Notice

This document is for a development version of Ceph.

Troubleshooting

Slow/stuck operations

If you are experiencing apparent hung operations, the first task is to identify where the problem is occurring: in the client, the MDS, or the network connecting them. Start by looking to see if either side has stuck operations (Slow requests (MDS), below), and narrow it down from there.

We can get hints about what’s going on by dumping the MDS cache

ceph daemon mds.<name> dump cache /tmp/dump.txt

Note

The file dump.txt is on the machine executing the MDS and for systemd controlled MDS services, this is in a tmpfs in the MDS container. Use nsenter(1) to locate dump.txt or specify another system-wide path.

If high logging levels are set on the MDS, that will almost certainly hold the information we need to diagnose and solve the issue.

Stuck during recovery

Stuck in up:replay

If your MDS is stuck in up:replay then it is likely that the journal is very long. Did you see MDS_HEALTH_TRIM cluster warnings saying the MDS is behind on trimming its journal? If the journal has grown very large, it can take hours to read the journal. There is no working around this but there are things you can do to speed things along:

Reduce MDS debugging to 0. Even at the default settings, the MDS logs some messages to memory for dumping if a fatal error is encountered. You can avoid this:

ceph config set mds debug_mds 0
ceph config set mds debug_ms 0
ceph config set mds debug_monc 0

Note if the MDS fails then there will be virtually no information to determine why. If you can calculate when up:replay will complete, you should restore these configs just prior to entering the next state:

ceph config rm mds debug_mds
ceph config rm mds debug_ms
ceph config rm mds debug_monc

Once you’ve got replay moving along faster, you can calculate when the MDS will complete. This is done by examining the journal replay status:

$ ceph tell mds.<fs_name>:0 status | jq .replay_status
{
  "journal_read_pos": 4195244,
  "journal_write_pos": 4195244,
  "journal_expire_pos": 4194304,
  "num_events": 2,
  "num_segments": 2
}

Replay completes when the journal_read_pos reaches the journal_write_pos. The write position will not change during replay. Track the progression of the read position to compute the expected time to complete.

Avoiding recovery roadblocks

When trying to urgently restore your file system during an outage, here are some things to do:

  • Deny all reconnect to clients. This effectively blocklists all existing CephFS sessions so all mounts will hang or become unavailable.

 ceph config set mds mds_deny_all_reconnect true

Remember to undo this after the MDS becomes active.

Note

This does not prevent new sessions from connecting. For that, see the refuse_client_session file system setting.

  • Extend the MDS heartbeat grace period. This avoids replacing an MDS that appears “stuck” doing some operation. Sometimes recovery of an MDS may involve an operation that may take longer than expected (from the programmer’s perspective). This is more likely when recovery is already taking a longer than normal amount of time to complete (indicated by your reading this document). Avoid unnecessary replacement loops by extending the heartbeat graceperiod:

ceph config set mds mds_heartbeat_grace 3600

Note

This has the effect of having the MDS continue to send beacons to the monitors even when its internal “heartbeat” mechanism has not been reset (beat) in one hour. The previous mechanism for achieving this was via the mds_beacon_grace monitor setting.

  • Disable open file table prefetch. Normally, the MDS will prefetch directory contents during recovery to heat up its cache. During long recovery, the cache is probably already hot and large. So this behavior can be undesirable. Disable using:

ceph config set mds mds_oft_prefetch_dirfrags false
  • Turn off clients. Clients reconnecting to the newly up:active MDS may cause new load on the file system when it’s just getting back on its feet. There will likely be some general maintenance to do before workloads should be resumed. For example, expediting journal trim may be advisable if the recovery took a long time because replay was reading a overly large journal.

    You can do this manually or use the new file system tunable:

 ceph fs set <fs_name> refuse_client_session true

That prevents any clients from establishing new sessions with the MDS.

Expediting MDS journal trim

If your MDS journal grew too large (maybe your MDS was stuck in up:replay for a long time!), you will want to have the MDS trim its journal more frequently. You will know the journal is too large because of MDS_HEALTH_TRIM warnings.

The main tunable available to do this is to modify the MDS tick interval. The “tick” interval drives several upkeep activities in the MDS. It is strongly recommended no significant file system load be present when modifying this tick interval. This setting only affects an MDS in up:active. The MDS does not trim its journal during recovery.

ceph config set mds mds_tick_interval 2

RADOS Health

If part of the CephFS metadata or data pools is unavailable and CephFS is not responding, it is probably because RADOS itself is unhealthy. Resolve those problems first (Troubleshooting).

The MDS

If an operation is hung inside the MDS, it will eventually show up in ceph health, identifying “slow requests are blocked”. It may also identify clients as “failing to respond” or misbehaving in other ways. If the MDS identifies specific clients as misbehaving, you should investigate why they are doing so.

