[PATCH bpf-next v1 00/13] MAC and Audit policy using eBPF (KRSI)

[Stephen Smalley]

On 1/14/20 9:48 PM, Alexei Starovoitov wrote:
> On Tue, Jan 14, 2020 at 12:42:22PM -0500, Stephen Smalley wrote:
>> On 1/14/20 11:54 AM, Stephen Smalley wrote:
>>> On 1/10/20 12:53 PM, Alexei Starovoitov wrote:
>>>> On Fri, Jan 10, 2020 at 04:27:58PM +0100, KP Singh wrote:
>>>>> On 09-Jan 14:47, Stephen Smalley wrote:
>>>>>> On 1/9/20 2:43 PM, KP Singh wrote:
>>>>>>> On 10-Jan 06:07, James Morris wrote:
>>>>>>>> On Thu, 9 Jan 2020, Stephen Smalley wrote:
>>>>>>>>
>>>>>>>>> On 1/9/20 1:11 PM, James Morris wrote:
>>>>>>>>>> On Wed, 8 Jan 2020, Stephen Smalley wrote:
>>>>>>>>>>
>>>>>>>>>>> The cover letter subject line and the
>>>>>>>>>>> Kconfig help text refer to it as a
>>>>>>>>>>> BPF-based "MAC and Audit policy".  It
>>>>>>>>>>> has an enforce config option that
>>>>>>>>>>> enables the bpf programs to deny access,
>>>>>>>>>>> providing access control. IIRC,
>>>>>>>>>>> in
>>>>>>>>>>> the earlier discussion threads, the BPF
>>>>>>>>>>> maintainers suggested that Smack
>>>>>>>>>>> and
>>>>>>>>>>> other LSMs could be entirely
>>>>>>>>>>> re-implemented via it in the future, and
>>>>>>>>>>> that
>>>>>>>>>>> such an implementation would be more optimal.
>>>>>>>>>>
>>>>>>>>>> In this case, the eBPF code is similar to a
>>>>>>>>>> kernel module, rather than a
>>>>>>>>>> loadable policy file.  It's a loadable
>>>>>>>>>> mechanism, rather than a policy, in
>>>>>>>>>> my view.
>>>>>>>>>
>>>>>>>>> I thought you frowned on dynamically loadable
>>>>>>>>> LSMs for both security and
>>>>>>>>> correctness reasons?
>>>>>>>
>>>>>>> Based on the feedback from the lists we've updated the design for v2.
>>>>>>>
>>>>>>> In v2, LSM hook callbacks are allocated dynamically using BPF
>>>>>>> trampolines, appended to a separate security_hook_heads and run
>>>>>>> only after the statically allocated hooks.
>>>>>>>
>>>>>>> The security_hook_heads for all the other LSMs (SELinux, AppArmor etc)
>>>>>>> still remains __lsm_ro_after_init and cannot be modified. We are still
>>>>>>> working on v2 (not ready for review yet) but the general idea can be
>>>>>>> seen here:
>>>>>>>
>>>>>>>       https://github.com/sinkap/linux-krsi/blob/patch/v1/trampoline_prototype/security/bpf/lsm.c
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> Evaluating the security impact of this is the next
>>>>>>>> step. My understanding
>>>>>>>> is that eBPF via BTF is constrained to read only access to hook
>>>>>>>> parameters, and that its behavior would be entirely restrictive.
>>>>>>>>
>>>>>>>> I'd like to understand the security impact more
>>>>>>>> fully, though.  Can the
>>>>>>>> eBPF code make arbitrary writes to the kernel, or
>>>>>>>> read anything other than
>>>>>>>> the correctly bounded LSM hook parameters?
>>>>>>>>
>>>>>>>
>>>>>>> As mentioned, the BPF verifier does not allow writes to BTF types.
>>>>>>>
>>>>>>>>> And a traditional security module would necessarily fall
>>>>>>>>> under GPL; is the eBPF code required to be
>>>>>>>>> likewise?  If not, KRSI is a
>>>>>>>>> gateway for proprietary LSMs...
>>>>>>>>
>>>>>>>> Right, we do not want this to be a GPL bypass.
>>>>>>>
>>>>>>> This is not intended to be a GPL bypass and the BPF verifier checks
>>>>>>> for license compatibility of the loaded program with GPL.
>>>>>>
>>>>>> IIUC, it checks that the program is GPL compatible if it
>>>>>> uses a function
>>>>>> marked GPL-only.  But what specifically is marked GPL-only
>>>>>> that is required
>>>>>> for eBPF programs using KRSI?
>>>>>
>>>>> Good point! If no-one objects, I can add it to the BPF_PROG_TYPE_LSM
>>>>> specific verification for the v2 of the patch-set which would require
>>>>> all BPF-LSM programs to be GPL.
>>>>
>>>> I don't think it's a good idea to enforce license on the program.
>>>> The kernel doesn't do it for modules.
>>>> For years all of BPF tracing progs were GPL because they have to use
>>>> GPL-ed helpers to do anything meaningful.
>>>> So for KRSI just make sure that all helpers are GPL-ed as well.
>>>
>>> IIUC, the example eBPF code included in this patch series showed a
>>> program that used a GPL-only helper for the purpose of reporting event
>>> output to userspace. But it could have just as easily omitted the use of
>>> that helper and still implemented its own arbitrary access control model
>>> on the LSM hooks to which it attached.  It seems like the question is
>>> whether the kernel developers are ok with exposing the entire LSM hook
>>> interface and all the associated data structures to non-GPLd code,
>>> irrespective of what helpers it may or may not use.
>>
>> Also, to be clear, while kernel modules aren't necessarily GPL, prior to
>> this patch series, all Linux security modules were necessarily GPLd in order
>> to use the LSM interface.
> 
> Because they use securityfs_create_file() GPL-ed api, right?
> but not because module license is enforced.

