Security

Description

How does Zope handle permissions, roles and users?

Much of Zope security is implemented in C, for speed, but there is a Python implementation in AccessControl.ImplPython, which can be enabled by setting security-policy-implementation python in zope.conf.

Note: We will not discuss RestrictedPython, used to apply security restrictions to through-the-web python scripts and page templates, here.

Declaring object roles and attribute permissions

The permissions required to access a given attribute are stored on classes and modules in a variable called __ac_permissions__. This contains a tuple of tuples that map a permission name to a list of attributes (e.g. methods) protected by that permission, e.g.:

__ac_permissions__=(

('View management screens',
 ['manage','manage_menu','manage_main','manage_copyright',
  'manage_tabs','manage_propertiesForm','manage_UndoForm']),
('Undo changes',       ['manage_undo_transactions']),
('Change permissions', ['manage_access']),
('Add objects',        ['manage_addObject']),
('Delete objects',     ['manage_delObjects']),
('Add properties',     ['manage_addProperty']),
('Change properties',  ['manage_editProperties']),
('Delete properties',  ['manage_delProperties']),
('Default permission', ['']),
)

The roles reuqired to access an object (e.g. a content object), are stored in a class or instance variable __roles__. This may contain a tuple or list of role names, an AccessControl.PermissionRole.PermissionRole object, or one of the following special variables:

AccessControl.SecurityInfo.ACCESS_NONE

Inaccessible from any context

AccessControl.SecurityInfo.ACCESS_PRIVATE

Accessible only from Python code

AccessControl.SecurityInfo.ACCESS_PUBLIC

Accessible from restricted Python code and publishable through the web (provided the object has a docstring)

For attributes (including methods), the roles are stored on the parent class in a variable called <name>__roles__, where <name> is the attribute name. Again, the special variables ACCESS_NONE, ACCESS_PRIVATE and ACCESS_PUBLIC can be used.

These variables are rarely set manually. Instead, declarative security info is typically used. For example:

from App.class_init import InitializeClass
from AccessControl.SecurityInfo import ClassSecurityInfo
from OFS.SimpleItem import Item

class SomeClass(Item):

  ...

  security = ClassSecurityInfo()
  security.declareObjectPublic() # like __roles__ = ACCESS_PUBLIC

  security.declareProtected('Some permission, 'someMethod')
  def someMethod(self):
    ...

  InitializeClass(SomeClass)

There is also security.declareObjectProtected(<permission>), security.declareObjectPrivate(), security.declarePrivate(<attribute>) and security.declarePublic(attribute), which do as their names suggest to make an object or attribute protected, private or public.

Attribute security can be set in ZCML using the <class /> directive with one or more <require /> sub-directives:

<class class=".someclass.SomeClass">
  <require
    permission="some.permission"
    attributes="someMethod"
    />
</class>

Behind the scenes, this simply creates a ClassSecurityInfo and invokes it on the attributes listed as applicable. This will also call InitializeClass on the given class.

Note that the <require /> directive, in common with all ZCML directives, uses ZTK-style permission names, not Zope 2-style permission strings. A ZTK permission is a named utility providing zope.security.interfaces.IPermission, with an id that is the short (usually dotted) name that is also the utility name, and a title that matches the Zope 2 name. New permissions can be registered using the <permission /> directive:

<permission
  id="some.permission"
  title="Some permission"
  />

Zope 2-style permission names spring into existence whenever used in a security declaration, which makes them susceptibly to typos (ZTK-style IPermission utilities must be explicitly registered before they can be used).

Permissions are also represented by “mangled” permission names, which simply turn the arbitrary string name of a permission into a valid Python identifier. For example, the permission "Access contents information" becomes _Access_contents_information_Permission. The mangling is done by the function AccessControl.Permission.pname.

ClassSecurityInfo does little except record information until the InitializeClass() call is made with the class as an arugment. This will:

  • Loop over all attributes and assign a __name__ attribute to the value of any attribute in the class’s __dict__ that has the _need__name__ marker set (this is used by through-the-web DTML and Zope Page Template objects that may not have a name until they are assigned to their parent).

  • Look for any function with the name manage() or a name starting with manage_. If this does not have a corresponding <name>__roles__ attribute, one is created with the roles ('Manager',), as a way to automatically protect such methods.

  • Look for any security info object (i.e. an attribute that has an attribute __security_info__). If one is found call its apply() method with the class as an argument, and then delete it.

    The apply() method of ClassSecurityInfo does this:

    • Collect any explicitly set __ac_permissions__ tuple and turn it into internal state, as if the ClassSecurityInfo had been used to set it, so that it is not lost.

