Re: rpc.lockd spinning; much breakage

Don Lewis wrote:
> >     Normally, lockf(1) gets back EAGAIN and polls for the lock to be
> >     released.  I'm not sure which case in the client rpc.lockd(8) it is.
> >     Since Solaris doesn't support O_EXLOCK and lockf(1) it wasn't easy for
> >     me to test, but since the server returns the same result from FreeBSD
> >     and Solaris, I'm guessing it's a client-side mapping problem.
> >     Presumably some instance of nlm_denied should return EAGAIN instead.
> 
> I think this gets an nlm4_denied response, which is handled by the
> following code in lock_answer():

Historically, this was implemented using O_EXCL to ensure only
a single process was allowed to open the file at a time.  It may
be that that's what's being expected over the wire.

I'm pretty sure AIX implemented this by putting an advisory
range lock over the entirety of the file; but AIX has a
couple of other quirks in the NFS processing that make it
less than ideal as a reference implementation.

As of Solaris 5.x, the man pages claim that O_EXCL without
O_CREAT has undefined behaviour.  The SCO manual pages are
pretty clear that it means exclusive use (FWIW; do a Google
search on "O_EXCL" and "exclusive use").

-- Terry