Re: [PATCH v7 1/3] random: add vgetrandom_alloc() syscall

From: Arnd Bergmann
Date: Mon Nov 28 2022 - 08:55:11 EST


On Sun, Nov 27, 2022, at 21:18, Jason A. Donenfeld wrote:
>>
>> config GENERIC_VDSO_RANDOM_WHATEVER
>> bool
>> select VGETRANDOM_ALLOC
>>
>> This gives a clear Kconfig dependency instead of the random
>> ADVISE_SYSCALLS select.
>
> That's much better indeed. I was trying to straddle the two conventions
> of `#define __ARCH_...` for syscalls and a Kconfig for vDSO functions,
> but doing it all together as you've suggested is nicer.
>
> I'll try to figure this out, though so far futzing around suggests there
> might have to be both, because of unistd.h being a userspace header.
> That is, include/uapi/asm-generic/unistd.h typically needs a `#if
> __ARCH_WANT..., #define ...` in it. I'll give it a spin and you'll see
> for v8. At the very least it should get rid of the more awkward
> `select ADVISE_SYSCALLS if X86_64` part, and will better separate the
> arch code from non-arch code.

I think you should not need an __ARCH_WANT_SYS_* symbol for this,
the only place we actually need them for is the asm-generic/unistd.h
header which is still used on a couple of architectures (I have
an experimental series for replacing it with a generic syscall.tbl
file, but it's not ready for 6.2). In most cases, the __ARCH_WANT_SYS_*
symbols are only used for syscalls that are part of the table for
old architectures but get skipped on newer targets that always had
a replacement syscalls (e.g. getrlimit getting replaced by prlimit64)

I think we should just reserve the syscall number for all architectures
right away and #define the __NR_* macro. libc will generally need
a runtime check anyway, and defining it now avoids the problem of
the tables getting out of sync.

The Kconfig symbol is fine in this case.

Arnd