Ceph Common Issues
Many of these problem cases are hard to summarize down to a short phrase that adequately describes the problem. Each problem will start with a bulleted list of symptoms. Keep in mind that all symptoms may not apply depending on the configuration of Rook. If the majority of the symptoms are seen there is a fair chance you are experiencing that problem.
If after trying the suggestions found on this page and the problem is not resolved, the Rook team is very happy to help you troubleshoot the issues in their Slack channel. Once you have registered for the Rook Slack, proceed to the #ceph
channel to ask for assistance.
See also the CSI Troubleshooting Guide.
Troubleshooting Techniques¶
There are two main categories of information you will need to investigate issues in the cluster:
- Kubernetes status and logs documented here
- Ceph cluster status (see upcoming Ceph tools section)
Ceph Tools¶
After you verify the basic health of the running pods, next you will want to run Ceph tools for status of the storage components. There are two ways to run the Ceph tools, either in the Rook toolbox or inside other Rook pods that are already running.
- Logs on a specific node to find why a PVC is failing to mount
- See the log collection topic for a script that will help you gather the logs
- Other artifacts:
- The monitors that are expected to be in quorum:
kubectl -n <cluster-namespace> get configmap rook-ceph-mon-endpoints -o yaml | grep data
Tools in the Rook Toolbox¶
The rook-ceph-tools pod provides a simple environment to run Ceph tools. Once the pod is up and running, connect to the pod to execute Ceph commands to evaluate that current state of the cluster.
Ceph Commands¶
Here are some common commands to troubleshoot a Ceph cluster:
ceph status
ceph osd status
ceph osd df
ceph osd utilization
ceph osd pool stats
ceph osd tree
ceph pg stat
The first two status commands provide the overall cluster health. The normal state for cluster operations is HEALTH_OK, but will still function when the state is in a HEALTH_WARN state. If you are in a WARN state, then the cluster is in a condition that it may enter the HEALTH_ERROR state at which point all disk I/O operations are halted. If a HEALTH_WARN state is observed, then one should take action to prevent the cluster from halting when it enters the HEALTH_ERROR state.
There are many Ceph sub-commands to look at and manipulate Ceph objects, well beyond the scope this document. See the Ceph documentation for more details of gathering information about the health of the cluster. In addition, there are other helpful hints and some best practices located in the Advanced Configuration section. Of particular note, there are scripts for collecting logs and gathering OSD information there.
Cluster failing to service requests¶
Symptoms¶
- Execution of the
ceph
command hangs - PersistentVolumes are not being created
- Large amount of slow requests are blocking
- Large amount of stuck requests are blocking
- One or more MONs are restarting periodically
Investigation¶
Create a rook-ceph-tools pod to investigate the current state of Ceph. Here is an example of what one might see. In this case the ceph status
command would just hang so a CTRL-C needed to be sent.
Another indication is when one or more of the MON pods restart frequently. Note the 'mon107' that has only been up for 16 minutes in the following output.
Solution¶
What is happening here is that the MON pods are restarting and one or more of the Ceph daemons are not getting configured with the proper cluster information. This is commonly the result of not specifying a value for dataDirHostPath
in your Cluster CRD.
The dataDirHostPath
setting specifies a path on the local host for the Ceph daemons to store configuration and data. Setting this to a path like /var/lib/rook
, reapplying your Cluster CRD and restarting all the Ceph daemons (MON, MGR, OSD, RGW) should solve this problem. After the Ceph daemons have been restarted, it is advisable to restart the rook-tools pod.
Monitors are the only pods running¶
Symptoms¶
- Rook operator is running
- Either a single mon starts or the mons start very slowly (at least several minutes apart)
- The crash-collector pods are crashing
- No mgr, osd, or other daemons are created except the CSI driver
Investigation¶
When the operator is starting a cluster, the operator will start one mon at a time and check that they are healthy before continuing to bring up all three mons. If the first mon is not detected healthy, the operator will continue to check until it is healthy. If the first mon fails to start, a second and then a third mon may attempt to start. However, they will never form quorum and the orchestration will be blocked from proceeding.
The crash-collector pods will be blocked from starting until the mons have formed quorum the first time.
There are several common causes for the mons failing to form quorum:
- The operator pod does not have network connectivity to the mon pod(s). The network may be configured incorrectly.
