Re: [PATCH v4 00/25] perf tool: AlderLake hybrid support series 1
Hi Arnaldo, Hi Jiri, Kan's patch series for AlderLake perf core support has been upstreamed, so the interface will not be changed any more. For this perf tool series (v4), do you have any comments? Thanks Jin Yao On 4/16/2021 10:04 PM, Jin Yao wrote: AlderLake uses a hybrid architecture utilizing Golden Cove cores (core cpu) and Gracemont cores (atom cpu). Each cpu has dedicated event list. Some events are available on core cpu, some events are available on atom cpu and some events can be available on both. Kernel exports new pmus "cpu_core" and "cpu_atom" through sysfs: /sys/devices/cpu_core /sys/devices/cpu_atom cat /sys/devices/cpu_core/cpus 0-15 cat /sys/devices/cpu_atom/cpus 16-23 In this example, core cpus are 0-15 and atom cpus are 16-23. To enable a core only event or atom only event: cpu_core// or cpu_atom// Count the 'cycles' event on core cpus. # perf stat -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 12,853,951,349 cpu_core/cycles/ 1.002581249 seconds time elapsed If one event is available on both atom cpu and core cpu, two events are created automatically. # perf stat -e cycles -a -- sleep 1 Performance counter stats for 'system wide': 12,856,467,438 cpu_core/cycles/ 6,404,634,785 cpu_atom/cycles/ 1.002453013 seconds time elapsed Group is supported if the events are from same pmu, otherwise a warning is displayed and disable grouping automatically. # perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}' -a -- sleep 1 Performance counter stats for 'system wide': 12,863,866,968 cpu_core/cycles/ 554,795,017 cpu_core/instructions/ 1.002616117 seconds time elapsed # perf stat -e '{cpu_core/cycles/,cpu_atom/instructions/}' -a -- sleep 1 WARNING: events in group from different hybrid PMUs! WARNING: grouped events cpus do not match, disabling group: anon group { cpu_core/cycles/, cpu_atom/instructions/ } Performance counter stats for 'system wide': 6,283,970 cpu_core/cycles/ 765,635 cpu_atom/instructions/ 1.003959036 seconds time elapsed Note that, since the whole patchset for AlderLake hybrid support is very large (40+ patches). For simplicity, it's splitted into several patch series. The patch series 1 only supports the basic functionality. The advanced supports for perf-c2c/perf-mem/topdown/metrics/topology header and others will be added in follow-up patch series. The perf tool codes can also be found at: https://github.com/yaoj/perf.git v4: --- - In Liang Kan's patch: '[PATCH V6 21/25] perf: Extend PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE', the user interface for hardware events and cache events are changed, so perf tool patches are changed as well. - Fix an issue when atom CPUs are offlined. "/sys/bus/event_source/devices/cpu_atom/cpus" exists but the content is empty. For this case, we can't enable the cpu_atom PMU. '[PATCH v4 05/25] perf pmu: Save detected hybrid pmus to a global pmu list' - Define 'ret' variable for return value in patch '[PATCH v4 09/25] perf parse-events: Create two hybrid cache events' - Directly return add_raw_hybrid() in patch '[PATCH v4 10/25] perf parse-events: Create two hybrid raw events' - Drop the patch 'perf pmu: Support 'cycles' and 'branches' inside hybrid PMU'. - Separate '[PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU' into two patches: '[PATCH v4 11/25] perf parse-events: Compare with hybrid pmu name' '[PATCH v4 12/25] perf parse-events: Support event inside hybrid pmu'. And these two patches are improved according to Jiri's comments. v3: --- - Drop 'perf evlist: Hybrid event uses its own cpus'. This patch is wide and actually it's not very necessary. The current perf framework has processed the cpus for evsel well even for hybrid evsel. So this patch can be dropped. - Drop 'perf evsel: Adjust hybrid event and global event mixed group'. The patch is a bit tricky and hard to understand. In v3, we will disable grouping when the group members are from different PMUs. So this patch would be not necessary. - Create parse-events-hybrid.c/parse-events-hybrid.h and evlist-hybrid.c/evlist-hybrid.h. Move hybrid related codes to these files. - Create a new patch 'perf pmu: Support 'cycles' and 'branches' inside hybrid PMU' to support 'cycles' and 'branches' inside PMU. - Create a new patch 'perf record: Uniquify hybrid event name' to tell user the pmu which the event belongs to for perf-record. - If group members are from different hybrid PMUs, shows warning and disable grouping. - Other refining and refactoring. v2: --- - Drop kernel patches (Kan posted the series "Add Alder Lake support for perf (kernel)" separately). - Drop the patch
[PATCH v4] perf Documentation: Document intel-hybrid support
Add some words and examples to help understanding of Intel hybrid perf support. Signed-off-by: Jin Yao --- v4: - Update due to PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE are extended to be PMU type aware. tools/perf/Documentation/intel-hybrid.txt | 214 ++ tools/perf/Documentation/perf-record.txt | 1 + tools/perf/Documentation/perf-stat.txt| 2 + 3 files changed, 217 insertions(+) create mode 100644 tools/perf/Documentation/intel-hybrid.txt diff --git a/tools/perf/Documentation/intel-hybrid.txt b/tools/perf/Documentation/intel-hybrid.txt new file mode 100644 index ..07f0aa3bf682 --- /dev/null +++ b/tools/perf/Documentation/intel-hybrid.txt @@ -0,0 +1,214 @@ +Intel hybrid support + +Support for Intel hybrid events within perf tools. + +For some Intel platforms, such as AlderLake, which is hybrid platform and +it consists of atom cpu and core cpu. Each cpu has dedicated event list. +Part of events are available on core cpu, part of events are available +on atom cpu and even part of events are available on both. + +Kernel exports two new cpu pmus via sysfs: +/sys/devices/cpu_core +/sys/devices/cpu_atom + +The 'cpus' files are created under the directories. For example, + +cat /sys/devices/cpu_core/cpus +0-15 + +cat /sys/devices/cpu_atom/cpus +16-23 + +It indicates cpu0-cpu15 are core cpus and cpu16-cpu23 are atom cpus. + +Quickstart + +List hybrid event +- + +As before, use perf-list to list the symbolic event. + +perf list + +inst_retired.any + [Fixed Counter: Counts the number of instructions retired. Unit: cpu_atom] +inst_retired.any + [Number of instructions retired. Fixed Counter - architectural event. Unit: cpu_core] + +The 'Unit: xxx' is added to brief description to indicate which pmu +the event is belong to. Same event name but with different pmu can +be supported. + +Enable hybrid event with a specific pmu +--- + +To enable a core only event or atom only event, following syntax is supported: + + cpu_core// +or + cpu_atom// + +For example, count the 'cycles' event on core cpus. + + perf stat -e cpu_core/cycles/ + +Create two events for one hardware event automatically +-- + +When creating one event and the event is available on both atom and core, +two events are created automatically. One is for atom, the other is for +core. Most of hardware events and cache events are available on both +cpu_core and cpu_atom. + +For hardware events, they have pre-defined configs (e.g. 0 for cycles). +But on hybrid platform, kernel needs to know where the event comes from +(from atom or from core). The original perf event type PERF_TYPE_HARDWARE +can't carry pmu information. So now this type is extended to be PMU aware +type. The PMU type ID is stored at attr.config[63:32]. + +PMU type ID is retrieved from sysfs. +/sys/devices/cpu_atom/type +/sys/devices/cpu_core/type + +The new attr.config layout for PERF_TYPE_HARDWARE: + +PERF_TYPE_HARDWARE: 0x00AA +AA: hardware event ID +: PMU type ID + +Cache event is similar. The type PERF_TYPE_HW_CACHE is extended to be +PMU aware type. The PMU type ID is stored at attr.config[63:32]. + +The new attr.config layout for PERF_TYPE_HW_CACHE: + +PERF_TYPE_HW_CACHE: 0x00DDCCBB +BB: hardware cache ID +CC: hardware cache op ID +DD: hardware cache op result ID +: PMU type ID + +When enabling a hardware event without specified pmu, such as, +perf stat -e cycles -a (use system-wide in this example), two events +are created automatically. + + + perf_event_attr: +size 120 +config 0x4 +sample_type IDENTIFIER +read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING +disabled 1 +inherit 1 +exclude_guest1 + + +and + + + perf_event_attr: +size 120 +config 0x8 +sample_type IDENTIFIER +read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING +disabled 1 +inherit 1 +exclude_guest1 + + +type 0 is PERF_TYPE_HARDWARE. +0x4 in 0x4 indicates it's cpu_core pmu. +0x8 in 0x8
[PATCH v4 25/25] perf tests: Skip 'perf stat metrics (shadow stat) test' for hybrid
Currently we don't support shadow stat for hybrid. root@ssp-pwrt-002:~# ./perf stat -e cycles,instructions -a -- sleep 1 Performance counter stats for 'system wide': 12,883,109,591 cpu_core/cycles/ 6,405,163,221 cpu_atom/cycles/ 555,553,778 cpu_core/instructions/ 841,158,734 cpu_atom/instructions/ 1.002644773 seconds time elapsed Now there is no shadow stat 'insn per cycle' reported. We will support it later and now just skip the 'perf stat metrics (shadow stat) test'. Signed-off-by: Jin Yao --- tools/perf/tests/shell/stat+shadow_stat.sh | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/perf/tests/shell/stat+shadow_stat.sh b/tools/perf/tests/shell/stat+shadow_stat.sh index ebebd3596cf9..e6e35fc6c882 100755 --- a/tools/perf/tests/shell/stat+shadow_stat.sh +++ b/tools/perf/tests/shell/stat+shadow_stat.sh @@ -7,6 +7,9 @@ set -e # skip if system-wide mode is forbidden perf stat -a true > /dev/null 2>&1 || exit 2 +# skip if on hybrid platform +perf stat -a -e cycles sleep 1 2>&1 | grep -e cpu_core && exit 2 + test_global_aggr() { perf stat -a --no-big-num -e cycles,instructions sleep 1 2>&1 | \ -- 2.17.1
[PATCH v4 22/25] perf tests: Support 'Parse and process metrics' test for hybrid
Some events are not supported. Only pick up some cases for hybrid. # ./perf test 67 67: Parse and process metrics : Ok Signed-off-by: Jin Yao --- tools/perf/tests/parse-metric.c | 10 -- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/tools/perf/tests/parse-metric.c b/tools/perf/tests/parse-metric.c index 4968c4106254..24e5ddff515e 100644 --- a/tools/perf/tests/parse-metric.c +++ b/tools/perf/tests/parse-metric.c @@ -11,6 +11,7 @@ #include "debug.h" #include "expr.h" #include "stat.h" +#include "pmu.h" static struct pmu_event pme_test[] = { { @@ -370,12 +371,17 @@ static int test_metric_group(void) int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused) { + perf_pmu__scan(NULL); + TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); - TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); - TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0); + + if (!perf_pmu__has_hybrid()) { + TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); + TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); + } return 0; } -- 2.17.1
[PATCH v4 24/25] perf tests: Support 'Convert perf time to TSC' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the second evsel in evlist also needs initialization. With this patch, # ./perf test 71 71: Convert perf time to TSC: Ok Signed-off-by: Jin Yao --- tools/perf/tests/perf-time-to-tsc.c | 16 1 file changed, 16 insertions(+) diff --git a/tools/perf/tests/perf-time-to-tsc.c b/tools/perf/tests/perf-time-to-tsc.c index 680c3cffb128..72f268c6cc5d 100644 --- a/tools/perf/tests/perf-time-to-tsc.c +++ b/tools/perf/tests/perf-time-to-tsc.c @@ -20,6 +20,7 @@ #include "tsc.h" #include "mmap.h" #include "tests.h" +#include "pmu.h" #define CHECK__(x) { \ while ((x) < 0) { \ @@ -66,6 +67,10 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe u64 test_tsc, comm1_tsc, comm2_tsc; u64 test_time, comm1_time = 0, comm2_time = 0; struct mmap *md; + bool hybrid = false; + + if (perf_pmu__has_hybrid()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); CHECK_NOT_NULL__(threads); @@ -88,6 +93,17 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe evsel->core.attr.disabled = 1; evsel->core.attr.enable_on_exec = 0; + /* +* For hybrid "cycles:u", it creates two events. +* Init the second evsel here. +*/ + if (hybrid) { + evsel = evsel__next(evsel); + evsel->core.attr.comm = 1; + evsel->core.attr.disabled = 1; + evsel->core.attr.enable_on_exec = 0; + } + CHECK__(evlist__open(evlist)); CHECK__(evlist__mmap(evlist, UINT_MAX)); -- 2.17.1
[PATCH v4 23/25] perf tests: Support 'Session topology' test for hybrid
Force to create one event "cpu_core/cycles/" by default, otherwise in evlist__valid_sample_type, the checking of 'if (evlist->core.nr_entries == 1)' would be failed. # ./perf test 41 41: Session topology: Ok Signed-off-by: Jin Yao --- tools/perf/tests/topology.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/topology.c b/tools/perf/tests/topology.c index 050489807a47..30b4acb08d35 100644 --- a/tools/perf/tests/topology.c +++ b/tools/perf/tests/topology.c @@ -8,6 +8,7 @@ #include "session.h" #include "evlist.h" #include "debug.h" +#include "pmu.h" #include #define TEMPL "/tmp/perf-test-XX" @@ -40,7 +41,14 @@ static int session_write_header(char *path) session = perf_session__new(, false, NULL); TEST_ASSERT_VAL("can't get session", !IS_ERR(session)); - session->evlist = evlist__new_default(); + if (!perf_pmu__has_hybrid()) { + session->evlist = evlist__new_default(); + } else { + struct parse_events_error err; + + session->evlist = evlist__new(); + parse_events(session->evlist, "cpu_core/cycles/", ); + } TEST_ASSERT_VAL("can't get evlist", session->evlist); perf_header__set_feat(>header, HEADER_CPU_TOPOLOGY); -- 2.17.1
[PATCH v4 21/25] perf tests: Support 'Track with sched_switch' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the number of events in evlist is not expected in next test steps. Now we just use one event "cpu_core/cycles:u/" for hybrid. # ./perf test 35 35: Track with sched_switch : Ok Signed-off-by: Jin Yao --- tools/perf/tests/switch-tracking.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/switch-tracking.c b/tools/perf/tests/switch-tracking.c index 3ebaa758df77..3a12176f8c46 100644 --- a/tools/perf/tests/switch-tracking.c +++ b/tools/perf/tests/switch-tracking.c @@ -18,6 +18,7 @@ #include "record.h" #include "tests.h" #include "util/mmap.h" +#include "pmu.h" static int spin_sleep(void) { @@ -340,6 +341,10 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ struct evsel *switch_evsel, *tracking_evsel; const char *comm; int err = -1; + bool hybrid = false; + + if (perf_pmu__has_hybrid()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); if (!threads) { @@ -371,7 +376,10 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ cpu_clocks_evsel = evlist__last(evlist); /* Second event */ - err = parse_events(evlist, "cycles:u", NULL); + if (!hybrid) + err = parse_events(evlist, "cycles:u", NULL); + else + err = parse_events(evlist, "cpu_core/cycles/u", NULL); if (err) { pr_debug("Failed to parse event cycles:u\n"); goto out_err; -- 2.17.1
[PATCH v4 20/25] perf tests: Skip 'Setup struct perf_event_attr' test for hybrid
For hybrid, the attr.type consists of pmu type id + original type. There will be much changes for this test. Now we temporarily skip this test case and TODO in future. Signed-off-by: Jin Yao --- tools/perf/tests/attr.c | 4 1 file changed, 4 insertions(+) diff --git a/tools/perf/tests/attr.c b/tools/perf/tests/attr.c index dd39ce9b0277..b37c35fb5a46 100644 --- a/tools/perf/tests/attr.c +++ b/tools/perf/tests/attr.c @@ -34,6 +34,7 @@ #include "event.h" #include "util.h" #include "tests.h" +#include "pmu.h" #define ENV "PERF_TEST_ATTR" @@ -184,6 +185,9 @@ int test__attr(struct test *test __maybe_unused, int subtest __maybe_unused) char path_dir[PATH_MAX]; char *exec_path; + if (perf_pmu__has_hybrid()) + return 0; + /* First try development tree tests. */ if (!lstat("./tests", )) return run_dir("./tests", "./perf"); -- 2.17.1
[PATCH v4 18/25] perf tests: Add hybrid cases for 'Parse event definition strings' test
Add basic hybrid test cases for 'Parse event definition strings' test. # perf test 6 6: Parse event definition strings : Ok Signed-off-by: Jin Yao --- tools/perf/tests/parse-events.c | 152 1 file changed, 152 insertions(+) diff --git a/tools/perf/tests/parse-events.c b/tools/perf/tests/parse-events.c index 026c54743311..40eb08049ab2 100644 --- a/tools/perf/tests/parse-events.c +++ b/tools/perf/tests/parse-events.c @@ -1512,6 +1512,110 @@ static int test__all_tracepoints(struct evlist *evlist) return test__checkevent_tracepoint_multi(evlist); } +static int test__hybrid_hw_event_with_pmu(struct evlist *evlist) +{ + struct evsel *evsel = evlist__first(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + return 0; +} + +static int test__hybrid_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0xc0 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_sw_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_hw_sw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_group_modifier1(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + TEST_ASSERT_VAL("wrong exclude_user", evsel->core.attr.exclude_user); + TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->core.attr.exclude_kernel); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0xc0 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + TEST_ASSERT_VAL("wrong exclude_user", !evsel->core.attr.exclude_user); + TEST_ASSERT_VAL("wrong exclude_kernel", evsel->core.attr.exclude_kernel); + return 0; +} + +static int test__hybrid_raw1(struct evlist *evlist) +{ + struct evsel *evsel = evlist__first(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x1a == evsel->core.attr.config); + + /* The ty
[PATCH v4 19/25] perf tests: Add hybrid cases for 'Roundtrip evsel->name' test
Since for one hw event, two hybrid events are created. For example, evsel->idx evsel__name(evsel) 0 cycles 1 cycles 2 instructions 3 instructions ... So for comparing the evsel name on hybrid, the evsel->idx needs to be divided by 2. # ./perf test 14 14: Roundtrip evsel->name : Ok Signed-off-by: Jin Yao --- tools/perf/tests/evsel-roundtrip-name.c | 19 --- 1 file changed, 12 insertions(+), 7 deletions(-) diff --git a/tools/perf/tests/evsel-roundtrip-name.c b/tools/perf/tests/evsel-roundtrip-name.c index f7f3e5b4c180..b74cf80d1f10 100644 --- a/tools/perf/tests/evsel-roundtrip-name.c +++ b/tools/perf/tests/evsel-roundtrip-name.c @@ -4,6 +4,7 @@ #include "parse-events.h" #include "tests.h" #include "debug.h" +#include "pmu.h" #include #include @@ -62,7 +63,8 @@ static int perf_evsel__roundtrip_cache_name_test(void) return ret; } -static int __perf_evsel__name_array_test(const char *names[], int nr_names) +static int __perf_evsel__name_array_test(const char *names[], int nr_names, +int distance) { int i, err; struct evsel *evsel; @@ -82,9 +84,9 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) err = 0; evlist__for_each_entry(evlist, evsel) { - if (strcmp(evsel__name(evsel), names[evsel->idx])) { + if (strcmp(evsel__name(evsel), names[evsel->idx / distance])) { --err; - pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx]); + pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx / distance]); } } @@ -93,18 +95,21 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) return err; } -#define perf_evsel__name_array_test(names) \ - __perf_evsel__name_array_test(names, ARRAY_SIZE(names)) +#define perf_evsel__name_array_test(names, distance) \ + __perf_evsel__name_array_test(names, ARRAY_SIZE(names), distance) int test__perf_evsel__roundtrip_name_test(struct test *test __maybe_unused, int subtest __maybe_unused) { int err = 0, ret = 0; - err = perf_evsel__name_array_test(evsel__hw_names); + if (perf_pmu__has_hybrid()) + return perf_evsel__name_array_test(evsel__hw_names, 2); + + err = perf_evsel__name_array_test(evsel__hw_names, 1); if (err) ret = err; - err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1); + err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1, 1); if (err) ret = err; -- 2.17.1
[PATCH v4 16/25] perf stat: Warn group events from different hybrid PMU
If a group has events which are from different hybrid PMUs, shows a warning: "WARNING: events in group from different hybrid PMUs!" This is to remind the user not to put the core event and atom event into one group. Next, just disable grouping. # perf stat -e "{cpu_core/cycles/,cpu_atom/cycles/}" -a -- sleep 1 WARNING: events in group from different hybrid PMUs! WARNING: grouped events cpus do not match, disabling group: anon group { cpu_core/cycles/, cpu_atom/cycles/ } Performance counter stats for 'system wide': 5,438,125 cpu_core/cycles/ 3,914,586 cpu_atom/cycles/ 1.004250966 seconds time elapsed Signed-off-by: Jin Yao --- v4: - No change. tools/perf/builtin-stat.c | 4 +++ tools/perf/util/evlist-hybrid.c| 47 ++ tools/perf/util/evlist-hybrid.h| 2 ++ tools/perf/util/evsel.c| 6 tools/perf/util/evsel.h| 1 + tools/perf/util/python-ext-sources | 2 ++ 6 files changed, 62 insertions(+) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 0351b99d17a7..c429aae6eeb6 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -48,6 +48,7 @@ #include "util/pmu.h" #include "util/event.h" #include "util/evlist.h" +#include "util/evlist-hybrid.h" #include "util/evsel.h" #include "util/debug.h" #include "util/color.h" @@ -240,6 +241,9 @@ static void evlist__check_cpu_maps(struct evlist *evlist) struct evsel *evsel, *pos, *leader; char buf[1024]; + if (evlist__has_hybrid(evlist)) + evlist__warn_hybrid_group(evlist); + evlist__for_each_entry(evlist, evsel) { leader = evsel->leader; diff --git a/tools/perf/util/evlist-hybrid.c b/tools/perf/util/evlist-hybrid.c index e11998526f2e..db3f5fbdebe1 100644 --- a/tools/perf/util/evlist-hybrid.c +++ b/tools/perf/util/evlist-hybrid.c @@ -7,6 +7,7 @@ #include "../perf.h" #include "util/pmu-hybrid.h" #include "util/evlist-hybrid.h" +#include "debug.h" #include #include #include @@ -39,3 +40,49 @@ int evlist__add_default_hybrid(struct evlist *evlist, bool precise) return 0; } + +static bool group_hybrid_conflict(struct evsel *leader) +{ + struct evsel *pos, *prev = NULL; + + for_each_group_evsel(pos, leader) { + if (!evsel__is_hybrid(pos)) + continue; + + if (prev && strcmp(prev->pmu_name, pos->pmu_name)) + return true; + + prev = pos; + } + + return false; +} + +void evlist__warn_hybrid_group(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel__is_group_leader(evsel) && + evsel->core.nr_members > 1 && + group_hybrid_conflict(evsel)) { + pr_warning("WARNING: events in group from " + "different hybrid PMUs!\n"); + return; + } + } +} + +bool evlist__has_hybrid(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel->pmu_name && + perf_pmu__is_hybrid(evsel->pmu_name)) { + return true; + } + } + + return false; +} diff --git a/tools/perf/util/evlist-hybrid.h b/tools/perf/util/evlist-hybrid.h index e25861649d8f..19f74b4c340a 100644 --- a/tools/perf/util/evlist-hybrid.h +++ b/tools/perf/util/evlist-hybrid.h @@ -8,5 +8,7 @@ #include int evlist__add_default_hybrid(struct evlist *evlist, bool precise); +void evlist__warn_hybrid_group(struct evlist *evlist); +bool evlist__has_hybrid(struct evlist *evlist); #endif /* __PERF_EVLIST_HYBRID_H */ diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c index 0ba4daa09453..0f64a32ea9c5 100644 --- a/tools/perf/util/evsel.c +++ b/tools/perf/util/evsel.c @@ -47,6 +47,7 @@ #include "memswap.h" #include "util.h" #include "hashmap.h" +#include "pmu-hybrid.h" #include "../perf-sys.h" #include "util/parse-branch-options.h" #include @@ -2797,3 +2798,8 @@ void evsel__zero_per_pkg(struct evsel *evsel) hashmap__clear(evsel->per_pkg_mask); } } + +bool evsel__is_hybrid(struct evsel *evsel) +{ + return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name); +} diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h index ff89196281bd..f6f90f68381b 100644 --- a/tools/perf/util/evsel.h +++ b/tools/perf/util/evsel.h @@ -453,4 +453,5 @@ struct perf_env *evsel__env(struct evsel *evsel); int evsel__sto
[PATCH v4 17/25] perf record: Uniquify hybrid event name
For perf-record, it would be useful to tell user the pmu which the event belongs to. For example, # perf record -a -- sleep 1 # perf report # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 106 of event 'cpu_core/cycles/' # Event count (approx.): 22043448 # # Overhead Command Shared ObjectSymbol # ... # ... Signed-off-by: Jin Yao --- v4: - No change. tools/perf/builtin-record.c | 28 1 file changed, 28 insertions(+) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 6af46c6a4fd8..3337b5f93336 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -1605,6 +1605,32 @@ static void hit_auxtrace_snapshot_trigger(struct record *rec) } } +static void record__uniquify_name(struct record *rec) +{ + struct evsel *pos; + struct evlist *evlist = rec->evlist; + char *new_name; + int ret; + + if (!perf_pmu__has_hybrid()) + return; + + evlist__for_each_entry(evlist, pos) { + if (!evsel__is_hybrid(pos)) + continue; + + if (strchr(pos->name, '/')) + continue; + + ret = asprintf(_name, "%s/%s/", + pos->pmu_name, pos->name); + if (ret) { + free(pos->name); + pos->name = new_name; + } + } +} + static int __cmd_record(struct record *rec, int argc, const char **argv) { int err; @@ -1709,6 +1735,8 @@ static int __cmd_record(struct record *rec, int argc, const char **argv) if (data->is_pipe && rec->evlist->core.nr_entries == 1) rec->opts.sample_id = true; + record__uniquify_name(rec); + if (record__open(rec) != 0) { err = -1; goto out_child; -- 2.17.1
[PATCH v4 14/25] perf stat: Add default hybrid events
Previously if '-e' is not specified in perf stat, some software events and hardware events are added to evlist by default. Before: # ./perf stat -a -- sleep 1 Performance counter stats for 'system wide': 24,044.40 msec cpu-clock # 23.946 CPUs utilized 99 context-switches #4.117 /sec 24 cpu-migrations#0.998 /sec 3 page-faults #0.125 /sec 7,000,244 cycles#0.000 GHz 2,955,024 instructions #0.42 insn per cycle 608,941 branches # 25.326 K/sec 31,991 branch-misses #5.25% of all branches 1.004106859 seconds time elapsed Among the events, cycles, instructions, branches and branch-misses are hardware events. One hybrid platform, two hardware events are created for one hardware event. cpu_core/cycles/, cpu_atom/cycles/, cpu_core/instructions/, cpu_atom/instructions/, cpu_core/branches/, cpu_atom/branches/, cpu_core/branch-misses/, cpu_atom/branch-misses/ These events would be added to evlist on hybrid platform. Since parse_events() has been supported to create two hardware events for one event on hybrid platform, so we just use parse_events(evlist, "cycles,instructions,branches,branch-misses") to create the default events and add them to evlist. After: # ./perf stat -a -- sleep 1 Performance counter stats for 'system wide': 24,048.60 msec task-clock# 23.947 CPUs utilized 438 context-switches # 18.213 /sec 24 cpu-migrations#0.998 /sec 6 page-faults #0.249 /sec 24,813,157 cpu_core/cycles/ #1.032 M/sec 8,072,687 cpu_atom/cycles/ # 335.682 K/sec 20,731,286 cpu_core/instructions/# 862.058 K/sec 3,737,203 cpu_atom/instructions/# 155.402 K/sec 2,620,924 cpu_core/branches/# 108.984 K/sec 381,186 cpu_atom/branches/# 15.851 K/sec 93,248 cpu_core/branch-misses/ #3.877 K/sec 36,515 cpu_atom/branch-misses/ #1.518 K/sec 1.004235472 seconds time elapsed We can see two events are created for one hardware event. One TODO is, the shadow stats looks a bit different, now it's just 'M/sec'. The perf_stat__update_shadow_stats and perf_stat__print_shadow_stats need to be improved in future if we want to get the original shadow stats. Signed-off-by: Jin Yao --- v4: - No change. tools/perf/builtin-stat.c | 21 + 1 file changed, 21 insertions(+) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 1255af4751c2..0351b99d17a7 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1145,6 +1145,13 @@ static int parse_stat_cgroups(const struct option *opt, return parse_cgroups(opt, str, unset); } +static int add_default_hybrid_events(struct evlist *evlist) +{ + struct parse_events_error err; + + return parse_events(evlist, "cycles,instructions,branches,branch-misses", ); +} + static struct option stat_options[] = { OPT_BOOLEAN('T', "transaction", _run, "hardware transaction statistics"), @@ -1626,6 +1633,12 @@ static int add_default_attributes(void) { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, +}; + struct perf_event_attr default_sw_attrs[] = { + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, }; /* @@ -1863,6 +1876,14 @@ static int add_default_attributes(void) } if (!evsel_list->core.nr_entries) { + if (perf_pmu__has_hybrid()) { + if (evlist__add_default_attrs(evsel_list, + default_sw_attrs) < 0) { + return -1; + } + return add_default_hybrid_events(evsel_list); + } + if (target__has_cpu()) default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; -- 2.17.1
