Remove the ability to dynamically update SELinux policy on the
device.
1) This functionality has never been used, so we have no idea if
it works or not.
2) If system_server is compromised, this functionality allows a
complete bypass of the SELinux policy on the device. In particular,
an attacker can force a regression of the following patch
* https://android-review.googlesource.com/138510
see also https://code.google.com/p/android/issues/detail?id=181826
3) Dynamic policy update can be used to bypass neverallow protections
enforced in CTS, by pushing a policy to the device after certification.
Such an updated policy could bring the device out of compliance or
deliberately introduce security weaknesses.
(cherrypicked from commit e827a8ab27)
Bug: 22885422
Bug: 8949824
Change-Id: I802cb61fd18a452a2bb71c02fe57cfce5b7e9dc8
Add a neverallow rule (compile time assertion) for /data/local/tmp
access. /data/local/tmp is intended entirely for the shell user, and
it's dangerous for other SELinux domains to access it. See, for example,
this commit from 2012:
f3ef1271f2
Change-Id: I5a7928ae2b51a574fad4e572b09e60e05b121cfe
https://android-review.googlesource.com/166419 changed the handling
of non-interactive adb shells to use a socket instead of a PTY.
When the stdin/stdout/stderr socket is received by /system/bin/sh,
the code runs isatty() (ioctl TCGETS) to determine how to handle the
file descriptor. This is denied by SELinux.
Allow it for all domains.
Addresses the following denial:
avc: denied { ioctl } for pid=4394 comm="sh" path="socket:[87326]" dev="sockfs" ino=87326 ioctlcmd=5401 scontext=u:r:shell:s0 tcontext=u:r:adbd:s0 tclass=unix_stream_socket permissive=0
TODO: When kernels are publicly available which support SELinux ioctl
filtering, limit this just to ioctl 5401 (TCGETS) instead of all ioctls.
Bug: 21215503
Change-Id: I5c9394f27b8f198d96df14eac4b0c46ecb9b0898
In Android 5.1, mediaserver couldn't execute any file on
/system. This slightly regressed due to
8a0c25efb0, which granted mediaserver
access to execute /system/bin/toolbox and /system/bin/toybox
Revoke that unneeded access and add a neverallow rule to prevent
regressions.
TODO: Remove toolbox_exec:file execute permissions from domain.te
and add it back to the specific domains that need it.
Change-Id: Ia7bc6028a9ffb723d4623d91cbe15c8c1bbb2eb9
Init never uses / add service manager services. It doesn't make
sense to allow these rules to init. Adding a rule of this type
is typically caused by a process inappropriately running in init's
SELinux domain, and the warning message:
Warning! Service %s needs a SELinux domain defined; please fix!
is ignored.
In addition, add neverallow rules to domain.te which prevent
nonsense SELinux service_manager rules from being added.
Change-Id: Id04a50d1826fe451a9ed216aa7ab249d0393cc57
Domains have the ability to read normal tmpfs files but not symlinks.
Grant this ability. In particular, allow domains to read /mnt/sdcard.
Addresses the following denial:
type=1400 audit(0.0:19):avc: denied { read } for comm=4173796E635461736B202333 name="sdcard" dev="tmpfs" ino=7475 scontext=u:r:untrusted_app:s0:c512,c768 tcontext=u:object_r:tmpfs:s0 tclass=lnk_file permissive=0
Bug: 20755029
Change-Id: I0268eb00e0eb43feb2d5bca1723b87b7a44f31a9
To help reduce code injection paths, a neverallow is placed
to prevent domain, sans untrusted_app and shell, execute
on data_file_type. A few data_file_type's are also exempt
from this rule as they label files that should be executable.
Additional constraints, on top of the above, are placed on domains
system_server and zygote. They can only execute data_file_type's
of type dalvikcache_data_file.
