> NTFS uses logging and B-trees almost everywhere. The log file
> can run out of space, in which case all filesystem activity
> must stop and incomplete transactions must be rolled back!
> Then the log file can be cleared and restarted. Nobody on the
> NTFS list knows how to make the Linux VFS layer undo and reissue
> filesystem operations. Logging filesystem experts are needed.
> (NTFS uses logging, but it is not a just a log) It seems that
> stipped disks will be hard to support.
>
> I doubt we'd want to get the Linux VFS layer involved. It sounds like
> it could be done entirely within the NTFS layer. If you run out of
> space, presumably what you'll need to do is to store transactions in
> memory during the rollback, clearing, and reissue process. And if the
> system crashes during this operation, it's hasta luego to all of the
> data hadn't been fully committed to the log.....
It might be quite a bit more efficient to avoid filling up the log in the
first place rather than rolling back. When the log hits a certain size,
stop accepting transaction requests, close all pending transaction, and
then shrink the log. Slightly shorter history, but if it's possible, it
should be faster than actually _undoing_ transactions.
-- "Love the dolphins," she advised him. "Write by W.A.S.T.E.."