Re: v5 of seccomp filter c/r patches

On Fri, Oct 2, 2015 at 3:06 PM, Kees Cook [off-list ref] wrote:

On Fri, Oct 2, 2015 at 3:04 PM, Andy Lutomirski [off-list ref] wrote:

quoted

On Fri, Oct 2, 2015 at 3:02 PM, Kees Cook [off-list ref] wrote:

quoted

On Fri, Oct 2, 2015 at 2:29 PM, Andy Lutomirski [off-list ref] wrote:

quoted

On Fri, Oct 2, 2015 at 2:10 PM, Kees Cook [off-list ref] wrote:

quoted

On Fri, Oct 2, 2015 at 9:27 AM, Tycho Andersen
[off-list ref] wrote:

quoted

Hi all,

Here's v5 of the seccomp filter c/r set. The individual patch notes have
changes, but two highlights are:

* This series is now based on http://patchwork.ozlabs.org/patch/525492/ and
  will need to be built with that patch applied. This gets rid of two incorrect
  patches in the previous series and is a nicer API.

* I couldn't figure out a nice way to have SECCOMP_GET_FILTER_FD return the
  same struct file across calls, so we still need a kcmp command. I've narrowed
  the scope of the one being added to only compare seccomp fds.

Thoughts welcome,

Hi, sorry I've been slow/busy. I'm finally reading through these threads.

Happy bit:
- avoiding eBPF and just saving the original filters makes things much easier.

Sad bit:
- inventing a new interface for seccompfds feels like massive overkill to me.

While Andy has big dreams, we're not presently doing seccompfd
monitoring, etc. There's no driving user for that kind of interface,
and accepting the maintenance burden of it only for CRIU seems unwise.

So, I'll go back to what I originally proposed at LSS (which it looks
like we're half way there now):

- save the original filter (done!)
- extract filters through a single special-purpose interface (looks
like ptrace is the way to go: root-only, stopped process, etc)
- compare filter content and issue TSYNCs to merge detected sibling
threads, since merging things that weren't merged before creates no
problems.

This means the parenting logic is heuristic, but it's entirely in
userspace, so the complexity burden doesn't live in seccomp which we,
by design, want to keep as simple as possible.

This is okay with me with a future-proofing caveat: I think that
whatever reads out the filter should be clearly documented as
returning some special error code that indicates that that filter it
tried to read wasn't in the expected form.  That would happen for
native eBPF filters, and it would also happen for seccomp monitors
even if those monitors use classic BPF.

As in, it should have something like "give me BPF" and that'll start
failing when it's only eBPF in the future?

Yes, but it might also start failing when if my dreams come true, it's
still classic BPF, but it's no longer a classic seccomp bpf filter
layer with the semantics we expect today.  (E.g. if it's classic bpf
but has a monitor attached, then the read should fail because
restoring it without restoring the monitor will cause all kinds of
mess.)

Ah-ha! Understood, and yeah, that seems fine.

Speaking of dreams -- what do you think about re-running seccomp in
the face of changed syscalls due to ptrace? Closing the ptrace hole
would be really nice.

Yes, absolutely!  We might even want to just move the seccomp check
after ptrace (except for seccomp-induced ptrace).

Unfortunately, I backed us into a corner with two-phase seccomp on
x86, and it's a big mess.  (I wrote the seccomp vs ptrace patches, and
I don't think they're acceptable.)  My big x86 low-level rewrite is an
attempt to get back out of that corner, and I'm hoping to resubmit the
bulk of it today or tomorrow.  Once that happens, I just need to fix
up the 64-bit native case (trivial, I know) and then revert two-phase
seccomp.

One nice outcome of all of this will be that the syscall tables will
contain bona fide C ABI compliant function pointers, which is
currently not the case.

--Andy