Authentication process design ============================= See for an overview of how the authentication process works. There are four major classes in the code: * 'struct mech_module': Authentication mechanism * 'struct password_scheme': Password scheme * 'struct passdb_module': Password database * 'struct userdb_module': User database There are many implementations for each of these, and it's simple to add more of them. They can also be added as plugins, although the current plugin loading code doesn't allow loading authentication mechanisms cleanly, and it's not possible to add new credentials (see below). The code flow usually goes like: * Dovecot-auth listens for new authentication client connections (the listener socket is created by master process and passed in MASTER_SOCKET_FD -> 'main.c:main_init()' -> 'auth-master-connection.c:auth_master_listener_add()') * A new authentication client connects via UNIX socket ('auth-master-connection.c:auth_master_listener_accept()' -> 'auth-client-connection.c:auth_client_connection_create()') * Authentication client begins an authentication ('auth-client-connection.c:auth_client_input()' -> 'auth_client_handle_line()' -> 'auth-request-handler.c:auth_request_handler_auth_begin()' [ -> 'auth-request.c:auth_request_new()']) * Authentication mechanism backend handles it ('mech->auth_initial()' and 'mech->auth_continue()' in 'mech-*.c') * The mechanism looks up the password from passdb ('auth-request.c:auth_request_verify_plain()' and 'auth_request_lookup_credentials()') and the password scheme code to verifies it ('password-scheme.c:password_verify()' and 'password_generate()') * If user is logging in, the user information is looked up from the userdb ('auth-master-connection.c:master_input()' -> 'master_input_request()' -> 'auth-request-handler.c:auth_request_handler_master_request()' -> 'auth-request.c:auth_request_lookup_user()') * The authentication may begin new authentication requests even before the existing ones are finished. It's also possible to request a userdb lookup directly, for example Dovecot's [LDA.txt] needs that. The code path for that goes 'auth-master-connection.c:master_input()' -> 'master_input_user()' -> 'auth-request.c:auth_request_lookup_user()'. Authentication mechanisms ------------------------- These are [Sasl.txt] authentication mechanism implementations. See for a list of mechanisms supported by Dovecot. A new mechanism is created by filling a 'struct mech_module' (in 'mech.h') and passing it to 'mech_register_module()'. The struct fields are: mech_name: The public name of the mechanism. This is shown to clients in the IMAP, POP3 and SMTP capability lists. If you create a new non-standard mechanism, please prefix it with "X-". flags: Describes how secure the mechanism is. Also 'MECH_SEC_PRIVATE' flag specifies that the mechanism shouldn't be advertised in the capability list. This is currently used only for APOP mechanism, which is defined by the POP3 protocol itself. passdb_need_plain: This mechanism uses passdb's 'verify_plain()' function to verify the password's validity. This means that the mechanism has access to the plaintext password. This is true only for plaintext mechanisms such as PLAIN and LOGIN. The main purpose of this flag is to make dovecot-auth complain at startup if there are no passdbs defined in the configuration file. Note that a configuration without any passdbs is valid with eg. GSSAPI mechanism which doesn't need a passdb at all. passdb_need_credentials: This mechanism uses passdb's 'lookup_credentials()' function. See below for description of the credentials. auth_new(): Allocates a new 'struct auth_request'. Typically with more complex mechanisms it really allocates a 'struct _auth_request' which contains 'struct auth_request' as the first field, followed by mechanism-specific fields. auth_initial(request, data, data_size): This begins the authentication, data and data_size containing the initial response sent by the client (decoded, not in base64). Call 'request->callback()' once you're done (see below). auth_continue(request, data, data_size): Continues the authentication. Works the same as 'auth_initial()'. auth_free(): Free the request. Usually all the memory allocations for the request should be allocated from 'request->pool', so you can use 'mech_generic_auth_free()' which simply frees the pool. 'auth_initial()' and 'auth_continue()' continue or finish the authentication by calling 'request->callback()': ---%<------------------------------------------------------------------------- typedef void mech_callback_t(struct auth_request *request, enum auth_client_result result, const void *reply, size_t reply_size); ---%<------------------------------------------------------------------------- The 'reply' and 'reply_size' contain the server's mechanism-specific reply to the client. If there is no need to return anything (which is usually the case with the "success" reply), the 'reply_size' can be 0. The 'result' parameter is one of: * AUTH_CLIENT_RESULT_CONTINUE: Client can continue the authentication. The reply contains the mechanism-specific reply sent to the client. * AUTH_CLIENT_RESULT_SUCCESS: Authentication successful. The