NFS - Network File System


System Administration



Original in fr Frédéric Raynal

fr to en Philippe Trbich and Emmanuel Bonnel


Frédéric Raynal is preparing a thesis in computer science about picture tattoo at the INRIA. He is reading a very good detective novel which implies Th. Roosevelt in the beginning of the century when he was a police prefect. The atmosphere is very dark. It's about the enquiry of a group of persons in order to find a serial killer who injures children. This group is supported by new technologies (psychology, fingerprint, etc...) to find the solution. This novel from Caleb Carr, L'ange des ténèbres, paints a surprising picture about the beginning of the last century.


A Network File System (NFS) allows to manage files on several computers inside a network as if they were on the local hard disk. So, there's no need to know where the files are physically located in order to access them.




NFS easily allows to share data between several computers. For example, a user logged on a network won't need to log on a specific computer: via NFS, he will access his home directory (we say exported) on the machine he is working at.

But NFS isn't a very efficient protocol and thus very slow over a modem connection. It's designed for a local network and is very flexible. It offers a lot of possibilities for users and administrators.

You'll have to administrate this service with care. Allowing everybody to write data into your net would not really be good policy ;-) Some essential actions may reduce risks.

This article starts with a very short introduction on file systems. Then we will see the NFS protocol. After that will go into the less theoretical part and do a NFS server and a client installation. We will also look at the minimum security precautions you should take. Then with an example, we will illustrate how to combine NFS, NIS and autofs.

General and non-exhaustive presentation of file systems

Before talking about NFS, you should understand the term file system. A file system is a way of storing data on a media, the way it is organized and managed. There are a lot of them, some much more used than the others (New Technology FileSystem (NTFS), High Performance FileSystem (HPFS), DOS, FAT 12/16/32, VFAT, Macintosh Hierarchical Filesystem (HFS), ISO 9660 (for CD-ROM), extended file systems (Ext, Ext2, Ext3), and many others).

For example, we can consider that every physical media for data (hard disk for example) is an array of small units holding information: we are talking about blocks. Every file system manages these blocks in a different way. For example, in figure 1 , we try to insert a file that will use two blocks. On the upper illustration, the file has been put after the last occupied block, leaving empty spaces at the beginning. In the lower part of the picture (a different files system), it has been put in the first free location. Such a policy has influence on how much the disk becomes fragmented. Some file system automatically avoid fragmentation while others need to be de-fragmented manually.

Fig. 1 : 2 different ways to place blocks

The most famous file system on Linux is called ext2fs (extended 2 file system). Every file is represented by an inode1. Directories hold the file list and the device access is done with operations like reading/writing on particular files.

The NFS server's task is to give its clients the inodes they want to access. However, a client wouldn't work very well only with the file's inode! An NFS server gives an additional net layer allowing remote machines to handle the inodes.

The NFS protocol

What we commonly name NFS is build from 4 distinct protocols. Each relies on the Remote Procedure Calls(RPC) and portmap (also called rpc.portmap). A portmapper converts RPC program numbers into port numbers. When a RPC server is starting, it tells portmap which port will be used and the managed RPC program number. When a client wishes to send a RPC request to a given program number, it first contacts the server portmap to get the port number giving access to the desired program. Then, it addresses the RPC packets to the corresponding port.

The 4 services allowing NFS to work are:
nfs This protocol is the base and allows for files creation, searching, reading or writing. This protocol also manages authentication and file statistics.
mountd This one is in charge of mounting exported systems in order to access them with nfs. The server receives requests like mount and umount and so must keep information about exported file systems.
(Network Status Monitor) 
It's used to monitor network nodes in order to know a machine's state (client or server). It informs, for example, about a reboot.
(Network Lock Manager) 
To avoid data modifications by several clients at the same time, this protocol manages a lock system. It knows which files are used. Thus, with the help of the Nsm protocol it is possible to know when a client is restarting. Nsm frees every client's locks before giving them back.

The daemon knfsd, available with latest kernel versions, directly supports the nfs and nlm protocols. On the other hand, mountd and nsm are not yet supported. When the NFS server is installed and started, we can verify that everything is working with the command:

>> ps auxwww | egrep "nfs|mount|lock|stat"
root      1370  0.0  0.2  1176  580 ?        S    22:28   0:00 rpc.mountd --no-nfs-version 3
root      1379  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1380  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1381  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1382  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1383  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1384  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1385  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1386  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [nfsd]
root      1399  0.0  0.0     0    0 pts/0    SW   22:28   0:00 [lockd]
root      1409  0.0  0.2  1156  560 ?        S    22:28   0:00 rpc.statd
root      1652  0.0  0.1  1228  484 pts/3    S    22:49   0:00 egrep nfs|mount|lock|stat
At the moment, 2 NFS versions are available (versions 2 and 3 - they will be respectively noted NFSv2 and NFSv3 to distinguish them). Linux's NFS servers only supports version 2 (hence the option on the mountd line in the previous example).

