Ceph » CephFS

This proposal is for working around https://tracker.ceph.com/issues/67102 ("mds: read-only file system due to large old_inodes"). Since hardlnks are disallowed across cephfs subvolumes (directory marked with ceph.dir.subvolume) xattr, it would suffice to use the sequence number of the subvolume directory snaprealm when deciding to COW old_inodes for files under the subvolume. The function of interest is MDCache::predirty_journal_parents() which traverse from the leaf inode to the root and for each ancestor path calls CInode::pre_cow_old_inode(). Currently, CInode::pre_cow_old_inode() uses the global snaprealm sequence number for deciding to COW the old_inode. So, for ancestors path component under the cephfs subvolume directory, use the subvolume directory snaprealm sequence number and once "outside" the subvolume directory, fallback using the global snaprealm sequence number. This would workaround the bug for users using the file system purely in subvolume mode, and preserve status-quo behaviour of others. Also, close the loop by disallowing marking a directory as a subvolume if the directory has existing files. This is bug BTW.

For now, it doesn't look like this functionality should be optional and only be enabled via a config switch - let's reassess that once we run tests and experiments.