Linux admin notes - users and permissions
Linux-related notes
Shell, admin, and both:
|
...particularly in the context of filesystems
Filesystem permission basics
Most people know the below to some degree, and should.
'nix file permissions are a simpler system ACLs.
Which sometimes makes them easier to deal with, and sometimes requires serious contortion - which is one reason ACLs slowly being introduced into unices here and there.
Each filesystem entry (files, directories, other) has a set of properties. The most interesting are:
- the user id (ususally shown as the name, as it is looked up on the current computer)
- the group id (ususally shown as the name, as it is looked up on the current computer)
- read, write, execute permissions bits for the owning user
- read, write, execute permissions bits for the owning group
- read, write, execute permissions bits for other - things that apply for everyone.
You can also get and set these numerically, as explained further down.
Permission bits are shown in various ways. The form form is the output used by ls -l. Its output for a regular file could look like:
-rw-r--r-- 1 sarah users 2414 Nov 21 2004 afile.txt
The first bit shows permission bits (and more, explained later). You should read this as a special flag/filetype field, followed by three group of rwx, one for each of user, group, and other. For example:
- rw- r-- r--
The permission is present if the letter is present. In this case, sarah can read and write, anyone in the users group can read, and anyone else can also read this file.
'Other' is used for minimal permissions, since it always applies (regardless of what user or group id a visiting user comes along with), so -rw----r-- has a very similar effect in to -rw-r--r--.
Details explained later:
- setgid and setuid are also be indicated in those letters
- rights work slightly differently on directories
Notes: Seeing a plus sign, e.g.
-rw-r--r--+ 1 sarah users 2414 Nov 21 2004 afile.txt
...means that there are ACLs on this listing. See them with things like getfacl
Changing permissions
When changing permissions on a file there are various ways of describing permissions. One common distinction is whether you specify what should change or whether you specify exactly what the new permission set would be.
There are also different ways of describing each permission, and different operations to the permission bits. Choose the one or two you like.
Common usage:
chmod o-rwx afile.txt # takes all permissions away from other
chmod +x afile.txt # gives everyone execute permission
# (short for a+x or the equivalent ugo+x)
chmod ug+w -R dataDir # Gives user and group write permission
It is technically also possible to combine operations:
chmod a-rwx,u+wr,g+r -R dataDir # A much-at-once style (not seen used much, but useful)
chmod u=rwx,go=rx dataDir # Set-to-exactly-this style, in this case rwxr-xr-x
The numeric way sets permission bits as octal numbers, and is the shortest way to set all bits in one go. Where r=4, w=2, x=1:
chmod 0777 somefile # gives full permission to all, like a=rwx.
chmod 0620 somefile # sets rw- r-- ---
chmod 0750 somefile # sets rwx r-x ---
...and any other combination. Most people remember just a few cases, such as 0777 as 'allow all', 600 (owner readwrite) or 644 (owner readwrite, others read-only) for files, 755 or 700 for directories, or such.
Changing ownership
You can change the user, group and each of these permissions, with chown and chmod respectively:
chown frank:goats afile.txt # sets user to frank, group to goats
chown frank afile.txt # sets user to frank, leave group unchanged
chown frank: afile.txt # sets user to frank, and group to frank's login group
chown :goats afile.txt # sets group to goats
chown -R frank dataDir #recursive: changes entire subtree (files, directories)
For more details on groups, see #Groups.
chown: changing ownership of `/path': Operation not permitted
Only the file/directorory's owning user (and root) can alter user or group membership.
You may expect that the file/dir's owning group can do so too, if it has write access to the file/dir. It can't. The reason seems to be practical one - it makes it easy to give away files that aren't yours, and/or that you can't get back. (is there some reason based in quota systems?(verify))
If you are the owning user (or root) and still can't, it's probably that the immutable attribute is set. See lsattr and chattr.
It may be that SELinux is prohibiting you (even if you're root).
You may be using a non-*nix filesystem, one which does not store file permissions (e.g. FAT).
Filesystem guts you can usually avoid
mode
Programmers may find it interesting that the permission bits mentioned are a subset of the mode.
Mode is a larger integer (now often 32-bit), which stores:
- permission bits (lowest 9 bits)
- setgid (S_ISGID, usually 2000). In text representation it's in the group-execute field, an s if the group execute bit is set, or S if the execute bit is missing
- setuid (S_ISUID, usually 4000). In text representation, see setgid, but for the user execute field/bit.
- sticky (S_ISVTX, usually 1000). In text presentation, it's in the other-execute field. It's t if the execute bit is present, T if it is missing.
- file type (see below, and these are read-only)
Most command line tools only show (and allow change of) permission, sticky, setuid, setgid,
The mode may also store OS-specific bits, which you can only get at via OS-specific libraries.
