[[!toc levels=3]] ## 2 Using AFS The Usage Section of the [[AFSFrequentlyAskedQuestions]]. - [[PreambleFAQ]] - [[GeneralFAQ]] - [[AdminFAQ]] - [[ResourcesFAQ]] - [[AboutTheFAQ]] - [[FurtherReading]] ### 2.01 What are the differences between AFS and a Unix filesystem? Essentially, from a user's point of view, there is little difference between AFS and local Unix filestore. Nearly all the commands normally used to access local files can be used to access files in `/afs`. In the following set of sections, I have attempted to "target" each section to an appropriate type of user by including to the right of each section heading one of: User, Programmer, [[SysAdmin]]. Here is a summary of the differences: **Authentication:** [ User ] Before a user can access protected AFS files (s)he needs to become authenticated to AFS using the `klog` command (Kerberos login) to get an AFS token. Or, a user can run `aklog` to convert existing krb5 tickets into an AFS token. Without a token, an unauthenticated user is given the AFS identity `system:anyuser`, and as such is only able to access files in directories that have ACLs granting `system:anyuser` access. Many systems have the `klog`/`aklog` functionality done integrated into their login procecedure. If you use a system where you have to issue the `klog`/`aklog` command after login, then you may want to run the `pagsh` command first (see below). AFS provides access control lists to give more precise control to users wishing to protect their files (see AFS ACL below). **File permissions:** [ User ] Unix mode bits on files are usually ignored. The exception to this are the "owner" bits, which can be used to remove access for almost everyone accessing the file (that is, if you remove the "w" bit from the owner permissions, almost nobody can write to the file). Instead of Unix mode bits, you should generally use AFS ACLs to protect your data (see below). **Data protection with AFS ACLs:** [ User ] Some versions of Unix (e.g. IBM's AIX version 3) allow ACLs on local files. In AFS, ACLs protect directories and used with AFS protection groups (see below) provide a finer granularity of protection than can be achieved with basic Unix file permissions. (AFS ACLs are described in more detail below.) **Protection groups:** [ User ] Users can create and maintain their own protection groups in AFS - as opposed to Unix where only sysadmins can manage protection groups. **Hard links:** [ User ] In AFS, hard links (`ln old new`) are only valid within a directory. This is because AFS ACLs protect directories (not individual files), and allowing hard links that span directories would subvert ACL protection. Symbolic links work in AFS because they reference a pathname and not an inode directly. (Hard links reference an inode directly.) **Changing file protection by moving a file:** [ User ] Moving a file to a different directory will change the protection of a file if the ACL on the new directory is different from the ACL on the original directory. **chown and chgrp:** [ User ] Only members of the AFS group `system:administrators` can use these commands on files in `/afs`. Note that these are only used by non-AFS-aware Unix utilities; AFS itself does not use these values, but instead uses the AFS ACL. **Save on close:** [ Programmer ] The AFS Cache Manager does not send file modifications to a file server until a `close()` or `fsync()` system call is performed. `write()` system calls only update the local cached copy on the client. Note the difference in semantics of writing a file:
local Unix file writes update the file immediately
AFS file local cached copy updated immediately but the server copy is only updated when the file is `close`d or `fsync`ed
It is important to understand that most applications (`vi`, `emacs`, `frame`, `interleaf`, `wingz`, `dogz`, etc.) issue the `close()` system call when the user chooses/issues the "save" command in the application. Users are not required to exit the application to save their changes back to the server. **Byte-range file locking:** [ Programmer ] AFS does not support byte-range locking within a file, although `lockf()` and `fcntl()` calls will return 0 (success). The first time a byte-range lock is attempted, AFS will display afs: byte-range lock/unlock ignored; make sure no one else else is running this program. There are a couple of platform-specific exceptions to this behavior. Currently (as of 1.6.2), Linux clients will enforce byte-range file locks for processes on the local client only. In addition, Windows clients attempt to simulate byte-range locks for local processes, and acquire full-file locks on the fileserver when a byte-range lock is requested. Server-side byte range locking requires a protocol change, which is currently under consideration. **Whole-file locking:** [ Programmer ] AFS supports advisory locking of an entire file with `flock()`. Processes on the same client workstation that attempt to lock a file obey the proper locking semantics. Processes on different AFS clients requesting a lock on the same file would get `EWOULDBLOCK` returned. **Character and block special files:** [ [[SysAdmin]] ] AFS does not support character and block special files. The `mknod` command does not create either character or block special files under `/afs`. ### 2.02 What is an AFS protection group? A named list of users. Group names are used in AFS ACLs to identify lists of users with particular access permissions. In AFS, users can create and maintain their own protection groups. This is different to unix where only the system administrator can