On Wed, Aug 21, 2024 at 03:43:48AM +0000, Song Liu wrote:

quoted

On Aug 20, 2024, at 2:11 PM, Paul Moore [off-list ref] wrote:

On Tue, Aug 20, 2024 at 1:43 PM Song Liu [off-list ref] wrote:

quoted

quoted

On Aug 20, 2024, at 5:45 AM, Mickaël Salaün [off-list ref] wrote:

...

quoted

quoted

What about adding BPF hooks to Landlock?  User space could create
Landlock sandboxes that would delegate the denials to a BPF program,
which could then also allow such access, but without directly handling
nor reimplementing filesystem path walks.  The Landlock user space ABI
changes would mainly be a new landlock_ruleset_attr field to explicitly
ask for a (system-wide) BPF program to handle access requests if no
Landlock rule allow them.  We could also tie a BPF data (i.e. blob) to
Landlock domains for consistent sandbox management.  One of the
advantage of this approach is to only run related BPF programs if the
sandbox policy would deny the request.  Another advantage would be to
leverage the Landlock user space interface to let any program partially
define and extend their security policy.

Given there is BPF LSM, I have never thought about adding BPF hooks to
Landlock or other LSMs. I personally would prefer to have a common API
to walk the path, maybe something like vma_iterator. But I need to read
more code to understand whether this makes sense?

I think it would not be an issue to use BPF Landlock hooks along with
BPF LSM hooks for the same global policy.  This could also use the
Landlock domain concept for your use case, including domain inheritance,
domain identification, cross-domain protections... to avoid
reimplementing the same semantic (and going through the same issues).
Limiting the BPF program calls could also improve performance.

quoted

Just so there isn't any confusion, I want to make sure that everyone
is clear that "adding BPF hooks to Landlock" should mean "add a new
Landlock specific BPF hook inside Landlock" and not "reuse existing
BPF LSM hooks inside Landlock".

I think we are on the same page. My understanding of Mickaël's idea is
to add some brand new hooks to Landlock code, so that Landlock can
use BPF program to make some decisions. 

Correct