Single-Sign-On using Doorkeeper and Warden

PROJECT STATE:

This project had some lift-offs in production environments, but has since been discontinued. It solves many problems, but adds a complexity which makes it hard to spread the knowledge across a team of developers. A central, shared datastore (with all its problems) might still be easier to maintain as a team.

The concept and the code remains for you to gather inspiration. There is a paid service called Auth0 which essentially does what you find here.

Philosophy

• There is no shared data store involved
• It works even for OAuth clients (consumers) on different top-level domains
• It works on native apps
• Support for single-sign-out (per device)
• OAuth Access Tokens are short lived, just like the RFC suggests (refresh Tokens are not used at all).

Requirements

• Client and server: Ruby 2.1.0
• Server: PostgreSQL (For the uuid, inet and hstore column types)

Setup

I refer to the separate README's for the server and the clients.

Terminology

It helps a lot using unambiguous words for different things. So I'll keep to the following convention thoughout all documentation.

• Our end user is called Emily, she has the user ID 42.
• Her browser is called Firefox.
• Our OAuth provider we call Bouncer and it runs on the domain bouncer.dev. Just like at a nightclub, he knows everything about the end users.
• We will refer to Alpha and Beta as our OAuth client web applications running on the domains alpha.dev and beta.dev. These are trusted OAuth clients we have full control over.
• iPhone and Android are our mobile OAuth client native applications. Even though we also have full control over their code base, they considered are untrusted OAuth clients, because othing on these devices can be kept secret from Emily.

How it works

I realize that you might want some sort of high-level overview first. Unfortunately, I never succeeded in visualizing the entire architecture and its timeline in a simple way. Also, there are many different cross-roads where either one or another thing may happen.

So bare with me as I will go through one big use case which includes all different parts of the system at one point or another.

Logging in for the first time

Let's assume Emily is not logged in. She requests a protected resource on Alpha.

Alpha checks Emilys cookie for alpha.dev by looking at session[:passport_id] and notices that she is not logged in.

At this point, Alpha initiates the standard OAuth 2.0 Authorization Grant flow by redirecting Emily to Bouncer.

Show me the code and what it's based on.

Now it's Bouncers turn to check for an existing session in the cookie of bouncer.dev and notices that session[:passport_id] is nil.

It's up to you to come up with an authentication mechanism on Bouncer. You could e.g. use Devise, but usually it will look something like the following.

At this point, Emily is presented with a login form and uses her credentials to authenticate herself.

Upon successful authentication, Bouncer creates a new record in its database in the following table. Don't worry, by and by you will come to understand what the columns mean.

# This table leverages Postgres-specific column types (uuid, inet, hstore).
# But there is no reason why this should not work with any other database.

enable_extension 'uuid-ossp'
enable_extension 'hstore'

create_table :passports, id: :uuid do |t|
  # Relationships with Doorkeeper-internal tables
  t.integer :oauth_access_grant_id     # OAuth Grant Token
  t.integer :oauth_access_token_id     # OAuth Access Token
  t.boolean :insider                   # Denormalized: Is the client app trusted?

  # Passport information
  t.integer :owner_id, null: false               # User ID
  t.string :secret, null: false, unique: true    # Random secret string

  # Passport activity
  t.datetime :activity_at, null: false   # Timestamp of most recent usage
  t.inet :ip, null: false                # Most recent IP which used this Passport
  t.string :agent                        # Post recent User Agent which used this Passport
  t.string :location                     # Human-readable city of the IP (geolocation)
  t.string :device                       # Mobile client hardware UUID (if applicable)
  t.hstore :stamps                       # Keeping track of *all* IPs which use(d) this Passport

  # Revocation
  t.datetime :revoked_at                 # If set, consider this record to be deleted
  t.string :revoke_reason                # Slug describing why deleted (logout, timeout, etc)
  t.timestamps null: false               # Internal Rails created_at and updated_at columns
end

# Doorkeeper is not guaranteed to create a new access token upon each login, it may just return an existing one
# That's why we need to check for `revoked_at`, only valid passports bear the constraint
add_index :passports, [:owner_id, :oauth_access_token_id], where: 'revoked_at IS NULL AND oauth_access_token_id IS NOT NULL', unique: true, name: :one_access_token_per_owner

add_index :passports, :oauth_access_grant_id
add_index :passports, :oauth_access_token_id
add_index :passports, :insider
add_index :passports, :owner_id
add_index :passports, :secret
add_index :passports, :activity_at
add_index :passports, :ip
add_index :passports, :location
add_index :passports, :device
add_index :passports, :revoked_at
add_index :passports, :revoke_reason

In the oauth_applicationns table, Alpha is defined to only allow the insider scope. This is how Doorkeeper knows which consumers are trusted and which are not.

id	name	scopes	...
1	Alpha	insider	...

