Troubleshooting Workload Clusters with Crash Diagnostics

This topic explains how to use Crash Diagnostics (Crashd) to diagnose unstable or unresponsive workload clusters based on Photon OS in Tanzu Kubernetes Grid with a standalone management cluster.

For how to use Crashd to diagnose workload clusters deployed by a vSphere with Tanzu Supervisor, see How to collect a diagnostic log bundle from a Tanzu Kubernetes Guest Cluster on vSphere with Tanzu in the VMware Knowledge Base.

Overview: Crashd

Crashd examines the bootstrap workload cluster that the tanzu cluster create process creates locally using kind before deploying the cluster to cloud infrastructure.

Crashd is an open source project that makes it easy troubleshoot problems with Kubernetes clusters.

Crashd uses a script file written in Starlark, a Python-like language, that interacts with your management or workload clusters to collect infrastructure and cluster information.

Crashd can collect diagnostics from supported infrastructures including:

  • AWS
  • Azure
  • vSphere

Crashd takes the output from the commands run by the script and adds the output to a tar file. The tar file is then saved locally for further analysis.

Tanzu Kubernetes Grid includes signed binaries for Crashd and a diagnostics script file for Photon OS workload clusters.

Install or Upgrade the Crashd Binary

To install or upgrade crashd, follow the instructions below.

  1. Go to the Broadcom Support Portal and log in with your VMware customer credentials.
  2. Visit the Tanzu Kubernetes Grid downloads page.
  3. In the version drop-down, select 2.2.0.
  4. Download Crashd for your platform.

    • Linux: crashd-linux-amd64-v0.3.7+vmware.5-4-g59b239d.tar.gz
    • macOS: crashd-darwin-amd64-v0.3.7+vmware.5-4-g59b239d.tar.gz
  5. Use the tar command to unpack the binary for your platform.

    • Linux:

      tar -xvf crashd-linux-amd64-v0.3.7-vmware.6.tar.gz
      
    • macOS:

      tar -xvf crashd-darwin-amd64-v0.3.7-vmware.6.tar.gz
      
  6. The previous step creates a directory named crashd with the following files:

    crashd
    crashd/args
    crashd/diagnostics.crsh
    crashd/crashd-PLATFORM-amd64-v0.3.7+vmware.6
    
  7. Move the binary into the /usr/local/bin folder.

    • Linux:

      mv ./crashd/crashd-linux-amd64-v0.3.7+vmware.6 /usr/local/bin/crashd
      
    • macOS:

      mv ./crashd/crashd-darwin-amd64-v0.3.7+vmware.6 /usr/local/bin/crashd
      

Run Crashd on Photon OS Workload Clusters

When Crashd runs, it takes argument values from an args file and passes them to a script file, diagnostics.crsh. The script runs commands to extract information that can help diagnose problems on Photon OS workload clusters.

Prerequisites

Prior to running Crashd script diagnostics.crsh, your local machine must have the following programs on its execution path:

  • kubectl
  • scp
  • ssh

    Note

    when investigating problems with a bootstrap cluster, you will need kind (v0.7.0 or higher) command installed locally.

Additionally, before you can run Crashd, you must follow these steps:

  • Configure Crashd with a SSH private/public key pair.
  • Ensure that your Tanzu Kubernetes Grid VMs are configured to use your SSH public key.
  • Extract the kubeconfig file for the management cluster by using command tanzu mc kubeconfig get <management-cluster-name>.
  • For a simpler setup, ensure that the kubeconfig, public-key file, the diagnostics.crsh file, and the args file are in the same location.
  • Delete any local kind clusters other than the one that was created to deploy the workload cluster that you are troubleshooting:

    • Run docker ps to identify the currently-running kind cluster
    • Delete any other kind clusters by running kind delete cluster --name CLUSTER-NAME

Configure Crashd

  1. Navigate to the location where you downloaded and unpacked the Crashd bundle.

  2. In a text editor, overwrite the existing arguments file args with the code below. This file contains the key/value pairs to pass to the CrashD script:

    # ######################################################
    # Crashd script argument file
    #
    # This file defines CLI argument values that are passed
    # Crashd when running scripts for troubleshooting TKG
    # clusters.
    # ######################################################
    
