Re: [PATCH v4 12/16] locking/rwsem: Enable time-based spinning on reader-owned rwsem

[Waiman Long <longman@redhat.com>]

On 04/19/2019 10:33 AM, Waiman Long wrote:
> On 04/19/2019 03:56 AM, Peter Zijlstra wrote:
>> On Thu, Apr 18, 2019 at 11:15:33AM -0400, Waiman Long wrote:
>>> On 04/18/2019 09:06 AM, Peter Zijlstra wrote:
>>>>> +			/*
>>>>> +			 * Check time threshold every 16 iterations to
>>>>> +			 * avoid calling sched_clock() too frequently.
>>>>> +			 * This will make the actual spinning time a
>>>>> +			 * bit more than that specified in the threshold.
>>>>> +			 */
>>>>> +			else if (!(++loop & 0xf) &&
>>>>> +				 (sched_clock() > rspin_threshold)) {
>>>> Why is calling sched_clock() lots a problem?
>>> Actually I am more concern about the latency introduced by the
>>> sched_clock() call. BTW, I haven't done any measurement myself. Do you
>>> know how much cost the sched_clock() call is?
>>>
>>> If the cost is relatively high, the average latency period after the
>>> lock is free and the spinner is ready to do a trylock will increase.
>> Totally depends on the arch or course :/ For 'sane' x86 it is: RDTSC,
>> MUL; SHRD; SHR; ADD, which is plenty fast.
>>
>> I know we have poll loops with sched_clock/local_clock in them, I just
>> can't seem to find any atm.
> Thanks, I will do some time measurement myself. If it is fast enough, I
> can change the code to do sched_clock on every iteration.
>
> Cheers,
> Longman
>
I had measured the time of doing 10 sched_clock() calls. On a 2.1GHz
skylake system, it was 83ns (~18 cycles per call). On a 2.5GHz Thunder
X2 arm system, it was 860ns (~215 cycles per call). On a 2.2GHz AMD EPYC
system, it was 200ns (~44 cycles per call). Intel is fastest, followed
by AMD and then the ARM64 chip.

Cheers,
Longman