Date: Mon, 1 May 2000 21:27:18 +0100
From: Chris Evans <[email protected]>
To: [email protected]Subject: Linux knfsd DoS issue
Hi,
I was going to wait until a 2.2 kernel incorporating the fix was released,
but that's taking a while, and details have already leaked by several
routes.
ISSUE
=====
A DoS condition exists in the Linux kernel knfsd server. Remote,
unauthenticated users (i.e. those with neither a directory mounted nor
permission to mount one) can OOPS the host kernel. The OOPS does not bring
down the target host, but it is possible to render the NFS service
inoperable until a reboot.
FIX
===
The fix is present in the following kernels
- 2.3.99pre7-pre1
- 2.2.15pre20 [1]
- The kernel update recently release by RedHat (AFAIK)
Fix is also included at the bottom of this post.
[1] Which, AFAIK, is only available as a patch on top of
2.2.15pre19. Please see
ftp://ftp.kernel.org/pub/linux/kernel/people/alan/pre-patch-2.2.15-19to20.bz2
DISCUSSION OF ISSUE
This issue is caused by signed/unsigned issues. A size check is performed
on an "int" (i.e. signed) variable. This size check does not check for
negative values. As a consequence it is possible to point an internal
buffer pointer _way_ before the beginning of the buffer. The kernel will
try to read a memory address in limbo, and OOPS.
The faulty code is in kernel/net/sunrpc/svcauth.c, svcauth_unix()
int slen;
...
slen = ntohl(*bufp++); /* gids length */
if (slen > 16 || (len -= slen + 2) < 0)
goto badcred;
...
bufp += (slen - i);
As a general note, signed/unsigned errors like the above are _almost_ as
prevalent as buffer overflows. Once you start looking for
them. Re-auditing previously audited software with signed/unsigned issues
in mind often yields new problems.
Note that the userland sunrpc implementation suffered from a very similar
flaw a while back. No, it certainly wasn't a copy/paste error.
Some signed/unsigned errors can lead to exploitable conditions rather than
just DoS. The following code fragment should illustrate this:
func()
{
char name[NAMELEN];
int len = get_from_network();
if (len > NAMELEN)
goto i_dont_think_so;
memcpy(name, get_from_network(), len); /* -1 implicitly converted to
4Gb */
}
An interesting variant on this (that HAS been observed) is
func()
{
char* bufp;
char* bufmax = bufp + 100; /* Or whatever */
unsigned int len = get_from_network();
if (bufp + len > bufmax)
goto i_dont_think_so;
...
bufp += len;
}
On machines with a small 4Gb address space, e.g. Intel x86, a suitable
large value of len will wrap bufp around the address space, leading to a
position outside of the buffer boundary, but satisfying the "bufp + len <=
bufmax" constraint.
Programs which have in their history been found to include signed/unsigned
issues include
- knfsd
- sunrpc part of glibc
- xfs
- libORBIT
- issues at syscall API of various kernels
- issues in networking stacks
PATCH TO FIX
============
--- net/sunrpc/svcauth.c.old Tue Apr 18 05:13:47 2000
+++ net/sunrpc/svcauth.c Tue Apr 18 06:36:20 2000
@@ -4,6 +4,9 @@
* The generic interface for RPC authentication on the server side.
*
* Copyright (C) 1995, 1996 Olaf Kirch <[email protected]>
+ *
+ * CHANGES
+ * 19-Apr-2000 Chris Evans - Security fix
*/
#include <linux/types.h>
@@ -117,7 +120,8 @@
struct svc_buf *resp = &rqstp->rq_resbuf;
struct svc_cred *cred = &rqstp->rq_cred;
u32 *bufp = argp->buf;
- int len = argp->len, slen, i;
+ int len = argp->len;
+ u32 slen, i;
if ((len -= 3) < 0) {
*statp = rpc_garbage_args;
@@ -127,7 +131,7 @@
bufp++; /* length */
bufp++; /* time stamp */
slen = (ntohl(*bufp++) + 3) >> 2; /* machname length */
- if (slen > 64 || (len -= slen) < 0)
+ if (slen > 64 || (len -= slen + 3) < 0)
goto badcred;
bufp += slen; /* skip machname */
EXPLOIT
=======
Not today! In fact, don't bother. DoS is very passe... very 90's :-)
Cheers
Chris