    # target: specifies cluster to target.
    # Valid targets are: {bootstrap, mgmt, workload}
    target=mgmt
    
    # infra: the underlying infrastructure used by the TKG cluster.
    # Valid values are: {vsphere, aws, azure}
    infra=vsphere
    
    # workdir: a local directory where collected files are staged.
    workdir=./workdir
    
    # ssh_user: the user ID used for SSH connections to cluster nodes.
    ssh_user=capv
    
    # ssh_pk_file: the path to the private key file created to SSH
    # into cluster nodes.
    ssh_pk_file=./capv.pem
    
    # ######################################################
    # Management Cluster
    # The following arguments are used to collect information
    # from a management cluster or named workload clusters.
    # ######################################################
    
    # mgmt_cluster_config: the kubeconfig file path for the management cluster.
    mgmt_cluster_config=./tkg_cluster_config
    
    # ######################################################
    # Workload Cluster
    # The following arguments are used to collect information
    # from one or more workload clusters that are managed
    # by the management cluster configured above.
    # ######################################################
    
    # workload_clusters: a comma separated list of workload cluster names
    # [uncomment below]
    #workload_clusters=tkg-cluster-wc-498
    
    # workload_cluster_ns: the namespace where the workload cluster
    # is hosted in the management plane.
    # Note: it's actually the namespace in which the secrets/${workload_cluster_name}-kubeconfig
    # is created in the management cluster.
    # [uncomment below]
    #workload_cluster_ns=default
    
  3. Record the local path to an SSH private key file. If you do not already have an SSH keypair, or want to generate a new one, run ssh-keygen as described in Create an SSH Key Pair. For example:

    ssh-keygen -t rsa -b 4096 -C "email@example.com"

    When prompted, enter a local path for the file location. For information about creating the SSH key pairs.

  4. Set the following arguments in the args file:

    • target: Set this value to:
      • bootstrap to diagnose a local bootstrap standalone management cluster
      • mgmt to diagnose a deployed standalone management cluster
      • workload to diagnose one or more workload clusters
    • infra: The underlying infrastructure for your cluster: aws, azure, or vsphere.
    • workdir: The location where files are collected.
    • ssh_user: The SSH user used to access cluster machines. For clusters running on vSphere, the user name is capv.
    • ssh_pk_file: The path to your SSH private key file.
    • mgmt_cluster_config The path of the kubeconfig file for the management cluster.
  5. To diagnose workload clusters, uncomment and set the following in addition to the arguments listed above:

    • workload_clusters: A comma-separated list of workload cluster names from which to collect diagnostics information.
    • workload_cluster_ns: The namespace in which the secrets/WORKLOAD-CLUSTER-NAME-kubeconfig is created in the management cluster.

Run Crashd

  1. Create a Crashd script file diagnostics.crsh containing the code in Diagnostics File diagnostics.crsh below.

  2. Run the crashd command from the location where the script file diagnostics.crsh and argument file args are located.

    crashd run --args-file args diagnostics.crsh
    
  3. Optionally, monitor Crashd output. By default, the crashd command runs silently until completion. However, you can use flag --debug to view log messages on the screen similar to the following:

    crashd run --debug --args-file args diagnostics.crsh
    
    DEBU[0003] creating working directory ./workdir/tkg-kind-12345
    DEBU[0003] kube_capture(what=objects)
    DEBU[0003] Searching in 20 groups
    ...
    DEBU[0015] Archiving [./workdir/tkg-kind-12345] in bootstrap.tkg-kind-12345.diagnostics.tar.gz
    DEBU[0015] Archived workdir/tkg-kind-12345/kind-logs/docker-info.txt
    DEBU[0015] Archived workdir/tkg-kind-12345/kind-logs/tkg-kind-12345-control-plane/alternatives.log
    DEBU[0015] Archived workdir/tkg-kind-12345/kind-logs/tkg-kind-12345-control-plane/containerd.log
    

