gitlab-org--gitlab-foss/doc/user/project/new_ci_build_permissions_model.md

New CI build permissions model

Introduced in GitLab 8.12.

GitLab 8.12 has a completely redesigned build permissions system. You can find all discussion and all our concerns when choosing the current approach in issue #18994.

---

Builds permissions should be tightly integrated with the permissions of a user who is triggering a build.

The reasons to do it like that are:

• We already have a permissions system in place: group and project membership of users.
• We already fully know who is triggering a build (using git push, using the web UI, executing triggers).
• We already know what user is allowed to do.
• We use the user permissions for builds that are triggered by the user.
• It opens a lot of possibilities to further enforce user permissions, like allowing only specific users to access runners or use secure variables and environments.
• It is simple and convenient that your build can access everything that you as a user have access to.
• Short living unique tokens are now used, granting access for time of the build and maximizing security.

With the new behavior, any build that is triggered by the user, is also marked with their permissions. When a user does a git push or changes files through the web UI, a new pipeline will be usually created. This pipeline will be marked as created be the pusher (local push or via the UI) and any build created in this pipeline will have the permissions of the pusher.

This allows us to make it really easy to evaluate the access for all projects that have Git submodules or are using container images that the pusher would have access too. The permission is granted only for time that build is running. The access is revoked after the build is finished.

Types of users

It is important to note that we have a few types of users:

• Administrators: CI builds created by Administrators will not have access to all GitLab projects, but only to projects and container images of projects that the administrator is a member of.That means that if a project is either public or internal users have access anyway, but if a project is private, the Administrator will have to be a member of it in order to have access to it via another project's build.

• External users: CI builds created by external users will have access only to projects to which user has at least reporter access. This rules out accessing all internal projects by default,

This allows us to make the CI and permission system more trustworthy. Let's consider the following scenario:

1. You are an employee of a company. Your company has a number of internal tools hosted in private repositories and you have multiple CI builds that make use of these repositories.

2. You invite a new external user. CI builds created by that user do not have access to internal repositories, because the user also doesn't have the access from within GitLab. You as an employee have to grant explicit access for this user. This allows us to prevent from accidental data leakage.

Build token

A unique build token is generated for each build and it allows the user to access all projects that would be normally accessible to the user creating that build.

We try to make sure that this token doesn't leak by:

1. Securing all API endpoints to not expose the build token.
2. Masking the build token from build logs.
3. Allowing to use the build token only when build is running.

However, this brings a question about the Runners security. To make sure that this token doesn't leak, you should also make sure that you configure your Runners in the most possible secure way, by avoiding the following:

1. Any usage of Docker's privileged mode is risky if the machines are re-used.
2. Using the shell executor since builds run on the same machine.

By using an insecure GitLab Runner configuration, you allow the rogue developers to steal the tokens of other builds.

Build triggers

Build triggers do not support the new permission model. They continue to use the old authentication mechanism where the CI build can access only its own sources. We plan to remove that limitation in one of the upcoming releases.

Before GitLab 8.12

In versions before GitLab 8.12, all CI builds would use the CI Runner's token to checkout project sources.

The project's Runner's token was a token that you could find under the project's Settings > CI/CD Pipelines and was limited to access only that project. It could be used for registering new specific Runners assigned to the project and to checkout project sources. It could also be used with the GitLab Container Registry for that project, allowing pulling and pushing Docker images from within the CI build.

---

GitLab would create a special checkout URL like:

https://gitlab-ci-token:<project-runners-token>/gitlab.com/gitlab-org/gitlab-ce.git

And then the users could also use it in their CI builds all Docker related commands to interact with GitLab Container Registry. For example:

docker login -u gitlab-ci-token -p $CI_BUILD_TOKEN registry.gitlab.com

Using single token had multiple security implications:

• The token would be readable to anyone who had developer access to a project that could run CI builds, allowing the developer to register any specific Runner for that project.
• The token would allow to access only the project's sources, forbidding from accessing any other projects.
• The token was not expiring and was multi-purpose: used for checking out sources, for registering specific runners and for accessing a project's container registry with read-write permissions.

All the above led to a new permission model for builds that was introduced with GitLab 8.12.

Making use of the new CI build permissions model

With the new build permissions model, there is now an easy way to access all dependent source code in a project. That way, we can:

1. Access a project's Git submodules
2. Access private container images
3. Access project's and submodule LFS objects

Below you can see the prerequisites needed to make use of the new permissions model and how that works with Git submodules and private Docker images hosted on the container registry.

