Re: [PATCH v4 2/2] perf/core: Fix incorrect time diff in tick adjust period

["Liang, Kan" <kan.liang@linux.intel.com>]

On 2024-08-27 9:10 p.m., Adrian Hunter wrote:
> On 27/08/24 23:06, Liang, Kan wrote:
>>
>>
>> On 2024-08-27 1:16 p.m., Adrian Hunter wrote:
>>> On 27/08/24 19:42, Liang, Kan wrote:
>>>>
>>>>
>>>> On 2024-08-21 9:42 a.m., Luo Gengkun wrote:
>>>>> Perf events has the notion of sampling frequency which is implemented in
>>>>> software by dynamically adjusting the counter period so that samples occur
>>>>> at approximately the target frequency.  Period adjustment is done in 2
>>>>> places:
>>>>>  - when the counter overflows (and a sample is recorded)
>>>>>  - each timer tick, when the event is active
>>>>> The later case is slightly flawed because it assumes that the time since
>>>>> the last timer-tick period adjustment is 1 tick, whereas the event may not
>>>>> have been active (e.g. for a task that is sleeping).
>>>>>
>>>>
>>>> Do you have a real-world example to demonstrate how bad it is if the
>>>> algorithm doesn't take sleep into account?
>>>>
>>>> I'm not sure if introducing such complexity in the critical path is
>>>> worth it.
>>>>
>>>>> Fix by using jiffies to determine the elapsed time in that case.
>>>>>
>>>>> Signed-off-by: Luo Gengkun <luogengkun@huaweicloud.com>
>>>>> ---
>>>>>  include/linux/perf_event.h |  1 +
>>>>>  kernel/events/core.c       | 11 ++++++++---
>>>>>  2 files changed, 9 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
>>>>> index 1a8942277dda..d29b7cf971a1 100644
>>>>> --- a/include/linux/perf_event.h
>>>>> +++ b/include/linux/perf_event.h
>>>>> @@ -265,6 +265,7 @@ struct hw_perf_event {
>>>>>  	 * State for freq target events, see __perf_event_overflow() and
>>>>>  	 * perf_adjust_freq_unthr_context().
>>>>>  	 */
>>>>> +	u64				freq_tick_stamp;
>>>>>  	u64				freq_time_stamp;
>>>>>  	u64				freq_count_stamp;
>>>>>  #endif
>>>>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>>>>> index a9395bbfd4aa..86e80e3ef6ac 100644
>>>>> --- a/kernel/events/core.c
>>>>> +++ b/kernel/events/core.c
>>>>> @@ -55,6 +55,7 @@
>>>>>  #include <linux/pgtable.h>
>>>>>  #include <linux/buildid.h>
>>>>>  #include <linux/task_work.h>
>>>>> +#include <linux/jiffies.h>
>>>>>  
>>>>>  #include "internal.h"
>>>>>  
>>>>> @@ -4120,7 +4121,7 @@ static void perf_adjust_freq_unthr_events(struct list_head *event_list)
>>>>>  {
>>>>>  	struct perf_event *event;
>>>>>  	struct hw_perf_event *hwc;
>>>>> -	u64 now, period = TICK_NSEC;
>>>>> +	u64 now, period, tick_stamp;
>>>>>  	s64 delta;
>>>>>  
>>>>>  	list_for_each_entry(event, event_list, active_list) {
>>>>> @@ -4148,6 +4149,10 @@ static void perf_adjust_freq_unthr_events(struct list_head *event_list)
>>>>>  		 */
>>>>>  		event->pmu->stop(event, PERF_EF_UPDATE);
>>>>>  
>>>>> +		tick_stamp = jiffies64_to_nsecs(get_jiffies_64());
>>>>
>>>> Seems it only needs to retrieve the time once at the beginning, not for
>>>> each event.
>>>>
>>>> There is a perf_clock(). It's better to use it for the consistency.
>>>
>>> perf_clock() is much slower, and for statistical sampling it doesn't
>>> have to be perfect.
>>
>> Because of rdtsc?
> 
> Yes

OK. I'm not worry about it too much as long as it's only invoked once in
each tick.

> 
>>
>> If it is only used here, it should be fine. What I'm worried about is
>> that someone may use it with other timestamp in perf later. Anyway, it's
>> not a big deal.
>>
>> The main concern I have is that do we really need the patch?
> 
> The current code is wrong.
> 
>> It seems can only bring us a better guess of the period for the sleep
>> test. Then we have to do all the calculate for each tick.
> 
> Or any workload that sleeps periodically.
> 
> Another option is to remove the period adjust on tick entirely.
> Although arguably the calculation at a tick is better because
> it probably covers a longer period.

Or we may remove the period adjust on overflow.

As my understanding, the period adjust on overflow is to handle the case
while the overflow happens very frequently (< 2 ticks). It is mainly
caused by the very low start period (1).
I'm working on a patch to set a larger start period, which should
minimize the usage of the period adjust on overflow.

Anyway, based on the current code, I agree that adding a new
freq_tick_stamp should be required. But it doesn't need to read the time
for each event. I think reading the time once at the beginning should be
good enough for the period adjust/estimate algorithm.

Thanks,
Kan

> 
>>
>> Thanks,
>> Kan
>>>
>>>>
>>>> Thanks,
>>>> Kan
>>>>> +		period = tick_stamp - hwc->freq_tick_stamp;
>>>>> +		hwc->freq_tick_stamp = tick_stamp;
>>>>> +
>>>>>  		now = local64_read(&event->count);
>>>>>  		delta = now - hwc->freq_count_stamp;
>>>>>  		hwc->freq_count_stamp = now;
>>>>> @@ -4157,9 +4162,9 @@ static void perf_adjust_freq_unthr_events(struct list_head *event_list)
>>>>>  		 * reload only if value has changed
>>>>>  		 * we have stopped the event so tell that
>>>>>  		 * to perf_adjust_period() to avoid stopping it
>>>>> -		 * twice.
>>>>> +		 * twice. And skip if it is the first tick adjust period.
>>>>>  		 */
>>>>> -		if (delta > 0)
>>>>> +		if (delta > 0 && likely(period != tick_stamp))
>>>>>  			perf_adjust_period(event, period, delta, false);>
>>>>>  		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
>>>
>>>
> 
>