the Chromium logo

The Chromium Projects

SELinux in ChromeOS

SELinux is a kernel security module that provides ability to write accessing policies to archive mandatory access control.

The SELinux talk (internal only) describes how SELinux is used in ChromeOS.

In this documentation, it will briefly introduce

Terms in This Documentation

This documentation contains many SELinux terms. They will be explained in this section.

For more details:

SELinux in ChromeOS boot process

SELinux has no presence until init (upstart) loads the policy, for example, in the bootloader. This section will not discuss the stage before initially loading the policy.

Please note: SELinux is only enabled on ARC-enabled boards, or amd64-generic based boards.

  1. init loads the selinux policy based on configs in /etc/selinux/config, and mounts the selinuxfs to /sys/fs/selinux. init will be assigned with initial contexts (kernel).

  2. init re-execs itself, to trigger type_transition type_transition kernel chromeos_init_exec:process cros_init, which will auto-transits init into cros_init domain.

  3. init starting up service. init will startup services, and kernel will auto-transits based on defined type transition rules. See next section for details.

init starting up services

init will start services based on the config files in /etc/init/ and their dependencies. The "service" here not only includes the daemon process service, but also some pre-startup, or short-lived script.

Simple startup script embedded in init config file

Simple service startups are simply written in <service-name>.conf like

exec /sbin/minijail0 -l --uts -i -v -e -t -P /mnt/empty -T static \
    -b / -b /dev,,1 -b /proc \
    -k tmpfs,/run,tmpfs,0xe -b /run/systemd/journal,,1 \
    -k tmpfs,/var,tmpfs,0xe -b /var/log,,1 -b /var/lib/timezone \
    /usr/sbin/rsyslogd -n -f /etc/rsyslog.chromeos -i /tmp/rsyslogd.pid

in syslog.conf, or

script
  ARGS=""
  case ${WPA_DEBUG} in
    excessive) ARGS='-ddd';;
    msgdump)   ARGS='-dd';;
    debug)     ARGS='-d';;
    info)      ARGS='';;
    warning)   ARGS='-q';;
    error)     ARGS='-qq';;
  esac
  exec minijail0 -u wpa -g wpa -c 3000 -n -i -- \
    /usr/sbin/wpa_supplicant -u -s ${ARGS} -O/run/wpa_supplicant
end script

in wpa_supplicant.conf.

The earlier one, without a small script, kernel SELinux subsystem will auto-transits the domain from u:r:cros_init:s0 to u:r:minijail:s0 upon executing /sbin/minijail0, and then auto-transits to u:r:cros_rsyslogd:s0 upon executing /usr/sbin/rsyslogd so the AVC for domain u:r:cros_rsyslogd:s0 can be applied to the process. From that point in the process, any file access, port usage, network usage, capability request, module load, etc. will be checked against AVC rules with scontext (source context) being u:r:cros_rsyslogd:s0.

While the latter one, with a small script between script and end script, init forks a subprocess (still under cros_init domain) to execute shell. Upon executing shell, this will transit the subprocess to u:r:cros_init_scripts:s0 domain, some simple AVCs could be added to cover all the simple embedded scripts together. Complex scripts or scripts needing permissions more than file; directory read, write, or creation; or exec, should be avoided in the simple script, and should use a separate script. Within the script, it will auto-transits to other domains upon executing the service program, directly (for example, exec /usr/sbin/rsyslogd) or indirectly (via minijail0 the same as above).

Separate script to start the service

But there are some more complex service startup scripts, which are written in a separate (shell) script. The init config file will look like exec /path/to/script.sh or exec /bin/sh /path/to/script.sh

The earlier one is always preferred since it tells the kernel the exact script being executed, so automatic domain transition can be feasible upon executing the script, to not mix up permission requirements of the single script to the whole init script.

The latter one should be avoided if the script has complex permission requirements, like special capabilities, create device, modify sysfs, mount filesystem, or load kernel modules. Using the latter approach will make the kernel not able to distinguish which script is being executed, since the exec syscall are always the same to exec /bin/sh, even the same with script ... end script simple embedded script.

If using the earlier one in the init config file, it will auto transits to a custom defined domain, let's say u:r:cros_service_a_startup:s0, AVC rules will be defined for this special domain, and finally a type_transition rule will transit it from u:r:cros_service_a_startup:s0 to the domain owning the service itself, like u:r:cros_service_a:s0. In the new domain, it will confine the rules for the service.

