Why Redhat 7.2 chose EXT3  http://linuxtoday.com/news_story.php3?ltsn=2001-08-22-004-20-NW-RH
Where to find a paper on the EXT3 design:  ftp://ftp.uk.linux.org/pub/linux/sct/fs/jfs/
EXT3 FAQ:  http://people.spoiled.org/jha/ext3-faq.html
How EXT2 works http://euclid.nmu.edu/~randy/Research/Papers/EXT2/

Goals of EXT3

• QUICK fscks
• NO/MINIMAL/SOME DATA LOSS on failure
  Predictable data loss
• Totally EXT2 compatable, with in place upgrades and downgrades
• Handle VERY large filesystems.
• Has a /.autofsck file

How a Log works

• There is NO data strucutre except the log
• On write
• Write to the log what you intend
• Put the 'commit record' which must fit into a single block
• Have a 'sweeper process' come along and clean ignorable parts of the log
• On recovery
• Everything with a commit record is valid
• Everything without can be cleaned

• Writing
• There are nornal data items on a normal disk
• create new data blocks holding the new data
• write all metadata changes to the log
• write the commit record
• free the old data blocks
• Then change the normal metadata items
• Recovery
• Scan the log
• For every comit record, do what was committed
• Ignore the other stuff

The Linux Implementation
What's in a journal

• Metadata blocks
• entire blocks of metadata that the file system would like to have inserted into the file systems collection of metadata
• Even if one byte has changed, we write the whole block
• Descriptor blocks
• An array of tuples (block in the journal, block on the filesystem to update)
• Headers
• Lots of them at fixed location
• Describe the current head and tail
• Have a sequence number
• Header not corrupted with the highest sequence number wins.

• On Write
• Write only to the log, both data and metadata (depending on options)
• If the log fills up, stop all opperations from continuing
• On commit
• Flush the journal log to the disk
• Update the headers (commit!!!)
• Flush the updates to their home locations
• Free the journal space
• On recovery
• Scan the headers for the highest sequence number
• Find the head and tail
• Every record, copy those blocks to their home locations

Performance boost... group many operations into a single transaction.  Then structures which updat all the time, like a directory where many files are created, or block bitmaps, or quota info, can be updated fewer times.

EXT3 Journalling modes

• data=journal   Log all changes to data and metadata.
• Needs a larger log
• Is slower
• Best integrity
• data=writeback  Updata data in place, then log metadata
• Can have some inconsistent data
• Can never have inconsistent metadata
• data=ordered  Log all metadata, then write all data
• Default
• All files will be consistent
• Some files might be old

The orphaned file story
How is the code written:  There is a journalling layer, sometimes called JFS or JDB.  The EXT3 just thinks in transactions of data blocks.