On Tue, Jul 29, 2014 at 07:49:47AM +0100, AKASHI Takahiro wrote:

On 07/25/2014 08:03 PM, Will Deacon wrote:

quoted

On Fri, Jul 25, 2014 at 11:36:49AM +0100, AKASHI Takahiro wrote:

quoted

On 07/25/2014 12:01 AM, Andy Lutomirski wrote:

quoted

quoted

quoted

If so, then you risk (at least) introducing

a nice user-triggerable OOPS if audit is enabled.

Can you please elaborate this?
Since I didn't find any definition of audit's behavior when syscall is
rewritten to -1, I thought it is reasonable to skip "exit tracing" of
"skipped" syscall.
(otherwise, "fake" seems to be more appropriate :)

The audit entry hook will oops if you call it twice in a row without
calling the exit hook in between.

Thank you, I could reproduce this problem which hits BUG(in_syscall) in
audit_syscall_entry(). Really bad, and I fixed it in my next version and
now a "skipped" system call is also traced by audit.

Can you reproduce this on arch/arm/ too? If so, we should also fix the code
there.

As far as I tried on arm with syscall auditing enabled,

1) Changing a syscall number to -1 under seccomp doesn't hit BUG_ON(in_syscall).
2) But, in fact, audit_syscall_entry() is NOT called in this case because
    __secure_computing() returns -1 and then it causes the succeeding tracing
    in syscall_trace_enter(), including audit_syscall_entry(), skipped.

What happens if CONFIG_SECCOMP=n?

3) On the other hand, calling syscall(-1) from userspace hits BUG_ON because
    the return path, ret_slow_syscall, doesn't contain syscall_trace_exit().
4) When we re-write a syscall number to -1 without seccomp, we will also see
    BUG_ON hit, although I didn't try yet.

Fixing case 3 is easy, but should we also fix case 2?

I think so.

Will