Message-ID: <b89f09fd-e9c4-9112-6a6a-16f9632ccbe3@arm.com>
Date:   Fri, 15 Nov 2019 11:37:29 +0000
From:   James Clark <James.Clark@....com>
To:     Tan Xiaojun <tanxiaojun@...wei.com>
CC:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-perf-users@...r.kernel.org" <linux-perf-users@...r.kernel.org>,
        nd <nd@....com>
Subject: Re: [RFC v2 4/4] perf tools: Support "branch-misses:pp" on arm64

Hi Xiaojun,

If the difference is not noticeable I think it would be better to leave it disabled. Presumably if the user
supplies the ":p" argument they are interested in the data being as precise as possible.

If they want to enable jitter, then can always configure the SPE event manually.

I have a question about what kind of approach you think we should take for multiple events that are provided with :p.
For example "perf record -e branch-misses:p -e cache-misses:p ...". In your current implementation this will
give the error "There may be only one SPE event". I think this is fine for a first implementation. But I wonder if there
is a way of supporting multiple SPE events?

From the documentation it seems like the filter events are ANDed together:

	PMSEVFR_EL1.
	Controls sample filtering by events. The overall filter is the logical AND of these filters. For example, if E[3] and E[5] are both set to 1,
	only samples that have both event 3 (Level 1 unified or data cache refill) and event 5 set (TLB walk) are recorded

Which means that if we kept adding filters for new event types, there would be no events received because they wouldn't satisfy the filter requirements
of being caused by a branch miss AND a cache miss for example. I have asked internally about whether this is a mistake or not.


Thanks
James

On 15/11/2019 02:59, Tan Xiaojun wrote:
> On 2019/11/13 22:47, James Clark wrote:
>> Hi Xiaojun,
>>
>>> I can't reproduce this problem. If the current system doesn't support spe, it shouldn't report an error. I use the latest codes of the mainline:
>>
>> I think the problem is related to the 'type' attribute of the event. To open the SPE PMU the event type on the platform I'm using is '7'. If I change
>> the code like this, the problem is fixed:
>>
>> @@ -914,13 +914,27 @@ void arm_spe_precise_ip_support(struct evlist *evlist, struct evsel *evsel)
>>                 pmu = perf_pmu__find("arm_spe_0");
>>                 if (pmu) {
>>                         evsel->pmu_name = pmu->name;
>> -                       evsel->core.attr.type = PERF_RECORD_AUXTRACE;
>> -                       evsel->core.attr.config = SPE_ATTR_TS_ENABLE
>> -                                               | SPE_ATTR_PA_ENABLE
>> -                                               | SPE_ATTR_JITTER
>> +                       evsel->core.attr.type = pmu->type;
>> +                       evsel->core.attr.config |= SPE_ATTR_TS_ENABLE
>>                                                 | SPE_ATTR_BRANCH_FILTER;
>>
> 
> Hi, James,
> OK. Thank you for your fix.
> 
>> Also do you think jitter should be enabled by default? I thought that it might make the data less precise, so I removed it here.
> 
> Since the interval for sampling without "jitter" is fixed (default 1024 on our server), I was worried that not adding it would result in the same result for each record, and some instructions could not be collected each time.
> 
> However, after many tests, it is not clear from the results that there is a significant difference between them (enable it or not).
> 
> So I am confused, whether to enable it or not.
> 
> Thanks.
> Xiaojun.
> 
>>
>> -James
>>
>>>
>>> commit f116b96685a046a89c25d4a6ba2da489145c8888 (mainline/master)
>>> Merge: f632bfaa33ed 603d9299da32
>>> Author: Linus Torvalds <torvalds@...ux-foundation.org>
>>> Date:   Thu Oct 24 06:13:45 2019 -0400
>>>
>>>     Merge tag 'mfd-fixes-5.4' of git://git.kernel.org/pub/scm/linux/kernel/git/lee/mfd
>>>
>>> I will go and see why this will be reported.
>>>
>>>>
>>>>
>>>> I would have expected to use the event name that is listed in the SPE documentation for branch misses which is br_mis_pred or br_mis_pred_retired:
>>>>
>>>>     E[7], byte 0 bit [7]
>>>>     Mispredicted. The defined values of this bit are:
>>>>     0 Did not cause correction to the predicted program flow.
>>>>     1 A branch that caused a correction to the predicted program flow.
>>>>
>>>>     If PMUv3 is implemented this Event is required to be implemented consistently with either BR_MIS_PRED or BR_MIS_PRED_RETIRED.
>>>>
>>>
>>> Do you mean that I can add these as new events to perf? If we think of them as new events, what should we do if the user does not add :pp for them?
>>> (Or for these events, users can only add :pp to use them?)
>>>
>>>>
>>>> +       if (!strcmp(perf_env__arch(evlist->env), "arm64")
>>>> +                       && evsel->core.attr.config == PERF_COUNT_HW_BRANCH_MISSES
>>>> +                       && evsel->core.attr.precise_ip) {
>>>>
>>>> As I mentioned above PERF_COUNT_HW_BRANCH_MISSESdoesn't seem to match up with the actual event counter that is associated with this SPE event (BR_MIS_PRED). The fix for this is probably as simple as adding an OR for the other aliases for branch mispredicts.
>>>
>>> What you mean is that we can filter with spe events(like BR_MIS_PRED) first, and if we have other events that are exactly the same(no more for now), then we can handle them by adding OR in the future?
>>>
>>>>
>>>> +               pmu = perf_pmu__find("arm_spe_0");
>>>> +               if (pmu) {
>>>> +                       evsel->pmu_name = pmu->name;
>>>> +                       evsel->core.attr.type = PERF_RECORD_AUXTRACE;
>>>> +                       evsel->core.attr.config = SPE_ATTR_TS_ENABLE
>>>> +                                               | SPE_ATTR_PA_ENABLE
>>>>
>>>> I wouldn't set physical addresses by default as this requires root. I would leave that to the user if they want to manually configure SPE.
>>>
>>> Yes. You are right. I got a error for this case. I will fix it.
>>>
>>> ------------------
>>> ./perf record -e branch-misses:p ls
>>> Error:
>>> You may not have permission to collect stats.
>>> ...
>>> ------------------
>>>
>>> Thanks.
>>> Xiaojun.
>>>
>>>>
>>>> I have only looked briefly and I will do some more testing.
>>>>
>>>>
>>>> Thanks
>>>> James
>>>>
>>>>
>>>
>>>
> 
> 

Powered by blists - more mailing lists