s390/cpumf: Handle events cycles and instructions identical

Events CPU_CYCLES and INSTRUCTIONS can be submitted with two different
perf_event attribute::type values:
 - PERF_TYPE_HARDWARE: when invoked via perf tool predefined events name
   cycles or cpu-cycles or instructions.
 - pmu->type: when invoked via perf tool event name cpu_cf/CPU_CYLCES/ or
   cpu_cf/INSTRUCTIONS/. This invocation also selects the PMU to which
   the event belongs.
Handle both type of invocations identical for events CPU_CYLCES and
INSTRUCTIONS. They address the same hardware.
The result is different when event modifier exclude_kernel is also set.
Invocation with event modifier for user space event counting fails.

Output before:

 # perf stat -e cpum_cf/cpu_cycles/u -- true

 Performance counter stats for 'true':

   <not supported>      cpum_cf/cpu_cycles/u

       0.000761033 seconds time elapsed

       0.000076000 seconds user
       0.000725000 seconds sys

 #

Output after:
 # perf stat -e cpum_cf/cpu_cycles/u -- true

 Performance counter stats for 'true':

           349,613      cpum_cf/cpu_cycles/u

       0.000844143 seconds time elapsed

       0.000079000 seconds user
       0.000800000 seconds sys
 #

Fixes: 6a82e23f45fe ("s390/cpumf: Adjust registration of s390 PMU device drivers")
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
[agordeev@linux.ibm.com corrected commit ID of Fixes commit]
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>

diff --git a/arch/s390/kernel/perf_cpum_cf.c b/arch/s390/kernel/perf_cpum_cf.c
index 483ab5e..f7dd3c8 100644 (file)
--- a/arch/s390/kernel/perf_cpum_cf.c
+++ b/arch/s390/kernel/perf_cpum_cf.c
@@ -516,6 +516,26 @@ static int __hw_perf_event_init(struct perf_event *event, unsigned int type)
        return err;
 }
 
+/* Events CPU_CYLCES and INSTRUCTIONS can be submitted with two different
+ * attribute::type values:
+ * - PERF_TYPE_HARDWARE:
+ * - pmu->type:
+ * Handle both type of invocations identical. They address the same hardware.
+ * The result is different when event modifiers exclude_kernel and/or
+ * exclude_user are also set.
+ */
+static int cpumf_pmu_event_type(struct perf_event *event)
+{
+       u64 ev = event->attr.config;
+
+       if (cpumf_generic_events_basic[PERF_COUNT_HW_CPU_CYCLES] == ev ||
+           cpumf_generic_events_basic[PERF_COUNT_HW_INSTRUCTIONS] == ev ||
+           cpumf_generic_events_user[PERF_COUNT_HW_CPU_CYCLES] == ev ||
+           cpumf_generic_events_user[PERF_COUNT_HW_INSTRUCTIONS] == ev)
+               return PERF_TYPE_HARDWARE;
+       return PERF_TYPE_RAW;
+}
+
 static int cpumf_pmu_event_init(struct perf_event *event)
 {
        unsigned int type = event->attr.type;
@@ -525,7 +545,7 @@ static int cpumf_pmu_event_init(struct perf_event *event)
                err = __hw_perf_event_init(event, type);
        else if (event->pmu->type == type)
                /* Registered as unknown PMU */
-               err = __hw_perf_event_init(event, PERF_TYPE_RAW);
+               err = __hw_perf_event_init(event, cpumf_pmu_event_type(event));
        else
                return -ENOENT;