NFS umount takes ages when no DNS available

[Rick Macklem]

Robert Schulze wrote:
> Hi,
> 
> during maintainance work, I realized that an umount -h ip.ad.re.ss
> takes
> very long time when there is no nameserver reachable by the client
> (and
> server). All nfs configuration is done without hostnames, so I wonder
> why there is this delay (about 1 minute per mountpoint).
> 
Well, here is the code snippet. (I'm not sure why the author felt that
the getaddinfo() needed to be done before the check for the need to do
an rpc?):
if (hostp != NULL) {
341 	*delimp = '\0';
342 	getaddrinfo(hostp, NULL, &hints, &ai);
343 	if (ai == NULL) {
344 	warnx("can't get net id for host");
345 	}
346 	}
347 	
348 	/*
349 	* Check if we have to start the rpc-call later.
350 	* If there are still identical nfs-names mounted,
351 	* we skip the rpc-call. Obviously this has to
352 	* happen before unmount(2), but it should happen
353 	* after the previous namecheck.
354 	* A non-NULL return means that this is the last
355 	* mount from mntfromname that is still mounted.
356 	*/
357 	if (getmntentry(sfs->f_mntfromname, NULL, NULL,
358 	CHECKUNIQUE) != NULL)
359 	do_rpc = 1; 

Just to clarify, umount(8) does do an rpc against the server,
but it isn't very important. All it does is tell the server
to remove an entry from the table it uses to generate replies
to showmount(8).

NFS itself doesn't care about this table. If you do a
# umount /mnt
and the server name can't be resolved via DNS, you can <ctrl>C
out and the mount point will be gone in the client. (As above,
I'm not sure why the author felt that the getaddrinfo() should
be done before the umount(2) for the "-h" case.

As such, I would be hesitant to change it..

Even if you change it, it would need to be done later, so the
RPC to the server can be performed. The <ctrl>C trick would then
work, but only if there was only one mount point on the server.

> client: 8.1-RELEASE-p1/amd64
> server: 8.2-STABLE/amd64
> 
> The hostname seems to be exchanged by client and server, this can be
> noticed by warnings like the following in /var/log/messages on the
> server:
> 
> rpc.statd: Failed to contact host client.foobar.net: RPC: Port mapper
> failure - RPC: Timed out
> 
> statd is running on the client, but bound to an interface with an ip
> address not matching the logged hostname.
> 
> Could NFS be tweaked in that way, that it does not use hostnames at
> all
> or at least will not provide hostnames on interfaces with don't carry
> the ip address the hostname matches to?
> 
I don't know if the hostnames are in arguments on the wire for the Network
Status Monitor (NSM) and Network Lock Manager (NLM) protocols. To find out,
you'll probably need to read the sources, since there isn't any RFCs for
these. (They were eventually published in an X/Open manual, but in the
old days, you had to buy it hardcopy, since that was a funding source for
X/Open.)

These protocols were done in the days when servers would have been in
/etc/host files, so the names always resolved. (mid to late 1980s)

rick