Change-Id: I15dafbce80ba2c85a03c23128eae4725703d5f02
Signed-off-by: William Roberts <william.c.roberts@intel.com>
Despite removing these from AOSP policy they seem to still be
present in device policies. Prohibit them via neverallow.
We would also like to minimize execmem to only app domains
and others using ART, but that will first require eliminating it
from device-specific service domains (which may only have it
due to prior incorrect handling of text relocations).
Change-Id: Id1f49566779d9877835497d8ec7537abafadadc4
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Allow vold, healthd, slideshow, and watchdogd access to /dev/kmsg.
These processes log to the kernel dmesg ring buffer, so they need
write access to that file.
Addresses the following denials:
avc: denied { write } for pid=134 comm="watchdogd" name="kmsg" dev="tmpfs" ino=9248 scontext=u:r:watchdogd:s0 tcontext=u:object_r:kmsg_device:s0 tclass=chr_file permissive=0
avc: denied { write } for pid=166 comm="healthd" name="kmsg" dev="tmpfs" ino=9248 scontext=u:r:healthd:s0 tcontext=u:object_r:kmsg_device:s0 tclass=chr_file permissive=0
avc: denied { write } for pid=180 comm="vold" name="kmsg" dev="tmpfs" ino=9248 scontext=u:r:vold:s0 tcontext=u:object_r:kmsg_device:s0 tclass=chr_file permissive=0
These denials were triggered by the change in
https://android-review.googlesource.com/151209 . Prior to that change,
any code which called klog_init would (unnecessarily) create the
device node themselves, rather than using the already existing device
node.
Drop special /dev/__null__ handling from watchdogd. As of
https://android-review.googlesource.com/148288 , watchdogd no longer
creates it's own /dev/null device, so it's unnecessary for us
to allow for it.
Drop mknod from healthd, slideshow, and watchdogd. healthd and slideshow
only needed mknod to create /dev/__kmsg__, which is now obsolete.
watchdogd only needed mknod to create /dev/__kmsg__ and /dev/__null__,
which again is now obsolete.
(cherry picked from e2651972c1)
Bug: 21242418
Change-Id: If01c8001084575e7441253f0fa8b4179ae33f534
To reduce the likelihood of malicious symlink attacks, neverallow
read access to shell- and app-writable symlinks.
Change-Id: I0dea1e6e4f0ce34531100696d230294e1b8a5500
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Prohibit all but a specific set of whitelisted domains
from writing to /data/dalvik-cache. This is to prevent
code injection into apps, zygote, or system_server.
Inspired by:
https://www.nowsecure.com/blog/2015/06/16/remote-code-execution-as-system-user-on-samsung-phones/
which depended on system UID apps having write access to
/data/dalvik-cache (not allowed in AOSP policy but evidently
in those device policies). Prevent this from recurring.
Change-Id: I282c7bf998421d794883e432b091ad1dcf9da67e
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Only a few daemons need transition to shell. Prevent
misuse and over-privileging of shell domain.
Change-Id: Ib1a5611e356d7a66c2e008232c565035e3fc4956
Signed-off-by: William Roberts <william.c.roberts@linux.intel.com>
Only a few system level components should be creating and writing
these files, force a type transition for shared files.
Change-Id: Ieb8aa8a36859c9873ac8063bc5999e9468ca7533
Signed-off-by: William Roberts <william.c.roberts@linux.intel.com>
Prevent defining any process types without the domain attribute
so that all allow and neverallow rules written on domain are
applied to all processes.
Prevent defining any app process types without the appdomain
attribute so that all allow and neverallow rules written on
appdomain are applied to all app processes.
Change-Id: I4cb565314fd40e1e82c4360efb671b175a1ee389
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
This reverts commit c450759e8e.
There was nothing wrong with this change originally --- the companion
change in init was broken.
