Common use of 'tmpwatch' utility and its counterparts triggers race conditions in many applications Michal Zalewski , 12/05/2002 Copyright (C) 2002 by Bindview Corporation 1) Scope and exposure info -------------------------- A common practice of installing 'tmpwatch' utility or similar software configured to sweep the /tmp directory on Linux and unix systems can compromise secure temporary file creation mechanisms in certain applications, creating a potential privilege escalation scenario. This document briefly discusses the exposure, providing some examples, and suggesting possible workarounds. It is believed that many unix operating systems using 'tmpwatch' or an equivalent are affected. Numerous Linux systems, such as Red Hat, that ship with cron daemon running and 'tmpwatch' configured to sweep /tmp are susceptible to the attack. 2) Application details ---------------------- 'Tmpwatch' is a handy utility that removes files which haven't been accessed for a period of time. It was developed by Erik Troan and Preston Brown of Red Hat Software, and, with time, has become a component of many Linux distributions, also ported to platforms such as Solaris, *BSD or HP/UX. By default, it is installed with a crontab entry that sweeps /tmp directory on a daily basis, deleting files that have not been accessed for the past few days. An alternative program, called 'stmpclean' and authored by Stanislav Shalunov, is shipped with *BSD systems and some Linux distributions to perform the same task, and some administrators deploy other tools or scripts for this purpose. 3) Vulnerability details ------------------------ Numerous applications rely either on mkstemp() or custom O_EXCL file creation mechanisms to store temporary data in the /tmp directory in a secure manner. Of those, certain programs run with elevated privileges, or simply at a different privilege level than the caller. The exposure is a result of a common misconception, promoted by almost all secure programming tutorials and manpages, that /tmp files created with mkstemp(), granted that umask() settings were proper, are safe against hijacking and common races. The file, since it is created in a sticky-bit directory, indeed cannot be removed or replaced by the attacker running with different non-root privileges, but since many operating systems feature 'tmpwatch'-alike solutions, the only thing that can and should be considered safe in /tmp is the descriptor returned by mkstemp() - the filename should not be relied upon. There are two major reasons for this: 1) unlink() races It is very difficult to remove a file without risking a potential race (see section 4). 'Tmpwatch' does not take any extra measures to prevent races, and probes file creation time using lstat(). Based on this data, it calls unlink() as root. Problem is, on a multitasking system, it is possible for the attacker to get some CPU time between those two system calls, remove the old "decoy" file that has been probed with lstat(), and let the application of his choice create its own temporary file under this name. While mkstemp() names are guaranteed to be unique, they shouldn't be expected to be unpredictable - in most implementations, the name is a function of process ID and time - so it is possible for the attacker to guess it and create a decoy in advance. Once the tmpwatch process is resumed, the file is immediately removed, based on the result of earlier lstat() on the old, no longer existing file. While this three-component race requires very precise timing, it is possible to try a number of times in a single 'tmpwatch' run if enough decoy files are created by the attacker. Additionally, since each step of the attack would result in a corresponding filesystem change, it is fairly easy to carefully measure timings and coordinate the attack. If the attacker cannot make the application run at the same time as 'tmpwatch' - for example, if the application is executed by hand by the administrator, or is running from cron - 'tmpwatch' itself can be artificially delayed for almost an arbitrary amount of time by creating and continuously expending an elaborate directory structure in /tmp using hard links (to preserve access times of files) and running other processes that demand disk access and cache space to slow down the process. 'Stmpclean' offers additional protection against races by not removing root-owned files and temporarily dropping privileges when removing the file to match the owner of lstat()ed resource. Unfortunately, not removing root files is a considerable drawback, and there is still a potential for a race using carefully crafted hard links to a file owned by the victim and two concurrent 'stmpclean' processes: - the attacker links /tmp/foo to ~victim/.bash_profile - tmpwatch #1 does lstat() on /tmp/foo and setuid victim - tmpwatch #2 does lstat() on /tmp/foo and setuid victim - tmpwatch #1 does unlink("/tmp/foo") - victim application creates /tmp/foo at uid==victim - tmpwatch #2 does unlink("/tmp/foo") and succeeds - the attacker creates /tmp/foo - victim application proceeds On certain systems such as Owl Linux, the attack will be not possible due to hardlink limits imposed on sticky-bit directories. 2) suspended processes and 'legitimate' file removal Here, all conventional measures that could be exercised by /tmp cleaners fail miserably. A vulnerable application can be often delayed or suspended after mkstemp() / open() - for example, a setuid program can be stopped with SIGSTOP and resumed with SIGCONT. If the application is suspended for long enough, its temporary files are likely to be removed. This method requires much less precision, but is also more time-consuming and has a more limited scope (interactive applications only). Note that it is sometimes possible to delay the execution of a daemon - client wait, considerable I/O or CPU loads, and subsequent mkstemp() calls can be all used to achieve the effect. The feasibility and efficiency is low, but the potential issue exists. Some client applications that are often left unattended and create temporary files - such as mail/news clients, web browsers, irc clients, etc - can also be used to compromise other accounts on the machine. Not all applications are prone to the problem just because mkstemp() is used to create files in /tmp; if the file name is not used to perform any sensitive operations with some extra privileges afterward (read, write, chown, chmod, link/rename, etc), and only the descriptor is being used, the application is safe. This practice is often exercised by programmers who want to avoid leaving dangling temporary files in case the program is aborted