2 StackingConfiningUsers
Steve Beattie edited this page 2017-11-10 19:10:02 +00:00

WARNING

This document is a work in progress and requires at a minimum the development version of apparmor 3.5

Requirements

  • AppArmor Kernel module - 3.5-beta1 (Ubuntu 16.04)
  • AppArmor Userspace - 2.11-beta (2.10.95, which was a beta release of 2.11 that is sufficient for stacking, shipped in Ubuntu 16.04)
  • pam_apparmor pam plug-in

Related Documentation

Introduction

The addition of stacking to AppArmor opens up the ability to provide user restrictions at the same time as applying application confinement. Previously AppArmor was limited to applying a single profile when confining a user, either the user profile had to confine all applications run by the user or for some applications the user confinement were dropped in favor of the application profile confinement.

Stacking allows both restrictions to be applied simultaneously with few changes to policy.

Stacking provides a few of ways of structuring policy to achieve simultaneous user and applications confinement in policy. Each method has its advantages and disadvantages.

TODO restructure so recommended methods are first, other options are an annotation of extra information.

  1. Using pam_apparmor ======================

The pam_apparmor module provides a way to confine users instead of applications. With minor modifications to policy designed for pam_apparmor stacking can be used to achieve both confinement of the user and application confinement. The recommend method requires modifying the pam_apparmor glue policy and the user role profiles but does not modify the application profiles.

Since adding policy namespaces complicate the set up, they will be discussed in section 3??? after the basics are established.

Limitations

This method

  • exposes the user confinement to the user via any label query; e.g.:

confined_user//&firefox ```

  • for programs that don't have an application profile the user exposed to stacking with the unconfined profile in its label; e.g.:

confined_user//&unconfined ```

  • if ignore rules are not supported requires the user confinement profile to allow any application directed transitions (change_hat, change_profile, ..) that the application may need.

Modify the role/user profile

The role or user profile needs to be modified to just use ix profile transitions, this prevents user application from having any path to escape the user confinement restrictions; e.g.:

 /bin/** Px,

would become

 /bin/** ix,

The modified role profiles for the example would look like

 #
 # This file contains the roles as referenced by pam/mappings
 #
 #include <tunables/global>
 # By default, allow users to read, lock and link to their own files anywhere,
 # but only write to files in their home directory. Only allow limited execution
 # of files.
 profile default_user {
    #include <abstractions/base>
    #include <abstractions/bash>
    #include <abstractions/consoles>
    #include <abstractions/nameservice>
    deny capability sys_ptrace,
    owner /** rkl,
    @{PROC}/** r,
 # old rules that have been modified to ix versions
 #     /bin/**  Pixmr,
 #     /usr/bin/** Pixmr,
     /bin/** ixmr,
     /usr/bin/** ixmr,
 
    owner @{HOMEDIRS}/ w,
    owner @{HOMEDIRS}/** w,
 }
 # Allow confined_users to read, write, lock and link to their own files
 # anywhere, and execute from some places.
 profile confined_user {
    #include <abstractions/base>
    #include <abstractions/bash>
    #include <abstractions/consoles>
    #include <abstractions/nameservice>
    deny capability sys_ptrace,
    owner /** rwkl,
    @{PROC}/** r,
 # old rules that have been modified to ix versions
 #   /bin/**  Pixmr,
 #   /usr/bin/** Pixmr,
    /bin/**  ixmr,
    /usr/bin/** ixmr,
    owner @{HOMEDIRS}/bin/** ixmr,
 }

Modifying /etc/apparmor.d/pam/mappings

 # gray is a confined user. The hat contains only the permissions necessary
 # to transition to gray's login shell. All other permissions have been
 # moved into the confined_user profile.
 ^gray {
   #include <abstractions/authentication>
   #include <abstractions/nameservice>
   capability dac_override,
   capability setgid,
   capability setuid,
   /etc/default/su r,
   /etc/environment r,
   @{HOMEDIRS}/.xauth* w,
 # old transition
 #  /bin/{,b,d,rb}ash Px -> confined_user,
 #  /bin/{c,k,tc}sh Px -> confined_user,
   /** Px -> confined_user//&unconfined,
 }

Now the user will begin a session with the confinement of

 confined_user//&unconfined

when an application is executed, the confined_user profile transitions to itself while the unconfined profile with transition to the application profile if present, and fall back to transitioning to itself if there is no matching profile present.

