})" commit=1d17f56b1aa7a880468b8e25d14467c92dca8eeb name=kata-check-markdown pid=9075 source=check-markdown version=0.0.1

Note: that is likely flagged because having two headings with the same
name, even under different sections, makes it impossible to create a
canonical heading link in Markdown.

This should eventually be squashed into the node-builder commit.

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>

docs: node-builder: Remove references to moby-containerd-cc

As we adopted containerd2, we remove references to our prior
forked containerd version.

Signed-off-by: Manuel Huber <mahuber@microsoft.com>

node-builder: 2Mb aligned guest image size

Build the mariner guest image using IMAGE_SIZE_ALIGNMENT_MB=2.

Signed-off-by: Dan Mihai <dmihai@microsoft.com>

to-squash: node-builder: add reference to README.md

This is needed to avoid the following static-checks error:

2025-08-05T21:27:20.0028337Z [static-checks.sh:808] ERROR: Document tools/osbuilder/node-builder/azure-linux/README.md is not referenced

This commit is to be squashed into the node-builder commit.

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>
After these changes:

1. The value of the K8s runtime class memory overhead:
   - Covers the memory usage from all the Host-side components (mainly
     the Kata Shim and the VMM).
   - Doesn't include the memory usage from any Guest-side components.

2. The value of a pod memory limit specified by the user:
   - Is equal to the memory size of the Pod VM.
   - Includes the memory usage from all the Guest-side components
     (mainly user's workload, the Guest kernel, and the Kata Agent)
   - Doesn't include the memory usage from any Host-side components.

Signed-off-by: Dan Mihai <dmihai@microsoft.com>

runtime: fix `make test`

This addresses the following errors from `make test` to allow us to require
that upstream CI:

https://github.com/microsoft/kata-containers/actions/runs/16656407213/job/47142422035?pr=392#step:13:53

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>
- similar to the static_sandbox_default_workload_mem option,
  assign a default number of vcpus to the VM when no limits
  are given, 1 vcpu in this case
- similar to commit c7b8ee9, do not allocate additional vcpus
  when limits are provided

Signed-off-by: Manuel Huber <mahuber@microsoft.com>
Point to msft-preview

Signed-off-by: Manuel Huber <mahuber@microsoft.com>
Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
For our Kata UVM, we know we need at least 128MB of memory to prevent instability in the guest.

Enforce this constraint with a descriptive error to prevent users from destabilizing the UVM with faulty k8s configurations.

Signed-off-by: Cameron Baird <cameronbaird@microsoft.com>
If memory limit is set and less than minimum, set it to minimum.

This is to to account for kata-containers@0ec3403

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
Add Microsoft mandatory file SECURITY.md

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
- Change Makefile to point to fork
- Change versions.yaml to point to proper version on fork

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
This change mirrors host networking into the guest as before, but now also
includes the default gateway neighbor entry for each interface.

Pods using overlay/synthetic gateways (e.g., 169.254.1.1) can hit a
first-connect race while the guest performs the initial ARP. Preseeding the
gateway neighbor removes that latency and makes early connections (e.g.,
to the API Service) deterministic.

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
This is a fork temporary measure to unblock CI required tests in our fork,
while we find a way to remove the 'main' hard codes from upstream.

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
Background:

 * `pull_request` runs on the PR branch code and has access to secrets
   ONLY if the PR is from microsoft/kata-containers (i.e. NOT from an external
   contributor who forked the repo).
 * `pull_request_target` runs on the trusted main branch code by default
   and has access to secrets for any PR.

Reference: https://docs.github.com/en/actions/reference/workflows-and-actions/events-that-trigger-workflows#pull_request

Upstream uses `pull_request_target` (and manually checks out the PR code)
to have access to secrets for PRs from external contributors, however we
don't expect external PRs, hence we can use `pull_request`.

Furthermore, since `pull_request_target` only runs from the default branch,
we need to use `pull_request` anyway as we have multiple leading branches
(i.e., msft-main, msft-preview, and release branches).

https://github.blog/changelog/2025-11-07-actions-pull_request_target-and-environment-branch-protections-changes/

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>
set default to msft-preview

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
use upstream cloud-hypervisor.

This is to unblock the CI and let CLH build

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
update target branch to msft-preview

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
This fixes a CI static check failure

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
- tests that deploy pods with too small of a memory limit
 - try to set a minimum memory limit for some containerd tests
- tests that use runners we don't have
- tests that depend on pushing to GHCR

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
Enable VFIO device pass-through at VM creation time on Cloud Hypervisor,
in addition to the existing hot-plug path.

