1.1.1. s_inodes_count

32bit value indicating the total number of inodes, both used and free, in the file system.

1.1.2. s_blocks_count

32bit value indicating the total number of blocks, both used and free, in the file system.

1.1.3. s_r_blocks_count

32bit value indicating the total number of blocks reserved for the usage of the super user. This is most useful if for some reason a user, maliciously or not, fill the file system to capacity; the super user will have this specified amount of free blocks at his disposal so he can edit and save configuration files.

1.1.4. s_free_blocks_count

32bit value indicating the total number of free blocks, including the number of reserved blocks (see s_r_blocks_count). This is a sum of all free blocks of all the block groups.

1.1.5. s_free_inodes_count

32bit value indicating the total number of free inodes. This is a sum of all free inodes of all the block groups.

1.1.6. s_first_data_block

32bit value identifying the first data block, in other word the id of the block containing the superblock structure.

Note that this value is always 0 for file systems with a block size larger than 1KB, and always 1 for file systems with a block size of 1KB. The superblock is always starting at the 1024th byte of the disk, which normally happens to be the first byte of the 3rd sector.

1.1.7. s_log_block_size

The block size is computed using this 32bit value as the number of bits to shift left the value 1024. This value may only be positive.

block size = 1024 << s_log_block_size;
    

1.1.8. s_log_frag_size

The fragment size is computed using this 32bit value as the number of bits to shift left the value 1024. Note that a negative value would shift the bit right rather than left.

if( positive )
  fragmnet size = 1024 << s_log_frag_size;
else
  framgnet size = 1024 >> -s_log_frag_size;
    

1.1.9. s_blocks_per_group

32bit value indicating the total number of blocks per group. This value in combination with s_first_data_block can be used to determine the block groups boundaries.

1.1.10. s_frags_per_group

32bit value indicating the total number of fragments per group. It is also used to determine the size of the block bitmap of each block group.

1.1.11. s_inodes_per_group

32bit value indicating the total number of inodes per group. This is also used to determine the size of the inode bitmap of each block group.

1.1.12. s_mtime

Unix time, as defined by POSIX, of the last time the file system was mounted.

1.1.13. s_wtime

Unix time, as defined by POSIX, of the last write access to the file system.

1.1.14. s_mnt_count

32bit value indicating how many time the file system was mounted since the last time it was fully verified.

1.1.15. s_max_mnt_count

32bit value indicating the maximum number of times that the file system may be mounted before a full check is performed.

1.1.16. s_magic

16bit value identifying the file system as Ext2. The value is currently fixed to 0xEF53.

1.1.17. s_state

16bit value indicating the file system state. When the file system is mounted, this state is set to EXT2_ERROR_FS. When the file system is not yet mounted, this value may be either EXT2_VALID_FS or EXT2_ERROR_FS in the event the file system was not cleanly unmounted.

1.1.18. s_errors

16bit value indicating what the file system driver should do when an error is detected. The following values have been defined:

1.1.19. s_minor_rev_level

16bit value identifying the minor revision level within its revision level.

1.1.20. s_lastcheck

Unix time, as defined by POSIX, of the last file system check.

1.1.21. s_checkinterval

Maximum Unix time interval, as defined by POSIX, allowed between file system checks.

1.1.22. s_creator_os

32bit identifier of the os that created the file system. Defined values are:

1.1.23. s_rev_level

32bit revision level value. There are currently only 2 values defined:

1.1.24. s_def_resuid

16bit value used as the default user id for reserved blocks.

1.1.25. s_def_resgid

16bit value used as the default group id for reserved blocks.

1.1.26. s_first_ino

32bit value used as index to the first inode useable for standard files. In the non-dynamic file system revisions, the first non-reserved inode was fixed to 11. With the introduction the dynamic revision of the file system it is now possible to modify this value.

1.1.27. s_inode_size

16bit value indicating the size of the inode structure. In non-dynamic file system revisions this value is assumed to be 128.

1.1.28. s_block_group_nr

16bit value used to indicate the block group number hosting this superblock structure. This can be used to rebuild the file system from any superblock backup.

1.1.29. s_feature_compat

32bit bitmask of compatible features. The file system implementation is free to support them or not without risk of damaging the meta-data. (more information will be added soon)

1.1.30. s_feature_incompat

32bit bitmask of incompatible features. The file system implementation should refuse to mount the file system if any of the indicated feature is unsupported. (more information will be added soon)

1.1.31. s_feature_ro_compat

32bit bitmask of "read-only" features. The file system implementation should mount as read-only if any of the indicated feature is unsupported. (more information will be added soon)

1.1.32. s_uuid

128bit value used as the volume id. This should, as much as possible, be unique for each file system formatted.

