Re: [PATCH, RFC] add a vfs_fsync helper

[Al Viro <viro@ZenIV.linux.org.uk>]

On Mon, Dec 22, 2008 at 01:56:22PM +0100, Christoph Hellwig wrote:
> On Mon, Dec 22, 2008 at 01:35:34PM +0100, Nick Piggin wrote:
> > Seems like an improvement.. I've wanted to know, though, what do
> > we need i_mutex for? Is it just convention, or is there some good
> > reason to have it in generic code?
> 
> At least XFS doesn't need it.  Same for the filemap_fdatawrite /
> filemap_fdatawait which at least for filesystems that want to provide
> integrity guarantees is in the wrong place.
> 
> This patch is a first one out my work to refactor fsync, and I'm trying
> to feed it to mainline in pieces.  The next one will be to make sure
> nfsd always has a struct file available when calling fsync, but I need
> to do some extensive benchmarking.  After that we can change the
> fsync prototype to drop the dentry paramters, and move the
> filemap_fdatawrite / filemap_fdatawait aswell as the i_mutex locking
> into the actual methods.

I don't think that anything generic needs it anymore (if it ever did).
That was before my time (2.1.56, so it's probably Bill Hawes) and I
suspect that the reason had been along the lines of "we really don't
want to think about fsync() vs. truncate() races, let's just serialize
it".