[PATCH v2 0/8] perf: Support perf-mem/perf-c2c for AlderLake

[PATCH v2 0/8] perf: Support perf-mem/perf-c2c for AlderLake

[Jin Yao]

AlderLake uses a hybrid architecture utilizing Golden Cove cores
(core CPU) and Gracemont cores (atom CPU). This patchset supports
perf-mem and perf-c2c for AlderLake.

v2:
---
- Use mem_loads_name__init to keep original behavior for non-hybrid platform.
- Move x86 specific perf_mem_events[] to arch/x86/util/mem-events.c.
- Move mem-store event to a new patch.
- Add a new patch to fix wrong verbose output for recording events
- Add a new patch to disable 'mem-loads-aux' group before reporting

Jin Yao (8):
  perf tools: Check mem-loads auxiliary event
  perf tools: Support pmu prefix for mem-load event
  perf tools: Support pmu prefix for mem-store event
  perf tools: Check if mem_events is supported for hybrid platform
  perf mem: Support record for hybrid platform
  perf mem: Fix wrong verbose output for recording events
  perf mem: Disable 'mem-loads-aux' group before reporting
  perf c2c: Support record for hybrid platform

 tools/perf/arch/arm64/util/mem-events.c   |   2 +-
 tools/perf/arch/powerpc/util/mem-events.c |   2 +-
 tools/perf/arch/x86/util/mem-events.c     |  54 ++++++++++--
 tools/perf/builtin-c2c.c                  |  40 +++++----
 tools/perf/builtin-mem.c                  |  51 ++++++-----
 tools/perf/builtin-report.c               |   2 +
 tools/perf/util/evlist.c                  |  25 ++++++
 tools/perf/util/evlist.h                  |   1 +
 tools/perf/util/mem-events.c              | 101 ++++++++++++++++++++--
 tools/perf/util/mem-events.h              |   4 +-
 10 files changed, 225 insertions(+), 57 deletions(-)

-- 
2.17.1

[PATCH v2 1/8] perf tools: Check mem-loads auxiliary event

[Jin Yao]

For some platforms, an auxiliary event has to be enabled
simultaneously with the load latency event.

For Alderlake, the auxiliary event is created in "cpu_core" pmu.

So first we need to check the existing of "cpu_core" pmu
and then check if this pmu has auxiliary event.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - No change.

 tools/perf/arch/x86/util/mem-events.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/tools/perf/arch/x86/util/mem-events.c b/tools/perf/arch/x86/util/mem-events.c
index 588110fd8904..e79232e3f2a0 100644
--- a/tools/perf/arch/x86/util/mem-events.c
+++ b/tools/perf/arch/x86/util/mem-events.c
@@ -11,8 +11,13 @@ static bool mem_loads_name__init;
 
 bool is_mem_loads_aux_event(struct evsel *leader)
 {
-	if (!pmu_have_event("cpu", "mem-loads-aux"))
-		return false;
+	if (perf_pmu__find("cpu")) {
+		if (!pmu_have_event("cpu", "mem-loads-aux"))
+			return false;
+	} else if (perf_pmu__find("cpu_core")) {
+		if (!pmu_have_event("cpu_core", "mem-loads-aux"))
+			return false;
+	}
 
 	return leader->core.attr.config == MEM_LOADS_AUX;
 }
-- 
2.17.1

[PATCH v2 2/8] perf tools: Support pmu prefix for mem-load event

[Jin Yao]

The perf_mem_events__name() can generate the mem-load event name.
It uses a variable 'mem_loads_name__init' to avoid generating the
event name every time (because perf_pmu__scan takes some time).

The perf_mem_events__name() assumes the pmu is "cpu" but it's not
correct for hybrid platform. For Alderlake, the pmu is "cpu_core" or
"cpu_atom"

Introduce a new parameter 'pmu_name' in perf_mem_events__name
to let the caller specify a pmu name.

Considering such event name is x86 specific, so move
perf_mem_events[] to arch/x86/util/mem-events.c.

We still keep the variable 'mem_loads_name__init' but it's only
used when pmu_name is NULL (compatible for original behavior). When
pmu_name is not NULL (e.g. "cpu_core"), this patch doesn't have
optimization. That can be implemented in follow up patch.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - Move perf_mem_events[] to x86 specific file.
 - Create x86 specific perf_mem_events__ptr().
 - Use mem_loads_name__init for keeping original behavior
   on non-hybrid platform.

 tools/perf/arch/arm64/util/mem-events.c   |  2 +-
 tools/perf/arch/powerpc/util/mem-events.c |  2 +-
 tools/perf/arch/x86/util/mem-events.c     | 35 ++++++++++++++++++-----
 tools/perf/builtin-c2c.c                  |  4 +--
 tools/perf/builtin-mem.c                  |  4 +--
 tools/perf/util/mem-events.c              |  4 +--
 tools/perf/util/mem-events.h              |  2 +-
 7 files changed, 37 insertions(+), 16 deletions(-)

diff --git a/tools/perf/arch/arm64/util/mem-events.c b/tools/perf/arch/arm64/util/mem-events.c
index 2a2497372671..be41721b9aa1 100644
--- a/tools/perf/arch/arm64/util/mem-events.c
+++ b/tools/perf/arch/arm64/util/mem-events.c
@@ -20,7 +20,7 @@ struct perf_mem_event *perf_mem_events__ptr(int i)
 	return &perf_mem_events[i];
 }
 
