Re: [RFC PATCH v4 0/2] Add capabilities file to securityfs

[Casey Schaufler]

On 8/17/2022 9:10 AM, Paul Moore wrote:
> On Wed, Aug 17, 2022 at 11:50 AM Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote:
>> On 8/17/2022 7:52 AM, Paul Moore wrote:
>>> On Wed, Aug 17, 2022 at 7:53 AM Francis Laniel
>>> <flaniel@xxxxxxxxxxxxxxxxxxx> wrote:
>>>> Le mardi 16 août 2022, 23:59:41 CEST Paul Moore a écrit :
>>>>> On Mon, Jul 25, 2022 at 8:42 AM Francis Laniel
>>>>>
>>>>> <flaniel@xxxxxxxxxxxxxxxxxxx> wrote:
>>>>>> Hi.
>>>>>>
>>>>>> First, I hope you are fine and the same for your relatives.
>>>>> Hi Francis :)
>>>>>
>>>>>> A solution to this problem could be to add a way for the userspace to ask
>>>>>> the kernel about the capabilities it offers.
>>>>>> So, in this series, I added a new file to securityfs:
>>>>>> /sys/kernel/security/capabilities.
>>>>>> The goal of this file is to be used by "container world" software to know
>>>>>> kernel capabilities at run time instead of compile time.
>>>>> ...
>>>>>
>>>>>> The kernel already exposes the last capability number under:
>>>>>> /proc/sys/kernel/cap_last_cap
>>>>> I'm not clear on why this patchset is needed, why can't the
>>>>> application simply read from "cap_last_cap" to determine what
>>>>> capabilities the kernel supports?
>>>> When you capabilities with, for example, docker, you will fill capabilities
>>>> like this:
>>>> docker run --rm --cap-add SYS_ADMIN debian:latest echo foo
>>>> As a consequence, the "echo foo" will be run with CAP_SYS_ADMIN set.
>>>>
>>>> Sadly, each time a new capability is added to the kernel, it means "container
>>>> stack" software should add a new string corresponding to the number of the
>>>> capabilities [1].
>>> Thanks for clarifying things, I thought you were more concerned about
>>> detecting what capabilities the running kernel supported, I didn't
>>> realize it was getting a string literal for each supported capability.
>>> Unless there is a significant show of support for this
>> I believe this could be a significant help in encouraging the use of
>> capabilities. An application that has to know the list of capabilities
>> at compile time but is expected to run unmodified for decades isn't
>> going to be satisfied with cap_last_cap. The best it can do with that
>> is abort, not being able to ask an admin what to do in the presence of
>> a capability that wasn't around before because the name isn't known.
> An application isn't going to be able to deduce the semantic value of
> a capability based solely on a string value,

True, but it can ask someone what to do, and in that case a string is
much better than a number:

  thwonkd: Unknown capability 42 - update thwonkd.conf policy section
  thwonkd: Unknown capability butter_toast - update thwonkd.conf policy section

The thwonkd configuration could be updated to use that capability correctly.
Yes, you could look capability 42 up in the system header files, but only
if they're installed and there's no guarantee that the header files match
the running kernel. That said, I can't think of a case where this would be
useful in real life except for systemd and chcap. I can't speak to the
container manager proposed, as I don't see containers being deployed with
finer granularity than "privileged" or "unprivileged".

>  an integer is just as
> meaningful in that regard.  What might be useful is if the application
> simply accepts a set of capabilities from the user and then checks
> those against the maximum supported by the kernel, but once again that
> doesn't require a string value, it just requires the application
> taking a set of integers and passing those into the kernel when a
> capability set is required.  I still don't see how adding the
> capability string names to the kernel is useful here.
>