Re: Proposal: Use hi-res clock for file timestamps

["J. Bruce Fields" <bfields@fieldses.org>]

On Wed, Aug 18, 2010 at 09:25:08PM +0200, Andi Kleen wrote:
> On Wed, Aug 18, 2010 at 02:54:56PM -0400, J. Bruce Fields wrote:
> > On Wed, Aug 18, 2010 at 07:50:40PM +0200, Andi Kleen wrote:
> > > >  - nfsd updates it whenever it reads an mtime out of an inode that matches
> > > >    current_fs_time to the granularity of 1/HZ.
> > > 
> > > That means you have a very very hot cache line on a larger system
> > > if there are a lot of mtime changes. Probably a bad idea.
> > 
> > Only if those mtime changes are also followed immediately by nfsd reads
> > of the mtime.
> 
> If multiple writers are changing the same location in quick succession 
> you have a hot cache line that gets bounced around.  It doesn't need reads,
> although reads make it even worse.

OK, at this point one of us is confused, and I'm not sure which.

Is the "same location" that you're referring to the current_nfsd_time?

Neil's suggestion is to only modify current_nfsd_time on nfsd getattr,
*not* on the write operation that modifies the file data.

Or are you talking about something else?

> There's a lot of effort currently to make the VFS more parallel
> and less synchronized and it would be bad again to regress here again.

Understood.

--b.