Skip to contentSkip to conversation

perf stat: Add arch-specific TopdownL1 check for the default mode

PATCH In review [email protected] 2 participants 2 messages Apr 27, 2023 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf stat: Add arch-specific TopdownL1 check for the default mode

From: Kan Liang <[email protected]>

The default of perf stat fails on several Intel platforms. Skylake:

$ perf stat true
Error:
Access to performance monitoring and observability operations is limited.
Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open
access to performance monitoring and observability operations for processes
without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.
More information can be found at 'Perf events and tool security' document:
https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html
perf_event_paranoid setting is 2:
  -1: Allow use of (almost) all events by all users
      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
Quoted 3 lines

= 0: Disallow raw and ftrace function tracepoint access = 1: Disallow CPU event access
= 2: Disallow kernel profiling

ADL (hybrid):

./perf stat
Segmentation fault (core dumped)

The default of perf stat was switched to TopdownL1 Json metric since commit 94b1a603fca7("perf stat: Add TopdownL1 metric as a default if present"). But the patch only checks whether the TopdownL1 is present in the event list. It doesn't check whether the hardware has the capability to provide a clean output for the default mode.

Add arch_has_topdown_metric() to check the hardware capability as well. Drop the TopdownL1 support in the defalut mode for pre-ICL and hybrid platforms. Users can still use -M TopdownL1 to access the TopdownL1 on pre-ICL platforms.

---

The patch tries to workaround the serious issues on pre-ICL and hybrid platforms with the default mode of perf stat. It could be a temporary fix for the upcoming 6.4. So we have more time to look for a proper fix for all metrics issues and output issues with 6.5.

Thanks,
Kan

 tools/perf/arch/x86/util/topdown.c | 14 ++++++++++++++
 tools/perf/builtin-stat.c          |  2 +-
 tools/perf/util/stat-display.c     |  2 +-
 tools/perf/util/topdown.c          |  6 ++++++
 tools/perf/util/topdown.h          |  2 ++
 5 files changed, 24 insertions(+), 2 deletions(-)
tools/perf/arch/x86/util/topdown.c +14 -0
index 9ad5e5c7bd27..5d861e851619 100644
@@ -3,6 +3,7 @@
 #include "util/evsel.h"
 #include "util/pmu.h"
 #include "util/topdown.h"
+#include "util/metricgroup.h"
 #include "topdown.h"
 #include "evsel.h"
 
@@ -48,3 +49,16 @@ bool arch_topdown_sample_read(struct evsel *leader)
 
 	return false;
 }
+
+bool arch_has_topdown_metric(const char *name)
+{
+	/*
+	 * Disable the Topdown events in the default mode
+	 * for hybrid platforms and old platform which
+	 * doesn't support the Topdown metric feature.
+	 */
+	if (!pmu_have_event("cpu", "slots"))
+		return false;
+
+	return metricgroup__has_metric(name);
+}
tools/perf/builtin-stat.c +1 -1
index efda63f6bf32..0b865155656d 100644
@@ -1885,7 +1885,7 @@ static int add_default_attributes(void)
 		 * Add TopdownL1 metrics if they exist. To minimize
 		 * multiplexing, don't request threshold computation.
 		 */
-		if (metricgroup__has_metric("TopdownL1") &&
+		if (arch_has_topdown_metric("TopdownL1") &&
 		    metricgroup__parse_groups(evsel_list, "TopdownL1",
 					    /*metric_no_group=*/false,
 					    /*metric_no_merge=*/false,
tools/perf/util/stat-display.c +1 -1
index e6035ecbeee8..73b2ff2ddf29 100644
@@ -747,7 +747,7 @@ static void uniquify_event_name(struct evsel *counter)
 	int ret = 0;
 
 	if (counter->uniquified_name || counter->use_config_name ||
-	    !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
+	    !counter->pmu_name || !strncmp(evsel__name(counter), counter->pmu_name,
 					   strlen(counter->pmu_name)))
 		return;
 
tools/perf/util/topdown.c +6 -0
index 18fd5fed5d1a..f3a9ebc52f8b 100644
@@ -1,8 +1,14 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "topdown.h"
+#include "metricgroup.h"
 #include <linux/kernel.h>
 
 __weak bool arch_topdown_sample_read(struct evsel *leader __maybe_unused)
 {
 	return false;
 }
+
+__weak bool arch_has_topdown_metric(const char *name)
+{
+	return metricgroup__has_metric(name);
+}
tools/perf/util/topdown.h +2 -0
index 1996c5fedcd7..7e83c8b247f2 100644
@@ -8,4 +8,6 @@ struct evsel;
 
 bool arch_topdown_sample_read(struct evsel *leader);
 
+bool arch_has_topdown_metric(const char *name);
+
 #endif
Signature · 1 line
2.35.1
reviewer
#1 lore raw
Quoted 71 lines · On Thu, Apr 27, 2023 at 11:31 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The default of perf stat fails on several Intel platforms. Skylake:

$ perf stat true
Error:
Access to performance monitoring and observability operations is limited.
Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open
access to performance monitoring and observability operations for processes
without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.
More information can be found at 'Perf events and tool security' document:
https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html
perf_event_paranoid setting is 2:
  -1: Allow use of (almost) all events by all users
      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
Quoted 3 lines

= 0: Disallow raw and ftrace function tracepoint access = 1: Disallow CPU event access
= 2: Disallow kernel profiling

ADL (hybrid):

./perf stat
Segmentation fault (core dumped)

The default of perf stat was switched to TopdownL1 Json metric since commit 94b1a603fca7("perf stat: Add TopdownL1 metric as a default if present"). But the patch only checks whether the TopdownL1 is present in the event list. It doesn't check whether the hardware has the capability to provide a clean output for the default mode.

