Permission development guidelines

There are multiple types of permissions across GitLab, and when implementing anything that deals with permissions, all of them should be considered.

Instance

User types

Each user can be one of the following types:

• Regular.
• External - access to groups and projects only if direct member.
• Internal users - system created.
• Auditor:
  • No access to projects or groups settings menu.
  • No access to Admin Area.
  • Read-only access to everything else.
• Administrator - read-write access.

See the permissions page for details on how each user type is used.

Groups and Projects

General permissions

Groups and projects can have the following visibility levels:

• public (20) - an entity is visible to everyone
• internal (10) - an entity is visible to authenticated users
• private (0) - an entity is visible only to the approved members of the entity

By default, subgroups can not have higher visibility levels. For example, if you create a new private group, it cannot include a public subgroup.

The visibility level of a group can be changed only if all subgroups and sub-projects have the same or lower visibility level. For example, a group can be set to internal only if all subgroups and projects are internal or private.

WARNING: If you migrate an existing group to a lower visibility level, that action does not migrate subgroups in the same way. This is a known issue.

Visibility levels can be found in the Gitlab::VisibilityLevel module.

Feature specific permissions

Additionally, the following project features can have different visibility levels:

• Issues
• Repository
  • Merge request
  • Forks
  • Pipelines
• Analytics
• Requirements
• Security and Compliance
• Wiki
• Snippets
• Pages
• Operations
• Metrics Dashboard

These features can be set to "Everyone with Access" or "Only Project Members". They make sense only for public or internal projects because private projects can be accessed only by project members by default.

Members

Users can be members of multiple groups and projects. The following access levels are available (defined in the Gitlab::Access module):

• No access (0)
• Minimal access (5)
• Guest (10)
• Reporter (20)
• Developer (30)
• Maintainer (40)
• Owner (50)

If a user is the member of both a project and the project parent groups, the highest permission is the applied access level for the project.

If a user is the member of a project, but not the parent groups, they can still view the groups and their entities (like epics).

Project membership (where the group membership is already taken into account) is stored in the project_authorizations table.

NOTE: In GitLab 14.9 and later, projects in personal namespaces have a maximum role of Owner. Because of a known issue in GitLab 14.8 and earlier, projects in personal namespaces have a maximum role of Maintainer.

Confidential issues

Confidential issues can be accessed only by project members who are at least reporters (they can't be accessed by guests). Additionally they can be accessed by their authors and assignees.

Licensed features

Some features can be accessed only if the user has the correct license plan.

Permission dependencies

Feature policies can be quite complex and consist of multiple rules. Quite often, one permission can be based on another.

Designing good permissions means reusing existing permissions as much as possible and making access to features granular.

In the case of a complex resource, it should be broken into smaller pieces of information and each piece should be granted a different permission.

A good example in this case is the Merge Request widget and the Security reports. Depending on the visibility level of the Pipelines, the Security reports are either visible in the widget or not. So, the Merge Request widget, the Pipelines, and the Security reports, have separate permissions. Moreover, the permissions for the Merge Request widget and the Pipelines are dependencies of the Security reports.

Permission dependencies of Secure features

Secure features have complex permissions since these features are integrated into different features like Merge Requests and CI flow.

Here is a list of some permission dependencies.

Where should permissions be checked?

We should typically apply defense-in-depth (implementing multiple checks at various layers) starting with low-level layers, such as finders and services, followed by high-level layers, such as GraphQL, public REST API, and controllers.

See Guidelines for reusing abstractions.

Protecting the same resources at many points means that if one layer of defense is compromised or missing, customer data is still protected by the additional layers.

See the permissions section in the Secure Coding Guidelines.

Considerations

Services or finders are appropriate locations because:

• Multiple endpoints share services or finders so downstream logic is more likely to be re-used.
• Sometimes authorization logic must be incorporated in DB queries to filter records.
• Permission checks at the display layer should be avoided except to provide better UX and not as a security check. For example, showing and hiding non-data elements like buttons.

The downsides to defense-in-depth are:

• DeclarativePolicy rules are relatively performant, but conditions may perform database calls.
• Higher maintenance costs.

Exceptions

Developers can choose to do authorization in only a single area after weighing the risks and drawbacks for their specific case.

Prefer domain logic (services or finders) as the source of truth when making exceptions.

Logic, like backend worker logic, might not need authorization based on the current user. If the service or finder's constructor does not expect current_user, then it typically won't check permissions.

Tips

If a class accepts current_user, then it may be responsible for authorization.

Example: Adding a new API endpoint

By default, we authorize at the endpoint. Checking an existing ability may make sense; if not, then we probably need to add one.

As an aside, most endpoints can be cleanly categorized as a CRUD (create, read, update, destroy) action on a resource. The services and abilities follow suit, which is why many are named like Projects::CreateService or :read_project.

Say, for example, we extract the whole endpoint into a service. The can? check will now be in the service. Say the service reuses an existing finder, which we are modifying for our purposes. Should we make the finder check an ability?

• If the finder doesn't accept current_user, and therefore doesn't check permissions, then probably no.
• If the finder accepts current_user, and doesn't check permissions, then it would be a good idea to double check other usages of the finder, and we might consider adding authorization.
• If the finder accepts current_user, and already checks permissions, then either we need to add our case, or the existing checks are appropriate.

