Date: Mon, 29 Jul 2002 10:51:50 -0400 (EDT)
From: Michal Zalewski <[email protected]>
To: [email protected], [email protected]Subject: RAZOR advisory: Linux util-linux chfn local root vulnerability
Linux util-linux chfn local root vulnerability
Issue Date: July 29, 2002
Contact: Michal Zalewski
CVE: CAN-2002-0638
CERT vulnerability note: http://www.kb.cert.org/vuls/id/405955
(the URL should be accessible soon)
Topic:
A locally exploitable vulnerability is present in the util-linux
package shipped with Red Hat Linux and numerous other Linux
distributions.
Affected Systems:
Red Hat Linux 7.3 and previous; potentially many other distributions
up to date that use util-linux to provide chfn and chsh utilities.
Please refer to the CERT vulnerability note for more information.
Systems that ship chfn within the shadow-utils package (for example
SuSE) are not vulnerable.
Details:
The vulnerability is present within the shared code of the util-linux
package. For the purpose of this discussion, we use the setuid
application 'chfn' as an attack vector. We used the "Fenris" utility,
available at http://razor.bindview.com/tools/fenris, to analyze the
issue.
The purpose of the 'chfn' utility is to allow users to modify personal
information stored within the system-wide password file, /etc/passwd.
In order to modify this file, this application must be installed
setuid root. It is a part of the base installation of most Linux
systems.
Under certain commonly met conditions (see "Mitigating Factors"), a
carefully crafted attack sequence can be performed to exploit a
complex file locking and modification race present in this utility,
and, as a result, alter /etc/passwd to escalate privileges in the
system.
Password file locking and modification code can be found in util-linux
within the login-utils/setpwnam.c file. The general logic used for
this process is as follows:
1. /etc/ptmptmp file is opened with O_WRONLY|O_CREAT, 0644 perms
2. the file is linked to /etc/ptmp, exit on failure
3. /etc/ptmptmp is removed
Later, the descriptor obtained in step 1 is used for writing to
construct the new /etc/passwd contents. This is done line by line, by
calling the fputs() routine. When the new file is ready, three more
steps are taken:
4. /etc/passwd.OLD is removed
5. /etc/passwd is linked to /etc/passwd.OLD
6. /etc/ptmp is renamed to /etc/passwd
At first sight, this logic is not flawed. There is no O_EXCL in step
1, but step 2 works as an accurate exclusive locking mechanism, and
prevents other instances of this application from making any writes
until the procedure is completed (step 6). Also, step 3 ensures that
no process will work on the hardlink of /etc/passwd after the
procedure is completed. This, while is not a standard way of handling
file locking on Linux, is an adequate protection against race
conditions. The only concern is that if chfn, chsh, or any other
process that uses /etc/ptmp is terminated abnormally (e.g. never exits
because of a system crash or is killed with SIGKILL), the lock has to
be removed manually in order for applications that rely on this
signaling method to work properly. This includes utilities such as
chsh, chfn, useradd, userdel, adduser, vipw and other software. Most
of seasoned Linux administrators have had to remove this lock file at
least once, and many utilities offer help troubleshooting this issue -
for example, adduser refuses to run when /etc/ptmp is present and
terminates with the following message:
# adduser
vipw lockfile (/etc/ptmp) is present!
This is exactly what causes the problem. Because this dangling lock
file makes it impossible to perform regular user management tasks, it
is trivial to force the administrator to remove it. This is a risk if
chfn is stopped (by SIGSTOP) between steps 2 and 3 - which can be
done. There is a very small window of opportunity here, but it is
possible to fit into it. Later, the application can be resumed with
SIGCONT.
Note that there is no easy way to diagnose who created /etc/ptmp and
for what reason they created it. The file is owned by root and has no
additional information whatsoever. The attack relies on small windows
of opportunity, making one-shot attempts not feasible on certain
systems, it is reasonable to expect that if the attacker repeatedly
runs chfn or chsh just to create dangling lock files several times a
day - and to make user management and maintenance tasks repeatedly
fail because of this lockfile - the administrator will most likely add
"rm -f /etc/ptmp" or equivalent to his crontab, instead of putting
much more effort into tracing the problem that cannot be easily
diagnosed using standard Linux tools - there is no log entry
associated with successful authentication for chfn service, ownership
of the file is set to root, the file is empty.
As soon as this or a similar counter-measure is deployed,
trial-and-error attacks become much simpler. None of the above
administrative tasks is to be viewed as requiring social engineering.
Such tasks are a part of standard system maintenance and operations,
and it is safe to assume that almost all system administrators will
follow this or a similar procedure to address the problem.
Once the attacker has a stopped chfn process in between steps 2 and 3,
and the /etc/ptmp file has been removed, he can easily execute another
instance of chfn. It will open the already existing /etc/ptmptmp (this
file is not considered a lock by any software, and most likely won't
be deleted). The second instance runs and step 2 succeeds because
there is no lock file present. The file is then written. As soon as
/etc/ptmp has the projected size (or /etc/passwd.OLD appears), the
instance should be killed to avoid renaming to /etc/passwd (this is a
timing concern for this attack, and we will address it later).
The First instance of chfn is still holding an open descriptor to
/etc/ptmptmp, which later became /etc/ptmp - and, if we send SIGCONT
to this process, will be renamed to /etc/passwd. Step 3 will fall
through because there is no error checking, and new information will
be written into a descriptor that will de facto become /etc/passwd.
