Re: [PATCH 3/3] io_uring: follow **iovec idiom in io_import_iovec

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

On 14/07/2020 00:16, Jens Axboe wrote:
> On 7/13/20 3:12 PM, Pavel Begunkov wrote:
>> On 14/07/2020 00:09, Jens Axboe wrote:
>>> On 7/13/20 1:59 PM, Pavel Begunkov wrote:
>>>> @@ -3040,8 +3040,7 @@ static int io_read(struct io_kiocb *req, bool force_nonblock,
>>>>  		}
>>>>  	}
>>>>  out_free:
>>>> -	if (!(req->flags & REQ_F_NEED_CLEANUP))
>>>> -		kfree(iovec);
>>>> +	kfree(iovec);
>>>>  	return ret;
>>>>  }
>>>
>>> Faster to do:
>>>
>>> if (iovec)
>>> 	kfree(iovec)
>>>
>>> to avoid a stupid call. Kind of crazy, but I just verified with this one
>>> as well that it's worth about 1.3% CPU in my stress test.
>>
>> That looks crazy indeed
> 
> I suspect what needs to happen is that kfree should be something ala:
> 
> static inline void kfree(void *ptr)
> {
> 	if (ptr)
> 		__kfree(ptr);
> }
> 
> to avoid silly games like this. Needs to touch all three slab
> allocators, though definitely in the trivial category.

Just thought the same, but not sure it's too common to have kfree(NULL).

The drop is probably because of extra call + cold jumps with unlikely().

void kfree() {
	trace_kfree(_RET_IP_, objp);

	if (unlikely(ZERO_OR_NULL_PTR(objp)))
		return;
}

-- 
Pavel Begunkov