glibc syslog() Heap-Based Buffer Overflow Exploit

CVE-2023-6246: Heap-based buffer overflow in the glibc's syslog()


========================================================================
Contents
========================================================================

Summary
Analysis
Proof of concept
Exploitation
Acknowledgments
Timeline


========================================================================
Summary
========================================================================

We discovered a heap-based buffer overflow in the GNU C Library's
__vsyslog_internal() function, which is called by both syslog() and
vsyslog(). This vulnerability was introduced in glibc 2.37 (in August
2022) by the following commit:

  https://sourceware.org/git?p=glibc.git;a=commit;h=52a5be0df411ef3ff45c10c7c308cb92993d15b1

and was also backported to glibc 2.36 because this commit was a fix for
another, minor vulnerability in __vsyslog_internal() (CVE-2022-39046, an
"uninitialized memory [read] from the heap"):

  https://sourceware.org/bugzilla/show_bug.cgi?id=29536

For example, we confirmed that Debian 12 and 13, Ubuntu 23.04 and 23.10,
and Fedora 37 to 39 are vulnerable to this buffer overflow. Furthermore,
we successfully exploited an up-to-date, default installation of Fedora
38 (on amd64): a Local Privilege Escalation, from any unprivileged user
to full root. Other distributions are probably also exploitable.

To the best of our knowledge, this vulnerability cannot be triggered
remotely in any likely scenario (because it requires an argv[0], or an
openlog() ident argument, longer than 1024 bytes to be triggered).

Last-minute note: in December 1997 Solar Designer published information
about a very similar vulnerability in the vsyslog() of the old Linux
libc (https://insecure.org/sploits/linux.libc.5.4.38.vsyslog.html).


========================================================================
Analysis
========================================================================

In the glibc, both syslog() and vsyslog() call the vulnerable function
__vsyslog_internal():

------------------------------------------------------------------------
122 __vsyslog_internal (int pri, const char *fmt, va_list ap,
123                     unsigned int mode_flags)
124 {
125   /* Try to use a static buffer as an optimization.  */
126   char bufs[1024];
127   char *buf = NULL;
128   size_t bufsize = 0;
...
171 #define SYSLOG_HEADER(__pri, __timestamp, __msgoff, pid) \
172   "<%d>%s%n%s%s%.0d%s: ",                                \
173   __pri, __timestamp, __msgoff,                          \
174   LogTag == NULL ? __progname : LogTag,                  \
175   "[" + (pid == 0), pid, "]" + (pid == 0)
...
182     l = __snprintf (bufs, sizeof bufs,
183                     SYSLOG_HEADER (pri, timestamp, &msgoff, pid));
...
187   if (0 <= l && l < sizeof bufs)
188     {
...
202     }
203 
204   if (buf == NULL)
205     {
206       buf = malloc ((bufsize + 1) * sizeof (char));
...
213             __snprintf (buf, l + 1,
214                         SYSLOG_HEADER (pri, timestamp, &msgoff, pid));
...
221           __vsnprintf_internal (buf + l, bufsize - l + 1, fmt, apc,
222                                 mode_flags);
------------------------------------------------------------------------

- at lines 182-183, SYSLOG_HEADER() includes __progname (the basename()
  of argv[0]) if LogTag is NULL (e.g., if openlog() was not called, or
  called with a NULL ident argument);

- because a local attacker fully controls argv[0] and hence __progname
  (even when executing a SUID-root program such as su), at line 187 l
  (the return value of __snprintf()) can be larger than sizeof bufs
  (1024), in which case the code block at lines 188-202 is skipped;

- consequently, at line 203 buf is still NULL and bufsize is still 0,
  and at line 206 a very small 1-byte buf is malloc()ated (because
  bufsize is 0);

- at lines 213-214 this small buf is overflowed with the attacker-
  controlled __progname (because l is larger than 1024), and at lines
  221-222 this small buf is further overflowed (because bufsize - l + 1
  is 0 - l + 1, a very large size_t).


