Re: [PATCH] perf print-events: make is_event_supported() more robust

[James Clark]

On 17/01/2024 09:05, Marc Zyngier wrote:
> Hi Mark,
> 
> On Tue, 16 Jan 2024 17:03:48 +0000,
> Mark Rutland <mark.rutland@xxxxxxx> wrote:
>>
>> Currently the perf tool doesn't deteect support for extneded event types
>> on Apple M1/M2 systems, and will not auto-expand plain PERF_EVENT_TYPE
>> hardware events into per-PMU events. This is due to the detection of
>> extended event types not handling mandatory filters required by the
>> M1/M2 PMU driver.
> 
> Thanks for looking into this.
> 
> I've given your patch a go on my M1 box, and it indeed makes things
> substantially better:
> 
> $ sudo ./perf stat -e cycles ~/hackbench 100 process 1000
> Running with 100*40 (== 4000) tasks.
> Time: 3.419
> 
>  Performance counter stats for '/home/maz/hackbench 100 process 1000':
> 
>    174,783,472,090      apple_firestorm_pmu/cycles/                                             (93.10%)
>     39,134,744,813      apple_icestorm_pmu/cycles/                                              (71.86%)
> 
>        3.568145595 seconds time elapsed
> 
>       12.203084000 seconds user
>       55.135271000 seconds sys
> 
> However, I'm seeing some slightly odd behaviours:
> 
> $ sudo ./perf stat -e cycles:k ~/hackbench 100 process 1000
> Running with 100*40 (== 4000) tasks.
> Time: 3.313
> 
>  Performance counter stats for '/home/maz/hackbench 100 process 1000':
> 
>    <not supported>      apple_firestorm_pmu/cycles:k/                                         
>    <not supported>      apple_icestorm_pmu/cycles:k/                                          
> 
>        3.467568841 seconds time elapsed
> 
>       13.080111000 seconds user
>       53.162099000 seconds sys
> 
> I would have expected it to count, but it didn't. For that to work, I
> have to add the 'H' modifier:
> 
> $ sudo ./perf stat -e cycles:Hk ~/hackbench 100 process 1000
> Running with 100*40 (== 4000) tasks.
> Time: 3.335
> 
>  Performance counter stats for '/home/maz/hackbench 100 process 1000':
> 
>    183,756,134,397      apple_firestorm_pmu/cycles:Hk/                                          (85.56%)
>     37,302,841,991      apple_icestorm_pmu/cycles:Hk/                                           (72.10%)
> 
>        3.490138958 seconds time elapsed
> 
>       13.376772000 seconds user
>       53.326289000 seconds sys
> 
> But my perf-foo is as basic as it gets, so it is likely that I'm
> missing something.
> 
> Thanks,
> 
> 	M.
> 

The default for exclude_guest=1 is always set unless you use "perf kvm
record". But unfortunately if you add _any_ modifier then all the
default values get overwritten, which is why adding k stops it from working.

It seems like a lot of the pain comes from the fact that the command
line modifiers are includes, and the kernel options are excludes.
Resulting in some complicated logic in get_event_modifier() to try to
work out what you wanted, which I suppose it got it wrong in this scenario.

Another thing is that evsel__detect_missing_features() only works to
remove arguments, where as for M1/M2 you want to _add_ exclude_guest
back in again. But maybe not if you started adding modifiers...

I don't really have an answer or a solution, but I was wondering if
there was a bug that could be fixed so I went to look.

My first thought was maybe we could ignore exclude_guest=0 on M1/M2
rather than returning an error. I don't think it would be too surprising
that guest samples are missing if they are never supported. Maybe the
exclude_guest argument is only useful if it actually has an effect,
rather than having an argument that always needs to be on anyway.

Or secondly maybe we could make get_event_modifier() interact with
evsel__detect_missing_features() but I can't imagine it being very
maintainable once we start adding a few interactions. And it's still
hard to predict what the user really wants given a single letter and
many exclude_* options.

James