Re: [RFC v2 08/20] sysdeps/wait: Use waitid if avaliable

[Arnd Bergmann <arnd at arndb dot de>]

On Tue, Jun 25, 2019 at 4:08 PM Florian Weimer <fweimer@redhat.com> wrote:
> * Arnd Bergmann:
> > On Tue, Jun 25, 2019 at 3:47 PM Florian Weimer <fweimer@redhat.com> wrote:
> >> This makes me rather unhappy.  I also don't see the benefit of renaming
> >> all time-related system calls for new architectures.
> >
> > What got renamed?
>
> futex to futex_time64 on RV32.  <asm/unistd.h> seems to expose only the
> latter.

But that's the point: futex() takes a timespec argument based on 'long',
so that won't work beyond y2038 on 32-bit architectures and we
cannot provide that.

One of the highest priorities in the conversion was always to ensure
that new architectures would behave exactly the same way as the
existing ones, except for leaving out the compatibility support for
old C libraries that never existed on those architectures.

Reusing the old __NR_futex() macro with a new ABI would require
having architecture specific hacks even in C libraries that never
supported 32-bit time_t on any architectures, or that dropped that
support.

> > I was trying very hard to make it as consistent
> > and easy as possible. There is a strict mapping of __NR_* macros
> > to argument types for each system call [1], so e.g. __kernel_timespec
> > will always be used together with system calls named *time64(),
> > while the old system calls always refer to the traditional
> > "struct timespec {long tv_sec; long tv_nsec;}" type. This is the
> > same way that loff_t gets handled, so I assumed that all C libraries
> > would know how to deal with this well.
>
> I'm afraid, but this setup makes little sense if the old system call
> does not exist for an architecture.  Instead it requires additional
> porting effort.

I think we can still add back the time32 syscalls for rv32 in linux-5.2 if
that helps, it should be a one-line patch in the kernel (and it would
mean that linux-5.1 would be broken for anyone actually using those).

        Arnd