File types:
what | constant (octal; may differ(verify)) | in text representation of mode |
---|---|---|
regular file | S_IFREG, 0100000 | - |
directory | S_IFDIR, 0040000 | d |
named pipe, fifo | S_IFIFO, 0010000 | p |
character special device | S_IFCHR, 0020000 | c |
block special device | S_IFBLK, 0060000 | b |
symbolic link | S_IFLNK, 0120000 | l |
socket | S_IFSOCK, 0140000 | s |
You usually only need to care about regular files, directories, and the occasional symlink.
You can get the full mode with a stat(), which is what library functions that test 'is this a directory' use, but even the command-line stat command masks out everything but the permissions bits unless you specifically ask for all of the mode bits.
It also shows it in hex for some reason, so to view the full mode in octal, you'ld have to do something like the following:
$ stat --format %f /tmp | tr [a-f] [A-F] | xargs --replace=@ echo "ibase=16; obase=8; @" | bc
41777
Extended file attributes
directory permissions, read and execute
tl;dr:
- In general, r and x are only useful in combination.
In more detail:
- directory read rights
- allows listing the directory's entry names
- directory write rights allow: create, rename, remove, and alter entries in it (including changing permissions)
- (...so o+w is scary)
- directory execute rights
- allow cd into the directory
- allow looking up (basically stat()ing) entry's details within this directory (verify)
If that seems weird, keep in mind that directories are just another entry within their parent directory.
They happen to be one specific kind due to being marked as a directory in its entry's mode)
For the examples below
mkdir -p a/b c/d e/f
chmod 0111 a
chmod 0444 c
chmod 0000 e
To have:
d--x--x--x 3 user user 4096 Mar 22 18:30 a dr--r--r-- 3 user user 4096 Mar 22 18:30 c d--------- 3 user user 4096 Mar 22 18:30 e
(b, d, f have typical permissions, according to your umask. You will typically have rwx. These contents don't need to be directories, but it can help your mental model of how this works in trees a bit)
x without r
You can go there (because x), but you can't listdir() (because no r)
~$ ls a
ls: cannot open directory a: Permission denied
~$ cd a
~a/$ ls
ls: cannot open directory .: Permission denied
Note that if you can guess entry names you can open them fine. If you can guess their name and they are directories (with x), change to them fine.
To continue the above example:
~a/$ ls b # says nothing only because our example gave b no contents
~a/$ stat b
# gives full output, omitted here
~a/$ cd b
~a/b$ ls # says nothing because no contents
r without x
You can list entry names (because r), but not cd to it or stat its contents (because no x)
~$ ls c
ls: cannot access c/d: Permission denied
d
When you can't stat entries in a directory (due to -x on its parent), you also can't do descending into them (because you can't even do the permission check)
~$ cd c
cd: c: Permission denied
~$ cd c/d
cd: c/d: Permission denied
neither r or x
Can't list its entries, can't go there or under it, for the reasons explained above.
in a tree
Mount points and permissions
directory permissions, write
o+w permission is scary since it means that anyone can alter directory entries, even if those entries don't list the user as their owner.
The same applies to g+w if the group is something like 'users' or 'students' and that includes people you may not want to trust implicitly.
Sometimes you may want a directory which can be used to give each other files, or a generally shared directory.
directory permissions, sticky
Without sticky, you can share a directory only if you are in the same group, or allow write permissions to world.
When a directory has the sticky bit set, then this parent directory's owner and each directory entry's actual owner can rename and remove entries (anyone can create files/directories).
With sticky, people can work in the same directory, but can shield each other from accidentally (or purposefully) messing with files and directories they don't own.
/tmp is often a stickied directory for this reason.
(The sticky bit was named for its historical use on regular files, where it was a signal to the OS to keep this around in memory (or in local rather than remote storage), as it would probably be used regularly. Modern caching and drive speed has made this mostly obsolete. (I'm not sure whether modern kernels still listen to the sticky bit(verify))
setuid, setgid, and sticky on files
tl;dr:
- the SUID and SGID bits in a file's mode will trigger seteuid() / setegid() syscalls at launch time, with the file's current user/group
- "SUID root" is the common case of that: chown as root, set suid bit, to have root permissions without need for sudo
- can only be set by root / through sudo (for obvious security reasons)
- often for daemons. Many of which will, after basic startup, use the same syscalls to give away root permissions again (switch to a normal user)
- (the point may e.g. be binding to a port under 1024, which normal users can't do)
These bits are stored in the mode, on top of basic rwx permissions, are:
- setuid (SUID), shown in the user's execute position, 4000 in octal, add with chmod u+s
- setgid (SGID), shown in the group's execute position, 2000 in octal, add with chmod g+s
- sticky bit, shown in other's execute position, 1000 in octal, add with chmod o+t
In ls -l they show up instead of the execute field field, e.g.
