Skip to contentSkip to conversation

Fix port-share option with TLS-Crypt v2

PATCH Superseded Arne Schwabe 3 participants 5 messages Nov 20, 2020 – Dec 4, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix port-share option with TLS-Crypt v2

The port-share option assumed that all openvpn initial reset packets are between 14 and 255 bytes long. This is not true for tls-crypt-v2.

---
 src/openvpn/ps.c | 34 ++++++++++++++++++++++++++++++----
 1 file changed, 30 insertions(+), 4 deletions(-)
src/openvpn/ps.c +30 -4
index 2089e6b9..cd00bc23 100644
@@ -983,10 +983,36 @@ is_openvpn_protocol(const struct buffer *buf)
     const int len = BLEN(buf);
     if (len >= 3)
     {
-        return p[0] == 0
-               && p[1] >= 14
-               && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT)
-                   || p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT));
+        if (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT))
+        {
+            /* WKc is at least 306 byte (not including metadata):
+             *
+             * 16 bit len + 256 bit HMAC + 128 bit IV + 2048 bit Kc = 2448 bit
+             *
+             * This is increased by the normal length of client handshake +
+             * tls-crypt overhead (32)
+             *
+             * For metadata tls-crypt-v2.txt does not explicitly specify
+             * an upper limit but we also have TLS_CRYPT_V2_MAX_WKC_LEN
+             * as 1024 bytes. We err on the safe side with 255 extra overhead
+             *
+             * We don't do the 2 byte check for tls-crypt-v2 because it is very
+             * unrealistic to have only 2 bytes available.
+             */
+            int plen = (p[0] << 8) | p[1];
+
+            return  (plen >= 352 && plen < (1024 + 255));
+        }
+        else
+        {
+            /* first two bytes are the size field. This is an openvpn packet
+             * between 14 bytes and 255 (otherwise p[1] would be 01 or larger.
+             * on the hard reset packet key_id is zero, so only the opcode
+             * part is non-null */
+            return p[0] == 0
+                && p[1] >= 14
+                && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT));
+        }
     }
     else if (len >= 2)
     {
Signature · 1 line
2.26.2
#1 lore raw

Hi,

Quoted 2 lines · Arne Schwabe

The port-share option assumed that all openvpn initial reset packets are between 14 and 255 bytes long. This is not true for tls-crypt-v2.

Good catch, I totally missed this in the tls-crypt-v2 patch set.

Quoted 11 lines
---
 src/openvpn/ps.c | 34 ++++++++++++++++++++++++++++++----
 1 file changed, 30 insertions(+), 4 deletions(-)
src/openvpn/ps.c +0 -0
index 2089e6b9..cd00bc23 100644
@@ -983,10 +983,36 @@ is_openvpn_protocol(const struct buffer *buf)
     const int len = BLEN(buf);
     if (len >= 3)

src/openvpn/ps.c is_openvpn_protocol(const struct buffer *buf) · line 990

Quoted 10 lines
src/openvpn/ps.c +5 -4
     {
-        return p[0] == 0
-               && p[1] >= 14
-               && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT)
-                   || p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT));
+        if (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT))
+        {
+            /* WKc is at least 306 byte (not including metadata):
+             *
+             * 16 bit len + 256 bit HMAC + 128 bit IV + 2048 bit Kc = 2448 bit

The IV is part of the HMAC, so does not need to be counted here. So this this be 2320 bits / 290 bytes.

Quoted 11 lines
src/openvpn/ps.c +11 -0
+             *
+             * This is increased by the normal length of client handshake +
+             * tls-crypt overhead (32)
+             *
+             * For metadata tls-crypt-v2.txt does not explicitly specify
+             * an upper limit but we also have TLS_CRYPT_V2_MAX_WKC_LEN
+             * as 1024 bytes. We err on the safe side with 255 extra overhead
+             *
+             * We don't do the 2 byte check for tls-crypt-v2 because it is very
+             * unrealistic to have only 2 bytes available.
+             */

src/openvpn/ps.c is_openvpn_protocol(const struct buffer *buf) · line 1011

Quoted 10 lines
src/openvpn/ps.c +10 -0
+            int plen = (p[0] << 8) | p[1];
+
+            return  (plen >= 352 && plen < (1024 + 255));
+        }
+        else
+        {
+            /* first two bytes are the size field. This is an openvpn packet
+             * between 14 bytes and 255 (otherwise p[1] would be 01 or larger.
+             * on the hard reset packet key_id is zero, so only the opcode
+             * part is non-null */

Instead of explaining the length field in the comment, why not just move the "int plen = (p[0] << 8) | p[1];" outside the if to have the code explain that by itself?

src/openvpn/ps.c is_openvpn_protocol(const struct buffer *buf) · line 1018

Quoted 8 lines
src/openvpn/ps.c +4 -0
+            return p[0] == 0
+                && p[1] >= 14
+                && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT));
+        }
     }
     else if (len >= 2)
     {

Thanks,
-Steffan

#2 lore raw

PATCH v2 1/1 Fix port-share option with TLS-Crypt v2

In reply to Arne Schwabe

The port-share option assumed that all openvpn initial reset packets are between 14 and 255 bytes long. This is not true for tls-crypt-v2.