Generally it will be the result of

  1. Overloading the system (if you have extra RAM, increase the “mds cache memory limit” config from its default 1GiB; having a larger active file set than your MDS cache is the #1 cause of this!).

  2. Running an older (misbehaving) client.

  3. Underlying RADOS issues.

Otherwise, you have probably discovered a new bug and should report it to the developers!

Slow requests (MDS)

You can list current operations via the admin socket by running:

ceph daemon mds.<name> dump_ops_in_flight

from the MDS host. Identify the stuck commands and examine why they are stuck. Usually the last “event” will have been an attempt to gather locks, or sending the operation off to the MDS log. If it is waiting on the OSDs, fix them. If operations are stuck on a specific inode, you probably have a client holding caps which prevent others from using it, either because the client is trying to flush out dirty data or because you have encountered a bug in CephFS’ distributed file lock code (the file “capabilities” [“caps”] system).

If it’s a result of a bug in the capabilities code, restarting the MDS is likely to resolve the problem.

If there are no slow requests reported on the MDS, and it is not reporting that clients are misbehaving, either the client has a problem or its requests are not reaching the MDS.

ceph-fuse debugging

ceph-fuse also supports dump_ops_in_flight. See if it has any and where they are stuck.

Debug output

To get more debugging information from ceph-fuse, try running in the foreground with logging to the console (-d) and enabling client debug (--debug-client=20), enabling prints for each message sent (--debug-ms=1).

If you suspect a potential monitor issue, enable monitor debugging as well (--debug-monc=20).

Kernel mount debugging

If there is an issue with the kernel client, the most important thing is figuring out whether the problem is with the kernel client or the MDS. Generally, this is easy to work out. If the kernel client broke directly, there will be output in dmesg. Collect it and any inappropriate kernel state.

Slow requests

Unfortunately the kernel client does not support the admin socket, but it has similar (if limited) interfaces if your kernel has debugfs enabled. There will be a folder in sys/kernel/debug/ceph/, and that folder (whose name will look something like 28f7427e-5558-4ffd-ae1a-51ec3042759a.client25386880) will contain a variety of files that output interesting output when you cat them. These files are described below; the most interesting when debugging slow requests are probably the mdsc and osdc files.

  • bdi: BDI info about the Ceph system (blocks dirtied, written, etc)

  • caps: counts of file “caps” structures in-memory and used

  • client_options: dumps the options provided to the CephFS mount

  • dentry_lru: Dumps the CephFS dentries currently in-memory

  • mdsc: Dumps current requests to the MDS

  • mdsmap: Dumps the current MDSMap epoch and MDSes

  • mds_sessions: Dumps the current sessions to MDSes

  • monc: Dumps the current maps from the monitor, and any “subscriptions” held

  • monmap: Dumps the current monitor map epoch and monitors

  • osdc: Dumps the current ops in-flight to OSDs (ie, file data IO)

  • osdmap: Dumps the current OSDMap epoch, pools, and OSDs

If the data pool is in a NEARFULL condition, then the kernel cephfs client will switch to doing writes synchronously, which is quite slow.

Disconnected+Remounted FS

Because CephFS has a “consistent cache”, if your network connection is disrupted for a long enough time, the client will be forcibly disconnected from the system. At this point, the kernel client is in a bind: it cannot safely write back dirty data, and many applications do not handle IO errors correctly on close(). At the moment, the kernel client will remount the FS, but outstanding file system IO may or may not be satisfied. In these cases, you may need to reboot your client system.

You can identify you are in this situation if dmesg/kern.log report something like:

Jul 20 08:14:38 teuthology kernel: [3677601.123718] ceph: mds0 closed our session
Jul 20 08:14:38 teuthology kernel: [3677601.128019] ceph: mds0 reconnect start
Jul 20 08:14:39 teuthology kernel: [3677602.093378] ceph: mds0 reconnect denied
Jul 20 08:14:39 teuthology kernel: [3677602.098525] ceph:  dropping dirty+flushing Fw state for ffff8802dc150518 1099935956631
Jul 20 08:14:39 teuthology kernel: [3677602.107145] ceph:  dropping dirty+flushing Fw state for ffff8801008e8518 1099935946707
Jul 20 08:14:39 teuthology kernel: [3677602.196747] libceph: mds0 172.21.5.114:6812 socket closed (con state OPEN)
Jul 20 08:14:40 teuthology kernel: [3677603.126214] libceph: mds0 172.21.5.114:6812 connection reset
Jul 20 08:14:40 teuthology kernel: [3677603.132176] libceph: reset on mds0

This is an area of ongoing work to improve the behavior. Kernels will soon be reliably issuing error codes to in-progress IO, although your application(s) may not deal with them well. In the longer-term, we hope to allow reconnect and reclaim of data in cases where it won’t violate POSIX semantics (generally, data which hasn’t been accessed or modified by other clients).