So in this case, the following records will be created (long random UUIDs are simplified for readability).

First, Doorkeeper creates the OAuth grant token in its internal oauth_access_grants table.

id	resource_owner_id	application_id	token	scopes
111	42	1	1g1g1g	insider

Then the sso gem creates a Passport in its passports table which is related to that grant via the oauth_access_grant_id 111.

Bouncer also sets the insider flag for this Passport, because Bouncer knows that Doorkeeper would not have handed out this insider Grant Token to an outsider OAuth consumer (such as the iPhone or Android).

id	owner_id	group_id	secret	oauth_access_grant_id	oauth_access_token_id	insider
aiaiai	42	agagag	s3same	111		true

We also store the IP (and geolocation derived from the IP) when Bouncer creates the Passport. The IP is taken directly from request.ip as it is Emilys Firefox which directly connects to Bouncer.

id	...	ip	agent	location
aiaiai	...	198.51.100.11	Firefox 1.0	Rome

Additionally, it is useful to keep track of all IPs which used this Passport by storing them in the hstore field called stamps. Much like you would get a stamp in your Passport when you land at an international airport.

id	...	stamps
aiaiai	...	{ "198.51.100.11" => "2015-12-24 20:00" }

While the ip column only contains the most recent IP and the activity_at column only contains the most recent activity, the stamps column contains both the latest IP and all previous IPs and their most recent activity timestamp.

Also, the sso gem persists the passport_id in the bouncer.dev cookie.

session[:passport_id] = 'aiaiai'

From here, the usual OAuth dance between Alpha and Bouncer continues as Doorkeeper directs. At the end of it, Alpha holds an OAuth access token (which is unknown to Emily).

Note that Alpha, being a trusted OAuth consumer, requests the OAuth scope insider. This will later be relevant when untrusted OAuth consumers come into play, for which the scope outsider is reserved.

So, at this point, Doorkeeper creates an access token in its internal oauth_access_tokens table.

id	resource_owner_id	application_id	token
222	42	1	2t2t2t

And the sso gem augments the existing passport with the access_token_id so that the passport record looks like the following. The oauth_access_grant_id is used to find the corresponding Passport.

id	...	oauth_access_grant_id	oauth_access_token_id	insider
aiaiai	...	111	222	true

Bouncer also seizes the moment to update the IP meta information which Alpha was so kind to provide in params[:ip] and params[:agent] on behalf of Firefox. Bouncer trusts the params to truthgully reflect Firefoxes IP, because the insider flag is true.

So, Doorkeeper hands over the new Access Token to Alpha.

Technically, this is where the OAuth flow ends. There is nothing more in the RFC after this point.

Alpha now possesses an Access Token. Per convention of the OmniAuth gem, this Access Token is now used to ask Bouncer for information about Emily.

Show me the code Alpha uses for this.

Bouncer will not only return general information about Emily. In our setup, the crucial information is the single-sign-on session information. We call this a Passport and this is what it looks like deserialized by Alpha.

# The ID and random secret of the corresponding Bouncer database record
passport.id     #=> "aiaiai"
passport.secret #=> "s3same"

# The user and a digest state of the user
passport.user   #=> <User @name="Emily" ...>
passport.state  #=> "asasas"

# An AES-encrypted blob which will be explained later (iPhone section)
passport.chip   #=> "¶§#&"

So what does Alpha do with this information? The information describes the authentication (and maybe even authorization) of the user and should obviously be persisted. Storing this information directly in a cookie would make you vulnerable to cookie replay attacks.

Instead, Alpha saves the passport ID in a alpha.dev cookie and stores the rest in some server side session database, e.g. Redis or ActiveRecord. Here is an example using Warden.

