On Tue, Oct 31, 2023 at 09:04:29AM -0500, Eric Blake wrote:
Another round of fuzz testing revealed that when a server negotiates
extended headers and replies with a 64-bit flag value where the client
used the 32-bit API command, we were correctly flagging the server's
response as being an EOVERFLOW condition, but then immediately failing
in an assertion failure instead of reporting it to the application.
The following one-byte change to qemu.git at commit fd9a38fd43 allows
the creation of an intentionally malicious server:
| diff --git i/nbd/server.c w/nbd/server.c
| index 859c163d19f..32e1e771a95 100644
| --- i/nbd/server.c
| +++ w/nbd/server.c
| @@ -2178,7 +2178,7 @@ static void nbd_extent_array_convert_to_be(NBDExtentArray *ea)
|
| for (i = 0; i < ea->count; i++) {
| ea->extents[i].length = cpu_to_be64(ea->extents[i].length);
| - ea->extents[i].flags = cpu_to_be64(ea->extents[i].flags);
| + ea->extents[i].flags = ~cpu_to_be64(ea->extents[i].flags);
| }
| }
and can then be detected with the following command line:
$ nbdsh -c - <<\EOF
> def f(a,b,c,d):
> pass
>
> h.connect_systemd_socket_activation(["/path/to/bad/qemu-nbd",
> "-r", "-f", "raw", "TODO"])
> h.block_staus(h.get_size(), 0, f)
> EOF
nbdsh: generator/states-reply-chunk.c:626: enter_STATE_REPLY_CHUNK_REPLY_RECV_BS_ENTRIES:
Assertion `(len | flags) <= UINT32_MAX' failed.
Aborted (core dumped)
whereas a fixed libnbd will give:
nbdsh: command line script failed: nbd_block_status: block-status: command failed: Value
too large for defined data type
We can either relax the assertion (by changing to 'assert ((len |
flags) <= UINT32_MAX || cmd->error)'), or intentionally truncate flags
to make the existing assertion reliable. This patch goes with the
latter approach.
Sadly, this crash is possible in all existing 1.18.x stable releases,
if they were built with assertions enabled (most distros do this by
default), meaning a malicious server has an easy way to cause a Denial
of Service attack by triggering the assertion failure in vulnerable
clients, so we have assigned this CVE-2023-5871. Mitigating factors:
the crash only happens for a server that sends a 64-bit status block
reply (no known production servers do so; qemu 8.2 will be the first
known server to support extended headers, but it is not yet released);
and as usual, a client can use TLS to guarantee it is connecting only
to a known-safe server. If libnbd is compiled without assertions,
there is no crash or other mistaken behavior; and when assertions are
enabled, the attacker cannot accomplish anything more than a denial of
service.
Reported-by: Richard W.M. Jones <rjones(a)redhat.com>
Fixes: 20dadb0e10 ("generator: Prepare for extent64 callback", v1.17.4)
Signed-off-by: Eric Blake <eblake(a)redhat.com>
---
I'll send a followup mail patching docs/libnbd-security.pod once we
have a stable archive link (my understanding is that we will not have
stable hyperkitty URLs until all messages have been migrated from
libguestfs(a)redhat.com to guestfs(a)lists.libguestfs.org).
generator/states-reply-chunk.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/generator/states-reply-chunk.c b/generator/states-reply-chunk.c
index 5a31c192..8ab7e8ba 100644
--- a/generator/states-reply-chunk.c
+++ b/generator/states-reply-chunk.c
@@ -600,6 +600,7 @@ REPLY.CHUNK_REPLY.RECV_BS_ENTRIES:
break; /* Skip this and later extents; we already made progress */
/* Expose this extent as an error; we made no progress */
cmd->error = cmd->error ? : EOVERFLOW;
+ flags = (uint32_t)flags;
}
}
Reviewed-by: Richard W.M. Jones <rjones(a)redhat.com>
However please remember to update docs/libnbd-security.pod :-)
Rich.
--
Richard Jones, Virtualization Group, Red Hat
http://people.redhat.com/~rjones
Read my programming and virtualization blog:
http://rwmj.wordpress.com
virt-builder quickly builds VMs from scratch
http://libguestfs.org/virt-builder.1.html