No, securityfs was a later addition and is not required by all LSMs 
either.  Originally LSMs had to register their hooks via 
register_security(), which was intentionally EXPORT_SYMBOL_GPL() to 
avoid exposing the LSM interface to non-GPLd modules because there were 
significant concerns with doing so when LSM was first merged.  Then in 
20510f2f4e2dabb0ff6c13901807627ec9452f98 ("security: Convert LSM into a 
static interface"), the ability for loadable modules to use 
register_security() at all was removed, limiting its use to built-in 
modules.  In commit b1d9e6b0646d0e5ee5d9050bd236b6c65d66faef ("LSM: 
Switch to lists of hooks"), register_security() was replaced by 
security_add_hooks(), but this was likewise not exported for use by 
modules and could only be used by built-in code.  The bpf LSM is 
providing a shim that allows eBPF code to attach to these hooks that 
would otherwise not be exposed to non-GPLd code, so if the bpf LSM does 
not require the eBPF programs to also be GPLd, then that is a change 
from current practice.

>> So allowing non-GPL eBPF-based LSMs would be a
>> change.
> 
> I don't see it this way. seccomp progs technically unlicensed. Yet they can
> disallow any syscall. Primitive KRSI progs like
> int bpf-prog(void*) { return REJECT; }
> would be able to do selectively disable a syscall with an overhead acceptable
> in production systems (unlike seccomp). I want this use case to be available to
> people. It's a bait, because to do real progs people would need to GPL them.
> Key helpers bpf_perf_event_output, bpf_ktime_get_ns, bpf_trace_printk are all
> GPL-ed. It may look that most networking helpers are not-GPL, but real life is
> different. To debug programs bpf_trace_printk() is necessary. To have
> communication with user space bpf_perf_event_output() is necssary. To measure
> anything or implement timestamps bpf_ktime_get_ns() is necessary. So today all
> meaninful bpf programs are GPL. Those that are not GPL probably exist, but
> they're toy programs. Hence I have zero concerns about GPL bypass coming from
> tracing, networking, and, in the future, KRSI progs too.

You have more confidence than I do about that.  I would anticipate 
developers of out-of-tree LSMs latching onto this bpf LSM and using it 
to avoid GPL.  I don't see that any of those helpers are truly needed to 
implement an access control model.