Ext4 Disk Layout
|Line 864:||Line 864:|
<code>struct ext4_extent_tail</code> 4 bytes long:
Revision as of 22:48, 17 October 2012
This document attempts to describe the on-disk format for ext4 filesystems. The same general ideas should apply to ext2/3 filesystems as well, though they do not support all the features that ext4 supports, and the fields will be shorter.
NOTE: This is a work in progress, based on notes that the author (djwong) made while picking apart a filesystem by hand. The data structure definitions were pulled out of Linux 3.0 and e2fsprogs-1.41. He welcomes all comments and corrections, since there is undoubtedly plenty of lore that doesn't necessarily show up on freshly created demonstration filesystems.
ext4 divides a storage device into an array of logical blocks both to reduce bookkeeping overhead and to increase throughput by forcing larger transfer sizes. Generally, the block size will be 4KiB (the same size as pages on x86 and the block layer's default block size), though the actual size is calculated as 2 ^ (10 +
sb.s_log_block_size) bytes. Throughout this document, disk locations are given in terms of these logical blocks, not raw LBAs, and not 1024-byte blocks. For the sake of convenience, the logical block size will be referred to as
$block_size throughout the rest of the document.
When referenced in
preformatted text blocks,
sb refers to fields in the super block, and
inode refers to fields in an inode table entry.
An ext4 file system is split into a series of block groups. To reduce performance difficulties due to fragmentation, the block allocator tries very hard to keep each file's blocks within the same group, thereby reducing seek times. The size of a block group is specified in
sb.s_blocks_per_group blocks, though it can also calculated as 8 *
block_size_in_bytes. With the default block size of 4KiB, each group will contain 32,768 blocks, for a length of 128MiB. The number of block groups is the size of the device divided by the size of a block group.
All fields in ext4 are written to disk in little-endian order. HOWEVER, all fields in jbd2 (the journal) are written to disk in big-endian order.
The layout of a standard block group is approximately as follows (each of these fields is discussed in a separate section below):
|Group 0 Padding||ext4 Super Block||Group Descriptors||Reserved GDT Blocks||Data Block Bitmap||inode Bitmap||inode Table||Data Blocks|
|1024 bytes||1 block||many blocks||many blocks||1 block||1 block||many blocks||many more blocks|
For the special case of block group 0, the first 1024 bytes are unused, to allow for the installation of x86 boot sectors and other oddities. The superblock will start at offset 1024 bytes, whichever block that happens to be (usually 0). However, if for some reason the block size = 1024, then block 0 is marked in use and the superblock goes in block 1. For all other block groups, there is no padding.
The ext4 driver primarily works with the superblock and the group descriptors that are found in block group 0. Redundant copies of the superblock and group descriptors are written to some of the block groups across the disk in case the beginning of the disk gets trashed, though not all block groups necessarily host a redundant copy (see following paragraph for more details). If the group does not have a redundant copy, the block group begins with the data block bitmap. Note also that when the filesystem is freshly formatted, mkfs will allocate "reserve GDT block" space after the block group descriptors and before the start of the block bitmaps to allow for future expansion of the filesystem. By default, a filesystem is allowed to increase in size by a factor of 1024x over the original filesystem size.
The location of the inode table is given by
grp.bg_inode_table_*. It is continuous range of blocks large enough to contain
As for the ordering of items in a block group, it is generally established that the super block and the group descriptor table, if present, will be at the beginning of the block group. The bitmaps and the inode table can be anywhere, and it is quite possible for the bitmaps to come after the inode table, or for both to be in different groups (flex_bg). Leftover space is used for file data blocks, indirect block maps, extent tree blocks, and extended attributes.
Flexible Block Groups
Starting in ext4, there is a new feature called flexible block groups (flex_bg). In a flex_bg, several block groups are tied together as one logical block group; the bitmap spaces and the inode table space in the first block group of the flex_bg are expanded to include the bitmaps and inode tables of all other block groups in the flex_bg. For example, if the flex_bg size is 4, then group 0 will contain (in order) the superblock, group descriptors, data block bitmaps for groups 0-3, inode bitmaps for groups 0-3, inode tables for groups 0-3, and the remaining space in group 0 is for file data. The effect of this is to group the block metadata close together for faster loading, and to enable large files to be continuous on disk. Backup copies of the superblock and group descriptors are always at the beginning of block groups, even if flex_bg is enabled. The number of block groups that make up a flex_bg is given by 2 ^
Meta Block Groups
Normally, a complete copy of the entire block group descriptor table is recorded after every copy of the superblock. Assuming the default group size of 2^27 bytes (128MiB) and 64-byte group descriptors, this imposes a limitation of 2^21 block groups, or 256TiB. With the meta block group feature enabled, each block group contains redundant copies of the block group descriptor for that group, thereby enabling the creation of the full 2^32 block groups, for a total size of 512EiB.
Lazy Block Group Initialization
New also for ext4, the inode bitmap and inode tables in a group are uninitialized if the corresponding flag is set in the group descriptor. This is to reduce mkfs time considerably. If the group descriptor checksum feature is enabled, then even the group descriptors can be uninitialized.
ext4 reserves some inode for special features, as follows:
|0||Doesn't exist; there is no inode 0.|
|1||List of defective blocks.|
|7||Reserved group descriptors inode.|
|11||First non-reserved inode. Usually this is the lost+found directory.|
Block and Inode Allocation Policy
ext4 recognizes (better than ext3, anyway) that data locality is generally a desirably quality of a filesystem. On a spinning disk, keeping related blocks near each other reduces the amount of movement that the head actuator and disk must perform to access a data block, thus speeding up disk IO. On an SSD there of course are no moving parts, but locality can increase the size of each transfer request while reducing the total number of requests. This locality may also have the effect of concentrating writes on a single erase block, which can speed up file rewrites significantly. Therefore, it is useful to reduce fragmentation whenever possible.
