9b2e0cbeea
In kernel 4.7, the capability and capability2 classes were split apart
from cap_userns and cap2_userns (see kernel commit
8e4ff6f228e4722cac74db716e308d1da33d744f). Since then, Android cannot be
run in a container with SELinux in enforcing mode.
This change applies the existing capability rules to user namespaces as
well as the root namespace so that Android running in a container
behaves the same on pre- and post-4.7 kernels.
This is essentially:
1. New global_capability_class_set and global_capability2_class_set
that match capability+cap_userns and capability2+cap2_userns,
respectively.
2. s/self:capability/self:global_capability_class_set/g
3. s/self:capability2/self:global_capability2_class_set/g
4. Add cap_userns and cap2_userns to the existing capability_class_set
so that it covers all capabilities. This set was used by several
neverallow and dontaudit rules, and I confirmed that the new
classes are still appropriate.
Test: diff new policy against old and confirm that all new rules add
only cap_userns or cap2_userns;
Boot ARC++ on a device with the 4.12 kernel.
Bug: crbug.com/754831
Change-Id: I4007eb3a2ecd01b062c4c78d9afee71c530df95f
52 lines
1.9 KiB
Text
52 lines
1.9 KiB
Text
# only HALs responsible for network hardware should have privileged
|
|
# network capabilities
|
|
neverallow {
|
|
halserverdomain
|
|
-hal_bluetooth_server
|
|
-hal_wifi_server
|
|
-hal_wifi_supplicant_server
|
|
-rild
|
|
} self:global_capability_class_set { net_admin net_raw };
|
|
|
|
# Unless a HAL's job is to communicate over the network, or control network
|
|
# hardware, it should not be using network sockets.
|
|
neverallow {
|
|
halserverdomain
|
|
-hal_tetheroffload_server
|
|
-hal_wifi_server
|
|
-hal_wifi_supplicant_server
|
|
-rild
|
|
} domain:{ tcp_socket udp_socket rawip_socket } *;
|
|
|
|
###
|
|
# HALs are defined as an attribute and so a given domain could hypothetically
|
|
# have multiple HALs in it (or even all of them) with the subsequent policy of
|
|
# the domain comprised of the union of all the HALs.
|
|
#
|
|
# This is a problem because
|
|
# 1) Security sensitive components should only be accessed by specific HALs.
|
|
# 2) hwbinder_call and the restrictions it provides cannot be reasoned about in
|
|
# the platform.
|
|
# 3) The platform cannot reason about defense in depth if there are
|
|
# monolithic domains etc.
|
|
#
|
|
# As an example, hal_keymaster and hal_gatekeeper can access the TEE and while
|
|
# its OK for them to share a process its not OK with them to share processes
|
|
# with other hals.
|
|
#
|
|
# The following neverallow rules, in conjuntion with CTS tests, assert that
|
|
# these security principles are adhered to.
|
|
#
|
|
# Do not allow a hal to exec another process without a domain transition.
|
|
# TODO remove exemptions.
|
|
neverallow {
|
|
halserverdomain
|
|
-hal_dumpstate_server
|
|
-rild
|
|
} { file_type fs_type }:file execute_no_trans;
|
|
# Do not allow a process other than init to transition into a HAL domain.
|
|
neverallow { domain -init } halserverdomain:process transition;
|
|
# Only allow transitioning to a domain by running its executable. Do not
|
|
# allow transitioning into a HAL domain by use of seclabel in an
|
|
# init.*.rc script.
|
|
neverallow * halserverdomain:process dyntransition;
|