[PATCH v15 01/11] LSM: Identify modules by more than name

[Casey Schaufler]

On 10/21/2023 5:20 AM, Tetsuo Handa wrote:
> On 2023/10/21 4:52, Casey Schaufler wrote:
>> On 10/5/2023 5:58 AM, Tetsuo Handa wrote:
>>> On 2023/09/13 5:56, Casey Schaufler wrote:
>>>> Create a struct lsm_id to contain identifying information about Linux
>>>> Security Modules (LSMs). At inception this contains the name of the
>>>> module and an identifier associated with the security module.  Change
>>>> the security_add_hooks() interface to use this structure.  Change the
>>>> individual modules to maintain their own struct lsm_id and pass it to
>>>> security_add_hooks().
>>> I came to worry about what purpose does the LSM ID value (or more precisely,
>>> "struct lsm_id") is used for. If the LSM ID value is used for only switch
>>> {reading,writing} /proc/self/attr/ of specific LSM module's information, only
>>> LSM modules that use /proc/self/attr/ will need the LSM ID value.
>>>
>>> But this series uses "struct lsm_id" as one of arguments for security_add_hooks(),
>>> and might be reused for different purposes.
>>>
>>> Then, BPF-based LSMs (which are not considered as in-tree LSM modules, for
>>> only BPF hook is considered as in-tree LSM module) might receive unfavorable
>>> treatment than non BPF-based LSMs? 
>>>
>>> [PATCH v15 05/11] says
>>>
>>>   Create a system call to report the list of Linux Security Modules
>>>   that are active on the system. The list is provided as an array
>>>   of LSM ID numbers.
>>>   
>>>   The calling application can use this list determine what LSM
>>>   specific actions it might take. That might include choosing an
>>>   output format, determining required privilege or bypassing
>>>   security module specific behavior.
>>>
>>> but, at least, name of BPF-based LSMs won't be shown up in lsm_list_modules()
>>> compared to non BPF-based LSMs? Then, the calling application can't use this
>>> list determine what BPF-based LSM specific actions it might take?
>> That is correct. Just as knowing that your system is using SELinux won't
>> tell you whether a specific action might be permitted because that's driven
>> by the loaded policy, so too knowing that your system is using BPF won't
>> tell you whether a specific action might be permitted because that's driven
>> by the eBPF programs in place.
> If the system call returning LSM ID value for SELinux but does not tell
> the caller of that system call whether a specific action might be permitted,
> what information does LSM ID value tell?

It tells the caller that the LSM is active on the system. That's it.
Just like reading /sys/kernel/security/lsm.

>
> The patch description lacks relationship between LSM ID value and data.
> In other words, why LSM ID values are needed (and are useful for doing what).
> If the only information the caller can know from the LSM ID value were
> what LSMs are enabled (i.e. the content of /sys/kernel/security/lsm ), why
> bother to use LSM ID values? (Yes, integer comparison is faster than string
> comparison. But that is not enough justification for not allowing out-of-tree
> LSMs and eBPF-based access control mechanisms to have stable LSM ID values.)
>
> What does "choosing an output format", "determining required privilege",
> "bypassing security module specific behavior" mean? How can they choose
> meaningful output format, determine appropriate privilege, bypass security
> module specific behavior (if the only information the caller can know from
> the LSM ID value were what LSMs are enabled) ?

If Smack and SELinux not enabled on the system there is no point in
setting up a netlabel configuration, for example.

>> I wish we could stop people from saying "BPF-based LSM". BPF is the LSM. The
>> eBPF programs that implement a "policy" are NOT a LSM. There needs to be a
>> name for that, but LSM  is  not  it.
> My understanding is that "BPF is not an LSM module but infrastructure for using
> LSM hooks".

As BPF is implemented as a LSM I suggest your statement is incorrect.

> Say, an access control implementation consists of two parts; "code" and "data".
> The "code" is written by developers and is determined at compile time and is
> interpreted by CPU, and the "data" is written by administrators and is interpreted
> by "code". The "data" part can be either built-in (determined at compile time) or
> loadable (configurable at run-time).
>
> eBPF-based access control implementations (which can be loaded via bpf() system
> call after boot) consists of "code" and "data". BPF will remain no-op unless
> eBPF-based access control implementations are loaded via bpf() system call.
> Thus, I believe that an eBPF-based access control implementation should be
> considered as an LSM module (like SELinux etc.)

And I say you're wrong. Your arguments are gibberish.

https://www.youtube.com/watch?v=-7cUnID7vFs