[PATCH v4 15/25] perf stat: Filter out unmatched aggregation for hybrid event
perf-stat has supported some aggregation modes, such as --per-core, --per-socket and etc. While for hybrid event, it may only available on part of cpus. So for --per-core, we need to filter out the unavailable cores, for --per-socket, filter out the unavailable sockets, and so on. Before: # perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2479,530 cpu_core/cycles/ S0-D0-C4 2175,007 cpu_core/cycles/ S0-D0-C8 2166,240 cpu_core/cycles/ S0-D0-C12 2704,673 cpu_core/cycles/ S0-D0-C16 2865,835 cpu_core/cycles/ S0-D0-C20 2 2,958,461 cpu_core/cycles/ S0-D0-C24 2163,988 cpu_core/cycles/ S0-D0-C28 2164,729 cpu_core/cycles/ S0-D0-C32 0cpu_core/cycles/ S0-D0-C33 0cpu_core/cycles/ S0-D0-C34 0cpu_core/cycles/ S0-D0-C35 0cpu_core/cycles/ S0-D0-C36 0cpu_core/cycles/ S0-D0-C37 0cpu_core/cycles/ S0-D0-C38 0cpu_core/cycles/ S0-D0-C39 0cpu_core/cycles/ 1.003597211 seconds time elapsed After: # perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2210,428 cpu_core/cycles/ S0-D0-C4 2444,830 cpu_core/cycles/ S0-D0-C8 2435,241 cpu_core/cycles/ S0-D0-C12 2423,976 cpu_core/cycles/ S0-D0-C16 2859,350 cpu_core/cycles/ S0-D0-C20 2 1,559,589 cpu_core/cycles/ S0-D0-C24 2163,924 cpu_core/cycles/ S0-D0-C28 2376,610 cpu_core/cycles/ 1.003621290 seconds time elapsed Signed-off-by: Jin Yao --- v4: - No change. tools/perf/util/stat-display.c | 20 1 file changed, 20 insertions(+) diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 5255d78b1c30..15eafd249e46 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -643,6 +643,20 @@ static void aggr_cb(struct perf_stat_config *config, } } +static bool aggr_id_hybrid_matched(struct perf_stat_config *config, + struct evsel *counter, struct aggr_cpu_id id) +{ + struct aggr_cpu_id s; + + for (int i = 0; i < evsel__nr_cpus(counter); i++) { + s = config->aggr_get_id(config, evsel__cpus(counter), i); + if (cpu_map__compare_aggr_cpu_id(s, id)) + return true; + } + + return false; +} + static void print_counter_aggrdata(struct perf_stat_config *config, struct evsel *counter, int s, char *prefix, bool metric_only, @@ -656,6 +670,12 @@ static void print_counter_aggrdata(struct perf_stat_config *config, double uval; ad.id = id = config->aggr_map->map[s]; + + if (perf_pmu__has_hybrid() && + !aggr_id_hybrid_matched(config, counter, id)) { + return; + } + ad.val = ad.ena = ad.run = 0; ad.nr = 0; if (!collect_data(config, counter, aggr_cb, )) -- 2.17.1
[PATCH v4 13/25] perf record: Create two hybrid 'cycles' events by default
When evlist is empty, for example no '-e' specified in perf record, one default 'cycles' event is added to evlist. While on hybrid platform, it needs to create two default 'cycles' events. One is for cpu_core, the other is for cpu_atom. This patch actually calls evsel__new_cycles() two times to create two 'cycles' events. # ./perf record -vv -a -- sleep 1 ... perf_event_attr: size 120 config 0x4 { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|CPU|PERIOD read_format ID disabled 1 inherit 1 freq 1 precise_ip 3 sample_id_all1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 5 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8 = 6 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8 = 7 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8 = 9 sys_perf_event_open: pid -1 cpu 4 group_fd -1 flags 0x8 = 10 sys_perf_event_open: pid -1 cpu 5 group_fd -1 flags 0x8 = 11 sys_perf_event_open: pid -1 cpu 6 group_fd -1 flags 0x8 = 12 sys_perf_event_open: pid -1 cpu 7 group_fd -1 flags 0x8 = 13 sys_perf_event_open: pid -1 cpu 8 group_fd -1 flags 0x8 = 14 sys_perf_event_open: pid -1 cpu 9 group_fd -1 flags 0x8 = 15 sys_perf_event_open: pid -1 cpu 10 group_fd -1 flags 0x8 = 16 sys_perf_event_open: pid -1 cpu 11 group_fd -1 flags 0x8 = 17 sys_perf_event_open: pid -1 cpu 12 group_fd -1 flags 0x8 = 18 sys_perf_event_open: pid -1 cpu 13 group_fd -1 flags 0x8 = 19 sys_perf_event_open: pid -1 cpu 14 group_fd -1 flags 0x8 = 20 sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 21 perf_event_attr: size 120 config 0x8 { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|CPU|PERIOD read_format ID disabled 1 inherit 1 freq 1 precise_ip 3 sample_id_all1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 22 sys_perf_event_open: pid -1 cpu 17 group_fd -1 flags 0x8 = 23 sys_perf_event_open: pid -1 cpu 18 group_fd -1 flags 0x8 = 24 sys_perf_event_open: pid -1 cpu 19 group_fd -1 flags 0x8 = 25 sys_perf_event_open: pid -1 cpu 20 group_fd -1 flags 0x8 = 26 sys_perf_event_open: pid -1 cpu 21 group_fd -1 flags 0x8 = 27 sys_perf_event_open: pid -1 cpu 22 group_fd -1 flags 0x8 = 28 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 29 We have to create evlist-hybrid.c otherwise due to the symbol dependency the perf test python would be failed. Signed-off-by: Jin Yao --- v4: - Use PERF_TYPE_HARDWARE (v3 uses PERF_TYPE_HARDWARE_PMU). v3: - Move the major code to new created evlist-hybrid.c. tools/perf/builtin-record.c | 19 +++ tools/perf/util/Build | 1 + tools/perf/util/evlist-hybrid.c | 41 + tools/perf/util/evlist-hybrid.h | 12 ++ tools/perf/util/evlist.c| 5 +++- tools/perf/util/evsel.c | 6 ++--- tools/perf/util/evsel.h | 2 +- 7 files changed, 77 insertions(+), 9 deletions(-) create mode 100644 tools/perf/util/evlist-hybrid.c create mode 100644 tools/perf/util/evlist-hybrid.h diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 5fb9665a2ec2..6af46c6a4fd8 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -47,6 +47,8 @@ #include "util/util.h" #include "util/pfm.h" #include "util/clockid.h" +#include "util/pmu-hybrid.h" +#include "util/evlist-hybrid.h" #include "asm/bug.h" #include "perf.h" @@ -2790,10 +2792,19 @@ int cmd_record(int argc, const char **argv) if (record.opts.overwrite) record.opts.tail_synthesize = true; - if (rec->evlist->core.nr_entries == 0 && - __evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) { - pr_err("Not enough memory for event selector list\n"); - goto out; + if (rec->evlist->core.nr_entries == 0) { +
[PATCH v4 11/25] perf parse-events: Compare with hybrid pmu name
On hybrid platform, user may want to enable event only on one pmu. Following syntax will be supported: cpu_core// cpu_atom// For hardware event, hardware cache event and raw event, two events are created by default. We pass the specified pmu name in parse_state and it would be checked before event creation. So next only the event with the specified pmu would be created. Signed-off-by: Jin Yao --- v4: - New in v4. tools/perf/util/parse-events-hybrid.c | 21 - tools/perf/util/parse-events-hybrid.h | 3 ++- tools/perf/util/parse-events.c| 5 +++-- tools/perf/util/parse-events.h| 4 +++- tools/perf/util/parse-events.y| 9 ++--- 5 files changed, 34 insertions(+), 8 deletions(-) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index e27b747080e1..10160ab126f9 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -59,6 +59,15 @@ static int create_event_hybrid(__u32 config_type, int *idx, return 0; } +static int pmu_cmp(struct parse_events_state *parse_state, + struct perf_pmu *pmu) +{ + if (!parse_state->hybrid_pmu_name) + return 0; + + return strcmp(parse_state->hybrid_pmu_name, pmu->name); +} + static int add_hw_hybrid(struct parse_events_state *parse_state, struct list_head *list, struct perf_event_attr *attr, char *name, struct list_head *config_terms) @@ -67,6 +76,9 @@ static int add_hw_hybrid(struct parse_events_state *parse_state, int ret; perf_pmu__for_each_hybrid_pmu(pmu) { + if (pmu_cmp(parse_state, pmu)) + continue; + ret = create_event_hybrid(PERF_TYPE_HARDWARE, _state->idx, list, attr, name, config_terms, pmu); @@ -103,6 +115,9 @@ static int add_raw_hybrid(struct parse_events_state *parse_state, int ret; perf_pmu__for_each_hybrid_pmu(pmu) { + if (pmu_cmp(parse_state, pmu)) + continue; + ret = create_raw_event_hybrid(_state->idx, list, attr, name, config_terms, pmu); if (ret) @@ -138,7 +153,8 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, int parse_events__add_cache_hybrid(struct list_head *list, int *idx, struct perf_event_attr *attr, char *name, struct list_head *config_terms, - bool *hybrid) + bool *hybrid, + struct parse_events_state *parse_state) { struct perf_pmu *pmu; int ret; @@ -149,6 +165,9 @@ int parse_events__add_cache_hybrid(struct list_head *list, int *idx, *hybrid = true; perf_pmu__for_each_hybrid_pmu(pmu) { + if (pmu_cmp(parse_state, pmu)) + continue; + ret = create_event_hybrid(PERF_TYPE_HW_CACHE, idx, list, attr, name, config_terms, pmu); if (ret) diff --git a/tools/perf/util/parse-events-hybrid.h b/tools/perf/util/parse-events-hybrid.h index 9ad33cd0cef4..f33bd67aa851 100644 --- a/tools/perf/util/parse-events-hybrid.h +++ b/tools/perf/util/parse-events-hybrid.h @@ -17,6 +17,7 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, int parse_events__add_cache_hybrid(struct list_head *list, int *idx, struct perf_event_attr *attr, char *name, struct list_head *config_terms, - bool *hybrid); + bool *hybrid, + struct parse_events_state *parse_state); #endif /* __PERF_PARSE_EVENTS_HYBRID_H */ diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 9b2588df22a4..f69475a158bb 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -453,7 +453,8 @@ static int config_attr(struct perf_event_attr *attr, int parse_events_add_cache(struct list_head *list, int *idx, char *type, char *op_result1, char *op_result2, struct parse_events_error *err, - struct list_head *head_config) + struct list_head *head_config, + struct parse_events_state *parse_state) { struct perf_event_attr attr; LIST_HEAD(config_terms); @@ -524,7 +525,7 @@ int parse_events_add_cache(struct list_head *list, int *idx, ret = parse_events__add_cache_hybrid(list, idx, , config_name ?
[PATCH v4 12/25] perf parse-events: Support event inside hybrid pmu
On hybrid platform, user may want to enable events on one pmu. Following syntax are supported: cpu_core// cpu_atom// But the syntax doesn't work for cache event. Before: # perf stat -e cpu_core/LLC-loads/ -a -- sleep 1 event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' Cache events are a bit complex. We can't create aliases for them. We use another solution. For example, if we use "cpu_core/LLC-loads/", in parse_events_add_pmu(), term->config is "LLC-loads". Then we create a new parser to scan "LLC-loads". The parse_events_add_cache() would be called during parsing. The parse_state->hybrid_pmu_name is used to identify the pmu where the event should be enabled on. After: # perf stat -e cpu_core/LLC-loads/ -a -- sleep 1 Performance counter stats for 'system wide': 24,593 cpu_core/LLC-loads/ 1.003911601 seconds time elapsed If the user sets the config name, we will not uniquify hybrid event name. # perf stat -e cpu_core/r3c/ -a -- sleep 1 Performance counter stats for 'system wide': 5,072,048 cpu_core/r3c/ 1.001989415 seconds time elapsed # perf stat -e cpu_core/r3c,name=EVENT/ -a -- sleep 1 Performance counter stats for 'system wide': 6,819,847 EVENT 1.001795630 seconds time elapsed Signed-off-by: Jin Yao --- v4: - New in v4. tools/perf/util/parse-events.c | 55 ++ 1 file changed, 55 insertions(+) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index f69475a158bb..bd3fd722b4ac 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -38,6 +38,7 @@ #include "util/event.h" #include "util/pfm.h" #include "util/parse-events-hybrid.h" +#include "util/pmu-hybrid.h" #include "perf.h" #define MAX_NAME_LEN 100 @@ -48,6 +49,9 @@ extern int parse_events_debug; int parse_events_parse(void *parse_state, void *scanner); static int get_config_terms(struct list_head *head_config, struct list_head *head_terms __maybe_unused); +static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, +const char *str, char *pmu_name, +struct list_head *list, bool *parsed); static struct perf_pmu_event_symbol *perf_pmu_events_list; /* @@ -1567,6 +1571,27 @@ int parse_events_add_pmu(struct parse_events_state *parse_state, if (pmu->default_config && get_config_chgs(pmu, head_config, _terms)) return -ENOMEM; + if (!parse_state->fake_pmu && head_config && + perf_pmu__is_hybrid(name)) { + struct parse_events_term *term; + bool parsed; + int ret; + + term = list_first_entry(head_config, struct parse_events_term, + list); + if (term && term->config && strcmp(term->config, "event")) { + ret = parse_events__with_hybrid_pmu(parse_state, + term->config, name, + list, ); + /* +* If the string inside the pmu can't be parsed, +* don't return, try next steps. +*/ + if (parsed) + return ret; + } + } + if (!parse_state->fake_pmu && perf_pmu__config(pmu, , head_config, parse_state->error)) { struct evsel_config_term *pos, *tmp; @@ -1585,6 +1610,9 @@ int parse_events_add_pmu(struct parse_events_state *parse_state, if (!evsel) return -ENOMEM; + if (evsel->name) + evsel->use_config_name = true; + evsel->pmu_name = name ? strdup(name) : NULL; evsel->use_uncore_alias = use_uncore_alias; evsel->percore = config_term_percore(>config_terms); @@ -2180,6 +2208,33 @@ int parse_events_terms(struct list_head *terms, const char *str) return ret; } +static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, +const char *str, char *pmu_name, +struct list_head *list, bool *parsed) +{ + struct parse_events_state ps = { + .list= LIST_HEAD_INIT(ps.list), + .stoken = PE_START_EVENTS, + .hybrid_pmu_name = pmu_name, + .idx = parse_state->idx, + }; + int ret; + + *parsed = false; + ret = parse_events_
[PATCH v4 10/25] perf parse-events: Create two hybrid raw events
On hybrid platform, same raw event is possible to be available on both cpu_core pmu and cpu_atom pmu. It's supported to create two raw events for one event encoding. For raw events, the attr.type is PMU type. # perf stat -e r3c -a -vv -- sleep 1 Control descriptor is not initialized perf_event_attr: type 4 size 120 config 0x3c sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 3 ... perf_event_attr: type 4 size 120 config 0x3c sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 19 perf_event_attr: type 8 size 120 config 0x3c sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 20 ... perf_event_attr: type 8 size 120 config 0x3c sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 27 r3c: 0: 434449 1001412521 1001412521 r3c: 1: 173162 1001482031 1001482031 r3c: 2: 231710 1001524974 1001524974 r3c: 3: 110012 1001563523 1001563523 r3c: 4: 191517 1001593221 1001593221 r3c: 5: 956458 1001628147 1001628147 r3c: 6: 416969 1001715626 1001715626 r3c: 7: 1047527 1001596650 1001596650 r3c: 8: 103877 1001633520 1001633520 r3c: 9: 70571 1001637898 1001637898 r3c: 10: 550284 1001714398 1001714398 r3c: 11: 1257274 1001738349 1001738349 r3c: 12: 107797 1001801432 1001801432 r3c: 13: 67471 1001836281 1001836281 r3c: 14: 286782 1001923161 1001923161 r3c: 15: 815509 1001952550 1001952550 r3c: 0: 95994 1002071117 1002071117 r3c: 1: 105570 1002142438 1002142438 r3c: 2: 115921 1002189147 1002189147 r3c: 3: 72747 1002238133 1002238133 r3c: 4: 103519 1002276753 1002276753 r3c: 5: 121382 1002315131 1002315131 r3c: 6: 80298 1002248050 1002248050 r3c: 7: 466790 1002278221 1002278221 r3c: 6821369 16026754282 16026754282 r3c: 1162221 8017758990 8017758990 Performance counter stats for 'system wide': 6,821,369 cpu_core/r3c/ 1,162,221 cpu_atom/r3c/ 1.002289965 seconds time elapsed Signed-off-by: Jin Yao --- v4: - Directly return add_raw_hybrid(). v3: - Raw event creation is moved to parse-events-hybrid.c. tools/perf/util/parse-events-hybrid.c | 38 ++- 1 file changed, 37 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index 7a7e065d2b5f..e27b747080e1 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -77,6 +77,41 @@ static int add_hw_hybrid(struct parse_events_state *parse_state, return 0; } +static int create_raw_event_hybrid(int *idx, struct list_head *list, + struct perf_event_attr *attr, char *name, + struct list_head *config_terms, + struct perf_pmu *pmu) +{ + struct evsel *evsel; + + attr->type = pmu->type; + evsel = parse_events__add_event_hybrid(list, idx, attr, name, + pmu, config_terms); + if
[PATCH v4 08/25] perf parse-events: Create two hybrid hardware events
Current hardware events has special perf types PERF_TYPE_HARDWARE. But it doesn't pass the PMU type in the user interface. For a hybrid system, the perf kernel doesn't know which PMU the events belong to. So now this type is extended to be PMU aware type. The PMU type ID is stored at attr.config[63:32]. PMU type ID is retrieved from sysfs. root@lkp-adl-d01:/sys/devices/cpu_atom# cat type 8 root@lkp-adl-d01:/sys/devices/cpu_core# cat type 4 When enabling a hybrid hardware event without specified pmu, such as, 'perf stat -e cycles -a', two events are created automatically. One is for atom, the other is for core. # perf stat -e cycles -a -vv -- sleep 1 Control descriptor is not initialized perf_event_attr: size 120 config 0x4 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 3 ... perf_event_attr: size 120 config 0x4 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 19 perf_event_attr: size 120 config 0x8 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 20 ... perf_event_attr: size 120 config 0x8 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 27 cycles: 0: 836272 1001525722 1001525722 cycles: 1: 628564 1001580453 1001580453 cycles: 2: 872693 1001605997 1001605997 cycles: 3: 70417 1001641369 1001641369 cycles: 4: 88593 1001726722 1001726722 cycles: 5: 470495 1001752993 1001752993 cycles: 6: 484733 1001840440 1001840440 cycles: 7: 1272477 1001593105 1001593105 cycles: 8: 209185 1001608616 1001608616 cycles: 9: 204391 1001633962 1001633962 cycles: 10: 264121 1001661745 1001661745 cycles: 11: 826104 1001689904 1001689904 cycles: 12: 89935 1001728861 1001728861 cycles: 13: 70639 1001756757 1001756757 cycles: 14: 185266 1001784810 1001784810 cycles: 15: 171094 1001825466 1001825466 cycles: 0: 129624 1001854843 1001854843 cycles: 1: 122533 1001840421 1001840421 cycles: 2: 90055 1001882506 1001882506 cycles: 3: 139607 1001896463 1001896463 cycles: 4: 141791 1001907838 1001907838 cycles: 5: 530927 1001883880 1001883880 cycles: 6: 143246 1001852529 1001852529 cycles: 7: 667769 1001872626 1001872626 cycles: 6744979 16026956922 16026956922 cycles: 1965552 8014991106 8014991106 Performance counter stats for 'system wide': 6,744,979 cpu_core/cycles/ 1,965,552 cpu_atom/cycles/ 1.001882711 seconds time elapsed 0x4 in 0x4 indicates the cpu_core pmu. 0x8 in 0x8 indicates the cpu_atom pmu. Signed-off-by: Jin Yao --- v4: - Use PERF_TYPE_HARDWARE (v3 uses PERF_TYPE_HARDWARE_PMU) v3: - Create new parse-events-hybrid.c/parse-events-hybrid.h - Refine the code tools/perf/util/Build | 1 + tools/perf/util/parse-events-hybrid.c | 100 ++ tools/perf/util/parse-events-hybrid.h | 17 + tools/perf/util/parse-events.c| 18 + tools/perf/util/parse-events.h| 5 ++ 5 files changed, 141 insertions(+) create mode 100644 tools/perf/util/parse-events-hybrid.c create mode 100644 tools/perf/util/parse-events-hybrid.h diff --git
[PATCH v4 06/25] perf pmu: Add hybrid helper functions
The functions perf_pmu__is_hybrid and perf_pmu__find_hybrid_pmu can be used to identify the hybrid platform and return the found hybrid cpu pmu. All the detected hybrid pmus have been saved in 'perf_pmu__hybrid_pmus' list. So we just need to search this list. perf_pmu__hybrid_type_to_pmu converts the user specified string to hybrid pmu name. This is used to support the '--cputype' option in next patches. perf_pmu__has_hybrid checks the existing of hybrid pmu. Note that, we have to define it in pmu.c (make pmu-hybrid.c no more symbol dependency), otherwise perf test python would be failed. Signed-off-by: Jin Yao --- v4: - No change. v3: - Move perf_pmu__has_hybrid from pmu-hybrid.c to pmu.c. We have to add pmu-hybrid.c to python-ext-sources to solve symbol dependency issue found in perf test python. For perf_pmu__has_hybrid, it calls perf_pmu__scan, which is defined in pmu.c. It's very hard to add pmu.c to python-ext-sources, too much symbol dependency here. tools/perf/util/pmu-hybrid.c | 40 tools/perf/util/pmu-hybrid.h | 4 tools/perf/util/pmu.c| 11 ++ tools/perf/util/pmu.h| 2 ++ 4 files changed, 57 insertions(+) diff --git a/tools/perf/util/pmu-hybrid.c b/tools/perf/util/pmu-hybrid.c index 8ed0e6e1776d..f51ccaac60ee 100644 --- a/tools/perf/util/pmu-hybrid.c +++ b/tools/perf/util/pmu-hybrid.c @@ -47,3 +47,43 @@ bool perf_pmu__hybrid_mounted(const char *name) return true; } + +struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name) +{ + struct perf_pmu *pmu; + + if (!name) + return NULL; + + perf_pmu__for_each_hybrid_pmu(pmu) { + if (!strcmp(name, pmu->name)) + return pmu; + } + + return NULL; +} + +bool perf_pmu__is_hybrid(const char *name) +{ + return perf_pmu__find_hybrid_pmu(name) != NULL; +} + +char *perf_pmu__hybrid_type_to_pmu(const char *type) +{ + char *pmu_name = NULL; + + if (asprintf(_name, "cpu_%s", type) < 0) + return NULL; + + if (perf_pmu__is_hybrid(pmu_name)) + return pmu_name; + + /* +* pmu may be not scanned, check the sysfs. +*/ + if (perf_pmu__hybrid_mounted(pmu_name)) + return pmu_name; + + free(pmu_name); + return NULL; +} diff --git a/tools/perf/util/pmu-hybrid.h b/tools/perf/util/pmu-hybrid.h index 35bed3714438..d0fa7bc50a76 100644 --- a/tools/perf/util/pmu-hybrid.h +++ b/tools/perf/util/pmu-hybrid.h @@ -15,4 +15,8 @@ extern struct list_head perf_pmu__hybrid_pmus; bool perf_pmu__hybrid_mounted(const char *name); +struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name); +bool perf_pmu__is_hybrid(const char *name); +char *perf_pmu__hybrid_type_to_pmu(const char *type); + #endif /* __PMU_HYBRID_H */ diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 6e49c7b8ad71..88c8ecdc60b0 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -40,6 +40,7 @@ int perf_pmu_parse(struct list_head *list, char *name); extern FILE *perf_pmu_in; static LIST_HEAD(pmus); +static bool hybrid_scanned; /* * Parse & process all the sysfs attributes located under @@ -1861,3 +1862,13 @@ void perf_pmu__warn_invalid_config(struct perf_pmu *pmu, __u64 config, "'%llx' not supported by kernel)!\n", name ?: "N/A", buf, config); } + +bool perf_pmu__has_hybrid(void) +{ + if (!hybrid_scanned) { + hybrid_scanned = true; + perf_pmu__scan(NULL); + } + + return !list_empty(_pmu__hybrid_pmus); +} diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 9a2f89eeab6f..a790ef758171 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -132,4 +132,6 @@ int perf_pmu__caps_parse(struct perf_pmu *pmu); void perf_pmu__warn_invalid_config(struct perf_pmu *pmu, __u64 config, char *name); +bool perf_pmu__has_hybrid(void); + #endif /* __PMU_H */ -- 2.17.1
[PATCH v4 07/25] perf stat: Uniquify hybrid event name
It would be useful to tell user the pmu which the event belongs to. perf-stat has supported '--no-merge' option and it can print the pmu name after the event name, such as: "cycles [cpu_core]" Now this option is enabled by default for hybrid platform but change the format to: "cpu_core/cycles/" If user configs the name, we still use the user specified name. Signed-off-by: Jin Yao --- v4: - If user configs the name, we still use the user specified name. v3: - No change. tools/perf/builtin-stat.c | 4 tools/perf/util/evsel.h| 1 + tools/perf/util/stat-display.c | 15 +-- 3 files changed, 18 insertions(+), 2 deletions(-) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 2a2c15cac80a..1255af4751c2 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -68,6 +68,7 @@ #include "util/affinity.h" #include "util/pfm.h" #include "util/bpf_counter.h" +#include "util/pmu-hybrid.h" #include "asm/bug.h" #include @@ -2378,6 +2379,9 @@ int cmd_stat(int argc, const char **argv) evlist__check_cpu_maps(evsel_list); + if (perf_pmu__has_hybrid()) + stat_config.no_merge = true; + /* * Initialize thread_map with comm names, * so we could print it out on output. diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h index eccc4fd5b3eb..d518da2fd2eb 100644 --- a/tools/perf/util/evsel.h +++ b/tools/perf/util/evsel.h @@ -115,6 +115,7 @@ struct evsel { boolmerged_stat; boolreset_group; boolerrored; + booluse_config_name; struct hashmap *per_pkg_mask; struct evsel*leader; struct list_headconfig_terms; diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index d3137bc17065..5255d78b1c30 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -17,6 +17,7 @@ #include "cgroup.h" #include #include "util.h" +#include "pmu-hybrid.h" #define CNTR_NOT_SUPPORTED "" #define CNTR_NOT_COUNTED "" @@ -532,6 +533,7 @@ static void uniquify_event_name(struct evsel *counter) { char *new_name; char *config; + int ret = 0; if (counter->uniquified_name || !counter->pmu_name || !strncmp(counter->name, counter->pmu_name, @@ -546,8 +548,17 @@ static void uniquify_event_name(struct evsel *counter) counter->name = new_name; } } else { - if (asprintf(_name, -"%s [%s]", counter->name, counter->pmu_name) > 0) { + if (perf_pmu__has_hybrid()) { + if (!counter->use_config_name) { + ret = asprintf(_name, "%s/%s/", + counter->pmu_name, counter->name); + } + } else { + ret = asprintf(_name, "%s [%s]", + counter->name, counter->pmu_name); + } + + if (ret) { free(counter->name); counter->name = new_name; } -- 2.17.1
[PATCH v4 09/25] perf parse-events: Create two hybrid cache events
For cache events, they have pre-defined configs. The kernel needs to know where the cache event comes from (e.g. from cpu_core pmu or from cpu_atom pmu). But the perf type PERF_TYPE_HW_CACHE can't carry pmu information. Now the type PERF_TYPE_HW_CACHE is extended to be PMU aware type. The PMU type ID is stored at attr.config[63:32]. When enabling a hybrid cache event without specified pmu, such as, 'perf stat -e LLC-loads -a', two events are created automatically. One is for atom, the other is for core. # perf stat -e LLC-loads -a -vv -- sleep 1 Control descriptor is not initialized perf_event_attr: type 3 size 120 config 0x40002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 3 ... perf_event_attr: type 3 size 120 config 0x40002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 19 perf_event_attr: type 3 size 120 config 0x80002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 20 ... perf_event_attr: type 3 size 120 config 0x80002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 27 LLC-loads: 0: 1507 1001800280 1001800280 LLC-loads: 1: 666 1001812250 1001812250 LLC-loads: 2: 3353 1001813453 1001813453 LLC-loads: 3: 514 1001848795 1001848795 LLC-loads: 4: 627 1001952832 1001952832 LLC-loads: 5: 4399 1001451154 1001451154 LLC-loads: 6: 1240 1001481052 1001481052 LLC-loads: 7: 478 1001520348 1001520348 LLC-loads: 8: 691 1001551236 1001551236 LLC-loads: 9: 310 1001578945 1001578945 LLC-loads: 10: 1018 1001594354 1001594354 LLC-loads: 11: 3656 1001622355 1001622355 LLC-loads: 12: 882 1001661416 1001661416 LLC-loads: 13: 506 1001693963 1001693963 LLC-loads: 14: 3547 1001721013 1001721013 LLC-loads: 15: 1399 1001734818 1001734818 LLC-loads: 0: 1314 1001793826 1001793826 LLC-loads: 1: 2857 1001752764 1001752764 LLC-loads: 2: 646 1001830694 1001830694 LLC-loads: 3: 1612 1001864861 1001864861 LLC-loads: 4: 2244 1001912381 1001912381 LLC-loads: 5: 1255 1001943889 1001943889 LLC-loads: 6: 4624 1002021109 1002021109 LLC-loads: 7: 2703 1001959302 1001959302 LLC-loads: 24793 16026838264 16026838264 LLC-loads: 17255 8015078826 8015078826 Performance counter stats for 'system wide': 24,793 cpu_core/LLC-loads/ 17,255 cpu_atom/LLC-loads/ 1.001970988 seconds time elapsed 0x4 in 0x40002 indicates the cpu_core pmu. 0x8 in 0x80002 indicates the cpu_atom pmu. Signed-off-by: Jin Yao --- v4: - Use PERF_TYPE_HW_CACHE (v3 uses PERF_TYPE_HW_CACHE_PMU) - Define 'ret' variable for return value. v3: - Raw event creation is moved to parse-events-hybrid.c. tools/perf/util/parse-events-hybrid.c | 23 +++ tools/perf/util/parse-events-hybrid.h | 5 + tools/perf/util/parse-events.c| 10 +- 3 files changed, 37 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index 8fd7f19a9865..7a7e065d2b5f
