[RFC PATCH v1 3/7] landlock: Log ruleset creation and release

[Mickaël Salaün]

On Fri, Dec 29, 2023 at 06:42:10PM +0100, Mickaël Salaün wrote:
> On Fri, Dec 22, 2023 at 05:42:35PM -0500, Paul Moore wrote:
> > On Thu, Dec 21, 2023 at 1:45 PM Mickaël Salaün <mic at digikod.net> wrote:
> > > On Wed, Dec 20, 2023 at 04:22:15PM -0500, Paul Moore wrote:
> > > > On Thu, Sep 21, 2023 at 2:17 AM Mickaël Salaün <mic at digikod.net> wrote:
> > > > >
> > > > > Add audit support for ruleset/domain creation and release ...
> > 
> > ...

> > > For rule addition, several records per landlock_add_rule(2) call.
> > > Example with a path_beneath rule:
> > > - AUDIT_LANDLOCK_RULESET: "id=[ruleset ID] op=add_rule"
> > > - AUDIT_LANDLOCK_PATH: "scope=beneath path=[file path] dev= ino="
> > > - AUDIT_LANDLOCK_ACCESS: "type=fs rights=[bitmask]"
> > 
> > I worry that LANDLOCK_PATH is too much of a duplicate for the existing
> > PATH record.  Assuming the "scope=" field is important, could it live
> > in the LANDLOCK_ACCESS record and then you could do away with the
> > dedicated LANDLOCK_PATH record?  Oh, wait ... this is to record the
> > policy, not a individual access request, gotcha.  If that is the case
> > and RULESET, PATH, ACCESS are all used simply to record the policy
> > information I might suggest creation of an AUDIT_LANDLOCK_POLICY
> > record that captures all of the above.  If you think that is too
> > cumbersome, then perhaps you can do the object/access-specific record
> > type, e.g. AUDIT_LANDLOCK_POLICY_FS and AUDIT_LANDLOCK_POLICY_NET.
> 
> OK, what about this records for *one* rule addition event?
> 
> - AUDIT_LANDLOCK_RULE: "ruleset=[ruleset ID] rule_type=path_beneath
>   allowed_access=[bitmask]"
> - AUDIT_PATH: "path=[file path] dev= ino= ..."
> 
> However, because struct landlock_path_beneath_attr can evolve and get
> new fields which might be differents than the landlock_net_port_attr's
> ones, wouldn't it be wiser to use a dedicated AUDIT_LANDLOCK_RULE_FS or
> AUDIT_LANDLOCK_RULE_PATH_BENEATH record type? These names are getting a
> bit long though, but types match the UAPI.

Hmm, AUDIT_PATH is used when a syscall's argument is a path, but in the
case of Landlock, the arguments are file descriptors.

I can still export audit_copy_inode() to create a synthetic audit_names
struct, and export/call audit_log_name() to create an AUDIT_PATH entry
but I'm not sure it is the best approach. What would you prefer?
Should I use AUDIT_TYPE_NORMAL or create a new one?

[...]

> > > For denied FS access:
> > > - AUDIT_LANDLOCK_DENIAL: "id=[domain ID] op=mkdir"
> > > - AUDIT_LANDLOCK_PATH: "scope=exact path=[file path] dev= ino="
> > 
> > I would use a single record type, i.e. AUDIT_LANDLOCK_ACCESS, to
> > capture both access granted and denied events.  I'd also omit the
> > dedicated LANDLOCK_PATH record here in favor of the generic PATH
> > record (see my comments above).
> 
> Makes sense for the generic PATH record. We would get this:
> 
> - AUDIT_LANDLOCK_ACCESS: "domain=[domain ID] op=mkdir result=denied"
> - AUDIT_PATH: "path=[file path] dev= ino= ..."