manage `/etc/group`. AFS groups are stored in the protection database on fileserver(s) and managed by using the `pts` command. An AFS group typically has the format `owner-id:group-name`. By default, only the owner of a group can change its members. It is possible to have both users and IP addresses as members of an AFS group. By using an IP address like this you can specify all the users from the host with that IP address. Note that IP address membership is insecure, due to the possibility of packet spoofing and the inability of current AFS protocols to protect server communications that do not involve a user-based security token; the `rxgk` security protocol currently under development will enable token-protected access at the client machine level as well as the user level. ### 2.03 What are the AFS-defined protection groups? - `system:anyuser` - Everyone who has access to an AFS client in any cell that is on the same network as your cell. - `system:authuser` - Everyone who has access to an AFS client in any cell that is on the same network as your cell **and** has valid tokens for your cell (i.e. has been authenticated in your cell). - `system:administrators` - Users who have privileges to execute some but not all system administrator commands. ### 2.04 What is an AFS access control list (ACL)? There is an ACL for every directory in AFS. The ACL specifies protection at the directory level (not file level) by listing permissions of users and/or groups to a directory. There is a maximum of 20 entries on an ACL. For example: An AFS ACL is displayed by using the `fs` command as shown below: tweety@toontown $ fs listacl . Access list for . is Normal rights: fac:coords rlidwka system:anyuser rl This ACL shows that members of the AFS protection group `fac:coords` have full access rights to the current directory and `system:anyuser` has only read and lookup rights. The members of `fac:coords` can be determined by accessing the protection group database using the `pts` command as shown below: tweety@toontown $ pts membership fac:coords Members of fac:coords (id: -1577) are: sylvester roadrunner yosemite.sam ### 2.05 What are the AFS access rights? In AFS, there are seven access rights that may be set or not set:
lookup l Permission to examine the ACL and traverse the directory (needed with most other access rights); permission to look up filenames in a directory
read r View the contents of files in the directory
insert i Add new files or sub-directories
write w Modify file contents, use `chmod`
delete d Remove file(s) in the directory
lock k Permission for programs to `flock()` files in the directory
administer a Ability to change the ACL
There are shorthand forms for some common permission combinations:
read rl read and lookup
write rlidwk all rights except administer
all rlidwka all rights
none removes all rights
### 2.06 What is `pagsh`? A command to get a new shell with a process authentication group (PAG). This is normally used if your system does not get AFS tokens on login. It is used to get a PAG prior to running `klog`/`aklog`. The PAG uniquely identifies the user to the Cache Manager. Without a PAG, the Cache Manager uses the Unix UID to identify a user and tokens will be shared across all processes owned by that UID. ### 2.07 Why use a PAG? There are two reasons: 1. Child processes inherit the PAG and the AFS token so they are AFS authenticated. 1. For security: if you don't have a PAG, then the Cache Manager identifies you by Unix UID. Another user with `root` access to the client could `su` to you and thereby use your token. ### 2.08 How can I tell if I have a PAG? Usually you can tell if you have a PAG by typing `id`. (Platforms which are derived directly from AT&T System V Release 4, such as Solaris, will not show the additional group vector by default; if there is no `groups=` section in the output of `id`, try `id -a`.) A PAG is indicated by the appearance of one or two large integers in the list of groups. For example: sylvester@toontown $ id uid=1000(sylvester) gid=20(staff) groups=33536,32533,20(staff),30(catz) On Linux clients, your PAG may not show up as such a group in the group list. An alternative way to check on Linux is to look at your kernel keyring with `keyctl show`: $ keyctl show Session Keyring -3 --alswrv 1000 1000 keyring: _ses.32603 819041549 ----s--v 0 0 \_ afs_pag: _pag If you see an `afs_pag` key in the output, then you are in a PAG. ### 2.09 Can I still run `cron` jobs with AFS? Yes, but remember that in order to fully access files in AFS you have to be AFS authenticated. If your `cron` job doesn't `aklog` then it only gets `system:anyuser` access. The preferred way to make use of Kerberos tickets and AFS tokens from a `cron` job is [[kstart|http://www.eyrie.org/~eagle/software/kstart/]]. If you are using the Heimdal implementation of Kerberos, you can also use its `kinit` with the `-afslog` option and a command. Note that you can still run a `cron` job without getting a token, if the task does not need to be AFS authenticated. In this case, you may get `stderr` from the `cron` job if your `.profile` is not accessible because of the ACL protecting your `$HOME`. Simply redirect to `/dev/null`: 0 7 * * * $sys_anyuser_readable_dir/7AMdaily 2>/dev/null ### 2.10 How much disk space does a 1-byte file occupy in AFS? This varies depending on the filesystem used by the fileserver containing that file. Some filesystems may only use up 1024 bytes for such a file, and others may use 4096; still others may use more or fewer bytes. ### 2.11 Is it possible to specify a user who is external to the