[PATCH v4 05/25] perf pmu: Save detected hybrid pmus to a global pmu list
We identify the cpu_core pmu and cpu_atom pmu by explicitly checking following files: For cpu_core, checks: "/sys/bus/event_source/devices/cpu_core/cpus" For cpu_atom, checks: "/sys/bus/event_source/devices/cpu_atom/cpus" If the 'cpus' file exists and it has data, the pmu exists. But in order not to hardcode the "cpu_core" and "cpu_atom", and make the code in a generic way. So if the path "/sys/bus/event_source/devices/cpu_xxx/cpus" exists, the hybrid pmu exists. All the detected hybrid pmus are linked to a global list 'perf_pmu__hybrid_pmus' and then next we just need to iterate the list to get all hybrid pmu by using perf_pmu__for_each_hybrid_pmu. Signed-off-by: Jin Yao --- v4: - Check if 'cpus' file is empty. If so, don't create pmu. v3: - No functional change. tools/perf/util/Build| 1 + tools/perf/util/pmu-hybrid.c | 49 tools/perf/util/pmu-hybrid.h | 18 + tools/perf/util/pmu.c| 9 ++- tools/perf/util/pmu.h| 4 +++ 5 files changed, 80 insertions(+), 1 deletion(-) create mode 100644 tools/perf/util/pmu-hybrid.c create mode 100644 tools/perf/util/pmu-hybrid.h diff --git a/tools/perf/util/Build b/tools/perf/util/Build index e3e12f9d4733..37a8a63c7195 100644 --- a/tools/perf/util/Build +++ b/tools/perf/util/Build @@ -69,6 +69,7 @@ perf-y += parse-events-bison.o perf-y += pmu.o perf-y += pmu-flex.o perf-y += pmu-bison.o +perf-y += pmu-hybrid.o perf-y += trace-event-read.o perf-y += trace-event-info.o perf-y += trace-event-scripting.o diff --git a/tools/perf/util/pmu-hybrid.c b/tools/perf/util/pmu-hybrid.c new file mode 100644 index ..8ed0e6e1776d --- /dev/null +++ b/tools/perf/util/pmu-hybrid.c @@ -0,0 +1,49 @@ +// SPDX-License-Identifier: GPL-2.0 +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "fncache.h" +#include "pmu-hybrid.h" + +LIST_HEAD(perf_pmu__hybrid_pmus); + +bool perf_pmu__hybrid_mounted(const char *name) +{ + char path[PATH_MAX]; + const char *sysfs; + FILE *file; + int n, cpu; + + if (strncmp(name, "cpu_", 4)) + return false; + + sysfs = sysfs__mountpoint(); + if (!sysfs) + return false; + + snprintf(path, PATH_MAX, CPUS_TEMPLATE_CPU, sysfs, name); + if (!file_available(path)) + return false; + + file = fopen(path, "r"); + if (!file) + return false; + + n = fscanf(file, "%u", ); + fclose(file); + if (n <= 0) + return false; + + return true; +} diff --git a/tools/perf/util/pmu-hybrid.h b/tools/perf/util/pmu-hybrid.h new file mode 100644 index ..35bed3714438 --- /dev/null +++ b/tools/perf/util/pmu-hybrid.h @@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef __PMU_HYBRID_H +#define __PMU_HYBRID_H + +#include +#include +#include +#include +#include "pmu.h" + +extern struct list_head perf_pmu__hybrid_pmus; + +#define perf_pmu__for_each_hybrid_pmu(pmu) \ + list_for_each_entry(pmu, _pmu__hybrid_pmus, hybrid_list) + +bool perf_pmu__hybrid_mounted(const char *name); + +#endif /* __PMU_HYBRID_H */ diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 44225838eb03..6e49c7b8ad71 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -25,6 +25,7 @@ #include "string2.h" #include "strbuf.h" #include "fncache.h" +#include "pmu-hybrid.h" struct perf_pmu perf_pmu__fake; @@ -613,7 +614,6 @@ static struct perf_cpu_map *__pmu_cpumask(const char *path) */ #define SYS_TEMPLATE_ID"./bus/event_source/devices/%s/identifier" #define CPUS_TEMPLATE_UNCORE "%s/bus/event_source/devices/%s/cpumask" -#define CPUS_TEMPLATE_CPU "%s/bus/event_source/devices/%s/cpus" static struct perf_cpu_map *pmu_cpumask(const char *name) { @@ -645,6 +645,9 @@ static bool pmu_is_uncore(const char *name) char path[PATH_MAX]; const char *sysfs; + if (perf_pmu__hybrid_mounted(name)) + return false; + sysfs = sysfs__mountpoint(); snprintf(path, PATH_MAX, CPUS_TEMPLATE_UNCORE, sysfs, name); return file_available(path); @@ -951,6 +954,7 @@ static struct perf_pmu *pmu_lookup(const char *name) pmu->is_uncore = pmu_is_uncore(name); if (pmu->is_uncore) pmu->id = pmu_id(name); + pmu->is_hybrid = perf_pmu__hybrid_mounted(name); pmu->max_precise = pmu_max_precise(name); pmu_add_cpu_aliases(, pmu); pmu_add_sys_aliases(, pmu); @@ -962,6 +966,9 @@ static struct perf_pmu *pmu_lookup(const char *name) list_splice(, >aliases); list_add_tail(>list,
[PATCH v4 02/25] perf jevents: Support unit value "cpu_core" and "cpu_atom"
For some Intel platforms, such as Alderlake, which is a hybrid platform and it consists of atom cpu and core cpu. Each cpu has dedicated event list. Part of events are available on core cpu, part of events are available on atom cpu. The kernel exports new cpu pmus: cpu_core and cpu_atom. The event in json is added with a new field "Unit" to indicate which pmu the event is available on. For example, one event in cache.json, { "BriefDescription": "Counts the number of load ops retired that", "CollectPEBSRecord": "2", "Counter": "0,1,2,3", "EventCode": "0xd2", "EventName": "MEM_LOAD_UOPS_RETIRED_MISC.MMIO", "PEBScounters": "0,1,2,3", "SampleAfterValue": "103", "UMask": "0x80", "Unit": "cpu_atom" }, The unit "cpu_atom" indicates this event is only available on "cpu_atom". In generated pmu-events.c, we can see: { .name = "mem_load_uops_retired_misc.mmio", .event = "period=103,umask=0x80,event=0xd2", .desc = "Counts the number of load ops retired that. Unit: cpu_atom ", .topic = "cache", .pmu = "cpu_atom", }, But if without this patch, the "uncore_" prefix is added before "cpu_atom", such as: .pmu = "uncore_cpu_atom" That would be a wrong pmu. Signed-off-by: Jin Yao --- v4: - No change. tools/perf/pmu-events/jevents.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tools/perf/pmu-events/jevents.c b/tools/perf/pmu-events/jevents.c index 33aa3c885eaf..ed4f0bd72e5a 100644 --- a/tools/perf/pmu-events/jevents.c +++ b/tools/perf/pmu-events/jevents.c @@ -285,6 +285,8 @@ static struct map { { "imx8_ddr", "imx8_ddr" }, { "L3PMC", "amd_l3" }, { "DFPMC", "amd_df" }, + { "cpu_core", "cpu_core" }, + { "cpu_atom", "cpu_atom" }, {} }; -- 2.17.1
[PATCH v4 03/25] perf pmu: Simplify arguments of __perf_pmu__new_alias
Simplify the arguments of __perf_pmu__new_alias() by passing the whole 'struct pme_event' pointer. Signed-off-by: Jin Yao --- v4: - No change. tools/perf/util/pmu.c | 36 1 file changed, 16 insertions(+), 20 deletions(-) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 286d5e415bdc..8214def7b0f0 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -306,18 +306,25 @@ static bool perf_pmu_merge_alias(struct perf_pmu_alias *newalias, } static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, -char *desc, char *val, -char *long_desc, char *topic, -char *unit, char *perpkg, -char *metric_expr, -char *metric_name, -char *deprecated) +char *desc, char *val, struct pmu_event *pe) { struct parse_events_term *term; struct perf_pmu_alias *alias; int ret; int num; char newval[256]; + char *long_desc = NULL, *topic = NULL, *unit = NULL, *perpkg = NULL, +*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL; + + if (pe) { + long_desc = (char *)pe->long_desc; + topic = (char *)pe->topic; + unit = (char *)pe->unit; + perpkg = (char *)pe->perpkg; + metric_expr = (char *)pe->metric_expr; + metric_name = (char *)pe->metric_name; + deprecated = (char *)pe->deprecated; + } alias = malloc(sizeof(*alias)); if (!alias) @@ -406,8 +413,7 @@ static int perf_pmu__new_alias(struct list_head *list, char *dir, char *name, FI /* Remove trailing newline from sysfs file */ strim(buf); - return __perf_pmu__new_alias(list, dir, name, NULL, buf, NULL, NULL, NULL, -NULL, NULL, NULL, NULL); + return __perf_pmu__new_alias(list, dir, name, NULL, buf, NULL); } static inline bool pmu_alias_info_file(char *name) @@ -798,11 +804,7 @@ void pmu_add_cpu_aliases_map(struct list_head *head, struct perf_pmu *pmu, /* need type casts to override 'const' */ __perf_pmu__new_alias(head, NULL, (char *)pe->name, (char *)pe->desc, (char *)pe->event, - (char *)pe->long_desc, (char *)pe->topic, - (char *)pe->unit, (char *)pe->perpkg, - (char *)pe->metric_expr, - (char *)pe->metric_name, - (char *)pe->deprecated); + pe); } } @@ -869,13 +871,7 @@ static int pmu_add_sys_aliases_iter_fn(struct pmu_event *pe, void *data) (char *)pe->name, (char *)pe->desc, (char *)pe->event, - (char *)pe->long_desc, - (char *)pe->topic, - (char *)pe->unit, - (char *)pe->perpkg, - (char *)pe->metric_expr, - (char *)pe->metric_name, - (char *)pe->deprecated); + pe); } return 0; -- 2.17.1
[PATCH v4 04/25] perf pmu: Save pmu name
On hybrid platform, one event is available on one pmu (such as, available on cpu_core or on cpu_atom). This patch saves the pmu name to the pmu field of struct perf_pmu_alias. Then next we can know the pmu which the event can be enabled on. Signed-off-by: Jin Yao --- v4: - No change. v3: - Change pmu to pmu_name in struct perf_pmu_alias. tools/perf/util/pmu.c | 10 +- tools/perf/util/pmu.h | 1 + 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 8214def7b0f0..44225838eb03 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -283,6 +283,7 @@ void perf_pmu_free_alias(struct perf_pmu_alias *newalias) zfree(>str); zfree(>metric_expr); zfree(>metric_name); + zfree(>pmu_name); parse_events_terms__purge(>terms); free(newalias); } @@ -297,6 +298,10 @@ static bool perf_pmu_merge_alias(struct perf_pmu_alias *newalias, list_for_each_entry(a, alist, list) { if (!strcasecmp(newalias->name, a->name)) { + if (newalias->pmu_name && a->pmu_name && + !strcasecmp(newalias->pmu_name, a->pmu_name)) { + continue; + } perf_pmu_update_alias(a, newalias); perf_pmu_free_alias(newalias); return true; @@ -314,7 +319,8 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, int num; char newval[256]; char *long_desc = NULL, *topic = NULL, *unit = NULL, *perpkg = NULL, -*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL; +*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL, +*pmu_name = NULL; if (pe) { long_desc = (char *)pe->long_desc; @@ -324,6 +330,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, metric_expr = (char *)pe->metric_expr; metric_name = (char *)pe->metric_name; deprecated = (char *)pe->deprecated; + pmu_name = (char *)pe->pmu; } alias = malloc(sizeof(*alias)); @@ -389,6 +396,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, } alias->per_pkg = perpkg && sscanf(perpkg, "%d", ) == 1 && num == 1; alias->str = strdup(newval); + alias->pmu_name = pmu_name ? strdup(pmu_name) : NULL; if (deprecated) alias->deprecated = true; diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 1f1749ba830f..4f100768c264 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -72,6 +72,7 @@ struct perf_pmu_alias { bool deprecated; char *metric_expr; char *metric_name; + char *pmu_name; }; struct perf_pmu *perf_pmu__find(const char *name); -- 2.17.1
[PATCH v4 01/25] tools headers uapi: Update tools's copy of linux/perf_event.h
To get the changes in: Liang Kan's patch [PATCH V6 21/25] perf: Extend PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE Kan's patch is in review at the moment. The following perf tool patches need this interface for hybrid support. This patch can be removed after Kan's patch is upstreamed. Signed-off-by: Jin Yao --- v4: - Updated by Kan's latest patch, '[PATCH V6 21/25] perf: Extend PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE' include/uapi/linux/perf_event.h | 15 +++ tools/include/uapi/linux/perf_event.h | 15 +++ 2 files changed, 30 insertions(+) diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h index ad15e40d7f5d..14332f4cf816 100644 --- a/include/uapi/linux/perf_event.h +++ b/include/uapi/linux/perf_event.h @@ -37,6 +37,21 @@ enum perf_type_id { PERF_TYPE_MAX, /* non-ABI */ }; +/* + * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE + * PERF_TYPE_HARDWARE: 0x00AA + * AA: hardware event ID + * : PMU type ID + * PERF_TYPE_HW_CACHE: 0x00DDCCBB + * BB: hardware cache ID + * CC: hardware cache op ID + * DD: hardware cache op result ID + * : PMU type ID + * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied. + */ +#define PERF_PMU_TYPE_SHIFT32 +#define PERF_HW_EVENT_MASK 0x + /* * Generalized performance event event_id types, used by the * attr.event_id parameter of the sys_perf_event_open() diff --git a/tools/include/uapi/linux/perf_event.h b/tools/include/uapi/linux/perf_event.h index ad15e40d7f5d..14332f4cf816 100644 --- a/tools/include/uapi/linux/perf_event.h +++ b/tools/include/uapi/linux/perf_event.h @@ -37,6 +37,21 @@ enum perf_type_id { PERF_TYPE_MAX, /* non-ABI */ }; +/* + * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE + * PERF_TYPE_HARDWARE: 0x00AA + * AA: hardware event ID + * : PMU type ID + * PERF_TYPE_HW_CACHE: 0x00DDCCBB + * BB: hardware cache ID + * CC: hardware cache op ID + * DD: hardware cache op result ID + * : PMU type ID + * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied. + */ +#define PERF_PMU_TYPE_SHIFT32 +#define PERF_HW_EVENT_MASK 0x + /* * Generalized performance event event_id types, used by the * attr.event_id parameter of the sys_perf_event_open() -- 2.17.1
[PATCH v4 00/25] perf tool: AlderLake hybrid support series 1
uot;. - Move command output two chars to the right. - Move pmu hybrid functions to new created pmu-hybrid.c/pmu-hybrid.h. This is to pass the perf test python case. Jin Yao (25): tools headers uapi: Update tools's copy of linux/perf_event.h perf jevents: Support unit value "cpu_core" and "cpu_atom" perf pmu: Simplify arguments of __perf_pmu__new_alias perf pmu: Save pmu name perf pmu: Save detected hybrid pmus to a global pmu list perf pmu: Add hybrid helper functions perf stat: Uniquify hybrid event name perf parse-events: Create two hybrid hardware events perf parse-events: Create two hybrid cache events perf parse-events: Create two hybrid raw events perf parse-events: Compare with hybrid pmu name perf parse-events: Support event inside hybrid pmu perf record: Create two hybrid 'cycles' events by default perf stat: Add default hybrid events perf stat: Filter out unmatched aggregation for hybrid event perf stat: Warn group events from different hybrid PMU perf record: Uniquify hybrid event name perf tests: Add hybrid cases for 'Parse event definition strings' test perf tests: Add hybrid cases for 'Roundtrip evsel->name' test perf tests: Skip 'Setup struct perf_event_attr' test for hybrid perf tests: Support 'Track with sched_switch' test for hybrid perf tests: Support 'Parse and process metrics' test for hybrid perf tests: Support 'Session topology' test for hybrid perf tests: Support 'Convert perf time to TSC' test for hybrid perf tests: Skip 'perf stat metrics (shadow stat) test' for hybrid include/uapi/linux/perf_event.h| 15 ++ tools/include/uapi/linux/perf_event.h | 15 ++ tools/perf/builtin-record.c| 47 +- tools/perf/builtin-stat.c | 29 tools/perf/pmu-events/jevents.c| 2 + tools/perf/tests/attr.c| 4 + tools/perf/tests/evsel-roundtrip-name.c| 19 ++- tools/perf/tests/parse-events.c| 152 ++ tools/perf/tests/parse-metric.c| 10 +- tools/perf/tests/perf-time-to-tsc.c| 16 ++ tools/perf/tests/shell/stat+shadow_stat.sh | 3 + tools/perf/tests/switch-tracking.c | 10 +- tools/perf/tests/topology.c| 10 +- tools/perf/util/Build | 3 + tools/perf/util/evlist-hybrid.c| 88 ++ tools/perf/util/evlist-hybrid.h| 14 ++ tools/perf/util/evlist.c | 5 +- tools/perf/util/evsel.c| 12 +- tools/perf/util/evsel.h| 4 +- tools/perf/util/parse-events-hybrid.c | 178 + tools/perf/util/parse-events-hybrid.h | 23 +++ tools/perf/util/parse-events.c | 86 +- tools/perf/util/parse-events.h | 9 +- tools/perf/util/parse-events.y | 9 +- tools/perf/util/pmu-hybrid.c | 89 +++ tools/perf/util/pmu-hybrid.h | 22 +++ tools/perf/util/pmu.c | 64 +--- tools/perf/util/pmu.h | 7 + tools/perf/util/python-ext-sources | 2 + tools/perf/util/stat-display.c | 35 +++- 30 files changed, 933 insertions(+), 49 deletions(-) create mode 100644 tools/perf/util/evlist-hybrid.c create mode 100644 tools/perf/util/evlist-hybrid.h create mode 100644 tools/perf/util/parse-events-hybrid.c create mode 100644 tools/perf/util/parse-events-hybrid.h create mode 100644 tools/perf/util/pmu-hybrid.c create mode 100644 tools/perf/util/pmu-hybrid.h -- 2.17.1
Re: [PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU
Hi Jiri, On 4/16/2021 3:39 AM, Jiri Olsa wrote: On Thu, Apr 15, 2021 at 10:53:33PM +0800, Jin, Yao wrote: SNIP With my current code, static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, const char *str, char *pmu_name, struct list_head *list) { struct parse_events_state ps = { .list = LIST_HEAD_INIT(ps.list), .stoken = PE_START_EVENTS, .pmu_name = pmu_name, .idx= parse_state->idx, }; int ret; ret = parse_events__scanner(str, ); perf_pmu__parse_cleanup(); if (!ret) { if (!list_empty()) { list_splice(, list); parse_state->idx = ps.idx; } } return ret; } The new created evsels are added to the tail of list (ps.list) and ps.list is joined to the list (the parameter 'list'). If we want to reuse the __parse_events(), we may need to: struct evlist *evlist = evlist__new(); there's the original evlist pointer passed to the initial parser that we should use no? Unfortunately the answer is no. :( For "cpu_core/LLC-loads/", if we do the parser twice by just calling __parse_events, actually the __parse_events will be called two times. int __parse_events(struct evlist *evlist, const char *str, struct parse_events_error *err, struct perf_pmu *fake_pmu, char *pmu_name) { struct parse_events_state parse_state = { .list = LIST_HEAD_INIT(parse_state.list), ... }; ret = parse_events__scanner(str, _state); perf_pmu__parse_cleanup(); if (!ret && list_empty(_state.list)) { WARN_ONCE(true, "WARNING: event parser found nothing\n"); return -1; } ... } When returning to the first __parse_events,'parse_state.list' is an empty list so it would return "WARNING: event parser found nothing". So in my patch, I pass a list pointer in and the new created evsels will be added to this list. __parse_events(evlist, str, NULL, NULL); Add the evsels in evlist to the tail of list (the parameter 'list') evlist__delete(evlist); Is my understanding correct? Yes, we have to change the interface of __parse_events() by adding a new parameter 'pmu_name', which will bring much more changes. I agree to make this change in follow-up patches. let's check on this over the next version That's fine, thanks. Thanks Jin Yao thanks, jirka
Re: [PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU
Hi Jiri, On 4/15/2021 10:11 PM, Jiri Olsa wrote: On Thu, Apr 15, 2021 at 09:36:16PM +0800, Jin, Yao wrote: SNIP + int n = 0; + + list_for_each(pos, list) + n++; + + return n; +} + +static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, +const char *str, char *pmu_name, +bool *found, struct list_head *list) +{ + struct parse_events_state ps = { + .list = LIST_HEAD_INIT(ps.list), + .stoken = PE_START_EVENTS, + .pmu_name = pmu_name, + .idx= parse_state->idx, + }; could we add this pmu_name directly to __parse_events? Do you suggest we directly call __parse_events()? int __parse_events(struct evlist *evlist, const char *str, struct parse_events_error *err, struct perf_pmu *fake_pmu) struct parse_events_state parse_state = { .list = LIST_HEAD_INIT(parse_state.list), .idx = evlist->core.nr_entries, .error= err, .evlist = evlist, .stoken = PE_START_EVENTS, .fake_pmu = fake_pmu, }; But for parse_events__with_hybrid_pmu, we don't have valid evlist. So if we switch to __parse_events, evlist processing may be a problem. you should use parse_state->evlist no? but we can chec/make this change in next itaration.. it's already lot of changes jirka With my current code, static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, const char *str, char *pmu_name, struct list_head *list) { struct parse_events_state ps = { .list = LIST_HEAD_INIT(ps.list), .stoken = PE_START_EVENTS, .pmu_name = pmu_name, .idx= parse_state->idx, }; int ret; ret = parse_events__scanner(str, ); perf_pmu__parse_cleanup(); if (!ret) { if (!list_empty()) { list_splice(, list); parse_state->idx = ps.idx; } } return ret; } The new created evsels are added to the tail of list (ps.list) and ps.list is joined to the list (the parameter 'list'). If we want to reuse the __parse_events(), we may need to: struct evlist *evlist = evlist__new(); __parse_events(evlist, str, NULL, NULL); Add the evsels in evlist to the tail of list (the parameter 'list') evlist__delete(evlist); Is my understanding correct? Yes, we have to change the interface of __parse_events() by adding a new parameter 'pmu_name', which will bring much more changes. I agree to make this change in follow-up patches. Thanks Jin Yao
Re: [PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU
Hi Jiri, On 4/15/2021 7:03 PM, Jiri Olsa wrote: On Mon, Mar 29, 2021 at 03:00:31PM +0800, Jin Yao wrote: SNIP --- v3: - Rename the patch: 'perf parse-events: Support hardware events inside PMU' --> 'perf parse-events: Support no alias assigned event inside hybrid PMU' - Major code is moved to parse-events-hybrid.c. - Refine the code. tools/perf/util/parse-events-hybrid.c | 18 +- tools/perf/util/parse-events-hybrid.h | 3 +- tools/perf/util/parse-events.c| 80 +-- tools/perf/util/parse-events.h| 4 +- tools/perf/util/parse-events.y| 9 ++- tools/perf/util/pmu.c | 4 +- tools/perf/util/pmu.h | 2 +- 7 files changed, 108 insertions(+), 12 deletions(-) please move the support to pass pmu_name and filter on it within hybrid code in to separate patch OK. diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index 8a630cbab8f3..5bf176b55573 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -64,6 +64,11 @@ static int add_hw_hybrid(struct parse_events_state *parse_state, int ret; perf_pmu__for_each_hybrid_pmu(pmu) { + if (parse_state->pmu_name && + strcmp(parse_state->pmu_name, pmu->name)) { + continue; please add this check to separate function if (pmu_cmp(parse_stat)) continue; OK. SNIP + if (!parse_state->fake_pmu && head_config && !found && + perf_pmu__is_hybrid(name)) { + struct parse_events_term *term; + int ret; + + list_for_each_entry(term, head_config, list) { + if (!term->config) + continue; + + ret = parse_events__with_hybrid_pmu(parse_state, + term->config, + name, , + list); + if (found) + return ret; what if there are more terms in head_config? should we make sure there's just one term and fail if there's more? Yes, it should have only one term in head_config. Now I change the code to: + if (!parse_state->fake_pmu && head_config && !found && + perf_pmu__is_hybrid(name)) { + struct parse_events_term *term; + + term = list_first_entry(head_config, struct parse_events_term, + list); + if (term->config) { + return parse_events__with_hybrid_pmu(parse_state, +term->config, +name, list); + } + } also we already know the perf_pmu__is_hybrid(name) is true, so can't we just call: return parse_events__with_hybrid_pmu() Yes, we can direct return parse_events__with_hybrid_pmu(). + } + } if (verbose > 1) { fprintf(stderr, "After aliases, add event pmu '%s' with '", @@ -1605,6 +1630,15 @@ int parse_events_multi_pmu_add(struct parse_events_state *parse_state, struct perf_pmu *pmu = NULL; int ok = 0; + if (parse_state->pmu_name) { + list = malloc(sizeof(struct list_head)); + if (!list) + return -1; + INIT_LIST_HEAD(list); + *listp = list; + return 0; + } hum, why is this needed? Hmm, it's not necessary in new code, sorry about that. + *listp = NULL; /* Add it for all PMUs that support the alias */ list = malloc(sizeof(struct list_head)); @@ -2176,6 +2210,44 @@ int parse_events_terms(struct list_head *terms, const char *str) return ret; } +static int list_entries_nr(struct list_head *list) +{ + struct list_head *pos; + int n = 0; + + list_for_each(pos, list) + n++; + + return n; +} + +static int parse_events__with_hybrid_pmu(struct parse_events_state *parse_state, +const char *str, char *pmu_name, +bool *found, struct list_head *list) +{ + struct parse_events_state ps = { + .list = LIST_HEAD_INIT(ps.list), + .stoken = PE_START_EVENTS, + .pmu_name = pmu_name, + .idx= parse_state->idx, + }; could we add this pmu_name directly to __parse_events? Do you suggest we directly call __parse_events()? int __parse_events(struct evlist *evlist, const
Re: [PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU
Hi Jiri, On 4/9/2021 9:47 PM, Jiri Olsa wrote: On Mon, Mar 29, 2021 at 03:00:31PM +0800, Jin Yao wrote: SNIP + struct parse_events_state *parse_state) { struct perf_event_attr attr; LIST_HEAD(config_terms); @@ -521,7 +526,7 @@ int parse_events_add_cache(struct list_head *list, int *idx, i = parse_events__add_cache_hybrid(list, idx, , config_name ? : name, _terms, - ); + , parse_state); if (hybrid) return i; @@ -1481,7 +1486,7 @@ int parse_events_add_pmu(struct parse_events_state *parse_state, struct perf_pmu *pmu; struct evsel *evsel; struct parse_events_error *err = parse_state->error; - bool use_uncore_alias; + bool use_uncore_alias, found = false; LIST_HEAD(config_terms); if (verbose > 1) { @@ -1530,8 +1535,28 @@ int parse_events_add_pmu(struct parse_events_state *parse_state, } } - if (!parse_state->fake_pmu && perf_pmu__check_alias(pmu, head_config, )) + if (!parse_state->fake_pmu && + perf_pmu__check_alias(pmu, head_config, , )) { return -EINVAL; + } + ok, let's not polute surronding functions and make strict check on what we want in here.. we are after following events: cpu_xxx/L1-dcache/ cpu_xxx/l1-d|/ ... right? Yes, we only focus on the cache events now. so we are after events with single term in head_config that has name in: L1-dcache|l1-d|l1d|L1-data | L1-icache|l1-i|l1i|L1-instruction | LLC|L2 | dTLB|d-tlb|Data-TLB | iTLB|i-tlb|Instruction-TLB | branch|branches|bpu|btb|bpc | node I think that with such direct check the code will be more straight forward, also let's move it to parse-events-hybrid Do you suggest we just use string comparison for doing the direct check? e.g. if (strstr(term->config, "L1-dcache")) ... Of course, we can define a string array first and use a loop for string comparison. + if (!parse_state->fake_pmu && head_config && !found && + perf_pmu__is_hybrid(name)) { + struct parse_events_term *term; + int ret; + + list_for_each_entry(term, head_config, list) { + if (!term->config) + continue; + + ret = parse_events__with_hybrid_pmu(parse_state, + term->config, + name, , + list); do we need to call the parsing again? could we just call parse_events__add_cache_hybrid? jirka If we do the direct check for cache events, I think we don't need the parsing again. As I mentioned above, we need to define a string array and compare with term->config one by one. I'm OK for this solution. :) Thanks Jin Yao + if (found) + return ret; + } + } if (verbose > 1) { fprintf(stderr, "After aliases, add event pmu '%s' with '", @@ -1605,6 +1630,15 @@ int parse_events_multi_pmu_add(struct parse_events_state *parse_state, struct perf_pmu *pmu = NULL; int ok = 0; SNIP
