Re: [PATCH 5/5] locking/atomic: add generic arch_*() bitops

From: Marco Elver <elver@google.com>
To: Mark Rutland <mark.rutland@arm.com>
Cc: linux-kernel@vger.kernel.org, Boqun Feng <boqun.feng@gmail.com>,
	Daniel Axtens <dja@axtens.net>, Ingo Molnar <mingo@kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Will Deacon <will@kernel.org>
Subject: Re: [PATCH 5/5] locking/atomic: add generic arch_*() bitops
Date: Fri, 16 Jul 2021 15:02:21 +0200	[thread overview]
Message-ID: <CANpmjNNVn4UxBCMg1ke9xaQNv52OMuQjr17GxUXojZKwiAFzzg@mail.gmail.com> (raw)
In-Reply-To: <20210716122114.GB78984@C02TD0UTHF1T.local>

On Fri, 16 Jul 2021 at 14:21, Mark Rutland <mark.rutland@arm.com> wrote:
[...]
> > Why not just:
> >
> >       #define test_bit arch_test_bit
> >
> > and similar for the ones that use plain accesses?
> >
> > And not include instrumented-non-atomic.h here nor do the
> > __is_defined(*_uses_plain_access) change to the instrumented header,
> > which seems to overcomplicate things as it effectively just aliases the
> > non-arch_ name to the arch_ name if *_uses_plain_access is defined.
>
> I'd done that to still permit the compiler to out-of-line the non-arch
> forms if it wanted to. That said, I see that for the atomics we forced
> those to be __always_inline anyway, so maybe that's not a concern.
>
> I'm happy either way.

I'd prefer simplicity. In an optimized build, I think even if it
decided to not inline, the perf and code size difference is a wash.
Intuition tells me that inlining would even be preferred either way,
but I've been wrong about that in the past.

I think originally we turned the atomics __always_inline because of
uaccess regions. Because debugging tools do work in uaccess regions,
arch_* aren't necessarily appropriate. So something like a test_bit()
in an uaccess region might actually generate a warning. This might be
a valid argument for __always_inline irrespective of optimization
potential (if there is any).

Thanks,
-- Marco