Re: [PATCH] audit: do a quick exit when syscall number is invalid

[Richard Guy Briggs <rgb@redhat.com>]

On 2022-04-06 01:19, CGEL wrote:
> On Mon, Apr 04, 2022 at 11:58:50AM -0400, Richard Guy Briggs wrote:
> > On 2022-04-02 08:06, CGEL wrote:
> > > On Fri, Apr 01, 2022 at 10:16:45AM -0400, Paul Moore wrote:
> > > > On Fri, Apr 1, 2022 at 9:39 AM Steve Grubb <sgrubb@redhat.com> wrote:
> > > > > On Thursday, March 31, 2022 9:57:05 PM EDT CGEL wrote:
> > > > > > On Thu, Mar 31, 2022 at 10:16:23AM -0400, Paul Moore wrote:
> > > > > > > On Wed, Mar 30, 2022 at 10:29 PM CGEL <cgel.zte@gmail.com> wrote:
> > > > > > > > On Wed, Mar 30, 2022 at 10:48:12AM -0400, Paul Moore wrote:
> > > > > > > > > If audit is not generating SYSCALL records, even for invalid/ENOSYS
> > > > > > > > > syscalls, I would consider that a bug which should be fixed.
> > > > > > > >
> > > > > > > > If we fix this bug, do you think audit invalid/ENOSYS syscalls better
> > > > > > > > be forcible or be a rule that can be configure? I think configure is
> > > > > > > > better.
> > > > > > >
> > > > > > > It isn't clear to me exactly what you are asking, but I would expect
> > > > > > > the existing audit syscall filtering mechanism to work regardless if
> > > > > > > the syscall is valid or not.
> > > > > >
> > > > > > Thanks, I try to make it more clear. We found that auditctl would only
> > > > > > set rule with syscall number (>=0 && <2047) ...
> > > > 
> > > > That is exactly why I wrote the warning below in my response ...
> > > >
> > > I think the question is more clear now.
> > > 
> > > 1) libaudit.c wants to forbid setting invalid syscall, but inconsistent
> > > Currently way(>=0 && <2047) is inconsistent, syscall with number 2000 and
> > > syscall with number 3000 are both invalid syscall. But 2000 can be set by
> > > auditctl, and 3000 cannot be set by auditctl.
> > > A better way to do this forbidden is to use __NR_syscalls(asm-generic/unistd.h).
> > > 
> > > 2) if libaudit.c do the right forbidden, kernel better ignore invalid syscall
> > > See this patch.
> > > 
> > > If we want audit invalid syscall as you said before. libaudit.c should not
> > > do the forbidden, auditctl should allow setting syscall rule with 'any' number.
> > > So do you think we should fix libaudit.c?
> > 
> > I'm having a bit of trouble understanding what you've said above.
> > 
> > The kernel ultimately must protect itself from malice and mistakes, so
> > it must verify all data sent to it.
> > 
> > Userspace can help by knowing what that kernel policy is so it can avoid
> > violating that policy or provide useful feedback if it can't.  Userspace
> > can be used to make things more efficient, but the kernel is the last
> > step for security.
> > 
> > If userspace and the kernel are mismatched or out of sync, then the
> > kernel enforces policy to protect itself.
>
> Much appreciate for your interpretation. Have you get any idea of how
> to solve the mismatched? From your viewpoint, I think it's better for
> kernel to not handle syscall of syscall number<0, because it's invaild
> of all arch, and has no value for attacker to probing for specific
> syscall numbers.

Going back to the very first quoted line above, if you can generate a
test case that shows that audit is missing an auditable event, that is a
bug that should be fixed.

> > > > > > > to the audit syscall filter, which are unfortunately baked into the
> > > > > > > current design/implementation, which may affect this to some extent.
> > > > 
> > > > -- 
> > > > paul-moore.com
> > 
> > - RGB

- RGB

--
Richard Guy Briggs <rgb@redhat.com>
Sr. S/W Engineer, Kernel Security, Base Operating Systems
Remote, Ottawa, Red Hat Canada
IRC: rgb, SunRaycer
Voice: +1.647.777.2635, Internal: (81) 32635