Re: [PATCH RFC 1/4] syscalls: Create unified partial table for all archs

On Fri, Jul 24, 2026, at 23:22, H. Peter Anvin wrote:
> On July 24, 2026 1:25:05 PM PDT, Arnd Bergmann <[email protected]> wrote:

>>> It will wrongly add
>>>
>>> __SYSCALL_WITH_COMPAT(441, sys_epoll_pwait2, compat_sys_epoll_pwait2)
>>>
>>> to arch/x86/include/generated/asm/syscalls_64.h.
>>>
>>> Maybe I could add a --ignore-compat to scripts/syscalltbl.sh, and add 
>>> this flag for 64 builds. What do you think?
>>
>>IIRC, we can just use compat_sys_epoll_pwait2 for both i386
>>and x32, because compat_sigset_t is compatible with sigset_t --
>>they are only really different on big-endian targets, which
>>count the bits in a u64 different from two u32.
>>
>>This problem also goes away once x32 is gone.
>
> No reason to introduce another entry point if there's no difference. 
>
> x86, after all, is littleendian and alignment-tolerant.

What I meant was the reverse: if the common table has the
compat_sys_epoll_pwait2 entry for compat targets and x32 is
the only outlier, we don't need a special case for that
to force it to use the sys_epoll_pwait2 variant since they
are identical on x86.

Looking again at the x86 table, I see that x32 is not even
described as a compat target and just takes the second (native
64-bit) argument of __SYSCALL_WITH_COMPAT() instead of third
anyway, so there wouldn't even be a problem if native
and compat were different.

     Arnd