Messages in this thread | | | Date | Thu, 26 Feb 1998 23:33:13 GMT | From | "Stephen C. Tweedie" <> | Subject | Re: soft-update/async write file systems |
| |
Hi,
On Tue, 24 Feb 1998 23:58:24 -0500, "Theodore Y. Ts'o" <tytso@MIT.EDU> said:
> Date: Tue, 24 Feb 1998 17:59:09 -0600 > From: Miguel de Icaza <miguel@nuclecu.unam.mx>
> The transaction routines already detect this scenario and they call a > transaction-merge routine (which is not coded, been busy with the > gnome-beta-1 release :-).
> Which transaction routines are these, out of curiosity? Are you working > on a logging filesystem?
Yep. :)
Mig has a lot of the underlying basic buffer/transaction management, and I've been doing a fairly detailed design of the commit and checkpoint logic. It is looking like a very elegant and effective design right now.
Basically, whenever a new transaction wants a currently locked buffer, it just takes over ownership of that buffer, and the inclusive range of transactions between the first and last use of the buffer get committed to the journal atomically. Since we will be doing such a commit every brief while (1 to 5 seconds, I think), that's no big problem.
The only special cases surround a transaction being committed having a buffer required by a transaction which has come in post-checkpoint, and in that case we just need to make sure that the log writer task can prioritise the writing of these buffers to unlock them as quickly as possible (pretty much the same as the current scenario when you have a buffer being written out to a busy write queue).
--Stephen
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu
| |