Here's a patch for an idea to address an issue we have with log
replay.  The problem stems from the fact that not all changes to
inodes result in transactions.  Some changes (such as file size
updates, timestamp updates) would generate too much log traffic
if we logged a transaction for every event.  So we update the inode
and flag it as dirty (ie set i_update_core/i_update_size).  If the
inode gets logged in a later transaction then the update gets
rolled into that transaction and the flag is cleared.  If the inode
gets flushed to disk before another transaction then the on-disk
version of the inode is newer than what is in the log.

On log replay we risk overwriting a newer inode on disk with an
older version of the inode from the log.  We try to prevent this with
the i_flushiter counter in the inode (the on-disk inode will have a
greater flushiter than the inode in the log) but this does not work
for newly created inode cluster buffers.  When log replay encounters
a newly allocated inode cluster buffer in the log it cannot determine
if the on-disk version of the cluster is older, newer or even valid.
Since we cannot determine if the on-disk inode cluster has been
initialised since it was logged we cannot read it to check the
i_flushiter values.  If we write out the log record anyway we risk
overwriting newer inode data, if we don't write out the log record
we risk exposing an uninitialised inode cluster.

The easy solution is to log everything so that log replay doesn't need
to check if the on-disk version is newer - it can just replay the log.
But logging everything would cause too much log traffic so this patch
is a compromise and it logs a transaction before we flush an inode to
disk only if it has changes that have not yet been logged.  With this
scheme we don't clear the i_update_core flag when flushing an inode -
we only do that when logging a transaction.  The flag is used to
determine if a transaction needs to be done.  It needs to be this way
because the log pushing code may need to flush an inode and we cannot
create a transaction at that point.

Anyway it's an idea and needs a little polishing so comments are welcome.

Lachlan