Note that in addition to the below table, block capabilities can be implemented below the file system layer in Linux (LVM, integritysetup, cryptsetup) or Windows (Volume Shadow Copy Service, SECURITY), etc.
While storage devices usually have their size expressed in powers of 10 (for instance a 1 TB Solid State Drive will contain at least 1,000,000,000,000 (1012, 10004) bytes), filesystem limits are invariably powers of 2, so usually expressed with IEC prefixes. For instance, a 1 TiB limit means 240, 10244 bytes. Approximations (rounding down) using power of 10 are also given below to clarify.
^IBM introduced JFS with the initial release of AIX Version 3.1 in 1990. This file system now called JFS1. The new JFS, on which the Linux port was based, was first shipped in OS/2 Warp Server for e-Business in 1999. The same sourcebase was also used for release JFS2 on AIX 5L.
^Microsoft first introduced FAT32 in MS-DOS 7.1 / Windows 95 OSR2 (OEM Service Release 2) and then later in Windows 98. NT-based Windows did not have any support for FAT32 up to Windows NT4; Windows 2000 was the first NT-based Windows OS that received the ability to work with it.
^Some FAT implementations, such as in Linux, show file modification timestamp (mtime) in the metadata change timestamp (ctime) field. This timestamp is however, not updated on file metadata change.
^Particular Installable File System drivers and operating systems may not support extended attributes on FAT12 and FAT16. The OS/2 and Windows NT filesystem drivers for FAT12 and FAT16 support extended attributes (using a "EA DATA. SF" pseudo-file to reserve the clusters allocated to them). Other filesystem drivers for other operating systems do not.
^The f-node contains a field for a user identifier. This is not used except by OS/2 Warp Server, however.
^NTFS access control lists can express any access policy possible using simple POSIX file permissions (and far more), but use of a POSIX-like interface is not supported without an add-on such as Services for UNIX or Cygwin.
^Initially, ReFS lacked support for ADS, but Server 2012 R2 and up add support for ADS on ReFS
^ abcdAccess-control lists and MAC labels are layered on top of extended attributes.
^Some operating systems implemented extended attributes as a layer over UFS1 with a parallel backing file (e.g., FreeBSD 4.x).
^ abcdefghijklmnSome Installable File System drivers and operating systems may not support extended attributes, access control lists or security labels on these filesystems. Linux kernels prior to 2.6.x may either be missing support for these altogether or require a patch.
^ Metadata is mostly checksummed,[13] however Direct/indirect/triple-indirect block maps are not protected by checksums[14]
^Creation time stored since June 2015, xfsprogs version 3.2.3
^ abcdefThe local time, time zone/UTC offset, and date are derived from the time settings of the reference/single timesync source in the NDS tree.
^ abNovell calls this feature "multiple data streams". Published specifications say that NWFS allows for 16 attributes and 10 data streams, and NSS allows for unlimited quantities of both.
^ abSome file and directory metadata is stored on the NetWare server irrespective of whether Directory Services is installed or not, like date/time of creation, file size, purge status, etc; and some file and directory metadata is stored in NDS/eDirectory, like file/object permissions, ownership, etc.
^Record Management Services (RMS) attributes include record type and size, among many others.
^File permission in 9P are a variation of the traditional Unix permissions with some minor changes, e.g. the suid bit is replaced by a new 'exclusive access' bit.
^Supported on FreeBSD and Linux implementations, support may not be available on all operating systems.
^Solaris "extended attributes" are really full-blown alternate data streams, in both the Solaris UFS and ZFS.
^Access times are preserved from the original file system at creation time, but Rock Ridge file systems themselves are read-only.
^libburnia can back up and restore ACLs with file system creation and extraction programs, but no kernel support exists.
^ ablibburnia can back up and restore extended attributes and MAC labels with file system creation and extraction programs, but no kernel support exists.
^System V Release 4, and some other Unix systems, retrofitted symbolic links to their versions of the Version 7 Unix file system, although the original version didn't support them.
^Context based symlinks were supported in GFS, GFS2 only supports standard symlinks since the bind mount feature of the Linux VFS has made context based symlinks obsolete
^As of Windows Vista, NTFS fully supports symbolic links.[15] NTFS 3.0 (Windows 2000) and higher can create junctions, which allow entire directories (but not individual files) to be mapped to elsewhere in the directory tree of the same partition (file system). These are implemented through reparse points, which allow the normal process of filename resolution to be extended in a flexible manner.
^ abNTFS stores everything, even the file data, as meta-data, so its log is closer to block journaling.
^ abWhile NTFS itself supports case sensitivity, the Win32 environment subsystem cannot create files whose names differ only by case for compatibility reasons. When a file is opened for writing, if there is any existing file whose name is a case-insensitive match for the new file, the existing file is truncated and opened for writing instead of a new file with a different name being created. Other subsystems like e. g. Services for Unix, that operate directly above the kernel and not on top of Win32 can have case-sensitivity.
