The following sections describe some of the complex functions of the NFS software.
Because NFS servers might be supporting clients that are not using the NFS version 3 software, part of the initiation procedure includes negotiation of the protocol level. If both the client and the server can support version 3, that version is used. If either the client or the server can only support version 2, that version is selected.
You can override the values that are determined by the negotiation by using the -vers option with the mount command (see the mount_nfs(1M) man page). Under most circumstances, you should not have to specify the version level, as the best level is selected by default.
During initiation, the transport protocol is also negotiated. By default, the first connection-oriented transport that is supported on both the client and the server is selected. If this selection does not succeed, the first available connectionless transport protocol is used. The transport protocols that are supported on a system are listed in /etc/netconfig. TCP is the connection-oriented transport protocol that is supported by the release. UDP is the connectionless transport protocol.
When both the NFS protocol version and the transport protocol are determined by negotiation, the NFS protocol version is given precedence over the transport protocol. The NFS version 3 protocol that uses UDP is given higher precedence than the NFS version 2 protocol using TCP. You can manually select both the NFS protocol version and the transport protocol with the mount command (see the mount_nfs(1M) man page). Under most conditions, allow the negotiation to select the best options.
The file transfer size establishes the size of the buffers that are used when transferring data between the client and the server. In general, larger transfer sizes are better. The NFS version 3 protocol has an unlimited transfer size, but starting with the Solaris 2.6 release, the software bids a default buffer size of 32 Kbytes. The client can bid a smaller transfer size at mount time if needed, but under most conditions this bid is not necessary.
The transfer size is not negotiated with systems that use the NFS version 2 protocol. Under this condition, the maximum transfer size is set to 8 Kbytes.
You can use the -rsize and -wsize options to set the transfer size manually with the mount command. You might need to reduce the transfer size for some PC clients. Also, you can increase the transfer size if the NFS server is configured to use larger transfer sizes.
When a client needs to mount a file system from a server, it must obtain a file handle from the server that corresponds to the file system. This process requires that several transactions occur between the client and the server. In this example, the client is attempting to mount /home/terry from the server. A snoop trace for this transaction follows.
client -> server PORTMAP C GETPORT prog=100005 (MOUNT) vers=3 proto=UDP server -> client PORTMAP R GETPORT port=33492 client -> server MOUNT3 C Null server -> client MOUNT3 R Null client -> server MOUNT3 C Mount /export/home9/terry server -> client MOUNT3 R Mount OK FH=9000 Auth=unix client -> server PORTMAP C GETPORT prog=100003 (NFS) vers=3 proto=TCP server -> client PORTMAP R GETPORT port=2049 client -> server NFS C NULL3 server -> client NFS R NULL3 client -> server NFS C FSINFO3 FH=9000 server -> client NFS R FSINFO3 OK client -> server NFS C GETATTR3 FH=9000 server -> client NFS R GETATTR3 OK
In this trace, the client first requests the mount port number from the portmap service on the NFS server. After the client received the mount port number (33492), that number is used to ping the service on the server. After the client has determined that a service is running on that port number, the client then makes a mount request. When the server responds to this request, it includes the file handle for the file system (9000) that is being mounted. The client then sends a request for the NFS port number. When the client receives the number from the server, it pings the NFS service (nfsd), and requests NFS information about the file system that uses the file handle.
In the following trace, the client is mounting the file system with the public option.
client -> server NFS C LOOKUP3 FH=0000 /export/home9/terry server -> client NFS R LOOKUP3 OK FH=9000 client -> server NFS C FSINFO3 FH=9000 server -> client NFS R FSINFO3 OK client -> server NFS C GETATTR3 FH=9000 server -> client NFS R GETATTR3 OK
By using the default public file handle (which is 0000), all of the transactions to obtain information from the portmap service and to determine the NFS port number are skipped.
Using the -public option can create conditions that cause a mount to fail. Adding an NFS URL can also confuse the situation. The following list describes the specifics of how a file system is mounted when you use these options.
Public option with NFS URL - Forces the use of the public file handle. The mount fails if the public file handle is not supported.
Public option with regular path - Forces the use of the public file handle. The mount fails if the public file handle is not supported.
NFS URL only - Use the public file handle if it is enabled on the NFS server. If the mount fails when using the public file handle, then try the mount with the MOUNT protocol.
Regular path only - Do not use the public file handle. The MOUNT protocol is used.
By using client-side failover, an NFS client can switch to another server if the server that supports a replicated file system becomes unavailable. The file system can become unavailable if the server it is connected to crashes, if the server is overloaded, or if a network fault occurs. The failover, under these conditions, is normally transparent to the user. After it is established, the failover can occur at any time without disrupting the processes that are running on the client.
