On Wed, Oct 17, 2012 at 02:39:55AM +0000, Thor Lancelot Simon wrote:
> On Mon, Oct 15, 2012 at 04:24:54PM +0200, Manuel Bouyer wrote:
> >
> > I wonder if we could reuse bits from the read-ahead code for write-behind ?
>
> I'm not sure the read-ahead code is working properly. When I read through
> the filesystem with cat or dd bs=1024k, on a device (wd on ahcisata)
> that can do 1M transfers, I see that most transfers are considerably
> smaller.
For me it grows up to 512k, which is the limit in the read-ahead code.
>
> What I'm more worried about is that it may _never_ be a good idea, in the
> general case, to put 1M transfers on the bus, particularly single writes,
> since they may delay things like log transactions in the disk's write
> cache. We might need a much safer I/O scheduler before we can safely do
> this. Or we might need to let the user tune it on a per-filesystem basis.
Note that with the current code, even if the write-behind code in
ufs_readwrite limits the transfer, the pagedaemon can schedule
writes of 1M when memory gets short. I've seen thoses when playing with
the write-behind code (basically when I disabled write-behind, so
writes got triggered by the pagedaemon).
--
Manuel Bouyer <bou...@antioche.eu.org>
NetBSD: 26 ans d'experience feront toujours la difference
--