# Example of Alpha storing session information on the server side, not in the Browser
class Warden::SessionSerializer
  def serialize(passport)
    Redis.set passport.id, passport.to_json
    passport.id
  end

  def deserialize(passport_id)
    json = Redis.get passport_id
    SSO::Client::Passport.new JSON.parse(json)
  end
end

The OAuth callback sends Emily back to the resource she requested in the beginning of this whole use case (it's up to you to remember the path she originally requested before the OAuth dance began).

This time, Emily has a alpha.dev cookie with a valid Passport in it.

Verifying authentication and authorization

But can Alpha really trust the Passport in the alpha.dev cookie? Alpha has no knowledge about users at all. So Alpha has to ask Bouncer at every request what the user looks like at the moment.

sso provides a Warden hook for just this functionality. Whenever Warden in Alpha deserializes Emilys Passport from the session store, it is automatically verified with Bouncer.

Show me the code.

This is a public API endpoint, so as long as you are in possession of a Passport id and secret you can call it.

Of course, we do not send the actual Passport secret over the Internet, but we sign the request with it. Since Bouncer knows the secret it can verify the validity of the request sent by Alpha.

# Pseudo code executed by Alpha to verify the Passport
params = {
  ip:    request.ip,
  agent: request.agent,
  state: passport.state,
}

url = "bouncer.dev/oauth/sso/v1/passports/#{passport.id}"
HTTParty.get url, params.sign(passport.secret)

So what is this passport.state? It is a digest describing the state of the user session. Bouncer will look up Emily in the database and calculate the digest of her current state with something like this.

# Pseudo code of Bouncer calculating the user state digest
key   = "secret string only bouncer knows"
value = [emily.name, emily.email, emily...].join

state = HMAC(key, value)

If the state is the same as the one Alpha sent in, Bouncer now knows that Alphas information about all critical attributes of Emily is up-to-date.

If that is the case, Bouncer will just reply "alright".

There is one more thing happening in this process. Alpha sent in meta information, such as IP and user agent in the params. Bouncer can rely on these params because the Passport was created via the insider Alpha. If you think about it, nobody but Alpha actually knows the Passport. So only Alpha, who established the Passport, can make this request to Bouncer.

Alpha is now free to trust the information it has about Emily and can hand out the resource to her.

Propagating user changes

Let's say, Emily changes her profile at this point. She changes her email address from [email protected] to [email protected]. Since Bouncer is the only one knowing about all users, that is where the change is persisted.

Let's say Alpha has a top navigation bar, which displays [email protected]. At this point, only Bouncer knows about the new email address. So how does this information propagate to Alpha?

As you can see, Bouncer simply told Alpha about it at a passport verification request. When Bouncer calculated Emilys user state, the digest was not the same any more (since emily.email is one of the user attributes included in the state calculation). Bouncer responds with the most recent Passport (including the new user and the new state digest, but without the secret). Alpha updates the local Passport accordingly.

By the way, only you as a developer can know which of Emilys attributes are included in the state digest calculation. You might not care if her email address changes, but need to make sure her authorization rights are propagated as soon as possible. So you would have something like the following pseudo code in Bouncer (the sso gem allows you to configure the state calculation).

# Fetch Emilys permissions in realtime from some database
emily.permissions = [:admin, :moderator, ...]

# Make the Passport state dependent on the set of permissions
HMAC(secret, "... #{emily.permissions} ...")

Just like before, Bouncer keeps track of the IP, geolocation, browser agent identifier and passport stamps for this Passport. Bouncer updates this meta information by looking at the params which Alpha provided to Bouncer. The activity_at timestamp is also updated.

Fallback if Bouncer goes down

Now you say:

Wait a minute, so Bouncer is really just like a shared data store for the single-sign-on sessions, only that it is accessed via HTTP, and it's a single point of failure web app that has to respond really really fast?.

That's correct.

However, we optionally provide a fallback mechanism if Bouncer happens to not respond fast enough (within 10ms or so). Let's assume that Emily wants to see another resource. This time, however, Bouncer is under a DOS attack and does not respond to Alphas attempt to verify the information in the alpha.dev Passport..

If Alpha has a previously verified Passport about Emily, Alpha will use that information instead.

This way, Bouncer is not such a critical single point of failure which will cause every user to be logged out whenever Bouncer cannot respond within 10ms or there is network congestion.