Bug: http://b/19702273
Change-Id: I9d806f6ac251734a61aa90c0741bec7118ea0387
Some devices still have pre-built binaries with text relocations
on them. As a result, it's premature to assert a neverallow rule
for files in /system
Bug: 20013628
Change-Id: I3a1e43db5c610164749dee6882f645a0559c789b
Android has long enforced that code can't compile with text
relocations present. Add a compile time assertion to prevent
regressions.
Change-Id: Iab35267ce640c1fad9dc82b90d22e70e861321b7
/system/xbin/procrank is a setuid program run by adb shell on
userdebug / eng devices. Allow it to work without running adb root.
Bug: 18342188
Change-Id: I18d9f743e5588c26661eaa26e1b7e6980b15caf7
Executing /system/xbin/su is only supported on userdebug builds
for a limited number of domains. On user builds, it should never
occur.
Add a compile time assertion (neverallow rule) that this is
always true.
Bug: 19647373
Change-Id: I231a438948ea2d47c1951207e117e0fb2728c532
Add neverallow rules to ensure that zygote commands are only taken from
system_server.
Also remove the zygote policy class which was removed as an object manager in
commit: ccb3424639821b5ef85264bc5836451590e8ade7
Bug: 19624279
Change-Id: I1c925d7facf19b3953b5deb85d992415344c4c9f
Add rules to allow /sbin/slideshow to access framebuffer and input
devices at early stages of boot, and rules to allow init to execute
the program (from init.rc using exec).
Needed by changes from
I58c79a7f3ac747eec0d73a10f018d3d8ade9df7d
Change-Id: I1d5018feb7025853f0bf81651f497fef8c3a6ab0
Revert the tightening of /proc/net access. These changes
are causing a lot of denials, and I want additional time to
figure out a better solution.
Addresses the following denials (and many more):
avc: denied { read } for comm="SyncAdapterThre" name="stats" dev="proc" ino=X scontext=u:r:untrusted_app:s0:c512,c768 tcontext=u:object_r:proc_net:s0 tclass=file
avc: denied { read } for comm="facebook.katana" name="iface_stat_fmt" dev="proc" ino=X scontext=u:r:untrusted_app:s0:c512,c768 tcontext=u:object_r:proc_net:s0 tclass=file
avc: denied { read } for comm="IntentService[C" name="if_inet6" dev="proc" ino=X scontext=u:r:untrusted_app:s0:c512,c768 tcontext=u:object_r:proc_net:s0 tclass=file
avc: denied { read } for comm="dumpstate" name="iface_stat_all" dev="proc" ino=X scontext=u:r:dumpstate:s0 tcontext=u:object_r:proc_net:s0 tclass=file
This reverts commit 0f0324cc82
and commit 99940d1af5
Bug: 9496886
Bug: 19034637
Change-Id: I436a6e3638ac9ed49afbee214e752fe2b0112868
The recovery partition has been assigned a recovery_block_device
type for the AOSP devices, so install_recovery should not need
rw access to the generic block_device type. Remove it.
Change-Id: I31621a8157998102859a6e9eb76d405caf6d5f0d
Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov>
Add a compile time assertion that no SELinux rule exists which
allows mounting on top of symbolic links, fifo files, or socket
files. Remove the capability from unconfined domains.
Change-Id: I6d7cc95cd17e2e5f165fa5948563800ed206bb71
Android doesn't want to support System V IPC classes.
Ensure that it isn't supported by adding a neverallow rule
(compile time assertion).
Change-Id: I278d45960ee557917584f9137323b4cabfe140a9
The shell domain is already allowed to list and find all service_manager
objects, so extra auditing is pointless.
Bug: 18106000
Change-Id: I8dbf674fa7ea7b05e48e5bbc352b0c9593f2b627
Add an SELinux neverallow rule (compile time assertion) that only
authorized SELinux domains are writing to files in /data/dalvik-cache.
Currently, SELinux policy only allows the following SELinux domains
to perform writes to files in /data/dalvik-cache
* init
* zygote
* installd
* dex2oat
For zygote, installd, and dex2oat, these accesses make sense.