For example, when executing firefox, which has a profile defined

 confined_user//&unconfined

transitions to

 confined_user//&firefox

but when executing bash, which has no profile defined the confinement remains

 confined_user//&unconfined

Using Policy namespaces to improve user confinement

The methods presented so far each have a few limitations. The limitations for application directed transitions (change_hat, change_profile, ...) and user visibility of stacking can be addressed by leveraging apparmor policy namespaces.

Namespace Ordering

AppArmor policy namespaces provide a couple of useful behaviors

  • In a stacking situation, when a task is confined by profiles from multiple policy namespaces. The confined task can only see the confinement from its current namespace and any children namespaces. This allows hiding of the application confinement or the user confinement for tasks.

  • Application directed transitions are applied only to the current namespace, meaning profiles from another namespace do not have to deal with the transitions in policy.

To use policy namespaces in user confinement stacking the policy is split into two namespaces, a user confinement namespace, and an application policy namespace. There are two options for the ordering of these namespaces, either

  • user confinement namespace is the parent of the application namespace
    • user see application confinement, not user confinement
    • works with application directed transitions
  • application namespace is the parent of the user confinement namespace
    • user only sees the role/user profile, not application confinement
    • does not work with application directed transitions, as they are applied against the user confinement profile instead of the application profile.

As such for most situations the user confinement namespace should be set up as the parent of the application namespace.

Setting up ???

The user profiles can be loaded into the root namespace, and a policy namespace (:application:) is created for the application profiles. The application profiles are then loaded into the policy namespace.

The pam_apparmor transitions are setup as in method one. That is user confinement profiles are modified to use ix for transitions, and the pam apparmor mapping profiles are modified to do the stacking. However the stacking for the application profile includes the policy namespace name.

Example:

 Px /** -> confined_user//&:application://unconfined

This results in applications being confined by the stack, but unlike in Method 1 the user doesn't see the confined_user profile, the user only sees

 unconfined

When the user executes an application with a profile (eg firefox) the user will see the label as

 firefox

but if the application label is inspected by task confined only by the root namespace will see the full confinement stack of

 confined_user//&:application://firefox

or in the case of the user application being unconfined

 confined_user//&:application://unconfined

Applications that leverage application directed transitions will only affect the application profile so the user confinement profile does not need to include hats or change_profile rules that an application may need.

Multiple application namespaces

If different versions of application profiles are desired, multiple application namespaces can be setup, and users can be stacked to the application namespace that makes sense.

Note: communication between tasks in different parallel namespaces is restricted so care must be taken when using them.

Limitations

  • currently it is not possible for tasks in the application namespace to view or manage user confinement.
  • applications in the root/user confinement namespace will not attach to the profiles in the application namespace.
  • pattern matching transitions can not use the pattern match to transition to a profile in the application namespace.

Working around the Limitations

Currently using policy namespaces imposes its own set of limitations. However with a little care they can be worked around.

Managing the root/user confinement namespace

To be able to load/manage user confinement policy some tasks/user need to not be transition to a stack with the application namespace. The admin user can be left in the root/user confinement namespace and manage policy in both the root/user confinement namespace and the application namespace. Admin processes will still be confined by profiles in the root/user confinement namespace, so root/admin processes do not have to even be unconfined.

Confining processes in the root/user confinement namespace

Processes that are not transitions to a stack with the application namespace will not have the application profiles applied against them. If the application profiles should be applied against processes in the root/user confinement namespace then a copy of those profiles can be loaded into the root/user confinement namespace. The set of application profiles loaded into the root/user confinement namespace does not need to be the same set or even the same version of a profile. This allows the root namespace application confinement to differ from the application confinement being applied to users.

Example:

The firefox for root could be different than the firefox profile in :application:

The root version of a profile will not be applied to users confined by the application namespace because the confined_user profile does ix transitions, thus never transitioning that portion of the stack to the application profiles in root.

Note: There is currently no means to clone profiles between different namespaces, so even if the profile is the same in each namespace it needs to be loaded separately.

pattern matching and stacking

Currently there is a limitation that the patten match name can not be applied to a profile in a subnamespace. That is in the transition

 Px /** -> XXX//&:subns://YYY

there is no way to specify a YYY based on what was matched by /**. XXX in the current namespace can be specified to be the value of /** by specifying the transition like

 Px /** -> &:subns://YYY

However this transition form does not work for the way we need to construct the application namespace. This means a specific profile in the namespace must be targeted. This is why the transition initially provided transitions to :application://unconfined. So that the users session starts unconfined in the application namespace (but still confined by user confinement restrictions), and the application profiles are attached as the user executes them.

