2021-01-05 02:40:56

by Jin Yao

[permalink] [raw]
Subject: [PATCH v4] 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 hashmap(socket,die) to decide if
the cpu counts needs to skip. Only considering socket 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
1.001617187 S0-D0 1 952768 Bytes llc_misses.mem_read
1.001617187 S0-D1 1 849152 Bytes llc_misses.mem_read
1.001617187 S1-D0 1 854912 Bytes llc_misses.mem_read
1.001617187 S1-D1 1 948672 Bytes llc_misses.mem_read
2.004138021 S0-D0 1 1200256 Bytes llc_misses.mem_read
2.004138021 S0-D1 1 899840 Bytes llc_misses.mem_read
2.004138021 S1-D0 1 898752 Bytes llc_misses.mem_read
2.004138021 S1-D1 1 1039424 Bytes llc_misses.mem_read
3.006463971 S0-D0 1 1029056 Bytes llc_misses.mem_read
3.006463971 S0-D1 1 894336 Bytes llc_misses.mem_read
3.006463971 S1-D0 1 905088 Bytes llc_misses.mem_read
3.006463971 S1-D1 1 1042880 Bytes llc_misses.mem_read
4.008831441 S0-D0 1 920576 Bytes llc_misses.mem_read
4.008831441 S0-D1 1 877248 Bytes llc_misses.mem_read
4.008831441 S1-D0 1 883968 Bytes llc_misses.mem_read
4.008831441 S1-D1 1 967168 Bytes llc_misses.mem_read
5.001441349 S0-D0 1 864704 Bytes llc_misses.mem_read
5.001441349 S0-D1 1 827456 Bytes llc_misses.mem_read
5.001441349 S1-D0 1 829632 Bytes llc_misses.mem_read
5.001441349 S1-D1 1 925760 Bytes llc_misses.mem_read

On no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior
is not changed.

Reported-by: Huang Ying <[email protected]>
Signed-off-by: Jin Yao <[email protected]>
---
v4:
v3 used unnecessary bool allocatioin. v4 just uses the hash value '(void *)1'.

v4 is compiled ok with tmp.perf/core.

v3:
Since for some cpumap functions, the return type is changed from 'int' to
'struct aggr_cpu_id', the patch needs to be updated as well.

before:
d = cpu_map__get_die()

after:
d = cpu_map__get_die().die

v3 is compiled ok with tmp.perf/core.

v2:
Use hashmap to check the used socket+die pair.

tools/perf/util/evsel.c | 4 +++-
tools/perf/util/evsel.h | 3 ++-
tools/perf/util/stat.c | 41 ++++++++++++++++++++++++++++++++++++-----
3 files changed, 41 insertions(+), 7 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index c26ea82220bd..9715ed9b03f6 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -45,6 +45,7 @@
#include "string2.h"
#include "memswap.h"
#include "util.h"
+#include "hashmap.h"
#include "../perf-sys.h"
#include "util/parse-branch-options.h"
#include <internal/xyarray.h>
@@ -1377,7 +1378,8 @@ void evsel__exit(struct evsel *evsel)
zfree(&evsel->group_name);
zfree(&evsel->name);
zfree(&evsel->pmu_name);
- zfree(&evsel->per_pkg_mask);
+ hashmap__free(evsel->per_pkg_mask);
+ evsel->per_pkg_mask = NULL;
zfree(&evsel->metric_events);
perf_evsel__object.fini(evsel);
}
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index cd1d8dd43199..951628943fd0 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -17,6 +17,7 @@ struct cgroup;
struct perf_counts;
struct perf_stat_evsel;
union perf_event;
+struct hashmap;

typedef int (evsel__sb_cb_t)(union perf_event *event, void *data);

@@ -110,7 +111,7 @@ struct evsel {
bool merged_stat;
bool reset_group;
bool errored;
- unsigned long *per_pkg_mask;
+ struct hashmap *per_pkg_mask;
struct evsel *leader;
struct list_head config_terms;
int err;
diff --git a/tools/perf/util/stat.c b/tools/perf/util/stat.c
index 8ce1479c98f0..a658e0ffaf2a 100644
--- a/tools/perf/util/stat.c
+++ b/tools/perf/util/stat.c
@@ -13,6 +13,7 @@
#include "evlist.h"
#include "evsel.h"
#include "thread_map.h"
+#include "hashmap.h"
#include <linux/zalloc.h>