The first tool that ext4 uses to combat fragmentation is the multi-block allocator. When a file is first created, the block allocator speculatively allocates 8KiB of disk space to the file on the assumption that the space will get written soon. When the file is closed, the unused speculative allocations are of course freed, but if the speculation is correct (typically the case for full writes of small files) then the file data gets written out in a single multi-block extent. A second related trick that ext4 uses is delayed allocation. Under this scheme, when a file needs more blocks to absorb file writes, the filesystem defers deciding the exact placement on the disk until all the dirty buffers are being written out to disk. By not committing to a particular placement until it's absolutely necessary (the commit timeout is hit, or sync() is called, or the kernel runs out of memory), the hope is that the filesystem can make better location decisions.
The third trick that ext4 (and ext3) uses is that it tries to keep a file's data blocks in the same block group as its inode. This cuts down on the seek penalty when the filesystem first has to read a file's inode to learn where the file's data blocks live and then seek over to the file's data blocks to begin I/O operations.
The fourth trick is that all the inodes in a directory are placed in the same block group as the directory, when feasible. The working assumption here is that all the files in a directory might be related, therefore it is useful to try to keep them all together.
The fifth trick is that the disk volume is cut up into 128MB block groups; these mini-containers are used as outlined above to try to maintain data locality. However, there is a deliberate quirk -- when a directory is created in the root directory, the inode allocator scans the block groups and puts that directory into the least heavily loaded block group that it can find. This encourages directories to spread out over a disk; as the top-level directory/file blobs fill up one block group, the allocators simply move on to the next block group. Allegedly this scheme evens out the loading on the block groups, though the author suspects that the directories which are so unlucky as to land towards the end of a spinning drive get a raw deal performance-wise.
Of course if all of these mechanisms fail, one can always use e4defrag to defragment files.
Starting in early 2012, metadata checksums were added to all major ext4 and jbd2 data structures. The associated feature flag is metadata_csum. The desired checksum algorithm is indicated in the superblock, though as of October 2012 the only supported algorithm is crc32c. Some data structures did not have space to fit a full 32-bit checksum, so only the lower 16 bits are stored. Enabling the 64bit feature increases the data structure size so that full 32-bit checksums can be stored; however, existing 32-bit filesystems cannot be extended to enable 64bit mode.
Existing filesystems can have checksumming added by running
tune2fs -O metadata_csum against the underlying device. If tune2fs encounters directory blocks that lack sufficient empty space to add a checksum, it will request that you run
e2fsck -D to have the directories rebuilt with checksums. This has the added benefit of removing slack space from the directory files and rebalancing the htree indexes. If you _ignore_ this step, your directories will not be protected by a checksum!
The Super Block
The superblock records various information about the enclosing filesystem, such as block counts, inode counts, supported features, maintenance information, and more.
If the sparse_super feature flag is set, redundant copies of the superblock and group descriptors are kept only in the groups whose group number is either 0 or a power of 3, 5, or 7. If the flag is not set, redundant copies are kept in all groups.
The superblock checksum is calculated against the superblock structure, which includes the FS UUID.
The ext4 superblock is laid out as follows in
|0x0||__le32||s_inodes_count||Total inode count.|
|0x4||__le32||s_blocks_count_lo||Total block count.|
|0x8||__le32||s_r_blocks_count_lo||Reserved block count.|
|0xC||__le32||s_free_blocks_count_lo||Free block count.|
|0x10||__le32||s_free_inodes_count||Free inode count.|
|0x14||__le32||s_first_data_block||First data block.|
|0x18||__le32||s_log_block_size||Block size is 2 ^ (10 + s_log_block_size).|
|0x1C||__le32||s_obso_log_frag_size||(Obsolete) fragment size.|
|0x20||__le32||s_blocks_per_group||Blocks per group.|
|0x24||__le32||s_obso_frags_per_group||(Obsolete) fragments per group.|
|0x28||__le32||s_inodes_per_group||Inodes per group.|
|0x2C||__le32||s_mtime||Mount time, in seconds since the epoch.|
|0x30||__le32||s_wtime||Write time, in seconds since the epoch.|
|0x34||__le16||s_mnt_count||Number of mounts since the last fsck.|
|0x36||__le16||s_max_mnt_count||Number of mounts beyond which a fsck is needed.|
|0x38||__le16||s_magic||Magic signature, 0xEF53|
|0x3A||__le16||s_state||File system state. Valid values are:
|0x3C||__le16||s_errors||Behaviour when detecting errors. One of:
|0x3E||__le16||s_minor_rev_level||Minor revision level.|
|0x40||__le32||s_lastcheck||Time of last check, in seconds since the epoch.|
|0x44||__le32||s_checkinterval||Maximum time between checks, in seconds.|
|0x48||__le32||s_creator_os||OS. One of:
|0x4C||__le32||s_rev_level||Revision level. One of:
|0x50||__le16||s_def_resuid||Default uid for reserved blocks.|
|0x52||__le16||s_def_resgid||Default gid for reserved blocks.|
|These fields are for EXT4_DYNAMIC_REV superblocks only.
Note: the difference between the compatible feature set and the incompatible feature set is that if there is a bit set in the incompatible feature set that the kernel doesn't know about, it should refuse to mount the filesystem.
e2fsck's requirements are more strict; if it doesn't know about a feature in either the compatible or incompatible feature set, it must abort and not try to meddle with things it doesn't understand...