    • For any attribute declared with declarePublic() or declarePrivate(), set <name>__roles__ to ACCESS_PUBLIC or ACCESS_PRIVATE as appropriate.

    • Build an __ac_permissions__ tuple from the saved declarations of any protected attributes.

      As a special case, a call to security.declareObjectProtected(<permission>) will result in a value stored with an empty attribute name, which later translates as setting __roles__ directly on the class.

  • Find any __ac_permissions__ on the class (probably created by the security info apply() call) and call AccessControl.Permission.registerPermissions with it as an argument. This will register the permission in a global list of known permissions with their default roles (usually ('Manager',)) held in that module under the variable _ac_permissions. The mangled permission name (see above) will also be set as a class attribute on the class AccessControl.Permission.ApplicationDefaultPermissions, which is a base class of the application root (OFS.Application.Application), hence making the mangled permission names available as (acquirable) class attributes on the application root. The value of this class variable is a tuple with the default roles for that permission.

  • For all permissions in __ac_permissions__ and for all attribute (method) names assigned to each permission, set a class attribute <name>__roles__ to a PermissionRole object. If a default list/tuple of roles was supplied, record this in the PermissionRole, otherwise default to ('Manager',).

Determining which roles have a given permission

To perform security checks, it is necessary to compare the roles a user has with the roles required for a given permission. The method to determine the roles of a permission on a given object is called rolesForPermissionOn(). It is found in AccessControl.ImplPython, though a C implementation may also be in use.

rolesForPermissionOn() can be called directly, but it should be imported from AccessControl.PermissionRole to ensure the correct implementation (C or Python) is used. Alternatively, the correct implementation can be accessed by using the rolesForPermissionOn() method of a PermissionRole object, which will supply the correct permission name and default roles.

The default rolesForPermissionOn() does the following:

  • Mangle the permission name (see above).

  • Walk from the object up the inner (containment) acquisition chain to find an object with the mangled permission name as an attribute. Then:

    • If the attribute is None, this is actually the ACCESS_PUBLIC marker. Return ('Anonymous',).

    • If the sequence of roles is a tuple, this is a signal to not acquire roles from parent objects. Stop and return any roles collected by walking the acquisition chain so far plus the roles at the current object.

    • If the sequence of roles is a list, this is a signal to acquire roles from parent objects. Hence, collect the roles at the current object and continue the walk up the acquisition chain.

    • If roles is a string, assumed to be a different mangled permission name, this is a signal to delegate to another permission. Continue acquisition from the parent, but discard any roles acquired so far.

  • If no object with the managled permission attribute is found, return the default roles. Applicable default roles are stored in each PermissionRole object, but for other types of roles, use ('Manager',).

  • In all cases, if the global variable _embed_permission_in_roles is true, include the mangled permission name in the list of roles returned (even if an empty list). This is used as a debugging aid.

Checking a permission in a context

The most basic permission check can be done using:

from AccessControl import getSecurityManager
sm = getSecurityManager()
sm.checkPermission('Some permission', someObject)

This returns either 1 or None to indicate whether the current user has such a permission.

The call to getSecurityManager() returns a security manager instance for the current request. A security manager is created using newSecurityManager() in the validated_hook at the end of traversal (hence note that it is not set during traversal itself; specifically it is not set when a view adapter is being looked up and instantiated and so there is no security information available in the __init__() of a view), which creates a new security manager with a context that is aware of the current authenticated user (or Anonymous if there is none).

Again, the security manager may use a C implementation, but the default one is defined in AccessControl.ImplPython. The two most important methods on this object are checkPermission() (seen above) and validate(), which is used during traversal to validate access to an object and will throw an Unauthorized exception if not valid. Both of these delegate to a security policy, which will invariably be the ZopeSecurityPolicy also found in ImplPython (or C code) and instantiated once with a module-level call to setDefaultBehaviors().

The checkPermission() implementation in ZopeSecurityPolicy is relatively simple. It uses rolesForPermissionOn() to discover the roles on the object, and then obtains the current user from the security context (passed as a parameter to its version of checkPermission()) and calls the user object’s allowed() method with the object and its roles.

Additionally, if the security policy allows for it (which it will by default), checks are made to ensure that if the “execution context” has an owner (e.g. it is a through-the-web Python script or template owned by a particular user), the owner as well as the current user has the appropriate roles, otherwise access is disallowed. Also, if proxy roles are set (again applicable to through-the-web scripts), these are allowed to be used in lieu of the user’s actual roles.