Re: [PATCH v3 11/27] perf pmu: Support 'cycles' and 'branches' inside hybrid PMU
Hi Jiri, On 4/9/2021 9:48 PM, Jiri Olsa wrote: On Mon, Mar 29, 2021 at 03:00:30PM +0800, Jin Yao wrote: On hybrid platform, user may want to enable the hardware event only on one PMU. So following syntax is supported: cpu_core// cpu_atom// # perf stat -e cpu_core/cpu-cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 6,049,336 cpu_core/cpu-cycles/ 1.003577042 seconds time elapsed It enables the event 'cpu-cycles' only on cpu_core pmu. But for 'cycles' and 'branches', the syntax doesn't work. because the alias is not there.. but there's: cpu/cpu-cycles/ cpu/branch-instructions/ doing the same thing.. what's wrong with that? I have a feeling we discussed this in the previous version.. did I give up? ;-) Yes, we discussed this in previous threads. :) Now I'm fine to keep the original behavior. Because the syntax 'cpu/cycles/' and 'cpu/branches/' are not supported by current perf. SNIP diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index beff29981101..72e5ae5e868e 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -916,6 +916,35 @@ static int pmu_max_precise(const char *name) return max_precise; } +static void perf_pmu__add_hybrid_aliases(struct list_head *head) +{ + static struct pmu_event pme_hybrid_fixup[] = { + { + .name = "cycles", + .event = "event=0x3c", + }, + { + .name = "branches", + .event = "event=0xc4", + }, + { + .name = 0, + .event = 0, + }, if you really need to access these 2 events with special name, why not add it through the json.. let's not have yet another place that defines aliases ... also this should be model specific no? Yes, defining in json is a good idea if we really need to support 'cpu/cycles/' and 'cpu/branches/'. Anyway, I will drop this patch in next version. Thanks Jin Yao jirka
Re: [PATCH v3 09/27] perf parse-events: Create two hybrid cache events
Hi Jiri, On 4/9/2021 9:48 PM, Jiri Olsa wrote: On Mon, Mar 29, 2021 at 03:00:28PM +0800, Jin Yao wrote: SNIP index 1bbd0ba92ba7..3692fa3c964a 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -458,6 +458,7 @@ int parse_events_add_cache(struct list_head *list, int *idx, int cache_type = -1, cache_op = -1, cache_result = -1; char *op_result[2] = { op_result1, op_result2 }; int i, n; + bool hybrid; /* * No fallback - if we cannot get a clear cache type @@ -517,6 +518,13 @@ int parse_events_add_cache(struct list_head *list, int *idx, if (get_config_terms(head_config, _terms)) return -ENOMEM; } + + i = parse_events__add_cache_hybrid(list, idx, , + config_name ? : name, _terms, + ); + if (hybrid) + return i; please define 'ret' for the return value, i is confusing thanks, jirka Previously I wanted to save a 'ret' variable, but yes it's confusing. I will define 'ret' in next version. Thanks Jin Yao + return add_event(list, idx, , config_name ? : name, _terms); } -- 2.17.1
Re: [PATCH v3 10/27] perf parse-events: Create two hybrid raw events
Hi Jiri, On 4/9/2021 9:49 PM, Jiri Olsa wrote: On Mon, Mar 29, 2021 at 03:00:29PM +0800, Jin Yao wrote: SNIP + name, config_terms, pmu); + if (ret) + return ret; + } + + return 0; +} + int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, struct list_head *list, struct perf_event_attr *attr, @@ -91,6 +126,9 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, if (attr->type != PERF_TYPE_RAW) { return add_hw_hybrid(parse_state, list, attr, name, config_terms); + } else { + return add_raw_hybrid(parse_state, list, attr, name, + config_terms); } return -1; no need for the return -1 jirka Yes, no need return -1 here. if (attr->type != PERF_TYPE_RAW) { return add_hw_hybrid(parse_state, list, attr, name, config_terms); } return add_raw_hybrid(parse_state, list, attr, name, config_terms); Thanks Jin Yao -- 2.17.1
[PATCH] perf report: Fix wrong LBR block sorting
[get-dynamic-info.h:102 -> get-dynamic-info.h:111]ld-2.27.so 0.08% 820.06% 11 [intel_pmu_drain_pebs_nhm+580 -> intel_pmu_drain_pebs_nhm+627] [kernel.kallsyms] 0.08% 770.42% 77 [lru_add_drain_cpu+0 -> lru_add_drain_cpu+133] [kernel.kallsyms] 0.08% 740.10% 18 [handle_pmi_common+271 -> handle_pmi_common+310] [kernel.kallsyms] 0.08% 740.40% 74 [get-dynamic-info.h:131 -> get-dynamic-info.h:157]ld-2.27.so 0.07% 690.09% 17 [intel_pmu_drain_pebs_nhm+432 -> intel_pmu_drain_pebs_nhm+468] [kernel.kallsyms] Now the hottest block is reported at the top of output. Fixes: b65a7d372b1a ("perf hist: Support block formats with compare/sort/display") Signed-off-by: Jin Yao --- tools/perf/util/block-info.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tools/perf/util/block-info.c b/tools/perf/util/block-info.c index 423ec69bda6c..5ecd4f401f32 100644 --- a/tools/perf/util/block-info.c +++ b/tools/perf/util/block-info.c @@ -201,7 +201,7 @@ static int block_total_cycles_pct_entry(struct perf_hpp_fmt *fmt, double ratio = 0.0; if (block_fmt->total_cycles) - ratio = (double)bi->cycles / (double)block_fmt->total_cycles; + ratio = (double)bi->cycles_aggr / (double)block_fmt->total_cycles; return color_pct(hpp, block_fmt->width, 100.0 * ratio); } @@ -216,9 +216,9 @@ static int64_t block_total_cycles_pct_sort(struct perf_hpp_fmt *fmt, double l, r; if (block_fmt->total_cycles) { - l = ((double)bi_l->cycles / + l = ((double)bi_l->cycles_aggr / (double)block_fmt->total_cycles) * 10.0; - r = ((double)bi_r->cycles / + r = ((double)bi_r->cycles_aggr / (double)block_fmt->total_cycles) * 10.0; return (int64_t)l - (int64_t)r; } -- 2.17.1
[PATCH] perf vendor events: Add missing model numbers
Kernel has supported COMETLAKE/COMETLAKE_L to use the SKYLAKE events and supported TIGERLAKE_L/TIGERLAKE/ROCKETLAKE to use the ICELAKE events. But pmu-events mapfile.csv is missing these model numbers. Now add the missing model numbers to mapfile.csv. Signed-off-by: Jin Yao --- tools/perf/pmu-events/arch/x86/mapfile.csv | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/perf/pmu-events/arch/x86/mapfile.csv b/tools/perf/pmu-events/arch/x86/mapfile.csv index 2f2a209e87e1..6455f06f35d3 100644 --- a/tools/perf/pmu-events/arch/x86/mapfile.csv +++ b/tools/perf/pmu-events/arch/x86/mapfile.csv @@ -24,6 +24,7 @@ GenuineIntel-6-1F,v2,nehalemep,core GenuineIntel-6-1A,v2,nehalemep,core GenuineIntel-6-2E,v2,nehalemex,core GenuineIntel-6-[4589]E,v24,skylake,core +GenuineIntel-6-A[56],v24,skylake,core GenuineIntel-6-37,v13,silvermont,core GenuineIntel-6-4D,v13,silvermont,core GenuineIntel-6-4C,v13,silvermont,core @@ -35,6 +36,8 @@ GenuineIntel-6-55-[01234],v1,skylakex,core GenuineIntel-6-55-[56789ABCDEF],v1,cascadelakex,core GenuineIntel-6-7D,v1,icelake,core GenuineIntel-6-7E,v1,icelake,core +GenuineIntel-6-8[CD],v1,icelake,core +GenuineIntel-6-A7,v1,icelake,core GenuineIntel-6-86,v1,tremontx,core AuthenticAMD-23-([12][0-9A-F]|[0-9A-F]),v2,amdzen1,core AuthenticAMD-23-[[:xdigit:]]+,v1,amdzen2,core -- 2.17.1
[PATCH v3 19/27] perf tests: Add hybrid cases for 'Parse event definition strings' test
Add basic hybrid test cases for 'Parse event definition strings' test. # ./perf test 6 6: Parse event definition strings : Ok Signed-off-by: Jin Yao --- v3: - Use PERF_TYPE_RAW for cpu_core/cycles/ tools/perf/tests/parse-events.c | 170 1 file changed, 170 insertions(+) diff --git a/tools/perf/tests/parse-events.c b/tools/perf/tests/parse-events.c index 026c54743311..1cd1d2778172 100644 --- a/tools/perf/tests/parse-events.c +++ b/tools/perf/tests/parse-events.c @@ -1512,6 +1512,123 @@ static int test__all_tracepoints(struct evlist *evlist) return test__checkevent_tracepoint_multi(evlist); } +static int test__hybrid_hw_event_with_pmu(struct evlist *evlist) +{ + struct evsel *evsel = evlist__first(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + return 0; +} + +static int test__hybrid_hw_event(struct evlist *evlist) +{ + struct evsel *evsel1 = evlist__first(evlist); + struct evsel *evsel2 = evlist__last(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel1->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel1->core.attr.config); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel2->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0xa == evsel2->core.attr.config); + return 0; +} + +static int test__hybrid_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0xc0 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_sw_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_hw_sw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_group_modifier1(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x3c == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + TEST_ASSERT_VAL("wrong exclude_user", evsel->core.attr.exclude_user); + TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->core.attr.exclude_kernel); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_RAW == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config&quo
[PATCH v3 24/27] perf tests: Support 'Session topology' test for hybrid
Force to create one event "cpu_core/cycles/" by default, otherwise in evlist__valid_sample_type, the checking of 'if (evlist->core.nr_entries == 1)' would be failed. # ./perf test 41 41: Session topology: Ok Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/topology.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/topology.c b/tools/perf/tests/topology.c index 050489807a47..30b4acb08d35 100644 --- a/tools/perf/tests/topology.c +++ b/tools/perf/tests/topology.c @@ -8,6 +8,7 @@ #include "session.h" #include "evlist.h" #include "debug.h" +#include "pmu.h" #include #define TEMPL "/tmp/perf-test-XX" @@ -40,7 +41,14 @@ static int session_write_header(char *path) session = perf_session__new(, false, NULL); TEST_ASSERT_VAL("can't get session", !IS_ERR(session)); - session->evlist = evlist__new_default(); + if (!perf_pmu__has_hybrid()) { + session->evlist = evlist__new_default(); + } else { + struct parse_events_error err; + + session->evlist = evlist__new(); + parse_events(session->evlist, "cpu_core/cycles/", ); + } TEST_ASSERT_VAL("can't get evlist", session->evlist); perf_header__set_feat(>header, HEADER_CPU_TOPOLOGY); -- 2.17.1
[PATCH v3 18/27] perf record: Uniquify hybrid event name
For perf-record, it would be useful to tell user the pmu which the event belongs to. For example, # perf record -a -- sleep 1 # perf report # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 106 of event 'cpu_core/cycles/' # Event count (approx.): 22043448 # # Overhead Command Shared ObjectSymbol # ... # ... Signed-off-by: Jin Yao --- v3: - New patch in v3. tools/perf/builtin-record.c | 28 1 file changed, 28 insertions(+) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 1b44b1a8636e..74cc9ffbd9ef 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -1605,6 +1605,32 @@ static void hit_auxtrace_snapshot_trigger(struct record *rec) } } +static void record__uniquify_name(struct record *rec) +{ + struct evsel *pos; + struct evlist *evlist = rec->evlist; + char *new_name; + int ret; + + if (!perf_pmu__has_hybrid()) + return; + + evlist__for_each_entry(evlist, pos) { + if (!evsel__is_hybrid(pos)) + continue; + + if (strchr(pos->name, '/')) + continue; + + ret = asprintf(_name, "%s/%s/", + pos->pmu_name, pos->name); + if (ret) { + free(pos->name); + pos->name = new_name; + } + } +} + static int __cmd_record(struct record *rec, int argc, const char **argv) { int err; @@ -1709,6 +1735,8 @@ static int __cmd_record(struct record *rec, int argc, const char **argv) if (data->is_pipe && rec->evlist->core.nr_entries == 1) rec->opts.sample_id = true; + record__uniquify_name(rec); + if (record__open(rec) != 0) { err = -1; goto out_child; -- 2.17.1
[PATCH v3 25/27] perf tests: Support 'Convert perf time to TSC' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the second evsel in evlist also needs initialization. With this patch, # ./perf test 71 71: Convert perf time to TSC: Ok Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/perf-time-to-tsc.c | 16 1 file changed, 16 insertions(+) diff --git a/tools/perf/tests/perf-time-to-tsc.c b/tools/perf/tests/perf-time-to-tsc.c index 680c3cffb128..72f268c6cc5d 100644 --- a/tools/perf/tests/perf-time-to-tsc.c +++ b/tools/perf/tests/perf-time-to-tsc.c @@ -20,6 +20,7 @@ #include "tsc.h" #include "mmap.h" #include "tests.h" +#include "pmu.h" #define CHECK__(x) { \ while ((x) < 0) { \ @@ -66,6 +67,10 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe u64 test_tsc, comm1_tsc, comm2_tsc; u64 test_time, comm1_time = 0, comm2_time = 0; struct mmap *md; + bool hybrid = false; + + if (perf_pmu__has_hybrid()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); CHECK_NOT_NULL__(threads); @@ -88,6 +93,17 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe evsel->core.attr.disabled = 1; evsel->core.attr.enable_on_exec = 0; + /* +* For hybrid "cycles:u", it creates two events. +* Init the second evsel here. +*/ + if (hybrid) { + evsel = evsel__next(evsel); + evsel->core.attr.comm = 1; + evsel->core.attr.disabled = 1; + evsel->core.attr.enable_on_exec = 0; + } + CHECK__(evlist__open(evlist)); CHECK__(evlist__mmap(evlist, UINT_MAX)); -- 2.17.1
[PATCH v3 22/27] perf tests: Support 'Track with sched_switch' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the number of events in evlist is not expected in next test steps. Now we just use one event "cpu_core/cycles:u/" for hybrid. # ./perf test 35 35: Track with sched_switch : Ok Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/switch-tracking.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/switch-tracking.c b/tools/perf/tests/switch-tracking.c index 3ebaa758df77..3a12176f8c46 100644 --- a/tools/perf/tests/switch-tracking.c +++ b/tools/perf/tests/switch-tracking.c @@ -18,6 +18,7 @@ #include "record.h" #include "tests.h" #include "util/mmap.h" +#include "pmu.h" static int spin_sleep(void) { @@ -340,6 +341,10 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ struct evsel *switch_evsel, *tracking_evsel; const char *comm; int err = -1; + bool hybrid = false; + + if (perf_pmu__has_hybrid()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); if (!threads) { @@ -371,7 +376,10 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ cpu_clocks_evsel = evlist__last(evlist); /* Second event */ - err = parse_events(evlist, "cycles:u", NULL); + if (!hybrid) + err = parse_events(evlist, "cycles:u", NULL); + else + err = parse_events(evlist, "cpu_core/cycles/u", NULL); if (err) { pr_debug("Failed to parse event cycles:u\n"); goto out_err; -- 2.17.1
[PATCH v3 27/27] perf Documentation: Document intel-hybrid support
Add some words and examples to help understanding of Intel hybrid perf support. Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/Documentation/intel-hybrid.txt | 228 ++ tools/perf/Documentation/perf-record.txt | 1 + tools/perf/Documentation/perf-stat.txt| 2 + 3 files changed, 231 insertions(+) create mode 100644 tools/perf/Documentation/intel-hybrid.txt diff --git a/tools/perf/Documentation/intel-hybrid.txt b/tools/perf/Documentation/intel-hybrid.txt new file mode 100644 index ..784f598dd36f --- /dev/null +++ b/tools/perf/Documentation/intel-hybrid.txt @@ -0,0 +1,228 @@ +Intel hybrid support + +Support for Intel hybrid events within perf tools. + +For some Intel platforms, such as AlderLake, which is hybrid platform and +it consists of atom cpu and core cpu. Each cpu has dedicated event list. +Part of events are available on core cpu, part of events are available +on atom cpu and even part of events are available on both. + +Kernel exports two new cpu pmus via sysfs: +/sys/devices/cpu_core +/sys/devices/cpu_atom + +The 'cpus' files are created under the directories. For example, + +cat /sys/devices/cpu_core/cpus +0-15 + +cat /sys/devices/cpu_atom/cpus +16-23 + +It indicates cpu0-cpu15 are core cpus and cpu16-cpu23 are atom cpus. + +Quickstart + +List hybrid event +- + +As before, use perf-list to list the symbolic event. + +perf list + +inst_retired.any + [Fixed Counter: Counts the number of instructions retired. Unit: cpu_atom] +inst_retired.any + [Number of instructions retired. Fixed Counter - architectural event. Unit: cpu_core] + +The 'Unit: xxx' is added to brief description to indicate which pmu +the event is belong to. Same event name but with different pmu can +be supported. + +Enable hybrid event with a specific pmu +--- + +To enable a core only event or atom only event, following syntax is supported: + + cpu_core// +or + cpu_atom// + +For example, count the 'cycles' event on core cpus. + + perf stat -e cpu_core/cycles/ + +Create two events for one hardware event automatically +-- + +When creating one event and the event is available on both atom and core, +two events are created automatically. One is for atom, the other is for +core. Most of hardware events and cache events are available on both +cpu_core and cpu_atom. + +For hardware events, they have pre-defined configs (e.g. 0 for cycles). +But on hybrid platform, kernel needs to know where the event comes from +(from atom or from core). The original perf event type PERF_TYPE_HARDWARE +can't carry pmu information. So a new type PERF_TYPE_HARDWARE_PMU is +introduced. + +The new attr.config layout for PERF_TYPE_HARDWARE_PMU: + +0xDD00AA +AA: original hardware event ID +DD: PMU type ID + +Cache event is similar. A new type PERF_TYPE_HW_CACHE_PMU is introduced. + +The new attr.config layout for PERF_TYPE_HW_CACHE_PMU: + +0xDD00CCBBAA +AA: original hardware cache ID +BB: original hardware cache op ID +CC: original hardware cache op result ID +DD: PMU type ID + +PMU type ID is retrieved from sysfs + +cat /sys/devices/cpu_atom/type +10 + +cat /sys/devices/cpu_core/type +4 + +When enabling a hardware event without specified pmu, such as, +perf stat -e cycles -a (use system-wide in this example), two events +are created automatically. + + +perf_event_attr: + type 6 + size 120 + config 0x4 + sample_type IDENTIFIER + read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING + disabled 1 + inherit 1 + exclude_guest1 + + +and + + +perf_event_attr: + type 6 + size 120 + config 0xa + sample_type IDENTIFIER + read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING + disabled 1 + inherit 1 + exclude_guest1 + + +type 6 is PERF_TYPE_HARDWARE_PMU. +0x4 in 0x4 indicates it's cpu_core pmu. +0xa in 0xa indicates it's cpu_atom pmu (atom pmu type id is random). + +The kernel creates 'cycles' (0x4) on cpu0-cpu15 (core cpus), +and create 'cycles' (0xa) on cpu16-cpu23 (atom cpus). + +For perf-stat result, it displays two events: + + Performance counter stats for 'system wide
[PATCH v3 23/27] perf tests: Support 'Parse and process metrics' test for hybrid
Some events are not supported. Only pick up some cases for hybrid. # ./perf test 67 67: Parse and process metrics : Ok Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/parse-metric.c | 10 -- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/tools/perf/tests/parse-metric.c b/tools/perf/tests/parse-metric.c index 4968c4106254..24e5ddff515e 100644 --- a/tools/perf/tests/parse-metric.c +++ b/tools/perf/tests/parse-metric.c @@ -11,6 +11,7 @@ #include "debug.h" #include "expr.h" #include "stat.h" +#include "pmu.h" static struct pmu_event pme_test[] = { { @@ -370,12 +371,17 @@ static int test_metric_group(void) int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused) { + perf_pmu__scan(NULL); + TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); - TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); - TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0); + + if (!perf_pmu__has_hybrid()) { + TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); + TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); + } return 0; } -- 2.17.1
[PATCH v3 26/27] perf tests: Skip 'perf stat metrics (shadow stat) test' for hybrid
Currently we don't support shadow stat for hybrid. root@ssp-pwrt-002:~# ./perf stat -e cycles,instructions -a -- sleep 1 Performance counter stats for 'system wide': 12,883,109,591 cpu_core/cycles/ 6,405,163,221 cpu_atom/cycles/ 555,553,778 cpu_core/instructions/ 841,158,734 cpu_atom/instructions/ 1.002644773 seconds time elapsed Now there is no shadow stat 'insn per cycle' reported. We will support it later and now just skip the 'perf stat metrics (shadow stat) test'. Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/shell/stat+shadow_stat.sh | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/perf/tests/shell/stat+shadow_stat.sh b/tools/perf/tests/shell/stat+shadow_stat.sh index ebebd3596cf9..e6e35fc6c882 100755 --- a/tools/perf/tests/shell/stat+shadow_stat.sh +++ b/tools/perf/tests/shell/stat+shadow_stat.sh @@ -7,6 +7,9 @@ set -e # skip if system-wide mode is forbidden perf stat -a true > /dev/null 2>&1 || exit 2 +# skip if on hybrid platform +perf stat -a -e cycles sleep 1 2>&1 | grep -e cpu_core && exit 2 + test_global_aggr() { perf stat -a --no-big-num -e cycles,instructions sleep 1 2>&1 | \ -- 2.17.1
[PATCH v3 21/27] perf tests: Skip 'Setup struct perf_event_attr' test for hybrid
For hybrid, kernel introduces new perf type PERF_TYPE_HARDWARE_PMU (6) and it's assigned to hybrid hardware events. # ./perf test 17 -vvv ... compare matching [event:base-stat] to [event-6-17179869184-4] [cpu] * 0 [flags] 0|8 8 [type] 0 6 ->FAIL match: [event:base-stat] matches [] event:base-stat does not match, but is optional matched compare matching [event-6-17179869184-4] to [event:base-stat] [cpu] 0 * [flags] 8 0|8 [type] 6 0 ->FAIL match: [event-6-17179869184-4] matches [] expected type=6, got 0 expected config=17179869184, got 0 FAILED './tests/attr/test-stat-C0' - match failure The type matching is failed because one type is 0 but the type of hybrid hardware event is 6. We temporarily skip this test case and TODO in future. Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/attr.c | 4 1 file changed, 4 insertions(+) diff --git a/tools/perf/tests/attr.c b/tools/perf/tests/attr.c index dd39ce9b0277..b37c35fb5a46 100644 --- a/tools/perf/tests/attr.c +++ b/tools/perf/tests/attr.c @@ -34,6 +34,7 @@ #include "event.h" #include "util.h" #include "tests.h" +#include "pmu.h" #define ENV "PERF_TEST_ATTR" @@ -184,6 +185,9 @@ int test__attr(struct test *test __maybe_unused, int subtest __maybe_unused) char path_dir[PATH_MAX]; char *exec_path; + if (perf_pmu__has_hybrid()) + return 0; + /* First try development tree tests. */ if (!lstat("./tests", )) return run_dir("./tests", "./perf"); -- 2.17.1
[PATCH v3 20/27] perf tests: Add hybrid cases for 'Roundtrip evsel->name' test
Since for one hw event, two hybrid events are created. For example, evsel->idx evsel__name(evsel) 0 cycles 1 cycles 2 instructions 3 instructions ... So for comparing the evsel name on hybrid, the evsel->idx needs to be divided by 2. # ./perf test 14 14: Roundtrip evsel->name : Ok Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/tests/evsel-roundtrip-name.c | 19 --- 1 file changed, 12 insertions(+), 7 deletions(-) diff --git a/tools/perf/tests/evsel-roundtrip-name.c b/tools/perf/tests/evsel-roundtrip-name.c index f7f3e5b4c180..b74cf80d1f10 100644 --- a/tools/perf/tests/evsel-roundtrip-name.c +++ b/tools/perf/tests/evsel-roundtrip-name.c @@ -4,6 +4,7 @@ #include "parse-events.h" #include "tests.h" #include "debug.h" +#include "pmu.h" #include #include @@ -62,7 +63,8 @@ static int perf_evsel__roundtrip_cache_name_test(void) return ret; } -static int __perf_evsel__name_array_test(const char *names[], int nr_names) +static int __perf_evsel__name_array_test(const char *names[], int nr_names, +int distance) { int i, err; struct evsel *evsel; @@ -82,9 +84,9 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) err = 0; evlist__for_each_entry(evlist, evsel) { - if (strcmp(evsel__name(evsel), names[evsel->idx])) { + if (strcmp(evsel__name(evsel), names[evsel->idx / distance])) { --err; - pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx]); + pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx / distance]); } } @@ -93,18 +95,21 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) return err; } -#define perf_evsel__name_array_test(names) \ - __perf_evsel__name_array_test(names, ARRAY_SIZE(names)) +#define perf_evsel__name_array_test(names, distance) \ + __perf_evsel__name_array_test(names, ARRAY_SIZE(names), distance) int test__perf_evsel__roundtrip_name_test(struct test *test __maybe_unused, int subtest __maybe_unused) { int err = 0, ret = 0; - err = perf_evsel__name_array_test(evsel__hw_names); + if (perf_pmu__has_hybrid()) + return perf_evsel__name_array_test(evsel__hw_names, 2); + + err = perf_evsel__name_array_test(evsel__hw_names, 1); if (err) ret = err; - err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1); + err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1, 1); if (err) ret = err; -- 2.17.1
[PATCH v3 15/27] perf stat: Filter out unmatched aggregation for hybrid event
perf-stat has supported some aggregation modes, such as --per-core, --per-socket and etc. While for hybrid event, it may only available on part of cpus. So for --per-core, we need to filter out the unavailable cores, for --per-socket, filter out the unavailable sockets, and so on. Before: # perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2479,530 cpu_core/cycles/ S0-D0-C4 2175,007 cpu_core/cycles/ S0-D0-C8 2166,240 cpu_core/cycles/ S0-D0-C12 2704,673 cpu_core/cycles/ S0-D0-C16 2865,835 cpu_core/cycles/ S0-D0-C20 2 2,958,461 cpu_core/cycles/ S0-D0-C24 2163,988 cpu_core/cycles/ S0-D0-C28 2164,729 cpu_core/cycles/ S0-D0-C32 0cpu_core/cycles/ S0-D0-C33 0cpu_core/cycles/ S0-D0-C34 0cpu_core/cycles/ S0-D0-C35 0cpu_core/cycles/ S0-D0-C36 0cpu_core/cycles/ S0-D0-C37 0cpu_core/cycles/ S0-D0-C38 0cpu_core/cycles/ S0-D0-C39 0cpu_core/cycles/ 1.003597211 seconds time elapsed After: # perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2210,428 cpu_core/cycles/ S0-D0-C4 2444,830 cpu_core/cycles/ S0-D0-C8 2435,241 cpu_core/cycles/ S0-D0-C12 2423,976 cpu_core/cycles/ S0-D0-C16 2859,350 cpu_core/cycles/ S0-D0-C20 2 1,559,589 cpu_core/cycles/ S0-D0-C24 2163,924 cpu_core/cycles/ S0-D0-C28 2376,610 cpu_core/cycles/ 1.003621290 seconds time elapsed Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/util/stat-display.c | 20 1 file changed, 20 insertions(+) diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 161826938a00..b7ce3c4ae5a8 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -635,6 +635,20 @@ static void aggr_cb(struct perf_stat_config *config, } } +static bool aggr_id_hybrid_matched(struct perf_stat_config *config, + struct evsel *counter, struct aggr_cpu_id id) +{ + struct aggr_cpu_id s; + + for (int i = 0; i < evsel__nr_cpus(counter); i++) { + s = config->aggr_get_id(config, evsel__cpus(counter), i); + if (cpu_map__compare_aggr_cpu_id(s, id)) + return true; + } + + return false; +} + static void print_counter_aggrdata(struct perf_stat_config *config, struct evsel *counter, int s, char *prefix, bool metric_only, @@ -648,6 +662,12 @@ static void print_counter_aggrdata(struct perf_stat_config *config, double uval; ad.id = id = config->aggr_map->map[s]; + + if (perf_pmu__has_hybrid() && + !aggr_id_hybrid_matched(config, counter, id)) { + return; + } + ad.val = ad.ena = ad.run = 0; ad.nr = 0; if (!collect_data(config, counter, aggr_cb, )) -- 2.17.1
[PATCH v3 17/27] perf script: Support PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU
