Re: [PATCH v3 13/13] epoll: implement epoll_create2() syscall

[Andrew Morton]

On Wed, 22 May 2019 13:14:41 +0200 Arnd Bergmann <arnd@xxxxxxxx> wrote:

> > I thought the preferred approach was to wire up the architectures on
> > which the submitter has tested the syscall, then allow the arch
> > maintainers to enable the syscall independently?
> 
> I'm hoping to change that practice now, as it has not worked well
> in the past:
> 
> - half the architectures now use asm-generic/unistd.h, so they are
>   already wired up at the same time, regardless of testing
> - in the other half, not adding them at the same time actually
>   made it harder to test, as it was significantly harder to figure
>   out how to build a modified kernel for a given architecture
>   than to run the test case
> - Not having all architectures add a new call at the same time caused
>   the architectures to get out of sync when some got added and others
>   did not. Now that we use the same numbers across all architectures,
>   that would be even more confusing.
>
> My plan for the long run is to only have one file to which new
> system calls get added in the future.

Fair enough.  We're adding code to architectures without having tested
it on those architectures but we do that all the time anyway - I guess
there's not a lot of point in special-casing new syscalls.