ceph/doc/cephfs/disaster-recovery.rst

Disaster recovery
=================

.. danger::

    The notes in this section are aimed at experts, making a best effort
    to recovery what they can from damaged filesystems.  These steps
    have the potential to make things worse as well as better.  If you
    are unsure, do not proceed.


Journal export
--------------

Before attempting dangerous operations, make a copy of the journal like so:

::

    cephfs-journal-tool journal export backup.bin

Note that this command may not always work if the journal is badly corrupted,
in which case a RADOS-level copy should be made (http://tracker.ceph.com/issues/9902).


Dentry recovery from journal
----------------------------

If a journal is damaged or for any reason an MDS is incapable of replaying it,
attempt to recover what file metadata we can like so:

::

    cephfs-journal-tool event recover_dentries summary

This command by default acts on MDS rank 0, pass --rank=<n> to operate on other ranks.

This command will write any inodes/dentries recoverable from the journal
into the backing store, if these inodes/dentries are higher-versioned
than the previous contents of the backing store.  If any regions of the journal
are missing/damaged, they will be skipped.

Note that in addition to writing out dentries and inodes, this command will update
the InoTables of each 'in' MDS rank, to indicate that any written inodes' numbers
are now in use.  In simple cases, this will result in an entirely valid backing
store state.

.. warning::

    The resulting state of the backing store is not guaranteed to be self-consistent,
    and an online MDS scrub will be required afterwards.  The journal contents
    will not be modified by this command, you should truncate the journal
    separately after recovering what you can.

Journal truncation
------------------

If the journal is corrupt or MDSs cannot replay it for any reason, you can
truncate it like so:

::

    cephfs-journal-tool journal reset

.. warning::

    Resetting the journal *will* lose metadata unless you have extracted
    it by other means such as ``recover_dentries``.  It is likely to leave
    some orphaned objects in the data pool.  It may result in re-allocation
    of already-written inodes, such that permissions rules could be violated.

MDS table wipes
---------------

After the journal has been reset, it may no longer be consistent with respect
to the contents of the MDS tables (InoTable, SessionMap, SnapServer).

To reset the SessionMap (erase all sessions), use:

::

    cephfs-table-tool all reset session

This command acts on the tables of all 'in' MDS ranks.  Replace 'all' with an MDS
rank to operate on that rank only.

The session table is the table most likely to need resetting, but if you know you
also need to reset the other tables then replace 'session' with 'snap' or 'inode'.

MDS map reset
-------------

Once the in-RADOS state of the filesystem (i.e. contents of the metadata pool)
is somewhat recovered, it may be necessary to update the MDS map to reflect
the contents of the metadata pool.  Use the following command to reset the MDS
map to a single MDS:

::

    ceph fs reset <fs name> --yes-i-really-mean-it

Once this is run, any in-RADOS state for MDS ranks other than 0 will be ignored:
as a result it is possible for this to result in data loss.

One might wonder what the difference is between 'fs reset' and 'fs remove; fs new'.  The
key distinction is that doing a remove/new will leave rank 0 in 'creating' state, such
that it would overwrite any existing root inode on disk and orphan any existing files.  In
contrast, the 'reset' command will leave rank 0 in 'active' state such that the next MDS
daemon to claim the rank will go ahead and use the existing in-RADOS metadata.

Recovery from missing metadata objects
--------------------------------------

Depending on what objects are missing or corrupt, you may need to
run various commands to regenerate default versions of the
objects.

::

    # Session table
    cephfs-table-tool 0 reset session
    # SnapServer
    cephfs-table-tool 0 reset snap
    # InoTable
    cephfs-table-tool 0 reset inode
    # Journal
    cephfs-journal-tool --rank=0 journal reset
    # Root inodes ("/" and MDS directory)
    cephfs-data-scan init

Finally, you can regenerate metadata objects for missing files
and directories based on the contents of a data pool.  This is
a two-phase process.  First, scanning *all* objects to calculate
size and mtime metadata for inodes.  Second, scanning the first
object from every file to collect this metadata and inject
it into the metadata pool.

::

    cephfs-data-scan scan_extents <data pool>
    cephfs-data-scan scan_inodes <data pool>

This command may take a *very long* time if there are many
files or very large files in the data pool.

To accelerate the process, run multiple instances of the tool.

Decide on a number of workers, and pass each worker a number within
the range 0-(worker_m - 1).