- One or more mon pods are in running state, but the operator log shows they are not able to form quorum
- A mon is using configuration from a previous installation. See the cleanup guide for cleaning the previous cluster.
- A firewall may be blocking the ports required for the Ceph mons to form quorum. Ensure ports 6789 and 3300 are enabled. See the Ceph networking guide for more details.
- There may be MTU mismatch between different networking components. Some networks may be more susceptible to mismatch than others. If Kubernetes CNI or hosts enable jumbo frames (MTU 9000), Ceph will use large packets to maximize network bandwidth. If other parts of the networking chain don't support jumbo frames, this could result in lost or rejected packets unexpectedly.
Operator fails to connect to the mon¶
First look at the logs of the operator to confirm if it is able to connect to the mons.
Likely you will see an error similar to the following that the operator is timing out when connecting to the mon. The last command is ceph mon_status
, followed by a timeout message five minutes later.
The error would appear to be an authentication error, but it is misleading. The real issue is a timeout.
Solution¶
If you see the timeout in the operator log, verify if the mon pod is running (see the next section). If the mon pod is running, check the network connectivity between the operator pod and the mon pod. A common issue is that the CNI is not configured correctly.
To verify the network connectivity:
- Get the endpoint for a mon
- Curl the mon from the operator pod
For example, this command will curl the first mon from the operator:
If "ceph v2" is printed to the console, the connection was successful. If the command does not respond or otherwise fails, the network connection cannot be established.
Failing mon pod¶
Second we need to verify if the mon pod started successfully.
If the mon pod is failing as in this example, you will need to look at the mon pod status or logs to determine the cause. If the pod is in a crash loop backoff state, you should see the reason by describing the pod.
See the solution in the next section regarding cleaning up the dataDirHostPath
on the nodes.
Solution¶
This is a common problem reinitializing the Rook cluster when the local directory used for persistence has not been purged. This directory is the dataDirHostPath
setting in the cluster CRD and is typically set to /var/lib/rook
. To fix the issue you will need to delete all components of Rook and then delete the contents of /var/lib/rook
(or the directory specified by dataDirHostPath
) on each of the hosts in the cluster. Then when the cluster CRD is applied to start a new cluster, the rook-operator should start all the pods as expected.
Caution
Deleting the dataDirHostPath
folder is destructive to the storage. Only delete the folder if you are trying to permanently purge the Rook cluster.
See the Cleanup Guide for more details.
PVCs stay in pending state¶
Symptoms¶
- When you create a PVC based on a rook storage class, it stays pending indefinitely
For the Wordpress example, you might see two PVCs in pending state.
Investigation¶
There are two common causes for the PVCs staying in pending state:
- There are no OSDs in the cluster
- The CSI provisioner pod is not running or is not responding to the request to provision the storage
Confirm if there are OSDs¶
To confirm if you have OSDs in your cluster, connect to the Rook Toolbox and run the ceph status
command. You should see that you have at least one OSD up
and in
. The minimum number of OSDs required depends on the replicated.size
setting in the pool created for the storage class. In a "test" cluster, only one OSD is required (see storageclass-test.yaml
). In the production storage class example (storageclass.yaml
), three OSDs would be required.
OSD Prepare Logs¶
If you don't see the expected number of OSDs, let's investigate why they weren't created. On each node where Rook looks for OSDs to configure, you will see an "osd prepare" pod.
See the section on why OSDs are not getting created to investigate the logs.
CSI Driver¶
The CSI driver may not be responding to the requests. Look in the logs of the CSI provisioner pod to see if there are any errors during the provisioning.
There are two provisioner pods:
Get the logs of each of the pods. One of them should be the "leader" and be responding to requests.
See also the CSI Troubleshooting Guide.
Operator unresponsiveness¶
Lastly, if you have OSDs up
and in
, the next step is to confirm the operator is responding to the requests. Look in the Operator pod logs around the time when the PVC was created to confirm if the request is being raised. If the operator does not show requests to provision the block image, the operator may be stuck on some other operation. In this case, restart the operator pod to get things going again.
