Skip to contentSkip to conversation

perf stat: Regression perf stat -T cpi fails on s390 z/VM guest systems

PATCH In review Thomas Richter 2 participants 2 messages Nov 13, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf stat: Regression perf stat -T cpi fails on s390 z/VM guest systems

On s390 z/VM systems (linux guest on s390 LPAR) this happens:

 # perf test 96 97
 96: perf all metricgroups test        : FAILED!
 97: perf all metrics test             : FAILED!

This test works on the linux repo:

 # perf test 95 96
 95: perf all metricgroups test        : Ok
 96: perf all metrics test             : Ok

On both systems the command

 # ./perf list --raw-dump metric
 cpi est_cpi finite_cpi l1mp l2p l3p l4lp l4rp memp prbstate scpl1m \
	tlb_miss tlb_percent transaction
shows the same list of available metrics.

On z/VM the CPU Measurement facilities do not exist:

 # ll  /sys/devices/cpum_cf/events
 ls: cannot access '/sys/devices/cpum_cf/events': No such file or directory
 #

The json files for s390 define the metric 'cpi' in arch/s390/cf_z16/transaction.json:

  {
    "BriefDescription": "Cycles per Instruction",
    "MetricName": "cpi",
    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS \
                                if has_event(INSTRUCTIONS) else 0"
  },

The macro has_event(INSTRUCTIONS) now refers to a legacy event which always exists. It is always true even when the hardware does not support this event. Change the has_event(xxx) to check for a hardware event not available, for example CPU_CYCLES.

Fixes 0012e0fa221b ("perf jevents: Add legacy-hardware and legacy-cache json")

Suggested by: Ian Rogers <[email protected]>

---
 tools/perf/pmu-events/arch/s390/cf_z16/transaction.json | 8 ++++----
 tools/perf/pmu-events/arch/s390/cf_z17/transaction.json | 8 ++++----
 2 files changed, 8 insertions(+), 8 deletions(-)
tools/perf/pmu-events/arch/s390/cf_z16/transaction.json +4 -4
index 3ab1d3a6638c..26c550621d6a 100644
@@ -7,17 +7,17 @@
   {
     "BriefDescription": "Cycles per Instruction",
     "MetricName": "cpi",
-    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Problem State Instruction Ratio",
     "MetricName": "prbstate",
-    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Level One Miss per 100 Instructions",
     "MetricName": "l1mp",
-    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Percentage sourced from Level 2 cache",
@@ -52,7 +52,7 @@
   {
     "BriefDescription": "Estimated Instruction Complexity CPI infinite Level 1",
     "MetricName": "est_cpi",
-    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Estimated Sourcing Cycles per Level 1 Miss",
tools/perf/pmu-events/arch/s390/cf_z17/transaction.json +4 -4
index 74df533c8b6f..4d296e0c8934 100644
@@ -7,17 +7,17 @@
   {
     "BriefDescription": "Cycles per Instruction",
     "MetricName": "cpi",
-    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Problem State Instruction Ratio",
     "MetricName": "prbstate",
-    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Level One Miss per 100 Instructions",
     "MetricName": "l1mp",
-    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Percentage sourced from Level 2 cache",
@@ -52,7 +52,7 @@
   {
     "BriefDescription": "Estimated Instruction Complexity CPI infinite Level 1",
     "MetricName": "est_cpi",
-    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Estimated Sourcing Cycles per Level 1 Miss",
Signature · 1 line
2.51.1
reviewer Reviewed-by on v1 1/1
#1 lore raw
Quoted 34 lines · Thomas Richter

On s390 z/VM systems (linux guest on s390 LPAR) this happens:

 # perf test 96 97
 96: perf all metricgroups test        : FAILED!
 97: perf all metrics test             : FAILED!

This test works on the linux repo:

 # perf test 95 96
 95: perf all metricgroups test        : Ok
 96: perf all metrics test             : Ok

On both systems the command

 # ./perf list --raw-dump metric
 cpi est_cpi finite_cpi l1mp l2p l3p l4lp l4rp memp prbstate scpl1m \
        tlb_miss tlb_percent transaction
shows the same list of available metrics.

On z/VM the CPU Measurement facilities do not exist:

 # ll  /sys/devices/cpum_cf/events
 ls: cannot access '/sys/devices/cpum_cf/events': No such file or directory
 #

The json files for s390 define the metric 'cpi' in arch/s390/cf_z16/transaction.json:

  {
    "BriefDescription": "Cycles per Instruction",
    "MetricName": "cpi",
    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS \
                                if has_event(INSTRUCTIONS) else 0"
  },
Quoted 8 lines

The macro has_event(INSTRUCTIONS) now refers to a legacy event which always exists. It is always true even when the hardware does not support this event. Change the has_event(xxx) to check for a hardware event not available, for example CPU_CYCLES.

Fixes 0012e0fa221b ("perf jevents: Add legacy-hardware and legacy-cache json")

Suggested by: Ian Rogers <[email protected]>

This fix is equivalent to:
https://lore.kernel.org/lkml/[email protected]/
I'm easy about which one is picked up, the commit message here has more detail.

Thanks,
Ian

Quoted 76 lines
---
 tools/perf/pmu-events/arch/s390/cf_z16/transaction.json | 8 ++++----
 tools/perf/pmu-events/arch/s390/cf_z17/transaction.json | 8 ++++----
 2 files changed, 8 insertions(+), 8 deletions(-)
tools/perf/pmu-events/arch/s390/cf_z16/transaction.json +4 -4
index 3ab1d3a6638c..26c550621d6a 100644
@@ -7,17 +7,17 @@
   {
     "BriefDescription": "Cycles per Instruction",
     "MetricName": "cpi",
-    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Problem State Instruction Ratio",
     "MetricName": "prbstate",
-    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Level One Miss per 100 Instructions",
     "MetricName": "l1mp",
-    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Percentage sourced from Level 2 cache",
@@ -52,7 +52,7 @@
   {
     "BriefDescription": "Estimated Instruction Complexity CPI infinite Level 1",
     "MetricName": "est_cpi",
-    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Estimated Sourcing Cycles per Level 1 Miss",
tools/perf/pmu-events/arch/s390/cf_z17/transaction.json +4 -5
index 74df533c8b6f..4d296e0c8934 100644
@@ -7,17 +7,17 @@
   {
     "BriefDescription": "Cycles per Instruction",
     "MetricName": "cpi",
-    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "CPU_CYCLES / INSTRUCTIONS if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Problem State Instruction Ratio",
     "MetricName": "prbstate",
-    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(PROBLEM_STATE_INSTRUCTIONS / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Level One Miss per 100 Instructions",
     "MetricName": "l1mp",
-    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "((L1I_DIR_WRITES + L1D_DIR_WRITES) / INSTRUCTIONS) * 100 if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Percentage sourced from Level 2 cache",
@@ -52,7 +52,7 @@
   {
     "BriefDescription": "Estimated Instruction Complexity CPI infinite Level 1",
     "MetricName": "est_cpi",
-    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(INSTRUCTIONS) else 0"
+    "MetricExpr": "(CPU_CYCLES / INSTRUCTIONS) - (L1C_TLB2_MISSES / INSTRUCTIONS) if has_event(CPU_CYCLES) else 0"
   },
   {
     "BriefDescription": "Estimated Sourcing Cycles per Level 1 Miss",
--

2.51.1