Refactoring permissions

Finding existing permissions checks

As mentioned above, permissions are often checked in multiple locations for a single endpoint or web request. As a result, finding the list of authorization checks that are run for a given endpoint can be challenging.

To assist with this, you can locally set GITLAB_DEBUG_POLICIES=true.

This outputs information about which abilities are checked in the requests made in any specs that you run. The output also includes the line of code where the authorization check was made. Caller information is especially helpful in cases where there is metaprogramming used because those cases are difficult to find by grepping for ability name strings.

Example:

# example spec run

GITLAB_DEBUG_POLICIES=true bundle exec rspec spec/controllers/groups_controller_spec.rb:162

# permissions debug output when spec is run; if multiple policy checks are run they will all be in the debug output.

POLICY CHECK DEBUG -> policy: GlobalPolicy, ability: create_group, called_from: ["/gitlab/app/controllers/application_controller.rb:245:in `can?'", "/gitlab/app/controllers/groups_controller.rb:255:in `authorize_create_group!'"]

This flag is meant to help learn more about authorization checks while refactoring and should not remain enabled for any specs on the default branch.

Understanding logic for individual abilities

References to an ability may appear in a DeclarativePolicy class many times and depend on conditions and rules which reference other abilities. As a result, it can be challenging to know exactly which conditions apply to a particular ability.

DeclarativePolicy provides a ability_map for each Policy class, which pulls all Rules for an ability into an array.

Example:

> GroupPolicy.ability_map.map.select { |k,v| k == :read_group_member }
=> {:read_group_member=>[[:enable, #<Rule can?(:read_group)>], [:prevent, #<Rule ~can_read_group_member>]]}

> GroupPolicy.ability_map.map.select { |k,v| k == :read_group }
=> {:read_group=>
  [[:enable, #<Rule public_group>],
   [:enable, #<Rule logged_in_viewable>],
   [:enable, #<Rule guest>],
   [:enable, #<Rule admin>],
   [:enable, #<Rule has_projects>],
   [:enable, #<Rule read_package_registry_deploy_token>],
   [:enable, #<Rule write_package_registry_deploy_token>],
   [:prevent, #<Rule all?(~public_group, ~admin, user_banned_from_group)>],
   [:enable, #<Rule auditor>],
   [:prevent, #<Rule needs_new_sso_session>],
   [:prevent, #<Rule all?(ip_enforcement_prevents_access, ~owner, ~auditor)>]]}

DeclarativePolicy also provides a debug method that can be used to understand the logic tree for a specific object and actor. The output is similar to the list of rules from ability_map. But, DeclarativePolicy stops evaluating rules once one prevents an ability, so it is possible that not all conditions are called.

Example:

policy = GroupPolicy.new(User.last,  Group.last)
policy.debug(:read_group)

- [0] enable when public_group ((@custom_guest_user1 : Group/139))
- [0] enable when logged_in_viewable ((@custom_guest_user1 : Group/139))
- [0] enable when admin ((@custom_guest_user1 : Group/139))
- [0] enable when auditor ((@custom_guest_user1 : Group/139))
- [14] prevent when all?(~public_group, ~admin, user_banned_from_group) ((@custom_guest_user1 : Group/139))
- [14] prevent when needs_new_sso_session ((@custom_guest_user1 : Group/139))
- [16] enable when guest ((@custom_guest_user1 : Group/139))
- [16] enable when has_projects ((@custom_guest_user1 : Group/139))
- [16] enable when read_package_registry_deploy_token ((@custom_guest_user1 : Group/139))
- [16] enable when write_package_registry_deploy_token ((@custom_guest_user1 : Group/139))
  [21] prevent when all?(ip_enforcement_prevents_access, ~owner, ~auditor) ((@custom_guest_user1 : Group/139))

=> #<DeclarativePolicy::Runner::State:0x000000015c665050
 @called_conditions=
  #<Set: {
   "/dp/condition/GroupPolicy/public_group/Group:139",
   "/dp/condition/GroupPolicy/logged_in_viewable/User:83,Group:139",
   "/dp/condition/BasePolicy/admin/User:83",
   "/dp/condition/BasePolicy/auditor/User:83",
   "/dp/condition/GroupPolicy/user_banned_from_group/User:83,Group:139",
   "/dp/condition/GroupPolicy/needs_new_sso_session/User:83,Group:139",
   "/dp/condition/GroupPolicy/guest/User:83,Group:139",
   "/dp/condition/GroupPolicy/has_projects/User:83,Group:139",
   "/dp/condition/GroupPolicy/read_package_registry_deploy_token/User:83,Group:139",
   "/dp/condition/GroupPolicy/write_package_registry_deploy_token/User:83,Group:139"}>,
 @enabled=false,
 @prevented=true>

Testing that individual policies are equivalent

You can use the 'equivalent project policy abilities' shared example to ensure that 2 project policy abilities are equivalent for all project visibility levels and access levels.

Example:

  context 'when refactoring read_pipeline_schedule and read_pipeline' do
    let(:old_policy) { :read_pipeline_schedule }
    let(:new_policy) { :read_pipeline }

    it_behaves_like 'equivalent policies'
  end