Re: [PATCH] remove bashisms from xfstests

[Dave Chinner <david@fromorbit.com>]

On Wed, Jan 06, 2010 at 11:24:10AM -0800, Christian Kujau wrote:
> On Wed, 6 Jan 2010 at 11:48, Christoph Hellwig wrote:
> > > 2) convert "let..." into something (hopefully) more portable
> > 
> > Dave converted these away from expr due to performance reasons.  I'd
> > like too see a prove that performance hasn't regressed due to this
> > change.
> 
> That's why I hesitated to just change all scripts to /bin/bash, but didn't 
> dare to say so, because of the inevitable "but today's computers are fast 
> enough" reactions :-)

FYI - the problem I came across was to do with running xfsqa under UML,
where forking a new process is an expensive operation (I measured it
at about 300ms) and so using bash built-in expressions for
incrementing variables is a major win in terms of test run time.
IIRC, one test loops 100,000 times and the runtime went from half an
hour to 10s just by using "let n=n+1" instead of expr...

> 
> > > 3) replace 'a == b' with 'a = b' in bourne shell scripts
> > 
> > This looks fine.  If you want feel free to submit these as a separate
> > first patch so that we have the large pile sorted out.
> 
> OK, will do.
> 
> > I fear this might cause some overhead in the shell.  What about the
> > following instead:
> > 
> > for i in `seq 0 39`; do
> >     ln -s $o symlink_$i
> >     o=symlink_$i
> > done
> 
> I didn't go for seq(1), as I feared this wouldn't be available on every 
> machine.

seq(1) is already used in several tests, so it is ok to use in
more. :)

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs