This is a PAM module that lets you log in via SSH to servers using OpenID Connect credentials, instead of SSH Keys or a username and password combination.
It uses the OAuth2 Device Flow, which means that during the login process, you will click a link and log in to your OpenID Connect Provider, which will then authenticate you for the SSH session.
This module will then check if you're in the right group(s) or have a specified username, and allow or deny access.
A demo video is avaliable here: https://drive.google.com/file/d/1WzDRL0RFDXfvUgabbXNzBppV-DKXyUN1/view?usp=sharing
This code was originally developed by Mazarykova Univerzita and this branch has been refactored by UKRI-STFC as a part of the IRIS activity. This work was funded by STFC, a part of UK Research and Innovation (UKRI).
The upstream build uses basic make and we have stuck with this for compatibility reasons.
The two basic targets are make and make test; the latter will build (some of) the tests and run them.
- Note that at present some of the integration tests are failing.
- Note you can build RPMs and DEBs, though currently they are designed for CentOS7 and Ubuntu 18.04 respectively.
- The build currently requires Docker (or compatible)
yum install epel-release
yum install openldap-devel
yum install libcurl-devel
yum install pam-devel
yum install libldb-devel
yum install http://ftp.scientificlinux.org/linux/scientific/7x/external_products/softwarecollections/yum-conf-softwarecollections-2.0-1.el7.noarch.rpm
yum install devtoolset-8 # this is needed as we need a more up to date g++ version than is supplied by default in SL repos.
scl enable devtoolset-8 bash
git clone https://github.com/stfc/pam_oauth2_device.git
cd pam_oauth2_device/
make
cp pam_oauth2_device.so /lib64/security/pam_oauth2_device.so
cp config_template.json config.json
To install the module, copy pam_oauth2_device.so into the PAM modules directory (usually with permissions 0755).
On Debian-based systems, this would be /lib/x86_64-linux-gnu/security whereas CentOS and related flavours would use /usr/lib64/security. If in doubt, check dpkg --L libpam-modules or rpm -ql pam respectively.
Although JSON is not ideal for configuration files (it's a bit picky about syntax and it's hard to add comments), we now support splitting the configuration file into several segments (all of which should be JSON "objects" at the outermost level). This is useful for maintaining parts of the file such as client secrets or user maps in separate files.
In any place where a JSON object (the curly braces construct) is expected in the configuration, a JSON string can be placed instead with a filename in the string; this file will be loaded and will serve as the object in question and must have the format expected in the configuration template. This can be nested to any depth required. It is strongly recommended to use full pathnames inside these strings.
For example, if the normal configuration file looks like this (obviously a real configuration file would have more stuff in it):
{
"oauth": {
"client": {
"id": "abcd",
"secret": "meetmeinstlouis"
},
"scope": "openid profile"
}
and the system administrator decides to store the client object in a separate file, they can do it as follows:
{
"oauth": {
"client": "/etc/pam_oauth2_device/client.json",
"scope": "openid profile"
}
The client.json file should then look like this:
{
"id": "abcd",
"secret": "meetmeinstlouis"
}
In contrast, the scope cannot be split off into a separate file as a string is expected as value for the scope parameter.
Usernames are mentioned several times in this document and could probably get a bit confusing. This section attempts to give a short explanation.
For every user there are three usernames, which can be distinct.
The local user name is the name the user uses in the ssh login, as in ssh fred@example.com where the local user name is fred. This is the name that is passed into the PAM module for authentication.
The remote user name is the corresponding name for the user as held by the IAM system. Once the user has successfully authenticated to IAM, IAM publishes a "userinfo" structure with the user's name and email address and other attributes that IAM can assert. Within this structure, the PAM module can pick an attribute to use as the remote user name (using the username_attribute option).
The account name is the name of the local Unix account that the user is mapped into once they have authenticated. By default, it is the same as the local user name.
The template config_template.json should give an outline of the configuration file.
The configuration should be installed at /etc/pam_oauth2_device/config.json (a future release should make it configurable, to allow different modules to coexist).
As the name suggests, the file is in JSON, so it is recommended to check it with a JSON validator like jq after editing it (the PAM module will refuse to load an invalid JSON file, but you will not see this error till runtime.)