For a hybrid system, the perf subsystem doesn't know which PMU the events belong to. So the PMU aware version PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU are introduced. Now define the new output[] entries for these two types. Signed-off-by: Jin Yao --- v3: - No change. tools/perf/builtin-script.c | 24 1 file changed, 24 insertions(+) diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c index 1280cbfad4db..627ec640d2e6 100644 --- a/tools/perf/builtin-script.c +++ b/tools/perf/builtin-script.c @@ -275,6 +275,30 @@ static struct { .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, }, + [PERF_TYPE_HARDWARE_PMU] = { + .user_set = false, + + .fields = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | + PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | + PERF_OUTPUT_EVNAME | PERF_OUTPUT_IP | + PERF_OUTPUT_SYM | PERF_OUTPUT_SYMOFFSET | + PERF_OUTPUT_DSO | PERF_OUTPUT_PERIOD, + + .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, + }, + + [PERF_TYPE_HW_CACHE_PMU] = { + .user_set = false, + + .fields = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | + PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | + PERF_OUTPUT_EVNAME | PERF_OUTPUT_IP | + PERF_OUTPUT_SYM | PERF_OUTPUT_SYMOFFSET | + PERF_OUTPUT_DSO | PERF_OUTPUT_PERIOD, + + .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, + }, + [OUTPUT_TYPE_SYNTH] = { .user_set = false, -- 2.17.1
[PATCH v3 14/27] perf stat: Add default hybrid events
Previously if '-e' is not specified in perf stat, some software events and hardware events are added to evlist by default. Before: # ./perf stat -a -- sleep 1 Performance counter stats for 'system wide': 24,044.40 msec cpu-clock # 23.946 CPUs utilized 99 context-switches #4.117 /sec 24 cpu-migrations#0.998 /sec 3 page-faults #0.125 /sec 7,000,244 cycles#0.000 GHz 2,955,024 instructions #0.42 insn per cycle 608,941 branches # 25.326 K/sec 31,991 branch-misses #5.25% of all branches 1.004106859 seconds time elapsed Among the events, cycles, instructions, branches and branch-misses are hardware events. One hybrid platform, two hardware events are created for one hardware event. cpu_core/cycles/, cpu_atom/cycles/, cpu_core/instructions/, cpu_atom/instructions/, cpu_core/branches/, cpu_atom/branches/, cpu_core/branch-misses/, cpu_atom/branch-misses/ These events would be added to evlist on hybrid platform. Since parse_events() has been supported to create two hardware events for one event on hybrid platform, so we just use parse_events(evlist, "cycles,instructions,branches,branch-misses") to create the default events and add them to evlist. After: # ./perf stat -a -- sleep 1 Performance counter stats for 'system wide': 24,048.60 msec task-clock# 23.947 CPUs utilized 438 context-switches # 18.213 /sec 24 cpu-migrations#0.998 /sec 6 page-faults #0.249 /sec 24,813,157 cpu_core/cycles/ #1.032 M/sec 8,072,687 cpu_atom/cycles/ # 335.682 K/sec 20,731,286 cpu_core/instructions/# 862.058 K/sec 3,737,203 cpu_atom/instructions/# 155.402 K/sec 2,620,924 cpu_core/branches/# 108.984 K/sec 381,186 cpu_atom/branches/# 15.851 K/sec 93,248 cpu_core/branch-misses/ #3.877 K/sec 36,515 cpu_atom/branch-misses/ #1.518 K/sec 1.004235472 seconds time elapsed We can see two events are created for one hardware event. One TODO is, the shadow stats looks a bit different, now it's just 'M/sec'. The perf_stat__update_shadow_stats and perf_stat__print_shadow_stats need to be improved in future if we want to get the original shadow stats. Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/builtin-stat.c | 21 + 1 file changed, 21 insertions(+) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 7b2dfe21c5a8..33fda8f55f66 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1140,6 +1140,13 @@ static int parse_stat_cgroups(const struct option *opt, return parse_cgroups(opt, str, unset); } +static int add_default_hybrid_events(struct evlist *evlist) +{ + struct parse_events_error err; + + return parse_events(evlist, "cycles,instructions,branches,branch-misses", ); +} + static struct option stat_options[] = { OPT_BOOLEAN('T', "transaction", _run, "hardware transaction statistics"), @@ -1619,6 +1626,12 @@ static int add_default_attributes(void) { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, +}; + struct perf_event_attr default_sw_attrs[] = { + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, }; /* @@ -1856,6 +1869,14 @@ static int add_default_attributes(void) } if (!evsel_list->core.nr_entries) { + if (perf_pmu__has_hybrid()) { + if (evlist__add_default_attrs(evsel_list, + default_sw_attrs) < 0) { + return -1; + } + return add_default_hybrid_events(evsel_list); + } + if (target__has_cpu()) default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; -- 2.17.1
[PATCH v3 16/27] perf stat: Warn group events from different hybrid PMU
If a group has events which are from different hybrid PMUs, shows a warning: "WARNING: events in group from different hybrid PMUs!" This is to remind the user not to put the core event and atom event into one group. Next, just disable grouping. # perf stat -e "{cpu_core/cycles/,cpu_atom/cycles/}" -a -- sleep 1 WARNING: events in group from different hybrid PMUs! WARNING: grouped events cpus do not match, disabling group: anon group { cpu_core/cycles/, cpu_atom/cycles/ } Performance counter stats for 'system wide': 5,438,125 cpu_core/cycles/ 3,914,586 cpu_atom/cycles/ 1.004250966 seconds time elapsed Signed-off-by: Jin Yao --- v3: - Change the processing logic. In v2, it just reported the warning and returned error. But in v3, we also disable grouping. tools/perf/builtin-stat.c | 4 +++ tools/perf/util/evlist-hybrid.c| 47 ++ tools/perf/util/evlist-hybrid.h| 2 ++ tools/perf/util/evsel.c| 6 tools/perf/util/evsel.h| 1 + tools/perf/util/python-ext-sources | 2 ++ 6 files changed, 62 insertions(+) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 33fda8f55f66..6f70a9f1971e 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -48,6 +48,7 @@ #include "util/pmu.h" #include "util/event.h" #include "util/evlist.h" +#include "util/evlist-hybrid.h" #include "util/evsel.h" #include "util/debug.h" #include "util/color.h" @@ -240,6 +241,9 @@ static void evlist__check_cpu_maps(struct evlist *evlist) struct evsel *evsel, *pos, *leader; char buf[1024]; + if (evlist__has_hybrid(evlist)) + evlist__warn_hybrid_group(evlist); + evlist__for_each_entry(evlist, evsel) { leader = evsel->leader; diff --git a/tools/perf/util/evlist-hybrid.c b/tools/perf/util/evlist-hybrid.c index 185f60ec4351..39f520372447 100644 --- a/tools/perf/util/evlist-hybrid.c +++ b/tools/perf/util/evlist-hybrid.c @@ -7,6 +7,7 @@ #include "../perf.h" #include "util/pmu-hybrid.h" #include "util/evlist-hybrid.h" +#include "debug.h" #include #include #include @@ -39,3 +40,49 @@ int evlist__add_default_hybrid(struct evlist *evlist, bool precise) return 0; } + +static bool group_hybrid_conflict(struct evsel *leader) +{ + struct evsel *pos, *prev = NULL; + + for_each_group_evsel(pos, leader) { + if (!evsel__is_hybrid(pos)) + continue; + + if (prev && strcmp(prev->pmu_name, pos->pmu_name)) + return true; + + prev = pos; + } + + return false; +} + +void evlist__warn_hybrid_group(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel__is_group_leader(evsel) && + evsel->core.nr_members > 1 && + group_hybrid_conflict(evsel)) { + pr_warning("WARNING: events in group from " + "different hybrid PMUs!\n"); + return; + } + } +} + +bool evlist__has_hybrid(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel->pmu_name && + perf_pmu__is_hybrid(evsel->pmu_name)) { + return true; + } + } + + return false; +} diff --git a/tools/perf/util/evlist-hybrid.h b/tools/perf/util/evlist-hybrid.h index e25861649d8f..19f74b4c340a 100644 --- a/tools/perf/util/evlist-hybrid.h +++ b/tools/perf/util/evlist-hybrid.h @@ -8,5 +8,7 @@ #include int evlist__add_default_hybrid(struct evlist *evlist, bool precise); +void evlist__warn_hybrid_group(struct evlist *evlist); +bool evlist__has_hybrid(struct evlist *evlist); #endif /* __PERF_EVLIST_HYBRID_H */ diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c index 0ba4daa09453..0f64a32ea9c5 100644 --- a/tools/perf/util/evsel.c +++ b/tools/perf/util/evsel.c @@ -47,6 +47,7 @@ #include "memswap.h" #include "util.h" #include "hashmap.h" +#include "pmu-hybrid.h" #include "../perf-sys.h" #include "util/parse-branch-options.h" #include @@ -2797,3 +2798,8 @@ void evsel__zero_per_pkg(struct evsel *evsel) hashmap__clear(evsel->per_pkg_mask); } } + +bool evsel__is_hybrid(struct evsel *evsel) +{ + return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name); +} diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h index e7dc3448ab2f..e56826bbb628 100644 --- a/tools/perf/util/evsel.h +++ b/t
[PATCH v3 12/27] perf parse-events: Support no alias assigned event inside hybrid PMU
On hybrid platform, similar to hardware event, user may want to enable other events only on one PMU. So following syntax should be supported: cpu_core// cpu_atom// But the syntax doesn't work for some events, such as cache event. Before: # perf stat -e cpu_core/LLC-loads/ -a -- sleep 1 event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' Cache events are much more complex than hardware events, so we can't create aliases for them. We use another solution. For example, if we use "cpu_core/LLC-loads/", in parse_events_add_pmu(), term->config is "LLC-loads". We create a new "parse_events_state" with the pmu_name and use parse_events__scanner to scan the term->config ("LLC-loads" in this example). The parse_events_add_cache() would be called during parsing. The parse_state->pmu_name is used to identify the pmu where the event should be enabled on. After: # ./perf stat -e cpu_core/LLC-loads/ -a -- sleep 1 Performance counter stats for 'system wide': 24,593 cpu_core/LLC-loads/ 1.003911601 seconds time elapsed Signed-off-by: Jin Yao --- v3: - Rename the patch: 'perf parse-events: Support hardware events inside PMU' --> 'perf parse-events: Support no alias assigned event inside hybrid PMU' - Major code is moved to parse-events-hybrid.c. - Refine the code. tools/perf/util/parse-events-hybrid.c | 18 +- tools/perf/util/parse-events-hybrid.h | 3 +- tools/perf/util/parse-events.c| 80 +-- tools/perf/util/parse-events.h| 4 +- tools/perf/util/parse-events.y| 9 ++- tools/perf/util/pmu.c | 4 +- tools/perf/util/pmu.h | 2 +- 7 files changed, 108 insertions(+), 12 deletions(-) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index 8a630cbab8f3..5bf176b55573 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -64,6 +64,11 @@ static int add_hw_hybrid(struct parse_events_state *parse_state, int ret; perf_pmu__for_each_hybrid_pmu(pmu) { + if (parse_state->pmu_name && + strcmp(parse_state->pmu_name, pmu->name)) { + continue; + } + ret = create_event_hybrid(PERF_TYPE_HARDWARE_PMU, _state->idx, list, attr, name, config_terms, pmu); @@ -100,6 +105,11 @@ static int add_raw_hybrid(struct parse_events_state *parse_state, int ret; perf_pmu__for_each_hybrid_pmu(pmu) { + if (parse_state->pmu_name && + strcmp(parse_state->pmu_name, pmu->name)) { + continue; + } + ret = create_raw_event_hybrid(_state->idx, list, attr, name, config_terms, pmu); if (ret) @@ -137,7 +147,8 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, int parse_events__add_cache_hybrid(struct list_head *list, int *idx, struct perf_event_attr *attr, char *name, struct list_head *config_terms, - bool *hybrid) + bool *hybrid, + struct parse_events_state *parse_state) { struct perf_pmu *pmu; int ret; @@ -148,6 +159,11 @@ int parse_events__add_cache_hybrid(struct list_head *list, int *idx, *hybrid = true; perf_pmu__for_each_hybrid_pmu(pmu) { + if (parse_state->pmu_name && + strcmp(parse_state->pmu_name, pmu->name)) { + continue; + } + ret = create_event_hybrid(PERF_TYPE_HW_CACHE_PMU, idx, list, attr, name, config_terms, pmu); if (ret) diff --git a/tools/perf/util/parse-events-hybrid.h b/tools/perf/util/parse-events-hybrid.h index 9ad33cd0cef4..f33bd67aa851 100644 --- a/tools/perf/util/parse-events-hybrid.h +++ b/tools/perf/util/parse-events-hybrid.h @@ -17,6 +17,7 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, int parse_events__add_cache_hybrid(struct list_head *list, int *idx, struct perf_event_attr *attr, char *name, struct list_head *config_terms, - bool *hybrid); + bool *hybrid, + struct parse_events_state *parse_state); #endif /* __PERF_PARSE_EVENTS_HYBRID_H */ diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-even
[PATCH v3 10/27] perf parse-events: Create two hybrid raw events
/ 1.003722876 seconds time elapsed type 4 is cpu_core pmu type. type 10 is cpu_atom pmu type. Signed-off-by: Jin Yao --- v3: - Raw event creation is moved to parse-events-hybrid.c. tools/perf/util/parse-events-hybrid.c | 38 +++ 1 file changed, 38 insertions(+) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index ff2909bfbf86..8a630cbab8f3 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -74,6 +74,41 @@ static int add_hw_hybrid(struct parse_events_state *parse_state, return 0; } +static int create_raw_event_hybrid(int *idx, struct list_head *list, + struct perf_event_attr *attr, char *name, + struct list_head *config_terms, + struct perf_pmu *pmu) +{ + struct evsel *evsel; + + attr->type = pmu->type; + evsel = parse_events__add_event_hybrid(list, idx, attr, name, + pmu, config_terms); + if (evsel) + evsel->pmu_name = strdup(pmu->name); + else + return -ENOMEM; + + return 0; +} + +static int add_raw_hybrid(struct parse_events_state *parse_state, + struct list_head *list, struct perf_event_attr *attr, + char *name, struct list_head *config_terms) +{ + struct perf_pmu *pmu; + int ret; + + perf_pmu__for_each_hybrid_pmu(pmu) { + ret = create_raw_event_hybrid(_state->idx, list, attr, + name, config_terms, pmu); + if (ret) + return ret; + } + + return 0; +} + int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, struct list_head *list, struct perf_event_attr *attr, @@ -91,6 +126,9 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, if (attr->type != PERF_TYPE_RAW) { return add_hw_hybrid(parse_state, list, attr, name, config_terms); + } else { + return add_raw_hybrid(parse_state, list, attr, name, + config_terms); } return -1; -- 2.17.1
[PATCH v3 13/27] perf record: Create two hybrid 'cycles' events by default
When evlist is empty, for example no '-e' specified in perf record, one default 'cycles' event is added to evlist. While on hybrid platform, it needs to create two default 'cycles' events. One is for cpu_core, the other is for cpu_atom. This patch actually calls evsel__new_cycles() two times to create two 'cycles' events. # ./perf record -vv -a -- sleep 1 ... perf_event_attr: type 6 size 120 config 0x4 { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|CPU|PERIOD read_format ID disabled 1 inherit 1 freq 1 precise_ip 3 sample_id_all1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 5 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8 = 6 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8 = 7 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8 = 9 sys_perf_event_open: pid -1 cpu 4 group_fd -1 flags 0x8 = 10 sys_perf_event_open: pid -1 cpu 5 group_fd -1 flags 0x8 = 11 sys_perf_event_open: pid -1 cpu 6 group_fd -1 flags 0x8 = 12 sys_perf_event_open: pid -1 cpu 7 group_fd -1 flags 0x8 = 13 sys_perf_event_open: pid -1 cpu 8 group_fd -1 flags 0x8 = 14 sys_perf_event_open: pid -1 cpu 9 group_fd -1 flags 0x8 = 15 sys_perf_event_open: pid -1 cpu 10 group_fd -1 flags 0x8 = 16 sys_perf_event_open: pid -1 cpu 11 group_fd -1 flags 0x8 = 17 sys_perf_event_open: pid -1 cpu 12 group_fd -1 flags 0x8 = 18 sys_perf_event_open: pid -1 cpu 13 group_fd -1 flags 0x8 = 19 sys_perf_event_open: pid -1 cpu 14 group_fd -1 flags 0x8 = 20 sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 21 perf_event_attr: type 6 size 120 config 0xa { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|CPU|PERIOD read_format ID disabled 1 inherit 1 freq 1 precise_ip 3 sample_id_all1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 22 sys_perf_event_open: pid -1 cpu 17 group_fd -1 flags 0x8 = 23 sys_perf_event_open: pid -1 cpu 18 group_fd -1 flags 0x8 = 24 sys_perf_event_open: pid -1 cpu 19 group_fd -1 flags 0x8 = 25 sys_perf_event_open: pid -1 cpu 20 group_fd -1 flags 0x8 = 26 sys_perf_event_open: pid -1 cpu 21 group_fd -1 flags 0x8 = 27 sys_perf_event_open: pid -1 cpu 22 group_fd -1 flags 0x8 = 28 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 29 We have to create evlist-hybrid.c otherwise due to the symbol dependency the perf test python would be failed. Signed-off-by: Jin Yao --- v3: - Move the major code to new created evlist-hybrid.c. tools/perf/builtin-record.c | 19 +++ tools/perf/util/Build | 1 + tools/perf/util/evlist-hybrid.c | 41 + tools/perf/util/evlist-hybrid.h | 12 ++ tools/perf/util/evlist.c| 5 +++- tools/perf/util/evsel.c | 6 ++--- tools/perf/util/evsel.h | 2 +- 7 files changed, 77 insertions(+), 9 deletions(-) create mode 100644 tools/perf/util/evlist-hybrid.c create mode 100644 tools/perf/util/evlist-hybrid.h diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 35465d1db6dd..1b44b1a8636e 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -47,6 +47,8 @@ #include "util/util.h" #include "util/pfm.h" #include "util/clockid.h" +#include "util/pmu-hybrid.h" +#include "util/evlist-hybrid.h" #include "asm/bug.h" #include "perf.h" @@ -2786,10 +2788,19 @@ int cmd_record(int argc, const char **argv) if (record.opts.overwrite) record.opts.tail_synthesize = true; - if (rec->evlist->core.nr_entries == 0 && - __evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) { - pr_err("Not enough memory for event selector list\n"); - goto out; + if (rec->evlist->core.nr_entries == 0) { +
[PATCH v3 11/27] perf pmu: Support 'cycles' and 'branches' inside hybrid PMU
On hybrid platform, user may want to enable the hardware event only on one PMU. So following syntax is supported: cpu_core// cpu_atom// # perf stat -e cpu_core/cpu-cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 6,049,336 cpu_core/cpu-cycles/ 1.003577042 seconds time elapsed It enables the event 'cpu-cycles' only on cpu_core pmu. But for 'cycles' and 'branches', the syntax doesn't work. Before: # perf stat -e cpu_core/cycles/ -a -- sleep 1 event syntax error: 'cpu_core/cycles/' \___ unknown term 'cycles' for pmu 'cpu_core' # perf stat -e cpu_core/branches/ -a -- sleep 1 event syntax error: 'cpu_core/branches/' \___ unknown term 'branches' for pmu 'cpu_core' For 'cpu-cycles', why it works is because the event is defined in /sys/devices/cpu_core/events/. It's added as alias by pmu_add_sys_aliases and it's treated as 'event' when parsing the term->config. We use a similar idea, create a pme_hybrid_fixup table for 'cycles' and 'branches' and add them as aliases. After: # perf stat -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 5,769,631 cpu_core/cycles/ 1.003833235 seconds time elapsed # perf stat -e cpu_core/branches/ -a -- sleep 1 Performance counter stats for 'system wide': 490,951 cpu_core/branches/ 1.003693946 seconds time elapsed Signed-off-by: Jin Yao --- v3: - New patch in v3. tools/perf/util/pmu.c | 32 1 file changed, 32 insertions(+) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index beff29981101..72e5ae5e868e 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -916,6 +916,35 @@ static int pmu_max_precise(const char *name) return max_precise; } +static void perf_pmu__add_hybrid_aliases(struct list_head *head) +{ + static struct pmu_event pme_hybrid_fixup[] = { + { + .name = "cycles", + .event = "event=0x3c", + }, + { + .name = "branches", + .event = "event=0xc4", + }, + { + .name = 0, + .event = 0, + }, + }; + int i = 0; + + while (1) { + struct pmu_event *pe = _hybrid_fixup[i++]; + + if (!pe->name) + break; + + __perf_pmu__new_alias(head, NULL, (char *)pe->name, NULL, + (char *)pe->event, NULL); + } +} + static struct perf_pmu *pmu_lookup(const char *name) { struct perf_pmu *pmu; @@ -955,6 +984,9 @@ static struct perf_pmu *pmu_lookup(const char *name) pmu_add_cpu_aliases(, pmu); pmu_add_sys_aliases(, pmu); + if (pmu->is_hybrid) + perf_pmu__add_hybrid_aliases(); + INIT_LIST_HEAD(>format); INIT_LIST_HEAD(>aliases); INIT_LIST_HEAD(>caps); -- 2.17.1
[PATCH v3 09/27] perf parse-events: Create two hybrid cache events
1001696907 1001696907 L1-dcache-loads: 12: 7723 1001716599 1001716599 L1-dcache-loads: 13: 10211 1001743285 1001743285 L1-dcache-loads: 14: 13023 1001765343 1001765343 L1-dcache-loads: 15: 8991 1001788673 1001788673 L1-dcache-loads: 0: 240163 1001800830 1001800830 L1-dcache-loads: 1: 7454 1001773983 1001773983 L1-dcache-loads: 2: 32323 1001686339 1001686339 L1-dcache-loads: 3: 11039 1001732430 1001732430 L1-dcache-loads: 4: 52867 1001753753 1001753753 L1-dcache-loads: 5: 7481 1001756879 1001756879 L1-dcache-loads: 6: 7471 1001814616 1001814616 L1-dcache-loads: 7: 29627 1001815092 1001815092 L1-dcache-loads: 453924 16027424826 16027424826 L1-dcache-loads: 388425 8014133922 8014133922 Performance counter stats for 'system wide': 453,924 cpu_core/L1-dcache-loads/ 388,425 cpu_atom/L1-dcache-loads/ 1.003644499 seconds time elapsed type 7 is PERF_TYPE_HW_CACHE_PMU. 0x4 in 0x4 indicates the cpu_core pmu. 0xa in 0xa indicates the cpu_atom pmu. Signed-off-by: Jin Yao --- v3: - Hybrid cache event creation is moved to parse-events-hybrid.c. tools/perf/util/parse-events-hybrid.c | 23 +++ tools/perf/util/parse-events-hybrid.h | 5 + tools/perf/util/parse-events.c| 8 3 files changed, 36 insertions(+) diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c index bd48563596e0..ff2909bfbf86 100644 --- a/tools/perf/util/parse-events-hybrid.c +++ b/tools/perf/util/parse-events-hybrid.c @@ -95,3 +95,26 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, return -1; } + +int parse_events__add_cache_hybrid(struct list_head *list, int *idx, + struct perf_event_attr *attr, char *name, + struct list_head *config_terms, + bool *hybrid) +{ + struct perf_pmu *pmu; + int ret; + + *hybrid = false; + if (!perf_pmu__has_hybrid()) + return 0; + + *hybrid = true; + perf_pmu__for_each_hybrid_pmu(pmu) { + ret = create_event_hybrid(PERF_TYPE_HW_CACHE_PMU, idx, list, + attr, name, config_terms, pmu); + if (ret) + return ret; + } + + return 0; +} diff --git a/tools/perf/util/parse-events-hybrid.h b/tools/perf/util/parse-events-hybrid.h index d81a76978480..9ad33cd0cef4 100644 --- a/tools/perf/util/parse-events-hybrid.h +++ b/tools/perf/util/parse-events-hybrid.h @@ -14,4 +14,9 @@ int parse_events__add_numeric_hybrid(struct parse_events_state *parse_state, char *name, struct list_head *config_terms, bool *hybrid); +int parse_events__add_cache_hybrid(struct list_head *list, int *idx, + struct perf_event_attr *attr, char *name, + struct list_head *config_terms, + bool *hybrid); + #endif /* __PERF_PARSE_EVENTS_HYBRID_H */ diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 1bbd0ba92ba7..3692fa3c964a 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -458,6 +458,7 @@ int parse_events_add_cache(struct list_head *list, int *idx, int cache_type = -1, cache_op = -1, cache_result = -1; char *op_result[2] = { op_result1, op_result2 }; int i, n; + bool hybrid; /* * No fallback - if we cannot get a clear cache type @@ -517,6 +518,13 @@ int parse_events_add_cache(struct list_head *list, int *idx, if (get_config_terms(head_config, _terms)) return -ENOMEM; } + + i = parse_events__add_cache_hybrid(list, idx, , + config_name ? : name, _terms, + ); + if (hybrid) + return i; + return add_event(list, idx, , config_name ? : name, _terms); } -- 2.17.1
[PATCH v3 08/27] perf parse-events: Create two hybrid hardware events
= 27 cycles: 0: 402953 1001514714 1001514714 cycles: 1: 137456 1001554851 1001554851 cycles: 2: 130449 1001575948 1001575948 cycles: 3: 61370 1001600256 1001600256 cycles: 4: 100084 1001614135 1001614135 cycles: 5: 62549 1001646802 1001646802 cycles: 6: 453760 1001703487 1001703487 cycles: 7: 66527 1001739138 1001739138 cycles: 8: 80526 1001823867 1001823867 cycles: 9: 74942 1001863884 1001863884 cycles: 10: 322356 1001952832 1001952832 cycles: 11: 1681751 1001846058 1001846058 cycles: 12: 97608 1001874578 1001874578 cycles: 13: 62060 1001899200 1001899200 cycles: 14: 546496 1001920608 1001920608 cycles: 15: 65631 1001939206 1001939206 cycles: 0: 59901 1002047096 1002047096 cycles: 1: 57304 1001995373 1001995373 cycles: 2: 781291 1002027732 1002027732 cycles: 3: 99656 1002058466 1002058466 cycles: 4: 95071 1002092749 1002092749 cycles: 5: 346827 1002142979 1002142979 cycles: 6: 183967 1002183046 1002183046 cycles: 7: 391779 1002218286 1002218286 cycles: 4346518 16028069564 16028069564 cycles: 2015796 8016765727 8016765727 Performance counter stats for 'system wide': 4,346,518 cpu_core/cycles/ 2,015,796 cpu_atom/cycles/ 1.003685897 seconds time elapsed type 6 is PERF_TYPE_HARDWARE_PMU. 0x4 in 0x4 indicates the cpu_core pmu. 0xa in 0xa indicates the cpu_atom pmu. Signed-off-by: Jin Yao --- v3: - Create new parse-events-hybrid.c/parse-events-hybrid.h - Refine the code tools/perf/util/Build | 1 + tools/perf/util/parse-events-hybrid.c | 97 +++ tools/perf/util/parse-events-hybrid.h | 17 + tools/perf/util/parse-events.c| 18 + tools/perf/util/parse-events.h| 5 ++ 5 files changed, 138 insertions(+) create mode 100644 tools/perf/util/parse-events-hybrid.c create mode 100644 tools/perf/util/parse-events-hybrid.h diff --git a/tools/perf/util/Build b/tools/perf/util/Build index 37a8a63c7195..00c9fb064ba6 100644 --- a/tools/perf/util/Build +++ b/tools/perf/util/Build @@ -23,6 +23,7 @@ perf-y += llvm-utils.o perf-y += mmap.o perf-y += memswap.o perf-y += parse-events.o +perf-y += parse-events-hybrid.o perf-y += perf_regs.o perf-y += path.o perf-y += print_binary.o diff --git a/tools/perf/util/parse-events-hybrid.c b/tools/perf/util/parse-events-hybrid.c new file mode 100644 index ..bd48563596e0 --- /dev/null +++ b/tools/perf/util/parse-events-hybrid.c @@ -0,0 +1,97 @@ +// SPDX-License-Identifier: GPL-2.0 +#include +#include +#include +#include +#include +#include +#include +#include "evlist.h" +#include "evsel.h" +#include "parse-events.h" +#include "parse-events-hybrid.h" +#include "debug.h" +#include "pmu.h" +#include "pmu-hybrid.h" +#include "perf.h" + +static void config_hybrid_attr(struct perf_event_attr *attr, + int type, int pmu_type) +{ + /* +* attr.config layout: +* PERF_TYPE_HARDWARE_PMU: 0xDD00AA +* AA: hardware event ID +* DD: PMU type ID +* PERF_TYPE_HW_CACHE_PMU: 0xDD00CCBBAA +* AA: hardware cache ID +* BB: hardware cache op ID +* CC: hardware cache op result ID +* DD: PMU type ID +*/ + attr->type = type; + attr->config = attr->config | ((__u64)pmu_type << PERF_PMU_TYPE_SHIFT); +} + +static int create_event_hybrid(__u32 config_type, int *idx, + struct list_head *list, + struct perf_event_attr *attr, char *name, + struct list_head *config_terms, + struct perf_pmu *pmu) +{ + struct evsel *evsel; + __u32 type = attr->type; + __u64 config = attr->config; + + config_hybrid_attr(attr, config_type, pmu->type); + evsel = parse_events__add_event_hybrid(list, idx, attr, name, + pmu, config_terms); + if (evsel) + evsel->pmu_name = strdup(pmu->name); + else + return -ENOMEM; + + attr->type = type; + attr->config = config; + return 0; +} + +static int add_hw_hybrid(struct parse_events_state *parse_state, +struct list_head *list, struct perf_event_attr *attr, +char *name, struct list_head *config_terms) +{ + struct perf_pmu *pmu; + int ret; + + perf_pmu__for_each_hybrid_pmu(pmu) { + ret = create_event_hybrid(PERF_TYPE_HARDWARE_PMU, + _state->idx, list, attr, name, + c
[PATCH v3 04/27] perf pmu: Save pmu name
