Re: [PATCH] perf vendor events arm64: Enable JSON events for ThunderX2 B0

From: William Cohen
Date: Tue Mar 20 2018 - 22:38:03 EST


This is a multi-part message in MIME format.On 03/15/2018 12:47 PM, John Garry wrote:
> On 15/03/2018 15:53, William Cohen wrote:
>> On 03/07/2018 11:14 PM, Ganapatrao Kulkarni wrote:
>>> On Thu, Mar 8, 2018 at 12:01 AM, William Cohen <wcohen@xxxxxxxxxx> wrote:
>>>> On 03/07/2018 12:35 PM, Ganapatrao Kulkarni wrote:
>>>>> Hi Will Cohen,
>>>>>
>>>>> On Wed, Mar 7, 2018 at 8:08 PM, Arnaldo Carvalho de Melo
>>>>> <acme@xxxxxxxxxx> wrote:
>>>>>> Em Wed, Mar 07, 2018 at 09:32:05AM -0500, William Cohen escreveu:
>>>>>>> On 03/07/2018 06:08 AM, Ganapatrao Kulkarni wrote:
>>>>>>>> There is MIDR change on ThunderX2 B0, adding an entry to mapfile
>>>>>>>> to enable JSON events for B0.
>>>>>>>>
>>>>>>>> Signed-off-by: Ganapatrao Kulkarni <ganapatrao.kulkarni@xxxxxxxxxx>
>>>>>>
>>>>>> Ganapatrao, can you please take this in consideration and if agreeing
>>>>>> send a v2 patch?
>>>>>>
>>>>>> With that I can add an Acked-by: wcohen, Right?
>>>>>>
>>>>>> - Arnaldo
>>>>>>>> ---
>>>>>>>> Âtools/perf/pmu-events/arch/arm64/mapfile.csv | 1 +
>>>>>>>> Â1 file changed, 1 insertion(+)
>>>>>>>>
>>>>>>>> diff --git a/tools/perf/pmu-events/arch/arm64/mapfile.csv b/tools/perf/pmu-events/arch/arm64/mapfile.csv
>>>>>>>> index e61c9ca..93c5d14 100644
>>>>>>>> --- a/tools/perf/pmu-events/arch/arm64/mapfile.csv
>>>>>>>> +++ b/tools/perf/pmu-events/arch/arm64/mapfile.csv
>>>>>>>> @@ -13,4 +13,5 @@
>>>>>>>> Â#
>>>>>>>> Â#Family-model,Version,Filename,EventType
>>>>>>>> Â0x00000000420f5160,v1,cavium,core
>>>>>>>> +0x00000000430f0af0,v1,cavium,core
>>>>>>>> Â0x00000000410fd03[[:xdigit:]],v1,cortex-a53,core
>>>>>>>>
>>>>>>>
>>>>>>> Hi,
>>>>>>> Like the cortex-a53 the last digit '0' of the match for the MIDR should be replaced with [[:xdigit:]] to allow for possible future revisions of chip:
>>>>>
>>>>> for arm64 implementation, bits 3:0(Revision) and bits 23:20(Variant)
>>>>> are ignored/dont-care.
>>>>
>>>> Thanks for pointing that out. See the code masking out those bits in linux/toos/perf/arch/util/header.c. For the ppc64 it just copies the equivalent of the MIDR including the revision bits. Thus, the need for regular expression matching to avoid having to create a new entry for each revision.
>>>
>>> It is same for arm64 too, there is no need to add an entry for every
>>> revision change, need to add when part number changes.
>>> This patch is not intended to add entry for revision change, the fact
>>> of the matter is that, there is complete MIDR change (vulcan to
>>> thunderx2) in B0.
>>> as per current arm64
>>> implementation(.tools/perf/arch/arm64/util/header.c), it is not
>>> required to have any dontcare marking in mapfile for revision/variant
>>> bits.
>>>
>>> thanks
>>> Ganapat
>>
>> Hi Ganapat,
>>
>> Would it make more sense to pass the MIDR value unmodified and then use regular expressions in mapfile.csv to match the values? If an event on a particular processor revision is broken or unusable it can be excluded from the list of events with a corrected list of events. There certainly have been errata listing events that do not work on specific revisions of armv8 processor implementations.
>>
>
> Then there are vendors who do not always properly implemenent MIDR or IIDRs (people who live in glass houses...).
>
> Btw, topic originally discussed here:
> https://lkml.org/lkml/2017/5/2/113
>
> Thanks,
> John
>
>> -Will

Hi John,

Attached is a patch that leaves the MIDR value unmodified and uses regular expressions in the mapfile.csv instead to ignore those bits. I have verified that the changes work on ARM Cortex a53 processor. Does it look reasonable?

-Will Cohen