Diagnostics File diagnostics.crsh

In the CrashD bundle download, overwrite the existing diagnostics.crsh file with the following code, as the script to pass to the crashd run command:

def capture_node_diagnostics(nodes):
    capture(cmd="sudo df -i", resources=nodes)
    capture(cmd="sudo crictl info", resources=nodes)
    capture(cmd="df -h /var/lib/containerd", resources=nodes)
    capture(cmd="sudo systemctl status kubelet", resources=nodes)
    capture(cmd="sudo systemctl status containerd", resources=nodes)
    capture(cmd="sudo journalctl -xeu kubelet", resources=nodes)
    capture(cmd="sudo journalctl -xeu containerd", resources=nodes)
    capture(cmd="sudo cat /var/log/cloud-init-output.log", resources=nodes)
    capture(cmd="sudo cat /var/log/cloud-init.log", resources=nodes)
def capture_windows_node_diagnostics(nodes):
    capture(cmd="Get-CimInstance -ClassName Win32_LogicalDisk", file_name="disk_info.out", resources=nodes)
    capture(cmd="(Get-ItemProperty -Path c:\\windows\\system32\\hal.dll).VersionInfo.FileVersion",file_name="windows_version_info.out", resources=nodes)
    capture(cmd="cat C:\\k\\StartKubelet.ps1 ; cat C:\\var\\lib\\kubelet\\kubeadm-flags.env", resources=nodes)
    capture(cmd="Get-Service Kubelet | select * ", resources=nodes)
    capture(cmd="Get-Service Containerd | select * ", resources=nodes)
    capture(cmd="Get-Service ovs* | select * ", resources=nodes)
    capture(cmd="Get-Service antrea-agent | select * ", resources=nodes)
    capture(cmd="Get-Service kube-proxy | select * ", resources=nodes)
    capture(cmd="Get-Service Kubelet | select * ", resources=nodes)
    capture(cmd="Get-HNSNetwork", resources=nodes)
    capture(cmd="& 'c:\\Program Files\\containerd\\crictl.exe' -r 'npipe:////./pipe/containerd-containerd' info", resources=nodes)
    capture(cmd="Get-MpPreference | select ExclusionProcess", resources=nodes)
    capture(cmd="cat c:\\var\\log\\kubelet\\kubelet.exe.INFO", resources=nodes)
    capture(cmd="cat c:\\var\\log\\antrea\\antrea-agent.exe.INFO", resources=nodes)
    capture(cmd="cat c:\\var\\log\\kube-proxy\\kube-proxy.exe.INFO", resources=nodes)
    capture(cmd="cat 'c:\\Program Files\\Cloudbase Solutions\\Cloudbase-Init\\log\\cloudbase-init-unattend.log'", resources=nodes)
    capture(cmd="cat 'c:\\Program Files\\Cloudbase Solutions\\Cloudbase-Init\\log\\cloudbase-init.log'", resources=nodes)
    copy_from(path="C:\\Windows\\System32\\Winevt\\Logs\\System.evtx", resources=nodes)
    copy_from(path="C:\\Windows\\System32\\Winevt\\Logs\\Application.evtx", resources=nodes)
    copy_from(path="c:\\openvswitch\\var\\log\\openvswitch\\ovs-vswitchd.log", resources=nodes)
    copy_from(path="c:\\openvswitch\\var\\log\\openvswitch\\ovsdb-server.log", resources=nodes)
# fetches a suitable capi provider, for either capa or others (capv/capz),
# to be used for enumerating cluster machines
def fetch_provider(iaas, workload_cluster_name, ssh_cfg, kube_cfg, namespace, filter_labels):
    # workaround: vsphere and azure use same provider as they work similarly (see issue #162)
    if iaas == "vsphere" or iaas == "azure":
        provider = capv_provider(
            workload_cluster=workload_cluster_name,
            namespace=namespace,
            ssh_config=ssh_cfg,
            mgmt_kube_config=kube_cfg,
            labels=filter_labels
        )
    else:
        provider = capa_provider(
            workload_cluster=workload_cluster_name,
            namespace=namespace,
            ssh_config=ssh_cfg,
            mgmt_kube_config=kube_cfg,
            labels=filter_labels
        )
    return provider
# retrieves linux management provider for linux nodes
def fetch_mgmt_provider_linux(infra, ssh_cfg, kube_cfg, ns):
    return fetch_provider(infra, '', ssh_cfg, kube_cfg, ns, ["kubernetes.io/os=linux"])
# retrieves windows mgmt provider for windows nodes
def fetch_mgmt_provider_windows(infra, ssh_cfg, kube_cfg, ns):
    return fetch_provider(infra, '', ssh_cfg, kube_cfg, ns, ["kubernetes.io/os=windows"])
# retrieves linux workload provider for linux nodes
def fetch_workload_provider_linux(infra, wc_cluster, ssh_cfg, kube_cfg, ns):
    return fetch_provider(infra, wc_cluster, ssh_cfg, kube_cfg, ns, ["kubernetes.io/os=linux"])
# retrieves windows workload provider for windodws nodes
def fetch_workload_provider_windows(infra, wc_cluster, ssh_cfg, kube_cfg, ns):
    return fetch_provider(infra, wc_cluster, ssh_cfg, kube_cfg, ns, ["kubernetes.io/os=windows"])
def diagnose_mgmt_cluster(infra):
    # validation
    args.ssh_user
    args.ssh_pk_file
    args.mgmt_cluster_config
    if len(infra) == 0:
        print("Infra argument not provided")
        return
    wd = "{}/tkg-mgmt-cluster".format(args.workdir)
    conf = crashd_config(workdir=wd)
    ssh_conf = ssh_config(username=args.ssh_user, private_key_path=args.ssh_pk_file)
