Re: 2.4.21-pre2 stalls out when running unixbench

[John Bradford]

> > This is because of differences in how sync() is handled between 2.4.20's
> > ext3 and 2.4.21-pre2's.
> > 
> > 2.4.21-pre2:
> > 
> >   sync() will start the commit, and will wait on it.  So you know that
> >   when it returns, everything which was dirty is now tight on disk.
> > 
> > So yes, running a looping sync while someone else is writing stuff
> > will take much longer in 2.4.21-pre2, because that kernel actually
> > waits on the writeout.
> 
> Actually, I'm wondering if we should back that particular bit out.  For
> a user with a hundred mounted filesystems, syncing each one in order,
> sequentially, is going to suck (and we don't currently have a simple way
> in 2.4 to detect which syncs are on separate spindles and so can be
> parallelised.)

What!?  I'm suprised that no userspace applications were broken by
sync returning before the data is safely on oxide, even though it
doesn't violate the POSIX spec.

What about userspace media-changers, (if such a thing exists)?
Presumably they would assume that they can eject the media after a sync.

I think sync should definitely wait until it's completed before it
returns.

John.