-char *perf_mem_events__name(int i)
+char *perf_mem_events__name(int i, char *pmu_name __maybe_unused)
 {
 	struct perf_mem_event *e = perf_mem_events__ptr(i);
 
diff --git a/tools/perf/arch/powerpc/util/mem-events.c b/tools/perf/arch/powerpc/util/mem-events.c
index 07fb5e049488..4120fafe0be4 100644
--- a/tools/perf/arch/powerpc/util/mem-events.c
+++ b/tools/perf/arch/powerpc/util/mem-events.c
@@ -3,7 +3,7 @@
 #include "mem-events.h"
 
 /* PowerPC does not support 'ldlat' parameter. */
-char *perf_mem_events__name(int i)
+char *perf_mem_events__name(int i, char *pmu_name __maybe_unused)
 {
 	if (i == PERF_MEM_EVENTS__LOAD)
 		return (char *) "cpu/mem-loads/";
diff --git a/tools/perf/arch/x86/util/mem-events.c b/tools/perf/arch/x86/util/mem-events.c
index e79232e3f2a0..f9e444a4fe70 100644
--- a/tools/perf/arch/x86/util/mem-events.c
+++ b/tools/perf/arch/x86/util/mem-events.c
@@ -7,7 +7,23 @@ static char mem_loads_name[100];
 static bool mem_loads_name__init;
 
 #define MEM_LOADS_AUX		0x8203
-#define MEM_LOADS_AUX_NAME	"{cpu/mem-loads-aux/,cpu/mem-loads,ldlat=%u/pp}:S"
+#define MEM_LOADS_AUX_NAME     "{%s/mem-loads-aux/,%s/mem-loads,ldlat=%u/}:P"
+
+#define E(t, n, s) { .tag = t, .name = n, .sysfs_name = s }
+
+static struct perf_mem_event perf_mem_events[PERF_MEM_EVENTS__MAX] = {
+	E("ldlat-loads",	"%s/mem-loads,ldlat=%u/P",	"%s/events/mem-loads"),
+	E("ldlat-stores",	"cpu/mem-stores/P",		"cpu/events/mem-stores"),
+	E(NULL,			NULL,				NULL),
+};
+
+struct perf_mem_event *perf_mem_events__ptr(int i)
+{
+	if (i >= PERF_MEM_EVENTS__MAX)
+		return NULL;
+
+	return &perf_mem_events[i];
+}
 
 bool is_mem_loads_aux_event(struct evsel *leader)
 {
@@ -22,7 +38,7 @@ bool is_mem_loads_aux_event(struct evsel *leader)
 	return leader->core.attr.config == MEM_LOADS_AUX;
 }
 
-char *perf_mem_events__name(int i)
+char *perf_mem_events__name(int i, char *pmu_name)
 {
 	struct perf_mem_event *e = perf_mem_events__ptr(i);
 
@@ -30,17 +46,22 @@ char *perf_mem_events__name(int i)
 		return NULL;
 
 	if (i == PERF_MEM_EVENTS__LOAD) {
-		if (mem_loads_name__init)
+		if (mem_loads_name__init && !pmu_name)
 			return mem_loads_name;
 
-		mem_loads_name__init = true;
+		if (!pmu_name) {
+			mem_loads_name__init = true;
+			pmu_name = (char *)"cpu";
+		}
 
-		if (pmu_have_event("cpu", "mem-loads-aux")) {
+		if (pmu_have_event(pmu_name, "mem-loads-aux")) {
 			scnprintf(mem_loads_name, sizeof(mem_loads_name),
-				  MEM_LOADS_AUX_NAME, perf_mem_events__loads_ldlat);
+				  MEM_LOADS_AUX_NAME, pmu_name, pmu_name,
+				  perf_mem_events__loads_ldlat);
 		} else {
 			scnprintf(mem_loads_name, sizeof(mem_loads_name),
-				  e->name, perf_mem_events__loads_ldlat);
+				  e->name, pmu_name,
+				  perf_mem_events__loads_ldlat);
 		}
 		return mem_loads_name;
 	}
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index e3b9d63077ef..a4fd375acdd1 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -2971,13 +2971,13 @@ static int perf_c2c__record(int argc, const char **argv)
 
 		if (!e->supported) {
 			pr_err("failed: event '%s' not supported\n",
-			       perf_mem_events__name(j));
+			       perf_mem_events__name(j, NULL));
 			free(rec_argv);
 			return -1;
 		}
 
 		rec_argv[i++] = "-e";
-		rec_argv[i++] = perf_mem_events__name(j);
+		rec_argv[i++] = perf_mem_events__name(j, NULL);
 	}
 
 	if (all_user)
diff --git a/tools/perf/builtin-mem.c b/tools/perf/builtin-mem.c
index cdd2b9f643f6..03795bf49d51 100644
--- a/tools/perf/builtin-mem.c
+++ b/tools/perf/builtin-mem.c
@@ -135,13 +135,13 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 
 		if (!e->supported) {
 			pr_err("failed: event '%s' not supported\n",
-			       perf_mem_events__name(j));
+			       perf_mem_events__name(j, NULL));
 			free(rec_argv);
 			return -1;
 		}
 
 		rec_argv[i++] = "-e";
-		rec_argv[i++] = perf_mem_events__name(j);
+		rec_argv[i++] = perf_mem_events__name(j, NULL);
 	}
 
 	if (all_user)
diff --git a/tools/perf/util/mem-events.c b/tools/perf/util/mem-events.c
index f93a852ad838..c736eaded06c 100644
--- a/tools/perf/util/mem-events.c
+++ b/tools/perf/util/mem-events.c
@@ -37,7 +37,7 @@ struct perf_mem_event * __weak perf_mem_events__ptr(int i)
 	return &perf_mem_events[i];
 }
 