    kube_conf = kube_config(path=args.mgmt_cluster_config)
    # fetch linux mgmt node diagnostics
    mgmt_provider_linux = fetch_mgmt_provider_linux(infra, ssh_conf, kube_conf, '')
    lin_nodes = resources(provider=mgmt_provider_linux)
    capture_node_diagnostics(lin_nodes)
    # fetch win mgmt node diagnostics
    mgmt_provider_win = fetch_mgmt_provider_windows(infra, ssh_conf, kube_conf, '')
    win_nodes = resources(provider=mgmt_provider_win)
    if len(win_nodes) > 0:
        capture_windows_node_diagnostics(win_nodes)
    #add code to collect pod info from cluster
    set_defaults(kube_config(capi_provider = mgmt_provider_linux))
    pods_ns=[
        "capi-kubeadm-bootstrap-system",
        "capi-kubeadm-control-plane-system",
        "capi-system",
        "capi-webhook-system",
        "cert-manager",
        "tkg-system",
        "kube-system",
        "tkr-system",
        "capa-system",
        "capv-system",
        "capz-system",
    ]
    if infra == "vsphere":
        pods_ns.append("tkg-system-networking")
        pods_ns.append("avi-system")
    kube_capture(what="logs", namespaces=pods_ns)
    kube_capture(what="objects", kinds=["pods", "services"], namespaces=pods_ns)
    kube_capture(what="objects", kinds=["deployments", "replicasets"], groups=["apps"], namespaces=pods_ns)
    kube_capture(what="objects", kinds=["apps"], groups=["kappctrl.k14s.io"], namespaces=["tkg-system"])
    kube_capture(what="objects", kinds=["tanzukubernetesreleases"], groups=["run.tanzu.vmware.com"])
    kube_capture(what="objects", kinds=["configmaps"], namespaces=["tkr-system"])
    kube_capture(what="objects", categories=["cluster-api"])
    kube_capture(what="objects", groups=["ipam.cluster.x-k8s.io"])
    if infra == "vsphere":
        kube_capture(what="objects", kinds=["akodeploymentconfigs"])
    archive(output_file="tkg-mgmt.diagnostics.tar.gz", source_paths=[conf.workdir])
def diagnose_workload_cluster(infra, name):
    # validation
    args.infra
    args.ssh_user
    args.ssh_pk_file
    args.mgmt_cluster_config
    workload_ns = args.workload_cluster_ns
    if len(infra) == 0:
        print("Infra argument not provided")
        return
    wd = "{}/{}".format(args.workdir, name)
    conf = crashd_config(workdir=wd)
    ssh_conf = ssh_config(username=args.ssh_user, private_key_path=args.ssh_pk_file)
    kube_conf = kube_config(path=args.mgmt_cluster_config)
    # fetch linux workload node diagnostics
    wc_provider_linux = fetch_workload_provider_linux(infra, name, ssh_conf, kube_conf, workload_ns)
    lin_nodes = resources(provider=wc_provider_linux)
    capture_node_diagnostics(lin_nodes)
    # fetch win workload node diagnostics
    wc_provider_win = fetch_workload_provider_windows(infra, name, ssh_conf, kube_conf, workload_ns)
    win_nodes = resources(provider=wc_provider_win)
    if len(win_nodes) > 0:
        capture_windows_node_diagnostics(win_nodes)
    #add code to collect pod info from cluster
    set_defaults(kube_config(capi_provider = wc_provider_linux))
    pods_ns=["default", "kube-system", "tkg-system"]
    if infra == "vsphere":
        pods_ns.append("tkg-system-networking")
        pods_ns.append("avi-system")
    kube_capture(what="logs", namespaces=pods_ns)
    kube_capture(what="objects", kinds=["pods", "services"], namespaces=pods_ns)
    kube_capture(what="objects", kinds=["deployments", "replicasets"], groups=["apps"], namespaces=pods_ns)
    kube_capture(what="objects", kinds=["apps"], groups=["kappctrl.k14s.io"], namespaces=["tkg-system"])
    if infra == "vsphere":
        kube_capture(what="objects", kinds=["akodeploymentconfigs"])
    archive(output_file="{}.diagnostics.tar.gz".format(name), source_paths=[conf.workdir])
# extract diagnostic info from local kind boostrap cluster
def diagnose_bootstrap_cluster():
    p = prog_avail_local("kind")
    if p == "":
        print("Error: kind is not available")
        return
    clusters=get_tkg_bootstrap_clusters()
    if len(clusters) == 0:
        print("No tkg-kind bootstrap cluster found")
        return
    pod_ns=[
        "caip-in-cluster-system",
        "capi-kubeadm-bootstrap-system",
        "capi-kubeadm-control-plane-system",
        "capi-system",
        "capi-webhook-system",
        "capv-system",
        "capa-system",
        "capz-system",
        "cert-manager",
        "tkg-system",
        "tkg-system-networking",
        "avi-system",
    ]
    # for each tkg-kind cluster:
    #  - capture kind logs, export kubecfg, and api objects
    for kind_cluster in clusters:
        wd = "{}/{}".format(args.workdir, kind_cluster)
        run_local("kind export logs --name {} {}/kind-logs".format(kind_cluster, wd))
        kind_cfg = capture_local(
            cmd="kind get kubeconfig --name {0}".format(kind_cluster),
            workdir="./",
            file_name="{}.kubecfg".format(kind_cluster)
        )