^Metadata-only journaling was introduced in the Mac OS X 10.2.2 HFS Plus driver; journaling is enabled by default on Mac OS X 10.3 and later.
^Although often believed to be case sensitive, HFS Plus normally is not. The typical default installation is case-preserving only. From Mac OS X 10.3 on the command newfs_hfs -s will create a case-sensitive new file system.[17] HFS Plus version 5 optionally supports case-sensitivity. However, since case-sensitivity is fundamentally different from case-insensitivity, a new signature was required so existing HFS Plus utilities would not see case-sensitivity as a file system error that needed to be corrected. Since the new signature is 'HX', it is often believed this is a new filesystem instead of a simply an upgraded version of HFS Plus.[18][19]
^Mac OS X Tiger (10.4) and late versions of Panther (10.3) provide file change logging (it's a feature of the file system software, not of the volume format, actually).[20]
^Particular Installable File System drivers and operating systems may not support case sensitivity for JFS. OS/2 does not, and Linux has a mount option for disabling case sensitivity.
^ abcdCase-sensitivity/Preservation depends on client. Windows, DOS, and OS/2 clients don't see/keep case differences, whereas clients accessing via NFS or AFP may.
^ abThe file change logs, last entry change timestamps, and other filesystem metadata, are all part of the extensive suite of auditing capabilities built into NDS/eDirectory called NSure Audit.[26]
^ abZFS is a transactional filesystem using copy-on-write semantics, guaranteeing an always-consistent on-disk state without the use of a traditional journal. However, it does also implement an intent log to provide better performance when synchronous writes are requested.
^Btrfs is a transactional filesystem using copy-on-write semantics, guaranteeing an always-consistent on-disk state without the use of a traditional journal. It keeps track of last five transactions and uses checksums to find problematic drives, making write intent logs unnecessary.
^Bcachefs is a transactional filesystem using copy-on-write semantics, guaranteeing an always-consistent on-disk state without the use of a traditional journal. Journal commits are fairly expensive operations as they require issuing FLUSH and FUA operations to the underlying devices. By default, a journal flush is issued one second after a filesystem update has been done, which primarily records btree updates ordered by when they occurred. This option may be useful on a personal workstation or laptop, and perhaps less appropriate on a server.
^ Since Windows 10 Enterprise Insider Preview build 19536
^ abA file system is self-healing if its capable to proactively autonomously detect and correct all but grave errors, faults and corruptions online both in internal metadata AND data. See US7694191B1 as example. This usually requires full checksumming as well as internal redundancy as well as corresponding logic.
^Supported only on Windows Server SKUs. However, partitions deduplicated on Server can be used on Client.
^HFS+ does not actually encrypt files: to implement FileVault, OS X creates an HFS+ filesystem in a sparse, encrypted disk image that is automatically mounted over the home directory when the user logs in.
^Reiser4 supports transparent compression and encryption with the cryptcompress plugin which is the default file handler in version 4.1.
^VxFS provides an optional feature called "Storage Checkpoints" which allows for advanced file system snapshots.
^Applies to proprietary ZFS release 30 and ZFS On Linux. Encryption support is not yet available in all OpenZFS ports.[37][38][39]
^LZJB (optimized for performance while providing decent data compression) LZ4 (faster & higher ratio than lzjb) gzip levels: 1 (fastest) to 9 (best), default is 6 zstd positive: 1 (fastest) to 19 (best), default is 3 zstd negative: 1(best & default)-10, 20, 30, …, 100, 500, 1000(fastest) zle: compresses runs of zeros.[40]
^ disabling copy-on-write (COW) to prevent fragmentation also disables data checksumming
^zlib levels: 1 to 9, default is 3 LZO (no levels) faster than ZLIB, worse ratio zstd levels: 1 to 15, default is 3 (higher levels are not available)[41]
^ none CRC-32C (default) crc64 chacha20/poly1305 (When encryption is enabled. Encryption can only be specified for the entire filesystem, not per file or directory)[42]
^ none (default) The three currently supported algorithms are gzip, LZ4, zstd. The compression level may also be optionally specified, as an integer between 0 and 15, e.g. lz4:15. 0 specifies the default compression level, 1 specifies the fastest and lowest compression ratio, and 15 the slowest and best compression ratio.[43]
^ abBy using the per-file "integrity stream" that internally stores a checksum per cluster. Those per cluster checksums are not accessible so it is actually a per file feature and not a per block feature. Integrity streams are not enabled by default.[45]
^* 3.9: Added post process compression with LZ4 and ZSTD and transparent decompression.
