Mellanox / network-operator / 35207940122

64%

Build Type

push

github

Committed by web-flow

Commit Message

feat: configure RDMA kubelet root via helm staticconfig (#3189)

## Summary

Configure an optional kubelet root for the RDMA shared device plugin
using the same helm / env / staticconfig flow as `CNI_BIN_DIR` /
`CNI_NETWORK_DIR`, instead of expanding the CRD API.

- Helm: `operator.kubeletRootDir` (empty by default)
- Operator Deployment env: `KUBELET_ROOT_DIR` (only when the helm value
is set)
- `staticconfig.StaticConfig.KubeletRootDir` consumed by the RDMA shared
device plugin state
- DaemonSet: hostPath / mountPath for `device-plugins` and
`plugins_registry` derived from that root; `--kubelet-root-dir` only
when the override is non-empty and not the default `/var/lib/kubelet`
- Unset preserves current `/var/lib/kubelet` mounts and omits the CLI
flag
- No CRD / webhook API field

Depends on upstream plugin support:
Mellanox/k8s-rdma-shared-dev-plugin#157

Fixes #2989

## Test plan

- [x] `make generate` / `make manifests`
- [x] `make lint` (0 issues)
- [x] `make lint-helm` / `make test-helm`
- [x] full `make test` (envtest; exit 0)
- [x] `make build`
- [x] helm template: default omits `KUBELET_ROOT_DIR`;
`operator.kubeletRootDir=/var/lib/k0s/kubelet` sets it
- [x] unit coverage for default / override / trailing-slash / combined
`--use-cdi`
- [ ] Kind nested cluster / live RDMA pods (no `pci-15b3` hardware here)
- [ ] NVIDIA GH Actions runners (waiting on maintainer `/ok-to-test` via
copy-pr-bot)

## Notes for reviewers

Addressing @rollandf feedback to follow the CNI_BIN_DIR /
CNI_NETWORK_DIR pattern rather than adding `kubeletRootDir` on the API.

Coverage Stats

25 of 32 new or added lines in 4 files covered. (78.13%)

70 existing lines in 7 files now uncovered.

5049 of 8058 relevant lines covered (62.66%)

45.33 hits per line