Skip to contentSkip to conversation

perf stat: Fix wrong skipping for per-die aggregation

PATCH Superseded Jin Yao 2 participants 3 messages Dec 16, 2020 – Dec 24, 2020 perf events

v1 of 9 Superseded Latest is v9 →

lore
Share & tools
lore

Jin Yao

author
#0 lore raw

PATCH 1/1 perf stat: Fix wrong skipping for per-die aggregation

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
     1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
     1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
     1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
     2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
     2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
     2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
     2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
     3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
     3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
     3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
     3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
     4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
     4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
     4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
     4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
     5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
     5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
     5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
     5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use a pair of die_id + socket_id to decide if this cpu counts needs to skip. Only considering socket_id is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001622560 S0-D0           1            1210176 Bytes llc_misses.mem_read
     1.001622560 S0-D1           1             950208 Bytes llc_misses.mem_read
     1.001622560 S1-D0           1             968704 Bytes llc_misses.mem_read
     1.001622560 S1-D1           1            1113280 Bytes llc_misses.mem_read
     2.004399430 S0-D0           1            1258560 Bytes llc_misses.mem_read
     2.004399430 S0-D1           1            1011008 Bytes llc_misses.mem_read
     2.004399430 S1-D0           1            1036544 Bytes llc_misses.mem_read
     2.004399430 S1-D1           1            1237376 Bytes llc_misses.mem_read
     3.006535657 S0-D0           1            1036608 Bytes llc_misses.mem_read
     3.006535657 S0-D1           1             906560 Bytes llc_misses.mem_read
     3.006535657 S1-D0           1             994112 Bytes llc_misses.mem_read
     3.006535657 S1-D1           1            1025472 Bytes llc_misses.mem_read
     4.008934402 S0-D0           1            1005120 Bytes llc_misses.mem_read
     4.008934402 S0-D1           1             882368 Bytes llc_misses.mem_read
     4.008934402 S1-D0           1             961408 Bytes llc_misses.mem_read
     4.008934402 S1-D1           1             988352 Bytes llc_misses.mem_read
     5.001605406 S0-D0           1            1930240 Bytes llc_misses.mem_read
     5.001605406 S0-D1           1            1134272 Bytes llc_misses.mem_read
     5.001605406 S1-D0           1            1531136 Bytes llc_misses.mem_read
     5.001605406 S1-D1           1            2230848 Bytes llc_misses.mem_read

On no-die system, die_id is 0, this patch keeps original behavior unchanged.

---
 tools/perf/util/cpumap.h |  2 ++
 tools/perf/util/stat.c   | 21 +++++++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)
tools/perf/util/cpumap.h +2 -0
index 3a442f021468..58604c047121 100644
@@ -7,6 +7,8 @@
 #include <internal/cpumap.h>
 #include <perf/cpumap.h>
 
+#define CPU_MAP_MAX_DIE	256
+
 struct perf_record_cpu_map_data;
 
 struct perf_cpu_map *perf_cpu_map__empty_new(int nr);
tools/perf/util/stat.c +17 -4
index 1e125e39ff84..b3fc6e261fe0 100644
@@ -282,9 +282,9 @@ static void zero_per_pkg(struct evsel *counter)
 static int check_per_pkg(struct evsel *counter,
 			 struct perf_counts_values *vals, int cpu, bool *skip)
 {
-	unsigned long *mask = counter->per_pkg_mask;
+	unsigned long *mask = counter->per_pkg_mask, *l;
 	struct perf_cpu_map *cpus = evsel__cpus(counter);
-	int s;
+	int s, d;
 
 	*skip = false;
 
@@ -295,7 +295,7 @@ static int check_per_pkg(struct evsel *counter,
 		return 0;
 
 	if (!mask) {
-		mask = zalloc(cpu__max_cpu());
+		mask = zalloc(cpu__max_cpu() * CPU_MAP_MAX_DIE);
 		if (!mask)
 			return -ENOMEM;
 
@@ -317,7 +317,20 @@ static int check_per_pkg(struct evsel *counter,
 	if (s < 0)
 		return -1;
 
-	*skip = test_and_set_bit(s, mask) == 1;
+	d = cpu_map__get_die(cpus, cpu, NULL);
+	if (d < 0)
+		return -1;
+
+	/*
+	 * On multi-die system, die_id < 256. We use a pair
+	 * of socket + die to identify the used bit.
+	 * On no-die system, die_id is 0, l = &mask[0]. It doesn't
+	 * change original "test_and_set_bit(s, mask)" behavior.
+	 */
+	d &= 0xff;
+	l = &mask[d];
+
+	*skip = test_and_set_bit(s, l) == 1;
 	return 0;
 }
 
