Re: [PATCH V32 01/27] Add the ability to lock down access to the running kernel image

From: Andy Lutomirski
Date: Thu May 02 2019 - 20:35:22 EST



> On May 2, 2019, at 4:19 PM, James Morris <jmorris@xxxxxxxxx> wrote:
>
>> On Thu, 2 May 2019, Matthew Garrett wrote:
>>
>>> On Thu, May 2, 2019 at 2:07 PM James Morris <jmorris@xxxxxxxxx> wrote:
>>> One possible direction is to (as previously mentioned) assign IDs to each
>>> callsite and be able to check this ID against a simple policy array
>>> (allow/deny). The default policy choices could be reduced to 'all' or
>>> 'none' during kconfig, and allow a custom policy to be loaded later if
>>> desired.
>>
>> Ok. My primary concern around this is that it's very difficult to use
>> correctly in anything other than the "all" or "none" modes. If a new
>> kernel feature is added with integrated lockdown support, if an admin
>> is simply setting the flags of things they wish to block then this
>> will be left enabled - and may violate the admin's expectations around
>> integrity. On the other hand, if an admin is simply setting the flags
>> of things they wish to permit, then adding lockdown support to an
>> existing kernel feature may result in that feature suddenly being
>> disabled, which may also violate the admin's expectations around the
>> flags providing a stable set of behaviour.
>
> Understood. Most uses will likely be either a distro or an embedded
> system, who I'm assuming would provide a useful policy by default, and
> perhaps a high-level abstraction for modification.
>
>> Given that, would you prefer such a policy expression to look like?
>
> Perhaps a write-once policy, injected from userspace during early boot?
>
> The policy could be simply a list of:
>
> lockdown_feature true|false
>

Iâm not convinced this is worthwhile. As I see it, there really are only two privileges here: root can read kernel memory, and root can corrupt kernel state. A policy that root canât corrupt kernel memory except using, say, eBPF is useless â it gives warm fuzzy feelings but nothing else.