bad bignum encoding for curve25519-sha256@libssh.org

Hi,

So I screwed up when writing the support for the curve25519 KEX method
that doesn't depend on OpenSSL's BIGNUM type - a bug in my code left
leading zero bytes where they should have been skipped. The impact of
this is that OpenSSH 6.5 and 6.6 will fail during key exchange with a
peer that implements curve25519-sha256@libssh.org properly about 0.2%
of the time (one in every 512ish connections).

We've fixed this for OpenSSH 6.7 by avoiding the curve25519-sha256
key exchange for previous versions, but I'd recommend distributors
of OpenSSH apply this patch so the affected code doesn't become
too entrenched in LTS releases.

The patch fixes the bug and makes OpenSSH identify itself as 6.6.1 so as
to distinguish itself from the incorrect versions so the compatibility
code to disable the affected KEX isn't activated.

I've committed this on the 6.6 branch too.

Apologies for the hassle.

-d

Origin: upstream, https://lists.mindrot.org/pipermail/openssh-unix-dev/2014-April/032494.html
Forwarded: not-needed
Last-Update: 2014-04-21

Patch-Name: curve25519-sha256-bignum-encoding.patch

1 files changed, 4 insertions, 1 deletions

diff --git a/bufaux.c b/bufaux.c
index e24b5fc0a..f6a6f2ab2 100644
--- a/bufaux.c
+++ b/bufaux.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: bufaux.c,v 1.56 2014/02/02 03:44:31 djm Exp $ */
+/* $OpenBSD: bufaux.c,v 1.57 2014/04/16 23:22:45 djm Exp $ */
 /*
  * Author: Tatu Ylonen <ylo@cs.hut.fi>
  * Copyright (c) 1995 Tatu Ylonen <ylo@cs.hut.fi>, Espoo, Finland
@@ -372,6 +372,9 @@ buffer_put_bignum2_from_string(Buffer *buffer, const u_char *s, u_int l)
 
         if (l > 8 * 1024)
                 fatal("%s: length %u too long", __func__, l);
+        /* Skip leading zero bytes */
+        for (; l > 0 && *s == 0; l--, s++)
+                ;
         p = buf = xmalloc(l + 1);
         /*
          * If most significant bit is set then prepend a zero byte to