setuid
—
checklist for security of setuid programs
Please note: This manual page was written long ago, and is in
need of updating to match today's systems. We think it is valuable enough to
include, even though parts of it are outdated. A carefully-researched updated
version would be very useful, if anyone is feeling enthusiastic...
Writing a secure setuid (or setgid) program is tricky. There are a
number of possible ways of subverting such a program. The most conspicuous
security holes occur when a setuid program is not sufficiently careful to
avoid giving away access to resources it legitimately has the use of. Most
of the other attacks are basically a matter of altering the program's
environment in unexpected ways and hoping it will fail in some
security-breaching manner. There are generally three categories of
environment manipulation: supplying a legal but unexpected environment that
may cause the program to directly do something insecure, arranging for error
conditions that the program may not handle correctly, and the specialized
subcategory of giving the program inadequate resources in hopes that it
won't respond properly.
The following are general considerations of security when writing
a setuid program.
- The program should run with the weakest userid possible, preferably one
used only by itself. A security hole in a setuid program running with a
highly-privileged userid can compromise an entire system.
Security-critical programs like
passwd(1) should always have
private userids, to minimize possible damage from penetrations
elsewhere.
- The result of getlogin(2)
or ttyname(3) may be wrong
if the descriptors have been meddled with. There is no
foolproof way to determine the controlling terminal or the login name (as
opposed to uid) on V7.
- On some systems, the setuid bit may not be honored if the program is run
by root, so the program may find itself running as root.
- Programs that attempt to use
creat(3) for locking can foul
up when run by root; use of
link(2) is preferred when
implementing locking. Using
chmod(2) for locking is an
obvious disaster.
- Breaking an existing lock is very dangerous; the breakdown of a locking
protocol may be symptomatic of far worse problems. Doing so on the basis
of the lock being ‘old’ is sometimes necessary, but programs
can run for surprising lengths of time on heavily-loaded systems.
- Care must be taken that user requests for I/O are checked for permissions
using the user's permissions, not the program's. Use of
access(2) is
recommended.
- Programs executed at user request (e.g. shell escapes) must not receive
the setuid program's permissions; use of daughter processes and
“setuid(getuid())” plus “setgid(getgid())”
after fork(2) but before
exec(3) is vital.
- Similarly, programs executed at user request must not receive other
sensitive resources, notably file descriptors. Use of
fcntl(2)
F_CLOSEM
, FILENO_STDERR +
1
(close all fd's greater than stderr) and/or
fcntl(2)
F_SETFD
, FD_CLOEXEC
(close-on-exec) arrangements on systems which have them is recommended.
Other resources should also be examined for sanity and
possibly set to desired settings, such as the current working directory,
signal disposition, resource limits, environment, umask, group
membership, chroot.
Programs activated by one user but handling traffic on behalf
of others (e.g. daemons) should avoid doing
“setuid(getuid())” or “setgid(getgid())”,
since the original invoker's identity is almost certainly inappropriate.
On systems which permit it, use of “setuid(geteuid())” and
“setgid(getegid())” is recommended when performing work on
behalf of the system as opposed to a specific user.
- There are inherent permission problems when a setuid program executes
another setuid program, since the permissions are not additive. Care
should be taken that created files are not owned by the wrong person. Use
of “setuid(geteuid())” and its gid counterpart can help, if
the system allows them.
- Care should be taken that newly-created files do not have the wrong
permission or ownership even momentarily. Permissions should be arranged
by using umask(2) in advance,
rather than by creating the file wide-open and then using
chmod(2). Ownership can get
sticky due to the limitations of the setuid concept, although using a
daughter process connected by a pipe can help.
- Setuid programs should be especially careful about error checking, and the
normal response to a strange situation should be termination, rather than
an attempt to carry on.
The following are ways in which the program may be induced to
carelessly give away its special privileges.
- The directory the program is started in, or directories it may plausibly
chdir(2) to, may contain
programs with the same names as system programs, placed there in hopes
that the program will activate a shell with a permissive
PATH
setting. PATH
should
always be standardized before invoking a shell (either
directly or via popen(3) or
execvp(3) or
execlp(3)).