|0x54||__le32||s_first_ino||First non-reserved inode.|
|0x58||__le16||s_inode_size||Size of inode structure, in bytes.|
|0x5A||__le16||s_block_group_nr||Block group # of this superblock.|
|0x5C||__le32||s_feature_compat||Compatible feature set flags. Kernel can still read/write this fs even if it doesn't understand a flag; fsck should not do that. Any of:
|0x60||__le32||s_feature_incompat||Incompatible feature set. If the kernel or fsck doesn't understand one of these bits, it should stop. Any of:
|0x64||__le32||s_feature_ro_compat||Readonly-compatible feature set. If the kernel doesn't understand one of these bits, it can still mount read-only. Any of:
|0x68||__u8||s_uuid||128-bit UUID for volume.|
|0x88||char||s_last_mounted||Directory where filesystem was last mounted.|
|Performance hints. Directory preallocation should only happen if the EXT4_FEATURE_COMPAT_DIR_PREALLOC flag is on.|
|0xCC||__u8||s_prealloc_blocks||# of blocks to try to preallocate for ... files?|
|0xCD||__u8||s_prealloc_dir_blocks||# of blocks to preallocate for directories.|
|0xCE||__le16||s_reserved_gdt_blocks||Number of reserved GDT entries for future filesystem expansion.|
|Journaling support valid if EXT4_FEATURE_COMPAT_HAS_JOURNAL set.|
|0xD0||__u8||s_journal_uuid||UUID of journal superblock|
|0xE0||__le32||s_journal_inum||inode number of journal file.|
|0xE4||__le32||s_journal_dev||Device number of journal file, if the external journal feature flag is set.|
|0xE8||__le32||s_last_orphan||Start of list of orphaned inodes to delete.|
|0xEC||__le32||s_hash_seed||HTREE hash seed.|
|0xFC||__u8||s_def_hash_version||Default hash algorithm to use for directory hashes. One of:
|0xFE||__le16||s_desc_size||Size of group descriptors, in bytes, if the 64bit incompat feature flag is set.|
|0x100||__le32||s_default_mount_opts||Default mount options. Any of:
|0x104||__le32||s_first_meta_bg||First metablock block group, if the meta_bg feature is enabled.|
|0x108||__le32||s_mkfs_time||When the filesystem was created, in seconds since the epoch.|
|0x10C||__le32||s_jnl_blocks||Backup copy of the first 68 bytes of the journal inode.|
|64bit support valid if EXT4_FEATURE_COMPAT_64BIT|
|0x150||__le32||s_blocks_count_hi||High 32-bits of the block count.|
|0x154||__le32||s_r_blocks_count_hi||High 32-bits of the reserved block count.|
|0x158||__le32||s_free_blocks_count_hi||High 32-bits of the free block count.|
|0x15C||__le16||s_min_extra_isize||All inodes have at least # bytes.|
|0x15E||__le16||s_want_extra_isize||New inodes should reserve # bytes.|
|0x160||__le32||s_flags||Miscellaneous flags. Any of:
|0x164||__le16||s_raid_stride||RAID stride. This is the number of logical blocks read from or written to the disk before moving to the next disk. This affects the placement of filesystem metadata, which will hopefully make RAID storage faster.|
|0x166||__le16||s_mmp_interval||# seconds to wait in multi-mount prevention (MMP) checking. In theory, MMP is a mechanism to record in the superblock which host and device have mounted the filesystem, in order to prevent multiple mounts. This feature does not seem to be implemented...|
|0x168||__le64||s_mmp_block||Block # for multi-mount protection data.|
|0x170||__le32||s_raid_stripe_width||RAID stripe width. This is the number of logical blocks read from or written to the disk before coming back to the current disk. This is used by the block allocator to try to reduce the number of read-modify-write operations in a RAID5/6.|
|0x174||__u8||s_log_groups_per_flex||Size of a flexible block group is 2 ^ |
|0x178||__le64||s_kbytes_written||Number of KiB written to this filesystem over its lifetime.|
|0x180||__le32||s_snapshot_inum||inode number of active snapshot.|
|0x184||__le32||s_snapshot_id||Sequential ID of active snapshot.|
|0x188||__le64||s_snapshot_r_blocks_count||Number of blocks reserved for active snapshot's future use.|
|0x190||__le32||s_snapshot_list||inode number of the head of the on-disk snapshot list.|
|0x194||__le32||s_error_count||Number of errors seen.|
|0x198||__le32||s_first_error_time||First time an error happened, in seconds since the epoch.|
|0x19C||__le32||s_first_error_ino||inode involved in first error.|
|0x1A0||__le64||s_first_error_block||Number of block involved of first error.|
|0x1A8||__u8||s_first_error_func||Name of function where the error happened.|
|0x1C8||__le32||s_first_error_line||Line number where error happened.|
|0x1CC||__le32||s_last_error_time||Time of most recent error, in seconds since the epoch.|
|0x1D0||__le32||s_last_error_ino||inode involved in most recent error.|
|0x1D4||__le32||s_last_error_line||Line number where most recent error happened.|
|0x1D8||__le64||s_last_error_block||Number of block involved in most recent error.|
|0x1E0||__u8||s_last_error_func||Name of function where the most recent error happened.|
|0x200||__u8||s_mount_opts||ASCIIZ string of mount options.|
|0x240||__le32||s_usr_quota_inum||Inode number of user quota file.|
|0x244||__le32||s_grp_quota_inum||Inode number of group quota file.|
|0x248||__le32||s_overhead_blocks||Overhead blocks/clusters in fs. (huh?)|
|0x24C||__le32||s_reserved||Padding to the end of the block.|
Total size is 1024 bytes.
Block Group Descriptors
Each block group on the filesystem has one of these descriptors associated with it. As noted in the Layout section above, the group descriptors (if present) are the second item in the block group. The standard configuration is for each block group to contain a full copy of the block group descriptor table unless the sparse_super feature flag is set.
Notice how the group descriptor records the location of both bitmaps and the inode table (i.e. they can float). This means that within a block group, the only data structures with fixed locations are the superblock and the group descriptor table. The flex_bg mechanism uses this property to group several block groups into a flex group and lay out all of the groups' bitmaps and inode tables into one long run in the first group of the flex group.
If the meta_bg feature flag is set, then several block groups are grouped together into a meta group. Note that in the meta_bg case, however, the first and last two block groups within the larger meta group contain only group descriptors for the groups inside the meta group.
flex_bg and meta_bg do not appear to be mutually exclusive features.
In ext2, ext3, and ext4 (when the 64bit feature is not enabled), the block group descriptor was only 32 bytes long and therefore ends at bg_used_dirs_count_lo. On an ext4 filesystem with the 64bit feature enabled, the block group descriptor expands to the full 64 bytes described below.
