Understanding NFS

02/14/2002

We've discussed sharing filesystems via SMB a few times. SMB lets you access files shared by a Windows system after jumping through only half a dozen loops. Sharing files with another Unix system is much, much simpler. FreeBSD supports the Unix standard Network File System out of the box. NFS intimidates many junior system administrators, but it's really quite simple once you know what's going on.

Each NFS connection works on a client-server model. One computer is the server and offers filesystems to other systems. This is called "NFS exporting," and the filesystems offered are called "exports." The clients can mount server exports in a manner almost identical to that used to mount local filesystems.

One interesting thing about NFS is its statelessness. You can reboot a server and the client won't crash. It won't be able to access files on the server's export while the server is down, but once it returns, you'll pick up right where things left off. Other network file sharing systems are not so resilient.

In this discussion, we're going to link multiple FreeBSD to 4-STABLE systems. Each NFS implementation is slightly different. You will find minor NFS variations between Solaris, Linux, BSD, and other Unixes. NFS should work between them all, but it might require the occasional tweak. If you have problems interoperating with some other Unix, check the freebsd-net mailing list archive; the issue has almost certainly been discussed there.

Let's configure some systems for NFS. Both server and clients require NFS support in the kernel, but the various NFS commands dynamically load the appropriate module into the kernel. FreeBSD's generic kernel supports NFS, but if you customize your kernel and don't like loading file system support as a module, be sure your kernel configuration includes:

options		NFS

First of all, we have the server side. You can enable basic NFS exports with the following rc.conf options:

portmap_enable="YES"
nfs_server_enable="YES"

Portmap provides, as you might guess, a mapping service for network ports. Different exports and clients require unique network ports. Clients ask portmap which port they should connect to for their actual mount. The nfs_server_enable option starts nfsd and mountd. mountd just listens for incoming NFS requests on assorted high-numbered network ports, and makes these port numbers available to portmap. When clients talk to portmap and mountd, nfsd actually handles their requests.

Once you reboot, your server should show something like the following amongst it's sockstat(1) output. This shows that server-side NFS is running more or less properly. If you don't see something resembling this, check /var/log/messages for log messages indicating your problem.

# sockstat -4
USER    COMMAND   PID  FD PROTO  LOCAL ADDRESS  FOREIGN ADDRESS
root    rpc.stat   77   3 udp4   *:1011         *:*
root    rpc.stat   77   4 tcp4   *:1022         *:*
root    nfsd       70   3 tcp4   *:2049         *:*
root    mountd     68   3 udp4   *:1023         *:*
root    mountd     68   5 tcp4   *:1023         *:*
daemon  portmap    66   3 udp4   *:111          *:*
daemon  portmap    66   4 tcp4   *:111          *:*

Your client won't show any special sockstat output before network shares are mounted, but a ps -ax will display several nfsiod processes.

Now that your systems are prepared to handle NFS, we need to tell the server which directories it can export. We could just export the entire server, but that's not generally a good thing. Clients should have little or no need to remote-mount the server's root filesystem, for example. Define allowed exports in the /etc/exports file. This file has a separate line for each hard-drive partition on the server. Each line has up to three components:

• the directories to be exported
• the options on that export
• the clients that can connect

Each combination of clients and server disk partition can only have one line in the exports file. This means that if /usr/ports and /usr/home are on the same partition, they must be exported in the same line to any one client. You don't have to export the entire partition, mind you; you can just as easily share out a single directory within a partition. This directory must be a true path, and it must not contain either symlinks or double or single dots (i.e., ".."). If I wanted to export my home directory to every host on the Internet, I could use an /etc/exports line that consisted entirely of this:

/usr/home/mwlucas

We have no options and no host restrictions. Now that we've edited the exports file, we have to tell mountd to re-read it.

killall -1 mountd

Any problems will appear in /var/log/messages. For example, when I was first testing this, I had a single entry in /etc/exports of /home/mwlucas. It was about this time that I learned that /etc/exports cannot contain symlinks. FreeBSD puts user home directories under /usr/home and uses a symlink to create the /home directory. The error log gave me a warning like this:

Jan 24 07:13:35 server mountd[68]: bad exports list line /home/mwlucas

This told me where the problem was; identifying the problem was my job.

Now, over on the client side, I create the directory $HOME/serverhome. I want to NFS-mount my home directory on the server onto this directory. This looks almost exactly like a standard mount(8) command. mount takes two arguments: the physical device you're using and the mount point. In this case, our physical device is a remote server and the exported filesystem:

# mount server:/usr/home/mwlucas serverhome
#

Once this finishes, test your mount with df(1):

# 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    21211   161317    12%    /var
procfs                           4        4        0   100%    /proc
server:/usr/home/mwlucas  34723447  3886523 28059049    12%    /usr/home/mwlucas/serverhome
#

One thing to note is that NFS uses the same usernames on each side of the connection. My files are owned by mwlucas on the server, so they are owned by mwlucas on the client. This can be a problem on a large network where users have root on their own machines. To create a central repository of authorized users, consider Kerberos or NIS. On a small network, or on a network with limited administrators, this usually isn't a problem.

Pages: 1, 2

Next Page