Failover requires that the file system be mounted read-only. The file systems must be identical for the failover to occur successfully. See "What Is a Replicated File System?" for a description of what makes a file system identical. A static file system or one that is not changed often is the best candidate for failover.
You cannot use file systems that are mounted by using CacheFS with failover. Extra information is stored for each CacheFS file system. This information cannot be updated during failover, so only one of these two features can be used when mounting a file system.
The number of replicas that need to be established for each file system depends on many factors. In general, the ideal is to have a minimum of two servers, with each supporting multiple subnets, rather than to have a unique server on each subnet. The process requires checking of each server in the list. Therefore, if more servers are listed, each mount will be slower.
To fully comprehend the process, you need to understand two terms.
failover - The process of selecting a server from a list of servers that support a replicated file system. Normally, the next server in the sorted list is used, unless it fails to respond.
remap - To making use of a new server. Through normal use, the clients store the path name for each active file on the remote file system. During the remap, these path names are evaluated to locate the files on the new server.
For the purposes of failover, a file system can be called a replica when each file is the same size and has the same vnode type as the original file system. Permissions, creation dates, and other file attributes are not considered. If the file size or vnode types are different, the remap fails and the process hangs until the old server becomes available.
You can maintain a replicated file system by using rdist, cpio, or another file transfer mechanism. Because updating the replicated file systems causes inconsistency, follow these suggestions for best results:
Rename the old version of the file before installing a new one.
Run the updates at night when client usage is low.
Keep the updates small.
Minimize the number of copies.
Some software packages require read locks on files. To prevent these products from breaking, read locks on read-only file systems are allowed, but are visible to the client side only. The locks persist through a remap because the server doesn't "know" about them. Because the files should not change, you do not need to lock the file on the server side.
Starting with 2.6, the Solaris release supports files that are over 2 Gbytes. By default, UFS file systems are mounted with the -largefiles option to support the new functionality. Previous releases cannot handle files of this size. See "How to Disable Large Files on an NFS Server" for instructions.
If the file system on the server is mounted with the -largefiles option, no changes need to occur on a Solaris 2.6 NFS client in order for it to be able to access a large file. However, not all 2.6 commands can handle these large files. See largefile(5) for a list of the commands that can handle the large files. Clients that cannot support the NFS version 3 protocol with the large file extensions cannot access any large files. Although clients that run the Solaris 2.5 release can use the NFS version 3 protocol, large file support was not included in that release.
NFS server logging provides records of NFS reads and writes, as well as operations that modify the file system. This data can be used to track access to information. In addition, the records can provide a quantitative way to measure interest in the information.
When a file system with logging enabled is accessed, the kernel writes raw data into a buffer file. This data includes the following:
The client IP address
The UID of the requester
The file handle of the file or directory object that is being accessed
The type of operation that occurred
The nfslogd daemon converts this raw data into ASCII records that are stored in log files. During the conversion, the IP addresses are modified to host names and the UIDs are modified to logins if the name service that is enabled can find matches. The file handles are also converted into path names. To accomplish the conversion, the daemon tracks the file handles and stores information in a separate file handle-to-path table. That way the path does not have to be re-identified each time a file handle is accessed. Because no changes to the mappings are made in the file handle-to-path table if nfslogd is turned off, you must keep the daemon running.
The WebNFS service makes files in a directory available to clients by using a public file handle. A file handle is an address that is generated by the kernel that identifies a file for NFS clients. The public file handle has a predefined value, so the server does not need to generate a file handle for the client. The ability to use this predefined file handle reduces network traffic by eliminating the MOUNT protocol and should accelerate processes for the clients.
By default, the public file handle on an NFS server is established on the root file system. This default provides WebNFS access to any clients that already have mount privileges on the server. You can change the public file handle to point to any file system by using the share command.
When the client has the file handle for the file system, a LOOKUP is run to determine the file handle for the file to be accessed. The NFS protocol allows the evaluation of only one path name component at a time. Each additional level of directory hierarchy requires another LOOKUP. A WebNFS server can evaluate an entire path name with a single transaction, called multicomponent lookup, when the LOOKUP is relative to the public file handle. With multicomponent lookup, the WebNFS server can deliver the file handle to the desired file without having to exchange the file handles for each directory level in the path name.
In addition, an NFS client can initiate concurrent downloads over a single TCP connection. This connection provides quick access without the additional load on the server that is caused by setting up multiple connections. Although Web browser applications support concurrent downloading of multiple files, each file has its own connection. By using one connection, the WebNFS software reduces the overhead on the server.