If gdt_csum is set and metadata_csum is not set, the block group checksum is the crc16 of the FS UUID, the group number, and the group descriptor structure. If metadata_csum is set, then the block group checksum is the lower 16 bits of the checksum of the FS UUID, the group number, and the group descriptor structure. Both block and inode bitmap checksums are calculated against the FS UUID, the group number, and the entire bitmap.
The block group descriptor is laid out in
|0x0||__le32||bg_block_bitmap_lo||Lower 32-bits of location of block bitmap.|
|0x4||__le32||bg_inode_bitmap_lo||Lower 32-bits of location of inode bitmap.|
|0x8||__le32||bg_inode_table_lo||Lower 32-bits of location of inode table.|
|0xC||__le16||bg_free_blocks_count_lo||Lower 16-bits of free block count.|
|0xE||__le16||bg_free_inodes_count_lo||Lower 16-bits of free inode count.|
|0x10||__le16||bg_used_dirs_count_lo||Lower 16-bits of directory count.|
|0x12||__le16||bg_flags||Block group flags. Any of:
|0x14||__le32||bg_exclude_bitmap_lo||Lower 32-bits of location of snapshot exclusion bitmap.|
|0x18||__le16||bg_block_bitmap_csum_lo||Lower 16-bits of the block bitmap checksum.|
|0x1A||__le16||bg_inode_bitmap_csum_lo||Lower 16-bits of the inode bitmap checksum.|
|0x1C||__le16||bg_itable_unused_lo||Lower 16-bits of unused inode count.|
|0x1E||__le16||bg_checksum||Group descriptor checksum; crc16(sb_uuid+group+desc). Probably only calculated if the rocompat bg_checksum feature flag is set.|
|These fields only exist if the 64bit feature is enabled and s_desc_size > 32.|
|0x20||__le32||bg_block_bitmap_hi||Upper 32-bits of location of block bitmap.|
|0x24||__le32||bg_inode_bitmap_hi||Upper 32-bits of location of inodes bitmap.|
|0x28||__le32||bg_inode_table_hi||Upper 32-bits of location of inodes table.|
|0x2C||__le16||bg_free_blocks_count_hi||Upper 16-bits of free block count.|
|0x2E||__le16||bg_free_inodes_count_hi||Upper 16-bits of free inode count.|
|0x30||__le16||bg_used_dirs_count_hi||Upper 16-bits of directory count.|
|0x32||__le16||bg_itable_unused_hi||Upper 16-bits of unused inode count.|
|0x34||__le32||bg_exclude_bitmap_hi||Upper 32-bits of location of snapshot exclusion bitmap.|
|0x38||__le16||bg_block_bitmap_csum_hi||Upper 16-bits of the block bitmap checksum.|
|0x3A||__le16||bg_inode_bitmap_csum_hi||Upper 16-bits of the inode bitmap checksum.|
|0x3C||__u32||bg_reserved||Padding to 64 bytes.|
Total size is 64 bytes.
Block and inode Bitmaps
The data block bitmap tracks the usage of data blocks within the block group.
The inode bitmap records which entries in the inode table are in use.
As with most bitmaps, one bit represents the usage status of one data block or inode table entry. This implies a block group size of 8 * number_of_bytes_in_a_logical_block.
In a regular UNIX filesystem, the inode stores all the metadata pertaining to the file (time stamps, block maps, extended attributes, etc), not the directory entry. To find the information associated with a file, one must traverse the directory files to find the directory entry associated with a file, then load the inode to find the metadata for that file. ext4 appears to cheat (for performance reasons) a little bit by storing a copy of the file type (normally stored in the inode) in the directory entry. (Compare all this to FAT, which stores all the file information directly in the directory entry, but does not support hard links and is in general more seek-happy than ext4 due to its simpler block allocator and extensive use of linked lists.)
The inode table is a linear array of
struct ext4_inode. The table is sized to have enough blocks to store at least
sb.s_inodes_per_group bytes. The number of the block group containing an inode can be calculated as (inode_number - 1) /
sb.s_inodes_per_group, and the offset into the group's table is (inode_number - 1) %
sb.s_inodes_per_group. There is no inode 0.
The inode checksum is calculated against the FS UUID, the inode number, and the inode structure itself.
The inode table entry is laid out in
|0x0||__le16||i_mode||File mode. Any of:
|0x2||__le16||i_uid||Lower 16-bits of Owner UID.|
|0x4||__le32||i_size_lo||Lower 32-bits of size in bytes.|
|0x8||__le32||i_atime||Last access time, in seconds since the epoch.|
|0xC||__le32||i_ctime||Last inode change time, in seconds since the epoch.|
|0x10||__le32||i_mtime||Last data modification time, in seconds since the epoch.|
|0x14||__le32||i_dtime||Deletion Time, in seconds since the epoch.|
|0x18||__le16||i_gid||Lower 16-bits of GID.|
|0x1A||__le16||i_links_count||Hard link count.|
|0x1C||__le32||i_blocks_lo||Lower 32-bits of block count.|
|0x20||__le32||i_flags||Inode flags. Any of:
|0x28||__le32||i_block[EXT4_N_BLOCKS=15]||Block map or extent tree. See the section "The Contents of i_block".|
|0x64||__le32||i_generation||File version (for NFS).|
|0x68||__le32||i_file_acl_lo||Lower 32-bits of extended attribute block. ACLs are of course one of many possible extended attributes; I think the name of this field is a result of the first use of extended attributes being for ACLs.|
|0x6C||__le32||i_size_high||Upper 32-bits of file size.|
|0x70||__le32||i_obso_faddr||(Obsolete) fragment address.|
|0x80||__le16||i_extra_isize||Size of this inode - 128.|
|0x82||__le16||i_checksum_hi||Upper 16-bits of the inode checksum.|
|0x84||__le32||i_ctime_extra||Extra change time bits. This provides sub-second precision.|
|0x88||__le32||i_mtime_extra||Extra modification time bits. This provides sub-second precision.|
|0x8C||__le32||i_atime_extra||Extra access time bits. This provides sub-second precision.|
|0x90||__le32||i_crtime||File creation time, in seconds since the epoch.|
|0x94||__le32||i_crtime_extra||Extra file creation time bits. This provides sub-second precision.|
|0x98||__le32||i_version_hi||Upper 32-bits for version number.|
Note that the size of the structure is 156 bytes, though the standard inode size in ext4 is 256 bytes. It was 128 previously. I think(?) the extra space can be used for extended attributes.