Signed-off-by: Roaa Sakr <romoh@microsoft.com>
Regenerate CH client against v51.1

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
The recently-added nested property is true by default, but is not
supported yet on MSHV.

See cloud-hypervisor/cloud-hypervisor#7408 for additional information.

Signed-off-by: Dan Mihai <dmihai@microsoft.com>
This cloud-hypervisor is a directory, so it needs "rm -rf" instead of
"rm -f".

Signed-off-by: Dan Mihai <dmihai@microsoft.com>
disable Kata Containers CI / kata-containers-ci-on-push / run-kata-deploy-tests / run-kata-deploy-tests (qemu, k3s)

Signed-off-by: Saul Paredes <saulparedes@microsoft.com>
`cargo check` was introduced in 3f1533a to check that Cargo.lock is in sync
with Cargo.toml. However, if there are uncommitted changes in the working
tree, the current invocation will immediately fail because of the `git
diff` call, which is frustrating for local development.

As it turns out, `cargo clippy` is a superset of `cargo check`, so we can
simply pass `--locked` to `cargo clippy` to detect Cargo.lock issues.

This is tested with the following change:

diff --git a/src/agent/Cargo.lock b/src/agent/Cargo.lock
index 96b6c67..e1963af 100644
--- a/src/agent/Cargo.lock
+++ b/src/agent/Cargo.lock
@@ -4305,6 +4305,7 @@ checksum = "8f50febec83f5ee1df3015341d8bd429f2d1cc62bcba7ea2076759d315084683"
 name = "test-utils"
 version = "0.1.0"
 dependencies = [
- "libc",
  "nix 0.26.4",
 ]

which results in the following output:

$ make -C src/agent check
make: Entering directory '/kata-containers/src/agent'
standard rust check...
cargo fmt -- --check
cargo clippy --all-targets --all-features --release --locked \
        -- \
        -D warnings
error: the lock file /kata-containers/src/agent/Cargo.lock needs to be updated but --locked was passed to prevent this
If you want to try to generate the lock file without accessing the network, remove the --locked flag and use --offline instead.
make: *** [../../utils.mk:184: standard_rust_check] Error 101
make: Leaving directory '/kata-containers/src/agent'

Signed-off-by: Aurélien Bombo <abombo@microsoft.com>
Signed-off-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Interfaces whose drivers do not register a specific netlink kind
(e.g. mlx5 Scalable Functions) are reported with the generic type
"device". The endpoint creation code did not handle this type,
causing sandbox creation to fail with:

  "Unsupported network interface: device"

This is particularly visible on arm64 with Mellanox ConnectX NICs
using Scalable Functions, where the ethtool BusInfo returns a
non-PCI identifier (e.g. "mlx5_core.sf.4") so isPhysicalIface()
cannot classify the interface as physical either.

Handle "device" type interfaces the same way as veth endpoints,
connecting them through a TAP + TC-filter bridge.

Additionally, relax getLinkForEndpoint() for VethEndpoint so it
accepts the concrete link type returned by the kernel instead of
asserting *netlink.Veth. A "device" type interface wrapped in a
VethEndpoint returns *netlink.Device from LinkByName(), which
would fail the strict type assertion. All callers only need
link.Attrs(), so accepting any link type is safe.

Signed-off-by: Fabiano Fidêncio <ffidencio@nvidia.com>
kata-runtime can only use physical NIC devices on the host that
support VFIO passthrough. For vmbus devices which don't support
VFIO passthrough, instead set up a tap/bridge pair.

Signed-off-by: Cameron Baird <cameronbaird@microsoft.com>
runtime-rs can only use physical NIC devices on the host that support
VFIO passthrough. For certain devices which don't support
VFIO passthrough, instead set up a tap/bridge pair.

Signed-off-by: Cameron Baird <cameronbaird@microsoft.com>
Copilot AI review requested due to automatic review settings May 11, 2026 18:18

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR extends Kata’s “physical endpoint” networking support to handle Hyper-V VMBus-backed NICs, which cannot be passed through via vfio-pci. It introduces a split behavior: PCI-backed physical NICs continue to use VFIO passthrough, while VMBus-backed physical NICs use a tap/bridge “network pair” (similar to veth endpoints). Parallel changes are made in both the Go runtime and runtime-rs to keep the implementations aligned.