- Similarly, a bizarre
IFS
setting may alter the
interpretation of a shell command in really strange ways, possibly causing
a user-supplied program to be invoked. IFS
too
should always be standardized before invoking a shell.
- Environment variables in general cannot be trusted. Their contents should
never be taken for granted.
- Setuid shell files (on systems which implement such) simply cannot cope
adequately with some of these problems. They also have some nasty problems
like trying to run a .profile when run under a
suitable name. They are terminally insecure, and must be avoided.
- Relying on the contents of files placed in publically-writable
directories, such as /tmp, is a nearly-incurable
security problem. Setuid programs should avoid using
/tmp entirely, if humanly possible. The
sticky-directories modification (sticky bit on for a directory means only
owner of a file can remove it) helps, but is not a complete solution.
- A related problem is that spool directories, holding information that the
program will trust later, must never be publically writable even if the
files in the directory are protected. Among other sinister manipulations
that can be performed, note that on many Unixes, a core dump of a setuid
program is owned by the program's owner and not by the user running
it.
The following are unusual but possible error conditions that the
program should cope with properly (resource-exhaustion questions are
considered separately, see below).
- The value of argc might be 0.
- The setting of the umask(2)
might not be sensible. In any case, it should be standardized when
creating files not intended to be owned by the user.
- One or more of the standard descriptors might be closed, so that an opened
file might get (say) descriptor 1, causing chaos if the program tries to
do a printf(3).
- The current directory (or any of its parents) may be unreadable and
unsearchable. On many systems
pwd(1) does not run
setuid-root, so it can fail under such conditions.
- Descriptors shared by other processes (i.e., any that are open on startup)
may be manipulated in strange ways by said processes.
- The standard descriptors may refer to a terminal which has a bizarre mode
setting, or which cannot be opened again, or which gives end-of-file on
any read attempt, or which cannot be read or written successfully.
- The process may be hit by interrupt, quit, hangup, or broken-pipe signals,
singly or in fast succession. The user may deliberately exploit the race
conditions inherent in catching signals; ignoring signals is safe, but
catching them is not.
- Although non-keyboard signals cannot be sent by ordinary users in V7, they
may perhaps be sent by the system authorities (e.g. to indicate that the
system is about to shut down), so the possibility cannot be ignored.
- On some systems there may be an
alarm(3) signal pending on
startup.
- The program may have children it did not create. This is normal when the
process is part of a pipeline.
- In some non-V7 systems, users can change the ownerships of their files.
Setuid programs should avoid trusting the owner identification of a
file.
- User-supplied arguments and input data must be checked
meticulously. Overly-long input stored in an array without proper bound
checking can easily breach security. When software depends on a file being
in a specific format, user-supplied data should never be inserted into the
file without being checked first. Meticulous checking includes allowing
for the possibility of non-ASCII characters.
- Temporary files left in public directories like
/tmp might vanish at inconvenient times.
The following are resource-exhaustion possibilities that the
program should respond properly to.
- The user might have used up all of his allowed processes, so any attempt
to create a new one (via
fork(2) or
popen(3)) will fail.
- There might be many files open, exhausting the supply of descriptors.
Running fcntl(2)
F_CLOSEM
on systems which have it, is
recommended.
- There might be many arguments.
- The arguments and the environment together might occupy a great deal of
space.
Systems which impose other resource limitations can open setuid
programs to similar resource-exhaustion attacks.
Setuid programs which execute ordinary programs without reducing
authority pass all the above problems on to such unprepared children.
Standardizing the execution environment is only a partial solution.
passwd(1),
pwd(1),
access(2),
chdir(2),
chroot(2),
execve(2),
fcntl(2),
fork(2),
getlogin(2),
link(2),
setegid(2),
seteuid(2),
setgid(2),
setgroups(2),
setrlimit(2),
setuid(2),
sigaction(2),
umask(2),
alarm(3),
creat(3),
execvp(3),
popen(3),
printf(3),
ttyname(3)
Written by Henry Spencer, and based on additional outside contributions.
The list really is rather long... and probably incomplete.