        conf = crashd_config(workdir=wd)
        set_defaults(kube_config(path=kind_cfg))
        kube_capture(what="objects", kinds=["pods", "services"], namespaces=pod_ns)
        kube_capture(what="objects", kinds=["deployments", "replicasets"], groups=["apps"], namespaces=pod_ns)
        kube_capture(what="objects", categories=["cluster-api"])
        kube_capture(what="objects", kinds=["akodeploymentconfigs"])
        archive(output_file="bootstrap.{}.diagnostics.tar.gz".format(kind_cluster), source_paths=[conf.workdir])
# return tkg clusters in kind (tkg-kind-xxxx)
def get_tkg_bootstrap_clusters():
    clusters = run_local("kind get clusters").split('\n')
    result = []
    for cluster in clusters:
        if cluster.startswith("tkg-kind"):
            result.append(cluster)

    return result
def check_prereqs():
    # validate args
    args.workdir
    p = prog_avail_local("ssh")
    if p == "":
        print("Error: ssh is not available")
        return False

    p = prog_avail_local("scp")
    if p == "":
        print("Error: scp is not available")
        return False

    p = prog_avail_local("kubectl")
    if p == "":
        print("Error: kubectl is not available")
        return False

    return True

def diagnose(target, infra):
    # validation
    if not check_prereqs():
        print("Error: One or more prerequisites are missing")
        return
    # run diagnostics
    if target == "bootstrap":
        diagnose_bootstrap_cluster()
    elif target == "mgmt":
        diagnose_mgmt_cluster(infra)
    elif target == "workload":
        for name in args.workload_clusters.split(","):
            diagnose_workload_cluster(infra, name)
    else:
        print("Error: unknown target {}".format(target))
diagnose(args.target, args.infra)
check-circle-line exclamation-circle-line close-line
Scroll to top icon