The NFS is about a data structure called a file handle. It's quite an esoteric bit series which allows to identify in a unique way each file system object (such as a file but not only files). It contains for example the file inode but also an entry representing the device where file is located. So we can look at NFS as a file system embedded within a file system.


The server

The first thing to do, as we already saw, is to start portmap since this protocol is needed by NFS .
root >>/usr/sbin/rpcinfo -p
rpcinfo: can't contact portmapper: RPC: Remote system error - Connection refused
root >>/sbin/portmap
root >>/usr/sbin/rpcinfo -p
   program vers proto   port
    100000    2   tcp    111  portmapper
    100000    2   udp    111  portmapper
The rpcinfo command shows the RPCs services working on the machine specified as the argument (-p option). We notice that portmap isn't working yet: we start it (most of Linux distributions provide scripts to automate this on startup) and we check that it's working. Another common reason for a negative answer to rpcinfo is that the portmapper is not allowed to answer because of security restrictions in /etc/hosts.{allow, deny} files. In that case add a "portmap: hosts" entry to the hosts.allow file.

Before starting NFS itself, it must be configured. There is only one configuration file and it is called /etc/exports. Each line shows the export location followed by a list of clients allowed to access it. It is possible to add options at the end of each client name. The man exports page explains the syntax for client names and options.

The accepted forms for client names are:

We won't detail here all the available mount options, but here are the most important ones: We now have to start rpc.mountd and rpc.nfs daemons to have a working NFS server. We check that everything is working again with the rpcinfo command. We can even initialize the server for the nsm and nlm protocols (respectively rpc.statd and rpc.lockd). They aren't essentials to run an NFS server... but strongly recommended in case a machine fails, reboots on its own, etc...

When we modify the /etc/exports configuration file, we must warn the affected daemons that changes have been made. The exportfs command transmits this information to our servers. The -r option synchronizes the /etc/mtab2 file with the /etc/exports file. The -v option shows the exported files systems together with theirs options.

After startup the following files are holding important information:

When a client wishes to access a file system, it starts asking mountd. It then searches in etab if the query is available. It checks the kernel to know if the client is allowed to query (check hosts.{allow, deny}, firewall rules, ...). The kernel uses exportfs for the check, allowing it to update the /var/lib/nfs/etab file. If, in this file, the exported system is allowed to be exported to the group the client is part of, then mountd informs the kernel which in turn updates xtab with this new host.

The client

Nothing to do ... usually. The access to a file system exported by NFS is directly managed from the kernel. It has to have compiled in support for NFS. The file /proc/filesystems lists all the file systems directly supported by the kernel. You then only have to tell the kernel that you want to access a system exported by NFS.

The mount command allows to access different file systems. It informs the kernel that a new file system is available indicating its type, its device and a mount point. The -t  option can be used to specify the type of file system to use. For NFS, we write: -t nfs.

mount has its own options for NFS. For example, the options rsize and wsize can be used to modify the blocks size for reading or writing. You can combine NFS specific options with more general options like intr, noexec or nosuid. The mount man page lists all these options.

Let's assume the machine charly has an NFS server and exports its /usr/local directory. When you want access it from the machine jill, then you just have to mount the exported directory from charly to jill:

root@jill >> mount -t nfs -o nosuid,hard,intr charly:/usr/local /usr/local
The command indicates we are mounting a NFS file system (-t nfs), with the nosuid, hard and intr options. The 2 last arguments are the most interesting. The first one specifies the device to mount. For NFS, the syntax is different from the usual mount command line where you specify device and directory. Here we specify server:exported_directory instead of a device. The last argument indicates the location of the file system on the client side. We just share charly's /usr/local with jill and can avoiding to install programs in /usr/local more than once. To make this set up permanent, we can specify it in the jill /etc/fstab file. fstab contains all the devices to be mounted on start up. The syntax for /etc/fstab is:
#    device           mount point   file system   options     dump  fsckorder
charly:/usr/local    /usr/local       nfs       nosuid,hard,intr    0      0

However be careful with a permanent entry. You can use it only when either the server (charly) is all the time on or when you switch on charly before jill.


