Issue with SCRAM-SHA Authorization

Wed Dec 1 23:43:55 UTC 2021

Hi,

I was debugging a bug report for my sieve editor implementation
concerning an issue with SCRAM-SHA Authorization.

But I am a bit unsure about my findings. It seems to be a server side
bug. But I am wondering if I missed something. What puzzles me a bit is
that SCRAM-SHA is out since ages and obviously no one ran into this
issue. Which feels odd.

When I try to connect to a dovecot server with enabled Authorization,
then SCRAM SHA1 or SHA256 fail with a channel binding error.
Authorization via SASL PLAIN works perfectly fine.

So I digged into the logs. The trace is below also my
findings/conclusion. The username and authorization in the trace are set
to "user" the password is "pencil".

Nov 24 14:00:06 auth: Debug: client in: AUTH    1       SCRAM-SHA-256
service=sieve   secured=tlssession=sI9yRonRZISsEQAB lip=172.17.0.2
rip=172.17.0.1  lport=4190      rport=33892
resp=bixhPXVzZXIsbj11c2VyLHI9ZmEzZGY1NjNlZjczZTY3YjQzZThhZWRiZjUzNGI1NWIxNzIwNzU3OThjZmJiMjExNmMyMTNkMWM4NDE4MmZkOQ==
(previous base64 data may contain sensitive data)
Nov 24 14:00:06 auth: Debug:
passwd-file(user,172.17.0.1,<sI9yRonRZISsEQAB>): Performing passdb lookup
Nov 24 14:00:06 auth: Debug:
passwd-file(user,172.17.0.1,<sI9yRonRZISsEQAB>): lookup: user=user
file=/etc/dovecot/users
Nov 24 14:00:06 auth: Debug:
passwd-file(user,172.17.0.1,<sI9yRonRZISsEQAB>): Generating
SCRAM-SHA-256 from user 'user', password 'pencil'
Nov 24 14:00:06 auth: Debug:
passwd-file(user,172.17.0.1,<sI9yRonRZISsEQAB>): Finished passdb lookup
Nov 24 14:00:06 auth: Debug: Credentials:
343039362c326b3934364452546c326431765a37434b6d556c32513d3d2c46444d364f7173577035464c446e394151666554316547585045427a47566d6965673842427331456d37633d2c79766e5a4169376250392f59706c4a71597872356a79794f7a6e4c664730667636485155325a7a6975376b3d
Nov 24 14:00:06 auth: Debug: client passdb out: CONT    1
cj1mYTNkZjU2M2VmNzNlNjdiNDNlOGFlZGJmNTM0YjU1YjE3MjA3NTc5OGNmYmIyMTE2YzIxM2QxYzg0MTgyZmQ5J2Q9NCskV2NuWEFVTnZxOU1EZENgM2UheVguXFw3O2JgKlg0X29gKDtBXnJrTi1xUmQ2IXgiQX5+eGU8ZlJHWSxzPTJrOTQ2RFJUbDJkMXZaN0NLbVVsMlE9PSxpPTQwOTY=
Nov 24 14:00:06 auth: Debug: client in: CONT    1
Yz1iaXhoUFhWelpYSXMscj1mYTNkZjU2M2VmNzNlNjdiNDNlOGFlZGJmNTM0YjU1YjE3MjA3NTc5OGNmYmIyMTE2YzIxM2QxYzg0MTgyZmQ5J2Q9NCskV2NuWEFVTnZxOU1EZENgM2UheVguXFw3O2JgKlg0X29gKDtBXnJrTi1xUmQ2IXgiQX5+eGU8ZlJHWSxwPUcvZGZWUUlGZy9NMkpSMlpEMXM0cms4MS9jTmFMa0tuMzZFYzFlTTN1enc9
(previous base64 data may contain sensitive data)
Nov 24 14:00:06 auth: Info:
scram-sha-256(user,172.17.0.1,<sI9yRonRZISsEQAB>): Invalid channel
binding data
Nov 24 14:00:06 auth: Debug: auth(user,172.17.0.1,<sI9yRonRZISsEQAB>):
Auth request finished
Nov 24 14:00:08 auth: Debug: client passdb out: FAIL    1       user=user
Nov 24 14:00:08 managesieve-login: Info: Disconnected: Aborted login by
logging out (auth failed, 1 attempts in 2 secs): user=<user>,
method=SCRAM-SHA-256, rip=172.17.0.1, lip=172.17.0.2, TLS,
session=<sI9yRonRZISsEQAB>
Nov 24 14:00:08 managesieve-login: Debug: SSL alert: close notify

The matching trace from the client, a sieve editor
(https://github.com/thsmi/sieve) is as follows:

