[PATCH V1 3/4] perf vendor events: Add JSON metrics for imx8mq DDR Perf

[John Garry]

On 20/01/2021 05:46, Joakim Zhang wrote:
>> -----Original Message-----
>> From: John Garry<john.garry at huawei.com>
>> Sent: 2021年1月19日 20:27
>> To: Joakim Zhang<qiangqing.zhang at nxp.com>;acme at kernel.org;
>> jolsa at redhat.com
>> Cc:will at kernel.org;linux-perf-users at vger.kernel.org; dl-linux-imx
>> <linux-imx at nxp.com>;mark.rutland at arm.com;
>> linux-arm-kernel at lists.infradead.org
>> Subject: Re: [PATCH V1 3/4] perf vendor events: Add JSON metrics for imx8mq
>> DDR Perf
>>
>> On 19/01/2021 11:00, Joakim Zhang wrote:
>>>> -----Original Message-----
>>>> From: John Garry<john.garry at huawei.com>
>>>> Sent: 2021年1月19日 18:40
>>>> To: Joakim Zhang<qiangqing.zhang at nxp.com>;acme at kernel.org;
>>>> jolsa at redhat.com
>>>> Cc:will at kernel.org;linux-perf-users at vger.kernel.org; dl-linux-imx
>>>> <linux-imx at nxp.com>;mark.rutland at arm.com;
>>>> linux-arm-kernel at lists.infradead.org
>>>> Subject: Re: [PATCH V1 3/4] perf vendor events: Add JSON metrics for
>>>> imx8mq DDR Perf
>>>>
>>>> On 19/01/2021 02:56, Joakim Zhang wrote:
>>>>> index 000000000000..c3e9f89a4a4d
>>>>> --- /dev/null
>>>>> +++ b/tools/perf/pmu-events/arch/arm64/freescale/imx8mq/sys/metrics.
>>>>> +++ js
>>>>> +++ on
>>>>> @@ -0,0 +1,26 @@
>>>>> +[
>>>>> +   {
>>>>> +	    "BriefDescription": "bandwidth usage for lpddr4 evk board",
>>>> Why mention the evk board specifically? Is this metric only specific
>>>> to that board, and not all i.MX8MQ platforms?
>>> Yes, John, we have lpddr4 evk board and ddr4 evk borad......
>>>
>> How is this supposed to work?
>>
>> So perf matches on the identifier file for system PMUs; however the contents of
>> this file are based on the HW implementation for imx DDR PMU, not the board.
> Yes, I know this feature is for system PMU, not board. You know DDR bandwidth calculation is based on DDR clock frequency, which is vary form boards.
> 
> e.g.
> lpddr4 board, DDR clock frequency is 750MHZ:
> "MetricExpr": "(( imx8_ddr0 at read\\-cycles@ + imx8_ddr0 at write\\-cycles@ ) * 4 * 4 / duration_time) / (750 * 1000000 * 4 * 4)"
> 
> ddr4 board, DDR clock frequency is 600MHZ.
> "MetricExpr": "(( imx8_ddr0 at read\\-cycles@ + imx8_ddr0 at write\\-cycles@ ) * 4 * 4 / duration_time) / (600 * 1000000 * 4 * 4)"
> 
> I give different metric names,
> "MetricName": "imx8mm_bandwidth_usage.lpddr4"
> "MetricName": "imx8mm_bandwidth_usage.ddr4"
> 

But that is just not how things are done. We - or I - would expect that 
perf tool would have the intelligence to know the appropriate metric for 
that specific platform. Or, more specifically, we should not present the 
user with many metrics and expect him/her to know the appropriate metric 
for the specific board.

Are there other cases for other architectures where things are done like 
this?

Thanks,
John

> So users can select correct bandwidth metric according to their own board types. To a certain extent, it can also support some board-level metrics.
> 
> If it is okay for you, I will send a V2 to replace ( imx8_ddr0 at read\\-cycles@ + imx8_ddr0 at write\\-cycles@ ) with ( imx8mm_ddr.read_cycles + imx8mm_ddr.write_cycles ).
> 
>>> John, change to your new framework, I found it can't support all features
>> which is supported in 5.10 kernel.
>>
>> Yes, so let me know about all the problems.
> It should have worked as 5.10 kernel at my side. Thanks.