Finding an Inode
Each block group contains
sb->s_inodes_per_group inodes. Because inode 0 is defined not to exist, this formula can be used to find the block group that an inode lives in:
bg = (inode_num - 1) / sb->s_inodes_per_group. The particular inode can be found within the block group's inode table at
index = (inode_num - 1) % sb->s_inodes_per_group. To get the byte address within the inode table, use
offset = index * sb->s_inode_size.
The Contents of inode.i_block
Depending on the type of file an inode describes, the 60 bytes of storage in
inode.i_block can be used in different ways. In general, regular files and directories will use it for file block indexing information, and special files will use it for special purposes.
The target of a symbolic link will be stored in this field if the target string is less than 60 bytes long. Otherwise, either extents or block maps will be used to allocate data blocks to store the link target.
Direct/Indirect Block Addressing
In ext2/3, file block numbers were mapped to logical block numbers by means of an (up to) three level 1-1 block map. To find the logical block that stores a particular file block, the code would navigate through this increasingly complicated structure. Notice that there is neither a magic number nor a checksum to provide any level of confidence that the block isn't full of garbage.
|i.i_block Offset||Where It Points|
|0 to 11||Direct map to file blocks 0 to 11.|
|12||Indirect block: (file blocks 12 to (|
|13||Double-indirect block: (file blocks |
|14||Triple-indirect block: (file blocks (|
Note that with this block mapping scheme, it is necessary to fill out a lot of mapping data even for a large contiguous file! This inefficiency led to the creation of the extent mapping scheme, discussed below.
Notice also that a file using this mapping scheme cannot be placed higher than 2^32 blocks.
In ext4, the file to logical block map has been replaced with an extent tree. Under the old scheme, allocating a contiguous run of 1,000 blocks requires an indirect block to map all 1,000 entries; with extents, the mapping is reduced to a single
struct ext4_extent with
ee_len = 1000. If flex_bg is enabled, it is possible to allocate very large files with a single extent, at a considerable reduction in metadata block use, and some improvement in disk efficiency. The inode must have the extents flag (0x80000) flag set for this feature to be in use.
Extents are arranged as a tree. Each node of the tree begins with a
struct ext4_extent_header. If the node is an interior node (
eh.eh_depth > 0), the header is followed by
eh.eh_entries instances of
struct ext4_extent_idx; each of these index entries points to a block containing more nodes in the extent tree. If the node is a leaf node (
eh.eh_depth == 0), then the header is followed by
eh.eh_entries instances of
struct ext4_extent; these instances point to the file's data blocks. The root node of the extent tree is stored in
inode.i_block, which allows for the first four extents to be recorded without the use of extra metadata blocks.
The extent tree header is recorded in
struct ext4_extent_header, which is 12 bytes long:
|0x0||__le16||eh_magic||Magic number, 0xF30A.|
|0x2||__le16||eh_entries||Number of valid entries following the header.|
|0x4||__le16||eh_max||Maximum number of entries that could follow the header.|
|0x6||__le16||eh_depth||Depth of this extent node in the extent tree. 0 = this extent node points to data blocks; otherwise, this extent node points to other extent nodes.|
|0x8||__le32||eh_generation||Generation of the tree. (Used by Lustre, but not standard ext4).|
Internal nodes of the extent tree, also known as index nodes, are recorded as
struct ext4_extent_idx, and are 12 bytes long:
|0x0||__le32||ei_block||This index node covers file blocks from 'block' onward.|
|0x4||__le32||ei_leaf_lo||Lower 32-bits of the block number of the extent node that is the next level lower in the tree. The tree node pointed to can be either another internal node or a leaf node, described below.|
|0x8||__le16||ei_leaf_hi||Upper 16-bits of the previous field.|
Leaf nodes of the extent tree are recorded as
struct ext4_extent, and are also 12 bytes long:
|0x0||__le32||ee_block||First file block number that this extent covers.|
|0x4||__le16||ee_len||Number of blocks covered by extent.|
|0x6||__le16||ee_start_hi||Upper 16-bits of the block number to which this extent points.|
|0x8||__le32||ee_start_lo||Lower 32-bits of the block number to which this extent points.|
Prior to the introduction of metadata checksums, the extent header + extent entries always left at least 4 bytes of unallocated space at the end of each extent tree data block (because (2^x % 12) >= 4). Therefore, the 32-bit checksum is inserted into this space. The 4 extents in the inode do not need checksumming, since the inode is already checksummed. The checksum is calculated against the FS UUID and the entire extent block leading up to (but not including) the checksum itself.
struct ext4_extent_tail is 4 bytes long:
|0x0||__le32||eb_checksum||Checksum of the extent block.|
In an ext4 filesystem, a directory is more or less a flat file that maps an arbitrary byte string (usually ASCII) to an inode number on the filesystem. There can be many directory entries across the filesystem that reference the same inode number--these are known as hard links, and that is why hard links cannot reference files on other filesystems. As such, directory entries are found by reading the data block(s) associated with a directory file for the particular directory entry that is desired.
