Re: [Nmh-workers] tempfile creation

[Mike O'Dell]

i'm well aware of Plan 9's decisions in this area.

having had this discussion with Rob and others in what
was still 1127 at Murray Hill, my conclusion is that the
decision hinged less on the desirability of the semantics
than the complexity of sorting out correct behavior in
the presence of snapshots or logs in the file server.

not everyone in 1127 was thrilled by the omission.

others have shown one can solve those problems quite nicely
so sacrificing the capabilities is unnecessary. eg Network Appliances
filesystem services with snapshots, various log-based filesystems for
Linux, XFS from SGI, and ZFS from Sun.

what is true is that relaxing those semantics makes it easier to
map "unix file semantics" onto more systems - ie, it increases
portability.  the clever hack pulled by NFS recognizes that the most
common use of creat()/unlink()/use is in fact, mk*temp() kind of things
and hence it's worth the trouble to make that work.

however, the various complaints about "never put mailboxes on NFS"
also arise precisely because the semantics have been blurred to allow
assorted significantly different filesystem implementations to be
vended by NFS.  that means precise locking and sequencing behavior
cannot be guaranteed with NFS. it's not the fault of NFS - that's
what comes with the increased portability (note i've avoided saying
"improved" portability).

According to Rusty Sandberg, I was the first person outside of the
authors of VFS first done in SunOS 3.5 to write a filesystem which
plugged into the system and supported an alien filesystem format.
and they had not written a new one, just put a VFS front-end on
the existing UFS code to keep it working.

That was quite a shake-down cruise and more than a few of these
issues got discussed at considerable length.

        cheers,
        -mo


Ralph Corderoy wrote:
> Hi mo,
>
>
> > the reference-count semantics are immensely useful, but not without
> > cost.  it complicates the design of NFS no small amount and is why
> > hard links cannot cross demountable volumes.  however, doing without
> > those semantics is very difficult to imagine.
>
>
> What you say is true, but it may be of interest that Bell Labs dropped
> the link count IIRC in Plan 9, Unix's successor.
>
>     http://plan9.bell-labs.com/magic/man2html/2/stat
>
> Cheers,
>
>
> Ralph.