(was slashes, now NFS devices)

Dan Bernstein brnstnd at kramden.acf.nyu.edu
Sat Mar 2 13:04:09 AEST 1991


Some personal comments: My files on this system happen to be mounted
on another system through NFS. At least several times a week I observe
ridiculous delays (ten seconds, sometimes even a minute), even while
both systems are up and the network is working perfectly. Sometimes the
operation succeeds, sometimes it doesn't---and I can't interrupt the
process in question during whatever awfully slow operations it's trying
to do. At least once a month the file handle for the filesystem goes
stale without any warning and for no apparent reason; it has to be
unmounted and remounted, and all processes accessing the disk have to be
killed. At least once a month I observe simply incorrect behavior of
flock() and unlink() and other calls. On two occasions in January ..
from the mount point disappeared, for no apparent reason; this wasn't a
problem for anything except pwd (and ofiles, which I was porting at the
time).

The client is a Sun 4/280 under SunOS 4.0.3 and the server is a VAX 8650
under MORE/BSD 4.3. Sure, I can believe that later releases work better,
but the above problems should have been eradicated before NFS ever hit
the streets. I don't particularly mind the delays (I just switch to
another job on another system until this one wakes up), and I haven't
lost any files (yet) because of the nastier bugs. But they're still
problems.

As for the design of NFS: Anyone who believes that he's writing a
stateless filesystem is deluding himself or doesn't understand what
``stateless'' means. (Whenever Sun wants to get out of making NFS work
right, it puts up the banner of ``statelessness'' and continues
developing some other buggy subsystem.) And I would be remiss if I
didn't mention how easy it is to spoof NFS as soon as you control any
machine on the same network (or, if you know what you're doing, as soon
as the server goes down). I imagine that Sun thinks security would
contradict NFS's ``stateless'' design.

(What next? Shall we explain why Sun's make is such a disaster?)

---Dan



More information about the Comp.unix.internals mailing list