linux-kernel - Re: [PATCH v2 11/27] perf parse-events: Support hardware events inside PMU

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <fd88f214-f0a4-87bc-ef52-ee750ca13a8d@linux.intel.com>
Date:   Tue, 16 Mar 2021 09:49:42 +0800
From:   "Jin, Yao" <yao.jin@...ux.intel.com>
To:     Jiri Olsa <jolsa@...hat.com>
Cc:     acme@...nel.org, jolsa@...nel.org, peterz@...radead.org,
        mingo@...hat.com, alexander.shishkin@...ux.intel.com,
        Linux-kernel@...r.kernel.org, ak@...ux.intel.com,
        kan.liang@...el.com, yao.jin@...el.com
Subject: Re: [PATCH v2 11/27] perf parse-events: Support hardware events
 inside PMU

Hi Jiri,

On 3/16/2021 1:37 AM, Jiri Olsa wrote:
> On Mon, Mar 15, 2021 at 10:28:12AM +0800, Jin, Yao wrote:
>> Hi Jiri,
>>
>> On 3/13/2021 3:15 AM, Jiri Olsa wrote:
>>> On Thu, Mar 11, 2021 at 03:07:26PM +0800, Jin Yao wrote:
>>>> On hybrid platform, some hardware events are only available
>>>> on a specific pmu. For example, 'L1-dcache-load-misses' is only
>>>> available on 'cpu_core' pmu. And even for the event which can be
>>>> available on both pmus, the user also may want to just enable
>>>> one event. So now following syntax is supported:
>>>>
>>>> cpu_core/<hardware event>/
>>>> cpu_core/<hardware cache event>/
>>>> cpu_core/<pmu event>/
>>>>
>>>> cpu_atom/<hardware event>/
>>>> cpu_atom/<hardware cache event>/
>>>> cpu_atom/<pmu event>/
>>>>
>>>> It limits the event to be enabled only on a specified pmu.
>>>>
>>>> The patch uses this idea, for example, if we use "cpu_core/LLC-loads/",
>>>> in parse_events_add_pmu(), term->config is "LLC-loads".
>>>
>>> hum, I don't understand how this doest not work even now,
>>> I assume both cpu_core and cpu_atom have sysfs device directory
>>> with events/ directory right?
>>>
>>
>> Yes, we have cpu_core and cpu_atom directories with events.
>>
>> root@...-pwrt-002:/sys/devices/cpu_atom/events# ls
>> branch-instructions  bus-cycles    cache-references  instructions
>> mem-stores  topdown-bad-spec topdown-fe-bound
>> branch-misses        cache-misses  cpu-cycles        mem-loads
>> ref-cycles  topdown-be-bound topdown-retiring
>>
>> root@...-pwrt-002:/sys/devices/cpu_core/events# ls
>> branch-instructions  cache-misses      instructions   mem-stores
>> topdown-bad-spec topdown-fe-bound   topdown-mem-bound
>> branch-misses        cache-references  mem-loads      ref-cycles
>> topdown-be-bound topdown-fetch-lat  topdown-retiring
>> bus-cycles           cpu-cycles        mem-loads-aux  slots
>> topdown-br-mispredict topdown-heavy-ops
>>
>>> and whatever is defined in events we allow in parsing syntax..
>>>
>>> why can't we treat them like 2 separated pmus?
>>>
>>
>> But if without this patch, it reports the error,
>>
>> root@...-pwrt-002:~# ./perf stat -e cpu_core/cycles/ -a -vv -- sleep 1
>> event syntax error: 'cpu_core/cycles/'
>>                                \___ unknown term 'cycles' for pmu 'cpu_core'
> 
> yep, because there's special care for 'cycles' unfortunately,
> but you should be able to run 'cpu_core/cpu-cycles/' right?
> 

Yes, cpu_core/cpu-cycles/ is OK.

# ./perf stat -e cpu_core/cpu-cycles/ -a -- sleep 1

  Performance counter stats for 'system wide':

     12,831,980,326      cpu_core/cpu-cycles/

        1.003132639 seconds time elapsed

>>
>> valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore
>>
>> Initial error:
>> event syntax error: 'cpu_core/cycles/'
>>                                \___ unknown term 'cycles' for pmu 'cpu_core'
>>
>> valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore
>> Run 'perf list' for a list of valid events
>>
>> The 'cycles' is treated as a unknown term, then it errors out.
> 
> yep, because it's not in events.. we could add special rule to
> treat cycles as cpu-cycles inside pmu definition ;-)
> 
> jirka
> 

But not only the cycles, the branches has error too.

# ./perf stat -e cpu_core/branches/ -a -- sleep 1
event syntax error: 'cpu_core/branches/'
                               \___ unknown term 'branches' for pmu 'cpu_core'

valid terms: 
event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore

Initial error:
event syntax error: 'cpu_core/branches/'
                               \___ unknown term 'branches' for pmu 'cpu_core'

Of course, branch-instructions runs OK.

# ./perf stat -e cpu_core/branch-instructions/ -a -- sleep 1

  Performance counter stats for 'system wide':

        136,655,302      cpu_core/branch-instructions/

        1.003171561 seconds time elapsed

So we need special rules for both cycles and branches.

The worse thing is, we also need to process the hardware cache events.

# ./perf stat -e cpu_core/LLC-loads/
event syntax error: 'cpu_core/LLC-loads/'
                               \___ unknown term 'LLC-loads' for pmu 'cpu_core'

valid terms: 
event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore

Initial error:
event syntax error: 'cpu_core/LLC-loads/'
                               \___ unknown term 'LLC-loads' for pmu 'cpu_core'

If we use special rules for establishing all event mapping, that looks too much. :(

Thanks
Jin Yao