Re: [PATCH] barrier patch set

[Chris Mason <mason@suse.com>]

On Tue, 2004-03-30 at 16:50, Stephen C. Tweedie wrote:
> Hi,
> 
> On Tue, 2004-03-30 at 20:19, Chris Mason wrote:
> 
> 
> > I think we're mixing a few concepts together.  submit_bh(WRITE_BARRIER,
> > bh) gives us an ordered write in whatever form the lower layers can
> > provide.  It also ensures that if you happen to call wait_on_buffer()
> > for the barrier buffer, the wait won't return until the data is on
> > media.
> 
> Right, but that's just how it works right now --- one doesn't _have_ to
> imply the other.  You could easily imagine an implementation that
> implements barriers and flushing separately, and which does not do
> automatic flushing on completion of WRITE_BARRIER IOs.  SCSI with
> writeback caching enabled might be one example of that.  NBD/DRBD would
> be another likely candidate --- if you've got network latencies in the
> way, then a flushing sync may be far more expensive than a barrier
> propagation.
> 
Yes, that's true, although the barriers don't really imply a flush, it
just implies that if you do use wait_on_* for flushing, it will report
things accurately.

> Unfortunately, a lot of the cases we care about really have to do the
> barrier via flushing, so the benefit of keeping them separate is
> limited.  For LVM/raid0, for example, we've got no way of preserving
> ordering between IOs on different drives, so a flush is necessary there
> unless we start journaling the low-level IOs to preserve order.
> 
Right.

> Yep.  It scares me to think what performance characteristics we'll start
> seeing once that gets used everywhere it's needed, though.  If every raw
> or O_DIRECT write needs a flush after it, databases are going to become
> very sensitive to flush performance.  I guess disabling the flushing and
> using disks which tell the truth about data hitting the platter is the
> sane answer there.

Most database benchmarks are done on scsi, and the blkdev_flush should
be a noop there.  For IDE based database and mail server benchmarks, the
results won't be pretty.  

The reiserfs fsync code tries hard to only flush once, so if a commit is
done then blkdev_flush isn't called.  We might have to do a few other
tricks to queue up multiple synchronous ios and only flush once.

-chris