Re: [RFC][PATCH 1/5] atomic: Introduce atomic_{inc,dec,dec_and_test}_ofl()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Dec 09, 2021 at 12:42:47PM +0000, Mark Rutland wrote:
> On Wed, Dec 08, 2021 at 07:36:56PM +0100, Peter Zijlstra wrote:
> > In order to facilitate architecture support for refcount_t, introduce
> > a number of new atomic primitives that have a uaccess style exception
> > for overflow.
> > 
> > Notably:
> > 
> >   atomic_inc_ofl(v, Label) -- increment and goto Label when
> > 			      v is zero or negative.
> > 
> >   atomic_dec_ofl(v, Label) -- decrement and goto Label when
> > 			      the result is zero or negative
> > 
> >   atomic_dec_and_test_ofl(v, Label) -- decrement and return true when
> > 				       the result is zero and goto Label
> > 				       when the result is negative
> 
> Just to check, atomic_inc_ofl() tests the *old* value of `v`, and the other
> cases check the *new* value of `v`?
> 
> For clarity, in the descriptions it might be worth:
> 
>   s/v/the old value of v/
>   s/the result/the new value of v/
> 
> ... which I think makes that clearer.

Right, I'll clarify.

> > Since the GCC 'Labels as Values' extention doesn't allow having the
> > goto in an inline function, these new 'functions' must in fact be
> > implemented as macro magic.
> 
> Oh; fun... :(

Yeah, I tried all sorta things, it's all >.< close to working but then
GCC refuses to do the sensible thing.

> > This meant extending the atomic generation scripts to deal with
> > wrapping macros instead of inline functions. Since
> > xchg/cmpxchg/try_cmpxchg were already macro magic, there was existant
> > code for that. While extending/improving that a few latent
> > 'instrumentation' bugs were uncovered and 'accidentally' fixed.
> 
> I assume for non-RFC we can split that out into a preparatory patch. :)

Sure, I can split it in two; one add the infra and fix bugs and two
introduce the new ops.