perf stat: Fix wrong skipping for per-die aggregation
PATCH v9 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_readFrom 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
1.001586691 S0-D0 1 1229440 Bytes llc_misses.mem_read
1.001586691 S0-D1 1 976832 Bytes llc_misses.mem_read
1.001586691 S1-D0 1 938304 Bytes llc_misses.mem_read
1.001586691 S1-D1 1 1227328 Bytes llc_misses.mem_read
2.003776312 S0-D0 1 1586752 Bytes llc_misses.mem_read
2.003776312 S0-D1 1 875392 Bytes llc_misses.mem_read
2.003776312 S1-D0 1 855616 Bytes llc_misses.mem_read
2.003776312 S1-D1 1 949376 Bytes llc_misses.mem_read
3.006512788 S0-D0 1 1338880 Bytes llc_misses.mem_read
3.006512788 S0-D1 1 920064 Bytes llc_misses.mem_read
3.006512788 S1-D0 1 877184 Bytes llc_misses.mem_read
3.006512788 S1-D1 1 1020736 Bytes llc_misses.mem_read
4.008895291 S0-D0 1 926592 Bytes llc_misses.mem_read
4.008895291 S0-D1 1 906368 Bytes llc_misses.mem_read
4.008895291 S1-D0 1 892224 Bytes llc_misses.mem_read
4.008895291 S1-D1 1 987712 Bytes llc_misses.mem_read
5.001590993 S0-D0 1 962624 Bytes llc_misses.mem_read
5.001590993 S0-D1 1 912512 Bytes llc_misses.mem_read
5.001590993 S1-D0 1 891200 Bytes llc_misses.mem_read
5.001590993 S1-D1 1 978432 Bytes llc_misses.mem_readOn no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior is not changed.
---
v9:
Rename zero_per_pkg to evsel__zero_per_pkg and move it to evsel.c. Then evsel__zero_per_pkg can be called under different code path.
Call evsel__zero_per_pkg in evsel__exit().
v8:
Jiri contributes idea and code to allocate 'uint64_t' type hash key for better supporting for 64 bits platform and 32 bits platform. I merge Jiri's code in this patch. Thanks Jiri!
Rebase the patch to latest perf/core branch.
v7:
It reported build error on 32-bit system (such as cross build by mipsel-linux-gnu-gcc).
In v7,
1. Use size_t to replace uint64_t.
2. The hash key is changed from 'die_id << 32 | socket_id' to 'die_id << 16 | socket_id', 16 bits is enough for socket id , right?
v6:
Fix the perf test python failure by adding hashmap.c to python-ext-sources.
root@kbl-ppc:~# ./perf test python 19: 'import perf' in python : Ok
v5:
Hash key is changed to die_id << 32 | socket.
In pkg_id_hash, return (int64_t)key & 0xffffffff; actually it's socket.
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 | 18 +++++++++++- tools/perf/util/evsel.h | 4 ++- tools/perf/util/python-ext-sources | 1 + tools/perf/util/stat.c | 47 ++++++++++++++++++++++++------ 4 files changed, 59 insertions(+), 11 deletions(-)
tools/perf/util/evsel.c +17 -1
@@ -46,6 +46,7 @@#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>
@@ -1385,7 +1386,9 @@ void evsel__exit(struct evsel *evsel)zfree (&evsel->group_name);zfree (&evsel->name);zfree (&evsel->pmu_name); -zfree (&evsel->per_pkg_mask); +evsel__zero_per_pkg (evsel); +hashmap__free (evsel->per_pkg_mask); + evsel->per_pkg_mask =NULL ;zfree (&evsel->metric_events); perf_evsel__object.fini (evsel); }