-char * __weak perf_mem_events__name(int i)
+char * __weak perf_mem_events__name(int i, char *pmu_name  __maybe_unused)
 {
 	struct perf_mem_event *e = perf_mem_events__ptr(i);
 
@@ -141,7 +141,7 @@ void perf_mem_events__list(void)
 		fprintf(stderr, "%-13s%-*s%s\n",
 			e->tag ?: "",
 			verbose > 0 ? 25 : 0,
-			verbose > 0 ? perf_mem_events__name(j) : "",
+			verbose > 0 ? perf_mem_events__name(j, NULL) : "",
 			e->supported ? ": available" : "");
 	}
 }
diff --git a/tools/perf/util/mem-events.h b/tools/perf/util/mem-events.h
index cacdebd65b8a..a3fa19093fd2 100644
--- a/tools/perf/util/mem-events.h
+++ b/tools/perf/util/mem-events.h
@@ -38,7 +38,7 @@ extern unsigned int perf_mem_events__loads_ldlat;
 int perf_mem_events__parse(const char *str);
 int perf_mem_events__init(void);
 
-char *perf_mem_events__name(int i);
+char *perf_mem_events__name(int i, char *pmu_name);
 struct perf_mem_event *perf_mem_events__ptr(int i);
 bool is_mem_loads_aux_event(struct evsel *leader);
 
-- 
2.17.1

[PATCH v2 3/8] perf tools: Support pmu prefix for mem-store event

[Jin Yao]

For enabling mem-store event, it doesn't need an auxiliary event.
So just build an event name string with the pmu prefix.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - New in v2.

 tools/perf/arch/x86/util/mem-events.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/tools/perf/arch/x86/util/mem-events.c b/tools/perf/arch/x86/util/mem-events.c
index f9e444a4fe70..5214370ca4e4 100644
--- a/tools/perf/arch/x86/util/mem-events.c
+++ b/tools/perf/arch/x86/util/mem-events.c
@@ -5,6 +5,7 @@
 
 static char mem_loads_name[100];
 static bool mem_loads_name__init;
+static char mem_stores_name[100];
 
 #define MEM_LOADS_AUX		0x8203
 #define MEM_LOADS_AUX_NAME     "{%s/mem-loads-aux/,%s/mem-loads,ldlat=%u/}:P"
@@ -13,7 +14,7 @@ static bool mem_loads_name__init;
 
 static struct perf_mem_event perf_mem_events[PERF_MEM_EVENTS__MAX] = {
 	E("ldlat-loads",	"%s/mem-loads,ldlat=%u/P",	"%s/events/mem-loads"),
-	E("ldlat-stores",	"cpu/mem-stores/P",		"cpu/events/mem-stores"),
+	E("ldlat-stores",	"%s/mem-stores/P",		"%s/events/mem-stores"),
 	E(NULL,			NULL,				NULL),
 };
 
@@ -66,5 +67,14 @@ char *perf_mem_events__name(int i, char *pmu_name)
 		return mem_loads_name;
 	}
 
+	if (i == PERF_MEM_EVENTS__STORE) {
+		if (!pmu_name)
+			pmu_name = (char *)"cpu";
+
+		scnprintf(mem_stores_name, sizeof(mem_stores_name),
+			  e->name, pmu_name);
+		return mem_stores_name;
+	}
+
 	return (char *)e->name;
 }
-- 
2.17.1

[PATCH v2 4/8] perf tools: Check if mem_events is supported for hybrid platform

[Jin Yao]

Check if the mem_events ('mem-loads' and 'mem-stores') exist
in the sysfs path.

For Alderlake, the hybrid cpu pmu are "cpu_core" and "cpu_atom".
Check the existing of following paths:

/sys/devices/cpu_atom/events/mem-loads
/sys/devices/cpu_atom/events/mem-stores
/sys/devices/cpu_core/events/mem-loads
/sys/devices/cpu_core/events/mem-stores

If the patch exists, the mem_event is supported.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - Use 'e->supported |= perf_mem_event__supported(mnt, sysfs_name);'

 tools/perf/util/mem-events.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)

diff --git a/tools/perf/util/mem-events.c b/tools/perf/util/mem-events.c
index c736eaded06c..69dcac730ada 100644
--- a/tools/perf/util/mem-events.c
+++ b/tools/perf/util/mem-events.c
@@ -12,6 +12,8 @@
 #include "mem-events.h"
 #include "debug.h"
 #include "symbol.h"
+#include "pmu.h"
+#include "pmu-hybrid.h"
 
 unsigned int perf_mem_events__loads_ldlat = 30;
 