If a specific profile other than unconfined in the application namespace is required, that can be specified instead. Multiple transition rules can be used, and specific matches can be mixed with a single generic match.

Examples:

 Px /** -> confined_user//&:application://unconfined            # generic transition
 Px /usr/bin/firefox -> confined_user//&:application://firefox  # specific transition

Using the user attachment conditional

NOTE: not yet available

The user attachment conditional provides several possibilities to improve user confinement. The user attachment conditional allows for confining users without pam_apparmor and allows for different user variants of an application profile to exist in the same namespace.

allows multiple version of profile for different users based on the conditional (don't need a separate namespace) use without namespaces

  • still has visibility and change_profile limitations use with namespaces can be used to avoid requiring pam_apparmor

To confine a user based on the user conditional a profile with a generic attachment, and the user conditional set to the user that the profile should apply to. The profile then contanins the stacking rule to allow transition to an application profile while stacking with the current profile.

 profile confined_user /** user=gray {

    Pix /** -> &@{profile},
 }

Note: same limitation as method2

Now when the first application for a given user is executed it will be confined by user profile and subsequent execs are covered by the stack if it applies.

If the very first attachment requires a stack then this can be done through pam_apparmor rules, or by a global profile.

Examination of alternate policy configuration methods

The above discussion covers the recommended way to setup AppArmor policy when using both application and user confinement. The flexibility of AppArmor policy allows for alternate methods for achieving this style of confinement. Below some variations are documented along with the reasons they are not the recommended method of achieving confinement.

using stacked transitions to remove the stacked unconfined profile

The stacked unconfined profile

 confined_user//&unconfined

looks a little weird and it can be desirable to remove it, and this is possible to do using stacking profile transitions. Using stacked profile transitions we can achieve

 confined_user

instead of the recommended

 confined_user//&unconfined

The exec transitions of the user/role profile are modified to have a form of either

 Pix /** -> &@{profile},

or

 Pix /** -> target//&@{profile}

dependent on whether a -> target transition was used in the original rule. The ix portion of the rule is required so that applications that do not have a profile defined will fall back to inheriting the current confinement of confined_user.

The user will begin a session with the confinement of

 confined_user

when an application is executed, the confined_user profile transitions to the matching profile and stacks itself @{profile} on top. Resulting in final transition with the correct stack.

For example, when executing firefox, which has a profile defined

 confined_user

transitions to

 confined_user//&firefox

but when executing bash, which has no profile defined the confinement remains

 confined_user

because the Pix transition falls back to inheriting the current confined_user profile and confined_user stacked with itself is just the confined user_profile.

Unfortunately this method has a major limitation around application profile exec transitions. If the application makes an exec transition to a program with a globally defined profile and the application profile defines a transition for that exec to a different profile, the resultant confinement will be wrong. The reason for this is that the user confinement profile will also make a transitions to another application profile.

Lets look at an example say we have application profiles

 firefox
 evince
 firefox//evince

and the firefox profile specifies a transition to a custom evince profile firefox//evince.

The user starts the session in the profile

 confined_user

and when starting evince will have a confinement of

 confined_user//&evince

and when starting firefox will have a confinement of

 confined_user//&firefox

however when firefox starts evince the confinement becomes

 confined_user//&evince//&firefox//evince

Notice how the confinement now includes both the evince and the firefox//evince profiles. This is because the confined_user profile does the stacking transition

 confined_user -> confined_user//&evince

while the firefox profile does the transition

 firefox -> firefox//evince

the final confinement stack is thus the combination of those two separate transition, confined_user//&evince//&firefox//evince

Fixing the transition problem

It is possible to fix the transition problem described above, however the solution has its own problems.??? It really isn't at least not yet.

The confined_user profile is duplicated into confined_user_profile1 and confined_user_profile2. The difference between the two profiles is in their domain transitions.

confined_user_profile1 uses a transition of

 Px /** -> &confined_user2,

which will transition to the correct stack of confined_user2//&application_profile

 confined_user_profile2

however Pix can not be used as that would result in

 confined_user1//&confined_user2

however currently apparmor does not have a means of specifying that if the targe

Limitations

This method

  • exposes the user confinement to the user via any label query; e.g.

confined_user//&firefox ```

  • if ignore rules are not supported requires the user confinement profile to allow any application directed transitions (change_hat, change_profile, ..) that the application may need.

  • results in odd even bad confinement when the application profile contains none Px domain transitions (ie. Cx, ix, ux, and even px ->).