Signature · 1 line
2.17.1
#1 lore raw
Quoted 124 lines · Jin Yao

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
     1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
     1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
     1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
     2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
     2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
     2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
     2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
     3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
     3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
     3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
     3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
     4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
     4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
     4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
     4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
     5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
     5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
     5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
     5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use a pair of die_id + socket_id to decide if this cpu counts needs to skip. Only considering socket_id is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001622560 S0-D0           1            1210176 Bytes llc_misses.mem_read
     1.001622560 S0-D1           1             950208 Bytes llc_misses.mem_read
     1.001622560 S1-D0           1             968704 Bytes llc_misses.mem_read
     1.001622560 S1-D1           1            1113280 Bytes llc_misses.mem_read
     2.004399430 S0-D0           1            1258560 Bytes llc_misses.mem_read
     2.004399430 S0-D1           1            1011008 Bytes llc_misses.mem_read
     2.004399430 S1-D0           1            1036544 Bytes llc_misses.mem_read
     2.004399430 S1-D1           1            1237376 Bytes llc_misses.mem_read
     3.006535657 S0-D0           1            1036608 Bytes llc_misses.mem_read
     3.006535657 S0-D1           1             906560 Bytes llc_misses.mem_read
     3.006535657 S1-D0           1             994112 Bytes llc_misses.mem_read
     3.006535657 S1-D1           1            1025472 Bytes llc_misses.mem_read
     4.008934402 S0-D0           1            1005120 Bytes llc_misses.mem_read
     4.008934402 S0-D1           1             882368 Bytes llc_misses.mem_read
     4.008934402 S1-D0           1             961408 Bytes llc_misses.mem_read
     4.008934402 S1-D1           1             988352 Bytes llc_misses.mem_read
     5.001605406 S0-D0           1            1930240 Bytes llc_misses.mem_read
     5.001605406 S0-D1           1            1134272 Bytes llc_misses.mem_read
     5.001605406 S1-D0           1            1531136 Bytes llc_misses.mem_read
     5.001605406 S1-D1           1            2230848 Bytes llc_misses.mem_read

On no-die system, die_id is 0, this patch keeps original behavior unchanged.

---
 tools/perf/util/cpumap.h |  2 ++
 tools/perf/util/stat.c   | 21 +++++++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)
tools/perf/util/cpumap.h +2 -0
index 3a442f021468..58604c047121 100644
@@ -7,6 +7,8 @@
 #include <internal/cpumap.h>
 #include <perf/cpumap.h>
 
+#define CPU_MAP_MAX_DIE	256
+
 struct perf_record_cpu_map_data;
 
 struct perf_cpu_map *perf_cpu_map__empty_new(int nr);
tools/perf/util/stat.c +5 -4
index 1e125e39ff84..b3fc6e261fe0 100644
@@ -282,9 +282,9 @@ static void zero_per_pkg(struct evsel *counter)
 static int check_per_pkg(struct evsel *counter,
 			 struct perf_counts_values *vals, int cpu, bool *skip)
 {
-	unsigned long *mask = counter->per_pkg_mask;
+	unsigned long *mask = counter->per_pkg_mask, *l;
 	struct perf_cpu_map *cpus = evsel__cpus(counter);
-	int s;
+	int s, d;
 
 	*skip = false;
 
@@ -295,7 +295,7 @@ static int check_per_pkg(struct evsel *counter,
 		return 0;
 
 	if (!mask) {
-		mask = zalloc(cpu__max_cpu());
+		mask = zalloc(cpu__max_cpu() * CPU_MAP_MAX_DIE);
 		if (!mask)
 			return -ENOMEM;
 
@@ -317,7 +317,20 @@ static int check_per_pkg(struct evsel *counter,
 	if (s < 0)
 		return -1;
 
-	*skip = test_and_set_bit(s, mask) == 1;
+	d = cpu_map__get_die(cpus, cpu, NULL);
+	if (d < 0)

tools/perf/util/stat.c static int check_per_pkg(struct evsel *counter, · line 331

Quoted 10 lines
tools/perf/util/stat.c +10 -0
+		return -1;
+
+	/*
+	 * On multi-die system, die_id < 256. We use a pair
+	 * of socket + die to identify the used bit.
+	 * On no-die system, die_id is 0, l = &mask[0]. It doesn't
+	 * change original "test_and_set_bit(s, mask)" behavior.
+	 */
+	d &= 0xff;
+	l = &mask[d];

James is fixing the aggregation, because we could get large socket numbers:
https://lore.kernel.org/lkml/[email protected]/

I think we should fix that in here as well
maybe just use hashmap (socket,die) -> bool ?

jirka

Quoted 8 lines
tools/perf/util/stat.c +2 -0
+
+	*skip = test_and_set_bit(s, l) == 1;
 	return 0;
 }
 
Signature · 1 line
2.17.1
author
#2 lore raw
Quoted 137 lines · Jiri Olsa
Quoted 124 lines · Jin Yao

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

      1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
      1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
      1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
      1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
      2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
      2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
      2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
      2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
      3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
      3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
      3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
      3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
      4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
      4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
      4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
      4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
      5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
      5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
      5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
      5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval
values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use a pair of die_id + socket_id to decide if this cpu counts needs to skip. Only considering socket_id is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