-rwsrwsr-t 1 root root 1191834 Nov 23 2013 thingie
(When they have these bits without execute, the S and T are shown in uppercase)
For the system side of that, see #The_various_UIDs.2F_GIDs.2C_related_syscalls.2C_and_SUID.2FGUID_on_execution
umask
The umask removes permission bits from files and directories (programs running as) you create, from mkdir to whatever fanciness you are running.
You can always change them later - the point here is what protection you get by default.
Example: If the umask is 0022:
- 0022 itself means --- -w- -w-
- ...which are the things that are removed from permissions
- in this case meaning that on newly created files, write permissions will never appear for group or other. Read and execute is not masked away, so other people can go into your directories and read files.
When people work in groups -- (but you still want at least processes to have a clear owner, so sharing a single account is not so useful), you might want to not mask anything from the group so that you can work with each others' data by default. Example: 0002 (or so),
If your group is something like 'students' or 'staff', you probably don't want group writability. You might use 0027 (no write on group, nothing on other), or perhaps 0077 if paranoid.
You can make this persistent by adding something like
umask 0002
to your ~/.bashrc (or whatever applies for your shell).
POSIX ACLs
See also:
Changing account's UID or GID
Accounts
By default, account information is backed by /etc/passwd (user list) and /etc/group (group list), and usually also the shadow files.
Note that filesystems store UIDs and GIDs, not usernames and group names, meaning filesystem permissions are inherently local to a computer unless those IDs are centralized / synchronized. This also matters to network filesystems such as NFS
useradd, usermod
useradd adds a user
adduser
- is a wrapper in Debian and derivatives (e.g. Ubuntu)
- which adds configurable site-specific configuration, making some things easier.
- as a sysadmin, you may wish to use only adduser to avoid confusion.
- (though I can never remember which name is which, but man is your friend)
usermod changes details for an existing user
Most people user them so rarely that we just look at the man page.
On homedir:
Notes:
- useradd isn't always configured to create the user's home directory.
- you'll want useradd -m (as it's more convenient than manually copying from /etc/skel/ and fiddling with the permissions)
- (Similarly, whether userdel deletes a home directory, mail spool and such depends on configuration)
- not having a home directory means the account cannot log in (as does not having a valid shell).
- Not having a password set usually also disallows logins, at least remotely(verify).
- (This can be useful to e.g. have users for samba to map to, but to not allow them to shell into the same server)
On groups:
Often, default is to create a group of the same name, and add the user to it.
If you want users to go to a shared group instead, you can control this via arguments and, often, configuration files (see the man page for your system to be sure) like:
/etc/login.defs /etc/defaults/useradd
Changing passwords
Use passwd
On systems that integrate with networked systems like NIS, you may need to use another executable (like yppasswd). A plain passwd may tell you so - or may be symlinked to what you need to use.
To not allow password login, use passwd -l.
This prepends something (!) at the start of the hash value, meaning no password will ever work. It also means you can unlock (passwd -u) by removing that character again (avoids having to set/remember the hash).
Keep in mind that -l applies only to this type of password login. SSH logins may have their own rules, and may effectively bypass it depending on how it is set up, and on how a user is trying to authenticate.
You may be able to set the hash of a blank password - as in 'accept pressing enter for a password'.
While passwd probably won't allow it, you can generate the appropriate hash value via mkpasswd.
This is generally a bad idea, though. If you want passwordless remote login, look at using a passphraseless keypair.
Deleting (passwd -d) seems to (usually? always?) mean a blank password (verify) - if so, you rarely want this.
Keep in mind that anyone with root (/general sudo) rights can su to any user, regardless of whether they are allowed remote or local login.
Groups
Theory: group logic
Some important distinctions:
- your primary group, a.k.a. initial login group
- part of your login info
- your additional group memberships
- part of your login info, can be queried, but not everything does (verify)
- a process's current effective group
- your shell also has a currently effective group - which likely is your primary group, unless it was explicitly forced to be something different.
The easiest command line tool to see a user's currently effective GID(verify), and collection of all group memberships is probably id.
For regular users it may be as simple as:
uid=1004(sarah) gid=100(users) groups=100(users)
Root gets around more, and may look like:
uid=0(root) gid=0(root) groups=0(root),1(bin),2(daemon),3(sys),4(adm), 6(disk),10(wheel),11(floppy),18(audio),20(dialout),26(tape),27(video)
Because of the history of how processes work, the mental model and the code is slightly hairy.
Lazy and older programs may test only for the current effective GID and UID. A process can checks for additional memberships, but they have to care, and have to know how to do it.
To change one's own currently effective group (in the shell), use newgrp groupname.
- If you are a member of that group. it will switch without question.
- If you are not a member, it will either
- deny you or
- ask for the group password, ...depending on how the group admin has set up the group.
There is a sg [1],
analogous to su but for groups:
it sets a different group for executing a particular command.