Add arch_has_topdown_metric() to check the hardware capability as well. Drop the TopdownL1 support in the defalut mode for pre-ICL and hybrid platforms. Users can still use -M TopdownL1 to access the TopdownL1 on pre-ICL platforms.

---

The patch tries to workaround the serious issues on pre-ICL and hybrid platforms with the default mode of perf stat. It could be a temporary fix for the upcoming 6.4. So we have more time to look for a proper fix for all metrics issues and output issues with 6.5.

Thanks,
Kan

 tools/perf/arch/x86/util/topdown.c | 14 ++++++++++++++
 tools/perf/builtin-stat.c          |  2 +-
 tools/perf/util/stat-display.c     |  2 +-
 tools/perf/util/topdown.c          |  6 ++++++
 tools/perf/util/topdown.h          |  2 ++
 5 files changed, 24 insertions(+), 2 deletions(-)
tools/perf/arch/x86/util/topdown.c +1 -0
index 9ad5e5c7bd27..5d861e851619 100644
@@ -3,6 +3,7 @@
 #include "util/evsel.h"
 #include "util/pmu.h"
 #include "util/topdown.h"
+#include "util/metricgroup.h"
 #include "topdown.h"
 #include "evsel.h"
 
@@ -48,3 +49,16 @@ bool arch_topdown_sample_read(struct evsel *leader)
 
        return false;
 }

tools/perf/arch/x86/util/topdown.c bool arch_topdown_sample_read(struct evsel *leader) · line 61

Quoted 10 lines
tools/perf/arch/x86/util/topdown.c +10 -0
+
+bool arch_has_topdown_metric(const char *name)
+{
+       /*
+        * Disable the Topdown events in the default mode
+        * for hybrid platforms and old platform which
+        * doesn't support the Topdown metric feature.
+        */
+       if (!pmu_have_event("cpu", "slots"))
+               return false;

The only platform crashing with this are hybrid ones, I think the test should be:

if (perf_pmu__has_hybrid())
return false;

Quoted 21 lines
tools/perf/arch/x86/util/topdown.c +3 -0
+
+       return metricgroup__has_metric(name);
+}
tools/perf/builtin-stat.c +1 -1
index efda63f6bf32..0b865155656d 100644
@@ -1885,7 +1885,7 @@ static int add_default_attributes(void)
                 * Add TopdownL1 metrics if they exist. To minimize
                 * multiplexing, don't request threshold computation.
                 */
-               if (metricgroup__has_metric("TopdownL1") &&
+               if (arch_has_topdown_metric("TopdownL1") &&
                    metricgroup__parse_groups(evsel_list, "TopdownL1",
                                            /*metric_no_group=*/false,
                                            /*metric_no_merge=*/false,
tools/perf/util/stat-display.c +0 -0
index e6035ecbeee8..73b2ff2ddf29 100644

tools/perf/util/stat-display.c static void uniquify_event_name(struct evsel *counter) · line 750

Quoted 5 lines
tools/perf/util/stat-display.c +1 -1
@@ -747,7 +747,7 @@ static void uniquify_event_name(struct evsel *counter)
        int ret = 0;
 
        if (counter->uniquified_name || counter->use_config_name ||
-           !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
+           !counter->pmu_name || !strncmp(evsel__name(counter), counter->pmu_name,

This fix is here:
https://lore.kernel.org/lkml/[email protected]/T/#mfce90d81aac130bbbf4743310b9ab918fc73d012

Thanks,
Ian

Quoted 36 lines
tools/perf/util/stat-display.c +0 -0
                                           strlen(counter->pmu_name)))
                return;
 
tools/perf/util/topdown.c +6 -0
index 18fd5fed5d1a..f3a9ebc52f8b 100644
@@ -1,8 +1,14 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "topdown.h"
+#include "metricgroup.h"
 #include <linux/kernel.h>
 
 __weak bool arch_topdown_sample_read(struct evsel *leader __maybe_unused)
 {
        return false;
 }
+
+__weak bool arch_has_topdown_metric(const char *name)
+{
+       return metricgroup__has_metric(name);
+}
tools/perf/util/topdown.h +2 -1
index 1996c5fedcd7..7e83c8b247f2 100644
@@ -8,4 +8,6 @@ struct evsel;
 
 bool arch_topdown_sample_read(struct evsel *leader);
 
+bool arch_has_topdown_metric(const char *name);
+
 #endif
--

2.35.1