On hybrid platform, one event is available on one pmu (such as, available on cpu_core or on cpu_atom). This patch saves the pmu name to the pmu field of struct perf_pmu_alias. Then next we can know the pmu which the event can be enabled on. Signed-off-by: Jin Yao --- v3: - Change pmu to pmu_name in struct perf_pmu_alias. tools/perf/util/pmu.c | 10 +- tools/perf/util/pmu.h | 1 + 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 9ed9a6a8b2d2..10709ec1cc3e 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -283,6 +283,7 @@ void perf_pmu_free_alias(struct perf_pmu_alias *newalias) zfree(>str); zfree(>metric_expr); zfree(>metric_name); + zfree(>pmu_name); parse_events_terms__purge(>terms); free(newalias); } @@ -297,6 +298,10 @@ static bool perf_pmu_merge_alias(struct perf_pmu_alias *newalias, list_for_each_entry(a, alist, list) { if (!strcasecmp(newalias->name, a->name)) { + if (newalias->pmu_name && a->pmu_name && + !strcasecmp(newalias->pmu_name, a->pmu_name)) { + continue; + } perf_pmu_update_alias(a, newalias); perf_pmu_free_alias(newalias); return true; @@ -314,7 +319,8 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, int num; char newval[256]; char *long_desc = NULL, *topic = NULL, *unit = NULL, *perpkg = NULL, -*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL; +*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL, +*pmu_name = NULL; if (pe) { long_desc = (char *)pe->long_desc; @@ -324,6 +330,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, metric_expr = (char *)pe->metric_expr; metric_name = (char *)pe->metric_name; deprecated = (char *)pe->deprecated; + pmu_name = (char *)pe->pmu; } alias = malloc(sizeof(*alias)); @@ -389,6 +396,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, } alias->per_pkg = perpkg && sscanf(perpkg, "%d", ) == 1 && num == 1; alias->str = strdup(newval); + alias->pmu_name = pmu_name ? strdup(pmu_name) : NULL; if (deprecated) alias->deprecated = true; diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 8164388478c6..8493b1719e10 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -72,6 +72,7 @@ struct perf_pmu_alias { bool deprecated; char *metric_expr; char *metric_name; + char *pmu_name; }; struct perf_pmu *perf_pmu__find(const char *name); -- 2.17.1
[PATCH v3 05/27] perf pmu: Save detected hybrid pmus to a global pmu list
We identify the cpu_core pmu and cpu_atom pmu by explicitly checking following files: For cpu_core, checks: "/sys/bus/event_source/devices/cpu_core/cpus" For cpu_atom, checks: "/sys/bus/event_source/devices/cpu_atom/cpus" If the 'cpus' file exists, the pmu exists. But in order not to hardcode the "cpu_core" and "cpu_atom", and make the code in a generic way. So if the path "/sys/bus/event_source/devices/cpu_xxx/cpus" exists, the hybrid pmu exists. All the detected hybrid pmus are linked to a global list 'perf_pmu__hybrid_pmus' and then next we just need to iterate the list to get all hybrid pmu by using perf_pmu__for_each_hybrid_pmu. Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/util/Build| 1 + tools/perf/util/pmu-hybrid.c | 35 +++ tools/perf/util/pmu-hybrid.h | 18 ++ tools/perf/util/pmu.c| 9 - tools/perf/util/pmu.h| 4 5 files changed, 66 insertions(+), 1 deletion(-) create mode 100644 tools/perf/util/pmu-hybrid.c create mode 100644 tools/perf/util/pmu-hybrid.h diff --git a/tools/perf/util/Build b/tools/perf/util/Build index e3e12f9d4733..37a8a63c7195 100644 --- a/tools/perf/util/Build +++ b/tools/perf/util/Build @@ -69,6 +69,7 @@ perf-y += parse-events-bison.o perf-y += pmu.o perf-y += pmu-flex.o perf-y += pmu-bison.o +perf-y += pmu-hybrid.o perf-y += trace-event-read.o perf-y += trace-event-info.o perf-y += trace-event-scripting.o diff --git a/tools/perf/util/pmu-hybrid.c b/tools/perf/util/pmu-hybrid.c new file mode 100644 index ..7316bf46e54b --- /dev/null +++ b/tools/perf/util/pmu-hybrid.c @@ -0,0 +1,35 @@ +// SPDX-License-Identifier: GPL-2.0 +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "fncache.h" +#include "pmu-hybrid.h" + +LIST_HEAD(perf_pmu__hybrid_pmus); + +bool perf_pmu__hybrid_mounted(const char *name) +{ + char path[PATH_MAX]; + const char *sysfs; + + if (strncmp(name, "cpu_", 4)) + return false; + + sysfs = sysfs__mountpoint(); + if (!sysfs) + return false; + + snprintf(path, PATH_MAX, CPUS_TEMPLATE_CPU, sysfs, name); + return file_available(path); +} diff --git a/tools/perf/util/pmu-hybrid.h b/tools/perf/util/pmu-hybrid.h new file mode 100644 index ..35bed3714438 --- /dev/null +++ b/tools/perf/util/pmu-hybrid.h @@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef __PMU_HYBRID_H +#define __PMU_HYBRID_H + +#include +#include +#include +#include +#include "pmu.h" + +extern struct list_head perf_pmu__hybrid_pmus; + +#define perf_pmu__for_each_hybrid_pmu(pmu) \ + list_for_each_entry(pmu, _pmu__hybrid_pmus, hybrid_list) + +bool perf_pmu__hybrid_mounted(const char *name); + +#endif /* __PMU_HYBRID_H */ diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 10709ec1cc3e..35e9660c3904 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -25,6 +25,7 @@ #include "string2.h" #include "strbuf.h" #include "fncache.h" +#include "pmu-hybrid.h" struct perf_pmu perf_pmu__fake; @@ -613,7 +614,6 @@ static struct perf_cpu_map *__pmu_cpumask(const char *path) */ #define SYS_TEMPLATE_ID"./bus/event_source/devices/%s/identifier" #define CPUS_TEMPLATE_UNCORE "%s/bus/event_source/devices/%s/cpumask" -#define CPUS_TEMPLATE_CPU "%s/bus/event_source/devices/%s/cpus" static struct perf_cpu_map *pmu_cpumask(const char *name) { @@ -645,6 +645,9 @@ static bool pmu_is_uncore(const char *name) char path[PATH_MAX]; const char *sysfs; + if (perf_pmu__hybrid_mounted(name)) + return false; + sysfs = sysfs__mountpoint(); snprintf(path, PATH_MAX, CPUS_TEMPLATE_UNCORE, sysfs, name); return file_available(path); @@ -946,6 +949,7 @@ static struct perf_pmu *pmu_lookup(const char *name) pmu->is_uncore = pmu_is_uncore(name); if (pmu->is_uncore) pmu->id = pmu_id(name); + pmu->is_hybrid = perf_pmu__hybrid_mounted(name); pmu->max_precise = pmu_max_precise(name); pmu_add_cpu_aliases(, pmu); pmu_add_sys_aliases(, pmu); @@ -957,6 +961,9 @@ static struct perf_pmu *pmu_lookup(const char *name) list_splice(, >aliases); list_add_tail(>list, ); + if (pmu->is_hybrid) + list_add_tail(>hybrid_list, _pmu__hybrid_pmus); + pmu->default_config = perf_pmu__get_default_config(pmu); return pmu; diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 8493b1719e10..29289e7c2649 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -5,6 +5,7
[PATCH v3 06/27] perf pmu: Add hybrid helper functions
The functions perf_pmu__is_hybrid and perf_pmu__find_hybrid_pmu can be used to identify the hybrid platform and return the found hybrid cpu pmu. All the detected hybrid pmus have been saved in 'perf_pmu__hybrid_pmus' list. So we just need to search this list. perf_pmu__hybrid_type_to_pmu converts the user specified string to hybrid pmu name. This is used to support the '--cputype' option in next patches. perf_pmu__has_hybrid checks the existing of hybrid pmu. Note that, we have to define it in pmu.c (try to make pmu-hybrid.c no more symbol dependency), otherwise perf test python would be failed. Signed-off-by: Jin Yao --- v3: - Move perf_pmu__has_hybrid from pmu-hybrid.c to pmu.c. We have to add pmu-hybrid.c to python-ext-sources to solve symbol dependency issue found in perf test python. For perf_pmu__has_hybrid, it calls perf_pmu__scan, which is defined in pmu.c. It's very hard to add pmu.c to python-ext-sources, too much symbol dependency here. tools/perf/util/pmu-hybrid.c | 40 tools/perf/util/pmu-hybrid.h | 4 tools/perf/util/pmu.c| 11 ++ tools/perf/util/pmu.h| 2 ++ 4 files changed, 57 insertions(+) diff --git a/tools/perf/util/pmu-hybrid.c b/tools/perf/util/pmu-hybrid.c index 7316bf46e54b..86ba84d9469c 100644 --- a/tools/perf/util/pmu-hybrid.c +++ b/tools/perf/util/pmu-hybrid.c @@ -33,3 +33,43 @@ bool perf_pmu__hybrid_mounted(const char *name) snprintf(path, PATH_MAX, CPUS_TEMPLATE_CPU, sysfs, name); return file_available(path); } + +struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name) +{ + struct perf_pmu *pmu; + + if (!name) + return NULL; + + perf_pmu__for_each_hybrid_pmu(pmu) { + if (!strcmp(name, pmu->name)) + return pmu; + } + + return NULL; +} + +bool perf_pmu__is_hybrid(const char *name) +{ + return perf_pmu__find_hybrid_pmu(name) != NULL; +} + +char *perf_pmu__hybrid_type_to_pmu(const char *type) +{ + char *pmu_name = NULL; + + if (asprintf(_name, "cpu_%s", type) < 0) + return NULL; + + if (perf_pmu__is_hybrid(pmu_name)) + return pmu_name; + + /* +* pmu may be not scanned, check the sysfs. +*/ + if (perf_pmu__hybrid_mounted(pmu_name)) + return pmu_name; + + free(pmu_name); + return NULL; +} diff --git a/tools/perf/util/pmu-hybrid.h b/tools/perf/util/pmu-hybrid.h index 35bed3714438..d0fa7bc50a76 100644 --- a/tools/perf/util/pmu-hybrid.h +++ b/tools/perf/util/pmu-hybrid.h @@ -15,4 +15,8 @@ extern struct list_head perf_pmu__hybrid_pmus; bool perf_pmu__hybrid_mounted(const char *name); +struct perf_pmu *perf_pmu__find_hybrid_pmu(const char *name); +bool perf_pmu__is_hybrid(const char *name); +char *perf_pmu__hybrid_type_to_pmu(const char *type); + #endif /* __PMU_HYBRID_H */ diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 35e9660c3904..beff29981101 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -40,6 +40,7 @@ int perf_pmu_parse(struct list_head *list, char *name); extern FILE *perf_pmu_in; static LIST_HEAD(pmus); +static bool hybrid_scanned; /* * Parse & process all the sysfs attributes located under @@ -1823,3 +1824,13 @@ int perf_pmu__caps_parse(struct perf_pmu *pmu) return nr_caps; } + +bool perf_pmu__has_hybrid(void) +{ + if (!hybrid_scanned) { + hybrid_scanned = true; + perf_pmu__scan(NULL); + } + + return !list_empty(_pmu__hybrid_pmus); +} diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 29289e7c2649..5ed2abab7fe0 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -128,4 +128,6 @@ int perf_pmu__convert_scale(const char *scale, char **end, double *sval); int perf_pmu__caps_parse(struct perf_pmu *pmu); +bool perf_pmu__has_hybrid(void); + #endif /* __PMU_H */ -- 2.17.1
[PATCH v3 07/27] perf stat: Uniquify hybrid event name
It would be useful to tell user the pmu which the event belongs to. perf-stat has supported '--no-merge' option and it can print the pmu name after the event name, such as: "cycles [cpu_core]" Now this option is enabled by default for hybrid platform but change the format to: "cpu_core/cycles/" Signed-off-by: Jin Yao --- v3: - No functional change. tools/perf/builtin-stat.c | 4 tools/perf/util/stat-display.c | 13 +++-- 2 files changed, 15 insertions(+), 2 deletions(-) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 4bb48c6b6698..7b2dfe21c5a8 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -68,6 +68,7 @@ #include "util/affinity.h" #include "util/pfm.h" #include "util/bpf_counter.h" +#include "util/pmu-hybrid.h" #include "asm/bug.h" #include @@ -2371,6 +2372,9 @@ int cmd_stat(int argc, const char **argv) evlist__check_cpu_maps(evsel_list); + if (perf_pmu__has_hybrid()) + stat_config.no_merge = true; + /* * Initialize thread_map with comm names, * so we could print it out on output. diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 7f09cdaf5b60..161826938a00 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -17,6 +17,7 @@ #include "cgroup.h" #include #include "util.h" +#include "pmu-hybrid.h" #define CNTR_NOT_SUPPORTED "" #define CNTR_NOT_COUNTED "" @@ -526,6 +527,7 @@ static void uniquify_event_name(struct evsel *counter) { char *new_name; char *config; + int ret; if (counter->uniquified_name || !counter->pmu_name || !strncmp(counter->name, counter->pmu_name, @@ -540,8 +542,15 @@ static void uniquify_event_name(struct evsel *counter) counter->name = new_name; } } else { - if (asprintf(_name, -"%s [%s]", counter->name, counter->pmu_name) > 0) { + if (perf_pmu__has_hybrid()) { + ret = asprintf(_name, "%s/%s/", + counter->pmu_name, counter->name); + } else { + ret = asprintf(_name, "%s [%s]", + counter->name, counter->pmu_name); + } + + if (ret) { free(counter->name); counter->name = new_name; } -- 2.17.1
[PATCH v3 02/27] perf jevents: Support unit value "cpu_core" and "cpu_atom"
For some Intel platforms, such as Alderlake, which is a hybrid platform and it consists of atom cpu and core cpu. Each cpu has dedicated event list. Part of events are available on core cpu, part of events are available on atom cpu. The kernel exports new cpu pmus: cpu_core and cpu_atom. The event in json is added with a new field "Unit" to indicate which pmu the event is available on. For example, one event in cache.json, { "BriefDescription": "Counts the number of load ops retired that", "CollectPEBSRecord": "2", "Counter": "0,1,2,3", "EventCode": "0xd2", "EventName": "MEM_LOAD_UOPS_RETIRED_MISC.MMIO", "PEBScounters": "0,1,2,3", "SampleAfterValue": "103", "UMask": "0x80", "Unit": "cpu_atom" }, The unit "cpu_atom" indicates this event is only availabe on "cpu_atom". In generated pmu-events.c, we can see: { .name = "mem_load_uops_retired_misc.mmio", .event = "period=103,umask=0x80,event=0xd2", .desc = "Counts the number of load ops retired that. Unit: cpu_atom ", .topic = "cache", .pmu = "cpu_atom", }, But if without this patch, the "uncore_" prefix is added before "cpu_atom", such as: .pmu = "uncore_cpu_atom" That would be a wrong pmu. Signed-off-by: Jin Yao --- v3: - No change. tools/perf/pmu-events/jevents.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tools/perf/pmu-events/jevents.c b/tools/perf/pmu-events/jevents.c index 33aa3c885eaf..ed4f0bd72e5a 100644 --- a/tools/perf/pmu-events/jevents.c +++ b/tools/perf/pmu-events/jevents.c @@ -285,6 +285,8 @@ static struct map { { "imx8_ddr", "imx8_ddr" }, { "L3PMC", "amd_l3" }, { "DFPMC", "amd_df" }, + { "cpu_core", "cpu_core" }, + { "cpu_atom", "cpu_atom" }, {} }; -- 2.17.1
[PATCH v3 03/27] perf pmu: Simplify arguments of __perf_pmu__new_alias
Simplify the arguments of __perf_pmu__new_alias() by passing the whole 'struct pme_event' pointer. Signed-off-by: Jin Yao --- v3: - No change. tools/perf/util/pmu.c | 36 1 file changed, 16 insertions(+), 20 deletions(-) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 88da5cf6aee8..9ed9a6a8b2d2 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -306,18 +306,25 @@ static bool perf_pmu_merge_alias(struct perf_pmu_alias *newalias, } static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, -char *desc, char *val, -char *long_desc, char *topic, -char *unit, char *perpkg, -char *metric_expr, -char *metric_name, -char *deprecated) +char *desc, char *val, struct pmu_event *pe) { struct parse_events_term *term; struct perf_pmu_alias *alias; int ret; int num; char newval[256]; + char *long_desc = NULL, *topic = NULL, *unit = NULL, *perpkg = NULL, +*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL; + + if (pe) { + long_desc = (char *)pe->long_desc; + topic = (char *)pe->topic; + unit = (char *)pe->unit; + perpkg = (char *)pe->perpkg; + metric_expr = (char *)pe->metric_expr; + metric_name = (char *)pe->metric_name; + deprecated = (char *)pe->deprecated; + } alias = malloc(sizeof(*alias)); if (!alias) @@ -406,8 +413,7 @@ static int perf_pmu__new_alias(struct list_head *list, char *dir, char *name, FI /* Remove trailing newline from sysfs file */ strim(buf); - return __perf_pmu__new_alias(list, dir, name, NULL, buf, NULL, NULL, NULL, -NULL, NULL, NULL, NULL); + return __perf_pmu__new_alias(list, dir, name, NULL, buf, NULL); } static inline bool pmu_alias_info_file(char *name) @@ -793,11 +799,7 @@ void pmu_add_cpu_aliases_map(struct list_head *head, struct perf_pmu *pmu, /* need type casts to override 'const' */ __perf_pmu__new_alias(head, NULL, (char *)pe->name, (char *)pe->desc, (char *)pe->event, - (char *)pe->long_desc, (char *)pe->topic, - (char *)pe->unit, (char *)pe->perpkg, - (char *)pe->metric_expr, - (char *)pe->metric_name, - (char *)pe->deprecated); + pe); } } @@ -864,13 +866,7 @@ static int pmu_add_sys_aliases_iter_fn(struct pmu_event *pe, void *data) (char *)pe->name, (char *)pe->desc, (char *)pe->event, - (char *)pe->long_desc, - (char *)pe->topic, - (char *)pe->unit, - (char *)pe->perpkg, - (char *)pe->metric_expr, - (char *)pe->metric_name, - (char *)pe->deprecated); + pe); } return 0; -- 2.17.1
[PATCH v3 01/27] tools headers uapi: Update tools's copy of linux/perf_event.h
To get the changes in: Liang Kan's patch ("perf: Introduce PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU") Kan's patch is in review at the moment, but the following perf tool patches need this interface for hybrid support. This patch can be removed after Kan's patch is upstreamed. Signed-off-by: Jin Yao --- v3: - No change. tools/include/uapi/linux/perf_event.h | 26 ++ 1 file changed, 26 insertions(+) diff --git a/tools/include/uapi/linux/perf_event.h b/tools/include/uapi/linux/perf_event.h index ad15e40d7f5d..c0a511eea498 100644 --- a/tools/include/uapi/linux/perf_event.h +++ b/tools/include/uapi/linux/perf_event.h @@ -33,6 +33,8 @@ enum perf_type_id { PERF_TYPE_HW_CACHE = 3, PERF_TYPE_RAW = 4, PERF_TYPE_BREAKPOINT= 5, + PERF_TYPE_HARDWARE_PMU = 6, + PERF_TYPE_HW_CACHE_PMU = 7, PERF_TYPE_MAX, /* non-ABI */ }; @@ -94,6 +96,30 @@ enum perf_hw_cache_op_result_id { PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ }; +/* + * attr.config layout for type PERF_TYPE_HARDWARE* and PERF_TYPE_HW_CACHE* + * PERF_TYPE_HARDWARE: 0xAA + * AA: hardware event ID + * PERF_TYPE_HW_CACHE: 0xCCBBAA + * AA: hardware cache ID + * BB: hardware cache op ID + * CC: hardware cache op result ID + * PERF_TYPE_HARDWARE_PMU: 0xDD00AA + * AA: hardware event ID + * DD: PMU type ID + * PERF_TYPE_HW_CACHE_PMU: 0xDD00CCBBAA + * AA: hardware cache ID + * BB: hardware cache op ID + * CC: hardware cache op result ID + * DD: PMU type ID + */ +#define PERF_HW_CACHE_ID_SHIFT 0 +#define PERF_HW_CACHE_OP_ID_SHIFT 8 +#define PERF_HW_CACHE_OP_RESULT_ID_SHIFT 16 +#define PERF_HW_CACHE_EVENT_MASK 0xff + +#define PERF_PMU_TYPE_SHIFT32 + /* * Special "software" events provided by the kernel, even if the hardware * does not support performance events. These events measure various -- 2.17.1
[PATCH v3 00/27] perf tool: AlderLake hybrid support series 1
AlderLake uses a hybrid architecture utilizing Golden Cove cores (core cpu) and Gracemont cores (atom cpu). Each cpu has dedicated event list. Some events are available on core cpu, some events are available on atom cpu and some events can be available on both. Kernel exports new pmus "cpu_core" and "cpu_atom" through sysfs: /sys/devices/cpu_core /sys/devices/cpu_atom cat /sys/devices/cpu_core/cpus 0-15 cat /sys/devices/cpu_atom/cpus 16-23 In this example, core cpus are 0-15 and atom cpus are 16-23. To enable a core only event or atom only event: cpu_core// or cpu_atom// Count the 'cycles' event on core cpus. # perf stat -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 12,853,951,349 cpu_core/cycles/ 1.002581249 seconds time elapsed If one event is available on both atom cpu and core cpu, two events are created automatically. # perf stat -e cycles -a -- sleep 1 Performance counter stats for 'system wide': 12,856,467,438 cpu_core/cycles/ 6,404,634,785 cpu_atom/cycles/ 1.002453013 seconds time elapsed Group is supported if the events are from same pmu, otherwise a warning is displayed and disable grouping automatically. # perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}' -a -- sleep 1 Performance counter stats for 'system wide': 12,863,866,968 cpu_core/cycles/ 554,795,017 cpu_core/instructions/ 1.002616117 seconds time elapsed # perf stat -e '{cpu_core/cycles/,cpu_atom/instructions/}' -a -- sleep 1 WARNING: events in group from different hybrid PMUs! WARNING: grouped events cpus do not match, disabling group: anon group { cpu_core/cycles/, cpu_atom/instructions/ } Performance counter stats for 'system wide': 6,283,970 cpu_core/cycles/ 765,635 cpu_atom/instructions/ 1.003959036 seconds time elapsed Note that, since the whole patchset for AlderLake hybrid support is very large (40+ patches). For simplicity, it's splitted into several patch series. The patch series 1 only supports the basic functionality. The advanced supports for perf-c2c/perf-mem/topdown/metrics/topology header and others will be added in follow-up patch series. The perf tool codes can also be found at: https://github.com/yaoj/perf.git v3: --- - Drop 'perf evlist: Hybrid event uses its own cpus'. This patch is wide and actually it's not very necessary. The current perf framework has processed the cpus for evsel well even for hybrid evsel. So this patch can be dropped. - Drop 'perf evsel: Adjust hybrid event and global event mixed group'. The patch is a bit tricky and hard to understand. In v3, we will disable grouping when the group members are from different PMUs. So this patch would be not necessary. - Create parse-events-hybrid.c/parse-events-hybrid.h and evlist-hybrid.c/evlist-hybrid.h. Move hybrid related codes to these files. - Create a new patch 'perf pmu: Support 'cycles' and 'branches' inside hybrid PMU' to support 'cycles' and 'branches' inside PMU. - Create a new patch 'perf record: Uniquify hybrid event name' to tell user the pmu which the event belongs to for perf-record. - If group members are from different hybrid PMUs, shows warning and disable grouping. - Other refining and refactoring. v2: --- - Drop kernel patches (Kan posted the series "Add Alder Lake support for perf (kernel)" separately). - Drop the patches for perf-c2c/perf-mem/topdown/metrics/topology header supports, which will be added in series 2 or series 3. - Simplify the arguments of __perf_pmu__new_alias() by passing the 'struct pme_event' pointer. - Check sysfs validity before access. - Use pmu style event name, such as "cpu_core/cycles/". - Move command output two chars to the right. - Move pmu hybrid functions to new created pmu-hybrid.c/pmu-hybrid.h. This is to pass the perf test python case. Jin Yao (27): tools headers uapi: Update tools's copy of linux/perf_event.h perf jevents: Support unit value "cpu_core" and "cpu_atom" perf pmu: Simplify arguments of __perf_pmu__new_alias perf pmu: Save pmu name perf pmu: Save detected hybrid pmus to a global pmu list perf pmu: Add hybrid helper functions perf stat: Uniquify hybrid event name perf parse-events: Create two hybrid hardware events perf parse-events: Create two hybrid cache events perf parse-events: Create two hybrid raw events perf pmu: Support 'cycles' and 'branches' inside hybrid PMU perf parse-events: Support no alias assigned event inside hybrid PMU perf record: Create two hybrid 'cycles' events by default perf stat: Add default hybrid events perf stat: Filter out unmatched aggregation for hybrid event perf stat: Warn group events from different hybrid PMU perf script: Support PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU perf
[PATCH v3 2/2] perf test: Add CVS summary test
The patch "perf stat: Align CSV output for summary mode" aligned CVS output and added "summary" to the first column of summary lines. Now we check if the "summary" string is added to the CVS output. If we set '--no-cvs-summary' option, the "summary" string would not be added, also check with this case. Signed-off-by: Jin Yao --- v3: - New in v3. tools/perf/tests/shell/stat+cvs_summary.sh | 31 ++ 1 file changed, 31 insertions(+) create mode 100755 tools/perf/tests/shell/stat+cvs_summary.sh diff --git a/tools/perf/tests/shell/stat+cvs_summary.sh b/tools/perf/tests/shell/stat+cvs_summary.sh new file mode 100755 index ..dd14f2ce7f6b --- /dev/null +++ b/tools/perf/tests/shell/stat+cvs_summary.sh @@ -0,0 +1,31 @@ +#!/bin/sh +# perf stat cvs summary test +# SPDX-License-Identifier: GPL-2.0 + +set -e + +# +# 1.001364330 9224197 cycles 8012885033 100.00 +# summary 9224197 cycles 8012885033 100.00 +# +perf stat -e cycles -x' ' -I1000 --interval-count 1 --summary 2>&1 | \ +grep -e summary | \ +while read summary num event run pct +do + if [ $summary != "summary" ]; then + exit 1 + fi +done + +# +# 1.001360298 9148534 cycles 8012853854 100.00 +#9148534 cycles 8012853854 100.00 +# +perf stat -e cycles -x' ' -I1000 --interval-count 1 --summary --no-cvs-summary 2>&1 | \ +grep -e summary | \ +while read num event run pct +do + exit 1 +done + +exit 0 -- 2.17.1
[PATCH v3 1/2] perf stat: Align CSV output for summary mode
,8013356501,100.00,0.36,insn per cycle 553564,,branches,8013366204,100.00,69.081,K/sec 54021,,branch-misses,8013375952,100.00,9.76,of all branches Signed-off-by: Jin Yao --- v3: - No change. v2: - Add new option '--no-cvs-summary'. - Add perf config variable 'stat.no-cvs-summary'. tools/perf/Documentation/perf-stat.txt | 9 + tools/perf/builtin-stat.c | 7 +++ tools/perf/util/config.c | 3 +++ tools/perf/util/stat-display.c | 6 ++ tools/perf/util/stat.h | 2 ++ 5 files changed, 27 insertions(+) diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt index 3055aad38d46..854597e70406 100644 --- a/tools/perf/Documentation/perf-stat.txt +++ b/tools/perf/Documentation/perf-stat.txt @@ -471,6 +471,15 @@ convenient for post processing. --summary:: Print summary for interval mode (-I). +--no-cvs-summary:: +Don't print 'summary' at the first column for CVS summary output. +This option must be used with -x and --summary. + +This option can be enabled in perf config by setting the variable +'stat.no-cvs-summary'. + +$ perf config stat.no-cvs-summary=true + EXAMPLES diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 2e2e4a8345ea..3823dd5fd6e8 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1083,6 +1083,11 @@ void perf_stat__set_big_num(int set) stat_config.big_num = (set != 0); } +void perf_stat__set_no_cvs_summary(int set) +{ + stat_config.no_cvs_summary = (set != 0); +} + static int stat__set_big_num(const struct option *opt __maybe_unused, const char *s __maybe_unused, int unset) { @@ -1235,6 +1240,8 @@ static struct option stat_options[] = { "threads of same physical core"), OPT_BOOLEAN(0, "summary", _config.summary, "print summary for interval mode"), + OPT_BOOLEAN(0, "no-cvs-summary", _config.no_cvs_summary, + "don't print 'summary' for CVS summary output"), OPT_BOOLEAN(0, "quiet", _config.quiet, "don't print output (useful with record)"), #ifdef HAVE_LIBPFM diff --git a/tools/perf/util/config.c b/tools/perf/util/config.c index 6984c77068a3..dbf585460791 100644 --- a/tools/perf/util/config.c +++ b/tools/perf/util/config.c @@ -457,6 +457,9 @@ static int perf_stat_config(const char *var, const char *value) if (!strcmp(var, "stat.big-num")) perf_stat__set_big_num(perf_config_bool(var, value)); + if (!strcmp(var, "stat.no-cvs-summary")) + perf_stat__set_no_cvs_summary(perf_config_bool(var, value)); + /* Add other config variables here. */ return 0; } diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 7f09cdaf5b60..2e7fec0bd8f3 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -439,6 +439,12 @@ static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int if (counter->cgrp) os.nfields++; } + + if (!config->no_cvs_summary && config->csv_output && + config->summary && !config->interval) { + fprintf(config->output, "%16s%s", "summary", config->csv_sep); + } + if (run == 0 || ena == 0 || counter->counts->scaled == -1) { if (config->metric_only) { pm(config, , NULL, "", "", 0); diff --git a/tools/perf/util/stat.h b/tools/perf/util/stat.h index 41107b8deac5..def0cdc84133 100644 --- a/tools/perf/util/stat.h +++ b/tools/perf/util/stat.h @@ -128,6 +128,7 @@ struct perf_stat_config { bool all_user; bool percore_show_thread; bool summary; + bool no_cvs_summary; bool metric_no_group; bool metric_no_merge; bool stop_read_counter; @@ -160,6 +161,7 @@ struct perf_stat_config { }; void perf_stat__set_big_num(int set); +void perf_stat__set_no_cvs_summary(int set); void update_stats(struct stats *stats, u64 val); double avg_stats(struct stats *stats); -- 2.17.1
Re: [PATCH v2] perf stat: Align CSV output for summary mode
Hi Arnaldo, On 3/18/2021 9:15 PM, Arnaldo Carvalho de Melo wrote: Em Wed, Mar 17, 2021 at 02:51:42PM -0700, Andi Kleen escreveu: If you care about not breaking existing scripts, then the output they get with what they use as command line options must continue to produce the same output. It's not clear there are any useful ones (except for tools that handle both). It's really hard to parse the previous mess. It's simply not valid CSV. That's why I'm arguing that keeping compatibility is not useful here. We would be stuck with the broken mess as default forever. Fair enough, lets fix the default then. Jin, can you please consider adding a 'perf test' shell entry to parse the CSV mode with/without that summary? This way we'll notice when the new normal gets broken. - Arnaldo Thanks Arnaldo! I will post v3 with the perf test script. Thanks Jin Yao