Patch V2: use correct length for TLS-Crypt v2, use length variable
non-tlscryptv2 test

---
 src/openvpn/ps.c | 34 +++++++++++++++++++++++++++++-----
 1 file changed, 29 insertions(+), 5 deletions(-)
src/openvpn/ps.c +29 -5
index 2089e6b9..5d760782 100644
@@ -983,14 +983,38 @@ is_openvpn_protocol(const struct buffer *buf)
     const int len = BLEN(buf);
     if (len >= 3)
     {
-        return p[0] == 0
-               && p[1] >= 14
-               && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT)
-                   || p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT));
+        int plen = (p[0] << 8) | p[1];
+
+        if (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT))
+        {
+            /* WKc is at least 290 byte (not including metadata):
+             *
+             * 16 bit len + 256 bit HMAC + 2048 bit Kc = 2320 bit
+             *
+             * This is increased by the normal length of client handshake +
+             * tls-crypt overhead (32)
+             *
+             * For metadata tls-crypt-v2.txt does not explicitly specify
+             * an upper limit but we also have TLS_CRYPT_V2_MAX_WKC_LEN
+             * as 1024 bytes. We err on the safe side with 255 extra overhead
+             *
+             * We don't do the 2 byte check for tls-crypt-v2 because it is very
+             * unrealistic to have only 2 bytes available.
+             */
+            return  (plen >= 336 && plen < (1024 + 255));
+        }
+        else
+        {
+            /* For non tls-crypt2 we assume the packet length to valid between
+             * 14 and 255 */
+            return plen >= 14 && plen <= 255
+                   && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT));
+        }
     }
     else if (len >= 2)
     {
-        return p[0] == 0 && p[1] >= 14;
+        int plen = (p[0] << 8) | p[1];
+        return plen >= 14 && plen <= 255;
     }
     else
     {
Signature · 1 line
2.26.2
Acked-by on v2 1/1
#3 lore raw

Hi,

Quoted 50 lines · Arne Schwabe

The port-share option assumed that all openvpn initial reset packets are between 14 and 255 bytes long. This is not true for tls-crypt-v2.

Patch V2: use correct length for TLS-Crypt v2, use length variable
non-tlscryptv2 test

---
 src/openvpn/ps.c | 34 +++++++++++++++++++++++++++++-----
 1 file changed, 29 insertions(+), 5 deletions(-)
src/openvpn/ps.c +26 -4
index 2089e6b9..5d760782 100644
@@ -983,14 +983,38 @@ is_openvpn_protocol(const struct buffer *buf)
     const int len = BLEN(buf);
     if (len >= 3)
     {
-        return p[0] == 0
-               && p[1] >= 14
-               && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT)
-                   || p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT));
+        int plen = (p[0] << 8) | p[1];
+
+        if (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V3 << P_OPCODE_SHIFT))
+        {
+            /* WKc is at least 290 byte (not including metadata):
+             *
+             * 16 bit len + 256 bit HMAC + 2048 bit Kc = 2320 bit
+             *
+             * This is increased by the normal length of client handshake +
+             * tls-crypt overhead (32)
+             *
+             * For metadata tls-crypt-v2.txt does not explicitly specify
+             * an upper limit but we also have TLS_CRYPT_V2_MAX_WKC_LEN
+             * as 1024 bytes. We err on the safe side with 255 extra overhead
+             *
+             * We don't do the 2 byte check for tls-crypt-v2 because it is very
+             * unrealistic to have only 2 bytes available.
+             */
+            return  (plen >= 336 && plen < (1024 + 255));
+        }
+        else
+        {
+            /* For non tls-crypt2 we assume the packet length to valid between
+             * 14 and 255 */
+            return plen >= 14 && plen <= 255
+                   && (p[2] == (P_CONTROL_HARD_RESET_CLIENT_V2 << P_OPCODE_SHIFT));

src/openvpn/ps.c is_openvpn_protocol(const struct buffer *buf) · line 1020

Quoted 10 lines
src/openvpn/ps.c +3 -1
+        }
     }
     else if (len >= 2)
     {
-        return p[0] == 0 && p[1] >= 14;
+        int plen = (p[0] << 8) | p[1];
+        return plen >= 14 && plen <= 255;
     }
     else
     {

Thanks. This looks good to me now.

Ran some tests to check that this correctly resolves the issue when using --tls-crypt-v2 in combination with --port-share.

-Steffan

#4 lore raw
In reply to Arne Schwabe

Magic heuristics... but if Arne and Steffan agree that this is correct, who am I to dispute it :-) - it passes client tests and is not "obviously wrong".

Your patch has been applied to the master and release/2.5 branch.

commit 1387f52682dcd3789c56c9979ccedca281ff88f4 (master) commit c27f97dc0ac99344659cca57bd048543a9899266 (release/2.5)

Author: Arne Schwabe
Date:   Mon Nov 30 13:38:13 2020 +0100
     Fix port-share option with TLS-Crypt v2

     Signed-off-by: Arne Schwabe <[email protected]...>
     Acked-by: Steffan Karger <[email protected]...>
     Message-Id: <[email protected]...>
     URL: https://www.mail-archive.com/[email protected]/msg21290.html
     Signed-off-by: Gert Doering <[email protected]...>

--
kind regards,

Gert Doering