lkml.org 
[lkml]   [2018]   [Apr]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v5 1/1] security: Add mechanism to safely (un)load LSMs after boot time
On Wed, Apr 11, 2018 at 10:17 AM, Stephen Smalley <sds@tycho.nsa.gov> wrote:
> On 04/10/2018 05:24 PM, Sargun Dhillon wrote:
>> On Sun, Apr 8, 2018 at 10:25 PM, Tetsuo Handa
>> <penguin-kernel@i-love.sakura.ne.jp> wrote:
>>> Sargun Dhillon wrote:
>>>>> Remove SECURITY_HOOK_COUNT and "struct security_hook_list"->owner and
>>>>> the exception in randomize_layout_plugin.c because preventing module
>>>>> unloading won't work as expected.
>>>>>
>>>>
>>>> Rather than completely removing the unloading code, might it make
>>>> sense to add a BUG_ON or WARN_ON, in security_delete_hooks if
>>>> allow_unload_module is false, and owner is not NULL?
>>>
>>> Do we need to check ->owner != NULL? Although it will be true that
>>> SELinux's ->owner == NULL and LKM-based LSM module's ->owner != NULL,
>>> I think we unregister SELinux before setting allow_unload_module to false.
>>> Thus, rejecting delete_security_hooks() if allow_unload_module == false will
>>> be sufficient. SELinux might want to call panic() if delete_security_hooks()
>>> did not unregister due to allow_unload_module == false. Also,
>>> allow_unload_module would be renamed to allow_unregister_module.
>>>
>>> By the way, please don't use BUG_ON() or WARN_ON() because syzbot would hit
>>> and call panic() because syzbot runs tests with panic_on_warn == true.
>>
>> I think my primary question is for the SELinux folks -- what do you
>> think the behaviour should be? If allow_unload_modules /
>> allow_unregister_module is set, do you want to be able to call
>> security_delete_hooks? What do you think the right
>> action should be if it fails?
>
> The one that avoids breakage for existing users ;)
>
> I personally am in favor of killing SELinux support for runtime disable aka
> CONFIG_SECURITY_SELINUX_DISABLE; the only reason it exists is that Red Hat
> originally insisted that bootloader configuration is too painful to modify/update on
> certain platforms and therefore the selinux=0 boot parameter is insufficient
> as a mechanism for disabling SELinux.

I too would like to remove the SELinux runtime disable code, and we
have looked at it briefly but there are a number of
userspace/bootloader upgrade concerns that need to be addressed first
(some of the issues have been captured in the BZ linked below).
Unfortunately it isn't as trivial a chance as it would initially
appear.

* https://bugzilla.redhat.com/show_bug.cgi?id=1430944

> However, we can't break existing users. Userspace should still attempt to proceed
> even if runtime disable fails, just with SELinux left in permissive mode and no
> policy loaded. That generally should work, but does retain the performance overhead
> of the SELinux hook function processing, unlike a real disable.
>
> I don't think we particularly care about allow_unload_modules / allow_unregister_module
> since there is no existing userspace or configurations relying on it.

--
paul moore
www.paul-moore.com

\
 
 \ /
  Last update: 2018-04-11 23:37    [W:0.059 / U:1.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site