1.1.33. s_volume_name

16 bytes volume name, mostly unusued. A valid volume name would consist of only ISO-Latin-1 characters and be 0 terminated.

1.1.34. s_last_mounted

64 bytes directory path where the file system was last mounted. While not normally used, it could serve for auto-finding the mountpoint when not indicated on the command line. Again the path should be zero terminated for compatibility reasons. Valid path is constructed from ISO-Latin-1 characters.

1.1.35. s_algo_bitmap

32bit value used by compression algorithms to determine the methods used. (I do not have any more detail about this field, if you do please do send me all the information you have, thanks).

1.2.1. bg_block_bitmap

32bit block id of the first block of the "block bitmap" for the group represented.

1.2.2. bg_inode_bitmap

32bit block id of the first block of the "inode bitmap" for the group represented.

1.2.3. bg_inode_table

32bit block id of the first block of the "inode table" for the group represented.

1.2.4. bg_free_blocks_count

16bit value indicating the total number of free blocks for the represented group.

1.2.5. bg_free_inodes_count

16bit value indicating the total number of free inodes for the represented group.

1.2.6. bg_used_dirs_count

16bit value indicating the number of inodes allocated to directories for the represented group.

1.2.7. bg_pad

16bit value used for padding the structure on a 32bit boundary.

1.2.8. bg_reserved

3 successive 32bit values reserved for future implementations.

1.5.1. i_mode

16bit value used to indicate the format of the described file and the access rights. Here are the possible values, which can be combined in various ways:

1.5.2. i_uid

16bit user id associated with the file.

1.5.3. i_size

32bit value indicating the size of the file in bytes.

1.5.4. i_atime

32bit value representing the number of seconds since january 1st 1970 of the last time this file was accessed.

1.5.5. i_ctime

32bit value representing the number of seconds since january 1st 1970 when the file was created.

1.5.6. i_mtime

32bit value representing the number of seconds since january 1st 1970 of the last time this file was modified.

1.5.7. i_dtime

32bit value representing the number of seconds since january 1st 1970 when the file was deleted. It is important that unless the file is deleted that this value is always 0.

1.5.8. i_gid

16bit value of the group having access to this file.

1.5.9. i_links_count

16bit value indicating how many times this particular inode is linked (referred to).

1.5.10. i_blocks

32bit value indicating the amount of blocks reserved for the associated file data. This includes both currently in used and currently reserved blocks in case the file grows in size.

A point worth of note is that this value indicate the number of 512 bytes block and not the number of blocks of the size indicated in the superblock. So if a file uses only 1 file system block and is 1024 bytes big, its .i_blocks value will be 2.

1.5.11. i_flags

32bit value indicating how the ext2 implementation should behave when accessing the data for this inode. (See the Behaviour flags section.)

1.5.12. i_osd1

32bit OS dependant value.

1.5.13. i_block

Array used to locate the blocks the particular file is stored on. Each entry is a 32bit block number. The first 12 entries in this array are block numbers, which can be used to fetch the first 12 blocks associated with the file.

The 13th entry is an indirect block number. Which means that at the specified data block, you will find an array of direct block numbers.

The 14th entry is an bi-indirect block number. Which means that at the specified data block, you will find an array of indirect block number, which in turn contains an array of block numbers that can be accessed directly.

The 15th entry is an tri-indirect block number. It is a block number which contains an array of bi-indirect block number, etc.

Each indirect/bi-indirect/tri-indirect block array contains as many entries of 32bit block numbers as possible (to fill one entire block).

1.5.14. i_generation

32bit value used to indicate the file version (used by NFS).

1.5.15. i_file_acl

32bit value indicating the block number containing the extended attributes. In previous revisions this value was always 0.

A general description of ACL for Digital UNIX can be found at this url for the moment: http://www.tru64unix.compaq.com/docs/base_doc/DOCUMENTATION/HTML/AA-Q0R2D-TET1_html/sec.c27.html

1.5.16. i_dir_acl

32bit value used to indicate the "high size" of the file. In previous revisions this value was always 0.

1.5.17. i_faddr

32bit value indicating the location of the last file fragment.