^ abcSome file system creation implementations reuse block references and support deduplication this way. This is not supported by the standard, but usually works well due to the file system's read-only nature.
^ abVariable block size refers to systems which support different block sizes on a per-file basis. (This is similar to extents but a slightly different implementational choice.) The current implementation in UFS2 is read-only.
^Each possible size (in sectors) of file tail has a corresponding suballocation block chain in which all the tails of that size are stored. The overhead of managing suballocation block chains is usually less than the amount of block overhead saved by being able to increase the block size but the process is less efficient if there is not much free disk space.
^ abFUSE based driver available that can eliminate need for iSCSI gateways or SMB shares, but the physical backend store BlueStore only runs on Linux.
^Filesystem driver "Dokany" available that can eliminate need for iSCSI gateways or SMB shares, but the physical backend store BlueStore only runs on Linux.
^ abcdeThese are the restrictions imposed by the on-disk directory entry structures themselves. Particular Installable File System drivers may place restrictions of their own on file and directory names; operating systems may also place restrictions of their own, across all filesystems. DOS, Windows, and OS/2 allow only the following characters from the current 8-bit OEM codepage in SFNs: A-Z, 0-9, characters ! # $ % & ' ( ) - @ ^ _ ` { } ~, as well as 0x80-0xFF and 0x20 (SPACE). Specifically, lowercase letters a-z, characters " * / : < > ? \ | + , . ; = [ ], control codes 0x00-0x1F, 0x7F and in some cases also 0xE5 are not allowed.) In LFNs, any UCS-2Unicode except \ / : ? * " > < | and NUL are allowed in file and directory names across all filesystems. Unix-like systems disallow the characters / and NUL in file and directory names across all filesystems.
^ abcdeFor filesystems that have variable allocation unit (block/cluster) sizes, a range of size are given, indicating the maximum volume sizes for the minimum and the maximum possible allocation unit sizes of the filesystem (e.g. 512 bytes and 128 KiB (131.0 KB) for FAT — which is the cluster size range allowed by the on-disk data structures, although some Installable File System drivers and operating systems do not support cluster sizes larger than 32 KiB (32.76 KB)).
^ abcdefghijklmnopqrstuvwxyzaaabacadaeafagahaiajakalIn these filesystems the directory entries named "." and ".." have special status. Directory entries with these names are not prohibited, and indeed exist as normal directory entries in the on-disk data structures. However, they are mandatory directory entries, with mandatory values, that are automatically created in each directory when it is created; and directories without them are considered corrupt.
^ abcdefghijklmnopqrstuvwxyzaaabacadaeafagahaiajakalamanThe on-disk structures have no inherent limit. Particular Installable File System drivers and operating systems may impose limits of their own, however. Limited by its Current Directory Structure (CDS), DOS does not support more than 32 directory levels (except for DR DOS 3.31-6.0) or full pathnames longer than 66 bytes for FAT, or 255 characters for LFNs. Windows NT does not support full pathnames longer than 32,767 bytes for NTFS. Older POSIX APIs which rely on the PATH_MAX constant have a limit of 4,096 bytes on Linux but this can be worked around. Linux itself has no hard path length limits.[122][123]
^Varies wildly according to block size and fragmentation of block allocation groups.
^ abcdefDepends on whether the FAT12, FAT16, and FAT32implementation has support for LFNs. Where it does not, as in OS/2, DOS, Windows 95, Windows 98 in DOS-only mode and the Linux "msdos" driver, file names are limited to 8.3 format of 8-bit OEM characters (space padded in both the basename and extension parts) and may not contain NUL (end-of-directory marker) or character 5 (replacement for character 229 which itself is used as deleted-file marker). Short names also must not contain lowercase letters. A few special device names (CON, NUL, AUX, PRN, LPT1, COM1, etc.) should be avoided, as some operating systems (notably DOS, OS/2 and Windows) reserve them.
^ abOn-disk structures would support up to 4 GiB (4.294 GB), but practical file size is limited by volume size.
^While FAT32 partitions this large work fine once created, some software won't allow creation of FAT32 partitions larger than 32 GiB (34.35 GB). This includes, notoriously, the Windows XP installation program and the Disk Management console in Windows 2000, XP, 2003 and Vista. Use FDISK from a Windows ME Emergency Boot Disk to avoid.[104]
^Depends on CPU arch. For 32bit kernels the max is 16 TiB (17.59 TB). [106]
^Depends on CPU arch. For 32bit kernels the max is 16 TiB (17.59 TB). [107]
^ abDepends on kernel version and arch. For 2.4 kernels the max is 2 TiB (2.199 TB). For 32-bit 2.6 kernels it is 16 TiB (17.59 TB). For 64-bit 2.6 kernels it is 8 EiB (9.223 EB).