Re: [PATCH v2 11/27] perf parse-events: Support hardware events inside PMU
Hi Jiri, On 3/17/2021 6:06 PM, Jiri Olsa wrote: On Wed, Mar 17, 2021 at 10:12:03AM +0800, Jin, Yao wrote: On 3/16/2021 10:04 PM, Jiri Olsa wrote: On Tue, Mar 16, 2021 at 09:49:42AM +0800, Jin, Yao wrote: SNIP Performance counter stats for 'system wide': 136,655,302 cpu_core/branch-instructions/ 1.003171561 seconds time elapsed So we need special rules for both cycles and branches. The worse thing is, we also need to process the hardware cache events. # ./perf stat -e cpu_core/LLC-loads/ event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' If we use special rules for establishing all event mapping, that looks too much. :( hmmm but wait, currently we do not support events like this: 'cpu/cycles/' 'cpu/branches/' the pmu style accepts only 'events' or 'format' terms within // we made hw events like 'cycles','instructions','branches' special to be used without the pmu so why do we need to support cpu_code/cycles/ ? jirka Actually we have to support pmu style event for hybrid platform. User may want to enable the events from specified pmus and also with flexible grouping. For example, perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}' -e '{cpu_atom/cycles/,cpu_atom/instructions/}' This usage is common and reasonable. So I think we may need to support pmu style events. sure, but we don't support 'cpu/cycles/' but we support 'cpu/cpu-cycles/' why do you insist on supporting cpu_core/cycles/ ? jirka I'm OK to only support 'cpu_core/cpu-cycles/' or 'cpu_atom/cpu-cycles/'. But what would we do for cache event? 'perf stat -e LLC-loads' is OK, but 'perf stat -e cpu/LLC-loads/' is not supported currently. For hybrid platform, user may only want to enable the LLC-loads on core CPUs or on atom CPUs. That's reasonable. While if we don't support the pmu style event, how to satisfy this requirement? If we can support the pmu style event, we can also use the same way for cpu_core/cycles/. At least it's not a bad thing, right? :) Thanks Jin Yao
[PATCH v2] perf stat: Align CSV output for summary mode
,8013356501,100.00,0.36,insn per cycle 553564,,branches,8013366204,100.00,69.081,K/sec 54021,,branch-misses,8013375952,100.00,9.76,of all branches Signed-off-by: Jin Yao --- v2: - Add new option '--no-cvs-summary'. - Add perf config variable 'stat.no-cvs-summary'. tools/perf/Documentation/perf-stat.txt | 9 + tools/perf/builtin-stat.c | 7 +++ tools/perf/util/config.c | 3 +++ tools/perf/util/stat-display.c | 6 ++ tools/perf/util/stat.h | 2 ++ 5 files changed, 27 insertions(+) diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt index 3055aad38d46..854597e70406 100644 --- a/tools/perf/Documentation/perf-stat.txt +++ b/tools/perf/Documentation/perf-stat.txt @@ -471,6 +471,15 @@ convenient for post processing. --summary:: Print summary for interval mode (-I). +--no-cvs-summary:: +Don't print 'summary' at the first column for CVS summary output. +This option must be used with -x and --summary. + +This option can be enabled in perf config by setting the variable +'stat.no-cvs-summary'. + +$ perf config stat.no-cvs-summary=true + EXAMPLES diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 2e2e4a8345ea..3823dd5fd6e8 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1083,6 +1083,11 @@ void perf_stat__set_big_num(int set) stat_config.big_num = (set != 0); } +void perf_stat__set_no_cvs_summary(int set) +{ + stat_config.no_cvs_summary = (set != 0); +} + static int stat__set_big_num(const struct option *opt __maybe_unused, const char *s __maybe_unused, int unset) { @@ -1235,6 +1240,8 @@ static struct option stat_options[] = { "threads of same physical core"), OPT_BOOLEAN(0, "summary", _config.summary, "print summary for interval mode"), + OPT_BOOLEAN(0, "no-cvs-summary", _config.no_cvs_summary, + "don't print 'summary' for CVS summary output"), OPT_BOOLEAN(0, "quiet", _config.quiet, "don't print output (useful with record)"), #ifdef HAVE_LIBPFM diff --git a/tools/perf/util/config.c b/tools/perf/util/config.c index 6984c77068a3..dbf585460791 100644 --- a/tools/perf/util/config.c +++ b/tools/perf/util/config.c @@ -457,6 +457,9 @@ static int perf_stat_config(const char *var, const char *value) if (!strcmp(var, "stat.big-num")) perf_stat__set_big_num(perf_config_bool(var, value)); + if (!strcmp(var, "stat.no-cvs-summary")) + perf_stat__set_no_cvs_summary(perf_config_bool(var, value)); + /* Add other config variables here. */ return 0; } diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 7f09cdaf5b60..2e7fec0bd8f3 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -439,6 +439,12 @@ static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int if (counter->cgrp) os.nfields++; } + + if (!config->no_cvs_summary && config->csv_output && + config->summary && !config->interval) { + fprintf(config->output, "%16s%s", "summary", config->csv_sep); + } + if (run == 0 || ena == 0 || counter->counts->scaled == -1) { if (config->metric_only) { pm(config, , NULL, "", "", 0); diff --git a/tools/perf/util/stat.h b/tools/perf/util/stat.h index 41107b8deac5..def0cdc84133 100644 --- a/tools/perf/util/stat.h +++ b/tools/perf/util/stat.h @@ -128,6 +128,7 @@ struct perf_stat_config { bool all_user; bool percore_show_thread; bool summary; + bool no_cvs_summary; bool metric_no_group; bool metric_no_merge; bool stop_read_counter; @@ -160,6 +161,7 @@ struct perf_stat_config { }; void perf_stat__set_big_num(int set); +void perf_stat__set_no_cvs_summary(int set); void update_stats(struct stats *stats, u64 val); double avg_stats(struct stats *stats); -- 2.17.1
Re: [PATCH v2 11/27] perf parse-events: Support hardware events inside PMU
On 3/16/2021 10:04 PM, Jiri Olsa wrote: On Tue, Mar 16, 2021 at 09:49:42AM +0800, Jin, Yao wrote: SNIP Performance counter stats for 'system wide': 136,655,302 cpu_core/branch-instructions/ 1.003171561 seconds time elapsed So we need special rules for both cycles and branches. The worse thing is, we also need to process the hardware cache events. # ./perf stat -e cpu_core/LLC-loads/ event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' If we use special rules for establishing all event mapping, that looks too much. :( hmmm but wait, currently we do not support events like this: 'cpu/cycles/' 'cpu/branches/' the pmu style accepts only 'events' or 'format' terms within // we made hw events like 'cycles','instructions','branches' special to be used without the pmu so why do we need to support cpu_code/cycles/ ? jirka Actually we have to support pmu style event for hybrid platform. User may want to enable the events from specified pmus and also with flexible grouping. For example, perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}' -e '{cpu_atom/cycles/,cpu_atom/instructions/}' This usage is common and reasonable. So I think we may need to support pmu style events. Thanks Jin Yao
Re: [PATCH] perf stat: Align CSV output for summary mode
On 3/17/2021 9:30 AM, Andi Kleen wrote: Is it serious or just a joke? :) I would prefer to not be compatible (at least not until someone complains), but if compatibility is required then yes opting in to the broken format would be better. Perhaps not with that name. And the option could be hidden in the perf config file instead of being on the command line. -Andi That makes sense, thanks Andi! Thanks Jin Yao
Re: [PATCH] perf stat: Align CSV output for summary mode
On 3/17/2021 5:55 AM, Jiri Olsa wrote: On Tue, Mar 16, 2021 at 01:02:20PM -0700, Andi Kleen wrote: On Tue, Mar 16, 2021 at 04:05:13PM -0300, Arnaldo Carvalho de Melo wrote: Em Tue, Mar 16, 2021 at 09:34:21AM -0700, Andi Kleen escreveu: looks ok, but maybe make the option more related to CVS, like: --x-summary, --cvs-summary ...? Actually I don't think it should be a new option. I doubt anyone could parse the previous mess. So just make it default with -x In these cases I always fear that people are already parsing that mess by considering the summary lines to be the ones not starting with spaces, and now we go on and change it to be "better" by prefixing it with "summary" and... break existing scripts. I think it was just one version or so? FWIW perf has broken CSV output several times, I added workarounds to toplev every time. Having a broken version for a short time shouldn't be too bad. I actually had a workaround for this one, but it can parse either way. Can we do this with a new option? I.e. like --cvs-summary? If you do it I would add an option for the old broken format --i-want-broken-csv. But not require the option forever just to get sane output. I like that.. also we'll find out how many people are actually parsing that ;-) jirka Is it serious or just a joke? :) Thanks Jin Yao Or maybe only a perf config option. -Andi
[PATCH] perf stat: Align CSV output for summary mode
perf-stat has supported the summary mode. But the summary lines break the CSV output so it's hard for scripts to parse the result. Before: # perf stat -x, -I1000 --interval-count 1 --summary 1.001323097,8013.48,msec,cpu-clock,8013483384,100.00,8.013,CPUs utilized 1.001323097,270,,context-switches,8013513297,100.00,0.034,K/sec 1.001323097,13,,cpu-migrations,8013530032,100.00,0.002,K/sec 1.001323097,184,,page-faults,8013546992,100.00,0.023,K/sec 1.001323097,20574191,,cycles,8013551506,100.00,0.003,GHz 1.001323097,10562267,,instructions,8013564958,100.00,0.51,insn per cycle 1.001323097,2019244,,branches,8013575673,100.00,0.252,M/sec 1.001323097,106152,,branch-misses,8013585776,100.00,5.26,of all branches 8013.48,msec,cpu-clock,8013483384,100.00,7.984,CPUs utilized 270,,context-switches,8013513297,100.00,0.034,K/sec 13,,cpu-migrations,8013530032,100.00,0.002,K/sec 184,,page-faults,8013546992,100.00,0.023,K/sec 20574191,,cycles,8013551506,100.00,0.003,GHz 10562267,,instructions,8013564958,100.00,0.51,insn per cycle 2019244,,branches,8013575673,100.00,0.252,M/sec 106152,,branch-misses,8013585776,100.00,5.26,of all branches The summary line loses the timestamp column, which breaks the CVS output. We add a column at the 'timestamp' position and it just says 'summary' for the summary line. After: # perf stat -x, -I1000 --interval-count 1 --summary 1.001196053,8012.72,msec,cpu-clock,8012722903,100.00,8.013,CPUs utilized 1.001196053,218,,context-switches,8012753271,100.00,0.027,K/sec 1.001196053,9,,cpu-migrations,8012769767,100.00,0.001,K/sec 1.001196053,0,,page-faults,8012786257,100.00,0.000,K/sec 1.001196053,15004518,,cycles,8012790637,100.00,0.002,GHz 1.001196053,7954691,,instructions,8012804027,100.00,0.53,insn per cycle 1.001196053,1590259,,branches,8012814766,100.00,0.198,M/sec 1.001196053,82601,,branch-misses,8012824365,100.00,5.19,of all branches summary,8012.72,msec,cpu-clock,8012722903,100.00,7.986,CPUs utilized summary,218,,context-switches,8012753271,100.00,0.027,K/sec summary,9,,cpu-migrations,8012769767,100.00,0.001,K/sec summary,0,,page-faults,8012786257,100.00,0.000,K/sec summary,15004518,,cycles,8012790637,100.00,0.002,GHz summary,7954691,,instructions,8012804027,100.00,0.53,insn per cycle summary,1590259,,branches,8012814766,100.00,0.198,M/sec summary,82601,,branch-misses,8012824365,100.00,5.19,of all branches Now it's easy for script to analyse the summary lines. Signed-off-by: Jin Yao --- tools/perf/util/stat-display.c | 4 1 file changed, 4 insertions(+) diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index 7f09cdaf5b60..c4183d3e87a4 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -439,6 +439,10 @@ static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int if (counter->cgrp) os.nfields++; } + + if (config->csv_output && config->summary && !config->interval) + fprintf(config->output, "%16s%s", "summary", config->csv_sep); + if (run == 0 || ena == 0 || counter->counts->scaled == -1) { if (config->metric_only) { pm(config, , NULL, "", "", 0); -- 2.17.1
Re: [PATCH v2 16/27] perf evlist: Warn as events from different hybrid PMUs in a group
Hi Jiri, On 3/16/2021 7:03 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:31PM +0800, Jin Yao wrote: SNIP goto try_again; } + + if (errno == EINVAL && perf_pmu__hybrid_exist()) + evlist__warn_hybrid_group(evlist); rc = -errno; evsel__open_strerror(pos, >target, errno, msg, sizeof(msg)); ui__error("%s\n", msg); diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 7a732508b2b4..6f780a039db0 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -239,6 +239,9 @@ static void evlist__check_cpu_maps(struct evlist *evlist) struct evsel *evsel, *pos, *leader; char buf[1024]; + if (evlist__hybrid_exist(evlist)) + return; this should be in separate patch and explained Now I have another idea. If a group consists of atom events and core events, we still follow current disabling group solution? I mean removing following code: if (evlist__hybrid_exist(evlist)) return; evlist__check_cpu_maps then continues running and disabling the group. But also report with a warning that says "WARNING: Group has events from different hybrid PMUs". Do you like this way? + evlist__for_each_entry(evlist, evsel) { leader = evsel->leader; @@ -726,6 +729,10 @@ enum counter_recovery { static enum counter_recovery stat_handle_error(struct evsel *counter) { char msg[BUFSIZ]; + + if (perf_pmu__hybrid_exist() && errno == EINVAL) + evlist__warn_hybrid_group(evsel_list); + /* * PPC returns ENXIO for HW counters until 2.6.37 * (behavior changed with commit b0a873e). diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c index f139151b9433..5ec891418cdd 100644 --- a/tools/perf/util/evlist.c +++ b/tools/perf/util/evlist.c @@ -2224,3 +2224,47 @@ void evlist__invalidate_all_cpus(struct evlist *evlist) perf_cpu_map__put(evlist->core.all_cpus); evlist->core.all_cpus = perf_cpu_map__empty_new(1); } + +static bool group_hybrid_conflict(struct evsel *leader) +{ + struct evsel *pos, *prev = NULL; + + for_each_group_evsel(pos, leader) { + if (!pos->pmu_name || !perf_pmu__is_hybrid(pos->pmu_name)) + continue; + + if (prev && strcmp(prev->pmu_name, pos->pmu_name)) + return true; + + prev = pos; + } + + return false; +} + +void evlist__warn_hybrid_group(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel__is_group_leader(evsel) && + evsel->core.nr_members > 1 && hm, could we just iterate all the members and make sure the first found hybrid event's pmu matches the pmu of the rest hybrid events in the list? '{cpu_core/event1/,cpu_core/event2/}','{cpu_atom/event3/,cpu_atom/event4/}' Two or more groups need to be supported. We get the first hybrid event's pmu (cpu_core in this example) but it doesn't match the cpu_atom/event3/ and cpu_atom/event4/. But actually this case should be supported, right? + group_hybrid_conflict(evsel)) { + WARN_ONCE(1, "WARNING: Group has events from " +"different hybrid PMUs\n"); + return; + } + } +} + +bool evlist__hybrid_exist(struct evlist *evlist) evlist__has_hybrid seems better Yes, agree. Thanks Jin Yao jirka
Re: [PATCH v2 17/27] perf evsel: Adjust hybrid event and global event mixed group
Hi Jiri, On 3/16/2021 7:04 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:32PM +0800, Jin Yao wrote: A group mixed with hybrid event and global event is allowed. For example, group leader is 'cpu-clock' and the group member is 'cpu_atom/cycles/'. e.g. perf stat -e '{cpu-clock,cpu_atom/cycles/}' -a The challenge is their available cpus are not fully matched. For example, 'cpu-clock' is available on CPU0-CPU23, but 'cpu_core/cycles/' is available on CPU16-CPU23. When getting the group id for group member, we must be very careful because the cpu for 'cpu-clock' is not equal to the cpu for 'cpu_atom/cycles/'. Actually the cpu here is the index of evsel->core.cpus, not the real CPU ID. e.g. cpu0 for 'cpu-clock' is CPU0, but cpu0 for 'cpu_atom/cycles/' is CPU16. Another challenge is for group read. The events in group may be not available on all cpus. For example the leader is a software event and it's available on CPU0-CPU1, but the group member is a hybrid event and it's only available on CPU1. For CPU0, we have only one event, but for CPU1 we have two events. So we need to change the read size according to the real number of events on that cpu. ugh, this is really bad.. do we really want to support it? ;-) I guess we need that for metrics.. Yes, it's a bit of pain but the user case makes sense. Some metrics need the event group which consists of global event + hybrid event. For example, CPU_Utilization = 'cpu_clk_unhalted.ref_tsc' / 'msr/tsc/'. 'msr/tsc/' is a global event. It's valid on all CPUs. But 'cpu_clk_unhalted.ref' is hybrid event. 'cpu_core/cpu_clk_unhalted.ref/' is valid on core CPUs 'cpu_atom/cpu_clk_unhalted.ref/' is valid on atom CPUs. So we have to support this usage. :) SNIP Performance counter stats for 'system wide': 24,059.14 msec cpu-clock # 23.994 CPUs utilized 6,406,677,892 cpu_atom/cycles/ # 266.289 M/sec 1.002699058 seconds time elapsed For cpu_atom/cycles/, cpu16-cpu23 are set with valid group fd (cpu-clock's fd on that cpu). For counting results, cpu-clock has 24 cpus aggregation and cpu_atom/cycles/ has 8 cpus aggregation. That's expected. But if the event order is changed, e.g. '{cpu_atom/cycles/,cpu-clock}', there leaves more works to do. root@ssp-pwrt-002:~# ./perf stat -e '{cpu_atom/cycles/,cpu-clock}' -a -vvv -- sleep 1 what id you add the other hybrid pmu event? or just cycles? Do you mean the config for cpu_atom/cycles/? Let's see the log. root@ssp-pwrt-002:~# perf stat -e '{cpu_atom/cycles/,cpu-clock}' -a -vvv -- sleep 1 Control descriptor is not initialized perf_event_attr: type 6 size 120 config 0xa sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING|ID|GROUP disabled 1 inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 16 group_fd -1 flags 0x8 = 3 sys_perf_event_open: pid -1 cpu 17 group_fd -1 flags 0x8 = 4 sys_perf_event_open: pid -1 cpu 18 group_fd -1 flags 0x8 = 5 sys_perf_event_open: pid -1 cpu 19 group_fd -1 flags 0x8 = 7 sys_perf_event_open: pid -1 cpu 20 group_fd -1 flags 0x8 = 8 sys_perf_event_open: pid -1 cpu 21 group_fd -1 flags 0x8 = 9 sys_perf_event_open: pid -1 cpu 22 group_fd -1 flags 0x8 = 10 sys_perf_event_open: pid -1 cpu 23 group_fd -1 flags 0x8 = 11 perf_event_attr: type 1 size 120 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING|ID|GROUP inherit 1 exclude_guest1 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 12 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8 = 13 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8 = 14 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8 = 15 sys_perf_event_open: pid -1 cpu 4 group_fd -1 flags 0x8 = 16 sys_perf_event_open: pid -1 cpu 5 group_fd -1 flags 0x8 = 17 sys_perf_event_open: pid -1 cpu 6 group_fd -1 flags 0x8 = 18 sys_perf_event_open: pid -1 cpu 7 group_fd -1 flags 0x8 = 19 sys_perf_event_open: pid -1 cpu 8 group_fd -1 flags 0x8 = 20 sys_perf_event_open: pid -1 cpu 9 group_fd -1 flags 0x8 = 21 sys_perf_event_open: pid -1 cpu 10 group_fd -1 flags 0x8 = 22 sys_perf_event_open: pid -1 cpu 11 group_fd -1 flags 0x8 = 23 sys_perf_event_open: pid -1 cpu 12 group_fd -1 flags 0x8 = 24 sys_perf_event_open: pid -1 cpu 13 group_fd -1 flags 0x8 =
Re: [PATCH v2 10/27] perf parse-events: Create two hybrid cache events
Hi Jiri, On 3/16/2021 7:05 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:25PM +0800, Jin Yao wrote: SNIP + config_terms, pmu); + if (ret) + return ret; + } + + return 0; +} + int parse_events_add_cache(struct list_head *list, int *idx, char *type, char *op_result1, char *op_result2, struct parse_events_error *err, @@ -474,7 +516,8 @@ int parse_events_add_cache(struct list_head *list, int *idx, char name[MAX_NAME_LEN], *config_name; int cache_type = -1, cache_op = -1, cache_result = -1; char *op_result[2] = { op_result1, op_result2 }; - int i, n; + int i, n, ret; + bool hybrid; /* * No fallback - if we cannot get a clear cache type @@ -534,6 +577,15 @@ int parse_events_add_cache(struct list_head *list, int *idx, if (get_config_terms(head_config, _terms)) return -ENOMEM; } + + if (!perf_pmu__hybrid_exist()) + perf_pmu__scan(NULL); actualy how about construct like: perf_pmu_is_hybrid() return hybrid_add_event_cache(...) return add_event(...) with: perf_pmu_is_hybrid() { static bool initialized; if (!initialized) { initialized = true; perf_pmu__scan(NULL) } return ... } jirka Thanks, that's a good solution. I will do that in v3. Thanks Jin Yao + + ret = add_hybrid_cache(list, idx, , config_name ? : name, + _terms, ); + if (hybrid) + return ret; + return add_event(list, idx, , config_name ? : name, _terms); } -- 2.17.1
Re: [PATCH v2 09/27] perf parse-events: Create two hybrid hardware events
Hi Jiri, On 3/16/2021 7:05 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:24PM +0800, Jin Yao wrote: For hardware events, they have pre-defined configs. The kernel needs to know where the event comes from (e.g. from cpu_core pmu or from cpu_atom pmu). But the perf type 'PERF_TYPE_HARDWARE' can't carry pmu information. So the kernel introduces a new type 'PERF_TYPE_HARDWARE_PMU'. The new attr.config layout for PERF_TYPE_HARDWARE_PMU is: 0xDD00AA AA: original hardware event ID DD: PMU type ID PMU type ID is retrieved from sysfs. For example, cat /sys/devices/cpu_atom/type 10 cat /sys/devices/cpu_core/type 4 When enabling a hybrid hardware event without specified pmu, such as, 'perf stat -e cycles -a', two events are created automatically. One is for atom, the other is for core. ok I think I understand the need for this (and the following) patch the perf_hw_id counters could be global, so when you specify only event like: -e cycles you want all the cycles, which on hybrid system means cycles from more than one pmus Yes, on hybrid system it means the cycles from two pmus. One cycle is from cpu_core pmu, another cycles is from cpu_atom pmu. SNIP @@ -1416,6 +1475,8 @@ int parse_events_add_numeric(struct parse_events_state *parse_state, { struct perf_event_attr attr; LIST_HEAD(config_terms); + bool hybrid; + int ret; memset(, 0, sizeof(attr)); attr.type = type; @@ -1430,6 +1491,18 @@ int parse_events_add_numeric(struct parse_events_state *parse_state, return -ENOMEM; } + /* +* Skip the software dummy event. +*/ + if (type != PERF_TYPE_SOFTWARE) { + if (!perf_pmu__hybrid_exist()) + perf_pmu__scan(NULL); this could be checked in the following add_hybrid_numeric call Yes, that should be OK. I will move the check in the next version. + + ret = add_hybrid_numeric(parse_state, list, , ); + if (hybrid) + return ret; + } could we add this to separate object.. hybrid.c or maybe parse-events-hybrid.c, there's already global __add_event wrapper - parse_events__add_event jirka Use a new parse-events-hybrid.c, hmm, well that's OK. Thanks Jin Yao + return add_event(list, _state->idx, , get_config_name(head_config), _terms); } -- 2.17.1
Re: [PATCH v2 11/27] perf parse-events: Support hardware events inside PMU
Hi Jiri, On 3/16/2021 1:37 AM, Jiri Olsa wrote: On Mon, Mar 15, 2021 at 10:28:12AM +0800, Jin, Yao wrote: Hi Jiri, On 3/13/2021 3:15 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:26PM +0800, Jin Yao wrote: On hybrid platform, some hardware events are only available on a specific pmu. For example, 'L1-dcache-load-misses' is only available on 'cpu_core' pmu. And even for the event which can be available on both pmus, the user also may want to just enable one event. So now following syntax is supported: cpu_core// cpu_core// cpu_core// cpu_atom// cpu_atom// cpu_atom// It limits the event to be enabled only on a specified pmu. The patch uses this idea, for example, if we use "cpu_core/LLC-loads/", in parse_events_add_pmu(), term->config is "LLC-loads". hum, I don't understand how this doest not work even now, I assume both cpu_core and cpu_atom have sysfs device directory with events/ directory right? Yes, we have cpu_core and cpu_atom directories with events. root@ssp-pwrt-002:/sys/devices/cpu_atom/events# ls branch-instructions bus-cyclescache-references instructions mem-stores topdown-bad-spec topdown-fe-bound branch-missescache-misses cpu-cyclesmem-loads ref-cycles topdown-be-bound topdown-retiring root@ssp-pwrt-002:/sys/devices/cpu_core/events# ls branch-instructions cache-misses instructions mem-stores topdown-bad-spec topdown-fe-bound topdown-mem-bound branch-missescache-references mem-loads ref-cycles topdown-be-bound topdown-fetch-lat topdown-retiring bus-cycles cpu-cyclesmem-loads-aux slots topdown-br-mispredict topdown-heavy-ops and whatever is defined in events we allow in parsing syntax.. why can't we treat them like 2 separated pmus? But if without this patch, it reports the error, root@ssp-pwrt-002:~# ./perf stat -e cpu_core/cycles/ -a -vv -- sleep 1 event syntax error: 'cpu_core/cycles/' \___ unknown term 'cycles' for pmu 'cpu_core' yep, because there's special care for 'cycles' unfortunately, but you should be able to run 'cpu_core/cpu-cycles/' right? Yes, cpu_core/cpu-cycles/ is OK. # ./perf stat -e cpu_core/cpu-cycles/ -a -- sleep 1 Performance counter stats for 'system wide': 12,831,980,326 cpu_core/cpu-cycles/ 1.003132639 seconds time elapsed valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/cycles/' \___ unknown term 'cycles' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Run 'perf list' for a list of valid events The 'cycles' is treated as a unknown term, then it errors out. yep, because it's not in events.. we could add special rule to treat cycles as cpu-cycles inside pmu definition ;-) jirka But not only the cycles, the branches has error too. # ./perf stat -e cpu_core/branches/ -a -- sleep 1 event syntax error: 'cpu_core/branches/' \___ unknown term 'branches' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/branches/' \___ unknown term 'branches' for pmu 'cpu_core' Of course, branch-instructions runs OK. # ./perf stat -e cpu_core/branch-instructions/ -a -- sleep 1 Performance counter stats for 'system wide': 136,655,302 cpu_core/branch-instructions/ 1.003171561 seconds time elapsed So we need special rules for both cycles and branches. The worse thing is, we also need to process the hardware cache events. # ./perf stat -e cpu_core/LLC-loads/ event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/LLC-loads/' \___ unknown term 'LLC-loads' for pmu 'cpu_core' If we use special rules for establishing all event mapping, that looks too much. :( Thanks Jin Yao
Re: [PATCH v2 04/27] perf pmu: Save pmu name
