|
COMMAND FreeBSD kernel integer overflow SYSTEMS AFFECTED FreeBSD: RELENG_4 (aka -STABLE) : not vulnerable after 20021111 RELENG_5_0 (aka -CURRENT) : vulnerable in lseek(2) (199) All current -RELEASE versions : vulnerable in fpathconf(2) (192) OpenBSD/NetBSD: not vulnerable PROBLEM In Joost Pol of Pine Digital Security advisory : http://www.pine.nl/press/pine-cert-20030101.txt --snap-- Inside the FreeBSD kernel each file (socket, device or regular file) opened is represented by a file structure (sys/file.h). Amongst other members this structure holds a reference counter (int f_count). This reference counter is increased by the fhold() function and decreased by the fdrop() function. (both in sys/file.h) For example, when a file is open(2)ed or dup(2)ed the reference counter is increased and when the file is close(2)ed again the reference counter is decreased. Once the reference counter reaches zero, the file structure itself is deallocated. Most system calls which perform (blocking) operations on a file will issue a fhold() call to prevent the file from being closed in the middle of an operation. Once the operation is finished the (extra) reference will be released again by issuing a fdrop() call. Inside the fpathconf(2) (192) system call we spotted a condition where a call to fdrop() is missing. When issueing a fpathconf(2) call on a socket it will return with an error condition but it will not release the extra file reference. Due to the missing fdrop() call inside the fpathconf(2) system call is it possible to overflow the reference counter of the file structure (int f_count). FreeBSD -CURRENT suffers from the same problem in another syscall. Exploitability ============== Causing a system panic and privilege escalation have been confirmed. Exploitation of this bug will take time, depending on machine speed and system limits this could vary between hours and days. 1. System Panic A system panic can be caused by issuing around 2^31 calls to fpathconf(2) with a filedescriptor which references a socket. The reference counter (int f_count) will wrap to a negative value and this will cause a panic in close(2). 2. Privilege Escalation It is also possible (although more difficult) to gain root access using this bug. One would open(2) a socket and dup(2) it. One would then issue around 2^32 - 1 calls to fpathconf(2) causing the reference counter to wrap to 1. After closing the original filedescriptor the file structure will be deallocated. At this point the dup(2)ed file descriptor is still hanging around with a file structure pointing to unallocated memory. The final step, which is left as an excercise for the reader, is to have a sensitive file (like /etc/skeykeys) opened and allocated at the previously freed location. Once this happens the dup(2)ed file descriptor is still hanging around providing access to this file. This could result in the escalation of user privileges. --snip-- Although the missing fdrop() call in fpathconf(2) was noticed before by Nakamura Takayuki <tak@st.rim.or.jp> its impact was severely underestimated. SOLUTION Update (08 January 2003) ====== The following patch has been verified to apply to FreeBSD 4.4, 4.5, 4.6, and 4.7 systems : Path Revision Branch - ------------------------------------------------------------------------- src/sys/kern/kern_descrip.c RELENG_4 1.81.2.15 RELENG_4_7 1.81.2.14.4.1 RELENG_4_6 1.81.2.14.2.1 RELENG_4_5 1.81.2.9.2.3 RELENG_4_4 1.81.2.8.2.3 RELENG_5_0 1.169.2.2 src/sys/kern/vfs_syscalls.c RELENG_5_0 1.297.2.2 - ------------------------------------------------------------------------- References http://www.pine.nl/press/pine-cert-20030101.txt file://usr/include/sys/file.h file://usr/src/sys/kern/kern_descrip.c file://usr/src/sys/kern/vfs_syscalls.c