[PATCH v6 0/4] Add perf interface to expose nvdimm

[Nageswara Sastry]

On 25/02/22 12:08 pm, kajoljain wrote:
> 
> 
> On 2/25/22 11:25, Nageswara Sastry wrote:
>>
>>
>> On 17/02/22 10:03 pm, Kajol Jain wrote:
>>> Patchset adds performance stats reporting support for nvdimm.
>>> Added interface includes support for pmu register/unregister
>>> functions. A structure is added called nvdimm_pmu to be used for
>>> adding arch/platform specific data such as cpumask, nvdimm device
>>> pointer and pmu event functions like event_init/add/read/del.
>>> User could use the standard perf tool to access perf events
>>> exposed via pmu.
>>>
>>> Interface also defines supported event list, config fields for the
>>> event attributes and their corresponding bit values which are exported
>>> via sysfs. Patch 3 exposes IBM pseries platform nmem* device
>>> performance stats using this interface.
>>>
>>> Result from power9 pseries lpar with 2 nvdimm device:
>>>
>>> Ex: List all event by perf list
>>>
>>> command:# perf list nmem
>>>
>>>     nmem0/cache_rh_cnt/                                [Kernel PMU event]
>>>     nmem0/cache_wh_cnt/                                [Kernel PMU event]
>>>     nmem0/cri_res_util/                                [Kernel PMU event]
>>>     nmem0/ctl_res_cnt/                                 [Kernel PMU event]
>>>     nmem0/ctl_res_tm/                                  [Kernel PMU event]
>>>     nmem0/fast_w_cnt/                                  [Kernel PMU event]
>>>     nmem0/host_l_cnt/                                  [Kernel PMU event]
>>>     nmem0/host_l_dur/                                  [Kernel PMU event]
>>>     nmem0/host_s_cnt/                                  [Kernel PMU event]
>>>     nmem0/host_s_dur/                                  [Kernel PMU event]
>>>     nmem0/med_r_cnt/                                   [Kernel PMU event]
>>>     nmem0/med_r_dur/                                   [Kernel PMU event]
>>>     nmem0/med_w_cnt/                                   [Kernel PMU event]
>>>     nmem0/med_w_dur/                                   [Kernel PMU event]
>>>     nmem0/mem_life/                                    [Kernel PMU event]
>>>     nmem0/poweron_secs/                                [Kernel PMU event]
>>>     ...
>>>     nmem1/mem_life/                                    [Kernel PMU event]
>>>     nmem1/poweron_secs/                                [Kernel PMU event]
>>>
>>> Patch1:
>>>           Introduces the nvdimm_pmu structure
>>> Patch2:
>>>           Adds common interface to add arch/platform specific data
>>>           includes nvdimm device pointer, pmu data along with
>>>           pmu event functions. It also defines supported event list
>>>           and adds attribute groups for format, events and cpumask.
>>>           It also adds code for cpu hotplug support.
>>> Patch3:
>>>           Add code in arch/powerpc/platform/pseries/papr_scm.c to expose
>>>           nmem* pmu. It fills in the nvdimm_pmu structure with pmu name,
>>>           capabilities, cpumask and event functions and then registers
>>>           the pmu by adding callbacks to register_nvdimm_pmu.
>>> Patch4:
>>>           Sysfs documentation patch
>>>
>>> Changelog
>>
>> Tested these patches with the automated tests at
>> avocado-misc-tests/perf/perf_nmem.py
>> URL:
>> https://github.com/avocado-framework-tests/avocado-misc-tests/blob/master/perf/perf_nmem.py
>>
>>
>> 1. On the system where target id and online id were different then not
>> seeing value in 'cpumask' and those tests failed.
>>
>> Example:
>> Log from dmesg
>> ...
>> papr_scm ibm,persistent-memory:ibm,pmemory at 44100003: Region registered
>> with target node 1 and online node 0
>> ...
> 
> Hi Nageswara Sastry,
>         Thanks for testing the patch set. Yes you right, incase target
> node id and online node id is different, it can happen when target
> node is not online and hence can cause this issue, thanks for pointing
> it.
> 
> Function dev_to_node will return node id for a given nvdimm device which
> can be offline in some scenarios. We should use numa node id return by
> numa_map_to_online_node function in that scenario. This function incase
> given node is offline, it will lookup for next closest online node and
> return that nodeid.
> 
> Can you try with below change and see, if you are still getting this
> issue. Please let me know.
> 
> diff --git a/arch/powerpc/platforms/pseries/papr_scm.c
> b/arch/powerpc/platforms/pseries/papr_scm.c
> index bdf2620db461..4dd513d7c029 100644
> --- a/arch/powerpc/platforms/pseries/papr_scm.c
> +++ b/arch/powerpc/platforms/pseries/papr_scm.c
> @@ -536,7 +536,7 @@ static void papr_scm_pmu_register(struct
> papr_scm_priv *p)
>                                  PERF_PMU_CAP_NO_EXCLUDE;
> 
>          /*updating the cpumask variable */
> -       nodeid = dev_to_node(&p->pdev->dev);
> +       nodeid = numa_map_to_online_node(dev_to_node(&p->pdev->dev));
>          nd_pmu->arch_cpumask = *cpumask_of_node(nodeid);
> 
> Thanks,
> Kajol Jain
> 

With the above patch all the tests are passing on the system where 
target id and online id were different. Here is the the result:

(1/9) perf_nmem.py:perfNMEM.test_pmu_register_dmesg: PASS (3.47 s)
(2/9) perf_nmem.py:perfNMEM.test_sysfs: PASS (1.15 s)
(3/9) perf_nmem.py:perfNMEM.test_pmu_count: PASS (1.08 s)
(4/9) perf_nmem.py:perfNMEM.test_all_events: PASS (18.15 s)
(5/9) perf_nmem.py:perfNMEM.test_all_group_events: PASS (2.22 s)
(6/9) perf_nmem.py:perfNMEM.test_mixed_events: CANCEL: With single PMU 
mixed events test is not possible. (1.18 s)
(7/9) perf_nmem.py:perfNMEM.test_pmu_cpumask: PASS (1.12 s)
(8/9) perf_nmem.py:perfNMEM.test_cpumask: PASS (1.17 s)
(9/9) perf_nmem.py:perfNMEM.test_cpumask_cpu_off: PASS (1.81 s)

Tested-by: Nageswara R Sastry <rnsastry at linux.ibm.com>

-- 
Thanks and Regards
R.Nageswara Sastry