Services/OpenId
This page discusses the implementation of an OpenID service.
Contents
Glossary
- OP: OpenId Provider. e.g. http://openid.ziade.org
- RP: Relying Party, the website that does an OpenID request
- FI: Firefox Identity Plugin
Flow
- FI detects forms that offer OpenID and inject a "Sign in using your Firefox Identity" in it so the user can use it to sign in.
- Clicking on the button initiates the process by sending the user's id to the RP
- the RP calls the OP "associate" API
- the RP calls the OP "checkid_setup" API
- FI intercepts the "checkid_setup" call and injects an Authentication header
- The OP redirects to the RP
Demo
You can try the current implementation by installing the custom account manager add-on here: http://ziade.org/account-manager-0.0.13-dev.xpi
This will work with my server at http://openid.ziade.org. The server uses a dummy authentication backend so it will always successfully authenticates you.
Once the add-on is installed, you can try connecting to websites that support openid.
Sites successfuly tested so far:
- http://pypi.python.org/pypi?:action=openid
- http://slashdot.org/my/login
- http://www.livejournal.com/identity/login.bml?type=openid
- https://bitbucket.org/account/signin
- http://stackoverflow.com/users/authenticate
The Identity Server
The server is a full OpenID 2.0 server that will authenticate the user against his Sync account.
The user identifier is: https://server/user-email
Where user-email is the user email used for Sync. The server will internally convert the email to a base32 hash to find the user.
Mine is currently: http://openid.ziade.org/tarek@mozilla.com
The other difference from a regular OpenId server is that authorizing EPs is done automatically in case the Sync Password is provided in an Authorization header, on the checkid_setup call. In that case, the server verifies the password and automatically adds the site, then performs the usual redirect.
If the add-on is deactivated for any reason, the user will still be able to use our openid provider, but will have to manually authorize sites through a screen that will display a "continue" button (after he has logged into our OP).
OpenID APIs
A user is identified by his e-mail.
GET or POST https://server/
Performs an OpenId operation. Depending on the openid_mode request parameter value, the server will perform a different process. * associate: creates an association with the EP. * check_authentication: checks that a given site is authorized * checkid_setup: checks if an EP owns the Claimed Identifier. * checkid_immediate: XXX points to checkid_setup for now See the OpenID doc for more details.
GET https://server/user_email
Discovery page for the user. Returns: * a Yadis page for clients that accept "application/xrds+xml" * an HTML page for other clients. If the user is logged in, and display his own page, he will get his list of active handles w/. associated sites.
Implementation details
The server prototype is here: http://bitbucket.org/tarek/server-openid
- The server is implemented using server-core, for the authentication APIs
- storage
- Users are stored in LDAP for Mozilla, but people can use alternative back-ends
- Association handles and Site tokens are stored in Redis, but people can use alternative back-ends
- The server uses the same node assignment library than Sync to associate a user to a server
- A new multi-value field is added in the LDAP User object: service-enabled
- to enable OpenID for the user, an OpenID value is added
- The Sync value is used to enable/disable Sync.
- The account-enabled field is kept for a global de-activation of a user
- Node assignment can be done on the first OpenID activation if Sync is disabled or not used yet.
The client prototype is here: http://hg.mozilla.org/labs/weave-identity/summary
See also, Account Manager specs: https://wiki.mozilla.org/Labs/Weave/Identity/Account_Manager/Spec/3
Performances
App profile:
- CPU intensive application (crypto)
- All storage in RAM (Redis)
- Storage: handles + sites associated w/ a handle
RAM Formula:
- HANDLE_SIZE = maxsize(handle) + maxsize(secret) + 10 * avgsize(url)
- HANDLE_SIZE = 40 + 32 + 10 * 150 = 1572 octets
- USER_DATA = 1572 * 10 = 15 ko
On a server w/ 8 gig for Redis, there's room for half a million USER_DATA.
An openid transaction is on average two requests:
- smart mode: association + checkid_setup
- dumb mode: checkid_setup + check_authentication
So a 8 gig. server can handle 250k transactions per hour when the RAM is full: 70 transactions/seconds.
Draft on quotas:
- Max TTL for a handle: 3600 seconds.
- Max number of sites / handle : 10
- Max number of handles per user: 10
XXX: bench a server to see how much transactions/seconds it can handle. (cpu work)
Impact on Sync
- If a user tries to create a new account in Firefox Sync via the wizard but already has OpenID activated
- we tell the user that there's an existing account and redirect him to the regular user authentication screen.
- we add Sync in service-enabled in LDAP on the first sync
- If a user deletes his Sync account
- we remove Sync from service-enabled but don't remove the LDAP user
- If a user tries to create a new account in the OpenID UI, and already has Sync activated, we warn that an account exists for that e-mail and propose to the user to activate OpenID through the user account management panel
Impact on Mozilla websites
- We need to implement OpenID support in all of them.
- XXX
Privacy concerns
XXX
- We store authorized sites in RAM. If someone breaks the server he could get them. One way to avoid this is to hash them. But that's more CPU work