On 7/13/20 3:18 PM, Pavel Begunkov wrote:
> 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).
Right, except the io_read/io_write path it'll be 100% common unless you
have more than the inline number of segments.
I see the same thing for eg the slab should_failslab() call, which isn't
inlined even if the kconfig isn't enabled. And for should_fail_bio()
as well. Those two add up to another ~1% or so of pointless overhead.
> 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;
> }
Must be, since the kfree() one adds up to more than the two above ones
that are just the call itself.
--
Jens Axboe