Pre-start, pre-shutdown, and post-stop

pre-start, pre-shutdown, or post-stop scripts are usually simple embedded scripts like

pre-start script
  mkdir -p -m 0750
  /run/wpa_supplicant
  chown wpa:wpa /run/wpa_supplicant
end script

in wpa_supplicant.conf.

This can be confined together with u:r:cros_init_scripts:s0 since

Like startup script, there are still a very small number of services using an external script file for the startup. For example pre-start exec /usr/share/cros/init/shill-pre-start.sh in shill.conf. This can be either separate domains if they involve complex permissions like mounting/unmounting filesystems, loading/unloading kernel modules, or special capabilities.

How to write SELinux policy for ChromeOS

SELinux policy contains definitions of classes, access vectors (list of permissions), security contexts (types, users, roles, and ranges), access vector rules, and transition rules.

For most developers the necessary classes, access vectors, users, roles, and ranges have already been defined. The most common workflow will be defining new rules only.

Defining Types

Types can be defined in the following syntax

type <type_name>[, <attribute1>, <attribute2>, ...];

This defines a type named <type_name>, and optionally adds the attributes: <attribute1>, <attribute2>, ...

Also, for an already defined type <type_name>, it can add an additional attribute <attributeN> by using the following syntax:

typeattribute <type_name>, <attributeN>;

While attributes can be defined in

attribute <attribute>;

File Contexts

System Image

File contexts for files in system image are defined in platform2/sepolicy/file_contexts/chromeos_file_contexts.

Each line defines a path and its security context. For example,

/sbin/init u:object_r:chromeos_init_exec:s0

It defines the security context (label) for file /sbin/init to be u:object_r:chromeos_init_exec:s0. The security context here must be a complete security context containing user, role, type, and range. Type-only will not work. ChromeOS files always use u as user, object_r as role, and s0 as range for files in system image.

The path can also be a regular expression. For example

/usr/share/zoneinfo(/.*)? u:object_r:cros_tz_data_file:s0

Defined file labels are labelled during cros build-image phase. A simple emerge invocation won't label files in the build directory. cros deploy doesn't label it either.

Runtime Files

Runtime files consist of persistent runtime files in stateful partition (for example, /var/lib, /var/log), and volatile runtime files in tmpfs (for example, /run).

At creation

Both runtime files need to be created at the correct security context. Relabeling is prohibited in general, except for policy upgrades.

Creation label can be handled by either setfscreatecon(3) from libselinux, or type transition rules.

Type transition is recommended since it doesn't require to modify the program to be SELinux-aware. Developers should try their best to make sure files that need type transition on creation, are created at a unique path to reduce the usage of file name in the type transition rules. For example, only the daemon process creates the corresponding directory like /var/lib/<service>, /var/log/<service>, etc, not the startup script, nor some random tests.

filetrans_pattern(<domain>, <contextA>, <contextB>, file|dir|...);
filetrans_pattern(<domain>, <contextA>, <contextB>, file|dir|..., <file name>);

On the above macros, when <domain> creates a file|dir|... under <contextA>, the created file|dir|... will be labelled as <contextB>. If <file name> is provided, only the created file|dir|... with the exact name will be labelled as <contextB>.

For example,

filetrans_pattern(cros_rsyslogd, cros_var_log, cros_syslog, file, "messages");

will label "messages" as cros_syslog, because it was created by a process in cros_rsyslogd domain, under directory with cros_var_log type. Thus, the created structure will look like

/var/log => u:object_r:cros_var_log:s0
/var/log/messages => u:object_r:cros_syslog:s0
Persistent label for upgrades or recoveries

For files under persistent path, e.g. /var/lib, or /var/log, the fullpath based file label MUST also be defined in chromeos_file_contexts together with system image.

This is to make sure when policy upgrades, the new label can be restored upon startup script without the need to recreate the files.

Access Vector Rules

Type Transition

Transition rules control auto type transition upon creation of an object (file, dir, sock_file, etc.), or executing an executable.

type_transition shares the same syntax for both file type and process type,

type_transition `source_type` `target_type` : `class` new_type [object name];

File Type Transition

For file type transition, when processes running in source_type create a class (file, dir, etc) under target_type, the created object is labeled as new_type by default.

The same example as above would be

type_transition cros_rsyslogd cros_var_log:file cros_syslog "messages";

filetrans_pattern macro wraps the type_transition rule and necessary AV rule to one single macro to let creating object as a different type more easily.