current AFS cell on an ACL? Yes. This requires setting up a cross-realm relationship with the Kerberos realm on the remote site, but this is possible. Typically you refer to "remote" users like `user@remote.cell`, and you can use them in ACLs, or add them to `pts` groups. ### 2.12 Are there any problems printing files in `/afs`? There are two common issues that come up with printing from AFS: - the print client passes pathnames to the print service, which does not have access to your AFS token; - if not using a PAG (UID-based tokens), `setuid` print clients don't have access to your AFS token. Both of these may be mitigated by using shell redirection to send the file to the print client: lpr < /afs/ny.toontown.com/home/elmer/private/wabbit-hunting (The `lp` command generally supports the `-c` option to force a file copy. Note that the `lp` command provided with [[CUPS|http://cups.org]] *always* works this way, and makes a connection to the server to transfer the file instead of copying it to a secure queue area directly, so should not have token issues at all.) Very few print services have the ability to manage token access in a way that allows an ACL-protected file to be printed by pathname without enabling access control to be subverted by users by means of the print service. It is **not** recommended to grant the print service general access to AFS by means of [[kstart|http://www.eyrie.org/software/kstart/]] or similar mechanisms. ### 2.13 Can I create a FIFO (a/k/a named pipe) in `/afs`? No. AFS does not support `mknod fifofile p`. AFS only supports normal files, directories, and symlinks; not Unix- or Windows-specific filesystem node types. ### 2.14 If an AFS server crashes, do I have to reboot my AFS client? No. Typically, if an AFS server becomes unavailable, the AFS Cache Manager on your AFS client will see you through the outage until the server returns. This robustness is dependent on the way your AFS cell has been configured including the following factors: - On the client side: - How big is the cache? - Are the files you need already in the cache? - On the server side: - How many servers? It's best to have a minimum of three. - Is the data you are accessing replicated? In AFS, replicas are [[ReadOnly]] copies. With replicated volumes, the AFS Cache Manager knows about all of the servers on which the replicas are located. Therefore, when the Cache Manager accesses a replicated volume, if the RPC times out, the Cache Manager automatically retrys the RPC, using a different file server. If necessary, the Cache Manager will attempt to contact all file servers on which a replica of the volume resides. If you are accessing [[ReadWrite]] volumes on a crashed server then you will not be able to save changes back to the server until it returns. You don't need to reboot, and the Cache Manager activity is "invisible" to the user. You may want to speed up recovery by issuing the command `fs checkservers`, but even this is unnecessary and will usually only improve recovery by a few seconds. ### 2.15 Can I use AFS on my diskless workstation? Yes. The AFS Cache Manager can be configured to work with either a disk based cache or a memory (RAM) based cache. Note that using a memory cache may not be as fast as you might think. Modern operating systems should cache disk data in memory when accessed, so using a disk cache should mean you are hitting RAM most of the time anyway. Disk caches are also much more common, and thus much more heavily tested, and so are better optimized. If you have a local disk and it's reasonably fast, usually going with a disk cache is preferable to memory cache. Additionally, most operating systems do a better job of optimizing a native RAM disk, so you might consider putting your AFS cache in a RAM disk instead of using [[OpenAFS]]'s own memory cache. ### 2.16 Can I test for AFS tokens from within my program? Yes. However, the mechanism for doing so varies depending on the platform. To see examples of how to do this, you can look at the source of any program that deals with AFS tokens. One such example is [pam-afs-session](http://www.eyrie.org/~eagle/software/pam-afs-session/) ### 2.17 What's the difference between `/afs/cellname` and `/afs/.cellname`? AFS has [[ReadOnly]] (RO) and [[ReadWrite]] (RW) volumes. The convention in AFS is to mount the RW volume `root.cell` as `/afs/.cellname` and the RO volume `root.cell.readonly` as `/afs/cellname`. This is so that when you travel down the `/afs/.cellname` link, AFS will always use the RW site of any volumes that have RO clones. This allows your administrator to update the RW copy of a volume and `vos release $volname` so that it will appear in `/afs/cellname`. ### 2.18 Can I `aklog` as two users on a machine in the same cell? Yes, *if* you use two different PAGs. The token store only supports one token per cell per authentication group; with UID-based PAGs, this means one token per cell per user, but with PAGs you can have multiple shell windows/sessions, each with its own PAG and associated AFS tokens. Note that most Kerberos implementations (the one on Mac OS X 10.7 and 10.8 being a notable exception) only allow a ticket-granting ticket (TGT) for a single principal (this restriction is even tighter than AFS's, as you cannot have a separate TGT per realm!), so if you plan to use separate PAGs you will also want to use separate credential caches ("ticket files" in older Kerberos parlance). An alternative to using multiple users in this way is to use ACLs to grant access on a shared directory to both users. ### 2.19 What are the `~/.