lkml.org 
[lkml]   [2020]   [Dec]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] perf stat: Fix wrong skipping for per-die aggregation
From
Date


On 12/24/2020 6:35 AM, Jiri Olsa wrote:
> On Wed, Dec 16, 2020 at 03:01:46PM +0800, Jin Yao wrote:
>> 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.
>>
>> Reported-by: Huang Ying <ying.huang@intel.com>
>> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
>> ---
>> tools/perf/util/cpumap.h | 2 ++
>> tools/perf/util/stat.c | 21 +++++++++++++++++----
>> 2 files changed, 19 insertions(+), 4 deletions(-)
>>
>> diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
>> index 3a442f021468..58604c047121 100644
>> --- a/tools/perf/util/cpumap.h
>> +++ b/tools/perf/util/cpumap.h
>> @@ -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);
>> diff --git a/tools/perf/util/stat.c b/tools/perf/util/stat.c
>> index 1e125e39ff84..b3fc6e261fe0 100644
>> --- a/tools/perf/util/stat.c
>> +++ b/tools/perf/util/stat.c
>> @@ -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];
>
> James is fixing the aggregation, because we could get large socket numbers:
> https://lore.kernel.org/lkml/20201126141328.6509-1-james.clark@arm.com/
>

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;
...
};

> 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.

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

Yes, hashmap should be better, saving more memory.

Thanks
Jin Yao

> jirka
>
>> +
>> + *skip = test_and_set_bit(s, l) == 1;
>> return 0;
>> }
>>
>> --
>> 2.17.1
>>
>

\
 
 \ /
  Last update: 2020-12-24 02:27    [W:0.032 / U:0.204 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site