@@ -100,6 +102,15 @@ int perf_mem_events__parse(const char *str)
 	return -1;
 }
 
+static bool perf_mem_event__supported(const char *mnt, char *sysfs_name)
+{
+	char path[PATH_MAX];
+	struct stat st;
+
+	scnprintf(path, PATH_MAX, "%s/devices/%s", mnt, sysfs_name);
+	return !stat(path, &st);
+}
+
 int perf_mem_events__init(void)
 {
 	const char *mnt = sysfs__mount();
@@ -110,9 +121,9 @@ int perf_mem_events__init(void)
 		return -ENOENT;
 
 	for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
-		char path[PATH_MAX];
 		struct perf_mem_event *e = perf_mem_events__ptr(j);
-		struct stat st;
+		struct perf_pmu *pmu;
+		char sysfs_name[100];
 
 		/*
 		 * If the event entry isn't valid, skip initialization
@@ -121,11 +132,20 @@ int perf_mem_events__init(void)
 		if (!e->tag)
 			continue;
 
-		scnprintf(path, PATH_MAX, "%s/devices/%s",
-			  mnt, e->sysfs_name);
+		if (!perf_pmu__has_hybrid()) {
+			scnprintf(sysfs_name, sizeof(sysfs_name),
+				  e->sysfs_name, "cpu");
+			e->supported = perf_mem_event__supported(mnt, sysfs_name);
+		} else {
+			perf_pmu__for_each_hybrid_pmu(pmu) {
+				scnprintf(sysfs_name, sizeof(sysfs_name),
+					  e->sysfs_name, pmu->name);
+				e->supported |= perf_mem_event__supported(mnt, sysfs_name);
+			}
+		}
 
-		if (!stat(path, &st))
-			e->supported = found = true;
+		if (e->supported)
+			found = true;
 	}
 
 	return found ? 0 : -ENOENT;
-- 
2.17.1

[PATCH v2 5/8] perf mem: Support record for hybrid platform

[Jin Yao]

Support 'perf mem record' for hybrid platform. On hybrid platform,
such as Alderlake, when executing 'perf mem record', it actually calls:

record -e {cpu_core/mem-loads-aux/,cpu_core/mem-loads,ldlat=30/}:P
       -e cpu_atom/mem-loads,ldlat=30/P
       -e cpu_core/mem-stores/P
       -e cpu_atom/mem-stores/P

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - For hybrid, rec_argc = argc + 9 * perf_pmu__hybrid_pmu_num().
 - Directly 'free(rec_tmp[i])', don't need to check NULL.
  
 tools/perf/builtin-mem.c     | 43 ++++++++++++++----------
 tools/perf/util/mem-events.c | 65 ++++++++++++++++++++++++++++++++++++
 tools/perf/util/mem-events.h |  2 ++
 3 files changed, 93 insertions(+), 17 deletions(-)

diff --git a/tools/perf/builtin-mem.c b/tools/perf/builtin-mem.c
index 03795bf49d51..6b633df458c2 100644
--- a/tools/perf/builtin-mem.c
+++ b/tools/perf/builtin-mem.c
@@ -18,6 +18,8 @@
 #include "util/dso.h"
 #include "util/map.h"
 #include "util/symbol.h"
+#include "util/pmu.h"
+#include "util/pmu-hybrid.h"
 #include <linux/err.h>
 
 #define MEM_OPERATION_LOAD	0x1
@@ -62,8 +64,9 @@ static const char * const *record_mem_usage = __usage;
 
 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 {
-	int rec_argc, i = 0, j;
+	int rec_argc, i = 0, j, tmp_nr = 0;
 	const char **rec_argv;
+	char **rec_tmp;
 	int ret;
 	bool all_user = false, all_kernel = false;
 	struct perf_mem_event *e;
@@ -87,11 +90,24 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 	argc = parse_options(argc, argv, options, record_mem_usage,
 			     PARSE_OPT_KEEP_UNKNOWN);
 
-	rec_argc = argc + 9; /* max number of arguments */
+	if (!perf_pmu__has_hybrid())
+		rec_argc = argc + 9; /* max number of arguments */
+	else
+		rec_argc = argc + 9 * perf_pmu__hybrid_pmu_num();
+
 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 	if (!rec_argv)
 		return -1;
 
+	/*
+	 * Save the allocated event name strings.
+	 */
+	rec_tmp = calloc(rec_argc + 1, sizeof(char *));
+	if (!rec_tmp) {
+		free(rec_argv);
+		return -1;
+	}
+
 	rec_argv[i++] = "record";
 
 	e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
@@ -128,21 +144,9 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 	if (mem->data_page_size)
 		rec_argv[i++] = "--data-page-size";
 
-	for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
-		e = perf_mem_events__ptr(j);
-		if (!e->record)
-			continue;
-
-		if (!e->supported) {
-			pr_err("failed: event '%s' not supported\n",
-			       perf_mem_events__name(j, NULL));
-			free(rec_argv);
-			return -1;
-		}
-
-		rec_argv[i++] = "-e";
-		rec_argv[i++] = perf_mem_events__name(j, NULL);
-	}
+	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
+	if (ret)
+		goto out;
 
 	if (all_user)
 		rec_argv[i++] = "--all-user";
@@ -164,6 +168,11 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 	}
 
 	ret = cmd_record(i, rec_argv);
+out:
+	for (i = 0; i < tmp_nr; i++)
+		free(rec_tmp[i]);
+
+	free(rec_tmp);
 	free(rec_argv);
 	return ret;
 }
diff --git a/tools/perf/util/mem-events.c b/tools/perf/util/mem-events.c
index 69dcac730ada..f38c0da9e698 100644
--- a/tools/perf/util/mem-events.c
+++ b/tools/perf/util/mem-events.c
@@ -166,6 +166,71 @@ void perf_mem_events__list(void)
 	}
 }
 
