Re: arc_usr_cmpxchg and preemption

[Alexey Brodkin]

Hi Peter,

On Thu, 2018-03-15 at 12:28 +0100, Peter Zijlstra wrote:
> On Thu, Mar 15, 2018 at 09:12:09AM +0000, Alexey Brodkin wrote:
> > On Thu, 2018-03-15 at 09:18 +0100, Peter Zijlstra wrote:
> > > Also, it might make sense to stuff this implementation in some lib/ file
> > > somewhere and make all platforms that need it use the same code, afaict
> > > there really isn't anything platform specific to it.
> > 
> > Not clear which part do you mean here.
> > Are you talking about entire cmpxchg syscall implementation?
> 
> Yep.

Hm... new generic syscall doing something sane people are not supposed to do?
Let's see who's going to express his/her excitement about that :)

But even introduction of that new syscall is obviously not enough
as we'll need to fix-up libc for affected arches accordingly...

> > Do you think there're many users of that quite an inefficient
> > [compared to proper HW version] atomic exchange?
> 
> I think there's a bunch of architectures that are in the same boat.
> m68k, arm, mips was mentioned. Sure, the moment an arch has hardware
> support you don't need the syscall anymore.

Here's a brief analysis:
ARM:  Looks like they got rid of that stuff in v4.4, see
      commit db695c0509d6 ("ARM: remove user cmpxchg syscall").

M68K: That's even uglier implementation which is really asking for
      a facelift, look at sys_atomic_cmpxchg_32() here:
      https://elixir.bootlin.com/linux/latest/source/arch/m68k/kernel/sys_m68k.c#L461

MIPS: They do it via special sysmips syscall which among other things
      might handle MIPS_ATOMIC_SET with mips_atomic_set()

I don't immediately see if there're others but really I'm not sure if it even worth trying to
clean-up all that since efforts might be spent pointlessly.

> I was just thinking it would be good to have a common implementation (if
> possible) rather than 4-5 different copies of basically the same thing.

From above I would conclude that only M68K might benefit from new library
implementation. BTW M68K uses a bit different ABI compared to ARC for that syscall so
it will be really atomic_cmpxchg_32() libfunction called from those syscalls,
but now I think that's exactly what you meant initially, correct?

-Alexey