The example below shows how to run 4 workers simultaneously:

::

    # Worker 0
    cephfs-data-scan scan_extents --worker_n 0 --worker_m 4 <data pool>
    # Worker 1
    cephfs-data-scan scan_extents --worker_n 1 --worker_m 4 <data pool>
    # Worker 2
    cephfs-data-scan scan_extents --worker_n 2 --worker_m 4 <data pool>
    # Worker 3
    cephfs-data-scan scan_extents --worker_n 3 --worker_m 4 <data pool>

    # Worker 0
    cephfs-data-scan scan_inodes --worker_n 0 --worker_m 4 <data pool>
    # Worker 1
    cephfs-data-scan scan_inodes --worker_n 1 --worker_m 4 <data pool>
    # Worker 2
    cephfs-data-scan scan_inodes --worker_n 2 --worker_m 4 <data pool>
    # Worker 3
    cephfs-data-scan scan_inodes --worker_n 3 --worker_m 4 <data pool>

It is **important** to ensure that all workers have completed the
scan_extents phase before any workers enter the scan_inodes phase.

After completing the metadata recovery, you may want to run cleanup
operation to delete ancillary data geneated during recovery.

::

    cephfs-data-scan cleanup <data pool>

Finding files affected by lost data PGs
---------------------------------------

Losing a data PG may affect many files.  Files are split into many objects,
so identifying which files are affected by loss of particular PGs requires
a full scan over all object IDs that may exist within the size of a file. 
This type of scan may be useful for identifying which files require
restoring from a backup.

.. danger::

    This command does not repair any metadata, so when restoring files in
    this case you must *remove* the damaged file, and replace it in order
    to have a fresh inode.  Do not overwrite damaged files in place.

If you know that objects have been lost from PGs, use the ``pg_files``
subcommand to scan for files that may have been damaged as a result:

::

    cephfs-data-scan pg_files <path> <pg id> [<pg id>...]

For example, if you have lost data from PGs 1.4 and 4.5, and you would like
to know which files under /home/bob might have been damaged:

::

    cephfs-data-scan pg_files /home/bob 1.4 4.5

The output will be a list of paths to potentially damaged files, one
per line.

Note that this command acts as a normal CephFS client to find all the
files in the filesystem and read their layouts, so the MDS must be
up and running.

Using an alternate metadata pool for recovery
---------------------------------------------

.. warning::

   There has not been extensive testing of this procedure. It should be
   undertaken with great care.

If an existing filesystem is damaged and inoperative, it is possible to create
a fresh metadata pool and attempt to reconstruct the filesystem metadata
into this new pool, leaving the old metadata in place. This could be used to
make a safer attempt at recovery since the existing metadata pool would not be
overwritten.

.. caution::

   During this process, multiple metadata pools will contain data referring to
   the same data pool. Extreme caution must be exercised to avoid changing the
   data pool contents while this is the case. Once recovery is complete, the
   damaged metadata pool should be deleted.

To begin this process, first create the fresh metadata pool and initialize
it with empty file system data structures:

::

    ceph fs flag set enable_multiple true --yes-i-really-mean-it
    ceph osd pool create recovery <pg-num> replicated <crush-ruleset-name>
    ceph fs new recovery-fs recovery <data pool> --allow-dangerous-metadata-overlay
    cephfs-data-scan init --force-init --filesystem recovery-fs --alternate-pool recovery
    ceph fs reset recovery-fs --yes-i-realy-mean-it
    cephfs-table-tool recovery-fs:all reset session
    cephfs-table-tool recovery-fs:all reset snap
    cephfs-table-tool recovery-fs:all reset inode

Next, run the recovery toolset using the --alternate-pool argument to output
results to the alternate pool:

::

    cephfs-data-scan scan_extents --alternate-pool recovery --filesystem <original filesystem name>
    cephfs-data-scan scan_inodes --alternate-pool recovery --filesystem <original filesystem name> --force-corrupt --force-init <original data pool name>

If the damaged filesystem contains dirty journal data, it may be recovered next
with:

::

    cephfs-journal-tool --rank=<original filesystem name>:0 event recover_dentries list --alternate-pool recovery
    cephfs-journal-tool --rank recovery-fs:0 journal reset --force

After recovery, some recovered directories will have incorrect link counts.
Ensure the parameter mds_debug_scatterstat is set to false (the default) to
prevent the MDS from checking the link counts, then run a forward scrub to
repair them. Ensure you have an MDS running and issue:

::

    ceph daemon mds.a scrub_path / recursive repair