For init, we could further restrict init to just relabelfrom
on /data/dalvik-cache files, and { create, write, setattr }
on /data/dalvik-cache directories. Currently init has full
write access, which can be reduced over time.
This change was motivated by the discussion
in https://android-review.googlesource.com/127582
Remove /data/dalvik-cache access from the unconfined domain.
This domain is only used by init, kernel, and fsck on user builds.
The kernel and fsck domains have no need to access files in
/data/dalvik-cache. Init has a need to relabel files, but
that rule is already granted in init.te.
The neverallow rule is intended to prevent regressions. Neverallow
rules are CTS tested, so regressions won't appear on our devices
or partner devices.
Change-Id: I15e7d17b1121c556463114d1c6c49557a57911cd
external/sepolicy commit 99940d1af5
(https://android-review.googlesource.com/123331) removed /proc/net
access from domain.te.
Around the same time, system/core commit
9a20e67fa62c1e0e0080910deec4be82ebecc922
(https://android-review.googlesource.com/123531) was checked in.
This change added libnl as a dependency of libsysutils.
external/libnl/lib/utils.c has a function called get_psched_settings(),
which is annotated with __attribute__((constructor)). This code
gets executed when the library is loaded, regardless of whether or
not other libnl code is executed.
By adding the libnl dependency, even code which doesn't use the
network (such as vold and logd) ends up accessing /proc/net/psched.
For now, allow this behavior. However, in the future, it would be
better to break this dependency so the additional code isn't loaded
into processes which don't need it.
Addresses the following denials:
avc: denied { read } for pid=148 comm="logd" name="psched" dev="proc" ino=4026536508 scontext=u:r:logd:s0 tcontext=u:object_r:proc_net:s0 tclass=file permissive=0
avc: denied { read } for pid=152 comm="vold" name="psched" dev="proc" ino=4026536508 scontext=u:r:vold:s0 tcontext=u:object_r:proc_net:s0 tclass=file permissive=0
avc: denied { read } for pid=930 comm="wpa_supplicant" name="psched" dev="proc" ino=4026536508 scontext=u:r:wpa:s0 tcontext=u:object_r:proc_net:s0 tclass=file permissive=0
Bug: 19079006
Change-Id: I1b6d2c144534d3f70f0028ef54b470a75bace1cf
SELinux domains wanting read access to /proc/net need to
explicitly declare it.
TODO: fixup the ListeningPortsTest cts test so that it's not
broken.
Bug: 9496886
Change-Id: Ia9f1214348ac4051542daa661d35950eb271b2e4
Temporarily give every system_server_service its own
domain in preparation for splitting it and identifying
special services or classes of services.
Change-Id: I81ffbdbf5eea05e0146fd7fd245f01639b1ae0ef
All domains are currently granted list and find service_manager
permissions, but this is not necessary. Pare the permissions
which did not trigger any of the auditallow reporting.
Bug: 18106000
Change-Id: Ie0ce8de2af8af2cbe4ce388a2dcf4534694c994a
tilapia's OTA code for updating the radio image needs to
create files on rootfs and create a character device in /dev.
Add an exception for recovery the the various neverallow rules
blocking this behavior.
(cherrypick, with modifications, from 0055ea904a)
Bug: 18281224
Change-Id: I5c57afe0a10b4598fea17f9c5c833bd39551907e
Recovery should never be accessing files from /data.
In particular, /data may be encrypted, and the files within
/data will be inaccessible to recovery, because recovery doesn't
know the decryption key.
Enforce write/execute restrictions on recovery. We can't tighten
it up further because domain.te contains some /data read-only
access rules, which shouldn't apply to recovery but do.
Create neverallow_macros, used for storing permission macros
useful for neverallow rules. Standardize recovery.te and
property_data_file on the new macros.
Change-Id: I02346ab924fe2fdb2edc7659cb68c4f8dffa1e88
