gitlab-org--gitlab-foss/doc/development/policies.md

stage	group	info
Manage	Access	To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

DeclarativePolicy framework

The DeclarativePolicy framework is designed to assist in performance of policy checks, and to enable ease of extension for EE. The DSL code in app/policies is what Ability.allowed? uses to check whether a particular action is allowed on a subject.

The policy used is based on the subject's class name - so Ability.allowed?(user, :some_ability, project) creates a ProjectPolicy and check permissions on that.

Managing Permission Rules

Permissions are broken into two parts: conditions and rules. Conditions are boolean expressions that can access the database and the environment, while rules are statically configured combinations of expressions and other rules that enable or prevent certain abilities. For an ability to be allowed, it must be enabled by at least one rule, and not prevented by any.

Conditions

Conditions are defined by the condition method, and are given a name and a block. The block is executed in the context of the policy object - so it can access @user and @subject, as well as call any methods defined on the policy. Note that @user may be nil (in the anonymous case), but @subject is guaranteed to be a real instance of the subject class.

class FooPolicy < BasePolicy
  condition(:is_public) do
    # @subject guaranteed to be an instance of Foo
    @subject.public?
  end

  # instance methods can be called from the condition as well
  condition(:thing) { check_thing }

  def check_thing
    # ...
  end
end

When you define a condition, a predicate method is defined on the policy to check whether that condition passes - so in the above example, an instance of FooPolicy also responds to #is_public? and #thing?.

Conditions are cached according to their scope. Scope and ordering is covered later.

Rules

A rule is a logical combination of conditions and other rules, that are configured to enable or prevent certain abilities. It is important to note that the rule configuration is static - a rule's logic cannot touch the database or know about @user or @subject. This allows us to cache only at the condition level. Rules are specified through the rule method, which takes a block of DSL configuration, and returns an object that responds to #enable or #prevent:

class FooPolicy < BasePolicy
  # ...

  rule { is_public }.enable :read
  rule { thing }.prevent :read

  # equivalently,
  rule { is_public }.policy do
    enable :read
  end

  rule { ~thing }.policy do
    prevent :read
  end
end

Within the rule DSL, you can use:

• A regular word mentions a condition by name - a rule that is in effect when that condition is truthy.
• ~ indicates negation, also available as negate.
• & and | are logical combinations, also available as all?(...) and any?(...).
• can?(:other_ability) delegates to the rules that apply to :other_ability. Note that this is distinct from the instance method can?, which can check dynamically - this only configures a delegation to another ability.

~, & and | operators are overridden methods in DeclarativePolicy::Rule::Base.

Do not use boolean operators such as && and || within the rule DSL, as conditions within rule blocks are objects, not booleans. The same applies for ternary operators (condition ? ... : ...), and if blocks. These operators cannot be overridden, and are hence banned via a custom cop.

Scores, Order, Performance

To see how the rules get evaluated into a judgment, it is useful in a console to use policy.debug(:some_ability). This prints the rules in the order they are evaluated.

For example, let's say you wanted to debug IssuePolicy. You might run the debugger in this way:

user = User.find_by(username: 'john')
issue = Issue.first
policy = IssuePolicy.new(user, issue)
policy.debug(:read_issue)

An example debug output would look as follows:

- [0] prevent when all?(confidential, ~can_read_confidential) ((@john : Issue/1))
- [0] prevent when archived ((@john : Project/4))
- [0] prevent when issues_disabled ((@john : Project/4))
- [0] prevent when all?(anonymous, ~public_project) ((@john : Project/4))
+ [32] enable when can?(:reporter_access) ((@john : Project/4))

Each line represents a rule that was evaluated. There are a few things to note:

1. The - or + symbol indicates whether the rule block was evaluated to be false or true, respectively.
2. The number inside the brackets indicates the score.
3. The last part of the line (for example, @john : Issue/1) shows the username and subject for that rule.

Here you can see that the first four rules were evaluated false for which user and subject. For example, you can see in the last line that the rule was activated because the user john had the Reporter role on Project/4.

When a policy is asked whether a particular ability is allowed (policy.allowed?(:some_ability)), it does not necessarily have to compute all the conditions on the policy. First, only the rules relevant to that particular ability are selected. Then, the execution model takes advantage of short-circuiting, and attempts to sort rules based on a heuristic of how expensive they are to calculate. The sorting is dynamic and cache-aware, so that previously calculated conditions are considered first, before computing other conditions.

Note that the score is chosen by a developer via the score: parameter in a condition to denote how expensive evaluating this rule would be relative to other rules.