Process Type Transition

For process transition, when a process running in source_type executes an executable labelled as target_type, the process automatically transits to new_type.

The example is:

type_transition minijail cros_anomaly_detector_exec:process cros_anomaly_detector;

When a process under minijail executes a file labelled as cros_anomaly_detector_exec, the after-exec process will be running under cros_anomaly_detector domain.

There's also a useful macro like filetrans_pattern for process type transition that wraps not only type_transition rule but also corresponding AV rules, called domain_auto_trans. The detail will be explained the later sections.

Besides type_transition rule, there are also some other type rules that are less frequently used, it can be referred from SELinux Project Wiki

Useful Macros in ChromeOS

Naming Conventions

ChromeOS policies will be combined with Android policies before compiling into a final monolithic policy. Therefore care should be taken to ensure names don't conflict with Android policies.

We use the following naming conventions to reduce the possibilities of conflicts.

  1. Labels created and used before June 2018, including files and directories in the stateful partition, are most likely being used in multiple Android branches, and can be difficult to remove. To reduce the chance of breakage, these files and directories remain on original label even if it doesn't fit into the naming conventions.

  2. minijail domain is u:r:minijail:s0 or u:r:<something>_minijail:s0. Most minijails should fall in the earlier one since that is the label used for minijail processes started from init or init scripts.

  3. All ChromeOS files or processes should have its type prefixed with cros_ unless it's described in previous rules.

  4. Individual executables that desire automatic domain transition on execution must have their type suffixed with _exec. Usually these executables are labelled as u:object_r:cros_<something>_exec:s0

  5. Regarding runtime files

  6. /var/a/b/c/... except /var/run/* (as /var/run is a bind-mount of /run), should be labelled in type cros_var_a_b_c. For example, /var/lib/chaps, should be labelled as u:object_r:cros_var_lib_chaps:s0.

  7. /run/a/b/c/... should be labelled as cros_run_a_b_c.

  8. The rule for choosing c at which level is not enforced. It's usually chosen by a level that you want to isolate access. c should be at least the same depth as /run/<service-name> or /var/{lib, spool}/<service-name>, but can be deeper if a special isolation of some files is necessary.

  9. A simple pid or temporary state file in random tmpfs, can be labelled with type cros_<something>_tmpfile or cros_<something>_pid_file, and the type must have an attribute cros_tmpfile_type.

  10. Regarding domains

In general, each service should have its own domain, named in format of u:r:cros_<service-name>:s0.

  1. Regarding attributes

  2. The prefix rule still applies to attributes.

  3. Attributes for files must suffix with _type, for example, cros_tmpfile_type, and cros_labeled_dev_type.

  4. Attributes for domains must suffix with _domain or domain, for example, cros_miscdomain, cros_bootstat_domain. Suffixing with _domain is preferred over domain.

  5. There's one special attribute for domain, named chromeos_domain. All domains outside ARC container should have this attribute.

Usage with minijail

Minijail is a tool combined with a library and a wrapper program to apply different kind of restrictions (cgroups, caps, seccomps, etc), and namespaces(mountns, pidns, IPCns, etc) in a correct way.

The major problem that the minijail wrapper program faces on SELinux is:

By default, minijail uses a preload library to hook `__libc_start_main` to
apply all kinds of restrictions.
This puts all the privileges that minijail needs into post-exec, where the
kernel can not distinguish the boundary between minijail and the main
program without an explicit setcon(3).
While setcon(3) is unlikely to be possible for our minijail usage since it
usually attempts to mount procfs readonly.

Of course, granting many unnecessary privileges to the domain is discouraged since an exploited process will be allowed to do what minijail could do (mknod, mount filesystems, etc). And we want to reduce the attack surface if possible.

/sbin/minijail0

Minijail wrapper has a static mode that does all the enforcement and lockdowns pre-exec. Due to lack of ambient caps in the past and seccomp issues, minijail developed a preload mode that runs as default for shared-linked ELFs that use a preload library to do the lockdowns post-exec.

But static mode introduces another issue: seccomp.

A common workaround for the two problems above is to install seccomp filter post-execve, just like what a preload library has been doing. A way to solve this could be to have embedded minijail, for quickly installing seccomp rules post-exec since installing seccomp rules doesn't need special capabilities on SELinux.

