fast_scram is a purely-functional Erlang implementation of the Salted Challenge Response Authentication Mechanism, where the challenge algorithm is implemented as a carefully-optimised NIF using fast_pbkdf2.
fast_scram is a rebar3-compatible OTP application, building is as easy as rebar3 compile, and using it in your projects as
{plugins, [pc]}.
{provider_hooks,
[{pre,
[{compile, {pc, compile}},
{clean, {pc, clean}}]}]}.
{deps,
[{fast_scram, "0.6.0"}]}.In SCRAM, a SaltedPassword is defined as
SaltedPassword := Hi(Normalize(password), salt, i)
This algorithm is precisely the one that pays the challenge, and it is the one we solve here with the best performance. Simply do:
SaltedPassword = fast_scram:hi(Hash, Password, Salt, IterationCount)where Hash is the underlying hash function chosen as described by
-type sha_type() :: crypto:sha1() | crypto:sha2().If you want to avoid reimplementing SCRAM again and again, you can use the extended API. The best example is that one of the tests. Given already configured states, the flow is as follows:
%% AUTH
{continue, ClientFirst, ClientState3} = fast_scram:mech_step(ClientState1, <<>>),
%% CHALLENGE
{continue, ServerFirst, ServerState4} = fast_scram:mech_step(ServerState2, ClientFirst),
%% RESPONSE
{continue, ClientFinal, ClientState5} = fast_scram:mech_step(ClientState3, ServerFirst),
%% SUCCESS
{ok, ServerFinal, ServerFinalState} = fast_scram:mech_step(ServerState4, ClientFinal),
%% Client successfully accepts the server's verifier
{ok, ClientFinal, ClientFinalState} = fast_scram:mech_step(ClientState5, ServerFinal).The API is simple: fast_scram:mech_step/2 takes a SCRAM state, and the last message it received
(in the case of the first step of the client, this is obviously, and necessarily, empty).
The return value is always a 3-tuple, tagged with either ok, continue or error.
The second element is always a binary, and the third is always the scram state.
-spec mech_step(fast_scram_state(), binary()) ->
{ok, final_message(), fast_scram_state()} |
{continue, next_message(), fast_scram_state()} |
{error, error_message(), fast_scram_state()}.oktagged-tuples mean that the algorithm has returned successfully. The message will be the last one to send to the peer, empty in the case of the client, containing the server verifier for the server. The state will not be needed anymore, so it can be ignored.continuemeans that the algorithm is not done yet. The message is what needs to be send to the peer, by whatever means the protocol chooses (encoded in a major packet through some network protocol, etc). The new state is the one that should be plugged into the next step, when the peer has answered.errormeans that the algorithm is over, unsuccessfully, where the message contains some explanation. The state might include parsed data or be return as it was.
How messages are delivered to peers is part of the protocol within which SCRAM is embedded:
for example, in XMPP, messages are delivered as special stanzas with the SCRAM payload encoded in
base64. So an XMPP client would do, for example, using exml
{continue, Message, NewState} = fast_scram:mech_step(State, <<>>),
Contents = #xmlcdata{content = base64:encode(Message)},
Stanza = #xmlel{name = <<"auth">>,
attrs = [{<<"xmlns">>, <<"urn:ietf:params:xml:ns:xmpp-sasl">>},
{<<"mechanism">>, <<"SCRAM-SHA-1">>}],
children = [Contents]},
%% send stanzaThis is the part that requires some knowledge of the SCRAM protocol.
A ready SCRAM state is built using fast_scram:mech_new/1,
which takes a map with the configuration parameters.
Example configurations are, for the client:
#{entity => client,
hash_method => sha,
username => <<"user">>,
auth_data => #{password => <<"somesupersafepassword">>}}And for a server:
#{entity => server,
hash_method => sha,
nonce => <<"3rfcNHYJY1ZVvWVs7j">>,
retrieve_mechanism => fun(Username) -> MoreConfig end}NOTE: SCRAM requires the username and password to be Normalized using the SASLprep profile of the stringprep algorithm. Stringprepping algorithms would introduce a dependency to this repository that I didn't want to, so it is left to the user of this plugin to provide stringprepped binaries from the get-go.
The first and most important key is the entity key,
which takes two values: client or server.