Mitigating replay attacks

Conceptually, however, this opens up a vector for a replay attack. For the duration that Bouncer is down, all Passports are "frozen in" yet remain valid.

So Firefox could send in the same old Cookie and it would always be interpreted as the most recent Passport Alpha has in store. As soon as Bouncer comes up again (and successfully responds to a Passport verification request by Alpha), the most recent Passport is immediately propagated to Alpha again.

The attack window, then, lies between the most recent succeeded Passport verification request between Alpha and Bouncer and the next succeeding Passport verification request between Alpha and Bouncer.

In other words, if you notice that Bouncer is down, you better fix it fast or shut down critical services to avoid attacks.

Additionally, you should expire insider Passports persisted in Alpha (i.e. the Redis session store) if they have not been updated for, say a week or two. This way you can further minizime the attack window to be between 2 weeks in the past and Bouncer coming back online. (Unfortunately this does not work well with long-term outsider Passports of the native iPhone app. Simply because Emily didn't use the app for a few months, should not destroy her session between the iPhone and Alpha.)

If you go the other way and log out every user immediately if Bouncer does not respond, you may have a whole other problem. If an attacker can bring down Bouncer, all your end-users are logged out. But you can be rather sure that no unauthorized leak of resources occured.

There is a middle way, too. Every time the Passport verification succeeded, the Passport will respond positively to passport.verified?. If it is not verified, you might still want to show Emilys name in the navigation bar, but deny her to see sensitive account information or buying products. So she would not be logged out but would still have to wait until Bouncer is back up for critical use cases.

You'll have to decide for yourself whether to use the fallback mechanism or not. I'm just trying to lay out the advantages and disadvantages.

Beta comes into play

Single-sign-on is just a concept which can look different from use case to use case. This becomes clear as soon as Emily (now already logged in on Alpha and Bouncer) surfs to Beta.

What would she see? Beta knows nothing about her.

1. Beta could simply always send Emily to Bouncer when she is not logged in. Since Firefox has a session with Bouncer, she would not even notice any redirect but would be immediately logged in on Bouncer as soon as this "automatic" OAuth dance with finishes.

But should Beta really be unreachable without authentication? After all, we did not do something like this when Emily came to Alpha earlier today.

2. Beta could present a login button. Whenever Emily clicks on it, she would suddenly be logged in without entering a password.

But is this maybe confusing to Emily?

A trade-off between these two solutions might be the following.

Whenever Emily surfs to Beta, Beta instructs her Firefox to make an AJAX call to something like GET bouncer.dev/am_i_logged_in.

If the answer is YES, i.e. there is a Passport ID in the bouncer.ev cookie, let JavaScript initiate the OAuth flow immediately.

window.location.href = '/auth/sso'

From Emilys point of view, she saw the page on Beta loading completely (saying "you are not logged in"). Then, suddenly the page disappears and comes back; this time she is logged in.

Additionally, whenver you cross-link from Alpha to Beta you might include a ?assume_logged_in=true flag in the URL so that Bouncer can skip the AJAX request right away and perform the 302 redirect to /auth/sso without rendering anything first.

Either way, Beta will have to send Firefox to Bouncer by some means. So what happens exactly when Firefox meets Bouncer when there already is a session?

At this point, Bouncer has access to the bouncer.dev cookie persisted in Firefox and can look up the ID of the Passport and find it in the database. When handing out the OAuth Grant token, Bouncer remembers the outgoing Grant by augmenting the Passport.

Doorkeeper oauth_access_grants Table

id	resource_owner_id	application_id	token	scopes
111	42	(Alpha) 1	1g1g1g	insider
333	42	(Beta) 2	2g2g2g	insider

Passports Table

id	...	oauth_access_grant_id	oauth_access_token_id
aiaiai	...	333

As you can see, this process is analogous to when Alpha established a session, only that the Passport now does not need to be created from scratch, we re-use the existing one.

Note that the IP, geolocation, browser agent identifier, passport stamps, and activity_at are also updated, but I omitted the details here. This time, Bouncer uses request.ip, since Firefox is directly talking with Bouncer.

So, Bouncer hands out the OAuth Grant Token to Firefox, who gives it to Beta, who exchanges it for an Access Token and uses that Access Token to retrieve Emilys Passport from Bouncer. Firefox will then re-attempt accessing the Beta resource, whereupon Bouncer verifies the Passport with Bouncer and delivers the resource to Firefox.