Linear (Classic) Directories
By default, directory files contained an almost-linear array of directory entries in that directory. I write "almost" because it's not a linear array in the memory sense because directory entries are not split across filesystem blocks. Therefore, it is more accurate to say that a directory is a series of data blocks and that each block contains a linear array of directory entries. The end of each per-block array is signified by reaching the end of the block; the last entry in the block has a record length that takes it all the way to the end of the block. The end of the entire directory is of course signified by reaching the end of the file. Unused directory entries are signified by inode = 0. By default the filesystem uses
struct ext4_dir_entry_2 for directory entries unless the "filetype" feature flag is not set, in which case it uses
The original directory entry format is
struct ext4_dir_entry, which is at most 263 bytes long, though on disk you'll need to reference
dirent.rec_len to know for sure.
|0x0||__le32||inode||Number of the inode that this directory entry points to.|
|0x4||__le16||rec_len||Length of this directory entry.|
|0x6||__le16||name_len||Length of the file name.|
Since file names cannot be longer than 255 bytes, the new directory entry format shortens the rec_len field and uses the space for a file type flag, probably to avoid having to load every inode during directory tree traversal. This format is
ext4_dir_entry_2, which is at most 263 bytes long, though on disk you'll need to reference
dirent.rec_len to know for sure.
|0x0||__le32||inode||Number of the inode that this directory entry points to.|
|0x4||__le16||rec_len||Length of this directory entry.|
|0x6||__u8||name_len||Length of the file name.|
|0x7||__u8||file_type||File type code, one of:
In order to add checksums to these classic directory blocks, a phony
struct ext4_dir_entry is placed at the end of each leaf block to hold the checksum. The directory entry is 12 bytes long. The inode number and name_len fields are set to zero to fool old software into ignoring an apparently empty directory entry, and the checksum is stored in the place where the name normally goes. The structure is
|0x0||__le32||det_reserved_zero1||Inode number, which must be zero.|
|0x4||__le16||det_rec_len||Length of this directory entry, which must be 12.|
|0x6||__u8||det_reserved_zero2||Length of the file name, which must be zero.|
|0x7||__u8||det_reserved_ft||File type, which must be 0xDE.|
|0x8||__le32||det_checksum||Directory leaf block checksum.|
The leaf directory block checksum is calculated against the FS UUID, the directory's inode number, the directory's inode generation number, and the entire directory entry block up to (but not including) the fake directory entry.
Hash Tree Directories
A linear array of directory entries isn't great for performance, so a new feature was added to ext3 to provide a faster (but peculiar) balanced tree keyed off a hash of the directory entry name. If the EXT4_INDEX_FL (0x1000) flag is set in the inode, this directory uses a hashed btree (htree) to organize and find directory entries. For backwards read-only compatibility with ext2, this tree is actually hidden inside the directory file, masquerading as "empty" directory data blocks! It was stated previously that the end of the linear directory entry table was signified with an entry pointing to inode 0; this is (ab)used to fool the old linear-scan algorithm into thinking that the rest of the directory block is empty so that it moves on.
The root of the tree always lives in the first data block of the directory. By ext2 custom, the '.' and '..' entries must appear at the beginning of this first block, so they are put here as two
struct ext4_dir_entry_2s and not stored in the tree. The rest of the root node contains metadata about the tree and finally a hash->block map to find nodes that are lower in the htree. If
dx_root.info.indirect_levels is non-zero then the htree has two levels; the data block pointed to by the root node's map is an interior node, which is indexed by a minor hash. Interior nodes in this tree contains a zeroed out
struct ext4_dir_entry_2 followed by a minor_hash->block map to find leafe nodes. Leaf nodes contain a linear array of all
struct ext4_dir_entry_2; all of these entries (presumably) hash to the same value. If there is an overflow, the entries simply overflow into the next leaf node, and the least-significant bit of the hash (in the interior node map) that gets us to this next leaf node is set.
To traverse the directory as a htree, the code calculates the hash of the desired file name and uses it to find the corresponding block number. If the tree is flat, the block is a linear array of directory entries that can be searched; otherwise, the minor hash of the file name is computed and used against this second block to find the corresponding third block number. That third block number will be a linear array of directory entries.
To traverse the directory as a linear array (such as the old code does), the code simply reads every data block in the directory. The blocks used for the htree will appear to have no entries (aside from '.' and '..') and so only the leaf nodes will appear to have any interesting content.
The root of the htree is in
struct dx_root, which is the full length of a data block:
|0x0||__le32||dot.inode||inode number of this directory.|
|0x4||__le16||dot.rec_len||Length of this record, 12.|
|0x6||u8||dot.name_len||Length of the name, 1.|
|0x7||u8||dot.file_type||File type of this entry, 0x2 (directory) (if the feature flag is set).|
|0xC||__le32||dotdot.inode||inode number of parent directory.|
|0x10||__le16||dotdot.rec_len||block_size - 12. The record length is long enough to cover all htree data.|
|0x12||u8||dotdot.name_len||Length of the name, 2.|
|0x13||u8||dotdot.file_type||File type of this entry, 0x2 (directory) (if the feature flag is set).|
|0x1C||u8||struct dx_root_info.hash_version||Hash version, one of:
|0x1D||u8||struct dx_root_info.info_length||Length of the tree information, 0x8.|
|0x1E||u8||struct dx_root_info.indirect_levels||Depth of the htree.|
|0x20||__le16||limit||Maximum number of dx_entries that can follow this header.|
|0x22||__le16||count||Actual number of dx_entries that follow this header.|
|0x24||__le32||block||The block number (within the directory file) that goes with hash=0.|
|0x28||struct dx_entry||entries||As many 8-byte |
Interior nodes of an htree are recorded as
struct dx_node, which is also the full length of a data block:
|0x0||__le32||fake.inode||Zero, to make it look like this entry is not in use.|
|0x4||__le16||fake.rec_len||The size of the block, in order to hide all of the dx_node data.|
|0x6||u8||name_len||Zero. There is no name for this "unused" directory entry.|
|0x7||u8||file_type||Zero. There is no file type for this "unused" directory entry.|
|0x8||__le16||limit||Maximum number of dx_entries that can follow this header.|
|0xA||__le16||count||Actual number of dx_entries that follow this header.|
|0xE||__le32||block||The block number (within the directory file) that goes with the lowest hash value of this block. This value is stored in the parent block.|
|0x12||struct dx_entry||entries||As many 8-byte |
The hash maps that exist in both
struct dx_root and
struct dx_node are recorded as
struct dx_entry, which is 8 bytes long:
|0x4||__le32||block||Block number (within the directory file, not filesystem blocks) of the next node in the htree.|
(If you think this is all quite clever and peculiar, so does the author.)
