Re: [PATCH] io_uring: kill io_cqring_ev_posted() and __io_cq_unlock_post()

[Pavel Begunkov <asml.silence@gmail.com>]

On 11/24/22 19:29, Jens Axboe wrote:
> On 11/24/22 12:17?PM, Pavel Begunkov wrote:
>> On 11/24/22 18:46, Jens Axboe wrote:
>>> On 11/24/22 9:16?AM, Pavel Begunkov wrote:
>>>> On 11/21/22 14:52, Jens Axboe wrote:
>>>>> __io_cq_unlock_post() is identical to io_cq_unlock_post(), and
>>>>> io_cqring_ev_posted() has a single caller so migth as well just inline
>>>>> it there.
>>>>
>>>> It was there for one purpose, to inline it in the hottest path,
>>>> i.e. __io_submit_flush_completions(). I'll be reverting it back
>>>
>>> The compiler is most certainly already doing that, in fact even
>>
>> .L1493:
>> # io_uring/io_uring.c:631:???? io_cq_unlock_post(ctx);
>> ????movq??? %r15, %rdi??? # ctx,
>> ????call??? io_cq_unlock_post??? #
> 
> Doubled checked here, and you're actually right:
> 
>      55bc:       94000000        bl      4760 <io_cq_unlock_post>
> 
> Huh, that's very odd that it doesn't inline it. It doesn't even it I
> mark it inline, __always_inline gets it done.

That's odd as well for a function of this size

>> Even more, after IORING_SETUP_CQE32 was added I didn't see
>> once __io_fill_cqe_req actually inlined even though it's marked
>> so.
> 
> Doesn't seem to be inlined here either. Compiler:
> 
> gcc (Debian 12.2.0-9) 12.2.0
> 
>>> __io_submit_flush_completions() is inlined in
>>> io_submit_flush_completions() for me here.
>>
>> And io_submit_flush_completions is inlined as well, right?
>> That would be quite odd, __io_submit_flush_completions() is not
>> small by any means and there are 3 call sites.
> 
> io_submit_flush_completions() doesn't get inlined,
> __io_submit_flush_completions() gets inlined in
> io_submit_flush_completions().

Then the compiler is drunk. It doesn't inline the function
explicitly marked inline but does it for a non-inline one.
Unless it's PGO'ed I can't think of a sane reason for it.

-- 
Pavel Begunkov