Re: [PATCHv2 1/1] Documentation: describe how to add a system call

[Josh Triplett <josh@joshtriplett.org>]

On Fri, Jul 31, 2015 at 03:54:56PM -0700, Andy Lutomirski wrote:
> On Fri, Jul 31, 2015 at 3:08 PM,  <josh@joshtriplett.org> wrote:
> > On Fri, Jul 31, 2015 at 02:19:29PM -0700, Andy Lutomirski wrote:
> >> On Fri, Jul 31, 2015 at 1:59 PM,  <josh@joshtriplett.org> wrote:
> >> > Agreed.  I think the proposal above would be a net improvement, but
> >> > ideally you'd want something that's annotated and generates automatic
> >> > marshalling code.
> >> >
> >>
> >> I assume this is idle musing.  If, however, we were to actually do
> >> this, I'd suggest we seriously consider speaking the Cap'n Proto
> >> serialization format.  It's quite nice, it encodes and decodes *very*
> >> quickly and, unlike TLV schemes, you don't have to read it in order,
> >> making the read-side code less awkward.
> >
> > That seems like *massive* overkill for a kernel<->userspace syscall
> > interface.  I was more thinking about having a few standardized marshal
> > types, and incrementally adding more when more patterns show up.  For a
> > first pass, just automatically running copy_from_user and
> > copy_param_struct on appropriate sets of __user parameters identified as
> > such in a structured text file seems quite sufficient.  (Plus
> > automatically generating syscalls.h from that.)
> 
> If a param struct does the trick, then I agree.  It's when you start
> having lists and other variable-size stuff that it gets messier.

Sure, agreed.  But I really hope we don't create new kernel ABIs that
involve constructs like that.

- Josh Triplett