There are various user implementations that can treat allowed() differently. The most common use in Plone is the PropertiedUser from Products.PluggableAuthService (PAS), though there is also a basic implementation in AccessControl.users.BasicUser, and a class called SpecialUser in the same module that is used for the Anonymous user.

The PAS version is only marginally more complex than the BasicUser implementation (it deals with roles obtained from groups a user belongs to), so we will describe the allowed() implementation from BasicUser here:

  • If the object’s required roles is the special variable _what_not_even_god_should_do (you couldn’t make this up), which corresponds to the ACCESS_NONE security declaration (as used by declareObjectPrivate()), immediately disallow access.

  • If the object’s required roles is None, which corresponds to the ACCESS_PUBLIC security declaration (as used by declareObjectPublic()), or if Anonymous is one of the roles (even if the user is not Anonymous), immediately allow access.

  • If Authenticated is one of the required roles and the user is not Anonymous, immediately allow access unless the object does not share an acquisition parent with the user folder (this is to avoid users with the same id in different user folders trying to steal each other’s access through acquisition tricks). This is referred to as the “context check” below.

  • Check if the user’s global roles intersect with the roles required to access the object, and allow access if the user passes the context check.

  • Check if there are any local roles, as defined in the attribute __ac_local_roles__, granted to the user and check these against the required roles (and perform the context check). __ac_local_roles__ may be a dict or a callable that returns a dict, containing a mapping of user (or group, if PAS is used,) ids to local roles granted. The local role check is performed iteratively by walking up the acquisition chain and checking the instances of bound methods, unti the root of the acquisition chain.

  • If none of the above succeed, return None to indicate that the user is not allowed to access the object.

Validating access to an object

The second type of security operation provided by the SecurityManager is to check whether the user should be able to access a particular context. This is most commonly used during traversal, by way of the user folder’s validate() method. The version in Products.PluggableAuthService.PluggableAuthService does this:

  • Get all applicable user ids from the request. Most likely, there is only one, but PAS’s modular nature means it is possible more than one plugin will supply a user id.

  • Extract the following information from the published object (REQUEST['published']):

    • accessed, the object the published object was accessed through, i.e. the first traversal parent (request['PARENTS'][0]).

    • container, the physical container of the object, i.e. the inner acquisition parent. If the published object is a method, the container is also set to be the method, but stripped of any outer acquisition chains by a call to aq_inner(). If the published object does not have an inner acquisition parent, the traversal parent is used in the same way as it is used to set accessed.

    • name, the name used to access the object, e.g. a traverasl path element.

    • value, the object we are validating access to, i.e. the published object.

  • If this is the top level user folder and the user is the emergency user, return the user immediately without further authorisation.

  • Otherwise, attempt to authorise the user by creating a new security manager for this user and calling its validate() method with``accessed``, container, name, and value as arguments.

The default security manager validate() method delegates to the equivalent method on the ZopeSecurityPolicy. This is a charming 200+ line bundle of if statements that does something like this:

  • If the name is an aq_* attribute other than aq_parent, aq_inner or aq_explicit, raise Unauthorized.

  • Obtain the aq_base’d version of container and accessed. If the accessed parent was not acquisition-wrapped, treat the aq_base’d container as the aq_base’d accessed.

  • The caller may have passed in the required roles already as an optimisation. If not, attempt to get the required roles by calling getRoles(container, name, value). The Python version of this is defined in AccessControl.ZopeSecurityPolicy. It does the following:

    • If the value has a __roles__ attribute, and it is None (ACCESS_PUBLIC) or a list or tuple of roles, return them. (This probably means the value is a content object or similar.)

    • If it is a PermissionRole object or another object with a rolesForPermissionOn() method (described above), call this with the value as an argument and return the results. (This probably means the value is a method.)

    • If there is no __roles__ attribute, check if we have a name. Return “no roles” if not.

    • Attempt to find a class for the value’s container. If value is a method, go via the im_self attribute to get an instance to use as the container. Then look for a <name>__roles__ attribute on the class. If this is a PermissionRole, call rolesForPermissionOn() as above; if it is a list, tuple or one of the sentinel values (ACCESS_PUBLIC, ACCESS_PRIVATE or ACCESS_NONE, return it directly.

  • If we still have no roles, we may have a primitive or other simple object

    that is not directly security-aware. We can still try to get security information from the container:

    • If there is no container passed in, we have no way of inferring one, so all bets are off. Raise Unauthorized.

    • Attempt to get a __roles__ value from the container. If it is acqusition-wrapped, also try to explicitly acquire __roles__ if it does not have a __roles__ attribute itself.

