Re: [nfsv4] OPEN_DOWNGRADE and posix byte range locking issue

[<[e-mail]>]

It isn't the server's problem.  There is no way the server has access to
the Vfs opens as visible objects or as having the assignment of locks to
such fine-grained opens.

I'm kind of thinking that this shows we (Bruce, me, and rest of the
working group) made a mistake in that sort of a design in which we do
not allow multiple distinguished open objects for a given fh-owner pair.
Anyway the problems that it caused are pretty minor and we don't know
what problems would have been generated with an alternate design.  I
think this is something to look at in NFSv5 or the next NFSv4.1-style
minor version, if any. 

-----Original Message-----
From: J. Bruce Fields [mailto:[e-mail]] 
Sent: Thursday, August 19, 2010 12:45 PM
To: Trond Myklebust
Cc: Noveck, David; [e-mail]; [e-mail]
Subject: Re: [nfsv4] OPEN_DOWNGRADE and posix byte range locking issue

> > -----Original Message-----
> > From: [e-mail] [mailto:[e-mail]] On
Behalf
> > Of Trond Myklebust
> > Sent: Wednesday, July 07, 2010 7:30 PM
> > To: [e-mail]
> > Subject: [nfsv4] OPEN_DOWNGRADE and posix byte range locking issue
> > 
> > Neither RFC3530, nor RFC5661 appear to list NFS4ERR_LOCKS_HELD as a
> > valid response when the client calls OPEN_DOWNGRADE.
> > 
> > The question is: what should the server then do if the NFS client
holds
> > a WRITE_LT lock, but then asks for an OPEN_DOWNGRADE to
> > OPEN4_SHARE_ACCESS_READ. I understand that this is sanctioned in
Windows
> > server environments, but it should definitely be forbidden in a
POSIX
> > environment, and NFS4ERR_LOCKS_HELD would appear to fit the bill...

A bizarre variation: the linux server associates vfs opens with
stateid's.  Locks are performed on vfs opens, and the vfs will complain
if you attempt to close a file that still has locks associated with it.

The sequence

	open RW
	lock R
	open R
	open downgrade to R

would therefore be implemented at the vfs level as:

	open RW -> f
	lock R on f
	open R -> g
	close f

Oops.  We're stuck with ditching the lock (or erroring out) even though
it's still compatible with the new config option.

Well, I suppose this is my problem: either I should get a new vfs open
for the use of the lock, or represent the original RW open by two vfs
open's.

It's not something a unix-like client could do, I think, but I don't
think it's safe for me to assume I can reject it?

--b.