On Wed, Feb 03, 2010 at 06:27:53AM -0500, Christoph Hellwig wrote:
> Still not entirely happy with this one. The first one is that I think
> the barriers in fsync are still too heavy for the normal sync use
> case. I'd be more happy with exporting the body of xfs_fsync without
> the cache flushes (and a ebtter name than xfs_fsync) and use that
> for write_inode.
> That leaves open the NFSD case thought. I'd prefer to have that fixed
> if possibly. Ben, any chance you could send your patch to use fsync
> to the nfs list ASAP? I think we'd be even better off to just force
> -o wsync and disable ->write_inode entirely for NFS, any chance you
> could test such a patch on your setup?
Thinking about it, we usually do cause a log buffer write from ->fsync
which means we submit the barrier anyway. That might be the reason
why you're not seeing the performance hit in your testing. With that
I'm okay with the patch as-is for now, we can micro-optimize it later.
Reviewed-by: Christoph Hellwig <hch@xxxxxx>