Thus, at the top level, there is a single object with a number of entries, described as "sections" in the table:
| Section | Entry | Type | Req'd | Description | Notes |
|---|---|---|---|---|---|
| oauth | Object | Y | |||
| oauth | client | Object | Y | Contains "id" and "secret" | |
| oauth | scope | String | Y | OIDC scope | Note 1 |
| oauth | device_endpoint | String | Y | Device endpoint | https://${url}/devicecode |
| oauth | token_endpoint | String | Y | Token endpoint | https://${url}/token |
| oauth | userinfo_endpoint | String | Y | Userinfo | https://${url}/userinfo |
| oauth | username_attribute | String | Y | Attribute for remote username | |
| oauth | local_username_suffix | String | Y | See usernames | |
| tls | Object | Y | |||
| tls | ca_bundle | String | N | Concatenated list of trust anchors | Note 2 |
| tls | ca_path | String | N | Directory with trust anchors | Note 2 |
| qr | Object | N | |||
| qr | error_correction_level | Int | Y | QR code | Note 3 |
Notes:
1 The string value should have a space separated list of scopes which must include offline_access
2 If present, the "ca_bundle" must be a file with PEM-formatted trust anchors (CA certificates) concatenated together.
- "ca_path" works only with OpenSSL
- On the target system, use
curl -Vto see whether curl uses OpenSSL or NSS - If both ca_path and ca_bundle are present, the latter takes precedence 3 The QR code section is optional but if present, it must have the error correction level defined. Permitted values are 1 (low), 2 (medium), 3 (high) or -1 (disabled). If the section is missing, the QR code is disabled. 4 The "${url}" above would be the URL (hostname) of your OpenID Provider. Its host certificate must be valid when checked against the CA bundle (see item 2)
No authorisation section is required, although if included there will be necessary entries in the section. However, if no authorisation section is provided, the user will not be able to log in.
This part of the module functionality carries a lot of legacy stuff; see the Authorisation section for discussion.
| Section | Entry | Type | Req'd | Description | Notes |
|---|---|---|---|---|---|
| ldap | Object | N | LDAP configuration | ||
| ldap | host | String | Y | LDAP URL | |
| ldap | basedn | String | Y | Base DN | Note 1 |
| ldap | user | String | Y | username | Note 2 |
| ldap | passwd | String | Y | password | Note 2 |
| ldap | scope | String | N | Note 3 | |
| ldap | preauth | String | N | ||
| ldap | filter | String | Y | ||
| ldap | attr | String | Y | ||
| group | Object | N | |||
| group | access | boolean | Y | Whether to use this section | |
| group | service_name | String | Y | Note 4 | |
| cloud | Object | N | |||
| cloud | access | boolean | Y | Whether to use this section | |
| cloud | endpoint | String | Y | Note 5 | |
| cloud | username | String | Y | endpoint username | Note 5 |
| cloud | metadata_file | String | Y | Note 5 | |
| users | Object | N | User Mapping section | Note 6 |
1 The base DN, least significant RDN first 2 Username and password are for authentication to the LDAP server, if used; if not used, just leave them as empty strings 3 scope is one of 'sub'/'subtree', 'one'/'onelevel' or 'base'/'baseobject'
- If the LDAP implementation supports 'subordinate' or 'children' (these are synonymous) then these are available as scopes as well
- The default is 'sub' 4 The service name is a string, which should name a group 5 The 'cloud' section implements a callout to a server to fetch a group membership file 6 The 'users' section provides mappings from the username attribute (selected with username_attribute) to a local user id.
The module provides a feature for bypassing authentication altogether, letting the process fall through to the next PAM module in the stack (provided the PAM authentication is configured correctly; see the HOWTO for further details.) A typical use case is to treat root logins separately.
There are currently two ways of bypassing; one is an LDAP lookup based on the "preauth" query, and the other is a
special 'users' section where the remote username is the magic string *bypass*; if the local username is in this
section, the PAM module is bypassed. See the HOWTO for further details.
- Future releases should change the
client_debugto loglevel.- Additionally, adding
debugto the PAM config should enable debug, as expected for a PAM module
- Additionally, adding
- The metadata file called
project_idcurrently has a backwards compatible default of/mnt/context/openstack/latest/meta_data.json
The major refactoring of the module in Sep 2021 preserved (and bugfixed) the existing authorisation functionality. However, the user should be warned that it is subject to revision, but generally preserving backward functionality if possible.
As above, user Fred Bloggs logs in with ssh fred@example.com. The host at example.com asks Fred to authenticate. Once successfully, it calls out to IAM to obtain the userinfo structure. From this it picks the attribute specified with username_attribute in the configuration, preferred_username, say. Let's say the value of preferred_username of Fred's userinfo structure is bloggs. Additionally, the userinfo structure contains the list of groups "users", "iris" and "cloud".
