Re: net/sctp: out-of-bounds access in sctp_add_bind_addr

On Mon, Jan 25, 2016 at 03:42:14PM +0100, Dmitry Vyukov wrote:

On Mon, Jan 25, 2016 at 3:31 PM, Neil Horman [off-list ref] wrote:

quoted

On Mon, Jan 25, 2016 at 03:02:38PM +0100, Dmitry Vyukov wrote:

quoted

Hello,

I've git the following error report while running syzkaller fuzzer:

==================================================================
BUG: KASAN: slab-out-of-bounds in memcpy+0x1d/0x40 at addr ffff88006c6361e8
Read of size 28 by task syz-executor/12551
=============================================================================
BUG kmalloc-16 (Not tainted): kasan: bad access detected
-----------------------------------------------------------------------------

INFO: Allocated in sctp_setsockopt_bindx+0xd2/0x3e0 age=12 cpu=2 pid=12551
[<     inline     >] kmalloc include/linux/slab.h:468
[<      none      >] sctp_setsockopt_bindx+0xd2/0x3e0 net/sctp/socket.c:975
[<      none      >] sctp_setsockopt+0x1493/0x3630 net/sctp/socket.c:3711
[<      none      >] sock_common_setsockopt+0x97/0xd0 net/core/sock.c:2620
[<     inline     >] SYSC_setsockopt net/socket.c:1752
[<      none      >] SyS_setsockopt+0x15b/0x250 net/socket.c:1731
[<      none      >] entry_SYSCALL_64_fastpath+0x16/0x7a
arch/x86/entry/entry_64.S:185

INFO: Slab 0xffffea0001b18d80 objects=16 used=4 fp=0xffff88006c6376e0
flags=0x5fffc0000004080
INFO: Object 0xffff88006c6361e8 @offset=488 fp=0x0000000000000002
Bytes b4 ffff88006c6361d8: 00 00 00 00 00 00 00 00 2f 98 34 88 ff ff
ff ff  ......../.4.....
Object ffff88006c6361e8: 02 00 00 00 00 00 00 00 02 00 ab 07 7f 00 00
01  ................
CPU: 2 PID: 12551 Comm: syz-executor Tainted: G    B           4.5.0-rc1+ #278
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
 00000000ffffffff ffff880036397928 ffffffff8299a02d ffff88003e807900
 ffff88006c6361e8 ffff88006c636000 ffff880036397958 ffffffff81752814
 ffff88003e807900 ffffea0001b18d80 ffff88006c6361e8 ffff88006c6361e8

Call Trace:
 [<ffffffff8175ad54>] __asan_loadN+0x124/0x1a0 mm/kasan/kasan.c:512
 [<ffffffff8175b2dd>] memcpy+0x1d/0x40 mm/kasan/kasan.c:297
 [<ffffffff85dcb249>] sctp_add_bind_addr+0xa9/0x270 net/sctp/bind_addr.c:162
 [<ffffffff85dcfd66>] sctp_do_bind+0x336/0x580 net/sctp/socket.c:389
 [<ffffffff85dd16ec>] sctp_bindx_add+0xac/0x1a0 net/sctp/socket.c:471
 [<ffffffff85dd5cc8>] sctp_setsockopt_bindx+0x2f8/0x3e0 net/sctp/socket.c:1010
 [<ffffffff85dde283>] sctp_setsockopt+0x1493/0x3630 net/sctp/socket.c:3711
 [<ffffffff851f5ae7>] sock_common_setsockopt+0x97/0xd0 net/core/sock.c:2620
 [<     inline     >] SYSC_setsockopt net/socket.c:1752
 [<ffffffff851f2c3b>] SyS_setsockopt+0x15b/0x250 net/socket.c:1731
 [<ffffffff863595f6>] entry_SYSCALL_64_fastpath+0x16/0x7a
arch/x86/entry/entry_64.S:185

Memory state around the buggy address:
 ffff88006c636080: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
 ffff88006c636100: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc

quoted

ffff88006c636180: fc fc fc fc fc fc fc fc fc fc fc fc fc 00 00 fc

                                                                ^
 ffff88006c636200: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
 ffff88006c636280: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
==================================================================


sctp_setsockopt_bindx verifies that the user-passed address has valid
len for the specified family, but then sctp_add_bind_addr copies whole
sctp_addr from there. This causes heap out-of-bounds access and can
crash kernel. Not sure if it is possible to copy out the trailing
garbage to user-space later.

It does more than that though.  sctp_setsockopt_bindx checks the following:
1) That passed addr_size is greater than zero
2) that the entire range of memory between addrs and addrs+addr_size is readable
3) That at least one address structure worth of data is available (implicit in
the while (walk_size < addr_size) loop).

Could one of the sockaddr_len fields in one of the addresses have been mangled
so that it appeared shorter in the the while loop from (3), so that a copy of
sizeof(sctp_addr in sctp_add_bind_addr overrun the allocated memory?

I may be missing something, but what I see is:

1. we check that there is at least family:
if (walk_size + sizeof(sa_family_t) > addrs_size) {

2. get family descriptor:
af = sctp_get_af_specific(sa_addr->sa_family);

3. check that the address size is enough to hold the declared family:
if (!af || (walk_size + af->sockaddr_len) > addrs_size) {

4. then we do sctp_add_bind_addr, which copies whole sctp_addr from addr:

int sctp_add_bind_addr(struct sctp_bind_addr *bp, union sctp_addr *new,
...
memcpy(&addr->a, new, sizeof(*new));

Now imagine that the addr is ipv4 (16 or so bytes, that's what we
checked) and we copy 28 bytes (ipv6) from addr.

Yes, that's pretty much it I think. That memcpy should be limited to
af->sockaddr_len, it's just that af is not readily available in that
function.

  Marcelo