      1.001622560 S0-D0           1            1210176 Bytes llc_misses.mem_read
      1.001622560 S0-D1           1             950208 Bytes llc_misses.mem_read
      1.001622560 S1-D0           1             968704 Bytes llc_misses.mem_read
      1.001622560 S1-D1           1            1113280 Bytes llc_misses.mem_read
      2.004399430 S0-D0           1            1258560 Bytes llc_misses.mem_read
      2.004399430 S0-D1           1            1011008 Bytes llc_misses.mem_read
      2.004399430 S1-D0           1            1036544 Bytes llc_misses.mem_read
      2.004399430 S1-D1           1            1237376 Bytes llc_misses.mem_read
      3.006535657 S0-D0           1            1036608 Bytes llc_misses.mem_read
      3.006535657 S0-D1           1             906560 Bytes llc_misses.mem_read
      3.006535657 S1-D0           1             994112 Bytes llc_misses.mem_read
      3.006535657 S1-D1           1            1025472 Bytes llc_misses.mem_read
      4.008934402 S0-D0           1            1005120 Bytes llc_misses.mem_read
      4.008934402 S0-D1           1             882368 Bytes llc_misses.mem_read
      4.008934402 S1-D0           1             961408 Bytes llc_misses.mem_read
      4.008934402 S1-D1           1             988352 Bytes llc_misses.mem_read
      5.001605406 S0-D0           1            1930240 Bytes llc_misses.mem_read
      5.001605406 S0-D1           1            1134272 Bytes llc_misses.mem_read
      5.001605406 S1-D0           1            1531136 Bytes llc_misses.mem_read
      5.001605406 S1-D1           1            2230848 Bytes llc_misses.mem_read

On no-die system, die_id is 0, this patch keeps original behavior unchanged.

---
  tools/perf/util/cpumap.h |  2 ++
  tools/perf/util/stat.c   | 21 +++++++++++++++++----
  2 files changed, 19 insertions(+), 4 deletions(-)
tools/perf/util/cpumap.h +2 -0
index 3a442f021468..58604c047121 100644
@@ -7,6 +7,8 @@
  #include <internal/cpumap.h>
  #include <perf/cpumap.h>
 
+#define CPU_MAP_MAX_DIE	256
+
  struct perf_record_cpu_map_data;
 
  struct perf_cpu_map *perf_cpu_map__empty_new(int nr);
tools/perf/util/stat.c +5 -4
index 1e125e39ff84..b3fc6e261fe0 100644
@@ -282,9 +282,9 @@ static void zero_per_pkg(struct evsel *counter)
  static int check_per_pkg(struct evsel *counter,
  			 struct perf_counts_values *vals, int cpu, bool *skip)
  {
-	unsigned long *mask = counter->per_pkg_mask;
+	unsigned long *mask = counter->per_pkg_mask, *l;
  	struct perf_cpu_map *cpus = evsel__cpus(counter);
-	int s;
+	int s, d;
 
  	*skip = false;
 
@@ -295,7 +295,7 @@ static int check_per_pkg(struct evsel *counter,
  		return 0;
 
  	if (!mask) {
-		mask = zalloc(cpu__max_cpu());
+		mask = zalloc(cpu__max_cpu() * CPU_MAP_MAX_DIE);
  		if (!mask)
  			return -ENOMEM;
 
@@ -317,7 +317,20 @@ static int check_per_pkg(struct evsel *counter,
  	if (s < 0)
  		return -1;
 
-	*skip = test_and_set_bit(s, mask) == 1;
+	d = cpu_map__get_die(cpus, cpu, NULL);
+	if (d < 0)

tools/perf/util/stat.c static int check_per_pkg(struct evsel *counter, · line 331

Quoted 10 lines
tools/perf/util/stat.c +10 -0
+		return -1;
+
+	/*
+	 * On multi-die system, die_id < 256. We use a pair
+	 * of socket + die to identify the used bit.
+	 * On no-die system, die_id is 0, l = &mask[0]. It doesn't
+	 * change original "test_and_set_bit(s, mask)" behavior.
+	 */
+	d &= 0xff;
+	l = &mask[d];

James is fixing the aggregation, because we could get large socket numbers:

Quoted 2 lines
   https://lore.kernel.org/lkml/[email protected]/

James is using "struct cpu_aggr_map" to replace "struct perf_cpu_map", but do we also need to replace the "struct perf_cpu_map" in "struct perf_evsel"? Otherwise the evsel->cpus will be inconsistent, right?

struct perf_evsel {
	...
	struct perf_cpu_map	*cpus;
	struct perf_cpu_map	*own_cpus;
	...
};
Quoted 1 line

I think we should fix that in here as well

Yes, we need to fix that in check_per_pkg() as well. As I said above, in check_per_pkg(),

struct perf_cpu_map *cpus = evsel__cpus(counter);

we should use "struct cpu_aggr_map" here.

Quoted 2 lines

maybe just use hashmap (socket,die) -> bool ?

Yes, hashmap should be better, saving more memory.

Thanks
Jin Yao

Quoted 11 lines

jirka

Quoted 8 lines
tools/perf/util/stat.c +2 -0
+
+	*skip = test_and_set_bit(s, l) == 1;
  	return 0;
  }
 
Signature · 1 line
2.17.1