Re: git: 4a69fc16a583 - main - Add membarrier(2)

On Thu, Aug 24, 2023 at 10:19 AM John Baldwin <jhb@freebsd.org> wrote:

> On 8/23/23 6:01 AM, Warner Losh wrote:
> > On Tue, Aug 22, 2023 at 11:37 PM Jessica Clarke <jrtc27@freebsd.org>
> wrote:
> >
> >>
> >>> The addition does not change any existing code path in the kernel.
> >>
> >> No, but it commits us to a new syscall being stable just days before 14
> >> branches and has its ABI frozen.
> >>
> >
> > I'd planned on committing timerfd later today. I didn't consider it an
> ABI
> > breakage, since it was just additive. It's one of the things that can be
> > MFC'd (we don't prohibit new system calls).
>
> (Not taking a side on the merits of the current membarrier(2)
> implementation)
>
> I think Jess's point here is not that a new syscall is not a valid ABI
> breakage, but more that once you add a new syscall that makes it into a
> release, now the ABI of that syscall is frozen and can't be changed.  That
> said, syscall numbers are relatively "cheap", so if we had to renumber
> membarrier(2) because its ABI was found to be a problem that could be done,
> albeit at the cost of keeping the old one around under COMPAT_FREEBSD<n>.
>

OK. I jumped to the wrong conclusion...  Good points all around.


Warner