If the final component in the path name is a symbolic link to another file system, the client can access the file if the client already has access through normal NFS activities.
Normally, an NFS URL is evaluated relative to the public file handle. The evaluation can be changed to be relative to the server's root file system by adding an additional slash to the beginning of the path. In this example, these two NFS URLs are equivalent if the public file handle has been established on the /export/ftp file system.
The Solaris 8 release includes a new protocol so a WebNFS client can negotiate a selected security mechanism with a WebNFS server. The new protocol uses security negotiation multicomponent lookup, which is an extension to the multicomponent lookup that was used in earlier versions of the WebNFS protocol.
The WebNFS client initiates the process by making a regular multicomponent lookup request by using the public file handle. Because the client has no knowledge of how the path is protected by the server, the default security mechanism is used. If the default security mechanism is not sufficient, the server replies with an AUTH_TOOWEAK error, indicating that the default mechanism is not valid and the client needs to use a stronger one.
When the client receives the AUTH_TOOWEAK error, it sends a request to the server to determine which security mechanisms are required. If the request succeeds, the server responds with an array of security mechanisms that are required for the specified path. Depending on the size of the array of security mechanisms, the client might have to make more requests to obtain the complete array. If the server does not support WebNFS security negotiation, the request fails.
After a successful request, the WebNFS client selects the first security mechanism from the array that it supports. The client then issues a regular multicomponent lookup request by using the selected security mechanism to acquire the file handle. All subsequent NFS requests are made by using the selected security mechanism and the file handle.
Several functions that a Web site that uses HTTP can provide are not supported by the WebNFS software. These differences stem from the fact that the NFS server only sends the file, so any special processing must be done on the client. If you need to have one web site configured for both WebNFS and HTTP access, consider the following issues:
NFS browsing does not run CGI scripts, so a file system with an active web site that uses many CGI scripts might not be appropriate for NFS browsing.
The browser might start different viewers, in order to handle files in different file formats. Accessing these files through an NFS URL starts an external viewer if the file type can be determined by the file name. The browser should recognize any file name extension for a standard MIME type when an NFS URL is used. The reason is that the WebNFS software does not check inside the file to determine the file type, so the only way to determine a file type is by the file name extension.
NFS browsing cannot utilize server-side image maps (clickable images). However, it can utilize client-side image maps (clickable images) because the URLs are defined with the location. No additional response is required from the document server.
The NFS environment is a powerful and convenient way to share file systems on a network of different computer architectures and operating systems. However, the same features that make sharing file systems through NFS operation convenient also pose some security problems. Historically, most NFS implementations have used UNIX (or AUTH_SYS) authentication, but stronger authentication methods such as AUTH_DH have also been available. When using UNIX authentication, an NFS server authenticates a file request by authenticating the computer making the request, but not the user. Therefore, a client user can run su and impersonate the owner of a file. If DH authentication is used, the NFS server authenticates the user, making this sort of impersonation much harder.
With root access and knowledge of network programming, anyone can introduce arbitrary data into the network and extract any data from the network. The most dangerous attacks are those that involve the introduction of data, such as impersonating a user by generating the right packets or recording "conversations" and replaying them later. These attacks affect data integrity. Attacks that involve passive eavesdropping-merely listening to network traffic without impersonating anybody-are not as dangerous, as data integrity is not compromised. Users can protect the privacy of sensitive information by encrypting data that is sent over the network.
A common approach to network security problems is to leave the solution to each application. A better approach is to implement a standard authentication system at a level that covers all applications.
The Solaris operating environment includes an authentication system at the level of remote procedure call (RPC)-the mechanism on which NFS operation is built. This system, known as Secure RPC, greatly improves the security of network environments and provides additional security to services such as the NFS system. When the NFS system uses the facilities that are provided by Secure RPC, it is known as a Secure NFS system.
Secure RPC is fundamental to the Secure NFS system. The goal of Secure RPC is to build a syste that is at minimum as secure as a time-sharing system (one in which all users share a single computer). A time-sharing system authenticates a user through a login password. With data encryption standard (DES) authentication, the same authentication process is completed. Users can log in on any remote computer just as they can on a local terminal, and their login passwords are their passports to network security. In a time-sharing environment, the system administrator has an ethical obligation not to change a password to impersonate someone. In Secure RPC, the network administrator is trusted not to alter entries in a database that stores public keys.
You need to be familiar with two terms to understand an RPC authentication system: credentials and verifiers. Using ID badges as an example, the credential is what identifies a person: a name, address, birthday, and so on. The verifier is the photo that is attached to the badge. You can be sure the badge has not been stolen by checking the photo on the badge against the person carrying it. In RPC, the client process sends both a credential and a verifier to the server with each RPC request. The server sends back only a verifier because the client already "knows" the server's credentials.
