Re: [RFC] Rename various 'IA32' uses in arch/x86/ code

[Ingo Molnar <mingo@kernel.org>]

* Brian Gerst <brgerst@gmail.com> wrote:

> > Ok, so your goal is to allow the x32 ABI, but not 32-bit user-space?
> 
> It just seems odd that x32 (which is really a 64-bit ABI with 32-bit pointers) 
> depended on enabling 32-bit support.  Other than both using the core compat 
> code, they are not really related.

Yeah.

> > I suppose that makes some sense, it might be a valid 'attack surface 
> > reduction' technique, while still allowing the x32 ABI.
> >
> > But I'm not sure we should bother and complicate things: 32-bit compat isn't 
> > going away anytime soon, and most of CONFIG_COMPAT is needed for x32.
> 
> Many of the compat syscalls are unused by x32.  It only needs to handle syscalls 
> with pointers embedded in data structures differently than native 64-bit.

Yeah, but in fact those are the 'most interesting' (read: most complex) aspects of 
the generic compat machinery. So most of the 'core compat' functionality is used - 
even though we don't use many of the (trivial) argument-converted syscall 
variants.

> 64-bit integer arguments (ie., loff_t) do not need special handling, since they 
> can be passed in a single register instead of a pair of 32-bit registers.  This 
> won't solve that particular issue yet, but it's something to be aware of for 
> future cleanups.

Yes, and I think 'X32' is a misnomer in that sense: in reality it's a 90% 64-bit 
ABI that just happens to have a handful of additional system calls that can deal 
with data pointers truncated to 32 bits.

So 'C64' would have been a better name: a compacted 64-bit ABI - but that 
particular name has its own problems ;-) Maybe S64 (small 64-bit memory model)?

'S64' would also have been an easier sell to distros: they generally resist adding 
anything that smells old, 32-bit ... but kernel hackers and marketing were always 
somewhat disjunct sets ;-)

I'm wondering whether we'll ever see a 48-bit user-space pointer model ;-) They 
are misaligned by 32 bits, but x86 CPUs generally handle 32-bit misalignment just 
fine. The killer would be to zero-extend from 48 bits to 64 bits I suspect - 
there's no natural instruction for that.

> > So maybe we could introduce CONFIG_X86_32_ABI=y or so, which would cover just 
> > the 32-bit entry code and the signal frame compatibility layer?
> 
> Yes.

Ok, then it sounds good to me!

Thanks,

	Ingo