@@ -2757,3 +2760,16 @@ int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)return store_evsel_ids (evsel, evlist); } + +void evsel__zero_per_pkg (struct evsel *evsel) +{ +struct hashmap_entry *cur; +size_t bkt; + +if (evsel->per_pkg_mask) { +hashmap__for_each_entry (evsel->per_pkg_mask, cur, bkt) +free ((char *)cur->key); + +hashmap__clear (evsel->per_pkg_mask); + } +}
tools/perf/util/evsel.h +3 -1
@@ -19,6 +19,7 @@ struct perf_stat_evsel;union perf_event ;struct bpf_counter_ops ;struct target ; +struct hashmap ;typedef int (evsel__sb_cb_t )(union perf_event *event,void *data);
@@ -112,7 +113,7 @@ 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;
@@ -430,4 +431,5 @@ struct perf_env *evsel__env(struct evsel *evsel);int evsel__store_ids (struct evsel *evsel,struct evlist *evlist); +void evsel__zero_per_pkg (struct evsel *evsel);#endif /* __PERF_EVSEL_H */
tools/perf/util/python-ext-sources +1 -0
@@ -36,3 +36,4 @@ util/symbol_fprintf.c util/units.c util/affinity.c util/rwsem.c +util/hashmap.c
tools/perf/util/stat.c +38 -9
84 lines not on this page · raw message
Signature · 1 line
2.17.1
Quoted 72 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-D1But 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 checkdie_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_readFrom 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 usehashmap(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
1.001586691 S0-D0 1 1229440 Bytes llc_misses.mem_read 1.001586691 S0-D1 1 976832 Bytes llc_misses.mem_read 1.001586691 S1-D0 1 938304 Bytes llc_misses.mem_read 1.001586691 S1-D1 1 1227328 Bytes llc_misses.mem_read 2.003776312 S0-D0 1 1586752 Bytes llc_misses.mem_read 2.003776312 S0-D1 1 875392 Bytes llc_misses.mem_read 2.003776312 S1-D0 1 855616 Bytes llc_misses.mem_read 2.003776312 S1-D1 1 949376 Bytes llc_misses.mem_read 3.006512788 S0-D0 1 1338880 Bytes llc_misses.mem_read 3.006512788 S0-D1 1 920064 Bytes llc_misses.mem_read 3.006512788 S1-D0 1 877184 Bytes llc_misses.mem_read 3.006512788 S1-D1 1 1020736 Bytes llc_misses.mem_read 4.008895291 S0-D0 1 926592 Bytes llc_misses.mem_read 4.008895291 S0-D1 1 906368 Bytes llc_misses.mem_read 4.008895291 S1-D0 1 892224 Bytes llc_misses.mem_read 4.008895291 S1-D1 1 987712 Bytes llc_misses.mem_read 5.001590993 S0-D0 1 962624 Bytes llc_misses.mem_read 5.001590993 S0-D1 1 912512 Bytes llc_misses.mem_read 5.001590993 S1-D0 1 891200 Bytes llc_misses.mem_read 5.001590993 S1-D1 1 978432 Bytes llc_misses.mem_readOn no-die system,
die_idis 0, actually it'shashmap(socket,0), original behavior is not changed.
Quoted 8 lines
Reported-by Huang Ying <[email protected]>Signed-off-by Jin Yao <[email protected]>---
v9:
Renamezero_per_pkgtoevsel__zero_per_pkgand move it toevsel.c. Thenevsel__zero_per_pkgcan be called under different code path.Call
evsel__zero_per_pkginevsel__exit().