^The "classic" Mac OS provides two sets of functions to retrieve file names from an HFS Plus volume, one of them returning the full Unicode names, the other shortened names fitting in the older 31 byte limit to accommodate older applications.
^HFS Plus mandates support for an escape sequence to allow arbitrary Unicode. Users of older software might see the escape sequences instead of the desired characters.
^The "." and ".." directory entries in HPFS that are seen by applications programs are a partial fiction created by the Installable File System drivers. The on-disk data structure for a directory does not contain entries by those names, but instead contains a special "start" entry. Whilst on-disk directory entries by those names are not physically prohibited, they cannot be created in normal operation, and a directory containing such entries is corrupt.
^This is the limit of the on-disk structures. The HPFS Installable File System driver for OS/2 uses the top 5 bits of the volume sector number for its own use, limiting the volume size that it can handle to 64 GiB (68.71 GB).
^Assuming the typical 2048 Byte sector size. The volume size is specified as a 32 bit value identifying the number of sectors on the volume.
^ abSparse files can be larger than the file system size, even though they can't contain more data.
^ abNSS allows files to have multiple names, in separate namespaces.
^ abThis is the limit of the on-disk structures. The NTFS driver for Windows NT limits the volume size that it can handle to 256 TiB (281.4 TB) and the file size to 16 TiB (17.59 TB) respectively; in Windows 10 version 1709, the limit is 8 PiB (9.007 PB) when using 2 MiB (2.097 MB) cluster size.
^Some namespaces had lower name length limits. "LONG" had an 80-byte limit, "NWFS" 80 bytes, "NFS" 40 bytes and "DOS" imposed 8.3 filename.
^Maximum combined filename/filetype length is 236 bytes; each component has an individual maximum length of 255 bytes.
^Maximum pathname length is 4,096 bytes, but quoted limits on individual components add up to 1,664 bytes.
^ abQFS allows files to exceed the size of disk when used with its integrated HSM, as only part of the file need reside on disk at any one time.
^ReiserFS has a theoretical maximum file size of 1 EiB (1.152 EB), but "page cache limits this to 8 Ti on architectures with 32 bit int"[119]
^This restriction might be lifted in newer versions.
^The file size in the inode is 1 8-bit byte followed by 1 16-bit word, for 24 bits. The actual maximum was 8,847,360 bytes, with 7 singly-indirect blocks and 1 doubly-indirect block; PWB/UNIX 1.0's variant had 8 singly-indirect blocks, making the maximum 524,288 bytes or half a MB.
^The actual maximum was 1,082,201,088 bytes, with 10 direct blocks, 1 singly-indirect block, 1 doubly-indirect block, and 1 triply-indirect block. The 4.0BSD and 4.1BSD versions, and the System V version, used 1,024-byte blocks rather than 512-byte blocks, making the maximum 4,311,812,608 bytes or approximately 4 GiB (4.294 GB).
^ abMaximum file size on a VMFS volume depends on the block size for that VMFS volume. The figures here are obtained by using the maximum block size.
^ abXFS has a limitation under Linux 2.4 of 64 TiB (70.36 TB) file size, but Linux 2.4 only supports a maximum block size of 2 TiB (2.199 TB). This limitation is not present under IRIX.
^See dinode structure on page 355 (FILESYS(5)) of "Unix Programmers Manual"(PDF) (Seventh ed.). Murray Hill, New Jersey: Bell Telephone Laboratories. January 1979. Retrieved 2016-02-21.
^Siracusa, John (2011-07-20). "Mac OS X 10.7 Lion: the Ars Technica review". Ars Technica. Retrieved 14 December 2017. To keep track of hard links, HFS+ creates a separate file for each hard link inside a hidden directory at the root level of the volume.
^"clonefile(2)". The cloned file dst shares its data blocks with the src file [..]
^ abc"DMSDOS CVF module" (dmsdoc.doc). 0.9.2.0. 1998-11-19. Archived from the original on 2016-11-02. Retrieved 2016-11-01. Usually all data for one cluster are stored in contiguous sectors, but if the filesystem is too fragmented there may not be a 'free hole' that is large enough for the data. […] Drivespace 3 and Stacker know a hack for that situation: they allow storing the data of one cluster in several fragments on the disk.
^ ab"FAQ". Ext2 Installable File System For Windows. (Provides kernel level read/write access to Ext2 and Ext3 volumes in Windows NT4, 2000, XP and Vista.)
^"Frequently Asked Questions for FreeBSD 9.X and 10.X". FreeBSD Documentation Project. Retrieved 2016-03-20. If there was not a fsck(8) memory limit the maximum filesystem size would be 2 ^ 64 (blocks) * 32 KiB (32.76 KB) => 16 Exa * 32 KiB (32.76 KB) => 512 ZettaBytes.