The next necessary key is the negotiated hash_method,
that is, which of the SHA algorithms will be executed.
Can be any of the OTP's crypto:sha1() | crypto:sha2().
Next keys depend on the chosen entity.
If you want to configure a client state, then a username key is required.
If you want to configure a server state, then retrieve_mechanism is required.
Next, for both cases, an auth_data key is required. The value for this key is a map containing the
minimum necessary information for executing a SCRAM algorithm: often just a password.
But often, to avoid the challenge penalty, servers and client cache certain keys,
considering that a server often gives the same salt and iteration count for a specific client.
So we can instead cache salted_password, or a pair stored_key-server_key,
or a pair client_key-server_key. All these pairs can be given with a password as a fallback,
if the algorithm was to need recalculation.
If the client is being given any cached configuration, it will simply attempt that data regardless
of the challenge that the server requests from him. If verification was desired instead of failing,
the main config map can take keys cached_it_count and cached_salt, and these will be verified
against the challenge requested by the server: if it matches, the cached data will be used. If it
doesn't, all data will be recalculated using the password key in the auth_data map, provided it
is available.
Channel binding specification can also be given by channel_binding => {Type, Data},
where Type is the channel binding name, and Data is its associated payload.
The default is {undefined, <<>>}, which will set the gs2 flag to no binding, that is, <<"n">>.
If for example a client had channel binding, but saw the server not offering any,
this client should set the flag to {none, <<>>}: this will send the gs2 flag as <<"y">>.
SCRAM requires that the server retrieves the user's data with the username as exactly given
in the client's first message. To configure this, a retrieve_mechanism key is required,
whose value is a function of the type:
-type retrieve_mechanism() :: fun((username()) -> configuration())
| fun((username(), fast_scram_state()) ->
{configuration(), fast_scram_state()}).That is, a function object that:
- Takes a username and returns more configuration to append to the state
- Takes a username and the current state, and returns a pair of the extended configuration and a possibly new state.
See examples below.
Fun = fun(Username) ->
%% Get scram data for this user from the database
...
%%% {StoredKey, ServerKey, Salt, ItCount} ->
...
#{salt => Salt,
it_count => ItCount,
auth_data => #{stored_key => StoredKey,
server_key => ServerKey}}
end,
{ok, State} = fast_scram:mech_new(
#{entity => server, hash_method => Sha, retrieve_mechanism => Fun}). Fun = fun(Username, State0) ->
%% Get scram data for this user from the database
...
%%% {StoredKey, ServerKey, Salt, ItCount} ->
...
Config = #{salt => Salt,
it_count => ItCount,
auth_data => #{
stored_key => StoredKey,
server_key => ServerKey}}
%% Custom data can also be stored in the state to be extracted later
State1 = fast_scram:mech_set(some_key, SomeData, State0),
{Config, State1}
end,
{ok, State} = fast_scram:mech_new(
#{entity => server, hash_method => Sha, retrieve_mechanism => Fun}).SCRAM is a challenge-response authentication method, that is, it forces the client to compute a challenge in order to authenticate him. But when the server implementation is slower than that of an attacker, it makes the server vulnerable to DoS by hogging itself with computations. We could see that on the CI and load-testing pipelines of MongooseIM for example.
Is partial. We don't expect to have the fastest implementation, as that would be purely C code on GPUs, so unfortunately an attacker will pretty much always have better chances there. But we can make the computation cheap enough for us that other computations —like the load of a session establishment— will be more relevant than that of the challenge; and also that other defence mechanisms like IP blacklisting or traffic shaping, will fire in good time.
It all boils down to the right PBKDF2 implementation, as done in fast_pbkdf2, which is on average 10x faster on the machines I've tested it. But while the erlang implementation consumes memory linearly to the iteration count, the NIF implementation does not allocate any more memory.
Note that since OTP24.2, the crypto application exposes a native pbkdf2_hmac function. However, the NIF implementation from fast_pbkdf2 is twice as fast, so that one is left as a default.
If for any particular reason you want to skip compiling and loading custom NIFs, you can override the dependency using rebar3's overrides as below, this way the dependency will not be fetched and code will be compiled to use the crypto callback instead.
{overrides,
[
{override, fast_scram, [{erl_opts, [{d, 'WITHOUT_NIFS'}]}, {deps, []}]},
]
}.