One major problem with NFS comes from the fact that there exists by default, a trusted relation between a client and a NFS server. In this case if the server's root account is compromised, the client's one will be compromised too. The NFS-HOWTO decribes a set of essential measures to be taken to get some security.

A client can't blindly trust a server, so we must specify constraining options when using the mount command. We already mentioned the first one: nosuid. It cancels the SUID and SGID bits effect. So, a root person on the server must first login as a user on the client and then become root. Another option, more restrictive, is noexec. It forbids to execute programs on the exported file system. This option is only usable on systems holding data only.

On the NFS server side, we can also specify that we don't trust the client's root account. We have to specify this in /etc/exports with the root_squash option. Then if a user with UID 0 (root) on the client accesses the file system exported by the server, it gets the nobody UID to query files. This option is active by default under Linux but can be canceled with the no_root_squash option. A set of UIDs can be specified to which the the option should apply. Remember also that anonuid and anongid options allow to change user's UID/GID from nobody to different one.

Some actions are more general and effect the portmapper. For example, we forbid the access to all machines with the following line in the /etc/hosts.deny file:

# hosts.deny : absolute prohibition for every one to
#              use the portmap
portmap: ALL

Then in the /etc/hosts.allow file, this strict prohibition is counterbalanced allowing the access to the desired machines.

Good firewall rules also contribute to a better protection. Watch the ports used by the different services and the used protocols:

Service RPC Port Protocols
portmap 111 upd / tcp
nfsd 2049 udp
mountd variable udp / tcp

Using NIS, NFS and autofs

Now we look at a more complicated network as you will find it e.g in a company. In a little network at home you will probably be able to live without NIS. NIS Network Information Service is a way to distribute configuration files (e.g in /etc) to other machines.
The main server in our network is called "charly" and 3 other machines of the network are "sabrina", "jill" and "kelly". We configure charly as NIS server for domain bosley. The other machines are only NIS clients of charly (we should have a NIS slave server but this isn't our purpose today).

First, let's see the configuration of our server charly. We start with defining some NIS maps holding all needed information.

The /etc/netgroup file holds machine groups having common characteristics (a same architecture for example). A NIS map is very useful for NFS. We just have to gather all the machines allowed to access the same exported file system. This group is then used in the /etc/exports file instead of specifying all clients one by one:

# /etc/netgroup
charlysangels (sabrina,,) (jill,,) (kelly)
 As far as NFS is concerned, we know the configuration is quite restricted. The /etc/exports file from charly holds:
# /etc/exports
/usr/local    @charlysangels(ro)
We decide to use automount to access the exported /usr/local directory. Instead of mounting this system at boot time, it's done automatically when a user accesses a file in this directory.  We create the /etc/ file to define what will be accessible both by automount and by NIS:
# /etc/
charly          charly:/usr/local
Since we want these informations (the new and netgroup files) to be integrated into the NIS database, we have to modify the Makefile before rebuilding it. We must be sure that netgroup will be added to the base. Concerning, this file isn't defined by default . We just have to add a new entry in the Makefile, with the associated rule (using the existing one as a model):
#To be added in the Yellow Pages Makefile
# ...
#... $(AUTO_MAP) $(YPDIR)/Makefile
            @echo "Updating $@..."
            -@sed -e "/^#/d" -e s/#.*$$// $(AUTO_MAP) | $(DBLOAD) \
            -i $(AUTO_MAP) -o $(YPMAPDIR)/$@ - $@
            -@$(NOPUSH) || $(YPPUSH) -d $(DOMAIN) $@
This production rule just removes comments, adds a new entry to the database and then transmits the information to every server.

We just have to run make from the /var/yp directory.

Now, our three clients sabrina, jill and kelly. Here, there is nothing to do :) We must tell autofs to manage a new map given by YPs. In each client's /etc/auto.master file the following line informs about the presence of a map obtained via YP's services.

/usr/local    yp    --intr,nosuid,nodev
After that we have to restart autofs to make this new map effective.

Now we have a unique /usr/local physical directory on charly. Then, when specific programs are installed on charly, all machines can use them.

This example could go further with the installation of only one /usr system, /usr/doc or others, but the practice shows it isn't a good idea. Installations often need to modify files in the /etc directory or others. We then should have to modify all the machines to update non-exported files and that becomes quickly boring.


Files Systems



... inode1
It's a file descriptor (a continuation of bits) holding, among other, file access permissions, owner, physical blocks addresses holding data, etc...
... /etc/mtab2
This file contains the list of all file systems mounted by the kernel, when it's hard (via mount, like the systems described in the fstab), or by means less continuous (via autofs/automount).