Example with minijail0 wrapper

For programs using minijail wrapper /sbin/minijail0, the following practice is recommended to isolate pre-minijail and post-minijail as much as possible.

  1. Use -T static static mode. This puts all the lockdown steps in the minijail0 process.

  2. Use --ambient if you have -c. Capabilities are not preserved across execve without ambient caps. You should use --ambient whenever possible, otherwise your caps won't be preserved to the actual process with -T static.

  3. Put seccomp into an inner preload minijail. It will look like this to launch your program

    minijail0 -T static --ambient -c 0x999 --somethingelse -- \
    /sbin/minijail0 -n -S rules.seccomp -- /path/to/your/program
    

SELinux domain remains on minijail when applying all other restrictions. Upon first execve to execute inner minijail0, there's no domain transition. And the inner minijail simply executes the target program with preload library, when the domain transition happens before setting nnp bit.

Libminijail

TODO

Writing SELinux policy for a daemon

In this section, we'll take an example of steps to confine and enforce tcsd.

  1. Define executables

    tcsd has one executable at /usr/sbin/tcsd.

    1. We define the type in sepolicy/policy/chromeos/file.te like this

      type cros_tcsd_exec, file_type, exec_type, cros_file_type, cros_system_file_type;

    2. We define the context for /usr/sbin/tcsd in sepolicy/file_contexts/chromeos_file_contexts like this

      /usr/sbin/tcsd u:object_r:cros_tcsd:exec:s0

  2. Define domains

    Define domains and transitions like this

    type cros_tcsd, chromeos_domain, domain;
    permissive cros_tcsd;
    domain_auto_trans(cros_init, cros_tcsd_exec, cros_tcsd);
    

    The 1st line defines a new type cros_tcsd to be a chromeos_domain and a domain. These two attributes are mandatory for ChromeOS SELinux policies.

    The 2nd line put the domain into permissive, so any actions performed by this domain will be audited, but not denied. Initially having the domain in permissive mode will allow you to see all the potentially denied actions instead of stopping at the first one. You'll need USE="selinux_develop" use flag to make ChromeOS kernel to print permissive audit logs.

    The 3rd line defines an automatic domain transition, when any process at domain cros_init, executes a binary labelled as cros_tcsd_exec, it automatically transits to domain cros_tcsd. The macro will create corresponding type_transition and allow rules.

    It's time to verify if your process is running under the correct domain.

    • cros-workon-$BOARD start selinux-policy for the sepolicy change to take effect. Also cros build-image and cros flash is needed to update the selinux contexts of the files.

    • If it's a daemon process, simply ps auxZ | grep tcsd | grep -v grep. It will display the process matching tcsd with its pid, user, command line, and domain, etc.

    • If it's not a daemon process, but shortlived, you can verify it by

      • printing /proc/self/attr/current in the process to know its domain, or
      • you can observe the audit log from either dmesg, /var/log/messages, or journald, to grep scontext=u:r:cros_tcsd:s0. Unless you program is simply doing some math, therefore doesn't violate any existing SELinux rules, you should be able to see some permissive audit logs if you have USE="selinux_develop" use flag enabled.
  3. Update SELinux tests

    New SELinux tests should be written in tast instead of autotest.

    Tests for SELinux file labels and process domains are located at src/platform/tast-tests/local_tests/security

    Usually, you'll need to update selinux_files_system.go and selinux/processes_test_internal.go.

    In the example case, we'll need to add {Path: "/usr/sbin/tcsd", Context: "cros_tcsd_exec", IgnoreErrors: true}, in selinux_files_system.go, and {exe, "/usr/sbin/tcsd", matchRegexp, "cros_tcsd", zeroProcs, ""}, in selinux/processes_test_internal.go

    The files check checks file at /usr/sbin/tcsd has labels u:object_r:cros_tcsd_exec:s0, without recursion, ignore if file doesn't exist.

    The domain check checks a process with /proc/<pid>/exe to be /usr/sbin/tcsd (a.k.a any process running tcsd binary) to be cros_tcsd domain. The domain check can also match by cmdline.

  4. Write actual rules

    After creating a permissive domain for your daemon, now you need to write actual rules before making the daemon domain SELinux enforcing. Follow these steps to add the needed access rules:

    • Build and flash a new image with the USE flag ‘selinux_develop’. You need this flag to log permissive denials.

    • Run all the tests that are related to your daemon on the test device.