+static void perf_mem_events__print_unsupport_hybrid(struct perf_mem_event *e,
+						    int idx)
+{
+	const char *mnt = sysfs__mount();
+	char sysfs_name[100];
+	struct perf_pmu *pmu;
+
+	perf_pmu__for_each_hybrid_pmu(pmu) {
+		scnprintf(sysfs_name, sizeof(sysfs_name), e->sysfs_name,
+			  pmu->name);
+		if (!perf_mem_event__supported(mnt, sysfs_name)) {
+			pr_err("failed: event '%s' not supported\n",
+			       perf_mem_events__name(idx, pmu->name));
+		}
+	}
+}
+
+int perf_mem_events__record_args(const char **rec_argv, int *argv_nr,
+				 char **rec_tmp, int *tmp_nr)
+{
+	int i = *argv_nr, k = 0;
+	struct perf_mem_event *e;
+	struct perf_pmu *pmu;
+	char *s;
+
+	for (int j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
+		e = perf_mem_events__ptr(j);
+		if (!e->record)
+			continue;
+
+		if (!perf_pmu__has_hybrid()) {
+			if (!e->supported) {
+				pr_err("failed: event '%s' not supported\n",
+				       perf_mem_events__name(j, NULL));
+				return -1;
+			}
+
+			rec_argv[i++] = "-e";
+			rec_argv[i++] = perf_mem_events__name(j, NULL);
+		} else {
+			if (!e->supported) {
+				perf_mem_events__print_unsupport_hybrid(e, j);
+				return -1;
+			}
+
+			perf_pmu__for_each_hybrid_pmu(pmu) {
+				rec_argv[i++] = "-e";
+				s = perf_mem_events__name(j, pmu->name);
+				if (s) {
+					s = strdup(s);
+					if (!s)
+						return -1;
+
+					rec_argv[i++] = s;
+					rec_tmp[k++] = s;
+				}
+			}
+		}
+	}
+
+	*argv_nr = i;
+	*tmp_nr = k;
+	return 0;
+}
+
 static const char * const tlb_access[] = {
 	"N/A",
 	"HIT",
diff --git a/tools/perf/util/mem-events.h b/tools/perf/util/mem-events.h
index a3fa19093fd2..916242f8020a 100644
--- a/tools/perf/util/mem-events.h
+++ b/tools/perf/util/mem-events.h
@@ -43,6 +43,8 @@ struct perf_mem_event *perf_mem_events__ptr(int i);
 bool is_mem_loads_aux_event(struct evsel *leader);
 
 void perf_mem_events__list(void);
+int perf_mem_events__record_args(const char **rec_argv, int *argv_nr,
+				 char **rec_tmp, int *tmp_nr);
 
 int perf_mem__tlb_scnprintf(char *out, size_t sz, struct mem_info *mem_info);
 int perf_mem__lvl_scnprintf(char *out, size_t sz, struct mem_info *mem_info);
-- 
2.17.1

[PATCH v2 6/8] perf mem: Fix wrong verbose output for recording events

[Jin Yao]

Current code:

for (j = 0; j < argc; j++, i++)
        rec_argv[i] = argv[j];

if (verbose > 0) {
        pr_debug("calling: record ");

        while (rec_argv[j]) {
                pr_debug("%s ", rec_argv[j]);
                j++;
        }
        pr_debug("\n");
}

The entries of argv[] are copied to the end of rec_argv[], not
copied to the beginning of rec_argv[]. So the index j at
rec_argv[] doesn't point to the first event.

Now we record the start index and end index for events in rec_argv[],
and print them if verbose is enabled.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - New in v2.

 tools/perf/builtin-mem.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/tools/perf/builtin-mem.c b/tools/perf/builtin-mem.c
index 6b633df458c2..0fd2a74dbaca 100644
--- a/tools/perf/builtin-mem.c
+++ b/tools/perf/builtin-mem.c
@@ -65,6 +65,7 @@ static const char * const *record_mem_usage = __usage;
 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 {
 	int rec_argc, i = 0, j, tmp_nr = 0;
+	int start, end;
 	const char **rec_argv;
 	char **rec_tmp;
 	int ret;
@@ -144,9 +145,11 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 	if (mem->data_page_size)
 		rec_argv[i++] = "--data-page-size";
 
+	start = i;
 	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
 	if (ret)
 		goto out;
+	end = i;
 
 	if (all_user)
 		rec_argv[i++] = "--all-user";
@@ -160,10 +163,9 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
 	if (verbose > 0) {
 		pr_debug("calling: record ");
 
-		while (rec_argv[j]) {
+		for (j = start; j < end; j++)
 			pr_debug("%s ", rec_argv[j]);
-			j++;
-		}
+
 		pr_debug("\n");
 	}
 
-- 
2.17.1

Re: [PATCH v2 6/8] perf mem: Fix wrong verbose output for recording events

[Jiri Olsa]