1.5.18. i_osd2

96bit OS dependant structure.

offset  size    description
------- ------- -----------
      0       1 h_i_frag
      1       1 h_i_fsize
      2       2 h_i_mode_high
      4       2 h_i_uid_high
      6       2 h_i_gid_high
      8       4 h_i_author
     

offset  size    description
------- ------- -----------
      0       1 l_i_frag
      1       1 l_i_fsize
      2       2 reserved
      4       2 l_i_uid_high
      6       2 l_i_gid_high
      8       4 reserved
     

offset  size    description
------- ------- -----------
      0       1 m_i_frag
      1       1 m_i_fsize
      2      10 reserved
     

2.1.1. inode

32bit inode number of the file entry. A value of 0 indicate that the entry is not used.

2.1.2. rec_len

16bit unsigned displacement to the next directory entry from the start of the current directory entry.

2.1.3. name_len

8bit unsigned value indicating how many characters are contained in the name.

2.1.4. file_type

8bit unsigned value used to indicate file type. As noted, this value may be 0 in earlier implementations. Currently defined values are:

2.1.5. name

Name of the entry. The allowed character set is the ISO-Latin-1.

2.3.1. Index Structure

The root of the index tree is in the 0th block of the file. Space is reserved for a second level of the index tree in blocks 1 though 511 (for 4K filesystem blocks). Directory leaf blocks are appended starting at block 512, thus the tail of the directory file looks like a normal Ext2 directory and can be processed directly by ext2_readdir. For directories with less than about 90K files there is a hole running from block 1 to block 511, so an empty directory has just two blocks in it, though its size appears to be about 2 Meg in a directory listing.

So a directory file looks like:

0: Root index block
1: Index block/0
2: Index block/0
...
511: Index block/0
512: Dirent block
513: Dirent block
...
    

Each index block consists of 512 index entries of the form:

	hash, block
    

where hash is a 32 bit hash with a collision flag in its least significant bit, and block is the logical block number of an index of leaf block, depending on the tree level.

The hash value of the 0th index entry isn't needed because it can always be obtained from the level about, so it is used to record the count of index entries in an index block. This gives a nice round branching factor of 512, the evenness being a nicety that mainly satisfies my need to seek regularity, rather than winning any real performance. (On the other hand, the largeness of the branching factor matters a great deal.)

The root index block has the same format as the other index blocks, with its first 8 bytes reserved for a small header:

1 byte header length (default: 8)
1 byte index type (default: 0)
1 byte hash version (default:0)
1 byte tree depth (default: 1)
    

The treatment of the header differs slightly in the attached patch. In particular, only a single level of the index tree (the root) is implemented here. This turns out to be sufficient to handle more than 90,000 entries, so it is enough for today. When a second level is added to the tree, capacity will incease to somewhere around 50 million entries, and there is nothing preventing the use of n levels, should there ever be a reason. It's doubtfull that a third level will ever be required, but if it is, the design provides for it.

2.3.2. Lookup Algorithm

Lookup is straightforword:

- Compute a hash of the name
- Read the index root
- Use binary search (linear in the current code) to find the
  first index or leaf block that could contain the target hash
  (in tree order)
- Repeat the above until the lowest tree level is reached
- Read the leaf directory entry block and do a normal Ext2
  directory block search in it.
- If the name is found, return its directory entry and buffer
- Otherwise, if the collision bit of the next directory entry is
  set, continue searching in the successor block
    

Normally, two logical blocks of the file will need to be accessed, and one or two metadata index blocks. The effect of the metadata index blocks can largely be ignored in terms of disk access time since these blocks are unlikely to be evicted from cache. There is some small CPU cost that can be addressed by moving the whole directory into the page cache.

2.3.3. Insert Algorithm

Insertion of new entries into the directory is considerably more complex than lookup, due to the need to split leaf blocks when they become full, and to satisfy the conditions that allow hash key collisions to be handled reliably and efficiently. I'll just summarize here:

- Probe the index as for lookup
- If the target leaf block is full, split it and note the block
  that will receive the new entry
- Insert the new entry in the leaf block using the normal Ext2
  directory entry insertion code.
    

The details of splitting and hash collision handling are somewhat messy, but I will be happy to dwell on them at length if anyone is interested.

2.3.4. Splitting

In brief, when a leaf node fills up and we want to put a new entry into it the leaf has to be split, and its share of the hash space has to be partitioned. The most straightforward way to do this is to sort the entrys by hash value and split somewhere in the middle of the sorted list. This operation is log(number_of_entries_in_leaf) and is not a great cost so long as an efficient sorter is used. I used Combsort for this, although Quicksort would have been just as good in this case since average case performance is more important than worst case.

