The RAPL events exposed under /sys/devices/power/events should only reflect
what the underlying hardware actually support. This is how it works on Intel
RAPL and Intel core/uncore PMUs in general.
But on AMD, this was not the case. All possible RAPL events were advertised.
This is what it showed on an AMD Fam17h:
$ ls /sys/devices/power/events/
energy-cores energy-gpu energy-pkg energy-psys
energy-ram energy-cores.scale energy-gpu.scale energy-pkg.scale
energy-psys.scale energy-ram.scale energy-cores.unit energy-gpu.unit
energy-pkg.unit energy-psys.unit energy-ram.unit
Yet, on AMD Fam17h, only energy-pkg is supported.
This patch fixes the problem. Given the way perf_msr_probe() works, the
amd_rapl_msrs[] table has to have all entries filled out and in particular
the group field, otherwise perf_msr_probe() defaults to making the event
visible.
With the patch applied, the kernel now only shows was is actually supported:
$ ls /sys/devices/power/events/
energy-pkg energy-pkg.scale energy-pkg.unit
The patch also uses the RAPL_MSR_MASK because only the 32-bits LSB of the
RAPL counters are relevant when reading power consumption.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220105185659.643355-1-eranian@google.com
0036fb00a7