Practice: Group administration
Mostly:
- groupadd groupname - create group
- groupdel groupname - remove group
- gpasswd and usermod for membership (see below)
Additional group membership is probably easiest via:
- gpasswd -a user group also appends to additional groups (verify) (but note argument order)
- gpasswd -d user group removes a user from a group
Further group membership stuff:
- usermod -g group user set primary group -- and remove all other group membership(verify)
- usermod -G group user set additional groups to exact list
- usermod -a -G group user append to additional groups
- note -a is a relatively recent feature. If you don't have it, you may need to use -G with a full list
Other
gpasswd is largely for the fancier group stuff, like:
- gpasswd group to set the group password
- gpasswd -r group to remove the password
- gpasswd -R group disables the ability for non-members to use password login
- These can be done by a group admin and root.
- Group admins are one or more (non-root) users who are allowed to change group membership.
- Group admins are added to groups by root, using gpasswd -A user group.)}}
Notes:
- id will not show the change within the same login session, because group membership is queried only at login time and fixed for that session. Running a new shell (even -l(verify)) won't do it.
- for a quick-ish check:
- id -Gn $USER (just happens to need to query /etc/group)
- or get a shell with new login sesstion via su - $USER or ssh localhost
- for a proper change, log out and back in
- (graphical) session manages may reuse session if you log in within ~30 seconds or so. To be sure, wait that long.
Duplicate UIDs
Change UID
/etc/passwd, /etc/shadow, /etc/groups
These are plain-text, colon-delimited databases containing all information about users, passwords, and group membership.
Password shadowing describes system that separate out publicly fetchable information, and store the actual password hash in a separate file.
It's generally best to modify these with system utils like
useradd, usermod, userdel,
groupadd, groupmod, groupdel,
chpasswd (changing password),
gpasswd (changing group membership),
chfn (change finger/CEGOS field),
chsh (change login shell),
...although there are a number of situations where you can get away with editing them directly.
There is also vigr and vipw.
/etc/passwd is a colon-delimited file that stores basic user information:
- username (up to 32 chars?(verify))
- password - in the old days this stored the password hash.
- classically stored a password hash, but these days it's typical to use shadow files - which means
- /etc/passwd is publicly readable, because the rest of the fields are public and useful.
- /etc/shadow contains the password hash, and is only readable by the system
- putting x in the password field in /etc/passwd to indicate the real password hash is stored in /etc/shadow
- other values in /etc/passwd's shadow field, like * and !, seem to effectively mean "not possible to log in", where apparently * suggests it was created as a service account, and starting with an exclamation suggests that it has been explicitly locked. This seems to be mirrored in /etc/shadow (verify)
- UID - user's numeric ID
- GID - primary group ID, refers to entry in /etc/group
- Sometimes called comment, sometimes called GECOS. (Used by finger, if you use it) May contain comma-separated sub-fields including the following (but there are deviations from this)
- Real name
- physical address (building, room)
- phone number
- other contact info - pager, fax, email address
- home directory for user
- login shell to be invoked
/etc/shadow stores the user's password hashes, and some further password-related information
- username
- password hash
- starting with $ indicates use of a non-default hash method, and typically signifying which hash function. See also significators, but e.g. $1$ is MD5, $2a$ is Blowfish, $3$ is NT hash $5$ is SHA-256, $6$ is SHA-512. (there may be another $ in the value, splitting a salt from the hash)
- date of last change (days since Jan 1, 1970)
- minimum days before a next password change is allowed (0, i.e. no limitations, seems common)
- days after which a password is expired, and change is forced on the next login (99999 basically means "won't expire")
- amount of days before the previous entry that you get warnings (7 seems common)
- number of days after expiry after which the account is actually disabled (rather than just forces password change)
- absolute date of expiry (days since Jan 1, 1970)
- reserved field
Notes:
- 99999 days after 1970 is somewhere in October of 2243.
- Dates in ~2021 are around 18800
- passwd to change the password
- chage ('change user password expiry information') to change most other things, or list the details in a more friendly way.
/etc/groups
- group name
- GID
- password (if any) - though if you use password shadowing, it will be in /etc/gshadow
- comma-separated list of group members (is this an exact mirror of /etc/gshadow?(verify))
/etc/gshadow
- group name
- password
- comma-separated list of group admins
- comma-separated list of group members (is this an exact mirror of /etc/groups?(verify))
Some details
For another sort of example, svgalib needs root access to access graphics hardware directly, which would mean anything that needs svgalib is only useful when root runs it, or you're allowed to run it via sudo.
With SUID and/or SGID, once a program starts executing, it reads the user and/or group ID from the file's directory entry, and hands those to setuid and/or setgid, and the related seteuid and setegid.
A process can always call those syscalls. The file permission flags are just a handy way to set this up without changing the executable's code, and that you can optionally protect.
