- Groups and Projects
- Permission dependencies
- Where should permissions be checked?
There are multiple types of permissions across GitLab, and when implementing anything that deals with permissions, all of them should be considered.
Each user can be one of the following types:
- External - access to groups and projects only if direct member.
- Internal users - system created.
- 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 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.
Visibility levels can be found in the
Additionally, the following project features can have different visibility levels:
- Merge request
- Security & Compliance
- 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.
Users can be members of multiple groups and projects. The following access
levels are available (defined in the
- No access (
Minimal access (
- Guest (
- Reporter (
- Developer (
- Maintainer (
- Owner (
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
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.
Some features can be accessed only if the user has the correct license plan.
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.
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.
|Activity level||Resource||Locations||Permission dependency|
|View||License information||Dependency list, License Compliance||Can view repository|
|View||Dependency information||Dependency list, License Compliance||Can view repository|
|View||Vulnerabilities information||Dependency list||Can view security findings|
|View||Black/Whitelisted licenses for the project||License Compliance, merge request||Can view repository|
|View||Security findings||merge request, CI job page, Pipeline security tab||Can read the project and CI jobs|
|View||Vulnerability feedback||merge request||Can read security findings|
|View||Dependency List page||Project||Can access Dependency information|
|View||License Compliance page||Project||Can access License information|
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.
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.
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:
DeclarativePolicyrules are relatively performant, but conditions may perform database calls.
- Higher maintenance costs.
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
If a class accepts
current_user, then it may be responsible for authorization.
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
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.