Solution¶
If the "osd prepare" logs didn't give you enough clues about why the OSDs were not being created, please review your cluster.yaml configuration. The common misconfigurations include:
- If
useAllDevices: true
, Rook expects to find local devices attached to the nodes. If no devices are found, no OSDs will be created. - If
useAllDevices: false
, OSDs will only be created ifdeviceFilter
is specified. - Only local devices attached to the nodes will be configurable by Rook. In other words, the devices must show up under
/dev
. - The devices must not have any partitions or filesystems on them. Rook will only configure raw devices. Partitions are not yet supported.
OSD pods are failing to start¶
Symptoms¶
- OSD pods are failing to start
- You have started a cluster after tearing down another cluster
Investigation¶
When an OSD starts, the device or directory will be configured for consumption. If there is an error with the configuration, the pod will crash and you will see the CrashLoopBackoff status for the pod. Look in the osd pod logs for an indication of the failure.
One common case for failure is that you have re-deployed a test cluster and some state may remain from a previous deployment. If your cluster is larger than a few nodes, you may get lucky enough that the monitors were able to start and form quorum. However, now the OSDs pods may fail to start due to the old state. Looking at the OSD pod logs you will see an error about the file already existing.
Solution¶
If the error is from the file that already exists, this is a common problem reinitializing the Rook cluster when the local directory used for persistence has not been purged. This directory is the dataDirHostPath
setting in the cluster CRD and is typically set to /var/lib/rook
. To fix the issue you will need to delete all components of Rook and then delete the contents of /var/lib/rook
(or the directory specified by dataDirHostPath
) on each of the hosts in the cluster. Then when the cluster CRD is applied to start a new cluster, the rook-operator should start all the pods as expected.
OSD pods are not created on my devices¶
Symptoms¶
- No OSD pods are started in the cluster
- Devices are not configured with OSDs even though specified in the Cluster CRD
- One OSD pod is started on each node instead of multiple pods for each device
Investigation¶
First, ensure that you have specified the devices correctly in the CRD. The Cluster CRD has several ways to specify the devices that are to be consumed by the Rook storage:
useAllDevices: true
: Rook will consume all devices it determines to be availabledeviceFilter
: Consume all devices that match this regular expressiondevices
: Explicit list of device names on each node to consume
Second, if Rook determines that a device is not available (has existing partitions or a formatted filesystem), Rook will skip consuming the devices. If Rook is not starting OSDs on the devices you expect, Rook may have skipped it for this reason. To see if a device was skipped, view the OSD preparation log on the node where the device was skipped. Note that it is completely normal and expected for OSD prepare pod to be in the completed
state. After the job is complete, Rook leaves the pod around in case the logs need to be investigated.
Here are some key lines to look for in the log:
Solution¶
Either update the CR with the correct settings, or clean the partitions or filesystem from your devices. To clean devices from a previous install see the cleanup guide.
After the settings are updated or the devices are cleaned, trigger the operator to analyze the devices again by restarting the operator. Each time the operator starts, it will ensure all the desired devices are configured. The operator does automatically deploy OSDs in most scenarios, but an operator restart will cover any scenarios that the operator doesn't detect automatically.
Node hangs after reboot¶
This issue is fixed in Rook v1.3 or later.
Symptoms¶
- After issuing a
reboot
command, node never returned online - Only a power cycle helps
Investigation¶
On a node running a pod with a Ceph persistent volume
When the reboot command is issued, network interfaces are terminated before disks are unmounted. This results in the node hanging as repeated attempts to unmount Ceph persistent volumes fail with the following error:
Solution¶
The node needs to be drained before reboot. After the successful drain, the node can be rebooted as usual.
Because kubectl drain
command automatically marks the node as unschedulable (kubectl cordon
effect), the node needs to be uncordoned once it's back online.
Drain the node:
Uncordon the node:
Using multiple shared filesystem (CephFS) is attempted on a kernel version older than 4.7¶
Symptoms¶
- More than one shared filesystem (CephFS) has been created in the cluster
- A pod attempts to mount any other shared filesystem besides the first one that was created
- The pod incorrectly gets the first filesystem mounted instead of the intended filesystem
Solution¶
The only solution to this problem is to upgrade your kernel to 4.7
or higher. This is due to a mount flag added in the kernel version 4.7
which allows to chose the filesystem by name.
For additional info on the kernel version requirement for multiple shared filesystems (CephFS), see Filesystem - Kernel version requirement.