Copying the file should go smoothly up to the point where attacker
account information is stored. Let's analyze what is going to happen
past this point if the second instance of chfn was called with
significantly longer GECOS parameters than first (minimum GECOS length
set by chfn is zero, maximum is 260 characters). Keep in mind that
there is no call to ftruncate() before fclose(), so the new file would
have a "tail" consisting of the last characters of the longer version
of /etc/passwd. If 'Write 1' is the first write to file, or the write
made by the second instance of chfn, 'Write 1 is the second write
(first instance), and '|' represents a record separator (newline),
then:
Write 1: user 1 | user 2 | ATTACKER LONG | user 3 | user 4 |
Write 2: user 1 | user 2 | ATTACKER | user 3 | user 4 |
------------------------------------------------------------
Result: user 1 | user 2 | ATTACKER | user 3 | user 4 |r 4 |
As a result, /etc/passwd has now a new line at the end, a "leftover"
part from last username. This, itself, can be considered an annoyance,
but is hardly a full-blown security issue. There is, however,
something that makes this problem more serious - that is, GNU libc.
As we mentioned earlier, records are being written to wannabe
/etc/passwd line by line, using fputs(). It is important to realize
that the text-file functions such as fputs() are buffered - that is,
lines are being actually written to disk in chunks, when internal
cache is full. Until this point, the OS's idea of the file and the
process's idea of the file are not synchronized. Cache block size, by
default, is four kilobytes. This means that if the password file has
more than four kilobytes (which is very often true on systems where
local attacks are performed, see "Mitigating Factors"), and the
attacker's account is not in the last cached that is going to be
written before fclose(), it should be possible to "cut" the file on
four kilobytes boundary by killing the first chfn process as soon as
/etc/passwd has the desired contents:
<--- attacker's account 4 kB boundary
|
Write 1: ...ser1:x:640:640:Foo Bar:/bin/bash\nuser2:x:641:641:Foo...
Write 2: ...user1:x:640:640:Foo Bar:/bin/bash\nuser2: ***CUT***
--------------------------------------------------------------------
Result: ...\nuser2::641:641:Foo...
Note that both the data written in "Write 1" and in "Write 2" can be
padded within a 260 byte range, which is typically much more than a
typical /etc/passwd line length, so it is possible to pad the record
exactly in the way described above - which happens to erase the
password for user2. All it takes is to determine the appropriate
initial padding, then invoke the first chfn with one more character in
GECOS than in the second chfn.
It is also possible to create root accounts or other privileged access
accounts that effectively lead to privilege escalation (kmem, mail,
etc). A root account with no password can be generated in two steps,
first truncate the UID ending with '0' to the last digit, and the
second step is to remove the password as described above. Note that
even if there is no account with UID 0 anywhere close to the 4 kB
boundary, data can be moved over much longer distance by several
subsequent attack cycles where the first instance is called with the
maximum length GECOS field, and the second with minimum.
The question of timing required for trimming is essential in
evaluating the feasibility of this attack - but, fortunately for the
attacker, timing is on his side. First of all, the copying procedure
is essentially pretty slow, being done in a read-compare-write cycle.
Then, the results are cached so the output appears in larger chunks,
but less frequently, giving the attacker plenty of CPU cycles.
Finally, the passwd file is world-readable, which gives the attacker a
tremendous advantage - it can be easily monitored e.g. by mapping a
single byte at the boundary to the memory space of the exploit.
Additionally, for files that consist of more than one segment past the
attacker's account, the problem of careful timing needed to avoid
renaming /etc/ptmp to /etc/passwd becomes less of an issue, because
the process can be killed before reaching the end of the file, just
after the desired boundary is modified.
Mitigating factors:
In order to successfully exploit the vulnerability and perform
privilege escalation, the following additional requirements have to be
met:
1. There is a need for a minimal administrator interaction that falls
into the category of standard system maintenance and operations.
2. The password file, /etc/passwd, must be over 4 kilobytes. This
requirement is typically satisfied on systems with approximately
50 accounts or more - that is, on a vast majority of systems that
where unprivileged accounts are being compromised / used by
malicious attackers.
3. Assume that /etc/passwd is divided into 4 kB chunks. The
attacker's account record in this file must NOT be located in the
last chunk. That is, if /etc/passwd has 9 kB and we have three
chunks, 4 kB, 4 kB and 1 kB, and attacker's account must be
located within the first or second chunk and cannot be located
further than 8th kilobyte of the file.
Other mitigating factors are similar to ones for other local
vulnerabilities in setuid binaries.
Workaround / fix:
The fast workaround is to remove setuid flags from /usr/bin/chfn and
/usr/bin/chsh. A more appropriate fix is to patch the sources as
follows:
--- util-linux-2.11n-old/login-utils/setpwnam.c Mon Jul 31 08:50:39 2000
+++ util-linux-2.11n/login-utils/setpwnam.c Wed Jun 12 21:37:12 2002
@@ -98,7 +98,8 @@
/* sanity check */
for (x = 0; x < 3; x++) {
if (x > 0) sleep(1);
- fd = open(PTMPTMP_FILE, O_WRONLY|O_CREAT, 0644);
+ // Never share the temporary file.
+ fd = open(PTMPTMP_FILE, O_WRONLY|O_CREAT|O_EXCL, 0644);
if (fd == -1) {
umask(oldumask);
return -1;
Vendor Response:
Please refer to CERT VU#405955 for more information.
Additional credits:
Thanks to Chris Coffin and Mark Loveless for their contributions
to this advisory.