Scope

Sometimes, a condition only uses data from @user or only from @subject. In this case, we want to change the scope of the caching, so that we don't recalculate conditions unnecessarily. For example, given:

class FooPolicy < BasePolicy
  condition(:expensive_condition) { @subject.expensive_query? }

  rule { expensive_condition }.enable :some_ability
end

Naively, if we call Ability.allowed?(user1, :some_ability, foo) and Ability.allowed?(user2, :some_ability, foo), we would have to calculate the condition twice - since they are for different users. But if we use the scope: :subject option:

  condition(:expensive_condition, scope: :subject) { @subject.expensive_query? }

then the result of the condition is cached globally only based on the subject - so it is not calculated repeatedly for different users. Similarly, scope: :user caches only based on the user.

DANGER: If you use a :scope option when the condition actually uses data from both user and subject (including a simple anonymous check!) your result is cached at too global of a scope and results in cache bugs.

Sometimes we are checking permissions for a lot of users for one subject, or a lot of subjects for one user. In this case, we want to set a preferred scope - that is, tell the system that we prefer rules that can be cached on the repeated parameter. For example, in Ability.users_that_can_read_project:

def users_that_can_read_project(users, project)
  DeclarativePolicy.subject_scope do
    users.select { |u| allowed?(u, :read_project, project) }
  end
end

This, for example, prefers checking project.public? to checking user.admin?.

Delegation

Delegation is the inclusion of rules from another policy, on a different subject. For example:

class FooPolicy < BasePolicy
  delegate { @subject.project }
end

includes all rules from ProjectPolicy. The delegated conditions are evaluated with the correct delegated subject, and are sorted along with the regular rules in the policy. Note that only the relevant rules for a particular ability are actually considered.

Overrides

We allow policies to opt-out of delegated abilities.

Delegated policies may define some abilities in a way that is incorrect for the delegating policy. Take for example a child/parent relationship, where some abilities can be inferred, and some cannot:

class ParentPolicy < BasePolicy
  condition(:speaks_spanish) { @subject.spoken_languages.include?(:es) }
  condition(:has_license) { @subject.driving_license.present? }
  condition(:enjoys_broccoli) { @subject.enjoyment_of(:broccoli) > 0 }

  rule { speaks_spanish }.enable :read_spanish
  rule { has_license }.enable :drive_car
  rule { enjoys_broccoli }.enable :eat_broccoli
  rule { ~enjoys_broccoli }.prevent :eat_broccoli
end

Here, if we delegated the child policy to the parent policy, some values would be incorrect - we might correctly infer that the child can speak their parent's language, but it would be incorrect to infer that the child can drive or would eat broccoli just because the parent can and does.

Some of these things we can deal with - we can forbid driving universally in the child policy, for example:

class ChildPolicy < BasePolicy
  delegate { @subject.parent }

  rule { default }.prevent :drive_car
end

But the food preferences one is harder - because of the prevent call in the parent policy, if the parent dislikes it, even calling enable in the child does not enable :eat_broccoli.

We could remove the prevent call in the parent policy, but that still doesn't help us, since the rules are different: parents get to eat what they like, and children eat what they are given, provided they are well behaved. Allowing delegation would end up with only children whose parents enjoy green vegetables eating it. But a parent may well give their child broccoli, even if they dislike it themselves, because it is good for their child.

The solution it to override the :eat_broccoli ability in the child policy:

class ChildPolicy < BasePolicy
  delegate { @subject.parent }

  overrides :eat_broccoli

  condition(:good_kid) { @subject.behavior_level >= Child::GOOD }

  rule { good_kid }.enable :eat_broccoli
end

With this definition, the ChildPolicy never looks in the ParentPolicy to satisfy :eat_broccoli, but it will use it for any other abilities. The child policy can then define :eat_broccoli in a way that makes sense for Child and not Parent.

Alternatives to using overrides

Overriding policy delegation is complex, for the same reason delegation is complex - it involves reasoning about logical inference, and being clear about semantics. Misuse of override has the potential to duplicate code, and potentially introduce security bugs, allowing things that should be prevented. For this reason, it should be used only when other approaches are not feasible.

Other approaches can include for example using different ability names. Choosing to eat a food and eating foods you are given are semantically distinct, and they could be named differently (perhaps chooses_to_eat_broccoli and eats_what_is_given in this case). It can depend on how polymorphic the call site is. If you know that we always check the policy with a Parent or a Child, then we can choose the appropriate ability name. If the call site is polymorphic, then we cannot do that.

Specifying Policy Class

You can also override the Policy used for a given subject:

class Foo

  def self.declarative_policy_class
    'SomeOtherPolicy'
  end
end

This uses and checks permissions on the SomeOtherPolicy class rather than the usual calculated FooPolicy class.