3.10. Migrating from ext4 to XFS

Ext3/4 runs e2fsck in userspace at boot time to recover the journal as needed. XFS, by comparison, performs journal recovery in kernelspace at mount time. An fsck.xfs shell script is provided but does not perform any useful action as it is only there to satisfy initscript requirements.

When an XFS file system repair or check is requested, use the xfs_repair command. Use the -n option for a read-only check.

The xfs_repair command will not operate on a file system with a dirty log. To repair such a file system mount and unmount must first be performed to replay the log. If the log is corrupt and cannot be replayed, the -L option can be used to zero out in the log.

For more information on file system repair of XFS file systems, see Section 12.2.2, “XFS”

The ext3/4 file system has configurable behavior when metadata errors are encountered, with the default being to simply continue. When XFS encounters a metadata error that is not recoverable it will shut down the file system and return a EFSCORRUPTED error. The system logs will contain details of the error encountered and will recommend running xfs_repair if necessary.

XFS quotas are not a remountable option. The -o quota option must be specified on the initial mount for quotas to be in effect.

While the standard tools in the quota package can perform basic quota administrative tasks (tools such as setquota and repquota), the xfs_quota tool can be used for XFS-specific features, such as Project Quota administration.

The quotacheck command has no effect on an XFS file system. The first time quota accounting is turned on XFS does an automatic quotacheck internally. Because XFS quota metadata is a first-class, journaled metadata object, the quota system will always be consistent until quotas are manually turned off.

The XFS file system has no utility to shrink a file system. XFS file systems can be grown online via the xfs_growfs command.

For file systems larger than 1 TB with 256-byte inodes, or larger than 2 TB with 512-byte inodes, XFS inode numbers might exceed 2^32. Such large inode numbers cause 32-bit stat calls to fail with the EOVERFLOW return value. The described problem might occur when using the default Red Hat Enterprise Linux 7 configuration: non-striped with four allocation groups. A custom configuration, for example file system extension or changing XFS file system parameters, might lead to a different behavior.

Applications usually handle such larger inode numbers correctly. If needed, mount the XFS file system with the -o inode32 parameter to enforce inode numbers below 2^32. Note that using inode32 does not affect inodes that are already allocated with 64-bit numbers.

XFS uses speculative preallocation to allocate blocks past EOF as files are written. This avoids file fragmentation due to concurrent streaming write workloads on NFS servers. By default, this preallocation increases with the size of the file and will be apparent in "du" output. If a file with speculative preallocation is not dirtied for five minutes the preallocation will be discarded. If the inode is cycled out of cache before that time, then the preallocation will be discarded when the inode is reclaimed.

If premature ENOSPC problems are seen due to speculative preallocation, a fixed preallocation amount may be specified with the -o allocsize=amount mount option.

Fragmentation is rarely a significant issue on XFS file systems due to heuristics and behaviors, such as delayed allocation and speculative preallocation. However, tools exist for measuring file system fragmentation as well as defragmenting file systems. Their use is not encouraged.

The xfs_db frag command attempts to distill all file system allocations into a single fragmentation number, expressed as a percentage. The output of the command requires significant expertise to understand its meaning. For example, a fragmentation factor of 75% means only an average of 4 extents per file. For this reason the output of xfs_db's frag is not considered useful and more careful analysis of any fragmentation problems is recommended.