Re: Journaling patches

[Alistair Crooks <agc%pkgsrc.org@localhost>]

On Fri, Apr 25, 2008 at 12:37:09AM +0100, Andrew Doran wrote:
> > A patch against today's -current is at:
> > 
> >    http://www.NetBSD.org/~simonb/wapbl-current-20080422.diff
> 
> Some time back you, Greg and I briefly discussed moving the journalling
> stuff towards a producer-consumer model where there is a kthread pushing
> out journal updates to the disk. That might not be the best model. With the
> patch as-is, updating threads take care of syncing the journal directly.
> 
> In the tests that I've done it works great as-is, but in a multiuser
> environment it doesn't perform very well because the flushing serializes
> everything. I'm concerned that if we commit this before 5.0, we could be
> setting negative expectations of the feature. How do you feel about adding
> a printf to ffs_mount() that indicates it's an experimental feature, until
> we have fixed the concurrency issue?

I think a warning would be in order, yes.

But I'd like to generalise this discussion, especially in light of the
SoC project to implement ext3 journalling.  I think we're going to run
into the same dilemma here - the serialising through flushing was
quite obvious on Greg's tests, and quite expensive (I can't remember
the numbers right now).  So is there anything we can do to keep things
from being serialised at one point in the kernel, and how extensive will
those diffs be?

Regards,
Alistair