On Thu, May 27, 2021 at 08:16:08AM +0800, Jin Yao wrote:
> Current code:
> 
> for (j = 0; j < argc; j++, i++)
>         rec_argv[i] = argv[j];
> 
> if (verbose > 0) {
>         pr_debug("calling: record ");
> 
>         while (rec_argv[j]) {
>                 pr_debug("%s ", rec_argv[j]);
>                 j++;
>         }
>         pr_debug("\n");
> }
> 
> The entries of argv[] are copied to the end of rec_argv[], not
> copied to the beginning of rec_argv[]. So the index j at
> rec_argv[] doesn't point to the first event.
> 
> Now we record the start index and end index for events in rec_argv[],
> and print them if verbose is enabled.
> 
> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
> ---
> v2:
>  - New in v2.
> 
>  tools/perf/builtin-mem.c | 8 +++++---

hi,
do we need the same in c2c as well?

jirka

>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/perf/builtin-mem.c b/tools/perf/builtin-mem.c
> index 6b633df458c2..0fd2a74dbaca 100644
> --- a/tools/perf/builtin-mem.c
> +++ b/tools/perf/builtin-mem.c
> @@ -65,6 +65,7 @@ static const char * const *record_mem_usage = __usage;
>  static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>  {
>  	int rec_argc, i = 0, j, tmp_nr = 0;
> +	int start, end;
>  	const char **rec_argv;
>  	char **rec_tmp;
>  	int ret;
> @@ -144,9 +145,11 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>  	if (mem->data_page_size)
>  		rec_argv[i++] = "--data-page-size";
>  
> +	start = i;
>  	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
>  	if (ret)
>  		goto out;
> +	end = i;
>  
>  	if (all_user)
>  		rec_argv[i++] = "--all-user";
> @@ -160,10 +163,9 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>  	if (verbose > 0) {
>  		pr_debug("calling: record ");
>  
> -		while (rec_argv[j]) {
> +		for (j = start; j < end; j++)
>  			pr_debug("%s ", rec_argv[j]);
> -			j++;
> -		}
> +
>  		pr_debug("\n");
>  	}
>  
> -- 
> 2.17.1
> 

Re: [PATCH v2 6/8] perf mem: Fix wrong verbose output for recording events

[Jin, Yao]

Hi Jiri,

On 6/1/2021 6:42 AM, Jiri Olsa wrote:
> On Thu, May 27, 2021 at 08:16:08AM +0800, Jin Yao wrote:
>> Current code:
>>
>> for (j = 0; j < argc; j++, i++)
>>          rec_argv[i] = argv[j];
>>
>> if (verbose > 0) {
>>          pr_debug("calling: record ");
>>
>>          while (rec_argv[j]) {
>>                  pr_debug("%s ", rec_argv[j]);
>>                  j++;
>>          }
>>          pr_debug("\n");
>> }
>>
>> The entries of argv[] are copied to the end of rec_argv[], not
>> copied to the beginning of rec_argv[]. So the index j at
>> rec_argv[] doesn't point to the first event.
>>
>> Now we record the start index and end index for events in rec_argv[],
>> and print them if verbose is enabled.
>>
>> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
>> ---
>> v2:
>>   - New in v2.
>>
>>   tools/perf/builtin-mem.c | 8 +++++---
> 
> hi,
> do we need the same in c2c as well?
> 
> jirka
> 

perf c2c is a bit different. It sets 'j = 0;' before 'while' loop so it prints all of rec_argv[].

In test,

# perf c2c record -vvv -a -- sleep 1

calling: record -W -d --phys-data --sample-cpu -e 
{cpu_core/mem-loads-aux/,cpu_core/mem-loads,ldlat=30/}:P -e cpu_atom/mem-loads,ldlat=30/P -e 
cpu_core/mem-stores/P -e cpu_atom/mem-stores/P -a sleep 1

The verbose output looks OK.

Thanks
Jin Yao

>>   1 file changed, 5 insertions(+), 3 deletions(-)
>>
>> diff --git a/tools/perf/builtin-mem.c b/tools/perf/builtin-mem.c
>> index 6b633df458c2..0fd2a74dbaca 100644
>> --- a/tools/perf/builtin-mem.c
>> +++ b/tools/perf/builtin-mem.c
>> @@ -65,6 +65,7 @@ static const char * const *record_mem_usage = __usage;
>>   static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>>   {
>>   	int rec_argc, i = 0, j, tmp_nr = 0;
>> +	int start, end;
>>   	const char **rec_argv;
>>   	char **rec_tmp;
>>   	int ret;
>> @@ -144,9 +145,11 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>>   	if (mem->data_page_size)
>>   		rec_argv[i++] = "--data-page-size";
>>   
>> +	start = i;
>>   	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
>>   	if (ret)
>>   		goto out;
>> +	end = i;
>>   
>>   	if (all_user)
>>   		rec_argv[i++] = "--all-user";
>> @@ -160,10 +163,9 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
>>   	if (verbose > 0) {
>>   		pr_debug("calling: record ");
>>   
>> -		while (rec_argv[j]) {
>> +		for (j = start; j < end; j++)
>>   			pr_debug("%s ", rec_argv[j]);
>> -			j++;
>> -		}
>> +
>>   		pr_debug("\n");
>>   	}
>>   
>> -- 
>> 2.17.1
>>
> 

