Ceph » CephFS

(formatting description)

Its rather easy to end up with an on-disk inode object that stays un-fragmented, with omap count exceeding `mds_bal_split_size' (default: 10000 entries). This can be achieved as follows:

1. mkdir d0
2. touch d0/file{0..11000}
3. mkdir d0/.snap/snap_a
4. rm -f d0/file*
5. #rmdir d0/.snap/snap_a

Since the MDS does not fragment directory snapshots, the omap count for the directory inode is 11000. Doing this over and over again increases the omap count to a point if it exceeds `osd_deep_scrub_large_omap_object_key_threshold' (default: 200000), the cluster experiences "large omap" warnings with state HEALTH_WARN. Removing the snapshot does not purge stale omap entries and those stay there until the MDS loads the dirfrag.

Provide an asok/tell command to force the MDS to clear stale omap entries.