[PATCH bpf-next v4 3/8] bpf: lsm: provide attachment points for BPF LSM programs

[Casey Schaufler]

On 2/24/2020 9:41 PM, Alexei Starovoitov wrote:
> On Mon, Feb 24, 2020 at 01:41:19PM -0800, Kees Cook wrote:
>> But the LSM subsystem doesn't want special cases (Casey has worked very
>> hard to generalize everything there for stacking). It is really hard to
>> accept adding a new special case when there are still special cases yet
>> to be worked out even in the LSM code itself[2].
>> [2] Casey's work to generalize the LSM interfaces continues and it quite
>> complex:
>> https://lore.kernel.org/linux-security-module/20200214234203.7086-1-casey@schaufler-ca.com/
> I think the key mistake we made is that we classified KRSI as LSM.
> LSM stacking, lsmblobs that the above set is trying to do are not necessary for KRSI.
> I don't see anything in LSM infra that KRSI can reuse.
> The only thing BPF needs is a function to attach to.
> It can be a nop function or any other.
> security_*() functions are interesting from that angle only.
> Hence I propose to reconsider what I was suggesting earlier.
> No changes to secruity/ directory.
> Attach to security_*() funcs via bpf trampoline.
> The key observation vs what I was saying earlier is KRSI and LSM are wrong names.
> I think "security" is also loaded word that should be avoided.

No argument there.

> I'm proposing to rename BPF_PROG_TYPE_LSM into BPF_PROG_TYPE_OVERRIDE_RETURN.
>
>> So, unless James is going to take this over Casey's objections, the path
>> forward I see here is:
>>
>> - land a "slow" KRSI (i.e. one that hooks every hook with a stub).
>> - optimize calling for all LSMs
> I'm very much surprised how 'slow' KRSI is an option at all.
> 'slow' KRSI means that CONFIG_SECURITY_KRSI=y adds indirect calls to nop
> functions for every place in the kernel that calls security_*().
> This is not an acceptable overhead. Even w/o retpoline
> this is not something datacenter servers can use.

In the universe I live in data centers will disable hyper-threading,
reducing performance substantially, in the face of hypothetical security
exploits. That's a massively greater performance impact than the handful
of instructions required to do indirect calls. Not to mention the impact
of the BPF programs that have been included. Have you ever looked at what
happens to system performance when polkitd is enabled?


>
> Another option is to do this:
> diff --git a/include/linux/security.h b/include/linux/security.h
> index 64b19f050343..7887ce636fb1 100644
> --- a/include/linux/security.h
> +++ b/include/linux/security.h
> @@ -240,7 +240,7 @@ static inline const char *kernel_load_data_id_str(enum kernel_load_data_id id)
>         return kernel_load_data_str[id];
>  }
>
> -#ifdef CONFIG_SECURITY
> +#if defined(CONFIG_SECURITY) || defined(CONFIG_BPF_OVERRIDE_RETURN)
>
> Single line change to security.h and new file kernel/bpf/override_security.c
> that will look like:
> int security_binder_set_context_mgr(struct task_struct *mgr)
> {
>         return 0;
> }
>
> int security_binder_transaction(struct task_struct *from,
>                                 struct task_struct *to)
> {
>         return 0;
> }
> Essentially it will provide BPF side with a set of nop functions.
> CONFIG_SECURITY is off. It may seem as a downside that it will force a choice
> on kernel users. Either they build the kernel with CONFIG_SECURITY and their
> choice of LSMs or build the kernel with CONFIG_BPF_OVERRIDE_RETURN and use
> BPF_PROG_TYPE_OVERRIDE_RETURN programs to enforce any kind of policy. I think
> it's a pro not a con.

Err, no. All distros use an LSM or two. Unless you can re-implement SELinux
in BPF (good luck with state transitions) you've built a warp drive without
ever having mined dilithium crystals.