Prerequisites to use the new permissions model

With the new permissions model in place, there may be times that your build will fail. This is most likely because your project tries to access other project's sources, and you don't have the appropriate permissions. In the build log look for information about 403 or forbidden access messages.

In short here's what you need to do should you encounter any issues.

As an administrator:

• 500 errors: You will need to update GitLab Workhorse to at least 0.8.2. This is done automatically for Omnibus installations, you need to check manually for installations from source.
• 500 errors: Check if you have another web proxy sitting in front of NGINX (HAProxy, Apache, etc.). It might be a good idea to let GitLab use the internal NGINX web server and not disable it completely. See this comment for an example.
• 403 errors: You need to make sure that your installation has HTTP(S) cloning enabled. HTTP(S) support is now a requirement by GitLab CI to clone all sources.

As a user:

• Make sure you are a member of the group or project you're trying to have access to. As an Administrator, you can verify that by impersonating the user and retry the failing build in order to verify that everything is correct.

Git submodules

It often happens that while working on one project, you need to use another project from within it; perhaps it’s a library that a third party developed or you’re developing a project separately and are using it in multiple parent projects. A common issue arises in these scenarios: you want to be able to treat the two projects as separate yet still be able to use one from within the other.

Excerpt from the Git website about submodules.

If dealing with submodules, your project will probably have a file named .gitmodules. And this is how it usually looks like:

[submodule "tools"]
	path = tools
	url = git@gitlab.com/group/tools.git

Note: If you are not using GitLab 8.12 or higher, you would need to work your way around this issue in order to access the sources of gitlab.com/group/tools (e.g., use SSH keys).

With GitLab 8.12 onward, your permissions are used to evaluate what a CI build can access. More information about how this system works can be found in the Build permissions model.

To make use of the new changes, you have to update your .gitmodules file to use a relative URL.

Let's consider the following example:

1. Your project is located at https://gitlab.com/secret-group/my-project.
2. To checkout your sources you usually use an SSH address like git@gitlab.com:secret-group/my-project.git.
3. Your project depends on https://gitlab.com/group/tools.
4. You have the .gitmodules file with above content.

Since Git allows the usage of relative URLs for your .gitmodules configuration, this easily allows you to use HTTP for cloning all your CI builds and SSH for all your local checkouts.

For example, if you change the url of your tools dependency, from git@gitlab.com/group/tools.git to ../../group/tools.git, this will instruct Git to automatically deduce the URL that should be used when cloning sources. Whether you use HTTP or SSH, Git will use that same channel and it will allow to make all your CI builds use HTTPS (because GitLab CI uses HTTPS for cloning your sources), and all your local clones will continue using SSH.

Given the above explanation, your .gitmodules file should eventually look like this:

[submodule "tools"]
	path = tools
	url = ../../group/tools.git

However, you have to explicitly tell GitLab CI to clone your submodules as this is not done automatically. You can achieve that by adding a before_script section to your .gitlab-ci.yml:

before_script:
  - git submodule update --init --recursive

test:
  script:
    - run-my-tests

This will make GitLab CI initialize (fetch) and update (checkout) all your submodules recursively.

If Git does not use the newly added relative URLs but still uses your old URLs, you might need to add git submodule sync --recursive to your .gitlab-ci.yml, prior to running git submodule update --init --recursive. This transfers the changes from your .gitmodules file into the .git folder, which is kept by runners between runs.

In case your environment or your Docker image doesn't have Git installed, you have to either ask your Administrator or install the missing dependency yourself:

# Debian / Ubuntu
before_script:
  - apt-get update -y
  - apt-get install -y git-core
  - git submodule update --init --recursive

# CentOS / RedHat
before_script:
  - yum install git
  - git submodule update --init --recursive

# Alpine
before_script:
  - apk add -U git
  - git submodule update --init --recursive

Container Registry

With the update permission model we also extended the support for accessing Container Registries for private projects.

Note: As GitLab Runner 1.6 doesn't yet incorporate the introduced changes for permissions, this makes the image: directive to not work with private projects automatically. The manual configuration by an Administrator is required to use private images. We plan to remove that limitation in one of the upcoming releases.

Your builds can access all container images that you would normally have access to. The only implication is that you can push to the Container Registry of the project for which the build is triggered.

This is how an example usage can look like:

test:
  script:
    - docker login -u gitlab-ci-token -p $CI_BUILD_TOKEN $CI_REGISTRY
    - docker pull $CI_REGISTRY/group/other-project:latest
    - docker run $CI_REGISTRY/group/other-project:latest