Since you can't really give away ownership (e.g. set a file to be owned by root, unless you are root), this isn't a security problem. However, you and particularly root should be careful of what they take ownership of; whatever files are owned by root, generally accessible, o+x and setuid or setgid will effectively run as root when executed by anyone, and therefore be able to do anything. There are a few fairly simple ways around this, such as to disallow directory access, which can be done by moving files you take ownership of to /root, when it has something like 0700 permissions - like it usually does.
UID and EUID
A process on most unices OSes manages three user and three group IDs:
- real UID (ruid), real GID (rgid)
- who started the process
- also relates to signal permissions: A non-superuser process can only signals if ruid or euid (or rgid or egid) matches. Note that this implies children can be default, because inheritance
- effective UID (euid), effective GID (egid)
- Most things check this one checking a process is allowed to do a thing.
- euid is assigned to files created by this process
- egid semantics often the same (though there are/were some variations with kernels)
- saved UID, more fully "saved set user ID" (suid; note potential confusion with SUID), saved GID (sgid)
- a placeholder for a UID so that it can be restored later
- e.g. allows a process to drop existing superuser rights temporarily, rather than permanently (and to avoid having to ask the system for permission when setting that later. Basically this encourages processes to drop them while not necessary. This isn't so much a security thing as a procetion against common mistakes)
- more precisely, an unprivileged process can set its euid to: its ruid, its suid (and technically its present euid)
- privileged processes - ?(verify)
- seems this is only set implicitly -- when switching from euid 0 to non-0 euid (verify) (the semantics of other parts take care of the rest)
- Under typical circumstances
When you don't need privileges (the bulk of cases):
- your process's ruid and euid are the same
- your process's suid is irrelevant
Which securitywise is sensible default behaviour, but doesn't serve all cases.
- altering these manually
There are a handful of syscalls relating to real/effective/saved, and their semantics - what they set, and in what cases they would be refused. Note that some of these details vary between OSes.
Most of them:
- geteuid()
- getuid()
- seteuid(uid_t euid)
- setuid(uid_t ruid)
Also:
- setreuid(uid_t ruid, uid_t euid) - can set both real and effective UID at once, or just one
There is analogous logic for the real and effective group (TODO: more explanation).
The relevant system calls for it:
- getegid()
- getgid()
- setegid(gid_t egid)
- setgid(gid_t rgid)
Also:
- setregid(gid_t rgid, gid_t egid) - can set both at once
Unprivileged processes may only set euid to ruid, sgid, or (current) euid. (and analogous for groups)
Privileged processes can set any (though there are cases that are considered errors). Note that this is not a permanent privilege drop, because it can be set back to rgid later. (verify)
- on privilege drops
permanent privilege drops are a little harder, in part because euid and suid are designed to allow temporary
privilege drops. Basically, you must make sure that none of ruid, euid and suid are privileged.
- Having effective IDs set via the filesystem
Alongside all that are some relevant filesystem entry's mode bits:
- SUID
- kernel reads file's UID and does a seteuid() (verify) when doing an execve()
- SGID
- kernel reads file's GID and does a setegid() (verify) when doing an execve()
- sticky (no longer relevant(verify))
A filesystem entry with these bits will often have UID (and maybe GID) 0, for root privileges.
Some system commands need this, including
- ping (for raw socket access)
- passwd (because the password files are root-owned)
- su
- mount and umount
...and more. For kicks, take a look at the output of:
find /usr/*bin /*bin -perm /6000 -type f -ls
Note that SUID root regularly sorta violates least privilege, particularly when we only needed it for very limited reasons (see e.g ping), which is roughly why capabilities became a thing.
Note also that this mechanism means user permissions system has now extended to involve filesystem state. (For root this usually isn't so bad, since even under DAC only root can alter things owned by root. ...or people with blanket sudo rights, which is why you want to be careful with that.)
- related processes
A fork() will inherit these three sets from the source process
An execve() will inherit the three sets - but SUID/SGID will override when they apply. (verify)
- fsuid
Linux also adds fsuid, which is a separate identity that is used when checking file access permissions. Unless setfsuid is used,
This is not used much.
See also http://www.cs.umd.edu/~jkatz/TEACHING/comp_sec_F04/downloads/setuid.pdf
See also:
- https://people.eecs.berkeley.edu/~daw/papers/setuid-usenix02.pdf (or here)
- https://yarchive.net/comp/setuid_mess.html
- https://cnitarot.github.io/courses/cs526_Spring_2015/s2014_526_os.pdf
- https://www.osso.nl/blog/setuid-seteuid-uid-euid/
GID and EGID
See previous section.
su, sudo, sudoers, sudo su, etc
- su can be remembered as 'switch user' (by default to root)
- starts a non-login shell as the target user
- when non-login breaks things (e.g. because no user profile), consider su -l
- for non-root users it's similar to login to that other user (requiring their password)
- root can switch to other users arbitrarily
- meaning you can also switch to accounts that do not have a password (so cannot be logged into normally), which is useful to administer accounts for services that preferably run as their own user - without giving them the ability to log in.