Set debug log level for all Ceph daemons¶
You can set a given log level and apply it to all the Ceph daemons at the same time. For this, make sure the toolbox pod is running, then determine the level you want (between 0 and 20). You can find the list of all subsystems and their default values in Ceph logging and debug official guide. Be careful when increasing the level as it will produce very verbose logs.
Assuming you want a log level of 1, you will run:
Once you are done debugging, you can revert all the debug flag to their default value by running the following:
Activate log to file for a particular Ceph daemon¶
They are cases where looking at Kubernetes logs is not enough for diverse reasons, but just to name a few:
- not everyone is familiar for Kubernetes logging and expects to find logs in traditional directories
- logs get eaten (buffer limit from the log engine) and thus not requestable from Kubernetes
So for each daemon, dataDirHostPath
is used to store logs, if logging is activated. Rook will bindmount dataDirHostPath
for every pod. Let's say you want to enable logging for mon.a
, but only for this daemon. Using the toolbox or from inside the operator run:
This will activate logging on the filesystem, you will be able to find logs in dataDirHostPath/$NAMESPACE/log
, so typically this would mean /var/lib/rook/rook-ceph/log
. You don't need to restart the pod, the effect will be immediate.
To disable the logging on file, simply set log_to_file
to false
.
A worker node using RBD devices hangs up¶
Symptoms¶
- There is no progress on I/O from/to one of RBD devices (
/dev/rbd*
or/dev/nbd*
). - After that, the whole worker node hangs up.
Investigation¶
This happens when the following conditions are satisfied.
- The problematic RBD device and the corresponding OSDs are co-located.
- There is an XFS filesystem on top of this device.
In addition, when this problem happens, you can see the following messages in dmesg
.
It's so-called hung_task
problem and means that there is a deadlock in the kernel. For more detail, please refer to the corresponding issue comment.
Solution¶
This problem will be solve by the following two fixes.
- Linux kernel: A minor feature that is introduced by this commit. It will be included in Linux v5.6.
- Ceph: A fix that uses the above-mentioned kernel's feature. The Ceph community will probably discuss this fix after releasing Linux v5.6.
You can bypass this problem by using ext4 or any other filesystems rather than XFS. Filesystem type can be specified with csi.storage.k8s.io/fstype
in StorageClass resource.
Too few PGs per OSD warning is shown¶
Symptoms¶
ceph status
shows "too few PGs per OSD" warning as follows.
Solution¶
The meaning of this warning is written in the document. However, in many cases it is benign. For more information, please see the blog entry. Please refer to Configuring Pools if you want to know the proper pg_num
of pools and change these values.
LVM metadata can be corrupted with OSD on LV-backed PVC¶
Symptoms¶
There is a critical flaw in OSD on LV-backed PVC. LVM metadata can be corrupted if both the host and OSD container modify it simultaneously. For example, the administrator might modify it on the host, while the OSD initialization process in a container could modify it too. In addition, if lvmetad
is running, the possibility of occurrence gets higher. In this case, the change of LVM metadata in OSD container is not reflected to LVM metadata cache in host for a while.
If you still decide to configure an OSD on LVM, please keep the following in mind to reduce the probability of this issue.
Solution¶
- Disable
lvmetad.
- Avoid configuration of LVs from the host. In addition, don't touch the VGs and physical volumes that back these LVs.
- Avoid incrementing the
count
field ofstorageClassDeviceSets
and create a new LV that backs an OSD simultaneously.
You can know whether the above-mentioned tag exists with the command: sudo lvs -o lv_name,lv_tags
. If the lv_tag
field is empty in an LV corresponding to the OSD lv_tags, this OSD encountered the problem. In this case, please retire this OSD or replace with other new OSD before restarting.
This problem doesn't happen in newly created LV-backed PVCs because OSD container doesn't modify LVM metadata anymore. The existing lvm mode OSDs work continuously even thought upgrade your Rook. However, using the raw mode OSDs is recommended because of the above-mentioned problem. You can replace the existing OSDs with raw mode OSDs by retiring them and adding new OSDs one by one. See the documents Remove an OSD and Add an OSD on a PVC.
OSD prepare job fails due to low aio-max-nr setting¶
If the Kernel is configured with a low aio-max-nr setting, the OSD prepare job might fail with the following error:
To overcome this, you need to increase the value of fs.aio-max-nr
of your sysctl configuration (typically /etc/sysctl.conf
). You can do this with your favorite configuration management system.
