Understanding NFS
Pages: 1, 2

Looking at the contents of $HOME/serverhome, we see that the files are mostly owned by mwlucas, with the occasional file owned by root. Those files shouldn't be there, really; I have some root-owned .png files that are screenshots of terminal windows that I took for my book. Since I happen to be thinking about it, I'll remove them.

# su
# cd serverhome
# rm usersetup.png 
override rw-r--r--  root/mwlucas for usersetup.png? y
rm: usersetup.png: Permission denied

But I'm root! Why would it not let me delete a file?

I'm root on the client, but not on the server. The server doesn't trust root on other machines to execute commands as root on the server. It does trust usernames, however. NFS has a special option for handling root; you can map requests from root to any other username. For example, you might say that all requests from "root" on a client will run as "nfsroot" on the server. With careful use of groups, you could allow this nfsroot user to have limited access to things. Use the -maproot option to map root to another user.

This is my home network and I don't have other users, however, so I can just trust the client. In this case, I map NFS root requests to uid 0, aka root. Here's the new exports file.

/usr/home/mwlucas -maproot=0

I have to restart mountd again. Since NFS is stateless, old mounts remain in effect. On the client, I'm even in the same directory; restarting mountd didn't affect my client at all. The rm runs flawlessly now.

So, what if I want to export another directory on the same partition? For example, suppose I want to export /usr/src to my laptop to save space on the hard drive? List the other directories in /etc/exports, right after the first exported directory, separated by a space. While I'm at it, I could also NFS-mount /usr/obj. This way, I could run make buildworld on my fast machine and run make installworld on my slow laptop, greatly accelerating upgrades. My /etc/exports now looks like this:

/usr/home/mwlucas /usr/src /usr/obj -maproot=0

There are no identifiers between the components of the line. Yes, it would be easier to read if we could put each shared directory on its own line, but we can't; they're all on the same partition. The FreeBSD team could rewrite this so that it had more structure, but then our /etc/exports would be incompatible with that from any other Unix. Restart mountd and mount these filesytems:

# mount server:/usr/obj /usr/obj
# mount server:/usr/src /usr/src

My client filesystems now look like this:

# df
Filesystem       1K-blocks     Used    Avail Capacity  Mounted on
/dev/ad0s1a          99183    67411    23838    74%    /
/dev/ad0s1f        5186362  3873110   898344    81%    /usr
/dev/ad0s1e         198399    21212   161316    12%    /var
procfs                   4        4        0   100%    /proc
server:/usr/obj   34723447  3886485 28059087    12%    /usr/obj
server:/usr/src   34723447  3886485 28059087    12%    /usr/src

A buildworld on my laptop takes two hours; on the server, it takes thirty minutes. I ran a make buildworld and a make buildkernel earlier on the server; if I wanted the same kernel, I could just go into /usr/src and type make installkernel && make installworld && mergemaster, and the FreeBSD built on my server would be installed on the laptop.

To make things slightly more difficult, my home system is on a cable modem. I'm using one of FreeBSD's built-in firewalls to protect the system, but it would be nice to add some security into NFS itself. You can easily restrict NFS mounts by IP address, allowing only certain clients to mount an exported share. Just specify the host name or IP address at the end of the partition's /etc/exports entry. My client has an IP address of, so my /etc/exports now looks like this:

/usr/home/mwlucas /usr/src /usr/obj -maproot=0

This quickly becomes very convenient for upgrades. What's nice is that you can also build ports on the server, and install them on the client, by exporting /usr/ports. This is starting to get a little ridiculous, however; eventually, /etc/exports will have entries for almost every directory in /usr. I might as well export the whole of /usr and get it over with. This isn't that great, either, because I want to mount the exported directories at different places. The exported ports tree should go over /usr/ports, for example, but I don't want my server's home directory overwriting mine. Fortunately, there's an easy solution. The -alldirs option allows you to export a partition, and all the directories beneath it. You must specify a partition when you use alldirs. Multiple options are separated by a comma.

/usr -alldirs,maproot=0

Now I hup mountd again, and all of /usr can be exported and mounted separately. For example, I still had /usr/src and /usr/obj NFS-mounted. Quick tests with df and ls show that they're still there, still accessible, and still serving files, even though I've completely redone the way they're exported. I can mount arbitrary directories from the server's /usr.

Finally, here's a tip on NFS performance. By default, FreeBSD uses conservative NFS mounting options. These work well when trying to interoperate with other Unixes; everybody speaks the lowest common denominator. You can use mount options to augment NFS performance but reduce interoperability somewhat. These options aren't necessary when you're working with one or two clients, but as your NFS installation grows, you'll find them helpful. They may or may not work with other operating systems; it depends on what those OSs support.

First of all, NFS runs over UDP by default. The tcp option tells the client to request a mount over TCP.

Then we want to make the mount interruptible. This means that if the server goes away, for any reason, client programs that are trying to access the export can be interrupted. Otherwise, the client will continue trying to access the export until the filesystem times out. Set this with the intr option.

NFS comes in many versions. The latest one in wide use is Version 3. You can request this with the nfsv3 option.

Managing NFS and NIS

Related Reading

Managing NFS and NIS
By Hal Stern, Mike Eisler, Ricardo Labiaga

Finally, you have the size of read and write requests. The defaults are rather small, being well suited to networks of the early nineties. You can set the read and write size to more modern values with the -r and -w options. 32768 is a good value for both.

So, putting this all together, I want to mount my exports in the following way.

# mount -o tcp,intr,nfsv3,-w=32768,-r=32768 server:/usr/home/mwlucas serverhome

This will give me close-to-optimal performance with minimal effort. Further fine-tuning NFS requires testing various options with your particular network equipment.

Now that I have NFS running at home, I find myself mounting from one machine to others for just about any reason. It's well worth the minor trouble to set up.

Michael W. Lucas

Read more Big Scary Daemons columns.

Return to the BSD DevCenter.