[15:000:06.192 kw53213a-qc7kf99m1lp] Client -> Server:
AUTHENTICATE "SCRAM-SHA-256"
"bixhPXVzZXIsbj11c2VyLHI9ZmEzZGY1NjNlZjczZTY3YjQzZThhZWRiZjUzNGI1NWIxNzIwNzU3OThjZmJiMjExNmMyMTNkMWM4NDE4MmZkOQ=="
[15:000:06.202 kw53213a-qc7kf99m1lp] Server -> Client
"cj1mYTNkZjU2M2VmNzNlNjdiNDNlOGFlZGJmNTM0YjU1YjE3MjA3NTc5OGNmYmIyMTE2YzIxM2QxYzg0MTgyZmQ5J2Q9NCskV2NuWEFVTnZxOU1EZENgM2UheVguXFw3O2JgKlg0X29gKDtBXnJrTi1xUmQ2IXgiQX5+eGU8ZlJHWSxzPTJrOTQ2RFJUbDJkMXZaN0NLbVVsMlE9PSxpPTQwOTY="
[15:000:06.209 kw53213a-qc7kf99m1lp] Client -> Server:
"Yz1iaXhoUFhWelpYSXMscj1mYTNkZjU2M2VmNzNlNjdiNDNlOGFlZGJmNTM0YjU1YjE3MjA3NTc5OGNmYmIyMTE2YzIxM2QxYzg0MTgyZmQ5J2Q9NCskV2NuWEFVTnZxOU1EZENgM2UheVguXFw3O2JgKlg0X29gKDtBXnJrTi1xUmQ2IXgiQX5+eGU8ZlJHWSxwPUcvZGZWUUlGZy9NMkpSMlpEMXM0cms4MS9jTmFMa0tuMzZFYzFlTTN1enc9"
[15:000:08.215 kw53213a-qc7kf99m1lp] Server -> Client
NO "Authentication failed."

According to the RFC (https://datatracker.ietf.org/doc/html/rfc5802) the
initial request is defined as:

(("p=" cb-name) / "n" / "y") "," [ authzid ] "," client-first-message-bare

This means the decoded message from the trace is:

n,a=user,n=user,r=fa3df563ef73e67b43e8aedbf534b55b172075798cfbb2116c213d1c84182fd9

The final request (which is rejected by the server with the binding
error) is defined as :

client-final-message-without-proof =
   channel-binding "," nonce [","extensions]

or

client-final-message =
  client-final-message-without-proof "," proof

So the request from the trace decodes into:

c=bixhPXVzZXIs,r=fa3df563ef73e67b43e8aedbf534b55b172075798cfbb2116c213d1c84182fd9'd=4+$WcnXAUNvq9MDdC`3e!yX.\\7;b`*X4_o`(;A^rkN-qRd6!x"A~~xe<fRGY,p=G/dfVQIFg/M2JR2ZD1s4rk81/cNaLkKn36Ec1eM3uzw=

The channel-binding and cbind-input part are defined as:

channel-binding = "c=" base64(cbind-input)

cbind-input = (("p=" cb-name) / "n" / "y") "," [ authzid ] "," [
cbind-data ]

where cbind-data has to be absent by definition as dovecot does not
support channel binding so no "p=" is possible and thus no cbind-data.

This means c=bixhPXVzZXI decodes into.
   n,a=user,

And this is where it gets strange. The dovecot code section which throws
the error, implements the following:

https://github.com/dovecot/core/blob/a5209c83c3a82386c94d466eec5fea394973e88f/src/auth/mech-scram.c#L313

   cbind_input = t_strconcat(request->gs2_cbind_flag, ",,", NULL);
   str = t_str_new(MAX_BASE64_ENCODED_SIZE(strlen(cbind_input)));
   str_append(str, "c=");
   base64_encode(cbind_input, strlen(cbind_input), str);

   if (strcmp(fields[0], str_c(str)) != 0) {
     *error_r = "Invalid channel binding data";
     return FALSE;
   }

As you can see it uses the bind flag which was saved to the request. But
instead of the authzid, it uses always an hardcoded empty string.

Thus cbind-input is "n,," instead of "n,a=user," and results in the
request being rejected with an "Invalid channel binding data".

Which is on the one hand a funny message because dovecot does not
support channel binding at all, as describe in
https://github.com/dovecot/core/blob/a5209c83c3a82386c94d466eec5fea394973e88f/src/auth/mech-scram.c#L164
but it also on the other hand somehow correct because cbind-input string
does not match. Which is an illegal state during channel binding
negotiation which should not happen.

As said previously, it looks to me like a server side bug. Or did I miss
something special case in the RFC?

Kind Regards

Thomas Schmid