Arvados

Pushed a branch with a test: 16736-expiring-tokens-limits

Updates at 386b60af1 - branch 16736-expiring-tokens-limits
Test run: developer-run-tests: #2078

• Adds tests exposing the bug.
• Adds checks related to token's expires_at on create/update.
• Fixes fixture removing expiration dates that were set far in the future.

Lucas Di Pentima wrote:

Updates at 386b60af1 - branch 16736-expiring-tokens-limits
Test run: developer-run-tests: #2078

• Adds tests exposing the bug.
• Adds checks related to token's expires_at on create/update.
• Fixes fixture removing expiration dates that were set far in the future.

  def permission_to_create
    current_user.andand.is_admin or
      ((current_user.andand.id == self.user_id) and
        (current_api_client_authorization.andand.expires_at.nil? or
          (self.expires_at and current_api_client_authorization.expires_at >= self.expires_at)))
  end

I don't think this is quite right, this will allow an admin to always create a token even if their token is about to expire. But that's not right. Admins can create tokens for other users, but the tokens are still capped by expires_at.

Also, instead of "current_api_client_authorization.andand", if it is possible for current_api_client_authorization to be nil I don't think we allow tokens to be manipulated (is there a special case, like tests?)

I think this should be:

  def permission_to_create
    (current_user.andand.is_admin ||
      current_user.andand.id == self.user_id) and
        !current_api_client_authorization.nil? and
        (current_api_client_authorization.expires_at.nil? ||
          (!self.expires_at.nil? && current_api_client_authorization.expires_at >= self.expires_at)))
  end

By the way, I looked this up recently, the difference between "and"/"or" and "&&"/"||" in Ruby has to do with operator precedence, "and"/"or" have lower precedence than "&&"/"||" so the convention is to use the and/or for the "outer" clauses and &&/|| for the "inner" clauses although when grouping with parenthesis it doesn't really matter.

Arvados::V1::ApiClientAuthorizationsController#find_objects_for_index should include a filter on expires_at when current_api_client_authorization is non-nil, and there should be a controller test for that.

update: I think you mentioned on gitter than current_api_client_authorization is nil in the case of the login process, which makes sense. We are creating a token, we don't have one yet.

However, if there was a bug that allowed some other way of calling token creation without setting current_api_client_authorization, that would be a serious security hole.

I wonder if there's a way to handle login as a special case. The value of current_api_client_authorization comes from Thread.current[:api_client_authorization], so the login code could assign a fake (not persisted to database) ApiClientAuthorization object.

When Login.TokenLifetime is set:

• Token from login uses "Login.TokenLifetime" expiration (existing behavior)
• (valid) token from login can be used to create new tokens
  • If no expiration time is given, it is the "maximum" expiration (new configuration option API.MaximumTokenLifetime)
  • SystemRootToken can create tokens with any expiration (or nil)
• Token created to run a container doesn't have a set expire time, because it will expire when the container ends
• Token used on a shell node is max lifetime, when about to expire it is rotated by the arvados-login-sync script
• "Show current token" dialog in workbench should generate a new token instead
  • Change the label to something like "Get new user token"
• Auto-logout should coordinate across browser tabs/windows (using the same token?) so that it only logs out when all of them have been idle.
  • Figure out which browser feature lets you do that

Updates at bb5d1bb61 - branch 16736-max-token-lifetime
Test run: developer-run-tests: #2301

• Adds API.MaxTokenLifetime configuration.
• Limits token's expires_at when the configuration is set and users don't provide a valid expiration date on creation or update.
• Adds options to arvados-login-sync to allow the admin to set it up in a way that periodically rotates expiring tokens on shell sessions.

• clamp_token_expiration:

max_token_expiration = Time.now + Rails.configuration.API.MaxTokenLifetime

I think we want to be using db_current_time here?

Related, I see another use of Time.now in the validate method which I believe should also be db_current_time

• I'm not sure I see the purpose of elsif here:

      if self.new_record? && (self.expires_at.nil? || self.expires_at > max_token_expiration)
        self.expires_at = max_token_expiration
      elsif !self.new_record? && self.expires_at_changed? && (self.expires_at.nil? || self.expires_at > max_token_expiration)
        self.expires_at = max_token_expiration
      end

Would this do the same thing?

if (self.new_record? || self.expires_at_changed?) && (self.expires_at.nil? || self.expires_at > max_token_expiration)
   self.expires_at = max_token_expiration
end

• Did you confirm that tokens created for containers do not have an expiration set? Is there a test to check that? (it is unclear if the container assign_auth method is guaranteed to be called by the system user).

Updates at 46a5e39 (rebased) - branch 16736-max-token-lifetime
Test run: developer-run-tests: #2327

Peter Amstutz wrote:

I think we want to be using db_current_time here?
Related, I see another use of Time.now in the validate method which I believe should also be db_current_time

Yes, thank you, fixed!

• I'm not sure I see the purpose of elsif here:
  Would this do the same thing?

Fixed too.

• Did you confirm that tokens created for containers do not have an expiration set? Is there a test to check that? (it is unclear if the container assign_auth method is guaranteed to be called by the system user).

As I mentioned on today's standup: the assign_auth case doesn't get a clamped expires_at because that code is run with dispatcher credentials when the container is locked (added comment for clarification).
In the case of a federated call, the controller creates the runtime token so it also doesn't get clamped.
I added a test for the assign_auth case and expanded a preexisting one for the federated call case.

Lucas Di Pentima wrote:

Updates at 46a5e39 (rebased) - branch 16736-max-token-lifetime
Test run: developer-run-tests: #2327

Peter Amstutz wrote:

I think we want to be using db_current_time here?
Related, I see another use of Time.now in the validate method which I believe should also be db_current_time

Yes, thank you, fixed!

• I'm not sure I see the purpose of elsif here:
  Would this do the same thing?

Fixed too.

• Did you confirm that tokens created for containers do not have an expiration set? Is there a test to check that? (it is unclear if the container assign_auth method is guaranteed to be called by the system user).

As I mentioned on today's standup: the assign_auth case doesn't get a clamped expires_at because that code is run with dispatcher credentials when the container is locked (added comment for clarification).
In the case of a federated call, the controller creates the runtime token so it also doesn't get clamped.
I added a test for the assign_auth case and expanded a preexisting one for the federated call case.

This LGTM, thanks.