An alternative approach would be just to guess a median value for the hash key, and the partition could be done in linear time, but the resulting poorer partitioning of hash key space outweighs the small advantage of the linear partition algorithm. In any event, the number of entries needing sorting is bounded by the number that fit in a leaf.

2.3.5. Key Collisions

Some complexity is introduced by the need to handle sequences of hash key collisions. It is desireable to avoid splitting such sequences between blocks, so the split point of a block is adjusted with this in mind. But the possibility still remains that if the block fills up with identically-hashed entries, the sequence may still have to be split. This situation is flagged by placing a 1 in the low bit of the index entry that points at the sucessor block, which is naturally interpreted by the index probe as an intermediate value without any special coding. Thus, handling the collision problem imposes no real processing overhead, just come extra code and a slight reduction in the hash key space. The hash key space remains sufficient for any conceivable number of directory entries, up into the billions.

2.3.6. Hash Function

The exact properties of the hash function critically affect the performance of this indexing strategy, as I learned by trying a number of poor hash functions, at times intentionally. A poor hash function will result in many collisions or poor partitioning of the hash space. To illustrate why the latter is a problem, consider what happens when a block is split such that it covers just a few distinct hash values. The probability of later index entries hashing into the same, small hash space is very small. In practice, once a block is split, if its hash space is too small it tends to stay half full forever, an effect I observed in practice.

After some experimentation I came up with a hash function that gives reasonably good dispersal of hash keys across the entire 31 bit key space. This improved the average fullness of leaf blocks considerably, getting much closer to the theoretical average of 3/4 full.

But the current hash function is just a place holder, waiting for an better version based on some solid theory. I currently favor the idea of using crc32 as the default hash function, but I welcome suggestions.

Inevitably, no matter how good a hash function I come up with, somebody will come up with a better one later. For this reason the design allows for additional hash functiones to be added, with backward compatibility. This is accomplished simply, by including a hash function number in the index root. If a new, improved hash function is added, all the previous versions remain available, and previously created indexes remain readable.

Of course, the best strategy is to have a good hash function right from the beginning. The initial, quick hack has produced results that certainly have not been disappointing.

2.3.7. Performance

OK, if you have read this far then this is no doubt the part you've been waiting for. In short, the performance improvement over normal Ext2 has been stunning. With very small directories performance is similar to standard Ext2, but as directory size increases standard Ext2 quickly blows up quadratically, while htree-enhanced Ext2 continues to scale linearly.

Uli Luckas ran benchmarks for file creation in various sizes of directories ranging from 10,000 to 90,000 files. The results are pleasing: total file creation time stays very close to linear, versus quadratic increase with normal Ext2.

Time to create:

		Indexed		Normal
		=======		======
10000 Files:	0m1.350s	0m23.670s
20000 Files:	0m2.720s	1m20.470s
30000 Files:	0m4.330s	3m9.320s
40000 Files:	0m5.890s	5m48.750s
50000 Files:	0m7.040s	9m31.270s
60000 Files:	0m8.610s	13m52.250s
70000 Files:	0m9.980s	19m24.070s
80000 Files:	0m12.060s	25m36.730s
90000 Files:	0m13.400s	33m18.550s
    

A graph is posted at: http://www.innominate.org/~phillips/htree/performance.png

All of these tests are CPU-bound, which may come as a surprise. The directories fit easily in cache, and the limiting factor in the case of standard Ext2 is the looking up of directory blocks in buffer cache, and the low level scan of directory entries. In the case of htree indexing there are a number of costs to be considered, all of them pretty well bounded. Notwithstanding, there are a few obvious optimizations to be done:

- Use binary search instead of linear search in the interior index
  nodes.

- If there is only one leaf block in a directory, bypass the index
  probe, go straight to the block.

- Map the directory into the page cache instead of the buffer cache.
    

Each of these optimizations will produce a noticeable improvement in performance, but naturally it will never be anything like the big jump going from N**2 to Log512(N), ~= N. In time the optimizations will be applied and we can expect to see another doubling or so in performance.

There will be a very slight performance hit when the directory gets big enough to need a second level. Because of caching this will be very small. Traversing the directories metadata index blocks will be a bigger cost, and once again, this cost can be reduced by moving the directory blocks into the page cache.

Typically, we will traverse 3 blocks to read or write a directory entry, and that number increases to 4-5 with really huge directories. But this is really nothing compared to normal Ext2, which traverses several hundred blocks in the same situation.

