Ext3 users - Oct 2003 - Writes ordering in Ext3

Hi,

I have a doubt in the ordering of writes in ext3. Any help is
appreciated.

Ext3 enforces certain ordering in writes to ensure data integrity. For
example, the journal writes are issued first and only after their
successful completion the actual writes are issued.

Sometimes we need to maintain ordering across transactions also. This is
needed because if the actual writes of one transaction and the journal
writes of another transaction are scheduled differently then data
integrity may be lost.

For example, assume we have two transactions T1 and T2 modifying
block b1 and b2 of a file respectively.


(1) T1: Modify block b1.
(2) T1: Issue the journal write for block b1.
(3) T1: Issue the actual write to block b1.
(4) T2: Modify block b2.
(5) T2: Issue the journal write for block b2.
(6) T2: Issue the actual write to block b2.

Now, notice that if the file system is not enforcing ordering across
transactions, then writes (3) and (5) may get reordered by the disk
scheduler (which doesn't know about journaling) such that (5) happens
first and then (3) happens.

In such a case, if the system crashes after writing the journal block
for b2 in transaction T2 but before updating the actual block b1 of
transaction T1, then the system will lose its data integrity. Will this
situation ever occur in ext3 ?

thanks,
Vijayan

On Thu, Oct 23, 2003 at 01:24:55PM -0500, Vijayan Prabhakaran
wrote:
> Ext3 enforces certain ordering in writes to ensure data integrity. For
> example, the journal writes are issued first and only after their
> successful completion the actual writes are issued.
> 
> Sometimes we need to maintain ordering across transactions also. This is
> needed because if the actual writes of one transaction and the journal
> writes of another transaction are scheduled differently then data
> integrity may be lost.
Ext3 enforces ordering in writes to ensure filesystem integrity.  If
you care about the order in which data blocks are written to files,
you must use fsync().  This is always been the case with Unix
applications, and ext3 is no different here.  

The one exception is if you enable data block journaling.  In that
case, data blocks are themselves written into the journal, so ordering
is guaranteed.  However, you *still* will need to use fsync() to force
the close of a transaction at times which make sense to the article.
Otherwise, you will have a guarantee of write ordering, so that when
writing the blocks b1, b2, b3, b4, and b5, you will know that b3 is
written, b1 and b2 are written, and so on, but there will no guarantee
how far along the sequence of b1..b5 if at all has been written.  Ext3
normally coaleses many filesystem operations into a single
transaction.

						- Ted