Re: [PATCH 2/2] seccomp.2: document userspace notification

[Tycho Andersen]

On Thu, Feb 28, 2019 at 02:25:55PM +0100, Michael Kerrisk (man-pages) wrote:
> > 7. The monitoring process can use the information in the
> >    'struct seccomp_notif' to make a determination about the
> >    system call being made by the target process. This
> >    structure includes a 'data' field that is the same
> >    'struct seccomp_data' that is passed to a BPF filter.
> > 
> >    In addition, the monitoring process may make use of other 
> >    information that is available from user space. For example, 
> >    it may inspect the memory of the target process (whose PID
> >    is provided in the 'struct seccomp_notif') using
> >    /proc/PID/mem, which includes inspecting the values
> >    pointed to by system call arguments (whose location is
> >    available 'seccomp_notif.data.args). However, when using
> >    the target process PID in this way, one must guard against
> >    PID re-use race conditions using the seccomp()
> >    SECCOMP_IOCTL_NOTIF_ID_VALID operation.
> > 
> > 8. Having arrived at a decision about the target process's
> >    system call, the monitoring process can inform the kernel
> >    of its decision using the operation
> > 
> >        ioctl(listenfd, SECCOMP_IOCTL_NOTIF_SEND, respptr)
> > 
> >    where the third argument is a pointer to a
> >    'struct seccomp_notif_resp'. [Some more details
> >    needed here, but I still don't yet understand fully
> >    the semantics of the 'error' and 'val' fields.]
> 
> So clearly, I misunderstood these last two steps.
> 
> (7) is something like: discover information in userspace
> as required; perform userspace actions if appropriate
> (perhaps doing the system call operation "on behalf of" the
> target process).
> 
> 
> (8) is something like:
>    set 'error' and 'val' to return info to the target process:
>     * error != 0 ==> make it look like the syscall failed,
>       with 'errno' set to that value
>     * error == 0 ==> make it look like the syscall succeeded 
>       and returned 'val'
> 
> Right?

Yep, exactly.

Tycho