    • Check the SELinux violations logged at /var/log/audit/audit.log on your test device and add the required access rules to the policy as explained in How to read the denials in audit logs.

    • You can use 'audit2allow' script instead to define the access rules. Run 'audit2allow’ script on the audit.log file in your chroot:

    audit2allow -xp /build/$BOARD/etc/selinux/arc/policy/policy.30 -i /path_to_audit_log/audit.log
    
    • Add the rules created by the audit2allow to the policy.

    • Use macros where possible and group the rules that apply to the same file context. Macros are defined in base/imported/global_macros, base/imported/te_macros and chromeos/te_macros. Find the appropriate macro that includes all the access vectors reported by audit2allow. For example if audit2allow generates the following allow rule:

    allow <DomainA> <ContextA>:file { getattr open read ioctl lock map }
    

    Use 'r_file_perms' macro instead:

    allow <DomainA> <ContextA>:file r_file_perms
    
    • emerge and deploy the updated policy and run the tests again. You might want to remove /var/log/audit/audit.log file before running the tests to avoid any confusion.

    If you receive the same SELinux violation log despite adding the required access rule this means you need to add an additional rule related to the same violation. Right place to start the investigation is the audit logs at '/var/log/audit/audit.log' and system log messages at '/var/log/messages' together. Identify the step causing the violation. Examples could be the need to define an ioctl or a key search permission requirement or you may need to change how the file context is defined.

    • Once the policy is complete for the specific board you are testing the policy on, you need to make sure policy doesn't fail for other boards as well. There are macros to restrict the scope of the access rules for different cases. 'has_arc' or 'is_arc_vm' are the two examples for these macros. For example, /home/root/hash/android-data/data directory only exists if ARC is enabled on the device. Therefore if you need to define an access rule for the file context 'media_rw_data_file' you need to use ‘has_arc’ macro.

    • If SELinux logs report a violation against an unlabeled or an unconfined context define the context properly. Run the specific test triggering this rule and identify the specific step to find out the unlabeled/unconfined object. Label them properly following the steps defined in File Contexts.

    Testing after adding new rules:

    • emerge and deploy selinux-policy after adding the new rules. Don't forget to run cros-workon-$BOARD start selinux-policy before emerging the selinux-policy. And repeat the tests.

    • Domain is ready to become enforcing once there are no more SELinux violations logged at /var/log/audit/audit.log for your domain (or when audit2allow doesn't generate any rules for your domain anymore).

    • Remove 'permissive cros_tcsd' line to make the domain enforcing.

    • If you've added or updated any file context you need to build a new image. Even if you haven't changed any file context you still need to build a new image without the USE flag 'selinux_develop' since the daemon will be running in an enforcing domain (remember this flag is to log permissive denials). if there is any missing access rule for the daemon that will be logged since the daemon will be running in an enforcing domain.

    • Repeat the tests. All the tests need to pass and no more violations should be logged.

    • If any of the tests fail and SELinux violations are logged for your daemon, that is possibly because some of the violation logs were overridden on audit.log when you were running the tests earlier. In this case you may need to run the tests in smallers test sets. Or you can use a simple script to filter out the logs you are looking for. A sample script could be like this:

    #!/bin/bash
    
    touch selinux_rules
    chmod 666 selinux_rules
    for (( i=1; i<=360; i++ ))
    do
    scp DUT://var/log/audit/audit.log ./audit.log
    audit2allow -p /build/puff/etc/selinux/arc/policy/policy.30 -i ./audit.log | grep 'allow cros_cryptohomed' | tee -a selinux_rules
    echo $i
    sleep 30
    done
    

    This script copies the audit.log file from a puff board, converts the violation logs to allowrules, greps only the rules that are part of cros_cryptohome domain policy and dumps them to a separate file. Repeat the previous steps until all the tests are passing and no more violations are logged on audit.log for your daemon.

    • Before uploading the updated policy, make sure unit tests pass as well. Android neverallow rules apply to ChromeOS since ChromeOS and Android share the same kernel when Android is running in the ARC container. These neverallow rules are not merged to the policy and don't take effect at compile time. They are rather injected into the selinux-policy unit tests and that's why you need to run the unit tests to test against the Android neverallow rules:
    FEATURES=test emerge-DUT-with-ARC-Container selinux-policy
    
    • Note that the neverallow rules are not imported by the VM unittests like amd64-generic or betty-pi-arc boards, so you need to use a physical ARC container device build to run the unit tests for the updated policy.