PROPOSED: djwong proposes reducing the dx_node/dx_root structures' limit field by one. This frees up one
struct dx_entry's worth of space to store the checksum.
Extended attributes (xattrs) are typically stored in a separate data block on the disk and referenced from inodes via
inode.i_file_acl*. The first use of extended attributes seems to have been for storing file ACLs and other security data (selinux). With the
user_xattr mount option it is possible for users to store extended attributes so long as all attribute names begin with "user"; this restriction seems to have disappeared as of Linux 3.0.
There are two places where extended attributes can be found. The first place is between the end of each inode entry and the beginning of the next inode entry. For example, if inode.i_extra_isize = 28 and sb.inode_size = 256, then there are 256 - (128 + 28) = 100 bytes available for in-inode extended attribute storage. The second place where extended attributes can be found is in the block pointed to by inode.i_file_acl. As of Linux 3.0, it is not possible for this block to contain a pointer to a second extended attribute block. In theory it is possible for each attribute's value to be stored in a separate data block, though as of Linux 3.0 the code does not permit this.
Extended attributes, when not stored after the inode, have a header
ext4_xattr_ibody_header that is 4 bytes long:
|0x0||__le32||h_magic||Magic number for identification, 0xEA020000. This value is set by the Linux driver, though e2fsprogs doesn't seem to check it(?)|
The beginning of an extended attribute block is in
struct ext4_xattr_header, which is 32 bytes long:
|0x0||__le32||h_magic||Magic number for identification, 0xEA020000.|
|0x8||__le32||h_blocks||Number of disk blocks used.|
|0xC||__le32||h_hash||Hash value of all attributes.|
|0x10||__le32||h_checksum||Checksum of the extended attribute block.|
The checksum is calculated against the FS UUID, the 64-bit block number of the extended attribute block, and the entire block (header + entries).
struct ext4_xattr_header or
struct ext4_xattr_ibody_header is an array of
struct ext4_xattr_entry; each of these entries is at least 16 bytes long.
|0x0||__u8||e_name_len||Length of name.|
|0x1||__u8||e_name_index||Attribute name index.|
|0x2||__le16||e_value_offs||Location of this attribute's value on the disk block where it is stored. Multiple attributes can share the same value.|
|0x4||__le32||e_value_block||The disk block where the value is stored. Zero indicates the value is in the same block as this entry.|
|0x8||__le32||e_value_size||Length of attribute value.|
|0xC||__le32||e_hash||Hash value of name and value.|
|0x10||char||e_name[e_name_len]||Attribute name. Does not include trailing NULL.|
Attribute values can follow the end of the entry table. There appears to be a requirement that they be aligned to 4-byte boundaries. The values are stored starting at the end of the block and grow towards the xattr_header/xattr_entry table. When the two collide, the overflow is put into a separate disk block. If the disk block fills up, the filesystem returns -ENOSPC.
Multiple Mount Protection
Multiple mount protection (MMP) is a feature that protects the filesystem against multiple hosts trying to use the filesystem simultaneously. When a filesystem is opened (for mounting, or fsck, etc.), the MMP code running on the node (call it node A) checks a sequence number. If the sequence number is EXT4_MMP_SEQ_CLEAN, the open continues. If the sequence number is EXT4_MMP_SEQ_FSCK, then fsck is (hopefully) running, and open fails immediately. Otherwise, the open code will wait for twice the specified MMP check interval and check the sequence number again. If the sequence number has changed, then the filesystem is active on another machine and the open fails. If the MMP code passes all of those checks, a new MMP sequence number is generated and written to the MMP block, and the mount proceeds.
While the filesystem is live, the kernel sets up a timer to re-check the MMP block at the specified MMP check interval. To perform the re-check, the MMP sequence number is re-read; if it does not match the in-memory MMP sequence number, then another node (node B) has mounted the filesystem, and node A remounts the filesystem read-only. If the sequence numbers match, the sequence number is incremented both in memory and on disk, and the re-check is complete.
The hostname and device filename are written into the MMP block whenever an open operation succeeds. The MMP code does not use these values; they are provided purely for informational purposes.
The checksum is calculated against the FS UUID and the MMP structure. The MMP structure (
struct mmp_struct) is as follows:
|0x0||__le32||mmp_magic||Magic number for MMP, 0x004D4D50 ("MMP").|
|0x4||__le32||mmp_seq||Sequence number, updated periodically.|
|0x8||__le64||mmp_time||Time that the MMP block was last updated.|
|0x10||char||mmp_nodename||Hostname of the node that opened the filesystem.|
|0x50||char||mmp_bdevname||Block device name of the filesystem.|
|0x70||__le16||mmp_check_interval||The MMP re-check interval, in seconds.|
|0x3FC||__le32||mmp_checksum||Checksum of the MMP block.|
Introduced in ext3, the ext4 filesystem employs a journal to protect the filesystem against corruption in the case of a system crash. A small continuous region of disk (default 128MiB) is reserved inside the filesystem as a place to land "important" data writes on-disk as quickly as possible. Once the important data transaction is fully written to the disk and flushed from the disk write cache, a record of the data being committed is also written to the journal. At some later point in time, the journal code writes the transactions to their final locations on disk (this could involve a lot of seeking or a lot of small read-write-erases) before erasing the commit record. Should the system crash during the second slow write, the journal can be replayed all the way to the latest commit record, guaranteeing the atomicity of whatever gets written through the journal to the disk. The effect of this is to guarantee that the filesystem does not become stuck midway through a metadata update.
For performance reasons, ext4 by default only writes filesystem metadata through the journal. This means that file data blocks are /not/ guaranteed to be in any consistent state after a crash. If this default guarantee level (
data=ordered) is not satisfactory, there is a mount option to control journal behavior. If
data=journal, all data and metadata are written to disk through the journal. This is slower but safest. If
data=writeback, dirty data blocks are not flushed to the disk before the metadata are written to disk through the journal.