Changes:

  • Detect physical NIC bus type (PCI vs VMBus) and choose VFIO passthrough vs tap-backed network pair accordingly.
  • Persist additional physical-endpoint state (network pair, bus type, VFIO flag) and update teardown / hotplug logic to support both paths.
  • Add unit tests and runtime-rs helpers for bus-type detection and physical endpoint construction.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 5 comments.

Show a summary per file
File Description
src/runtime/virtcontainers/physical_endpoint.go Adds VFIO vs non-VFIO branching for physical endpoints; introduces bus-type detection, persistence fields, and tap-backed attach/detach for non-VFIO.
src/runtime/virtcontainers/physical_endpoint_test.go Adds tests for VFIO/non-VFIO attach/detach paths, bus-type helpers, and new persistence fields.
src/runtime/virtcontainers/persist/api/network.go Extends persisted PhysicalEndpoint with NetPair, BusType, and IsVFIO.
src/runtime/virtcontainers/network_linux.go Resolves missing netlink.Link for hotplug, routes physical endpoints through new createPhysicalEndpoint signature, and improves link handling for “device” netdev type.
src/runtime-rs/crates/resource/src/network/utils/link/mod.rs Adds bus-type detection and sysfs path resolution for PCI vs VMBus NICs.
src/runtime-rs/crates/resource/src/network/network_with_netns.rs Updates physical interface detection and passes additional params to PhysicalEndpoint::new().
src/runtime-rs/crates/resource/src/network/network_pair.rs Adds new_for_physical() helper to build a minimal/stub network pair for VFIO physical endpoints.
src/runtime-rs/crates/resource/src/network/endpoint/physical_endpoint.rs Implements VFIO vs tap-backed attach/detach for physical endpoints; wires in network pair and bus type.
src/runtime-rs/crates/resource/src/network/endpoint/endpoint_persist.rs Extends PhysicalEndpointState with is_vfio, iface_name, and bus_type fields.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

endpoint.Driver = s.Physical.Driver
endpoint.VendorDeviceID = s.Physical.VendorDeviceID
endpoint.BusType = s.Physical.BusType
endpoint.IsVFIO = s.Physical.IsVFIO
Comment on lines +218 to +226
return nil
}
span, ctx := vethTrace(ctx, "HotDetach", endpoint)
defer span.End()
if err := doNetNS(netNsPath, func(_ ns.NetNS) error {
return xDisconnectVMNetwork(ctx, endpoint)
}); err != nil {
networkLogger().WithError(err).Warn("Error un-bridging virtual ep")
}
Comment on lines 95 to +101
// Attach for physical endpoint binds the physical network interface to
// vfio-pci and adds device to the hypervisor with vfio-passthrough.
func (endpoint *PhysicalEndpoint) Attach(ctx context.Context, s *Sandbox) error {
span, ctx := physicalTrace(ctx, "Attach", endpoint)
defer span.End()

// Unbind physical interface from host driver and bind to vfio
// so that it can be passed to qemu.
vfioPath, err := bindNICToVFIO(endpoint)
if err != nil {
return err
}

c, err := resCtrl.DeviceToCgroupDeviceRule(vfioPath)
if err != nil {
if endpoint.IsVFIO {
// Unbind physical interface from host driver and bind to vfio
Comment on lines +160 to +161
isFVIODisabled := (s.config.HypervisorConfig.ColdPlugVFIO == config.NoPort)
endpoint, err = createPhysicalEndpoint(idx, netInfo, isFVIODisabled, n.interworkingModel)
Comment on lines +344 to +354
endpoint := &PhysicalEndpoint{
IfaceName: "eth0",
HardAddr: "aa:bb:cc:dd:ee:ff",
EndpointType: PhysicalEndpointType,
BDF: "0000:01:00.0",
Driver: "mlx5_core",
VendorDeviceID: "0x8086 0x1572",
IsVFIO: false,
NetPair: netPair,
BusType: "pci",
}
@sprt sprt marked this pull request as draft May 11, 2026 21:22
@sprt

sprt commented May 11, 2026

Copy link
Copy Markdown

@Camelron Could you link to the upstream PR in this PR's description?

Comment thread src/runtime/virtcontainers/network_linux.go
Comment thread src/runtime/virtcontainers/network_linux.go
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants