Re: [PATCH] arm64: make atomic helpers __always_inline

[Will Deacon]

On Fri, Jan 08, 2021 at 11:26:53AM +0100, Arnd Bergmann wrote:
> On Fri, Jan 8, 2021 at 10:33 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > On Fri, Jan 08, 2021 at 10:19:56AM +0100, Arnd Bergmann wrote:
> > > From: Arnd Bergmann <arnd@xxxxxxxx>
> > >
> > > With UBSAN enabled and building with clang, there are occasionally
> > > warnings like
> > >
> > > WARNING: modpost: vmlinux.o(.text+0xc533ec): Section mismatch in reference from the function arch_atomic64_or() to the variable .init.data:numa_nodes_parsed
> > > The function arch_atomic64_or() references
> > > the variable __initdata numa_nodes_parsed.
> > > This is often because arch_atomic64_or lacks a __initdata
> > > annotation or the annotation of numa_nodes_parsed is wrong.
> > >
> > > for functions that end up not being inlined as intended but operating
> > > on __initdata variables. Mark these as __always_inline, along with
> > > the corresponding asm-generic wrappers.
> >
> > Hmm, I don't fully grok this. Why does it matter if a non '__init' function
> > is called with a pointer to some '__initdata'? Or is the reference coming
> > from somewhere else? (where?).
> 
> There are (at least) three ways for gcc to deal with a 'static inline'
> function:
> 
> a) fully inline it as the __always_inline attribute does
> b) not inline it at all, treating it as a regular static function
> c) create a specialized version with different calling conventions
> 
> In this case, clang goes with option c when it notices that all
> callers pass the same constant pointer. This means we have a
> synthetic
> 
> static noinline long arch_atomic64_or(long i)
> {
>         return __lse_ll_sc_body(atomic64_fetch_or, i, &numa_nodes_parsed);
> }
> 
> which is a few bytes shorter than option b as it saves a load in the
> caller. This function definition however violates the kernel's rules
> for section references, as the synthetic version is not marked __init.

Ah, I was hoping the compiler would've sorted that out, but then again, how
would it know? But doesn't this mean that whenever we get one caller passing
something like an __initdata pointer to a function, then that function needs
to be __always_inline for everybody? It feels like a slippery slope
considering the incentive to go back and replace it with 'inline' if the
caller goes away is very small.

Didn't we used to #define inline as __always_inline to avoid this situation?

Will