In this process, the Passport is also augmented with the newly created Access Token.

Doorkeeper oauth_access_tokens Table

id	resource_owner_id	application_id	token	scopes
222	42	(Alpha) 1	2t2t2t	insider
444	42	(Beta) 2	4t4t4t	insider

Passports Table

id	...	oauth_access_grant_id	oauth_access_token_id
aiaiai	...	333	444

Congratulations. Emily is now logged in on Alpha and Beta. This works by keeping up the session between Firefox and Bouncer. You could now proceed with logging in on Gamma, Delta, etc.

Single-Sign-Out

This is simple. Just invalidate the Passport by setting the following flags.

Passports Table

id	...	revoked_at	revoke_reason
aiaiai	...	2015-12-24 21:00	logout

Both Alpha and Beta know the passport ID so you could just create a logout link to GET bouncer.dev/logout/aiaiai and perform the revocation there.

Alternatively, Alpha could make a server-to-server request to DELETE bouncer.dev/oauth/sso/v1/passports/aiaiai and Bouncer executes the revocation then. In this scenario, Emily would not end up seeing Bouncer telling her "you are logged out", but Alpha being able to tell her so.

Which one you use depends on whichever is more desirable in your use case.

As soon as Emily makes subsequent requests to Alpha, the Passport verification request from Alpha to Bouncer will inform Alpha to delete the local Passport entirely. The same is true for Beta.

This way we effectively logged out every session which was created using Firefox but leave those alive which were created using e.g. Safari.

Just a gentle recapitulation: if Bouncer is down, the logout will not work. So you better ensure it is up and running to minimize the exploitation window.

Native Clients (aka iPhone/Android)

In essence, a native client is like any other OAuth Client. Yet there are a few differences:

1. This client needs to authenticate to other clients, say, your API backend provided by Alpha.
2. There is no way you can hide any information on the device from the end user (this includes OAuth client credentials shared between iPhone and Bouncer).
3. Android needs to read the plain Passport to display Emilys name in the app. Remember that, so far, the Passport was located in an encrypted alpha.dev cookie unintelligible to Firefox.

Also, we really want to avoid HTML web views for login forms. The whole point of native apps is to not have to fallback to browser technology.

The iPhone is known to Bouncer as a Doorkeeper::Application in the internal Doorkeeper oauth_applications table. Only the scope outsider is allowed for native applications.

id	name	scopes	...
1	Alpha	insider	...
2	Beta	insider	...
3	iPhone	outsider	...

The corresponding OAuth client credentials are hardcoded in the iPhone app, and are considered public information (they can easily be extracted by the end-user).

So let's do this.

Logging in for the first time

It start's with the iPhone sending Emilys username and password to Bouncer. In return, the iPhone will get an OAuth Access Token (this is the OAuth Resource Owner Password Credentials Grant). That Token is exchanged for a Passport.

Basically, the iPhone acts as both, Firefox and Alpha.

During this process, Doorkeeper will create the Access Token in the internal oauth_access_tokens table (I omit showing any previous records here. Anything that happened earlier only concerned the session between Firefox and Bouncer).

id	resource_owner_id	application_id	token	scopes
555	42	(iPhone) 3	5t5t5t	outsider

Passports Table

Bouncer creates a new Passport for this new single-sign-on session of this device.

id	...	oauth_access_grant_id	oauth_access_token_id
bibibi	...		555

Earlier, Bouncer trusted the params provided by Alpha to update the Passport IP meta information. This time, Bouncer recognized that this is an outsider request and retrieves that information directly by inspecting the incoming request.ip object.

The agent and device UUID, however, are still retrieved from the params (provided by the iPhone), since these are not reliable either way.

id	...	ip	agent	location	device
bibibi	...	198.51.100.22	iPhone	Venice	dedede

Now the iPhone has a Passport, Emily is logged in.

You probably want to persist the user information in some way. The Passport secret will typically not change and should be persisted in secure storage (e.g. KeyChain). The user object will be updated more frequently (i.e. whenever the user state changes).

Cross-client authentication and authorization

Now let the iPhone use its Passport to request a resource from Alpha.

