Re: [PATCH] force inlining of spinlock ops

From: Denys Vlasenko
Date: Tue May 12 2015 - 05:45:43 EST


On 05/12/2015 12:19 AM, Andrew Morton wrote:
> On Mon, 11 May 2015 19:57:22 +0200 Denys Vlasenko <dvlasenk@xxxxxxxxxx> wrote:
>
>> With both gcc 4.7.2 and 4.9.2, sometimes gcc mysteriously doesn't inline
>> very small functions we expect to be inlined. In particular,
>> with this config: http://busybox.net/~vda/kernel_config
>> there are more than a thousand copies of tiny spinlock-related functions:
>>
>> $ nm --size-sort vmlinux | grep -iF ' t ' | uniq -c | grep -v '^ *1 ' | sort -rn | grep ' spin'
>> 473 000000000000000b t spin_unlock_irqrestore
>> 292 000000000000000b t spin_unlock
>> 215 000000000000000b t spin_lock
>> 134 000000000000000b t spin_unlock_irq
>> 130 000000000000000b t spin_unlock_bh
>> 120 000000000000000b t spin_lock_irq
>> 106 000000000000000b t spin_lock_bh
>>
>> Disassembly:
>>
>> ffffffff81004720 <spin_lock>:
>> ffffffff81004720: 55 push %rbp
>> ffffffff81004721: 48 89 e5 mov %rsp,%rbp
>> ffffffff81004724: e8 f8 4e e2 02 callq <_raw_spin_lock>
>> ffffffff81004729: 5d pop %rbp
>> ffffffff8100472a: c3 retq
>>
>> This patch fixes this via s/inline/__always_inline/ in spinlock.h.
>> This decreases vmlinux by about 30k:
>>
>> text data bss dec hex filename
>> 82375570 22255544 20627456 125258570 7774b4a vmlinux.before
>> 82335059 22255416 20627456 125217931 776ac8b vmlinux
>
> See also https://lkml.org/lkml/2015/4/23/598 ("enforce function
> inlining for hot functions").
>
> Presumably Hagen didn't see the issue with spinlock functions. I
> wonder why not.
>
> I suppose we should get both these consolidated into a coherent whole.
>
> It's a bit irritating to have to do this: presumably gcc will get fixed
> and the huge sprinkling of __always_inline will become less and less
> relevant over time and people will have trouble distinguishing "real
> __always_inline which was put here for a purpose" from "dopey
> __always_inline to work around a short-term gcc glitch".

In my patches, I put __always_inline *only* on functions
where my measurements show a large size decrease from doing so.
*Not* on functions where "I think it may be a good idea".

So far, all such functions were so trivial that inlining decision there
is a no-brainer.

> and then use inline_for_broken_gcc everywhere. That way, the reason
> for the marker is self-explanatory and we can later hunt all these
> things down and remvoe them.
>
> Also, the inline_for_broken_gcc definition can be made dependent on
> particular gcc versions, which will allow us to easily keep an eye on
> the behaviour of later gcc versions.

I've seen it on gcc-4.7.2 and gcc-4.9.2, so this behavior is not
limited to a narrow range of gcc versions. I'd say by now about half
of running kernels can easily be affected.

--
vda

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/