FreeBSD kernel buffer overflow

[gerarra@xxxxxx]

Topic: Buffer Overflow in FreeBSD
Versions: All the versions of FreeBSD are broken (4.x, 5.x, 6.0)
Arch: x86
Date: 16/09/2004

All discussion refers to CURRENT-6.0, for other versions some things could
change (btw bugged).
Discussion involves a lot of arch x32 dependant mechanisms, so, in some
points, could sound a little bit dark.


A buffer overflow has been found in i386/i386/trap.c syscall() function
of FreeBSD official
source tree.
In order to rule syscalls mechanism, the 'particular' interrupt 128 (0x80)
is provided in the
IDT vector. To serve this interrupt, i386/i386/exception.s int0x80_syscall()
function is
done and, in the end, it calls syscall().
syscall() is responsible for loading arguments from a syscall and copying
them in a kspace
pointer in order to accessing them. The code to do that is the following:

void
syscall(frame)
        struct trapframe frame;
{
        caddr_t params;
        struct sysent *callp;
        struct thread *td = curthread;
        struct proc *p = td->td_proc;
        register_t orig_tf_eflags;
        u_int sticks;
        int error;
        int narg;
        int args[8];
        u_int code;


        ...


        narg = callp->sy_narg & SYF_ARGMASK;  (<- you can see it's the only one
check)

        if (params != NULL && narg != 0)
                error = copyin(params, (caddr_t)args,
                    (u_int)(narg * sizeof(int)));
        else
                error = 0;


        ...


and:

> grep SYF_ARGMASK /usr/src/sys/sys/sysent.h
#define SYF_ARGMASK     0x0000FFFF

It's obvious that the amount of selectable memory is beyond the (8 * 
sizeof(int))
limit of
args array, so it would overwrite the saved eip by syscall() (it's invoked
through a call) or
making an interesting pointer corruption overwriting   struct proc *p .

It's exploitable, but the only one way I discovered is to link a new syscall
to the sysent
array and to do this you need to be root; I've no time to work on this 
vulnerability,
but i think another way could be found. However it could give serious problems
(e.g. kernel
crashes).

A good patch could be a dinamyc memory allocation for args, but it's not
a good solution
in order to mantain a well performanced system; another one could be a strongest
check, but
it's not a good solution in order to set a good flexibility.

You would get an attach containing proof of concept code (4.x, 5.x/6.0 
versions).


greetings

rookie


P.S: in order to try the code, compile and link module to the kernel, later
do 'make test' and start ./poc