Credentials not fully initialized before bprm_check LSM hook

[sergeh at kernel.org]

On Thu, Apr 10, 2025 at 01:47:07PM +0200, Roberto Sassu wrote:
> Hi everyone
> 
> recently I discovered a problem in the implementation of our IMA
> bprm_check hook, in particular when the policy is matched against the
> bprm credentials (to be committed later during execve().
> 
> Before commit 56305aa9b6fab ("exec: Compute file based creds only
> once"), bprm_fill_uid() was called in prepare_binprm() and filled the
> euid/egid before calling security_bprm_check(), which in turns calls
> IMA.
> 
> After that commit, bprm_fill_uid() was moved to begin_new_exec(), which
> is when the last interpreter is found.
> 
> The consequence is that IMA still sees the not yet ready credentials
> and an IMA rule like:
> 
> measure func=CREDS_CHECK euid=0

"IMA still sees" at which point exactly?

Do I understand right that the problem is that ima's version of
security_bprm_creds_for_exec() needs to run after
bprm_creds_from_file()?

Given that Eric's commit message said that no bprm handlers use
the uid, it seems it should be safe to just move that?

> will not be matched for sudo-like applications.
> 
> It does work however with SELinux, because it computes the transition
> before IMA in the bprm_creds_for_exec hook.
> 
> Since IMA needs to be involved for each execution in the chain of
> interpreters, we cannot move to the bprm_creds_from_file hook.
> 
> How do we solve this problem? The commit mentioned that it is an
> optimization, so probably would not be too hard to partially revert it
> (and keeping what is good).
> 
> Thanks
> 
> Roberto
>