- su by default does not preserve the environment that it was called from (security reasons)
- sudo, 'switch user and do command' (by default as root)
- runs a single command as another user (e.g. does not start a new shell)
- sudo only works when sudoers (see below) gives your user the right (or you are root)
- which means you get asked for your password, not root's.
- it is possible to configure sudo to not ask for a password (for a specific command)
- By default leaves most of the environment (but not all(verify)) intact.
- who is allowed to do what, as who (and further features like audit logging) is configurable, see sudoers below
sudo su is the combination:
- "get root rights via sudo, then tell su to run a shell"
Not for security reasons (rights to run sudo su carry essentially all the risks for giving out root) but more for administrative reasons: you can selectively allow specific users to run a root shell, based on their own password (and based on configuring them via sudoers) rather tham sharing the root password.
And yes, sudo bash would be similar to sudo su but not indentical in details, like when bashrc gets sourced(verify). In fact, for a more predictable environment, consider the below over sudo su (or sudo bash):
sudo -i means 'simulate a login shell as root'
- Some argue this is preferable to sudo su as your environment will be more predictable (arguably only matters when root is an account that you've ever set up a profile for, not so much when you're just using it for the privileges).
sudo -s means 'give me an interactive shell as root'
- So yes, su - is similar to sudo -s
- not all variants/versions of sudo have -i (or -s(verify)
Also, consider (consider also ssh to localhost)
su examples
Examples:
su # switch to root su someuser # switch to user someuser # force login shell with -l, or - after the last option su -l # switch to root with login shell su - someuser # switch to user someuser with login shell
sudo examples
# typical examples are software installation sudo apt install boinc-client # non-root example: # it's easier to run psql (CLI) as the postgres user # (because pg_hba typically has the postgres (=an admin) role trusted to the postgres system user) sudo -u postgres psql
Notes:
- you may prefer to explain sudo rights as "sudo specific commands".
- it is a good habit to use sudo on all admin commands, to associate using sudo with serious business -- rather than geting an always-root shell (via sudo su or similar)
- sudo is often configured to allow shell users to do successive commands without giving a password each time (for up to a few minutes at a time). This trades a little security for a bunch of convenience when you're doing admin work.
Differences (Su versus regular login and ssh)
- su does not (by default) preserve the source environment.
- su does not start a login shell, other login and ssh usually do
- so among the largest difference is that su doesn't source the target user's ~/.profile
- PAM config may hook in different things to su and ssh
sudoers
Sudoers is the configuration file that sudo checks.
It allows you to configure things like
- everyone can run a specific command
Say you have a classroom full of workstations, in which you want users to be able to mount CDs (and you do not have an auto-mounter doing that for them. So, on both counts, ten to twenty years ago).
Since mounting a volume requires root permissions, you could e.g. allow users to execute only two precise commands as root, mount /cdrom and umount /cdrom (so that it allows it for this and for no other devices).
- a specific user can run a specific command
Example: I wanted a web-facing port-scan, which means my apache user (and no one else in particular) needed to run nmap (and nothing else) as root. Something like:
apache ALL=(root) NOPASSWD:/usr/bin/nmap
- specific people can run anything
Usually, anyone who is considered an administrator for a host is given either root access or sudo rights.
When you trust specific people on a host, it is still a good idea to do things via sudo, so that they are only ever intentionally dangerous, and have a visual reminder (you put sudo before a command).
(Yes, you can sudo su to get out of that. But it's good habit not to.)
How this is done depends a little on the exact setup. /etc/sudoers often has a line that allows members of a specific user group sudo rights for everything. This is often sudo or wheel. In sudoers that means
%sudo ALL=(ALL) ALL # or %wheel ALL=(ALL) ALL
...and it would mean adding the user to the group.
editing
To edit the sudo configuration, run visudo.
It mostly just invokes an editor on /etc/sudoers, but adds checking that you probably want for things to not break (some locking to avoid editing conflicts).
The default editor it uses is vi, but you can configure this in sudoers itself (the env_editor flag in the Default section), or tell visudo to not clear the environment (which it does for slightly paranoid security reasons) using sudo -E visudo).
Remember you can check the effects of the current sudoers file using sudo -l
Various types of rules
The basic form of these rules is:
who where=(aswho) options: /path/to/what, /path/to/what
- 'who' is who is allowed to do something, and can be
- username
- %systemgroup
- ALIAS
- ALL
- 'where' describes the hostname. For local rules this is usually usually usually 'ALL' or omitted, unless the sysadmin decided to write one sudoers file for use across many hosts.
- 'aswho' is the user that should be executed as. Usually root.
