Is PEBS-via-PT expected to work on Alder Lake Gracemont cores?

Is PEBS-via-PT expected to work on Alder Lake Gracemont cores?

[Tudor Brindus]

Hi linux-perf-users,

I recently obtained an Alder Lake machine, and was hoping to try out
PEBS-via-PT on its E-cores.

The patchset[0] adding support for Alder Lake to perf suggests this should be
possible:

> - The "cpu_core" PMU supports Topdown metrics.
>   The "cpu_atom" PMU supports PEBS-via-PT.

However, I haven't been able to get it to work, despite spending the better
part of a day trying. I am on a 5.19 kernel.

$ taskset -c 15 ./demo &
$ perf record -e
'{intel_pt//,cpu_atom/name=br_misp,event=0xC5,umask=0x0,aux-output/ppp}:u'
-p $(pidof demo)
Error:
cpu_atom/br_misp/: PMU Hardware doesn't support 'aux_output' feature

dmesg reports that the cpu_atom PMU should support PEBS-via-PT:

[    0.005511] core: cpu_atom PMU driver: PEBS-via-PT
[    0.005511] ... version:                5
[    0.005511] ... bit width:              48
[    0.005511] ... generic registers:      6
[    0.005511] ... value mask:             0000ffffffffffff
[    0.005511] ... max period:             00007fffffffffff
[    0.005511] ... fixed-purpose events:   3
[    0.005511] ... event mask:             000000070000003f
[    0.060399]   #9 #10 #11 #12 #13 #14 #15

Am I doing something silly here? As far as I can tell from reading the patch,
PERF_PMU_CAP_AUX_OUTPUT is a function of intel_cap.pebs_output_pt_available,
which should be set (otherwise dmesg would not be printing the "PEBS-via-PT"
string).

Thanks,
Tudor

[0]: https://lore.kernel.org/lkml/20210310165358.GI23521@zn.tnic/T/

Re: Is PEBS-via-PT expected to work on Alder Lake Gracemont cores?

[Adrian Hunter]

On 13/06/22 22:45, Tudor Brindus wrote:
> Hi linux-perf-users,
> 
> I recently obtained an Alder Lake machine, and was hoping to try out
> PEBS-via-PT on its E-cores.
> 
> The patchset[0] adding support for Alder Lake to perf suggests this should be
> possible:
> 
>> - The "cpu_core" PMU supports Topdown metrics.
>>   The "cpu_atom" PMU supports PEBS-via-PT.
> 
> However, I haven't been able to get it to work, despite spending the better
> part of a day trying. I am on a 5.19 kernel.

I realise this is an old email but FWIW the Linux driver currently
supports only Intel PT features that are common to all CPUs.

> 
> $ taskset -c 15 ./demo &
> $ perf record -e
> '{intel_pt//,cpu_atom/name=br_misp,event=0xC5,umask=0x0,aux-output/ppp}:u'
> -p $(pidof demo)
> Error:
> cpu_atom/br_misp/: PMU Hardware doesn't support 'aux_output' feature
> 
> dmesg reports that the cpu_atom PMU should support PEBS-via-PT:
> 
> [    0.005511] core: cpu_atom PMU driver: PEBS-via-PT
> [    0.005511] ... version:                5
> [    0.005511] ... bit width:              48
> [    0.005511] ... generic registers:      6
> [    0.005511] ... value mask:             0000ffffffffffff
> [    0.005511] ... max period:             00007fffffffffff
> [    0.005511] ... fixed-purpose events:   3
> [    0.005511] ... event mask:             000000070000003f
> [    0.060399]   #9 #10 #11 #12 #13 #14 #15
> 
> Am I doing something silly here? As far as I can tell from reading the patch,
> PERF_PMU_CAP_AUX_OUTPUT is a function of intel_cap.pebs_output_pt_available,
> which should be set (otherwise dmesg would not be printing the "PEBS-via-PT"
> string).
> 
> Thanks,
> Tudor
> 
> [0]: https://lore.kernel.org/lkml/20210310165358.GI23521@zn.tnic/T/
>