thanks,
jirka
Hi Arnaldo,
Quoted 80 lines · Jiri Olsa
Quoted 79 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-D1But 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 checkdie_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_readFrom above output, we can see S0-D1 and S1-D1 don't report the interval
values, they are continued to grow. That's becausecheck_per_pkg()wrongly decides to use zero counts for S0-D1 and S1-D1.So in
check_per_pkg(), we should usehashmap(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
1.001586691 S0-D0 1 1229440 Bytes llc_misses.mem_read 1.001586691 S0-D1 1 976832 Bytes llc_misses.mem_read 1.001586691 S1-D0 1 938304 Bytes llc_misses.mem_read 1.001586691 S1-D1 1 1227328 Bytes llc_misses.mem_read 2.003776312 S0-D0 1 1586752 Bytes llc_misses.mem_read 2.003776312 S0-D1 1 875392 Bytes llc_misses.mem_read 2.003776312 S1-D0 1 855616 Bytes llc_misses.mem_read 2.003776312 S1-D1 1 949376 Bytes llc_misses.mem_read 3.006512788 S0-D0 1 1338880 Bytes llc_misses.mem_read 3.006512788 S0-D1 1 920064 Bytes llc_misses.mem_read 3.006512788 S1-D0 1 877184 Bytes llc_misses.mem_read 3.006512788 S1-D1 1 1020736 Bytes llc_misses.mem_read 4.008895291 S0-D0 1 926592 Bytes llc_misses.mem_read 4.008895291 S0-D1 1 906368 Bytes llc_misses.mem_read 4.008895291 S1-D0 1 892224 Bytes llc_misses.mem_read 4.008895291 S1-D1 1 987712 Bytes llc_misses.mem_read 5.001590993 S0-D0 1 962624 Bytes llc_misses.mem_read 5.001590993 S0-D1 1 912512 Bytes llc_misses.mem_read 5.001590993 S1-D0 1 891200 Bytes llc_misses.mem_read 5.001590993 S1-D1 1 978432 Bytes llc_misses.mem_readOn no-die system,
die_idis 0, actually it'shashmap(socket,0), original behavior is not changed.Reported-by Huang Ying <[email protected]>Signed-off-by Jin Yao <[email protected]>---
v9:
Renamezero_per_pkgtoevsel__zero_per_pkgand move it toevsel.c. Thenevsel__zero_per_pkgcan be called under different code path.
Quoted 7 lines
Quoted 1 line
Call
evsel__zero_per_pkginevsel__exit().Acked-by Jiri Olsa <[email protected]>thanks,
jirka
Can this fix be accepted or anything else I need to improve?
Thanks
Jin Yao
Em Sun, Jan 31, 2021 at 11:27:08PM +0100, Jiri Olsa escreveu:
Quoted 80 lines · Jin Yao
Quoted 79 lines
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-D1But 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 checkdie_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_readFrom 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 usehashmap(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
1.001586691 S0-D0 1 1229440 Bytes llc_misses.mem_read 1.001586691 S0-D1 1 976832 Bytes llc_misses.mem_read 1.001586691 S1-D0 1 938304 Bytes llc_misses.mem_read 1.001586691 S1-D1 1 1227328 Bytes llc_misses.mem_read 2.003776312 S0-D0 1 1586752 Bytes llc_misses.mem_read 2.003776312 S0-D1 1 875392 Bytes llc_misses.mem_read 2.003776312 S1-D0 1 855616 Bytes llc_misses.mem_read 2.003776312 S1-D1 1 949376 Bytes llc_misses.mem_read 3.006512788 S0-D0 1 1338880 Bytes llc_misses.mem_read 3.006512788 S0-D1 1 920064 Bytes llc_misses.mem_read 3.006512788 S1-D0 1 877184 Bytes llc_misses.mem_read 3.006512788 S1-D1 1 1020736 Bytes llc_misses.mem_read 4.008895291 S0-D0 1 926592 Bytes llc_misses.mem_read 4.008895291 S0-D1 1 906368 Bytes llc_misses.mem_read 4.008895291 S1-D0 1 892224 Bytes llc_misses.mem_read 4.008895291 S1-D1 1 987712 Bytes llc_misses.mem_read 5.001590993 S0-D0 1 962624 Bytes llc_misses.mem_read 5.001590993 S0-D1 1 912512 Bytes llc_misses.mem_read 5.001590993 S1-D0 1 891200 Bytes llc_misses.mem_read 5.001590993 S1-D1 1 978432 Bytes llc_misses.mem_readOn no-die system,
die_idis 0, actually it'shashmap(socket,0), original behavior is not changed.Reported-by Huang Ying <[email protected]>Signed-off-by Jin Yao <[email protected]>---
v9:
Renamezero_per_pkgtoevsel__zero_per_pkgand move it toevsel.c. Thenevsel__zero_per_pkgcan be called under different code path.
Quoted 3 lines
Quoted 1 line
Call
evsel__zero_per_pkginevsel__exit().Acked-by Jiri Olsa <[email protected]>
Thanks, applied.
- Arnaldo