- options
- NOPASSWD is a common one, meaning you don't have to enter your password, useful for scripted and cronned stuff
- actually a choice between PASSWD and NOPASSWD, but PASSWD is default unless configured otherwise
- path to the executable
- note that since you usually use absolute paths in sudoers, you should also do so on the invocation, or the rule won't match
- essentially a prefix match, so you can force certain arguments
Examples:
#Everyone can run synergyc (with any options) as root, without having to type their password ALL ALL=(root) /usr/bin/synergyc #wilma can do anything as the database user (including starting a shell) # without needing a password: wilma ALL = (dbuser) NOPASSWD: ALL # Note: To use aliases in that field, use Runas_Alias # Anyone can mount and umount removable devices ALL NOPASSWD: /bin/mount /mnt/floppy, \ /bin/umount /mnt/floppy, \ /bin/mount /mnt/cdrom, \ /bin/umount /mnt/cdrom, #(are these globs or EBNF/regex style wildcards?) # Wildcards can be useful in various ways, for example: ALL NOPASSWD: /bin/u?mount /mnt/cd*, \ /bin/u?mount /mnt/dvd*, \ /bin/u?mount /mnt/flop* # john can su to everyone except root (negative match) john !/usr/bin/su *root* # PBIS users can mount some personal shares (should probably be more restricted) %domain^users ALL=(root) NOPASSWD:/bin/mount %domain^users ALL=(root) NOPASSWD:/bin/umount #Root can do everything everywhere, as anyone root ALL = (ALL) ALL #same for wheel, except it can't run lilo %wheel ALL = (ALL) ALL,!/sbin/lilo
Note that blacklist setups are, as in general, not very secure.
Also, note that it is hard to limit people from changing user.
Setting Defaults
Setting Aliases
You can add aliases to add your own groups of...
#...users: User_Alias DEVELOPERS = alice, blob User_Alias TEMPS = mike, wendy # You may prefer using only the system's groups, though.
# ...commands: Cmnd_Alias PRINT = /usr/sbin/lpc, /usr/bin/lprm Cmnd_Alias BOOT = /usr/sbin/shutdown, /usr/sbin/halt, /usr/sbin/fasthalt Cmnd_Alias SU = /usr/bin/su
# ...hosts and nets: Host_Alias WEBHOSTS = www1, www2 Host_Alias OURNET = 5.5.0.0/255.255.0.0
See also
- http://www.sudo.ws/sudo/man/sudoers.html
- http://gentoo-wiki.com/Sudo_config
- http://www.gratisoft.us/sudo/intro.html
- http://wiki.zimbra.com/index.php?title=Sudoers
- http://www.afp548.com/article.php?story=20051025103428232
Debugging sudoers
Sudo group
As there is usually a sudoers line like
%sudo ALL=(ALL:ALL) ALL
...any user in the group called sudo can sudo any command, and one short version of checking who's in that group would be
getent group sudo
It's common to be stricter and whitelist specific commands for specific users.
Relatedly when there are multiple matches for a given sudo command, the first applies. That means you generally want the broader rules first in sudoers. This can be a little harder to see whith extensive use of includes.
Both are reasons you may want to check rules do what you think for a given user:
sudo -l -U thatuser
(or sudo -l as that user)
Notes:
- visudo -cs does some stricter checking of the sudoers setup
- can help catch a few bordercases that may be biting you, and things in included files.
CHECKME
- password expiry (if applicable) will make an account fail to sudo, and for functional accounts you may not notice that.
- check with chage -l username
- https://www.2daygeek.com/linux-check-user-password-expiration-date/
- apparently PAM can be in the way, e.g. failing on a password for entries that are NOPASSWD
- TODO: read up
Related errors
X and su; "MIT-MAGIC-COOKIE-1 data did not match"
See X_notes#.22MIT-MAGIC-COOKIE-1_data_did_not_match.22
error: parse error in /etc/sudoers near line -1
Usually means the user that is trying to sudo is not allowed to.
This can happen when there is no applicable rule for them, or when the user for an applicable rule comes from an external system (say, LDAP) and that system did not respond at the time of the sudo command.
no tty present and no askpass program specified
You'll get this error when sudo needs to ask for a password, and the way you run it means there is no terminal on which to do so.
The likely reasons are cron entries, and some cases of running from scripts.
Late 2009(verify) saw sudo becoming more strict.
It will refuse to ask for a password if no tty is present, because it can't disable echo(ing the password).
You can make it act like it did before this time by using the 'visiblepw' flag.
If this is cron and you thought you had set up a NOPASSWD sudo line to avoid this, usually it means it is getting run at a different user, e.g. nobody instead of munin.
Sorry, user username is not allowed to execute 'something' as root on host
When it mentions bash, in response to su
In my case, I had allowed a user /bin/su as specific other user via sudoers, but not to run anything else as that user.
...most relevantly the shell (that sudo implies when invoked like sudo -i or -l). In this specific case, the easy-enough fix was to allow the shell in the same sudoers line.