      If this fails, then we may still be able to get some security assertions from the container (see below), but we only allow this if the accessed parent is the container. If the value was accessed through a more convoluted acquisition chain, say, we cannot rely solely on container assertions, so we raise Unauthorized.

    • At this point, there are two possibilities: we have some roles required to access the container, or we have no roles at all, but we accessed the value directly from its parent container. In both cases, we check container security assertions:

      • If the container is a tuple or string, and we have gotten this far, we consider access to be allowed and return true. (This can’t really happen through URL traversal, but could occur with path traversal).

      • If the container is an object with an attribute __allow_access_to_unprotected_subobjects__, obtain this, which can be of three things:

        • An integer or boolean: if set to a truth value, allow access and return true, otherwise raise Unauthorized.

        • A dictionary: Attempt to look up a truth value in this dictionary by using the accessed name as a key. If not found or false, raise Unauthorized, otherwise allow access and return true. If the name is not found, default to allowing access.

        • A callable: Call it with the name and value as arguments, and use the return value to determine whether to allow access or raise Unauthorized.

      • If there is no __allow_access_to_unprotected_subobjects__, raise Unauthorized.

    • If we did manage to get some roles from the container, we still check __allow_access_to_unprotected_subobjects__ as above, but only as a negative: we raise Unauthorized if access is not allowed, and continue security checking against the roles we found otherwise. In this case, we use the container (probably a content object) as the value to check.

    • At this point, we have roles, and we know the container in theory allows access to the attribute that did not have its own security assertions. We set value to be the container so that we can check whether we are in fact allowed to access the container.

    • We can now check whether the user has the appropriate roles. This is essentially the same logic as in checkPermission() above, although stated slightly differently.

      • If __roles__ is None (ACCESS_PUBLIC) or contains Anonymous, allow access immediately.

      • If the execution context is something like a through-the-web Python script owned by a user, we raise Unauthorized if the owner does not have any of the required roles.

      • If the execution context has proxy roles, these are allowed to be used to validate access intead of the user’s actual roles.

      • Otherwise, call user.allowed() to validate access and either return true or raise Unauthorized.

The remainder of the logic in validate() concerns the case where verbose-security is enabled in zope.conf. Various checks are made in an attempt to raise Unauthorized exceptions with meaningful descriptions about where in the validation logic access was denied.

Changing permissions

The mapping of permissions to roles can be managed persistently at any object by setting the mangled permission attribute (see the description of rolesForPermissionOn() above) to a list of roles as an instance variable.

The most basic API to do so is the class AccessControl.Permission.Permission. This is a transient helper class initialised with a (non-mangled) permission name (i.e. the first element in an __ac_permissions__ tuple), a tuple of attributes the permission applies to (i.e. the second element in an __ac_permissions__ item) - referred to as the variable data - and an object where the permission is being managed.

The methods getRoles(), setRoles() and setRole() on the Permission class allow roles to be obtained and changed.

getRoles() will first attempt to get the mangled permission name attribute and return its value.

If it is not set, it will fall back to looping over all the listed attributes (data) and obtaining the roles from the first one found, taking into account the various ways in which __roles__ can be stored. Note that an empty string in the tuple of attributes means “check the object itself for a __roles__ attribute”. If __roles__ is a list, it is returned, though if it contains the legacy role Shared, this is removed first. The sentinel None (ACCESS_PUBLIC) is turned into ['Manager', 'Anonymous']. If no roles are set, the default return value is ['Manager'], though another default can be supplied as the optional last parameter to getRoles().

setRoles() will set or delete (if setting to an empty list of roles) the mangled permission name as an instance variable on the object. Next, it will ensure no other __roles__ or <name>__roles__ instance variables have been set (class variables are left alone, of course), so that the managled permission name attribute is the unambiguous statement of the permission-to- role mapping.

Note that for both getRoles() and setRoles(), the difference between a tuple (don’t acquire roles) and a list (do acquire) is significant, and preserved.

setRole() is used to manage a single role. It takes a role name and a boolean to decide whether the role should be set or not. It simply builds the appropriate list or tuple based on the current value of getRoles() and then calls setRoles().

