From: Rafal Wojtczuk

To: bugtraq@securityfocus.com

Subject: Flaws in recent Linux kernels

Date: 18 Oct 2001 17:35:40 +0000



Hello,

There are two bugs present in Linux kernels 2.2.x, x<=19 and 2.4.y,

y<=9. The first vulnerability results in local DoS. The second one,

involving ptrace, can be used to gain root privileges locally (in case of

default install of most popular distributions). Linux 2.0.x is not vulnerable

to the ptrace bug mentioned.



I. Local DoS via deep symlinks

An attacker can force the kernel to spend almost

arbitrary amount of time on dereferencing a single symlink, which prevents

other processes from running. The attached

script, mklink.sh, takes a single

parameter N. The script creates 5 symlinks, each of

them containing 2*N+1 path elements. When N=3, the symlinks look this way:

$ ls -lG

drwxr-xr-x 2 nergal 4096 wrz 21 14:46 l

lrwxrwxrwx 1 nergal 53 wrz 21 14:46 l0 ->

l1/../l1/../l1/../l/../../../../../../../etc/services

lrwxrwxrwx 1 nergal 19 wrz 21 14:46 l1 -> l2/../l2/../l2/../l

lrwxrwxrwx 1 nergal 19 wrz 21 14:46 l2 -> l3/../l3/../l3/../l

lrwxrwxrwx 1 nergal 19 wrz 21 14:46 l3 -> l4/../l4/../l4/../l

lrwxrwxrwx 1 nergal 19 wrz 21 14:46 l4 -> l5/../l5/../l5/../l

drwxr-xr-x 2 nergal 4096 wrz 21 14:46 l5

drwxr-xr-x 2 rybagowa 4096 lut 27 1999 still_here



The amount of time the command "head l0" consumes (measured with time(1))

follows:

N system time

10: sys 0m0.050s

20: sys 0m1.400s

30: sys 0m10.150s

40: sys 0m41.840s



When "head l0" is being executed, other processes are not scheduled to

run. Thus the possibility of local DoS (in case of SMP you may need to spawn

one mklink.sh process per cpu). The time spent on dereferencing "l0" is

proportional to the number of path elements in normalized "l0". So, when

N=120, the scheduler should be locked out for about three hours. One can

reach N=600, in case of 2.4.9; also in case of 2.4.9, one can create even more

(up to eight) levels of symlinks.

2.4.10 fixed this problem, but not completely. Under 2.4.10 "head

l0" command would not block the scheduler, but it cannot be killed. The

problem is fully solved in 2.4.12.



II. Root compromise by ptrace(3)

In order for this flaw to be exploitable, /usr/bin/newgrp must be

setuid root and world-executable. Additionally, newgrp, when run with no

arguments, should not prompt for password. This

conditions are satisfied in case of most popular Linux distributions (but

not Openwall GNU/*/Linux).

Suppose the following flow of execution (initially, Process 1 and

Process 2 are unprivileged):

Time Process 1 Process 2

0 ptrace(PTRACE_ATTACH, pid of Process 2,...)

1 execve /usr/bin/newgrp

2 execve /any/thing/suid

3 execve default user shell

4 execve ./insert_shellcode



The unexpected happens at moment 2. Process 2 is still traced, execve

/any/thing/suid succeeds, and the setuid bit is honored ! This is so

because

1) the property of "having an ptrace-attached child" survives the execve

2) at moment 2, the tracer (process 1) has CAP_SYS_PTRACE set (well, has all

root privs), therefore it is allowed to trace even execve of setuid binary.

In moment 3, newgrp executes a shell, which is an usual behavior.

This shell is still able to control the process 2 with ptrace. Therefore, the

"./insert_shellcode" binary is able to insert arbitrary code into the address

space of Process 2. Game over.

In order to exploit this kernel vulnerability, one needs a setuid

root binary which execs an user-defined binary (or a shell). Newgrp is

appropriate on most distributions. On default install of slackware it does

not work (the password fields in /etc/group are empty, and newgrp demands a

password). However, one can use "su" on this distribution. "su"

binary is compiled without PAM support on slackware, therefore it execs an

user shell.

Do you remember the exploit against *BSD procfs, published in

January 2000 (http://www.securityfocus.com/cgi-bin/archive.pl?id=1&mid=43189) ?

This one is very similar; a setuid binary is spawned so that the system treats

it as a tracing process. Observe that in case of newgrp, only CAP_SYS_SETGID

is required (plus probably some reserved egid E to read gshadow; provided that

gshadow would be readable by gid E). If the file system supported granting

capabilities to programs (not only +s bit), this bug could have been benign.

Similarly, "su" needs only CAP_SYS_SETUID+CAP_SYS_SETGID (and egid shadow).

The "least privilege" rule, strictly applied, can save from a lot of

unexpected trouble.

This bug seems to be Linux-specific. I have tested FreeBSD, OpenBSD

and [older versions of] Irix and Solaris. None of the tested systems

honored setuid bit when an executing process was traced, even when the

tracer was root.



III. Vendor status

The kernel developers were notified on 18th September.

vendor-sec at lists dot de was notified on 9th October.



IV. Availability of patches.

2.4.12 kernel fixes both presented problems. The attached patches,

2.2.19-deep-symlink.patch and 2.2.19-ptrace.patch, both blessed by Linus,

can be used to close the vulnerability in 2.2.19. The (updated)

Openwall GNU/*/Linux kernel patches can be retrieved from

http://www.openwall.com/linux/

Note that the default Owl installation is not vulnerable to the ptrace bug

described.



V. The exploits

The attached mklink.sh script creates malicious symlinks.

ptrace-exp.c and insert_shellcode.c exploit the ptrace bug on i386

architecture. You will probably need to adjust #define in the latter. Note

that ptrace-exp uses LD_DEBUG variable to force a setuid program to generate

output. This technique (stderr redirected to a pipe, LD_DEBUG set, especially

LD_DEBUG=symbols) allows for forced suspending of a setuid binary in a

precisely determined moments, which may be helpful to build exploits which

rely on race-conditions. And finally, notice that under Owl LD_DEBUG is

ignored in case of suid binaries.



Save yourself,

Nergal

http://www.7bulls.com