Note that in the same situation, sudo su won't complain as it's the already-elevated su that invokes that shell.(verify)
In other cases
You may have written rules that don't apply (can be transient failure, e.g. hostname confusion), or confused sudoers in another way.
gksu, kdesu
These are wrapper for GUI programs that may want root permissions. gksu is from/for GNOME, kdesu for/from KDE.
They do little more than ask for the root login when you are not root, but make it much easier for e.g. menu entries to easily deal with system configuration apps to apply to root and non-root users.
Quota
Quota
- keeps track of, and limits, users's and/or groups's files
- both their size (units of 1K blocks) and the number of inodes (usually less relevant)
- There is a soft limit that you can cross with only a warning, for a configured grace period.
- The hard limit cannot be crossed. Using both is nicer to your users, and usually you only care about order of magnitude anyway
- applies per ext[234] filesystem
- can work through NFS - clients will ask for quota at the server end (what you'ld want)
- other filesystems vary:
- XFS is slightly different.
- ZFS has its own separate system
- requires kernel support
- isn't useful without its management tools
- stores a record of the current use, in what it calls the index
- (There is an old variant, but pretty much everyone will have be using v2 by now)
Setting up:
- Make sure your kernel supports it (on modern *nix you can assume it)
- Install the quota tools. Usually called quota. (There are additional tools, such as quotatool, that you may not need)
- edit /etc/fstab to add usrquota,grpquota to the mount options of the filesystems you want a quotum on
- you can use journaled quota when supported. This basically means you don't need to run quotacheck after an unclean shutdown. It would involve adding usrjquota=aquota.user,grpjquota=aquota.group,jqfmt=vfsv0
- make sure the files to store the index exist in the root of each relevant filesystem. Either:
- touch aquota.user aquota.group, or:
- use -c when you manually run quotacheck the first time
- remount, to make usrquot,grpquota apply
- For filesystems currently in use (e.g. root filesystem), using mount -vo remount is nice
- Less-subtle variants: umount and mount, or reboot.
- create the quota index, to account for existing contents: quotacheck -vgum mountpoint
- note: -a on this and other tools means '...for all mountpoints for which we use quota', which can save typing
- -v for verbose may be nice for feedback to you.
- -u for users, -g for groups
- -m without this, the tool will remount the filesystem read-only, which you may not like on a running filesystem (particularly your root filesystem). This option tells it to leave it read-write, but keep in mind that any file currently changing size will be incorrectly indexed.
- -c for create the indices (alternative is creating empty files beforehand)
- if it tells you it can't guess format from filename, use -F vfsv0 (or vfsv1 if supported and wanted)
- Optional: run repquota -a to get a summary of the index we just made
quotaon -v filesystem
- you'll probably want to set that at bootup, e.g. in your init script
- You may want to put quotacheck -mug mountpoint in your crontab to make sure the index doesn't go off much(verify)
- There are warnings about using that on mounted filesystem. Not because it can corrupt things, but because if a user was handling a huge file just when the index was re-made, the quota can be off by that file's size until the next rescan
- ...so try to schedule for time of light use, like 4AM
Using - users:
- quota display (-s makes for human-readable sizes)
Using - admin:
- edquota user or edquota -g group - interactively alter quota for user or group.
- edquota -p user1 user2 user3 ... - copy quota settings from user1 to user2, user3, etc. Also works on groups. Sometimes quite convenient.
- soft: max blocks/inodes the user may have before warning is issued and grace countdown begins.
- hard: max blocks/inodes
- 0 means no enforced limit
- the grace period is the same for both
Example:
edquota -t 7 days edquota user_or_uid
- setquota is a non-interactive alternative, useful when automating things. Example: setquota -u -F vfsv0 quotauser 1048576 2097152 0 0 /
- repquota mountpoint - report quota use per user. Use -g for groups. Use -s for human-readable sizes.
- TODO: you probably want to automate quota setting for new users.
See also:
Quota over NFS
NFS supports it, though under some conditions you may "Input/output error" instead of "Disk quota exceeded"(verify) (I got the latter, though, so that seems to work too)
Basically:
- Server side:
- Configure quota as usual
- set up rquotad to run (...if you want client-side quota commands to work). This service may be called something else, such as rpcquota (verify)
- Client side:
- mount NFS as usual (no quota related mount options, they won't work)
- add the -r option to edquota to make it contact the relevant NFS server (note: many other quota tools are local-only, so a habit of logging into the server may be easier)
http://chschneider.eu/linux/server/nfs.shtml
Running with different credentials / reduced permissions
Know the permission details above.
See also su, sudo, sudoers
SUID, SGID
chroot jails
The graphical run-as-root wrapper is gksu, or kdesu (GNOME and KDE variants). May be injected into the command when the system can guess you'll need it (particularly in Alt-F2 style launching).