Re: [PATCH 10/10] arm64: mm: Re-implement the __flush_tlb_range_op macro in C

[Will Deacon <will@kernel.org>]

On Fri, Jul 11, 2025 at 11:16:32AM -0700, Linus Torvalds wrote:
> On Fri, 11 Jul 2025 at 09:18, Will Deacon <will@kernel.org> wrote:
> >
> > The __flush_tlb_range_op() macro is horrible and has been a previous
> > source of bugs thanks to multiple expansions of its arguments (see
> > commit f7edb07ad7c6 ("Fix mmu notifiers for range-based invalidates")).
> >
> > Rewrite the thing in C.
> 
> So I do think this is better than the old case, but I think you could
> go one step further...
> 
> > +static __always_inline void __flush_tlb_range_op(const enum tlbi_op op,
> > +                                                u64 start, size_t pages,
> > +                                                u64 stride, u16 asid,
> > +                                                u32 level, bool lpa2)
> 
> If you replaced that "enum tlbi_op op" with two function pointers
> instead (for "the invalidate range" and "invalidate one" cases
> respectively), I think you could make this code much more obvious.
> 
> And exactly like how you depend on that 'op' value being
> constant-folded because all the different levels are inline functions,
> the same thing ends up happening with function pointers where inlining
> will result in a constant function pointer becoming just a static call
> (and in turn inlined as well).

So I don't _strictly_ rely on the constant-folding and replacing that
BUILD_BUG_ON() with a BUG_ON() would still give functionally correct
code if inlining didn't occur. I just much preferred catching a wonky
TLBI op at compile-time, which is why I ended up with this but I hadn't
considered that this would allow us to inline indirect function calls.

> And then the actual *callers* would use the "look up op" thing, but I
> suspect that in many cases those could then be in turn also simplified
> to not use that op-number at all, but just end up using the op-name.

Right, I think we'd drop the enum entirely if we went down this route.

> I didn't try to actually create that series - and I think you'd want
> to do it in multiple steps just to make each individual step small and
> obvious - but I think it would end up looking nicer.

I'll have a play...

Will