or crashes. Similarly, if the application uses temporary files improperly, but does not rely on their contents and does not attempt to access them with higher privileges, the application is secure in that regard. Applications that run with higher privileges and reopen their /tmp temporary files for reading or writing, call chown(), chmod() on them, rename or link the file to replace some sensitive information, and so on, are exposed. It is worth mentioning that a popular 'mktemp' utility coming from OpenBSD passes only the filename to the caller shell script, thus rendering almost all scripts using it fundamentally flawed. If the script is being run as a cron job or other administrative task, and mktemp is used, the system can be likely compromised by replacing the file after mktemp and prior to any write to the file. In the example quoted in the documentation for mktemp(1): TMPFILE=`mktemp /tmp/$0.XXXXXX` || exit 1 echo "program output" >> $TMPFILE ...the attacker would want to replace temporary file right before 'echo', causing the text "program output" to be appended to a target file of his choice using symlinks or hardlinks; or, if it is more desirable, he'd spoof file contents to cause the program to misbehave. Another example of the problem is a popular logrotate utility, coded - ironically - by Erik Troan, one of co-authors of 'tmpwatch' itself. The program suffered /tmp races in the past, but later switched to mkstemp(). The following sequence is used to handle post-rotation shell commands specified in config files: open("/tmp/logrotate.wvpNmP", O_WRONLY|O_CREAT|O_EXCL, 0700) = 6 ... write(6, "#!/bin/sh\n\n", 11) = 11 write(6, "\n\t/bin/kill -HUP `cat /var/lock/"..., 79) = 79 close(6) = 0 ... fork, etc ... execve("/bin/sh", ["sh", "-c", "/bin/sh /tmp/logrotate.wvpNmP" ... Obviously, if the attacker can have /tmp/logrotate.* replaced in between mkstemp() (represented as open() syscall above) and the point where another process is spawned, a shell interpreter is invoked, then executes another copy of the shell interpreter (apparent programmer's mistake) and finally reads the input file - which is a considerable chunk of time - the shell will be called with attacker-supplied commands to be executed with root privileges. On Red Hat, logrotate is executed from crontab on a daily basis, in a sequence before 'tmpwatch', and the easiest option for the attacker is to maintain a still-running tmpwatch process from the previous day to exploit the condition. On systems where those programs are not executed sequentially - for example, when both programs are listed directly in /etc/crontab - the attack requires less precision. 4) Workarounds and fixes: ------------------------- Recommended immediate workaround is to discontinue the use of 'tmpwatch' or equivalent to sweep /tmp directory if this service is not necessary. For applications that rely on TMPDIR or a similar environment variable, setting it to a separate, not publicly writable directory is often a viable solution. Note that not all applications honor this setting. In terms of a permanent solution, two different attack vectors have to be addressed, as discussed in section 3: 1) unlink() race The proper way to remove files in sticky-bit directories while minimizing the risk is as follows: a) lstat() the file to be removed b) if owned by root, do not remove c) if st_nlink > 1, do not remove d) if owned by user, temporarily change privileges to this user e) attempt unlink() f) if failed, warn about a possible race condition g) switch privileges back to root With the exception of step c, this is implemented in 'stmpclean'. Unfortunately, step c is crucial on systems that do not have restricted /tmp kernel patches from Openwall (http://www.openwall.com), otherwise, there is a potential for fooling the algorithm by supplying a hard link to a file owned by the victim, as discussed in section 3. This approach has several drawbacks - such as the fact root-owned files will not be removed. Other solution is to modify applications that generate filenames on their own, and to modify mkstemp(), to generate names that are not only unique, but not feasible to predict. Another suggestion is to implement a funlink() capability in the kernel of the operating system in question, to allow race-free file removal, thus removing the non-root ownership requirement for the method described above, and simplifying the approach. A skeleton patch to implement funlink() semantics and make sure the file being removed is the file opened and fstat()ed previously is available at: http://lcamtuf.coredump.cx/soft/linux-2.4-funlink.diff (this and other patches are not endorsed by RAZOR in any way). 2) suspended process and 'legitimate' file removal This issue is fairly difficult to address. The most basic idea is to use a special naming scheme for temporary files to avoid deletion - unfortunately, this seems to defeat the purpose of using tmpwatch-alike solutions in the first place. An alternative approach, which is to enforce separate temporary directories for certain applications, either process-, session- or uid- based, is generally fairly controversial, and raises some concerns. Advisory separation is generally acceptable, but there are a number of applications that do not accept TMPDIR setting, and a widespread practice of using /tmp in in-house applications. Mandatory separation (kernel modification) raises compatibility concerns and is generally approached with skepticism - no implementation has become particularly popular. Ideally, implementators should carefully audit their sources. It is recommended for privileged applications to use private temporary directories for sensitive files, if possible; if using /tmp is necessary, extra caution has to be exercised to avoid referencing the file by name. Note that comparing the descriptor and a reopened file to verify inode numbers, creation times or file ownership is not sufficient - please refer to "Symlinks and Cryogenic Sleep" by Olaf Kirch, available at http://www.opennet.ru/base/audit/17.txt.html . It's worth noticing that 'tmpwatch' offers a -s option, which causes the program to run the 'fuser' command to prevent removal of files that are currently open. At first sight, this could be an effective way to solve the problem. Unfortunately, this is not true, since many applications close the file for a period of time before reopening (including logrotate and mktemp(1)). 5) Credits and thanks --------------------- Thanks to Solar Designer for interesting discussions on the subject, to Matt Power for useful feedback, and to RAZOR team in general for making this publication possible.