reply is usually empty. * AUTH_CLIENT_RESULT_FAILURE: Authentication failed. The reply is always ignored. The 'request->callback()' should actually be called directly only for continuation requests (a new function should probably be added for this as well). For success and failure replies, you should instead use one of these functions: * 'auth_request_success()' * 'auth_request_fail()' * 'auth_request_internal_failure()': Use this if you couldn't figure out if the authentication succeeded or failed, for example because passdb lookup returned internal failure. SASL authentication in general works like: 1. Client begins the authentication, optionally sending an "initial response", meaning some data that the mechanism sees in 'auth_initial()'. * Note that not all protocols support the initial response. For example IMAP supports it only if the server implements SASL-IR extension. Because of this mechanisms, such as PLAIN, support doing the authentication either in 'auth_initial()' or in 'auth_continue()'. * If the client initiates the authentication (ie. server's initial reply is empty, such as with PLAIN mechanism) you can use 'mech_generic_auth_initial()' instead of implementing your own. 2. Server processes the authentication request and replies back with 'request->callback()'. * If the authentication failed, it's placed into 'auth_failures_buf' unless 'request->no_failure_delay=TRUE'. The failures are flushed from the buffer once every 2 seconds to clients and 'mechanism->auth_free()' is called. * If the authentication succeeded and * there is a master connection associated with the request (IMAP/POP3 login), the authentication now waits for master connection to do a verification request. If this for some reason doesn't happen in 'AUTH_REQUEST_TIMEOUT' seconds (3,5 mins), it's freed. * there isn't a master connection (SMTP AUTH), the authentication is freed immediately. 3. Client processes the reply: * If the authentication continues, it sends back more data which is processed in 'auth_continue()'. Goto 2. * If the authentication failed, it's done. 4. If the authentication succeeded, the client requests a login from the master process, which in turn requests verification from the auth process. * Besides verifying the authentication, dovecot-auth also does a userdb lookup to return the userdb information to master. * If the verification fails (normally because userdb lookup fails), the client gets "internal authentication failure" * If the verification succeeds, the user is now logged in * In either case, 'mechanism->auth_free()' is called now. Credentials ----------- Most of the non-plaintext mechanisms need to verify the authentication by using a special hash of the user's password. So either the passdb credentials lookup returns a plaintext password from which the hash can be created, or the hash directly. The plaintext to hash conversion is done by calling 'password_generate' function of the password scheme. Unfortunately the list of allowed credentials is currently hardcoded in 'enum passdb_credentials'. The enum values are mapped to password scheme strings in 'passdb_credentials_to_str()'. Some day the enum will be removed so plugins can add new mechanisms. Besides the mechanism-specific credentials, the enum contains: _PASSDB_CREDENTIALS_INTERNAL: I don't remember why this really exists. It should probably be called _PASSDB_CREDENTIALS_INVALID or something and used only by some asserts.. PASSDB_CREDENTIALS_PLAINTEXT: Request a plaintext password. PASSDB_CREDENTIALS_CRYPT: Request the password in any scheme. This is especially useful if you only want to verify a user's existence in a passdb. Used by [UserDatabase.Static.txt] in userdb lookups. Password schemes ---------------- 'struct password_scheme' has fields: name: Name of the scheme. This only shows up in configuration files and maybe in the passwords stored in passdb ("{scheme_name}password_hash"). password_verify(plaintext, password, user): Returns TRUE if 'password' hash matches the plaintext password given in 'plaintext' parameter. If the password hash depends on the username (eg. with DIGEST-MD5), the 'user' parameter can also be used. password_generate(plaintext, user): Returns the password hash for given plaintext password and username. You can create a new password scheme by simply creating a 'struct password_scheme' named '_scheme', compiling a shared object and placing it to '$moduledir/auth/' directory. Password databases ------------------ See for a description of passdbs and a list of already implemented ones. 'struct passdb_module' contains fields: cache_key: A string containing [Variables.txt]. When expanded, it uniquely identifies a passdb lookup. This is '%u' when the passdb lookup validity depends only on the username. With more complex databases such as SQL and LDAP this is created dynamically based on the password query in the configuration file. If there are multiple variables, they should be separated so that their contents don't get mixed, for example '%u%r%l'. 