Re: [PATCH locking/atomic 18/19] locking/atomic: Refrain from generating duplicate fallback kernel-doc

[Mark Rutland <mark.rutland@arm.com>]

On Thu, May 11, 2023 at 09:38:56PM +0200, Peter Zijlstra wrote:
> On Thu, May 11, 2023 at 06:10:00PM +0100, Mark Rutland wrote:
> > Hi Paul
> > 
> > On Wed, May 10, 2023 at 11:17:16AM -0700, Paul E. McKenney wrote:
> > > The gen-atomics.sh script currently generates 42 duplicate definitions:
> > > 
> > > 	arch_atomic64_add_negative
> > > 	arch_atomic64_add_negative_acquire
> > > 	arch_atomic64_add_negative_release
> > 
> > 	[...]
> > 
> > > These duplicates are presumably to handle different architectures
> > > generating hand-coded definitions for different subsets of the atomic
> > > operations.
> > 
> > Yup, for each FULL/ACQUIRE/RELEASE/RELAXED variant of each op, we allow the
> > archtiecture to choose between:
> > 
> > * Providing the ordering variant directly
> > * Providing the FULL ordering variant only
> > * Providing the RELAXED ordering variant only
> > * Providing an equivalent op that we can build from
> > 
> > > However, generating duplicate kernel-doc headers is undesirable.
> > 
> > Understood -- I hadn't understood that duplication was a problem when this was
> > originally written.
> > 
> > The way this is currently done is largely an artifact of our ifdeffery (and the
> > kerneldoc for fallbacks living inthe fallback templates), and I think we can
> > fix both of those.
> > 
> > > Therefore, generate only the first kernel-doc definition in a group
> > > of duplicates.  A comment indicates the name of the function and the
> > > fallback script that generated it.
> > 
> > I'm not keen on this approach, especially with the chkdup.sh script -- it feels
> > like we're working around an underlying structural issue.
> > 
> > I think that we can restructure the ifdeffery so that each ordering variant
> > gets its own ifdeffery, and then we could place the kerneldoc immediately above
> > that, e.g.
> > 
> > 	/**
> > 	 * arch_atomic_inc_return_release()
> > 	 *
> > 	 * [ full kerneldoc block here ]
> > 	 */
> > 	#if defined(arch_atomic_inc_return_release)
> > 	/* defined in arch code */
> > 	#elif defined(arch_atomic_inc_return_relaxed)
> > 	[ define in terms of arch_atomic_inc_return_relaxed ]
> > 	#elif defined(arch_atomic_inc_return)
> > 	[ define in terms of arch_atomic_inc_return ]
> > 	#else
> > 	[ define in terms of arch_atomic_fetch_inc_release ]
> > 	#endif
> > 
> > ... with similar for the mandatory ops that each arch must provide, e.g.
> > 
> > 	/**
> > 	 * arch_atomic_or()
> > 	 *
> > 	 * [ full kerneldoc block here ]
> > 	 */
> > 	/* arch_atomic_or() is mandatory -- architectures must define it! */
> > 
> > I had a go at that restructuring today, and while local build testing indicates
> > I haven't got it quite right, I think it's possible:
> > 
> >   https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=atomics/fallback-rework
> > 
> > Does that sound ok to you?
> 
> If the end result is simpler scripts, sure.

FWIW, regardless of the comments, I'd like to make this restructuring as it
makes it easier to add some more fallback cases, and I find the generated
ifdeffery a bit easier to follow when it's a chain of of-elif-elif-else-end
rather than a few nested cases.

> I'm not at all keen to complicate the scripts for something daft like
> kernel-doc. The last thing we need is documentation style weenies making
> an unholy mess of things.

Sure. I agree we don't want to bend over backwards for it at the cost of
maintainability, but I think it can be made pretty simple and self-contained,
and hopefully we can prove that with a v2 or v3. ;)

If nothing else, handling this centrally means that we'll have *one* set of
comments for this rather than a tonne of randomly managed copies in arch
code, which seems like a win...

Thanks,
Mark.