========================================================================
Proof of concept
========================================================================

$ (exec -a "`printf '%0128000x' 1`" /usr/bin/su < /dev/null)
Password: Segmentation fault (core dumped)


========================================================================
Exploitation
========================================================================

We decided to exploit this vulnerability through su (the most common
SUID-root program) on Fedora 38. To authenticate a user, su calls the
PAM library, and if the password provided by the user is incorrect, then
PAM calls the glibc's syslog() function without calling openlog() first,
thus allowing us to trigger the buffer overflow in __vsyslog_internal():

------------------------------------------------------------------------
782                         pam_syslog(pamh, LOG_NOTICE,
783                                  "authentication failure; "
784                                  "logname=%s uid=%d euid=%d "
785                                  "tty=%s ruser=%s rhost=%s "
786                                  "%s%s",
787                                  new->name, new->uid, new->euid,
788                                  tty ? (const char *)tty : "",
789                                  ruser ? (const char *)ruser : "",
790                                  rhost ? (const char *)rhost : "",
791                                  (new->user && new->user[0] != '\0')
792                                   ? " user=" : "",
793                                  new->user
794                         );
------------------------------------------------------------------------
107 pam_syslog (const pam_handle_t *pamh, int priority,
108             const char *fmt, ...)
109 {
...
113   pam_vsyslog (pamh, priority, fmt, args);
------------------------------------------------------------------------
 73 pam_vsyslog (const pam_handle_t *pamh, int priority,
 74              const char *fmt, va_list args)
 75 {
 ..
 81       if (asprintf (&msgbuf1, "%s(%s:%s):", pamh->mod_name,
 82                     pamh->service_name?pamh->service_name:"<unknown>",
 83                     _pam_choice2str (pamh->choice)) < 0)
 ..
 91   if (vasprintf (&msgbuf2, fmt, args) < 0)
 ..
 99   syslog (LOG_AUTHPRIV|priority, "%s %s",
100           (msgbuf1 ? msgbuf1 : _PAM_SYSTEM_LOG_PREFIX), msgbuf2);
------------------------------------------------------------------------

But what should we overwrite in the heap to successfully exploit this
buffer overflow? Initially, because su calls setlocale(LC_ALL, ""); at
the very beginning of its su_main() function, we tried to reuse the key
idea from our Baron Samedit exploits (CVE-2021-3156 in Sudo): we wrote a
rudimentary fuzzer to execute su with a random argv[0] and random locale
environment variables and automatically inspect the resulting crashes in
gdb. Unfortunately this fuzzer failed to produce interesting results: we
only obtained a handful of unique crashes, and they did not look very
promising.

However, we did not investigate the reasons for this failure, because
while browsing through su's source code we noticed that su_main() calls
env_whitelist_from_string() to parse the argument of the -w command-line
option:

------------------------------------------------------------------------
1118                 case 'w':
1119                         env_whitelist_from_string(su, optarg);
1120                         break;
------------------------------------------------------------------------
 692 static int env_whitelist_from_string(struct su_context *su, const char *str)
 693 {
 694         char **all = strv_split(str, ",");
 ...
 703         STRV_FOREACH(one, all)
 704                 env_whitelist_add(su, *one);
 705         strv_free(all);
 706         return 0;
 707 }
------------------------------------------------------------------------
 662 static int env_whitelist_add(struct su_context *su, const char *name)
 663 {
 664         const char *env = getenv(name);
 665 
 666         if (!env)
 667                 return 1;
 668         if (strv_extend(&su->env_whitelist_names, name))
 669                 err_oom();
 670         if (strv_extend(&su->env_whitelist_vals, env))
 671                 err_oom();
 672         return 0;
 673 }
------------------------------------------------------------------------

Conveniently, env_whitelist_from_string() allows us (attackers) to
malloc()ate and free() an arbitrary number of arbitrary strings at the
very beginning of su's execution: an almost perfect heap feng shui. We
therefore rewrote our fuzzer to execute su with a random argv[0] and a
random whitelist option (instead of random locale environment variables)
and immediately observed numerous unique crashes; among these, three in
particular caught our attention.