[PATCH v2 7/8] perf mem: Disable 'mem-loads-aux' group before reporting

[Jin Yao]

For some platforms, such as Alderlake, the 'mem-loads' event is required
to use together with 'mem-loads-aux' within a group and 'mem-loads-aux'
must be the group leader. Now we disable this group before reporting
because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
any valid memory load result. If we show the 'mem-loads-aux' +
'mem-loads' as a group in report, it needs many of changes but they
are totally unnecessary.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - New in v2.

 tools/perf/builtin-report.c |  2 ++
 tools/perf/util/evlist.c    | 25 +++++++++++++++++++++++++
 tools/perf/util/evlist.h    |  1 +
 3 files changed, 28 insertions(+)

diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index 36f9ccfeb38a..bc5c393021dc 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -934,6 +934,8 @@ static int __cmd_report(struct report *rep)
 		return ret;
 	}
 
+	evlist__check_mem_load_aux(session->evlist);
+
 	if (rep->stats_mode)
 		return stats_print(rep);
 
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 6ea3e677dc1e..6ba9664089bd 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -2161,3 +2161,28 @@ int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
 
 	return printed;
 }
+
+void evlist__check_mem_load_aux(struct evlist *evlist)
+{
+	struct evsel *leader, *evsel, *pos;
+
+	/*
+	 * For some platforms, the 'mem-loads' event is required to use
+	 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
+	 * must be the group leader. Now we disable this group before reporting
+	 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
+	 * any valid memory load information.
+	 */
+	evlist__for_each_entry(evlist, evsel) {
+		leader = evsel->leader;
+		if (leader == evsel)
+			continue;
+
+		if (leader->name && strstr(leader->name, "mem-loads-aux")) {
+			for_each_group_evsel(pos, leader) {
+				pos->leader = pos;
+				pos->core.nr_members = 0;
+			}
+		}
+	}
+}
diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
index a8b97b50cceb..2073cfa79f79 100644
--- a/tools/perf/util/evlist.h
+++ b/tools/perf/util/evlist.h
@@ -367,4 +367,5 @@ int evlist__ctlfd_ack(struct evlist *evlist);
 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx);
 
 int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf);
+void evlist__check_mem_load_aux(struct evlist *evlist);
 #endif /* __PERF_EVLIST_H */
-- 
2.17.1

[PATCH v2 8/8] perf c2c: Support record for hybrid platform

[Jin Yao]

Support 'perf c2c record' for hybrid platform. On hybrid platform,
such as Alderlake, when executing 'perf c2c record', it actually calls:

record -W -d --phys-data --sample-cpu
-e {cpu_core/mem-loads-aux/,cpu_core/mem-loads,ldlat=30/}:P
-e cpu_atom/mem-loads,ldlat=30/P
-e cpu_core/mem-stores/P
-e cpu_atom/mem-stores/P

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
v2:
 - For hybrid, rec_argc = argc + 11 * perf_pmu__hybrid_pmu_num().
 - Directly 'free(rec_tmp[i])', don't need to check NULL.

 tools/perf/builtin-c2c.c | 40 +++++++++++++++++++++++-----------------
 1 file changed, 23 insertions(+), 17 deletions(-)

diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index a4fd375acdd1..6dea37f141b2 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -42,6 +42,8 @@
 #include "ui/ui.h"
 #include "ui/progress.h"
 #include "../perf.h"
+#include "pmu.h"
+#include "pmu-hybrid.h"
 
 struct c2c_hists {
 	struct hists		hists;
@@ -2907,8 +2909,9 @@ static const char * const *record_mem_usage = __usage_record;
 
 static int perf_c2c__record(int argc, const char **argv)
 {
-	int rec_argc, i = 0, j;
+	int rec_argc, i = 0, j, rec_tmp_nr = 0;
 	const char **rec_argv;
+	char **rec_tmp;
 	int ret;
 	bool all_user = false, all_kernel = false;
 	bool event_set = false;
@@ -2932,11 +2935,21 @@ static int perf_c2c__record(int argc, const char **argv)
 	argc = parse_options(argc, argv, options, record_mem_usage,
 			     PARSE_OPT_KEEP_UNKNOWN);
 
-	rec_argc = argc + 11; /* max number of arguments */
+	if (!perf_pmu__has_hybrid())
+		rec_argc = argc + 11; /* max number of arguments */
+	else
+		rec_argc = argc + 11 * perf_pmu__hybrid_pmu_num();
+
 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 	if (!rec_argv)
 		return -1;
 
+	rec_tmp = calloc(rec_argc + 1, sizeof(char *));
+	if (!rec_tmp) {
+		free(rec_argv);
+		return -1;
+	}
+
 	rec_argv[i++] = "record";
 
 	if (!event_set) {
@@ -2964,21 +2977,9 @@ static int perf_c2c__record(int argc, const char **argv)
 	rec_argv[i++] = "--phys-data";
 	rec_argv[i++] = "--sample-cpu";
 
