Subject: Re: comparing raid-like filesystems
To: Chuck Silvers <email@example.com>
From: Manuel Bouyer <firstname.lastname@example.org>
Date: 02/02/2003 21:38:54
On Sun, Feb 02, 2003 at 12:26:11PM -0800, Chuck Silvers wrote:
> the question here is, what layer should know how to do i/o breakup?
> currently file systems and other callers have to do it, and having
> a per-device MAXPHYS wouldn't change that. since this is really a
> device (or possibly bus) limitation, it seems better to do it below
> the file system.
Well, what may happen is that the upper layer will do some work to
merge requests into a large one, and this will be split again
at the next level. This looks like unecessary work for me.
Passing the info to various users will allow them to optimize their
work for the underlying MAXPHYS. Things like raidframe, the page daemon,
etc ... may work better if they have this information.
We'll probably have to set a upper bound to I/O request size anyway,
why not use the limitation of the underlying device ?
> you're saying you don't want each disk driver to have
> custom code for this, and that would also be good. so this is really
> the job of some middle layer between the file system and the disk driver.
> there are other things that should probably go in such an intermediate
> layer as well, such as partitioning.
Agreed, there is already too much code duplication in the various disk drivers.
Manuel Bouyer <email@example.com>
NetBSD: 24 ans d'experience feront toujours la difference