__afsXXXX` files? They are temporary reference files used by the AFS Cache Manager. With most Unix filesystems, when you a remove a file that is kept open by a process, the file stays around physically while it is no longer referenced in any directory (which you will see as a mismatch between disk space usage according to `df` and `du`). Some applications rely on that feature, e.g. they create a temporary file and remove it immediatley while keeping the file descriptor open. The file then disappears from the filesystem automatically when the process terminates or the file descriptor gets closed otherwise. Such applications could get into trouble with older versions of AFS, where the file could really disappear while it was held open. Newer versions of AFS rename such files to `.__afsXXXX`, thus making sure that the data stays around as expected by the application. As soon as the file gets closed, the associated `.__afsXXXX` should disappear. ### 2.20 How do you set up IP-based ACLs? See [[IPAccessControl]]. ### 2.21 What meaning do the owner, group, and mode bits have in AFS? In order to appear more like a local filesystem, AFS will faithfully store the numeric UID (owner), GID (group), for both files and directories, as well as the permission bits (read, write, and execute for user, group, and other, plus `setuid`, `setgid`, and "sticky" bits) for files. Note that permission bits for directories are not stored. For the most part, these values are simply recorded and reported back when requested. However, in some instances the fileserver and/or cache manager will make access control decisions based in part on these values. The following is believed to be a complete list of those circumstances. Below, "owner" refers to the user whose numeric `pts` identity is equal to the owner of the file or directory; this might not bear any relationship to the UNIX UID associated with the client process that created the file. - implicit ACLs - the owner of the root directory of a volume has implicit _administer_ (`a`) rights on all directories in the volume - the owner of a file has implicit _read_ (`r`) and _write_ (`w`) rights on a file if that user has _insert_ (`i`) rights on its parent directory - to **read** from a file, you must have _read_ (`r`) rights _and_ at least one of the following must be true: - the file's `u+r` (user read, `0400` octal) bit is set - you are the owner of the file - you are a member of `system:administrators` - to **write** to a file, you must have _write_ (`w`) rights _and_ at least one of the following must be true: - the file's `u+w` (user write, `0200` octal) bit is set - you are the owner of the file - you are a member of `system:administrators` - changing mode bits and owner/group: - the fileserver will only allow the mode bits on a file (`ugo+rwx`) to be changed if the user has _write_ (`w`) and _lookup_ (`l`) rights on the file's parent directory - the fileserver will only allow the mode bits on a directory to be changed if the user has _delete_ (`d`), _insert_ (`i`), and _lookup_ (`l`) rights on the directory - only members of `system:administrators` can change the owner or group of a file - only members of `system:administrators` can change the `setuid` and `setgid` bits on a file The "sticky" bit, group of a file, `g+rwx` (octal `0070`), and `o+rwx` (octal `0007`) bits are completely ignored by all AFS components. Additionally, the `u+rwx` (octal `0700`) bits are ignored on directories. Newly created files and directories are given an owner numerically equal to the `pts` identity of the user who created the file or directory. Initial mode bits are assigned by the AFS cilent, typically based on the creating user's `umask`. ### 2.22 What are "dropboxes"? When the ACL on a directory is set to `irl` (_read_, _list_, _insert_), this creates what is called a "dropbox". In theory, users should be able to deposit files in the directory, but not modify them once deposited. In practice, the "not modify them once deposited" part is not enforced by the fileserver; only the [[OpenAFS]] client enforces this restriction. Thus, you should not depend on this for security. Also, note that a `system:anyuser irl` ACL has an additional problems: because dropbox semantics are based on `pts` identities (see question 2.21), the fileserver cannot distinguish between two unauthenticated users. So, not only can a user come back days later and modify the "dropped" file, but **any** user can modify a file dropped by an unauthenticated user, at any time. ### 2.23 Can I access a RW volume using the RO path? Depends. Once you have RO-Volumes released, a mountpoint pointing to the RO will bring you to the RO volume. To change that behavior, you have to change the corresponding mountpoint with `fs rmmount` and `fs mkmount -rw`. However, for some situations, like software installations, it might be useful to reach the RW volume through the RO path. You can do that for a single client with a special setup. The trick is to break the convention described in 2.17 for a single client: mount the RW volume `root.cell` (instead of `root.cell.readonly`) as `/afs/cellname`. This can be done by creating an alternative `root.afsrw` volume which is identical to `root.afs` except that it has an RW mount for `root.cell`, then add `-rootvol root.afsrw` to the `afsd` command options on startup (either in `/etc/init.d/afs` or wherever your system stores service configuation; this is often `/usr/vice/etc/config/afsd.options` on most Unixes and `/etc/sysconfig/afs` on many Linux distributions) and ensure that the `-dynroot` option is *not* specified.