Re: [PATCH v2 1/2] padata: Mark padata_work_init() as __ref

[Nathan Chancellor <nathan@kernel.org>]

On Mon, Dec 12, 2022 at 01:06:16PM -0700, Nathan Chancellor wrote:
> On Mon, Dec 12, 2022 at 02:21:57PM -0500, Daniel Jordan wrote:
> > On Mon, Dec 12, 2022 at 10:05:02AM -0700, Nathan Chancellor wrote:
> > > On Mon, Dec 12, 2022 at 10:07:24PM +0900, Masahiro Yamada wrote:
> > > > I am not sure if the compiler should do this level of optimization
> > > > because kernel/padata.c does not seem to be a special case.
> > > > Perhaps, we might be hit with more cases that need __ref annotation,
> > > > which is only required by LTO.
> > > 
> > > That's possible. I did only see this once instance in all my builds but
> > > allmodconfig + ThinLTO might not be too interesting of a case,
> > > since the sanitizers will be enabled, which makes optimization more
> > > difficult. I could try to enable ThinLTO with some distribution
> > > configurations to see if there are any more instances that crop up.
> > 
> > Yes, if there were many more instances of this problem it might be worth
> > thinking about an LTO-specific solution to fix it closer to the source.
> 
> Ack, I will wire up some build tests to see if this optimization occurs
> frequently enough to warrant a wider fix.

Turns out this does not appear to happen often. I built several
distribution configurations for arm64 and x86_64 with
CONFIG_LTO_CLANG_THIN=y and saw no modpost warnings. So I think this is
sufficiently odd to keep the fix local to this one instance. I will send
a v3 later today.

> > > > One note is that, we could discard padata_work_init()
> > > > because (1) and (3) are both annotated as __init.
> > > > So, another way of fixing is
> > > >    static __always_inline void padata_work_init(...)
> > > > because the compiler would determine padata_work_init()
> > > > would be small enough if the caller and callee belonged to
> > > > the same section.
> > > > 
> > > > I do not have a strong opinion.
> > 
> > I'm right there with you.  :-)
> > 
> > > > Honestly, I do not know what the best approach would be to fix this.
> > 
> > Either approach works, either can include an explanatory comment.
> > __ref seems more targeted to the problem at hand.
> 
> Right, I suspect __ref is the right way to go for this particular issue.
> I will add a comment regardless.
> 
> > > > If we go with the __ref annotation, I can pick this, but
> > > > at least can you add some comments?
> > > > 
> > > > 
> > > > include/linux/init.h says:
> > > > "optimally document why the __ref is needed and why it's OK"
> > > > 
> > > > 
> > > > I think this is the case that needs some comments
> > > > because LTO optimization looks too tricky to me.
> > > 
> > > Sure thing, I will send a v3 either Tuesday or Wednesday with an updated
> > > commit message and code comment if we end up going this route.
> > 
> > A nitpick, but as long as you're respinning, if we stay with this
> > approach, could you put __ref just before the function name?  init.h
> > says "The markers follow same syntax rules as __init / __initdata" and
> > for those it says "You should add __init immediately before the function
> > name" though there are plenty of places in the tree that don't do this.
> 
> Sure thing!
> 
> Cheers,
> Nathan
>