    After updating the policy, go back to previous step to update SELinux tests for stateful files too.

    If your process relates to after-login behavior, you may also need to update selinux_files_arc.go and selinux_files_non_arc.go. Non-ARC specified process should present in both. If your process relates to files in home directory, simply modify selinux/files_test_internal_home.go instead.

  5. Enforce your domain

    When you're sure your policy fully covers the expected behavior (only behavior used by ChromeOS). You can remove the permissive cros_tcsd from the policy.

Troubleshooting

How to rule out SELinux a possible cause of a problem

Let's assume you write your own program, or modified a program. Suppose it doesn't work along with the system, you're sure you didn't implement anything wrong, and suspect it could be SELinux denying some operations.

There're three approaches to identify an potential SELinux problem.

  1. From M76, ChromeOS uses auditd to receive audit events from the kernel, and write corresponding audit messages to /var/log/audit/audit.log. During the early boot stages before auditd starts, [kauditd] will still write audit messages into syslog. Developers are supposed to examine both /var/log/messages and /var/log/audit/audit.log if they don't know at which stage a denial could occur.

    • pre-auditd: syslog will be logged to /var/log/messages and systemd-journald. You can read the log by reading /var/log/messages or by executing journalctl. Please note the messages file could be rotated to /var/log/messages.{1,2,3,4,...} if the system is running long term.
    • post-auditd: auditd will receive audit events from kernel via audit netlink socket, and write to /var/log/audit/audit.log. Auditd will handle log rotation to rotate the logs to /var/log/audit/audit.log.{1,2,3,4}.

    You should be able to find permissive=0 in above log locations. If you saw some denials with permissive=1, it doesn't mean it's denied. permissive=1 only mean this access it not allowed by policy, but SELinux is still allowing it because the domain is not enforced.

  2. A quick command could test whether your program works by putting the whole system permissive. By executing setenforce 0 as root in developer mode, you can put the whole system permissive. You'll be able to test if your program comes to work.

  3. If you program is a daemon process which fails so early before you can have a console access, you could change the SELinux config file located at /etc/selinux/config to

    SELINUX=permissive
    SELINUXTYPE=arc
    

    Please keep SELINUXTYPE=arc unchanged, and only changing SELINUX= line to permissive. Please don't change it to disabled otherwise your system may fail to boot since init will halt when it fails to load an SELinux policy.

Approach 2 and 3 to put the whole system permissive won't give you any useful information on what's wrong. Audit log won't print even the policy says to deny this because the whole system is permissive. It only gives you a true or false answer, you'll need the audit logs to find out exact what the problem is. We'll talk more at the how to debug section.

How to read the denials in audit logs

An AVC audit message looks like

audit: type=1400 audit(1550558262.594:5434): avc:  denied  { read } for
pid=26768 comm="cat" name="messages" dev="dm-0" ino=40
scontext=u:r:cros_ssh_session:s0 tcontext=u:object_r:cros_syslog:s0 tclass=file
permissive=1

We'll walk through this audit log as an example.

For more details, it can be referred from

SELinux Project Wiki: NB AL

Red Hat Documentation: Raw Audit Messages

Gentoo Wiki: Where to find SELinux permission denial details

CentOS Wiki: Troubleshooting SELinux

How to debug your SELinux policy

Analyzing audit logs

The most important and fundamental way to debug your policy is to read the audit log.

In the previous audit log example, we know it's "cat" process in cros_ssh_session domain was denied to read file "messages" in device dm-0 labelled as cros_syslog.

There're the main things to look at in the audit logs.

By looking at the log, the main thinkabout would be:

selinux_violation files in /var/spool/crash

/var/spool/crash contains crash reports to be uploaded by crash_sender to crash.corp. Usually it stores data like core dumps and metadata when a program has crashed. But some other anomalies (e.g. selinux violation, service death, kernel warnings, etc) also take advantage of the existing crash reporting mechanism. See Crash Reporter to know more about how crash reporting works.

As mentioned, we take advantage of existing crash reporting mechanism for SELinux violation collection for enforced domain. Since any unpermitted access will trigger an audit event, to reduce the chance it fills up 32 reports pool, we sample audit message at a probability of 0.1% before writing to crash pool for acknowledged users on user build. But for developer build, we still write all audit events dumped to syslog to crash pool.

