Re: [PATCH] audit: Report suspicious O_CREAT usage

From: Kees Cook
Date: Mon Sep 30 2019 - 17:23:40 EST


On Mon, Sep 30, 2019 at 09:50:00AM -0400, Steve Grubb wrote:
> On Thursday, September 26, 2019 11:31:32 AM EDT Paul Moore wrote:
> > On Wed, Sep 25, 2019 at 5:02 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > > This renames the very specific audit_log_link_denied() to
> > > audit_log_path_denied() and adds the AUDIT_* type as an argument. This
> > > allows for the creation of the new AUDIT_ANOM_CREAT that can be used to
> > > report the fifo/regular file creation restrictions that were introduced
> > > in commit 30aba6656f61 ("namei: allow restricted O_CREAT of FIFOs and
> > > regular files"). Without this change, discovering that the restriction
> > > is enabled can be very challenging:
> > > https://lore.kernel.org/lkml/CA+jJMxvkqjXHy3DnV5MVhFTL2RUhg0WQ-XVFW3ngDQO
> > > dkFq0PA@xxxxxxxxxxxxxx
> > >
> > > Reported-by: Jérémie Galarneau <jeremie.galarneau@xxxxxxxxxxxx>
> > > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > > ---
> > > This is not a complete fix because reporting was broken in commit
> > > 15564ff0a16e ("audit: make ANOM_LINK obey audit_enabled and
> > > audit_dummy_context")
> > > which specifically goes against the intention of these records: they
> > > should _always_ be reported. If auditing isn't enabled, they should be
> > > ratelimited.
> > >
> > > Instead of using audit, should this just go back to using
> > > pr_ratelimited()?
> >
> > I'm going to ignore the rename and other aspects of this patch for the
> > moment so we can focus on the topic of if/when/how these records
> > should be emitted by the kernel.
> >
> > Unfortunately, people tend to get very upset if audit emits *any*
> > records when they haven't explicitly enabled audit, the significance
> > of the record doesn't seem to matter, which is why you see patches
> > like 15564ff0a16e ("audit: make ANOM_LINK obey audit_enabled and
> > audit_dummy_context"). We could consider converting some records to
> > printk()s, rate-limited or not, but we need to balance this with the
> > various security certifications which audit was created to satisfy.
> > In some cases a printk() isn't sufficient.
> >
> > Steve is probably the only one who really keeps track of the various
> > auditing requirements of the different security certifications; what
> > say you Steve on this issue with ANOM_CREAT records?
>
> Common Criteria and other security standards I track do not call out for
> anomoly detection. So, there are no requirements on this. That said, we do
> have other anomaly detections because they give early warning that something
> strange is happening. I think adding this event is a nice improvement as long
> as it obeys audit_enabled before emitting an event - for example, look at the
> AUDIT_ANOM_ABEND event.

Okay, so the patch is good as-is? (The "report things always" issue I
will deal with separately. For now I'd just like to gain this anomaly
detection corner case...)

Paul, what do you see as next steps here?

--
Kees Cook