-	for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
-		e = perf_mem_events__ptr(j);
-		if (!e->record)
-			continue;
-
-		if (!e->supported) {
-			pr_err("failed: event '%s' not supported\n",
-			       perf_mem_events__name(j, NULL));
-			free(rec_argv);
-			return -1;
-		}
-
-		rec_argv[i++] = "-e";
-		rec_argv[i++] = perf_mem_events__name(j, NULL);
-	}
+	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &rec_tmp_nr);
+	if (ret)
+		goto out;
 
 	if (all_user)
 		rec_argv[i++] = "--all-user";
@@ -3002,6 +3003,11 @@ static int perf_c2c__record(int argc, const char **argv)
 	}
 
 	ret = cmd_record(i, rec_argv);
+out:
+	for (i = 0; i < rec_tmp_nr; i++)
+		free(rec_tmp[i]);
+
+	free(rec_tmp);
 	free(rec_argv);
 	return ret;
 }
-- 
2.17.1

Re: [PATCH v2 0/8] perf: Support perf-mem/perf-c2c for AlderLake

[Jiri Olsa]

On Thu, May 27, 2021 at 08:16:02AM +0800, Jin Yao wrote:
> AlderLake uses a hybrid architecture utilizing Golden Cove cores
> (core CPU) and Gracemont cores (atom CPU). This patchset supports
> perf-mem and perf-c2c for AlderLake.
> 
> v2:
> ---
> - Use mem_loads_name__init to keep original behavior for non-hybrid platform.
> - Move x86 specific perf_mem_events[] to arch/x86/util/mem-events.c.
> - Move mem-store event to a new patch.
> - Add a new patch to fix wrong verbose output for recording events
> - Add a new patch to disable 'mem-loads-aux' group before reporting

Acked-by: Jiri Olsa <jolsa@redhat.com>

thanks,
jirka

> 
> Jin Yao (8):
>   perf tools: Check mem-loads auxiliary event
>   perf tools: Support pmu prefix for mem-load event
>   perf tools: Support pmu prefix for mem-store event
>   perf tools: Check if mem_events is supported for hybrid platform
>   perf mem: Support record for hybrid platform
>   perf mem: Fix wrong verbose output for recording events
>   perf mem: Disable 'mem-loads-aux' group before reporting
>   perf c2c: Support record for hybrid platform
> 
>  tools/perf/arch/arm64/util/mem-events.c   |   2 +-
>  tools/perf/arch/powerpc/util/mem-events.c |   2 +-
>  tools/perf/arch/x86/util/mem-events.c     |  54 ++++++++++--
>  tools/perf/builtin-c2c.c                  |  40 +++++----
>  tools/perf/builtin-mem.c                  |  51 ++++++-----
>  tools/perf/builtin-report.c               |   2 +
>  tools/perf/util/evlist.c                  |  25 ++++++
>  tools/perf/util/evlist.h                  |   1 +
>  tools/perf/util/mem-events.c              | 101 ++++++++++++++++++++--
>  tools/perf/util/mem-events.h              |   4 +-
>  10 files changed, 225 insertions(+), 57 deletions(-)
> 
> -- 
> 2.17.1
> 

Re: [PATCH v2 0/8] perf: Support perf-mem/perf-c2c for AlderLake

[Arnaldo Carvalho de Melo]

Em Tue, Jun 01, 2021 at 12:49:34AM +0200, Jiri Olsa escreveu:
> On Thu, May 27, 2021 at 08:16:02AM +0800, Jin Yao wrote:
> > AlderLake uses a hybrid architecture utilizing Golden Cove cores
> > (core CPU) and Gracemont cores (atom CPU). This patchset supports
> > perf-mem and perf-c2c for AlderLake.
> > 
> > v2:
> > ---
> > - Use mem_loads_name__init to keep original behavior for non-hybrid platform.
> > - Move x86 specific perf_mem_events[] to arch/x86/util/mem-events.c.
> > - Move mem-store event to a new patch.
> > - Add a new patch to fix wrong verbose output for recording events
> > - Add a new patch to disable 'mem-loads-aux' group before reporting
> 
> Acked-by: Jiri Olsa <jolsa@redhat.com>

Thanks, applied.

- Arnaldo

 
> thanks,
> jirka
> 
> > 
> > Jin Yao (8):
> >   perf tools: Check mem-loads auxiliary event
> >   perf tools: Support pmu prefix for mem-load event
> >   perf tools: Support pmu prefix for mem-store event
> >   perf tools: Check if mem_events is supported for hybrid platform
> >   perf mem: Support record for hybrid platform
> >   perf mem: Fix wrong verbose output for recording events
> >   perf mem: Disable 'mem-loads-aux' group before reporting
> >   perf c2c: Support record for hybrid platform
> > 
> >  tools/perf/arch/arm64/util/mem-events.c   |   2 +-
> >  tools/perf/arch/powerpc/util/mem-events.c |   2 +-
> >  tools/perf/arch/x86/util/mem-events.c     |  54 ++++++++++--
> >  tools/perf/builtin-c2c.c                  |  40 +++++----
> >  tools/perf/builtin-mem.c                  |  51 ++++++-----
> >  tools/perf/builtin-report.c               |   2 +
> >  tools/perf/util/evlist.c                  |  25 ++++++
> >  tools/perf/util/evlist.h                  |   1 +
> >  tools/perf/util/mem-events.c              | 101 ++++++++++++++++++++--
> >  tools/perf/util/mem-events.h              |   4 +-
> >  10 files changed, 225 insertions(+), 57 deletions(-)
> > 
> > -- 
> > 2.17.1
> > 
> 

-- 

- Arnaldo