In most cases, it is easier to use the API provided by AccessControl.rolemanager.RoleManager to manipulate roles in a particular context, rather than using Permission directly. This class, usually via the more specific OFS.roles.RoleManager, is a mixin to most persistent objects in Zope. It contains a number of relevant methods:

ac_inherited_permissions(all=0)

Returns a list of permissions applicable to this class, but not defined on this class directly, by walking the __bases__ of the class. (Note that this not inheritance in the persitent acquisition sense!). If all is set to a truth value, the permissions on this class are included as well. The return value is an __ac_permissions__-like tuple of tuples. For inherited permissions, the attribute list of each permission entry will be an empty tuple.

permission_settings(permission=None)

Returns the settings for a single or all permissions, returning a list of dicts. Used mainly by ZMI screens.

manage_role(role, permissions=[])

Uses the Permission API to grant the role to the permissions passed in, and take it away from any other permissions where the role may be set.

manage_acquiredPermissions(permissions=[])

Uses the Permission API to set the roles lists for each of the passed-in permissions to a list (acquire), and for all other permissions to a tuple (don’t acquire).

manage_permission(permission, roles=[], acquire=0)

Uses the Permission API to set roles for the given permission to either a tuple or list (it does not matter what type of sequence the roles parameter contains, the acquire parameter is used), but only if the permission is known to this object.

permissionsOfRole(role)

Uses the Permission API to get the permissions of the given role. Returns a list of dicts with keys name and selected (set to either an empty string or the string SELECTED).

rolesOfPermission(permission)

The inverse of permissionsOfRole(), returning a similar data structure.

acquiredRolesAreUsedBy(permission)

Returns either CHECKED or an empty string, depending on whether the roles sequence of the given permission is a list or tuple.

The use of the strings CHECKED or SELECTED as booleans is an unfortunate side-effect of these methods being used quite literally by ZMI templates.

Global and local roles

The list of known (valid) roles in any context is set in the attribute __ac_roles__. On the initialisation of the application root during startup, in install_required_roles() in OFS.Application.AppInitializer, this is made to include at least Owner and Authenticated. The RoleManager base class set it as a class variable to contain ('Manager', 'Owner', 'Anonymous', 'Authenticated').

In AccessControl.rolemanager.RoleManager, the method valid_roles() can be used to obtain the list of valid roles in any given context. It will also include roles from any parent objects referenced via a __parent__ attribute.

User defined roles can be set through the ZMI or the method _addRole() in the OFS.roles.RoleManager specialisation, which simply manipulates the __ac_roles__ tuple as an instance variable. There is also _delRoles() to delete roles. The method userdefined_roles() on the base AccessControl.rolemanager.RoleManager class will return a list of all roles set as instance variables instead of class variables.

The global roles of a given user is determined by the getRoles() function on the user object (see the description of the allowed() method above). The default ZODBRoleManager plugin for PAS stores a mapping of users and roles persistently in the ZODB, though other implementations are possible, e.g. querying an LDAP repository.

Users may also have local roles, granted in a particular container and its children. These can be discovered for a given user most easily by calling the getRolesInContext() function on a user object, which takes a context object as a parameter.

Local roles are stored in the instance variable __ac_local_roles__. This may be a dict or a callable that returns a dict, containing a mapping of user (or group) ids to local roles granted. The local role check is performed iteratively by walking up the acquisition chain and checking the instances of bound methods, until the root of the acquisition chain is reached.

The API to manage local role assignments in a given context is found in AccessControl.rolemanager.RoleManager, through the following methods:

get_local_roles()

Return a tuple of local roles, each represented as a tuple of user ids and a tuple of local roles for that user id. With PAS, this may also include group ids.

users_with_local_role(role)

Inspect __ac_local_roles__ to get a list of all users with the given local role.

get_local_roles_for_userid(userid)

Inspect __ac_local_roles__ to get a tuple of all local roles for the given user id.

manage_addLocalRoles(userid, roles)

Modify __ac_local_roles__ to add the given roles to the given user id. Any existing roles are kept.

manage_setLocalRoles(userid, roles)

Modify __ac_local_roles__ to add the given roles to the given user id. Any existing roles are replaced.

manage_delLocalRoles(userids)

Remove all local roles for the given user ids.

Emergency users

On startup, at import time of AccessControl.users, the function readUserAccessFile() is called to look for a file called accesss in the Zope INSTANCE_HOME (an environment variable) directory. If found, it reads the first line and parses it to return a tuple (name, password, domains, remote_user_mode,).

If set, the module variable emergency_user is set to an UnrestrictedUser, a special type of user where the allowed() method always returns true. If not, it is set to a NullUnrestrictedUser, which acts in reverse and disallows everything.

The user folder implementations in AccessControl and PAS make specific checks for this user during authentication and permission validation to ensure this user can always log in and has virtually any permission, with the exception of _what_not_even_god_should_do (ACCESS_NONE).