Alternatively, you can have a DaemonSet to apply the configuration for you on all your nodes.
Unexpected partitions created¶
Symptoms¶
Users running Rook versions v1.6.0-v1.6.7 may observe unwanted OSDs on partitions that appear unexpectedly and seemingly randomly, which can corrupt existing OSDs.
Unexpected partitions are created on host disks that are used by Ceph OSDs. This happens more often on SSDs than HDDs and usually only on disks that are 875GB or larger. Many tools like lsblk
, blkid
, udevadm
, and parted
will not show a partition table type for the partition. Newer versions of blkid
are generally able to recognize the type as "atari".
The underlying issue causing this is Atari partition (sometimes identified as AHDI) support in the Linux kernel. Atari partitions have very relaxed specifications compared to other partition types, and it is relatively easy for random data written to a disk to appear as an Atari partition to the Linux kernel. Ceph's Bluestore OSDs have an anecdotally high probability of writing data on to disks that can appear to the kernel as an Atari partition.
Below is an example of lsblk
output from a node where phantom Atari partitions are present. Note that sdX1
is never present for the phantom partitions, and sdX2
is 48G on all disks. sdX3
is a variable size and may not always be present. It is possible for sdX4
to appear, though it is an anecdotally rare event.
You can see GitHub rook/rook - Issue 7940 unexpected partition on disks >= 1TB (atari partitions) for more detailed information and discussion.
Solution¶
Recover from corruption (v1.6.0-v1.6.7)¶
If you are using Rook v1.6, you must first update to v1.6.8 or higher to avoid further incidents of OSD corruption caused by these Atari partitions.
An old workaround suggested using deviceFilter: ^sd[a-z]+$
, but this still results in unexpected partitions. Rook will merely stop creating new OSDs on the partitions. It does not fix a related issue that ceph-volume
that is unaware of the Atari partition problem. Users who used this workaround are still at risk for OSD failures in the future.
To resolve the issue, immediately update to v1.6.8 or higher. After the update, no corruption should occur on OSDs created in the future. Next, to get back to a healthy Ceph cluster state, focus on one corrupted disk at a time and remove all OSDs on each corrupted disk one disk at a time.
As an example, you may have /dev/sdb
with two unexpected partitions (/dev/sdb2
and /dev/sdb3
) as well as a second corrupted disk /dev/sde
with one unexpected partition (/dev/sde2
).
- First, remove the OSDs associated with
/dev/sdb
,/dev/sdb2
, and/dev/sdb3
. There might be only one, or up to 3 OSDs depending on how your system was affected. Again see the OSD management doc. - Use
dd
to wipe the first sectors of the partitions followed by the disk itself. E.g.,dd if=/dev/zero of=/dev/sdb2 bs=1M
dd if=/dev/zero of=/dev/sdb3 bs=1M
dd if=/dev/zero of=/dev/sdb bs=1M
- Then wipe clean
/dev/sdb
to prepare it for a new OSD. See the teardown document for details. - After this, scale up the Rook operator to deploy a new OSD to
/dev/sdb
. This will allow Ceph to use/dev/sdb
for data recovery and replication while the next OSDs are removed. - Now Repeat steps 1-4 for
/dev/sde
and/dev/sde2
, and continue for any other corrupted disks.
If your Rook cluster does not have any critical data stored in it, it may be simpler to uninstall Rook completely and redeploy with v1.6.8 or higher.
Operator environment variables are ignored¶
Symptoms¶
Configuration settings passed as environment variables do not take effect as expected. For example, the discover daemonset is not created, even though ROOK_ENABLE_DISCOVERY_DAEMON="true"
is set.
Investigation¶
Inspect the rook-ceph-operator-config
ConfigMap for conflicting settings. The ConfigMap takes precedence over the environment. The ConfigMap must exist, even if all actual configuration is supplied through the environment.
Look for lines with the op-k8sutil
prefix in the operator logs. These lines detail the final values, and source, of the different configuration variables.
Verify that both of the following messages are present in the operator logs:
Solution¶
If it does not exist, create an empty ConfigMap:
If the ConfigMap exists, remove any keys that you wish to configure through the environment.