Throughout the rest of this section, it is assumed that Fred has authenticated successfully to IAM.
If the cloud section is configured and access is true, a local file configured as the metadata_file is read. This file should contain the structure
{"project\_id": "fleeps"}
The module adds the string fleeps to the endpoint (with a slash) and calls the server (with no client authentication) to see what is at the endpoint. It expects a JSON structure as response, structured like
{"groups": ["wop", "fap", "foo", "users"]}
If one of these groups matches Fred's groups as returned by the userinfo structure (it does here, "users"), then Fred is considered authorised. An additional check is made whether the local username plus suffix equals the remote username. If this check and the cloud group membership both pass, then Fred is considered authorised (the username check would fail in this example, because no suffix can make 'fred' equal to 'bloggs'.)
If the group section is configured and access is set to true, a check will be made whether the configured value for service_name is one of Fred's groups. Note the service name is single valued. Additionally, as for the cloud section, the local username plus suffix must equal the remote username.
If Fred's remote username were fred_fleeps then it would match the local username (fred) if the suffix were configured as _fleeps.
If Fred is not authorised through the cloud or group sections, either because the check failed or they were not enabled, then a configured usermap is consulted. This is written straight into the configuration file (it should probably be in its own file at some point), so would be suitable only for a smallish number of users.
This usermap is in the *users section which expects a JSON object mapping the remote username to an array of permissible local usernames. Thus, the same user could have multiple local logins using this method. If the local username is found here, Fred is considered authorised. No suffix is used in this section.
If Fred is not authorised through any of these methods, the module falls back to an LDAP lookup (if the ldap section is configured). The LDAP query takes a configured filter and substitutes the remote username for a %s part of the filter, and queries a configured attribute.
If the filter is (&(objectClass=user)(cn=%s)) then bloggs is substituted in our example, and a target attribute (configured with attr) is queried from the LDAP server. For example, if attr has the value uid then the equivalent of
ldapsearch -x -H ldap://host -b base '(&(objectClass=user)(cn=bloggs))' uid
is run and the result is compared with the local username, fred. If these match (no suffix is used), then Fred is again considered authorised.
You MUST edit the configuration before this module will work!
Make sure the module works correctly before changing your SSH config or you may be locked out! See Testing below.
Edit /etc/pam.d/sshd and comment out the other auth sections (unless you need MFA or something else).
auth required pam_oauth2_device.so
Edit /etc/ssh/sshd_config and make sure that the following configuration options are set
ChallengeResponseAuthentication yes
UsePAM yes
systemctl restart sshd
You are advised to do this before making changes to your main SSH config. There are two tests to do which are recommended to do in the order described here.
It is recommended that you create a (hard or sym) link to your sshd called (for example) pamsshd, e.g. /usr/local/sbin/pamsshd. This means you can have a PAM configuration for pamsshd which is different from the normal sshd.
In this case you can copy /etc/pam.d/sshd to /etc/pam.d/pamsshd and edit the latter, leaving the former to log you back into the system if something goes wrong. Also copy /etc/ssh/sshd_config to /etc/ssh/pamsshd_config so you can edit the configuration for pamsshd separately.
Note that testing requires that you install the module in the system location and you have the configuration set up in /etc/pam_oauth2_device/config.json and /etc/pam.d/ and /etc/ssh/.
Follow instructions above, and additionally install pamtester.
Run
pamtester -v pamtester localusername authenticate
and follow the onscreen prompts. Here, localusername refers to your local user name so replace it with whatever your name is.
You can check /var/log/secure or /var/log/auth.log to find what's wrong if there are errors authenticating. While the module
uses syslog, syslog is normally set up to log PAM stuff into one of these files.
While pamtester tests the authenticate section, you should try a proper ssh login from another host. If you created pamsshd as above (and copied configuration as described above), start it with
/usr/local/sbin/pamsshd -f /etc/ssh/pamsshd_config -p 2222 -d
This should start sshd with the name pamsshd listening on port 2222. Now try to log in from another host (bearing in mind the port should be open for incoming tcp). On the other host, run ssh -p 2222 localusername@myhost where localusername is the local user name and myhost is the host running pamsshd.
Again check the logs as in the previous tests.
- v1.03 support for segmented configuration files (as described above)
- v1.00 (commit efd50f6376a65319a2834b957a2cf6bbe6b7aa2d) - first version for full scale DiRAC test