If you see any selinux_violation* in /var/spool/crash, it doesn't mean that something has crashed. It only means an audit event has occurred. SELinux doesn't kill any process violating the policy, it just forces the corresponding syscall to return -EACCESS (permission denied). In most cases, you don't need to care about what's being stored in /var/spool/crash for SELinux violations. If you need, /var/log/messages or /var/log/audit/audit.log should provide you the information you need.

Inspecting the runtime state

Update the policy

After understanding why the denials occurred by reading the log, policy may need updating to fix the problem.

Locate the policy

In general, ChromeOS policy lives in sepolicy directory in chromiumos/platform2 project, which is src/platform2/sepolicy in the repo tree checkout.

A quick grep on the scontext will locate the where it is defined, and most of its policies.

For example, if we want to change cros_ssh_session:

$ grep cros_ssh_session . -R
./policy/chromeos/dev/cros_ssh_session.te:type cros_ssh_session, domain, chromeos_domain;
./policy/chromeos/dev/cros_ssh_session.te:permissive cros_ssh_session;
./policy/chromeos/dev/cros_ssh_session.te:typeattribute cros_ssh_session netdomain;
./policy/chromeos/dev/cros_sshd.te:domain_auto_trans(cros_sshd, sh_exec, cros_ssh_session);
./policy/chromeos/file.te:filetrans_pattern(cros_ssh_session, cros_etc, cros_ld_conf_cache, file, "ld.so.cache~");
./policy/chromeos/log-and-errors/cros_crash.te:-cros_ssh_session

We can see it's defined in cros_ssh_session.te, which means most of our changes should live in that file.

Searching the compiled policy file

sesearch is an excellent tool to search inside a compiled policy. You can use this tool to search what is allowed, what denials are not logged, what grants are logged, and type transitions, etc.

on Debian-based systems (or gLinux), sudo apt install setools will install this tool.

You can search a policy file, say, policy.30, in following examples:

# Search all allow rule with scontext to be cros_ssh_session or attributes
cros_ssh_session attributes to, tcontext to be cros_sshd or attributes cros_sshd
attributes to with class to be process
$ sesearch policy.30 -A -s cros_ssh_session -t cros_sshd -c process
# Search all type transitions with scontext to be exactly minijail
$ sesearch policy.30 -T -s minijail -ds

man sesearch will provide all the options to search allow, auditallow, dontaudit, allowxperm, etc, with filters on scontext, tcontext, class, permissions.

Put domain to permissive

Sometimes, during debugging, you may not to want to put the system permissive. You can put only one domain permissive.

  1. Locate the policy file as above.

  2. Simply add permissive <domain type>, for example, permissive cros_ssh_session will put cros_ssh_session to permissive.

This will only put the given domain to permissive, and everything with the permissive domain (scontext) will not actually be denied.

But please note: some operations may indicate other permissions at runtime. For example, file creation will check { associate } scontext=file_type tcontext=fs_type class=filesystem , so these kinds of denials may occur.

Writing policy fix
  1. Identify whether labeling files is needed. If yes, label the files either in file_contexts or via type_transition.

  2. Fix the program or add dontaudit rule to prevent from spamming logs if it shouldn't be allowed.

  3. Write allow rule or allowxperm rule based on denials seen, and the behavior analysis of the program.

    1. for one-time program-specific permission requests, simply allow[xperm] scontext tcontext:class perms [args for allowxperm]; scontext, tcontext, and perms can be plural in format like { a b c }

    2. for permission requests that may apply to other programs, create an attribute and attribute current domain to it. And write corresponding rules for that attribute.

    3. use m4 macros wisely, we have many useful macros like r_file_perms, rw_file_perms, create_file_perms, filetrans_pattern, and domain_auto_trans.

For more details in writing policy, please refer to previous sections about writing policies.

Useful build flags for debug

  1. USE="selinux_develop": log permissive denials and make sure log is almost not suppressed by printk limit.

  2. USE="selinux_experimental": build with SELinux mode in permissive by default. This is equivalent to manually changing SELINUX=permissive in /etc/selinux/config

  3. USE="selinux_audit_all": remove all the dontaudit rule before compiling to the final monolithic policy. There're some should-be-denied access with dontaudit rules, so denials don't spam the log. But you may want to see them sometimes during development or debugging process.

For Googlers, there's a nice introduction presentation slides how debugging SELinux policies to refer to though it's for Android, at go/sepolicy-debug