'auth_cache_parse_key()' can be used to easily create a cache key from a query string. default_pass_scheme: The default scheme to use when it's not explicitly specified with a "{scheme}" prefix. blocking: If TRUE, the lookup is done in dovecot-auth worker process. This should be used if the lookup may block. iface.preinit(auth_passdb, args): Allocate 'struct passdb_module' and return it. This function is called before chrooting and before privileges are dropped from dovecot-auth process, so if should do things like read the configuration file.'auth_passdb' is typically used for getting a memory pool and looking up some global settings such as 'auth_passdb->auth->verbose_debug'. 'args' contains the args parameter in configuration file. iface.init(module, args): The privileges have been dropped before calling this. 'module' contains the structure returned by 'preinit()'. 'args' is the same as in 'preinit()'. Typically this function will do things like connect to the database. iface.deinit(module): Close the connection to the password database and free all the memory you used. iface.verify_plain(auth_request, password, callback): Check if the given plaintext password matches. 'auth_request->credentials = -1' always. When the verification is done, call the given callback with the result in 'result' parameter. iface.lookup_credentials(auth_request, callback): Look up the password credentials. 'auth_request->credentials' contains the credentials that the mechanism wants. When the lookup is finished, call the given callback with the result in 'result' parameter, and if the lookup was successful the credentials in 'password' parameter. Plaintext authentication mechanisms typically call 'verify_plain()', which is possible to implement with all the passdbs. Non-plaintext mechanisms typically call 'lookup_credentials()', which isn't possible to implement always (eg. PAM). If it's not possible to implement 'lookup_credentials()', you can leave the pointer to it NULL. If the passdb uses connections to external services, it's preferred that they use non-blocking connections. Dovecot does this whenever possible (PostgreSQL and LDAP for example). If it's not possible, set 'blocking = TRUE'. With both functions 'auth_request->passdb->passdb' contains the passdb_module returned by your 'preinit()' function. 'auth_request->user' contains the username whose password we're verifying. You don't need to worry about [MasterUsers.txt] here. It's also possible to use any other fields in 'auth_request' to do the lookup, such as 'service', 'local_ip' or 'remote_ip' if they exist. Often you want to let user to configure the lookup with [Variables.txt] (eg. SQL query). In that case you can use 'auth_request_get_var_expand_table()' to retrieve the variable table for 'var_expand()'. The passdb lookup can return one of the following results: PASSDB_RESULT_INTERNAL_FAILURE: The lookup failed. For example SQL server is down. PASSDB_RESULT_SCHEME_NOT_AVAILABLE: 'lookup_credentials()' requested a scheme which isn't in the passdb PASSDB_RESULT_USER_UNKNOWN: The user doesn't exist in the database. PASSDB_RESULT_USER_DISABLED: The user is disabled either entirely, or for this specific login (eg. only POP3 logins allowed). This isn't commonly implemented in passdbs. PASSDB_RESULT_PASS_EXPIRED: The user's password had expired. This isn't commonly implemented in passdbs. PASSDB_RESULT_PASSWORD_MISMATCH: The password given in 'verify_plain()' wasn't valid. PASSDB_RESULT_OK: Success. User databases -------------- See for a description of userdbs and a list of already implemented ones. 'struct userdb_module' is very similar to 'struct passdb_module'. The lookup callback is a bit different though: ---%<------------------------------------------------------------------------- typedef void userdb_callback_t(enum userdb_result result, struct auth_stream_reply *reply, struct auth_request *request); ---%<------------------------------------------------------------------------- 'result' contains one of: USERDB_RESULT_INTERNAL_FAILURE: The lookup failed. For example SQL server is down. USERDB_RESULT_USER_UNKNOWN: The user doesn't exist in the database. USERDB_RESULT_OK: Success. There is no equivalent for PASSDB_RESULT_USER_DISABLED currently. Practically the userdb result is used only by Dovecot's [LDA.txt] to figure out if the user exists or not. When logging in with IMAP or POP3, the user's existence was already checked in passdb lookup, so only in rare conditions when a user is logging in at the same time as it's being deleted, the userdb result is USER_UNKNOWN. The 'reply' parameter contains the username (it's allowed to be different from the looked up username) and a list of key=value pairs that were found from the userdb. The userdb should make sure that at least "uid" and "gid" keys were returned. Here's an example code based on passwd userdb: ---%<------------------------------------------------------------------------- reply = auth_stream_reply_init(auth_request); auth_stream_reply_add(reply, NULL, pw->pw_name); auth_stream_reply_add(reply, "uid", dec2str(pw->pw_uid)); auth_stream_reply_add(reply, "gid", dec2str(pw->pw_gid)); auth_stream_reply_add(reply, "home", pw->pw_dir); callback(USERDB_RESULT_OK, reply, auth_request); ---%<------------------------------------------------------------------------- (This file was created from the wiki on 2019-03-05 04:42)