[RFC][PATCH 4/4] security: Enforce limitations on return values from LSMs

[Paul Moore]

On Tue, Nov 15, 2022 at 12:58 PM Roberto Sassu
<roberto.sassu at huaweicloud.com> wrote:
>
> From: Roberto Sassu <roberto.sassu at huawei.com>
>
> LSMs should not be able to return arbitrary return values, as the callers
> of the LSM infrastructure might not be ready to handle unexpected values
> (e.g. positive values that are first converted to a pointer with ERR_PTR,
> and then evaluated with IS_ERR()).
>
> Modify call_int_hook() to call is_ret_value_allowed(), so that the return
> value from each LSM for a given hook is checked. If for the interval the
> return value falls into the corresponding flag is not set, change the
> return value to the default value, just for the current LSM.
>
> A misbehaving LSM would not have impact on the decision of other LSMs, as
> the loop terminates whenever the return value is not zero.
>
> Signed-off-by: Roberto Sassu <roberto.sassu at huawei.com>
> ---
>  security/security.c | 34 ++++++++++++++++++++++++++++++++++
>  1 file changed, 34 insertions(+)

Casey touched on some of this in his reply to patch 0/4, but basically
I see this as a BPF LSM specific problem and not a generalized LSM
issue that should be addressed at the LSM layer.  Especially if the
solution involves incurring additional processing for every LSM hook
instantiation, regardless if a BPF LSM is present.  Reading your
overall patchset description I believe that you understand this too.

If you want to somehow instrument the LSM hook definitions (what I
believe to be the motivation behind patch 3/4) to indicate valid
return values for use by the BPF verifier, I think we could entertain
that, or at least discuss it further, but I'm not inclined to support
any runtime overhead at the LSM layer for a specific LSM.

-- 
paul-moore.com