Re: [PATCH 3/4] nfsd: Add a super simple flex file server

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

On Wed, May 25, 2016 at 05:57:29PM -0400, Jeff Layton wrote:
> On Wed, 2016-05-25 at 13:42 -0400, J. Bruce Fields wrote:
> > On Wed, May 25, 2016 at 08:30:44AM -0400, Jeff Layton wrote:
> > > 
> > > Maybe it's time to start thinking about how to support multiple layout
> > > types per export?
> > It looks like nobody would want this flex file code in production.  The
> > only users will be testers and developers.
> > 
> > And the scsi layout is really just a replacement for the block layout,
> > nobody should be supporting both of those at once either.
> > 
> 
> Well...unless you have a mix of clients that just support block and
> some that support scsi. Is that plausible?

Maybe so.  I don't think it would be useful to support, though.

(I'd rather people skipped straight to scsi.  And block never got much
use, so I don't think that should be hard.  But if you're really stuck
with some block clients, I suspect you may as well use block for all of
them--the scsi clients could probably do block layout too, and I think
you only get all the advantages of the scsi layout if all your clients
are using it.)

> > Would it be too much of a burden just to make flexfiles and developers
> > build their own kernels?
> >
> > > It doesn't look like it would be that hard. I think
> > > we could convert ex_layout_type into a bitmap that shows which types
> > > are supported.
> > ex_layout_type is only used internally, the only external interface is
> > an export flag, so we'd need some new interface.
> > 
> 
> I was thinking that with the "pnfs" export option you'd just enable any
> layouts that the fs supports. So here, you could theoretically allow
> nfsd to offer up block, scsi and flexfiles layouts given the right fs,
> and leave the decision of the layout type to actually use up to the
> client.

OK.

In this particular case that seems less interesting than just the
ability to turn flexfiles on and off from user space, so testers can
enable it without having to build a new kernel.

Hopefully we can figure out how to make this useful in production some
day, and then maybe the multiple-layout support becomes more
interesting.

--b.