Re: [2.6.18-rc7-mm1] slow boot

[Andrew Morton]

On Sun, 24 Sep 2006 23:36:41 +0100
Alan Cox <alan@xxxxxxxxxxxxxxxxxxx> wrote:

> Ar Sul, 2006-09-24 am 14:53 -0700, ysgrifennodd Andrew Morton:
> > I've *never* seen any reports of any problems being caused by disk
> > writeback caching.  Yes, it's a theoretical problem but for some reason it
> > just doesn't seem to be a problem in practice.  Hence I'm really reluctant
> > to go and slow everyone's machines down so much in this manner.
> 
> It happens in some cases, the usual one is sudden loss of power. In the
> crashed box cases the disk still gets to write data back and in the
> Linux power off sanely cases we explicitly cache flush. Its the sudden
> loss of power case that is nasty.

I don't know about reiserfs, but for ext3 writeback caching delays aren't a
problem per-se.  It's write *reordering* which matters.

And given that the jounal tends to be a single contiguous hunk of disk, the
probability that a journal block at LBA #N gets written before the commit
block at LBA #N+20 is probably fairly low.  There's block remapping of
course, but software journal wrapping might be a more likely cause of write
reordering.

And of course the time window is small - a few milliseconds per five
seconds, and not every five seconds at that.

Hand wavy, I know.  But I wouldn't pay 15% throughput for it..

> We are also helped of course by the fact the cache is never more than
> can be flushed in about 7 seconds because of other-os features.

Well, as I say, the absolute value of any delay doesn't really matter,
except you'd lose an additional seven seconds worth of work.  It's
write reordering which can damage the fs.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/