4.1.1. SUID, SGID and -rwxrwxrwx

There isn't much to say about those, they are located with the SGID and SUID bits in ext2_inode.i_mode.

4.1.2. File Size

The size of a file can be determined by looking at the ext2_inode.i_size field.

4.1.3. Owner and Group

Under most implementations, the owner and group are 16bit values, but on some recent Linux and Hurd implementations the owner and group id are 32bit. When 16bit values are used, only the "low" part should be used as valid, while when using 32bit value, both the "low" and "high" part should be used, the high part being shifted left 16 places then added to the low part.

The low part of owner and group are located in ext2_inode.i_uid and ext2_inode.i_gid respectively.

The high part of owner and group are located in ext2_inode.osd2.hurd.h_i_uid_high and ext2_inode.osd2.hurd.h_i_gid_high, respectively, for Hurd and located in ext2_inode.osd2.linux.l_i_uid_high and ext2_inode.osd2.linux.l_i_gid_high, respectively, for Linux.

4.2.1. Attribute Block Header

offset  size    description
------- ------- -----------
      0       4 h_magic
      4       4 h_refcount
      8       4 h_blocks
     12       4 h_hash
     16      16 reserved
    

4.2.2. Attribute Entry Header

offset  size    description
------- ------- -----------
      0       1 e_name_len
      1       1 e_name_index
      2       2 e_value_offs
      4       4 e_value_block
      8       4 e_value_size
     12       4 e_hash
     16     ... e_name
    

The total size of an attribute entry is always rounded to the next 4-bytes boundary.

4.3.1. EXT2_SECRM_FL - Secure Deletion

Enabling this bit will cause random data to be written over the flie's content several time before the blocks are unlinked. Note that this is highly implementation dependant and as such, it should not be assumed to be 100% secure. Make sure to study the implementation notes before relying on this option.

4.3.2. EXT2_UNRM_FL - Record for Undelete

When supported by the implementation, setting this bit will cause the deleted data to be moved to a temporary location, where the user can restore the original file without any risk of data lost. This is most useful when using ext2 on a desktop or workstation.

4.3.3. EXT2_COMPR_FL - Compressed File

The file's content is compressed. There is no note about the particular algorithm used other than maybe the s_algo_bitmap field of the superblock structure.

4.3.4. EXT2_SYNC_FL - Synchronous Updates

The file's content in memory will be constantly synchronized with the content on disk. This is mostly used for very sensitive boot files or encryption keys that you do not want to lose in case of a crash.

4.3.5. EXT2_IMMUTABLE_FL - Immutable File

The blocks associated with the file will not be exchanged. If for any reason a file system defragmentation is launched, such files will not be moved. Mostly used for stage2 and stage1.5 boot loaders.

4.3.6. EXT2_APPEND_FL - Append Only

Writing can only be used to append content at the end of the file and not modify the current content. Example of such use could be mailboxes, where anybody could send a message to a user but not modify any already present.

4.3.7. EXT2_NODUMP_FL - Do No Dump/Delete

Setting this bit will protect the file from deletion. As long as this bit is set, even if the i_links_count is 0, the file will not be removed.

4.3.8. EXT2_NOATIME_FL - Do Not Update .i_atime

The i_atime field of the inode structure will not be modified when the file is accessed if this bit is set. The only good use I can think of that are related to security.

4.3.9. EXT2_DIRTY_FL - Dirty

I do not have information at this moment about the use of this bit.

4.3.10. EXT2_COMPRBLK_FL - Compressed Blocks

This flag is set if one or more blocks are compressed. You can have more information about compression on ext2 at http://www.netspace.net.au/~reiter/e2compr/ Note that the project has not been updated since 1999.

4.3.11. EXT2_NOCOMPR_FL - Access Raw Compressed Data

When this flag is set, the file system implementation will not uncompress the data before fowarding it to the application but will rather give it as is.

4.3.12. EXT2_ECOMPR_FL - Compression Error

This flag is set if an error was detected when trying to uncompress the file.

4.3.13. EXT2_BTREE_FL - B-Tree Format Directory

4.3.14. EXT2_INDEX_FL - Hash Indexed Directory

When this bit is set, the format of the directory file is hash indexed. This is covered in details in Section 2.3.

4.3.15. EXT2_IMAGIC_FL -

4.3.16. EXT2_JOURNAL_DATA_FL - Journal File Data

4.3.17. EXT2_RESERVED_FL - Reserved