Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

[Bug]: container-toolkit DaemonSet: containerd-socket hostPath volume lacks mountPropagation: HostToContainer, pins host ZFS mounts and blocks zfs destroy ("dataset is busy")

オープン
  • 何か問題がありますか？
#3,083 コメント 1 件 リアクション 1 件 担当者 1 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@rahulait がすでに取り組んでいます。

2026年10月19日 から。

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
45/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
go, kubernetes
領域
devops, infrastructure

調査の方向性

The fix sits in the toolkit DaemonSet's volume definitions, specifically the containerd-socket hostPath volume, which lacks mountPropagation: HostToContainer. Find where the operator generates that DaemonSet and compare it with the host-root and driver-install-dir volumes. To reproduce, mount a ZFS dataset, restart the toolkit pod, unmount, then check /proc/<pid>/mountinfo and zfs destroy. Done when the destroy succeeds without a container restart and the setting survives operator reconciliation.

索引モデルが issue の本文から書いたものです。

説明

bug needs-triage

Problem

The nvidia-container-toolkit-daemonset in GPU Operator v26.7.1 mounts these hostPath volumes:

volume hostPath mountPath mountPropagation
host-root / /host HostToContainer
driver-install-dir /run/nvidia/driver /driver-root HostToContainer
containerd-socket /run/containerd /runtime/sock-dir/ not set (private)

The toolkit container's own rootfs lives under /run/containerd. Its /host sub-mount carries every host mount, including all ZFS datasets. Because containerd-socket is a private recursive bind, the whole mount tree is copied into the container's mount namespace when the container starts. After that, the copy never receives unmount events from the host.

Impact

  • Any ZFS dataset that is mounted on the host when the toolkit container starts stays pinned in the container's mount namespace.
  • Later, the host unmounts the dataset, and it reports mounted=no. But zfs destroy still fails with cannot destroy '': dataset is busy.
  • The dataset stays undeletable until the toolkit container restarts. The restart re-pins every dataset that exists at that moment.
  • This affects every container restart, including node reboots. Datasets created after the container starts are not affected.

Evidence

After the dataset is unmounted on the host, its mount entries are still found in the toolkit process's mount namespace:
$ grep -l /proc/[0-9]*/mountinfo
/proc//mountinfo # comm=nvidia-toolkit
.../runtime/sock-dir/io.containerd.runtime.v2.task/k8s.io//rootfs/host/ - zfs

Environment

  • NVIDIA GPU Operator v26.7.1 (Helm chart gpu-operator-v26.7.1)
  • OS: Ubuntu 24.04
  • Kernel Version: 6.8.0-142-generic
  • Container Runtime Version: containerd v2.1.4
  • Container-toolkit image: nvcr.io/nvidia/k8s/container-toolkit:v1.20.1
  • GPU Driver: 580.105.08
  • Host filesystems: ZFS datasets mounted on the host
  • Kubernetes Distro and Version: Kubelet v1.33.5

Steps to reproduce

  1. Create and mount a ZFS dataset on a GPU node.
  2. Restart the nvidia-container-toolkit pod on that node, or reboot the node.
  3. Unmount the dataset on the host and run zfs destroy -r .
  4. The destroy fails with dataset is busy. The dataset's mount is still listed in the toolkit process's /proc//mountinfo.

Requested fix

Set mountPropagation: HostToContainer on the containerd-socket volume, the same as host-root and driver-install-dir, so that host unmounts propagate into the container. A setting in ClusterPolicy to configure it would also work. Patching the DaemonSet by hand isn't a durable workaround because the operator reconciles it back.

主要言語
Go
スター
2.9k
フォーク
576
平均マージ
1日 5時間
マージ済み PR（30日）
65

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

NVIDIA/gpu-operator のほかの issue

NVIDIA/gpu-operator の issue をすべて見る

似ている issue

Go の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。