========================================================================
1/ Corruption of PAM structures
========================================================================

Surprisingly, our fuzzer directly overwrote two PAM function pointers
(in struct pam_data and struct handler):

------------------------------------------------------------------------
Thread 2.1 "su" received signal SIGSEGV, Segmentation fault.
0x00007fa7d3b0e3ac in _pam_free_data (status=7, pamh=0x56211242ec10) at /usr/src/debug/pam-1.5.2-16.fc38.x86_64/libpam/pam_data.c:161
161                 last->cleanup(pamh, last->data, status);
...
=> 0x7fa7d3b0e3ac <pam_end+92>: call   *%rax
rax            0x4141414141414141  4702111234474983745
------------------------------------------------------------------------
Thread 2.1 "su" received signal SIGSEGV, Segmentation fault.
0x00007f928b5e5781 in _pam_dispatch_aux (use_cached_chain=<optimized out>, resumed=<optimized out>, h=0x55f2e374aae0, flags=0, pamh=0x55f2e374aae0) at /usr/src/debug/pam-1.5.2-16.fc38.x86_64/libpam/pam_dispatch.c:110
110                 retval = h->func(pamh, flags, h->argc, h->argv);
...
=> 0x7f928b5e5781 <_pam_dispatch+465>:  call   *%rax
rax            0x4545454545454545  4991471925827290437
------------------------------------------------------------------------

Although this sounds exciting at first (a call to 0x4141414141414141!)
we decided to not pursue this avenue of exploitation:

- we cannot overwrite such a function pointer with null bytes (because
  we overflow __vsyslog_internal()'s buffer with a null-terminated
  string), but userland addresses contain at least two null bytes;

- we could try to partially overwrite such a function pointer, but we do
  not control the end of the string that overflows __vsyslog_internal()'s
  buffer (the end of the aforementioned pam_syslog() format string), and
  such an uncontrolled, partially overwritten function pointer is very
  unlikely to miraculously point to a useful ROP gadget.


========================================================================
2/ Corruption of heap metadata
========================================================================

Unsurprisingly, our fuzzer also overwrote various pieces of heap
metadata (chunk headers managed internally by the glibc's malloc), and
therefore triggered all kinds of assertion failures and security checks:

------------------------------------------------------------------------
$ grep -A1 __libc_message fuzzer.out | cut -d'"' -f2 | sort -u
...
chunk_main_arena (bck->bk)
chunk_main_arena (fwd)
corrupted double-linked list
corrupted double-linked list (not small)
corrupted size vs. prev_size
corrupted size vs. prev_size in fastbins
double free or corruption (out)
free(): corrupted unsorted chunks
free(): invalid next size (fast)
free(): invalid pointer
free(): invalid size
malloc_consolidate(): invalid chunk size
malloc(): corrupted top size
malloc(): invalid size (unsorted)
malloc(): smallbin double linked list corrupted
malloc(): unaligned tcache chunk detected
malloc(): unsorted double linked list corrupted
munmap_chunk(): invalid pointer
------------------------------------------------------------------------

Although some of these corruptions might be exploitable, we decided to
not pursue this avenue of exploitation either:

- we cannot overwrite a chunk header with a size field and an fd or bk
  pointer that are both valid (they must both contain null bytes to be
  valid), which severely limits our exploitation options;

- in any case, we would probably need a specific heap, mmap, or stack
  address to exploit such a corruption, but we do not have the luxury of
  an information leak, and all these addresses are too heavily
  randomized by ASLR to be brute forced.