The journal inode is typically inode 8. The first 68 bytes of the journal inode are replicated in the ext4 superblock. The journal itself is normal (but hidden) file within the filesystem. The file usually consumes an entire block group, though mke2fs tries to put it in the middle of the disk.
All fields in jbd2 are written to disk in big-endian order. This is the opposite of ext4.
NOTE: Both ext4 and ocfs2 use jbd2.
Generally speaking, the journal has this format:
Superblock [(descriptor_block data_blocks|revocation_block) [more data or revocations] commmit_block] [more transactions...] |<---------------------------------- one transaction ----------------------------------->|
Notice that a transaction begins with either a descriptor and some data, or a block revocation list. A finished transaction always ends with a commit. If there is no commit record (or the checksums don't match), the transaction will be discarded during replay.
Every block in the journal starts with a common 12-byte header
|0x0||__be32||h_magic||jbd2 magic number, 0xC03B3998.|
|0x4||__be32||h_blocktype||Description of what this block contains. One of:
|0x8||__be32||h_sequence||The transaction ID that goes with this block.|
The super block for the journal is much simpler as compared to ext4's. The key data kept within are size of the journal, and where to find the start of the log of transactions.
The journal superblock is recorded as
struct journal_superblock_s, which is 1024 bytes long:
|0x0||journal_header_t (12 bytes)||s_header||Common header identifying this as a superblock.|
|Static information describing the journal.|
|0xC||__be32||s_blocksize||Journal device block size.|
|0x10||__be32||s_maxlen||Total number of blocks in this journal.|
|0x14||__be32||s_first||First block of log information.|
|Dynamic information describing the current state of the log.|
|0x18||__be32||s_sequence||First commit ID expected in log.|
|0x1C||__be32||s_start||Block number of the start of log. If zero, the journal is clean.|
|0x20||__be32||s_errno||Error value, as set by jbd2_journal_abort().|
|The remaining fields are only valid in a version 2 superblock.|
|0x24||__be32||s_feature_compat;||Compatible feature set. Any of:
|0x28||__be32||s_feature_incompat||Incompatible feature set. Any of:
|0x2C||__be32||s_feature_ro_compat||Read-only compatible feature set. There aren't any of these currently.|
|0x30||__u8||s_uuid||128-bit uuid for journal. This is compared against the copy in the ext4 super block at mount time.|
|0x40||__be32||s_nr_users||Number of file systems sharing this journal.|
|0x44||__be32||s_dynsuper||Location of dynamic super block copy. (Not used?)|
|0x48||__be32||s_max_transaction||Limit of journal blocks per transaction. (Not used?)|
|0x4C||__be32||s_max_trans_data||Limit of data blocks per transaction. (Not used?)|
|0x100||__u8||s_users[16*48]||ids of all file systems sharing the log. (Not used?)|
The descriptor block contains an array of journal block tags that describe the final locations of the data blocks that follow in the journal. Descriptor blocks are open-coded instead of being completely described by a data structure, but here is the block structure anyway. Descriptor blocks consume at least 36 bytes, but use a full block:
|0x0||journal_header_t||(open coded)||Common block header.|
|0xC||struct journal_block_tag_s||open coded array||Enough tags either to fill up the block or to describe all the data blocks that follow this descriptor block.|
Journal block tags have the following format, as recorded by
struct journal_block_tag_s. They can be 8, 12, 24, or 38 bytes:
|0x0||__be32||t_blocknr||Lower 32-bits of the location of where the corresponding data block should end up on disk.|
|0x4||__be32||t_flags||Flags that go with the descriptor. Any of:
|This next field is only present if the super block indicates support for 64-bit block numbers.|
|0x8||__be32||t_blocknr_high||Upper 32-bits of the location of where the corresponding data block should end up on disk.|
|This field appears to be open coded. It always comes at the end of the tag, after |
|0x8 or 0xC||char||uuid||A UUID to go with this tag. This field appears to be copied from a field in |
In general, the data blocks being written to disk through the journal are written verbatim into the journal file after the descriptor block. However, if the first four bytes of the block match the jbd2 magic number then those four bytes are replaced with zeroes and the "escaped" flag is set in the descriptor block.
A revocation block is used to record a list of data blocks in this transaction that supersede any older copies of those data blocks that might still be lurking in the journal. This can speed up recovery because those older copies don't have to be written out to disk.
Revocation blocks are described in
struct jbd2_journal_revoke_header_s, are at least 16 bytes in length, but use a full block:
|0x0||journal_header_t||r_header||Common block header.|
|0xC||__be32||r_count||Number of bytes used in this block.|
|0x10||__be32 or __be64||blocks||Blocks to revoke.|
After r_count is a linear array of block numbers that are effectively revoked by this transaction. The size of each block number is 8 bytes if the superblock advertises 64-bit block number support, or 4 bytes otherwise.
The commit block is a sentry that indicates that a transaction has been completely written to the journal. Once this commit block reaches the journal, the data stored with this transaction can be written to their final locations on disk.
The commit block is described by
struct commit_header, which is 32 bytes long (but uses a full block):
|0x0||journal_header_s||(open coded)||Common block header.|
|0xC||unsigned char||h_chksum_type||The type of checksum to use to verify the integrity of the data blocks in the transaction. One of:
|0xD||unsigned char||h_chksum_size||The number of bytes used by the checksum. Most likely 4.|
|0x10||__be32||h_chksum[JBD2_CHECKSUM_BYTES]||32 bytes of space to store checksums.|
|0x30||__be64||h_commit_sec||The time that the transaction was committed, in seconds since the epoch.|
|0x38||__be32||h_commit_nsec||Nanoseconds component of the above timestamp.|
Areas in Need of Work
New patchsets to track with regards to changes in on-disk formats (in no particular order):
- Darrick's metadata checksumming patchset.
- Ted's bigalloc patch
- Amir's ext4 snapshot work.
Also see http://www.nongnu.org/ext2-doc/ for quite a collection of information about ext2/3.