Mounting

Mount 5 Error

A mount 5 error typically occurs if a MDS server is laggy or if it crashed. Ensure at least one MDS is up and running, and the cluster is active + healthy.

Mount 12 Error

A mount 12 error with cannot allocate memory usually occurs if you have a version mismatch between the Ceph Client version and the Ceph Storage Cluster version. Check the versions using:

ceph -v

If the Ceph Client is behind the Ceph cluster, try to upgrade it:

sudo apt-get update && sudo apt-get install ceph-common

You may need to uninstall, autoclean and autoremove ceph-common and then reinstall it so that you have the latest version.

Dynamic Debugging

You can enable dynamic debug against the CephFS module.

Please see: https://github.com/ceph/ceph/blob/master/src/script/kcon_all.sh

In-memory Log Dump

In-memory logs can be dumped by setting mds_extraordinary_events_dump_interval during a lower level debugging (log level < 10). mds_extraordinary_events_dump_interval is the interval in seconds for dumping the recent in-memory logs when there is an Extra-Ordinary event.

The Extra-Ordinary events are classified as:

  • Client Eviction

  • Missed Beacon ACK from the monitors

  • Missed Internal Heartbeats

In-memory Log Dump is disabled by default to prevent log file bloat in a production environment. The below commands consecutively enables it:

$ ceph config set mds debug_mds <log_level>/<gather_level>
$ ceph config set mds mds_extraordinary_events_dump_interval <seconds>

The log_level should be < 10 and gather_level should be >= 10 to enable in-memory log dump. When it is enabled, the MDS checks for the extra-ordinary events every mds_extraordinary_events_dump_interval seconds and if any of them occurs, MDS dumps the in-memory logs containing the relevant event details in ceph-mds log.

Note

For higher log levels (log_level >= 10) there is no reason to dump the In-memory Logs and a lower gather level (gather_level < 10) is insufficient to gather In-memory Logs. Thus a log level >=10 or a gather level < 10 in debug_mds would prevent enabling the In-memory Log Dump. In such cases, when there is a failure it’s required to reset the value of mds_extraordinary_events_dump_interval to 0 before enabling using the above commands.

The In-memory Log Dump can be disabled using:

$ ceph config set mds mds_extraordinary_events_dump_interval 0

Filesystems Become Inaccessible After an Upgrade

Note

You can avoid operation not permitted errors by running this procedure before an upgrade. As of May 2023, it seems that operation not permitted errors of the kind discussed here occur after upgrades after Nautilus (inclusive).

IF

you have CephFS file systems that have data and metadata pools that were created by a ceph fs new command (meaning that they were not created with the defaults)

OR

you have an existing CephFS file system and are upgrading to a new post-Nautilus major version of Ceph

THEN

in order for the documented ceph fs authorize... commands to function as documented (and to avoid ‘operation not permitted’ errors when doing file I/O or similar security-related problems for all users except the client.admin user), you must first run:

ceph osd pool application set <your metadata pool name> cephfs metadata <your ceph fs filesystem name>

and

ceph osd pool application set <your data pool name> cephfs data <your ceph fs filesystem name>

Otherwise, when the OSDs receive a request to read or write data (not the directory info, but file data) they will not know which Ceph file system name to look up. This is true also of pool names, because the ‘defaults’ themselves changed in the major releases, from:

data pool=fsname
metadata pool=fsname_metadata

to:

data pool=fsname.data and
metadata pool=fsname.meta

Any setup that used client.admin for all mounts did not run into this problem, because the admin key gave blanket permissions.

A temporary fix involves changing mount requests to the ‘client.admin’ user and its associated key. A less drastic but half-fix is to change the osd cap for your user to just caps osd = "allow rw" and delete tag cephfs data=....

Reporting Issues

If you have identified a specific issue, please report it with as much information as possible. Especially important information:

  • Ceph versions installed on client and server

  • Whether you are using the kernel or fuse client

  • If you are using the kernel client, what kernel version?

  • How many clients are in play, doing what kind of workload?

  • If a system is ‘stuck’, is that affecting all clients or just one?

  • Any ceph health messages

  • Any backtraces in the ceph logs from crashes

If you are satisfied that you have found a bug, please file it on the bug tracker. For more general queries, please write to the ceph-users mailing list.

Brought to you by the Ceph Foundation

The Ceph Documentation is a community resource funded and hosted by the non-profit Ceph Foundation. If you would like to support this and our other efforts, please consider joining now.