========================================================================
3/ Corruption of nss structures
========================================================================

Our fuzzer also produced two crashes that immediately caught our
attention because they are directly related to one of the techniques
that we used to exploit Baron Samedit:

------------------------------------------------------------------------
Thread 2.1 "su" received signal SIGSEGV, Segmentation fault.
__GI___nss_lookup (ni=ni@entry=0x7ffe876a05e8, fct_name=fct_name@entry=0x7fbba214e4e7 "getpwnam_r", fct2_name=fct2_name@entry=0x0, fctp=fctp@entry=0x7ffe876a05f0) at nsswitch.c:67
67        *fctp = __nss_lookup_function (*ni, fct_name);
...
=> 0x7fbba20ec50e <__GI___nss_lookup+30>:       mov    (%rax),%rdi
rax            0x4141414141414141  4702111234474983745
------------------------------------------------------------------------
Thread 2.1 "su" received signal SIGSEGV, Segmentation fault.
__nss_module_get_function (module=0x4141414141414141, name=name@entry=0x7f0aed9034e7 "getpwnam_r") at nss_module.c:328
328       if (!__nss_module_load (module))
...
=> 0x7f0aed8a34b7 <__nss_module_get_function+39>:       mov    (%rdi),%eax
rdi            0x4141414141414141  4702111234474983745
------------------------------------------------------------------------

As discussed in the "2/ struct service_user overwrite" subsection of our
Baron Samedit advisory, if we overwrite the name[] field of a heap-based
struct nss_module with a string of characters that contains a slash (for
example "A/B/C"), then at lines 180-181 the name of a shared library is
constructed ("libnss_A/B/C.so.2"), and at line 187 this shared library
is loaded from our current working directory (because its name contains
a slash, but does not start with a slash) and executed as root (because
su is a SUID-root program):

------------------------------------------------------------------------
170 module_load (struct nss_module *module)
171 {
...
180     if (__asprintf (&shlib_name, "libnss_%s.so%s",
181                     module->name, __nss_shlib_revision) < 0)
...
187     handle = __libc_dlopen (shlib_name);
------------------------------------------------------------------------

Unfortunately, the __progname part (which we control) of the string that
overflows __vsyslog_internal()'s buffer cannot contain a slash (because
__progname is the basename() of argv[0]). Luckily, however, the part of
the overflowing string that we do not control (the pam_syslog() format
string) includes the absolute path of our tty, which contains a slash.
For example, if:

- our tty is /dev/pts/23 (we use forkpty() in our exploit);

- our unprivileged local user is nobody (uid 65534);

- the argv[0] (and hence __progname) that we use to execute su is a long
  string of 'A' characters (longer than 1024);

then we can overwrite the name[] field of a heap-based struct nss_module
with a string of the form:

  "AAAAAAAAAA: pam_unix(su:auth): authentication failure; logname= uid=65534 euid=0 tty=/dev/pts/23 ruser=nobody rhost=  user=root"

Consequently, if we first create the following three directories (in our
current working directory):

  "libnss_AAAAAAAAAA: pam_unix(su:auth): authentication failure; logname= uid=65534 euid=0 tty="
  "libnss_AAAAAAAAAA: pam_unix(su:auth): authentication failure; logname= uid=65534 euid=0 tty=/dev"
  "libnss_AAAAAAAAAA: pam_unix(su:auth): authentication failure; logname= uid=65534 euid=0 tty=/dev/pts"

and also create the following shared library (in our current working
directory):

  "libnss_AAAAAAAAAA: pam_unix(su:auth): authentication failure; logname= uid=65534 euid=0 tty=/dev/pts/23 ruser=nobody rhost=  user=root.so.2"

then this shared library will eventually be loaded and executed with
full root privileges. In our tests, it takes a few 10,000s of tries to
successfully brute force the exploit parameters (the length of argv[0],
and the whitelist option and its associated environment variables).

Note: this exploit could certainly be made much more efficient; in
theory, it could even be a one-shot exploit, because we do not need to
brute force the ASLR, only the heap layout.