Re: perf record doesn't work on rtd129x SoC

[Robin Murphy <robin.murphy@arm.com>]

On 04/12/2019 11:20 am, Robin Murphy wrote:
> On 2019-12-04 7:28 am, Andreas Färber wrote:
>> Hi YanQing,
>>
>> + LAKML + Mark + Will
>>
>> Am 04.12.19 um 05:55 schrieb Wang YanQing:
>>> I use "perf record" to debug performance issue on RTD1296 SOC, it 
>>> does't work, but
>>> the "perf stat" is ok!
>>
>> Thanks for the report - which board, branch and (base) tag are you
>> testing against? And are you building perf yourself from kernel sources,
>> or are you using some distro package?
>>
>> I only have Busybox in my initrd on DS418; I have not tested perf.
>>
>>> After some dig in the kernel, I find the reason is no pmu overflow 
>>> interrupt, I think
>>> below pmu configuration isn't right for RTD1296:
>>> "
>>>          arm_pmu: arm-pmu {
>>>                  compatible = "arm,cortex-a53-pmu";
>>>                  interrupts = <GIC_SPI 48 IRQ_TYPE_LEVEL_HIGH>;
>>>          };
>>> "
>>>
>>> We need 4 PMU SPI for RTD1296 (4 cores), and I guess the 48 isn't 
>>> right too.
>>
>> Note that above rtd129x.dtsi snippet is not complete. See rtd1296.dtsi:
>>
>> &arm_pmu {
>>     interrupt-affinity = <&cpu0>, <&cpu1>, <&cpu2>, <&cpu3>;
>> };
> 
> That doesn't help much, since 4 affinities for one SPI is rather 
> nonsensical.
> 
>> 48 and high/4 match what I see in the latest BSP:
>>
>> https://github.com/BPI-SINOVOIP/BPI-M4-bsp/blob/master/linux-rtk/arch/arm64/boot/dts/realtek/rtd129x/rtd-1296.dtsi#L116 
>>
>>
>>> Any suggestion is welcome.
>>>
>>> Thanks!
>>
>> The only difference I see is "arm,cortex-a53-pmu" vs. "arm,armv8-pmuv3".
>> By my reading of arch/arm64/kernel/perf_event.c the only difference
>> between the two should be the name and an extra cache_map. You could try
>> the other compatible string in your .dts, but I doubt it'll help.
>>
>> Hopefully the Realtek or Arm guys can shed some light.
> 
> If the SoC really has all 4 overflow interrupts combined into a single 
> SPI line, then sampling just isn't going to be supported - it's 
> unreasonably difficult to handle overflow when the IRQ may be taken on 
> the wrong CPU.

On closer inspection, that BSP kernel implements a whole hrtimer-based 
bodge in arm_pmu to apparently work around not having usable interrupts, 
so yeah, this isn't going to be usable, sorry.

Robin.