[CASSANDRA-4875] Revert IAuthority2 interface

• Type: Improvement
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 1.1.7
• Labels: security
• Assignee: Aleksey Yeschenko Assign to me
• Reporter: Aleksey Yeschenko

Authors

Aleksey Yeschenko

Reviewers

Jonathan Ellis

• Created: 30/Oct/12 04:05
• Updated: 16/Apr/19 09:32
• Resolved: 17/Nov/12 14:35

Description

CASSANDRA-4874 is about general improvements to authorization handling, this one is about IAuthority[2] in particular.

• 'LIST GRANTS OF user should' become 'LIST PERMISSIONS [on resource] [of user]'.
  Currently there is no way to see all the permissions on the resource, only all the permissions of a particular user.
• IAuthority2.listPermissions() should return a generic collection of ResoucePermission or something, not CQLResult or ResultMessage.
  That's a wrong level of abstraction. I know this issue has been raised here - https://issues.apache.org/jira/browse/CASSANDRA-4490?focusedCommentId=13449732&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13449732com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13449732, but I think it's possible to change this. Returning a list of {resource, user, permission, grant_option}

  tuples should be possible.

• We should get rid of Permission.NO_ACCESS. An empty list of permissions should mean absence of any permission, not some magical Permission.NO_ACCESS value.
  It's insecure and error-prone and also ambiguous (what if a user has both FULL_ACCESS and NO_ACCESS permissions? If it's meant to be a way to strip a user
  of all permissions on the resource, then it should be replaced with some form of REVOKE statement. Something like 'REVOKE ALL PERMISSIONS' sounds more logical than GRANT NO_ACCESS to me.
• Previous point will probably require adding revokeAllPermissions() method to make it explicit, special-casing IAuthority2.revoke() won't do
• IAuthorize2.grant() and IAuthorize2.revoke() accept CFName instance for a resource, which has its ks and cf fields swapped if cf is omitted. This may cause a real security issue if IAuthorize2 implementer doesn't know about the issue. We must pass the resouce as a collection of strings ([cassandra, keyspaces[, ks_name][, cf_name]]) instead, the way we pass it to IAuthorize.authorize().
• We should probably get rid of FULL_ACCESS as well, at least as a valid permission value (but maybe allow it in the CQL statement) and add an equivalent IAuthority2.grantAllPermissions(), separately. Why? Imagine the following sequence: GRANT FULL_ACCESS ON resource FOR user; REVOKE SELECT ON resource FROM user; should the user be allowed to SELECT anymore?
  I say no, he shouldn't. Full access should be represented by a list of all permissions, not by a magical special value.
• P.DELETE probably should go in favour of P.UPDATE even for TRUNCATE. Presence of P.DELETE will definitely confuse users, who might think that it is somehow required to delete data, when it isn't. You can overwrite every value if you have P.UPDATE with TTL=1 and get the same result. We should also drop P.INSERT. Leave P.UPDATE (or rename it to P.MODIFY). P.MODIFY_DATA + P.READ_DATA should replace P.UPDATE, P.SELECT and P.DELETE.
• I suggest new syntax to allow setting permissions on cassandra/keyspaces resource: GRANT <permission> ON * FOR <user>.

The interface has to change because of the CFName argument to grant() and revoke(), and since it's going to be broken anyway (and has been introduced recently), I think we are in a position to make some other improvements while at it.

Issue Links

• relates to: CASSANDRA-4874 Possible authorizaton handling impovements