Re: neverallow rules

[Stephen Smalley <sds@tycho.nsa.gov>]

On 2/27/19 8:57 AM, Stephen Smalley wrote:
> On 2/27/19 8:54 AM, Stephen Smalley wrote:
>> On 2/27/19 8:48 AM, Stephen Smalley wrote:
>>> On 2/27/19 8:29 AM, Stephen Smalley wrote:
>>>> On 2/26/19 10:38 PM, Chris PeBenito wrote:
>>>>> On Tue, 2019-02-26 at 19:29 -0600, Joe Nall wrote:
>>>>>> Looking at neverallow rules, the semanage.conf file says
>>>>>> "# expand-check check neverallow rules when executing all semanage
>>>>>> commands.
>>>>>>   # Large penalty in time if you turn this on. "
>>>>>>
>>>>>> If I don't set expand-check=1, are the neverallow rules actually
>>>>>> enforced?
>>>>>
>>>>> Nope.
>>>>>
>>>>>> If so, when?
>>>>>>
>>>>>> An semodule -i of a policy module with neverallow rules that are
>>>>>> violated by the existing binary policy succeeds without complaint
>>>>>> unless expand-check=1 in RHEL 7.6. This is not what I expected.
>>>>>>
>>>>>> The time taken by a trivial module installation goes from ~.3 seconds
>>>>>> to ~14 seconds, so the time hit for expand-check is pretty serious.
>>>>>
>>>>> The reason for adding the expand-check option is because the 
>>>>> neverallow
>>>>> checking is so expensive.
>>>>>
>>>>>> We are trying to establish some policy invariants to protect against
>>>>>> unexpected/unnoticed RHEL upstream policy changes, some of which have
>>>>>> bitten us recently. Any suggestions are welcome.
>>>>>
>>>>> One alternative would be to use the setools API to code up some policy
>>>>> searches in Python, then process the results to find things you
>>>>> do/don't want in your policy.
>>>>
>>>> The approach taken in Fedora/RHEL is to run make validate as part of 
>>>> the selinux-policy.spec recipe, which uses the validate target in 
>>>> the refpolicy Makefile, which manually runs semodule_link and 
>>>> semodule_expand to manually link and expand the modules and performs 
>>>> full checking.  Then the cost is only paid at policy build time and 
>>>> is only applied for the neverallow rules and allow rules in the 
>>>> Fedora/RHEL policy, not for third party modules.  If you are 
>>>> rebuilding the RHEL policy from source, you could add your 
>>>> neverallows to it there and the make validate should catch the 
>>>> violations.
>>>>
>>>> In Android, neverallow checking is done both at policy build time 
>>>> and as part of device certification.  The latter is done using a 
>>>> sepolicy-analyze tool we originally contributed; it can take a 
>>>> separate neverallow rule configuration in source form and an already 
>>>> compiled kernel binary policy and check them against each other. 
>>>> That's useful for cases where you don't have policy sources.
>>>>
>>>> You can have libsemanage apply your own checker against policy as 
>>>> part of module transactions by specifying a [verify kernel] stanza 
>>>> in semanage.conf, see:
>>>> http://selinuxproject.org/page/PolicyValidate
>>>
>>> Also, FWIW, you can obtain, build, and use sepolicy-analyze outside 
>>> of the Android build system as follows. If you are on RHEL 7, you 
>>> likely need/want to obtain the upstream selinux userspace and build 
>>> its version of libsepol.a for use below instead of using the 
>>> libsepol-static package from RHEL since that is likely too old.
>>>
>>> # Get source and build.
>>> $ git clone https://android.googlesource.com/platform/system/sepolicy
>>> $ cd sepolicy/tools/sepolicy-analyze
>>> $ gcc -o sepolicy-analyze *.c /path/to/libsepol.a
>>>
>>> # Run with neverallows specified on the command-line.
>>> $ sepolicy-analyze /etc/selinux/targeted/policy/policy.31 neverallow 
>>> -n "neverallow user_t self:process signal;"
>>>
>>> # Run with neverallows from a separate file.
>>> $ cat > neverallows.conf <<EOF
>>> neverallow user_t self:process signal;
>>> EOF
>>> $ sepolicy-analyze /etc/selinux/targeted/policy/policy.31 neverallow 
>>> -f neverallows.conf
>>>
>>> There have been improvements to neverallow checking time upstream 
>>> since RHEL 7, but it will still be a function of policy size and 
>>> obviously RHEL has a much larger policy than Android...
>>
>> For reference, on Fedora 29:
>>
>> $ time sepolicy-analyze /etc/selinux/targeted/policy/policy.31 
>> neverallow -n "neverallow user_t self:capability mac_admin;"
>>
>> real    0m0.145s
>> user    0m0.133s
>> sys    0m0.012s
> 
> Also, you can put multiple neverallow statements (semicolon separated) 
> in the -n option value and have them all checked at once, or similarly 
> multiple ones in the file.

There is an error in the example above; mac_admin is a permission in 
capability2 rather than capability.  By default, sepolicy-analyze 
doesn't warn about undefined classes/permissions/types because a given 
binary policy might not define everything used in a neverallow 
configuration and that is ok (i.e. if there is no matching type, class 
or permission, then it isn't possible for the neverallow to be 
violated). If you want it to detect such things, pass -w to neverallow, ala:

$ sepolicy-analyze /etc/selinux/targeted/policy/policy.31 neverallow -w 
-n "neverallow user_t self:capability mac_admin;"
Warning!  Permission mac_admin used in neverallow undefined in class 
capability in policy being checked.
Warning!  Empty permission set

These are only warnings however; they do not yield an error exit.