void update_stats(struct stats *stats, u64 val)
@@ -276,15 +277,29 @@ void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
static void zero_per_pkg(struct evsel *counter)
{
if (counter->per_pkg_mask)
- memset(counter->per_pkg_mask, 0, cpu__max_cpu());
+ hashmap__clear(counter->per_pkg_mask);
+}
+
+static size_t pkg_id_hash(const void *key, void *ctx __maybe_unused)
+{
+ int socket = (int64_t)key >> 32;
+
+ return socket;
+}
+
+static bool pkg_id_equal(const void *key1, const void *key2,
+ void *ctx __maybe_unused)
+{
+ return (int64_t)key1 == (int64_t)key2;
}

static int check_per_pkg(struct evsel *counter,
struct perf_counts_values *vals, int cpu, bool *skip)
{
- unsigned long *mask = counter->per_pkg_mask;
+ struct hashmap *mask = counter->per_pkg_mask;
struct perf_cpu_map *cpus = evsel__cpus(counter);
- int s;
+ int s, d, ret;
+ uint64_t key;

*skip = false;

@@ -295,7 +310,7 @@ static int check_per_pkg(struct evsel *counter,
return 0;

if (!mask) {
- mask = zalloc(cpu__max_cpu());
+ mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
if (!mask)
return -ENOMEM;

@@ -317,7 +332,23 @@ static int check_per_pkg(struct evsel *counter,
if (s < 0)
return -1;

- *skip = test_and_set_bit(s, mask) == 1;
+ /*
+ * On multi-die system, 0 < die_id < 256. On no-die system, die_id = 0.
+ * We use hashmap(socket, die) to check the used socket+die pair.
+ */
+ d = cpu_map__get_die(cpus, cpu, NULL).die;
+ if (d < 0)
+ return -1;
+
+ key = (uint64_t)s << 32 | (d & 0xff);
+ if (hashmap__find(mask, (void *)key, NULL)) {
+ *skip = true;
+ } else {
+ ret = hashmap__add(mask, (void *)key, (void *)1);
+ if (ret)
+ return -1;
+ }
+
return 0;
}

--
2.17.1


2021-01-13 01:29:41

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH v4] perf stat: Fix wrong skipping for per-die aggregation

On Tue, Jan 05, 2021 at 10:36:15AM +0800, Jin Yao wrote:

SNIP

> diff --git a/tools/perf/util/stat.c b/tools/perf/util/stat.c
> index 8ce1479c98f0..a658e0ffaf2a 100644
> --- a/tools/perf/util/stat.c
> +++ b/tools/perf/util/stat.c
> @@ -13,6 +13,7 @@
> #include "evlist.h"
> #include "evsel.h"
> #include "thread_map.h"
> +#include "hashmap.h"
> #include <linux/zalloc.h>
>
> void update_stats(struct stats *stats, u64 val)
> @@ -276,15 +277,29 @@ void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
> static void zero_per_pkg(struct evsel *counter)
> {
> if (counter->per_pkg_mask)
> - memset(counter->per_pkg_mask, 0, cpu__max_cpu());
> + hashmap__clear(counter->per_pkg_mask);
> +}
> +
> +static size_t pkg_id_hash(const void *key, void *ctx __maybe_unused)
> +{
> + int socket = (int64_t)key >> 32;
> +
> + return socket;

could be just simple return right, why the variable?

> +}
> +
> +static bool pkg_id_equal(const void *key1, const void *key2,
> + void *ctx __maybe_unused)
> +{
> + return (int64_t)key1 == (int64_t)key2;
> }
>
> static int check_per_pkg(struct evsel *counter,
> struct perf_counts_values *vals, int cpu, bool *skip)
> {
> - unsigned long *mask = counter->per_pkg_mask;
> + struct hashmap *mask = counter->per_pkg_mask;
> struct perf_cpu_map *cpus = evsel__cpus(counter);
> - int s;
> + int s, d, ret;
> + uint64_t key;
>
> *skip = false;
>
> @@ -295,7 +310,7 @@ static int check_per_pkg(struct evsel *counter,
> return 0;
>
> if (!mask) {
> - mask = zalloc(cpu__max_cpu());
> + mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
> if (!mask)
> return -ENOMEM;
>
> @@ -317,7 +332,23 @@ static int check_per_pkg(struct evsel *counter,
> if (s < 0)
> return -1;
>
> - *skip = test_and_set_bit(s, mask) == 1;
> + /*
> + * On multi-die system, 0 < die_id < 256. On no-die system, die_id = 0.
> + * We use hashmap(socket, die) to check the used socket+die pair.
> + */
> + d = cpu_map__get_die(cpus, cpu, NULL).die;
> + if (d < 0)
> + return -1;
> +
> + key = (uint64_t)s << 32 | (d & 0xff);

why the masking for d? also I thought it could be bigger than 256

would it better to do it the other way around?