What does the iPhone have to offer to Alpha so that Alpha would trust, or even be able to verify the signature of the request? Nothing. Since Alpha does not even know the Passport secret. And thus has to rely on Bouncer being able to tell Alpha whether the Passport is valid or not.

If Alpha would have some (even unreliable) idea of what the secret is, Alpha could at least verify the signature of the iPhone request, and Alpha could then use that secret to sign the Passport verification request to Bouncer.

Do you remember the Passport chip attribute? It's a synchronously encrypted data store which Bouncer creates and the iPhone cannot decipher. If we introduce a simple shared secret between Bouncer and Alpha, we could "transport" some information from Bouncer to Alpha. This is how the Passport secret will come to Alpha.

When Bouncer handed out the Passport to the iPhone, it performed the following operation to put something in the Passport chip. This value is not stored in any database, it is simply "attached" to the Passport.

# Pseudo code of Bouncer setting the chip of the Passport
shared_secret = "something only Bouncer, Alpha and Beta know (i.e. trusted clients)"
# Including the ID in the plaintext ensures the chip is only valid for this Passport
plaintext = [passport.id, passport.secret].join('|')
passport.chip = AES.encrypt(plaintext).with(shared_secret)

The iPhone simply passed on the chip to Alpha. Upon receiving it, Alpha decrypts the chip and now knows the passport secret. With that secret, Alpha is able to determine whether the iPhone properly signed the request with it. Of course, Alpha cannot trust the secret (yet), but it is a practical approach for Alpha to be able to verify it with Bouncer).

So, the iPhone signs the request with the Passport secret and also sends along the Passport user state digest and the Passport chip.

At this point, Alpha decrypts the secret from the chip, verifies the request and makes its own verification request to Bouncer.

Since this is the first request by iPhone to Alpha, Alpha has no information about Emily in some local datastore. So when Alpha makes its Passport verification request to Bouncer, Alpha will simply omit the state so as to guarantee to receive a user object.

But there is one more thing. Bouncer would like to update the Passport IP activity meta information, which Alpha has to provide by sending iPhones IP to Bouncer in the params . But Bouncer will not trust the params of this outsider Passport - after all the iPhone could perform the following request itself and forge params[:ip] to be something fake.

Alpha has to do some effort to have Bouncer trust in the params[:ip]. Signing the IP should suffice for this purpose.

# Pseudo code of Alpha signing the IP proxied by Alpha from the iPhone to Bouncer
alpha_client_id     = "Alphas OAuth Client ID"
alpha_client_secret = "Alphas OAuth Client Secret"

params[:ip]         = "198.51.100.22"   # iPhone IP
params[:insider_id] = alpha_client_id   # So that Bouncer knows who Alpha is

params[:insider_signature] = HMAC.calculate(params[:ip]).with(alpha_client_secret)

Bouncer can lookup Alphas client secret in its Doorkeeper oauth_applications table and verify the signature to see if params[:ip] really is to be trusted.

Now Alpha can persist the Passport, including the user, in its local data store. From now on the chip sent by the iPhone can be ignored (but the iPhone needs to keep sending the chip because the iPhone does not know wether Alpha or Beta needs it or not).

Any subsequent verification request from Alpha to Bouncer includes the state as usual. The response may not contain a user if the user state did not change meanwhile.

The response by Bouncer can even fail while Alpha will fallback to the previously verified Passport (along with all security implications previously explained).

However, if it is the first time the iPhone talks to Alpha, the request from Alpha to Bouncer must succeed - in order to fetch a user. You need to be prepared for these kinds of errors in the iPhone app and retry or inform the user to try again later.

If the user object changed, Alpha informs the iPhone by delivering the new Passport state and user as a param

So, to summarize, the iPhone always sends in Passport id, state and chip and any trusted OAuth client (Alpha, Beta) can receive this information at any time and get the authentication/authorization information from Bouncer.

Are you dizzy? Me too.

Development

How to run the specs:

# RAILS_ENV is "test" by default.
# If you want "development", you have to `cd specs/dummy` first.
bundle exec rake db:create
bundle exec rake db:migrate
bundle exec rspec

Good to know:

• You can always git grep POI to see some points of interest. They will be properly documented as development progresses.
• You should tail spec/dummy/log/test.log because it's really helpful

Contributing

• The CHANGELOG follows this format.

License

MIT 2016 halo, see LICENSE