Neohapsis is currently accepting applications for employment. For more information, please visit our website www.neohapsis.com or email hr@neohapsis.com
Subject: klogd format bug
From: Jouko Pynnönen (joukoSOLUTIONS.FI)
Date: Sun Sep 17 2000 - 23:13:53 CDT


Kernel logging daemon klogd in the sysklogd package for Linux contains a
"format bug" making it vulnerable to local root compromise (successfully
tested on Linux/x86). There's also a possibility for remote vulnerability
under certain (rather unprobable) circumstances and a more probable
semi-remote exploitableness with knfsd.


Normally klogd reads kernel messages from /proc/kmsg and after some
processing passes them to syslogd, which in turn writes the messages to
log file(s) as defined in syslog.conf.

The erroneous function calls are located in function LogLine(), klogd.c
lines 680 and 707:

                Syslog( LOG_INFO, line_buff );

Each newline-postfixed kernel message goes to Syslog() as a format
string. Before passing the string to Syslog(), the LogLine() function
does some simple processing. Kernel addresses of form [<address>] are
translated to symbol names and '%' chars are checked and duplicated to
avoid format problems. The logics however fail to handle %-symbols
inside a [<foo>] symbol marker, so a malicious user can forge a kernel
message like "[<%s %s %s %s>]" which would be enough to kill klogd with
a segmentation fault. Possibilities of executing arbitrary code are
discussed below.


The vulnerability is exploitable for anyone who can feed strings in the
kernel log. There are many ways to do this locally; for example a
connect() call with an improperly initialized sockaddr structure
generates a warning to the kernel log containing the program name.
However the program name is taken from task_struct->comm which is a
16-byte buffer. The required "[<" takes 2 bytes so the space gets short
for a conventional shellcode and return address. The warning is printed
only once so exploiting with just this message is probably difficult, but
other warnings in the kernel make it possible.

Other possibilities include special kernel modules that call printk() to
log their things to kernel log, or various device drivers. The driver for
/dev/mixer for instance can be trivially made log a string containing a
31-byte user defineable substring if /dev/mixer is openable (as it is per
default in many systems).

The exploitation process itself isn't so trivial. The buffer in question
is declared static so it's mapped to a low memory address; there's
nothing user-defineable or otherwise usable in stack. I haven't been able
to produce a program doing a standard "%n" exploitation, but it is
possible to create a buffer overflow condition elsewhere in the program
in conjunction with the format bug. If the format string contains
something like "%2000d", a local buffer in vsyslog() overflows. This
function is called from Syslog() and does no bound checking but trusts
Syslog() checks. In addition, "%m" in the format string expands to the
error message (from strerror()) which can be used to "push" the end of
the string further in the memory, making it possible to build a longer
shellcode using multiple klogd messages if needed.

According to my quick test with a knfsd-enabled system, the driver can be
tricked to printk() path names with some race conditioning, which is
probably one of the easiest exploit methods where available.


Linux distribution vendors have been informed about the problem via
vendor-seclst.de a week ago; patches and advisories will be out today.


Vulnerability discovered by:
     Jouko Pynnönen

Greets and thanks to:
     Esa Etelävuori, vendor-sec team, cc-opersIRCNet

Jouko Pynnönen           Online Solutions Ltd       Secure your Linux -
joukosolutions.fi                                  http://www.secmod.com