die << 32 | socket


> + if (hashmap__find(mask, (void *)key, NULL)) {
> + *skip = true;
> + } else {
> + ret = hashmap__add(mask, (void *)key, (void *)1);
> + if (ret)
> + return -1;

you could 'return ret' at the end and skip this extra check

thanks,
jirka

> + }
> +
> return 0;
> }
>
> --
> 2.17.1
>

2021-01-13 03:35:14

by Jin Yao

[permalink] [raw]
Subject: Re: [PATCH v4] perf stat: Fix wrong skipping for per-die aggregation

Hi Jiri,

On 1/12/2021 6:04 PM, Jiri Olsa wrote:
> On Tue, Jan 05, 2021 at 10:36:15AM +0800, Jin Yao wrote:
>
> SNIP
>
>> diff --git a/tools/perf/util/stat.c b/tools/perf/util/stat.c
>> index 8ce1479c98f0..a658e0ffaf2a 100644
>> --- a/tools/perf/util/stat.c
>> +++ b/tools/perf/util/stat.c
>> @@ -13,6 +13,7 @@
>> #include "evlist.h"
>> #include "evsel.h"
>> #include "thread_map.h"
>> +#include "hashmap.h"
>> #include <linux/zalloc.h>
>>
>> void update_stats(struct stats *stats, u64 val)
>> @@ -276,15 +277,29 @@ void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
>> static void zero_per_pkg(struct evsel *counter)
>> {
>> if (counter->per_pkg_mask)
>> - memset(counter->per_pkg_mask, 0, cpu__max_cpu());
>> + hashmap__clear(counter->per_pkg_mask);
>> +}
>> +
>> +static size_t pkg_id_hash(const void *key, void *ctx __maybe_unused)
>> +{
>> + int socket = (int64_t)key >> 32;
>> +
>> + return socket;
>
> could be just simple return right, why the variable?
>

Sure, just simple return should be OK.

>> +}
>> +
>> +static bool pkg_id_equal(const void *key1, const void *key2,
>> + void *ctx __maybe_unused)
>> +{
>> + return (int64_t)key1 == (int64_t)key2;
>> }
>>
>> static int check_per_pkg(struct evsel *counter,
>> struct perf_counts_values *vals, int cpu, bool *skip)
>> {
>> - unsigned long *mask = counter->per_pkg_mask;
>> + struct hashmap *mask = counter->per_pkg_mask;
>> struct perf_cpu_map *cpus = evsel__cpus(counter);
>> - int s;
>> + int s, d, ret;
>> + uint64_t key;
>>
>> *skip = false;
>>
>> @@ -295,7 +310,7 @@ static int check_per_pkg(struct evsel *counter,
>> return 0;
>>
>> if (!mask) {
>> - mask = zalloc(cpu__max_cpu());
>> + mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
>> if (!mask)
>> return -ENOMEM;
>>
>> @@ -317,7 +332,23 @@ static int check_per_pkg(struct evsel *counter,
>> if (s < 0)
>> return -1;
>>
>> - *skip = test_and_set_bit(s, mask) == 1;
>> + /*
>> + * On multi-die system, 0 < die_id < 256. On no-die system, die_id = 0.
>> + * We use hashmap(socket, die) to check the used socket+die pair.
>> + */
>> + d = cpu_map__get_die(cpus, cpu, NULL).die;
>> + if (d < 0)
>> + return -1;
>> +
>> + key = (uint64_t)s << 32 | (d & 0xff);
>
> why the masking for d? also I thought it could be bigger than 256
>

I remember in previous perf tool it would show a warning "The die id number is too big" if die_id >
256 and cpu_map__get_die() returned -1. See commit db5742b6849e ("perf stat: Support per-die
aggregation").

But now the die_id checking is removed.

> would it better to do it the other way around?
>
> die << 32 | socket
>

I think this way is OK, thanks! I will use it in v5.

>
>> + if (hashmap__find(mask, (void *)key, NULL)) {
>> + *skip = true;
>> + } else {
>> + ret = hashmap__add(mask, (void *)key, (void *)1);
>> + if (ret)
>> + return -1;
>
> you could 'return ret' at the end and skip this extra check
>

Yes, define int ret = 0 and just 'return ret' at the end of function.

Thanks
Jin Yao

> thanks,
> jirka
>
>> + }
>> +
>> return 0;
>> }
>>
>> --
>> 2.17.1
>>
>