RPC's authentication is open ended, which means that a variety of authentication systems can be plugged into it. Currently, there are several systems: UNIX, DH, and KERB.
When UNIX authentication is used by a network service, the credentials contain the client's host name, UID, GID, and group-access list, but the verifier contains nothing. Because no verifier exists, a superuser could falsify appropriate credentials by using commands such as su. Another problem with UNIX authentication is that it assumes all computers on a network are UNIX computers. UNIX authentication breaks down when applied to other operating systems in a heterogeneous network.
To overcome the problems of UNIX authentication, Secure RPC uses DH authentication.
DH authentication uses the Data Encryption Standard (DES) and Diffie-Hellman public-key cryptography to authenticate both users and computers in the network. DES is a standard encryption mechanism. Diffie-Hellman public-key cryptography is a cipher system that involves two keys: one public and one secret. The public keys and secret keys are stored in the name space. NIS stores the keys in the publickey map. These maps contain the public key and secret key for all potential users. See the System Administration Guide: Naming and Directory Services (DNS, NIS, and LDAP) for more information on how to set up the maps.
The security of DH authentication is based on a sender's ability to encrypt the current time, which the receiver can then decrypt and check against its own clock. The timestamp is encrypted with DES. The requirements for this scheme to work are as follows:
The two agents must agree on the current time.
The sender and receiver must be using the same encryption key.
If a network runs a time-synchronization program, the time on the client and the server is synchronized automatically. If a time-synchronization program is not available, timestamps can be computed by using the server's time instead of the network time. The client asks the server for the time before starting the RPC session, then computes the time difference between its own clock and the server's. This difference is used to offset the client's clock when computing timestamps. If the client and server clocks get out of synchronization to the point where the server begins to reject the client's requests, the DH authentication system on the client resynchronizes with the server.
The client and server arrive at the same encryption key by generating a random conversation key, also known as the session key, and by using public-key cryptography to deduce a common key. The common key is a key that only the client and server are capable of deducing. The conversation key is used to encrypt and decrypt the client's timestamp. The common key is used to encrypt and decrypt the conversation key.
Kerberos is an authentication system developed at MIT. Encryption in Kerberos is based on DES. Kerberos support is no longer supplied as part of Secure RPC, but a server-side and client-side implementation is included with the Solaris 9 release. See "Introduction to SEAM" in System Administration Guide: Security Services for more information about the Solaris 9 implementation of Kerberos Authentication.
Be aware of the following points if you plan to use Secure RPC:
If a server crashes when no one is around (after a power failure, for example), all the secret keys that are stored on the system are deleted. Now no process can access secure network services or mount an NFS file system. The important processes during a reboot are usually run as root. Therefore, these processes would work if root's secret key were stored away, but nobody is available to type the password that decrypts it. keylogin -r allows root to store the clear secret key in /etc/.rootkey, which keyserv reads.
Some systems boot in single-user mode, with a root login shell on the console and no password prompt. Physical security is imperative in such cases.
Diskless computer booting is not totally secure. Somebody could impersonate the boot server and boot a devious kernel that, for example, makes a record of your secret key on a remote computer. The Secure NFS system provides protection only after the kernel and the key server are running. Otherwise, no way exists to authenticate the replies that are given by the boot server. This limitation could be a serious problem, but it requires a sophisticated attack, using kernel source code. Also, the crime would leave evidence. If you polled the network for boot servers, you would discover the devious boot server's location.
Most setuid programs are owned by root. If the secret key for root is stored in /etc/.rootkey, these programs behave as they always have. If a setuid program is owned by a user, however, it might not always work. For example, if a setuid program is owned by dave and dave has not logged into the computer since it booted, the program would not be able to access secure network services.
If you log in to a remote computer (using login, rlogin, or telnet) and use keylogin to gain access, you give access to your account. The reason is that your secret key is passed to that computer's key server, which then stores it. This process is only a concern if you do not trust the remote computer. If you have doubts, however, do not log in to a remote computer if it requires a password. Instead, use the NFS environment to mount file systems that are shared by the remote computer. As an alternative, you can use keylogout to delete the secret key from the key server.
If a home directory is shared with the -o sec=dh option, remote logins can be a problem. If the /etc/hosts.equiv or ~/.rhosts files are not set to prompt for a password, the login succeeds. However, the users cannot access their home directories because no authentication has occurred locally. If the user is prompted for a password, the user has access to his or her home directory if the password matches the network password.