Hi Jiri, On 3/16/2021 7:03 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:19PM +0800, Jin Yao wrote: On hybrid platform, one event is available on one pmu (such as, available on cpu_core or on cpu_atom). This patch saves the pmu name to the pmu field of struct perf_pmu_alias. Then next we can know the pmu which the event can be available on. Signed-off-by: Jin Yao --- tools/perf/util/pmu.c | 10 +- tools/perf/util/pmu.h | 1 + 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 54e586bf19a5..45d8db1af8d2 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -283,6 +283,7 @@ void perf_pmu_free_alias(struct perf_pmu_alias *newalias) zfree(>str); zfree(>metric_expr); zfree(>metric_name); + zfree(>pmu); parse_events_terms__purge(>terms); free(newalias); } @@ -297,6 +298,10 @@ static bool perf_pmu_merge_alias(struct perf_pmu_alias *newalias, list_for_each_entry(a, alist, list) { if (!strcasecmp(newalias->name, a->name)) { + if (newalias->pmu && a->pmu && + !strcasecmp(newalias->pmu, a->pmu)) { + continue; + } perf_pmu_update_alias(a, newalias); perf_pmu_free_alias(newalias); return true; @@ -314,7 +319,8 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, int num; char newval[256]; char *long_desc = NULL, *topic = NULL, *unit = NULL, *perpkg = NULL, -*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL; +*metric_expr = NULL, *metric_name = NULL, *deprecated = NULL, +*pmu = NULL; if (pe) { long_desc = (char *)pe->long_desc; @@ -324,6 +330,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, metric_expr = (char *)pe->metric_expr; metric_name = (char *)pe->metric_name; deprecated = (char *)pe->deprecated; + pmu = (char *)pe->pmu; } alias = malloc(sizeof(*alias)); @@ -389,6 +396,7 @@ static int __perf_pmu__new_alias(struct list_head *list, char *dir, char *name, } alias->per_pkg = perpkg && sscanf(perpkg, "%d", ) == 1 && num == 1; alias->str = strdup(newval); + alias->pmu = pmu ? strdup(pmu) : NULL; if (deprecated) alias->deprecated = true; diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h index 8164388478c6..0e724d5b84c6 100644 --- a/tools/perf/util/pmu.h +++ b/tools/perf/util/pmu.h @@ -72,6 +72,7 @@ struct perf_pmu_alias { bool deprecated; char *metric_expr; char *metric_name; + char *pmu; please use pmu_name thanks, jirka OK, I will use pmu_name in next version. Thanks Jin Yao
Re: [PATCH v2 11/27] perf parse-events: Support hardware events inside PMU
Hi Jiri, On 3/13/2021 3:15 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:26PM +0800, Jin Yao wrote: On hybrid platform, some hardware events are only available on a specific pmu. For example, 'L1-dcache-load-misses' is only available on 'cpu_core' pmu. And even for the event which can be available on both pmus, the user also may want to just enable one event. So now following syntax is supported: cpu_core// cpu_core// cpu_core// cpu_atom// cpu_atom// cpu_atom// It limits the event to be enabled only on a specified pmu. The patch uses this idea, for example, if we use "cpu_core/LLC-loads/", in parse_events_add_pmu(), term->config is "LLC-loads". hum, I don't understand how this doest not work even now, I assume both cpu_core and cpu_atom have sysfs device directory with events/ directory right? Yes, we have cpu_core and cpu_atom directories with events. root@ssp-pwrt-002:/sys/devices/cpu_atom/events# ls branch-instructions bus-cyclescache-references instructions mem-stores topdown-bad-spec topdown-fe-bound branch-missescache-misses cpu-cyclesmem-loads ref-cycles topdown-be-bound topdown-retiring root@ssp-pwrt-002:/sys/devices/cpu_core/events# ls branch-instructions cache-misses instructions mem-stores topdown-bad-spec topdown-fe-bound topdown-mem-bound branch-missescache-references mem-loads ref-cycles topdown-be-bound topdown-fetch-lat topdown-retiring bus-cycles cpu-cyclesmem-loads-aux slots topdown-br-mispredict topdown-heavy-ops and whatever is defined in events we allow in parsing syntax.. why can't we treat them like 2 separated pmus? But if without this patch, it reports the error, root@ssp-pwrt-002:~# ./perf stat -e cpu_core/cycles/ -a -vv -- sleep 1 event syntax error: 'cpu_core/cycles/' \___ unknown term 'cycles' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Initial error: event syntax error: 'cpu_core/cycles/' \___ unknown term 'cycles' for pmu 'cpu_core' valid terms: event,pc,edge,offcore_rsp,ldlat,inv,umask,frontend,cmask,config,config1,config2,name,period,percore Run 'perf list' for a list of valid events The 'cycles' is treated as a unknown term, then it errors out. So we have to create another parser to scan the term. Thanks Jin Yao thanks, jirka We create a new "parse_events_state" with the pmu_name and use parse_events__scanner to scan the term->config (the string "LLC-loads" in this example). The parse_events_add_cache() will be called during parsing. The parse_state->pmu_name is used to identify the pmu where the event is enabled. Let's see examples: root@ssp-pwrt-002:~# ./perf stat -e cpu_core/cycles/,cpu_core/LLC-loads/ -vv -- ./triad_loop Control descriptor is not initialized perf_event_attr: type 6 size 120 config 0x4 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7267 cpu -1 group_fd -1 flags 0x8 = 3 perf_event_attr: type 7 size 120 config 0x40002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7267 cpu -1 group_fd -1 flags 0x8 = 4 cycles: 0: 449252097 29724 29724 LLC-loads: 0: 1857 29724 29724 cycles: 449252097 29724 29724 LLC-loads: 1857 29724 29724 Performance counter stats for './triad_loop': 449,252,097 cpu_core/cycles/ 1,857 cpu_core/LLC-loads/ 0.298898415 seconds time elapsed root@ssp-pwrt-002:~# ./perf stat -e cpu_atom/cycles/,cpu_atom/LLC-loads/ -vv -- taskset -c 16 ./triad_loop Control descriptor is not initialized perf_event_attr: type 6 size 120 config 0xa sample_type
Re: [PATCH v2 09/27] perf parse-events: Create two hybrid hardware events
Hi Jiri, On 3/13/2021 3:15 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:24PM +0800, Jin Yao wrote: SNIP cycles: 4: 800933425 1002536659 1002536659 cycles: 5: 800928573 1002528386 1002528386 cycles: 6: 800924347 1002520527 1002520527 cycles: 7: 800922009 1002513176 1002513176 cycles: 8: 800919624 1002507326 1002507326 cycles: 9: 800917204 1002500663 1002500663 cycles: 10: 802096579 1002494280 1002494280 cycles: 11: 802093770 1002486404 1002486404 cycles: 12: 803284338 1002479491 1002479491 cycles: 13: 803277609 1002469777 1002469777 cycles: 14: 800875902 1002458861 1002458861 cycles: 15: 800873241 1002451350 1002451350 cycles: 0: 800837379 1002444645 1002444645 cycles: 1: 800833400 1002438505 1002438505 cycles: 2: 800829291 1002433698 1002433698 cycles: 3: 800824390 1002427584 1002427584 cycles: 4: 800819360 1002422099 1002422099 cycles: 5: 800814787 1002415845 1002415845 cycles: 6: 800810125 1002410301 1002410301 cycles: 7: 800791893 1002386845 1002386845 cycles: 12855737722 16040169029 16040169029 cycles: 6406560625 8019379522 8019379522 Performance counter stats for 'system wide': 12,855,737,722 cpu_core/cycles/ 6,406,560,625 cpu_atom/cycles/ so we do that no_merge stuff for uncore pmus, why can't we do that in here? that'd seems like generic way jirka We have set the "stat_config.no_merge = true;" in "[PATCH v2 08/27] perf stat: Uniquify hybrid event name". For hybrid hardware events, they have different configs. The config is 0xDD00AA (0x4 for core vs. 0xa for atom in this example) We use perf_pmu__for_each_hybrid_pmu() to iterate all hybrid PMUs, generate the configs and create the evsels for each hybrid PMU. This logic and the code are not complex and easy to understand. Uncore looks complicated. It has uncore alias concept which is for different PMUs but with same prefix. Such as "uncore_cbox" for "uncore_cbox_0" to "uncore_cbox_9". But the uncore alias concept doesn't apply to hybrid pmu (we just have "cpu_core" and "cpu_atom" here). And actually I also don't want to mix the core stuff with uncore stuff, that would be hard for understanding. Perhaps I misunderstand, correct me if I'm wrong. Thanks Jin Yao
Re: [PATCH v2 07/27] perf evlist: Hybrid event uses its own cpus
Hi Jiri, On 3/13/2021 3:15 AM, Jiri Olsa wrote: On Thu, Mar 11, 2021 at 03:07:22PM +0800, Jin Yao wrote: On hybrid platform, atom events can be only enabled on atom CPUs. Core events can be only enabled on core CPUs. So for a hybrid event, it can be only enabled on it's own CPUs. But the problem for current perf is, the cpus for evsel (via PMU sysfs) have been merged to evsel_list->core.all_cpus. It might be all CPUs. So we need to figure out one way to let the hybrid event only use it's own CPUs. The idea is to create a new evlist__invalidate_all_cpus to invalidate the evsel_list->core.all_cpus then evlist__for_each_cpu returns cpu -1 for hybrid evsel. If cpu is -1, hybrid evsel will use it's own cpus. that's wild.. I don't understand when you say we don't have cpus for evsel, because they have been merged.. each evsel has evsel->core.own_cpus coming from pmu->cpus, right? why can't you just filter out cpus that are in there? jirka Yes, you're right. This patch is wide and actually it's not very necessary. The current framework has processed the cpus for evsel well even for hybrid evsel. So this patch can be dropped. Thanks Jin Yao
[PATCH v2 25/27] perf tests: Support 'Convert perf time to TSC' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the second evsel in evlist also needs initialization. With this patch, root@otcpl-adl-s-2:~# ./perf test 71 71: Convert perf time to TSC: Ok Signed-off-by: Jin Yao --- tools/perf/tests/perf-time-to-tsc.c | 16 1 file changed, 16 insertions(+) diff --git a/tools/perf/tests/perf-time-to-tsc.c b/tools/perf/tests/perf-time-to-tsc.c index 680c3cffb128..b472205ec8e3 100644 --- a/tools/perf/tests/perf-time-to-tsc.c +++ b/tools/perf/tests/perf-time-to-tsc.c @@ -66,6 +66,11 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe u64 test_tsc, comm1_tsc, comm2_tsc; u64 test_time, comm1_time = 0, comm2_time = 0; struct mmap *md; + bool hybrid = false; + + perf_pmu__scan(NULL); + if (perf_pmu__hybrid_exist()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); CHECK_NOT_NULL__(threads); @@ -88,6 +93,17 @@ int test__perf_time_to_tsc(struct test *test __maybe_unused, int subtest __maybe evsel->core.attr.disabled = 1; evsel->core.attr.enable_on_exec = 0; + /* +* For hybrid "cycles:u", it creates two events. +* Init the second evsel here. +*/ + if (hybrid) { + evsel = evsel__next(evsel); + evsel->core.attr.comm = 1; + evsel->core.attr.disabled = 1; + evsel->core.attr.enable_on_exec = 0; + } + CHECK__(evlist__open(evlist)); CHECK__(evlist__mmap(evlist, UINT_MAX)); -- 2.17.1
[PATCH v2 26/27] perf tests: Skip 'perf stat metrics (shadow stat) test' for hybrid
Currently we don't support shadow stat for hybrid. root@ssp-pwrt-002:~# ./perf stat -e cycles,instructions -a -- sleep 1 Performance counter stats for 'system wide': 12,883,109,591 cpu_core/cycles/ 6,405,163,221 cpu_atom/cycles/ 555,553,778 cpu_core/instructions/ 841,158,734 cpu_atom/instructions/ 1.002644773 seconds time elapsed Now there is no shadow stat 'insn per cycle' reported. We will support it later and now just skip the 'perf stat metrics (shadow stat) test'. Signed-off-by: Jin Yao --- tools/perf/tests/shell/stat+shadow_stat.sh | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/perf/tests/shell/stat+shadow_stat.sh b/tools/perf/tests/shell/stat+shadow_stat.sh index ebebd3596cf9..e6e35fc6c882 100755 --- a/tools/perf/tests/shell/stat+shadow_stat.sh +++ b/tools/perf/tests/shell/stat+shadow_stat.sh @@ -7,6 +7,9 @@ set -e # skip if system-wide mode is forbidden perf stat -a true > /dev/null 2>&1 || exit 2 +# skip if on hybrid platform +perf stat -a -e cycles sleep 1 2>&1 | grep -e cpu_core && exit 2 + test_global_aggr() { perf stat -a --no-big-num -e cycles,instructions sleep 1 2>&1 | \ -- 2.17.1
[PATCH v2 27/27] perf Documentation: Document intel-hybrid support
Add some words and examples to help understanding of Intel hybrid perf support. Signed-off-by: Jin Yao --- tools/perf/Documentation/intel-hybrid.txt | 228 ++ tools/perf/Documentation/perf-record.txt | 1 + tools/perf/Documentation/perf-stat.txt| 2 + 3 files changed, 231 insertions(+) create mode 100644 tools/perf/Documentation/intel-hybrid.txt diff --git a/tools/perf/Documentation/intel-hybrid.txt b/tools/perf/Documentation/intel-hybrid.txt new file mode 100644 index ..ff641d9ac81b --- /dev/null +++ b/tools/perf/Documentation/intel-hybrid.txt @@ -0,0 +1,228 @@ +Intel hybrid support + +Support for Intel hybrid events within perf tools. + +For some Intel platforms, such as AlderLake, which is hybrid platform and +it consists of atom cpu and core cpu. Each cpu has dedicated event list. +Part of events are available on core cpu, part of events are available +on atom cpu and even part of events are available on both. + +Kernel exports two new cpu pmus via sysfs: +/sys/devices/cpu_core +/sys/devices/cpu_atom + +The 'cpus' files are created under the directories. For example, + +cat /sys/devices/cpu_core/cpus +0-15 + +cat /sys/devices/cpu_atom/cpus +16-23 + +It indicates cpu0-cpu15 are core cpus and cpu16-cpu23 are atom cpus. + +Quickstart + +List hybrid event +- + +As before, use perf-list to list the symbolic event. + +perf list + +inst_retired.any + [Fixed Counter: Counts the number of instructions retired. Unit: cpu_atom] +inst_retired.any + [Number of instructions retired. Fixed Counter - architectural event. Unit: cpu_core] + +The 'Unit: xxx' is added to brief description to indicate which pmu +the event is belong to. Same event name but with different pmu can +be supported. + +Enable hybrid event with a specific pmu +--- + +To enable a core only event or atom only event, following syntax is supported: + + cpu_core// +or + cpu_atom// + +For example, count the 'cycles' event on core cpus. + + perf stat -e cpu_core/cycles/ + +Create two events for one hardware event automatically +-- + +When creating one event and the event is available on both atom and core, +two events are created automatically. One is for atom, the other is for +core. Most of hardware events and cache events are available on both +cpu_core and cpu_atom. + +For hardware events, they have pre-defined configs (e.g. 0 for cycles). +But on hybrid platform, kernel needs to know where the event comes from +(from atom or from core). The original perf event type PERF_TYPE_HARDWARE +can't carry pmu information. So a new type PERF_TYPE_HARDWARE_PMU is +introduced. + +The new attr.config layout for PERF_TYPE_HARDWARE_PMU: + +0xDD00AA +AA: original hardware event ID +DD: PMU type ID + +Cache event is similar. A new type PERF_TYPE_HW_CACHE_PMU is introduced. + +The new attr.config layout for PERF_TYPE_HW_CACHE_PMU: + +0xDD00CCBBAA +AA: original hardware cache ID +BB: original hardware cache op ID +CC: original hardware cache op result ID +DD: PMU type ID + +PMU type ID is retrieved from sysfs + +cat /sys/devices/cpu_atom/type +10 + +cat /sys/devices/cpu_core/type +4 + +When enabling a hardware event without specified pmu, such as, +perf stat -e cycles -a (use system-wide in this example), two events +are created automatically. + + +perf_event_attr: + type 6 + size 120 + config 0x4 + sample_type IDENTIFIER + read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING + disabled 1 + inherit 1 + exclude_guest1 + + +and + + +perf_event_attr: + type 6 + size 120 + config 0xa + sample_type IDENTIFIER + read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING + disabled 1 + inherit 1 + exclude_guest1 + + +type 6 is PERF_TYPE_HARDWARE_PMU. +0x4 in 0x4 indicates it's cpu_core pmu. +0xa in 0xa indicates it's cpu_atom pmu (atom pmu type id is random). + +The kernel creates 'cycles' (0x4) on cpu0-cpu15 (core cpus), +and create 'cycles' (0xa) on cpu16-cpu23 (atom cpus). + +For perf-stat result, it displays two events: + + Performance counter stats for 'system wide': + +12,869,720,529 cpu_core/cycles
[PATCH v2 19/27] perf tests: Add hybrid cases for 'Parse event definition strings' test
Add basic hybrid test cases for 'Parse event definition strings' test. root@otcpl-adl-s-2:~# ./perf test 6 6: Parse event definition strings : Ok Signed-off-by: Jin Yao --- tools/perf/tests/parse-events.c | 171 1 file changed, 171 insertions(+) diff --git a/tools/perf/tests/parse-events.c b/tools/perf/tests/parse-events.c index a7f6661e6112..aec929867020 100644 --- a/tools/perf/tests/parse-events.c +++ b/tools/perf/tests/parse-events.c @@ -1512,6 +1512,123 @@ static int test__all_tracepoints(struct evlist *evlist) return test__checkevent_tracepoint_multi(evlist); } +static int test__hybrid_hw_event_with_pmu(struct evlist *evlist) +{ + struct evsel *evsel = evlist__first(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 1 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel->core.attr.config); + return 0; +} + +static int test__hybrid_hw_event(struct evlist *evlist) +{ + struct evsel *evsel1 = evlist__first(evlist); + struct evsel *evsel2 = evlist__last(evlist); + + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel1->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel1->core.attr.config); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel2->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0xa == evsel2->core.attr.config); + return 0; +} + +static int test__hybrid_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x40001 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_sw_hw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_hw_sw_group_event(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_SOFTWARE == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + return 0; +} + +static int test__hybrid_group_modifier1(struct evlist *evlist) +{ + struct evsel *evsel, *leader; + + evsel = leader = evlist__first(evlist); + TEST_ASSERT_VAL("wrong number of entries", 2 == evlist->core.nr_entries); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_PMU == evsel->core.attr.type); + TEST_ASSERT_VAL("wrong config", 0x4 == evsel->core.attr.config); + TEST_ASSERT_VAL("wrong leader", evsel->leader == leader); + TEST_ASSERT_VAL("wrong exclude_user", evsel->core.attr.exclude_user); + TEST_ASSERT_VAL("wrong exclude_kernel", !evsel->core.attr.exclude_kernel); + + evsel = evsel__next(evsel); + TEST_ASSERT_VAL("wrong type", PERF_TYPE_HARDWARE_P
[PATCH v2 21/27] perf tests: Skip 'Setup struct perf_event_attr' test for hybrid
For hybrid, kernel introduces new perf type PERF_TYPE_HARDWARE_PMU (6) and it's assigned to hybrid hardware events. root@otcpl-adl-s-2:~# ./perf test 17 -vvv ... compare matching [event:base-stat] to [event-6-17179869184-4] [cpu] * 0 [flags] 0|8 8 [type] 0 6 ->FAIL match: [event:base-stat] matches [] event:base-stat does not match, but is optional matched compare matching [event-6-17179869184-4] to [event:base-stat] [cpu] 0 * [flags] 8 0|8 [type] 6 0 ->FAIL match: [event-6-17179869184-4] matches [] expected type=6, got 0 expected config=17179869184, got 0 FAILED './tests/attr/test-stat-C0' - match failure The type matching is failed because one type is 0 but the type of hybrid hardware event is 6. We temporarily skip this test case and TODO in future. Signed-off-by: Jin Yao --- tools/perf/tests/attr.c | 5 + 1 file changed, 5 insertions(+) diff --git a/tools/perf/tests/attr.c b/tools/perf/tests/attr.c index dd39ce9b0277..fc7f74159764 100644 --- a/tools/perf/tests/attr.c +++ b/tools/perf/tests/attr.c @@ -34,6 +34,7 @@ #include "event.h" #include "util.h" #include "tests.h" +#include "pmu-hybrid.h" #define ENV "PERF_TEST_ATTR" @@ -184,6 +185,10 @@ int test__attr(struct test *test __maybe_unused, int subtest __maybe_unused) char path_dir[PATH_MAX]; char *exec_path; + perf_pmu__scan(NULL); + if (perf_pmu__hybrid_exist()) + return 0; + /* First try development tree tests. */ if (!lstat("./tests", )) return run_dir("./tests", "./perf"); -- 2.17.1
[PATCH v2 24/27] perf tests: Support 'Session topology' test for hybrid
Force to create one event "cpu_core/cycles/" by default, otherwise in evlist__valid_sample_type, the checking of 'if (evlist->core.nr_entries == 1)' would be failed. root@otcpl-adl-s-2:~# ./perf test 41 41: Session topology: Ok Signed-off-by: Jin Yao --- tools/perf/tests/topology.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/topology.c b/tools/perf/tests/topology.c index 74748ed75b2c..0f6e73baab2d 100644 --- a/tools/perf/tests/topology.c +++ b/tools/perf/tests/topology.c @@ -40,7 +40,15 @@ static int session_write_header(char *path) session = perf_session__new(, false, NULL); TEST_ASSERT_VAL("can't get session", !IS_ERR(session)); - session->evlist = evlist__new_default(); + perf_pmu__scan(NULL); + if (!perf_pmu__hybrid_exist()) { + session->evlist = evlist__new_default(); + } else { + struct parse_events_error err; + + session->evlist = evlist__new(); + parse_events(session->evlist, "cpu_core/cycles/", ); + } TEST_ASSERT_VAL("can't get evlist", session->evlist); perf_header__set_feat(>header, HEADER_CPU_TOPOLOGY); -- 2.17.1
[PATCH v2 23/27] perf tests: Support 'Parse and process metrics' test for hybrid
Some events are not supported. Only pick up some cases for hybrid. root@otcpl-adl-s-2:~# ./perf test 67 67: Parse and process metrics : Ok Signed-off-by: Jin Yao --- tools/perf/tests/parse-metric.c | 11 --- 1 file changed, 8 insertions(+), 3 deletions(-) diff --git a/tools/perf/tests/parse-metric.c b/tools/perf/tests/parse-metric.c index 55bf52e588be..149b18f1f96a 100644 --- a/tools/perf/tests/parse-metric.c +++ b/tools/perf/tests/parse-metric.c @@ -370,12 +370,17 @@ static int test_metric_group(void) int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused) { + perf_pmu__scan(NULL); + TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); - TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); - TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0); - return 0; + + if (!perf_pmu__hybrid_exist()) { + TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); + TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); + } +return 0; } -- 2.17.1
[PATCH v2 20/27] perf tests: Add hybrid cases for 'Roundtrip evsel->name' test
Since for one hw event, two hybrid events are created. For example, evsel->idx evsel__name(evsel) 0 cycles 1 cycles 2 instructions 3 instructions ... So for comparing the evsel name on hybrid, the evsel->idx needs to be divided by 2. root@otcpl-adl-s-2:~# ./perf test 14 14: Roundtrip evsel->name : Ok Signed-off-by: Jin Yao --- tools/perf/tests/evsel-roundtrip-name.c | 19 --- 1 file changed, 12 insertions(+), 7 deletions(-) diff --git a/tools/perf/tests/evsel-roundtrip-name.c b/tools/perf/tests/evsel-roundtrip-name.c index f7f3e5b4c180..2b938a15901e 100644 --- a/tools/perf/tests/evsel-roundtrip-name.c +++ b/tools/perf/tests/evsel-roundtrip-name.c @@ -62,7 +62,8 @@ static int perf_evsel__roundtrip_cache_name_test(void) return ret; } -static int __perf_evsel__name_array_test(const char *names[], int nr_names) +static int __perf_evsel__name_array_test(const char *names[], int nr_names, +int distance) { int i, err; struct evsel *evsel; @@ -82,9 +83,9 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) err = 0; evlist__for_each_entry(evlist, evsel) { - if (strcmp(evsel__name(evsel), names[evsel->idx])) { + if (strcmp(evsel__name(evsel), names[evsel->idx / distance])) { --err; - pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx]); + pr_debug("%s != %s\n", evsel__name(evsel), names[evsel->idx / distance]); } } @@ -93,18 +94,22 @@ static int __perf_evsel__name_array_test(const char *names[], int nr_names) return err; } -#define perf_evsel__name_array_test(names) \ - __perf_evsel__name_array_test(names, ARRAY_SIZE(names)) +#define perf_evsel__name_array_test(names, distance) \ + __perf_evsel__name_array_test(names, ARRAY_SIZE(names), distance) int test__perf_evsel__roundtrip_name_test(struct test *test __maybe_unused, int subtest __maybe_unused) { int err = 0, ret = 0; - err = perf_evsel__name_array_test(evsel__hw_names); + perf_pmu__scan(NULL); + if (perf_pmu__hybrid_exist()) + return perf_evsel__name_array_test(evsel__hw_names, 2); + + err = perf_evsel__name_array_test(evsel__hw_names, 1); if (err) ret = err; - err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1); + err = __perf_evsel__name_array_test(evsel__sw_names, PERF_COUNT_SW_DUMMY + 1, 1); if (err) ret = err; -- 2.17.1
[PATCH v2 22/27] perf tests: Support 'Track with sched_switch' test for hybrid
Since for "cycles:u' on hybrid platform, it creates two "cycles". So the number of events in evlist is not expected in next test steps. Now we just use one event "cpu_core/cycles:u/" for hybrid. root@otcpl-adl-s-2:~# ./perf test 35 35: Track with sched_switch : Ok Signed-off-by: Jin Yao --- tools/perf/tests/switch-tracking.c | 10 +- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tools/perf/tests/switch-tracking.c b/tools/perf/tests/switch-tracking.c index 3ebaa758df77..13a11ce51a1a 100644 --- a/tools/perf/tests/switch-tracking.c +++ b/tools/perf/tests/switch-tracking.c @@ -340,6 +340,11 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ struct evsel *switch_evsel, *tracking_evsel; const char *comm; int err = -1; + bool hybrid = false; + + perf_pmu__scan(NULL); + if (perf_pmu__hybrid_exist()) + hybrid = true; threads = thread_map__new(-1, getpid(), UINT_MAX); if (!threads) { @@ -371,7 +376,10 @@ int test__switch_tracking(struct test *test __maybe_unused, int subtest __maybe_ cpu_clocks_evsel = evlist__last(evlist); /* Second event */ - err = parse_events(evlist, "cycles:u", NULL); + if (!hybrid) + err = parse_events(evlist, "cycles:u", NULL); + else + err = parse_events(evlist, "cpu_core/cycles:u/", NULL); if (err) { pr_debug("Failed to parse event cycles:u\n"); goto out_err; -- 2.17.1
[PATCH v2 16/27] perf evlist: Warn as events from different hybrid PMUs in a group
If a group has events which are from different hybrid PMUs, shows a warning. This is to remind the user not to put the core event and atom event into one group. root@ssp-pwrt-002:~# ./perf stat -e "{cpu_core/cycles/,cpu_atom/cycles/}" -- sleep 1 WARNING: Group has events from different hybrid PMUs Performance counter stats for 'sleep 1': cpu_core/cycles/ cpu_atom/cycles/ 1.002585908 seconds time elapsed Signed-off-by: Jin Yao --- tools/perf/builtin-record.c | 3 +++ tools/perf/builtin-stat.c | 7 ++ tools/perf/util/evlist.c| 44 + tools/perf/util/evlist.h| 2 ++ 4 files changed, 56 insertions(+) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 363ea1047148..188a1198cd4b 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -929,6 +929,9 @@ static int record__open(struct record *rec) pos = evlist__reset_weak_group(evlist, pos, true); goto try_again; } + + if (errno == EINVAL && perf_pmu__hybrid_exist()) + evlist__warn_hybrid_group(evlist); rc = -errno; evsel__open_strerror(pos, >target, errno, msg, sizeof(msg)); ui__error("%s\n", msg); diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 7a732508b2b4..6f780a039db0 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -239,6 +239,9 @@ static void evlist__check_cpu_maps(struct evlist *evlist) struct evsel *evsel, *pos, *leader; char buf[1024]; + if (evlist__hybrid_exist(evlist)) + return; + evlist__for_each_entry(evlist, evsel) { leader = evsel->leader; @@ -726,6 +729,10 @@ enum counter_recovery { static enum counter_recovery stat_handle_error(struct evsel *counter) { char msg[BUFSIZ]; + + if (perf_pmu__hybrid_exist() && errno == EINVAL) + evlist__warn_hybrid_group(evsel_list); + /* * PPC returns ENXIO for HW counters until 2.6.37 * (behavior changed with commit b0a873e). diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c index f139151b9433..5ec891418cdd 100644 --- a/tools/perf/util/evlist.c +++ b/tools/perf/util/evlist.c @@ -2224,3 +2224,47 @@ void evlist__invalidate_all_cpus(struct evlist *evlist) perf_cpu_map__put(evlist->core.all_cpus); evlist->core.all_cpus = perf_cpu_map__empty_new(1); } + +static bool group_hybrid_conflict(struct evsel *leader) +{ + struct evsel *pos, *prev = NULL; + + for_each_group_evsel(pos, leader) { + if (!pos->pmu_name || !perf_pmu__is_hybrid(pos->pmu_name)) + continue; + + if (prev && strcmp(prev->pmu_name, pos->pmu_name)) + return true; + + prev = pos; + } + + return false; +} + +void evlist__warn_hybrid_group(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel__is_group_leader(evsel) && + evsel->core.nr_members > 1 && + group_hybrid_conflict(evsel)) { + WARN_ONCE(1, "WARNING: Group has events from " +"different hybrid PMUs\n"); + return; + } + } +} + +bool evlist__hybrid_exist(struct evlist *evlist) +{ + struct evsel *evsel; + + evlist__for_each_entry(evlist, evsel) { + if (evsel__is_hybrid_event(evsel)) + return true; + } + + return false; +} diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h index 0da683511d98..33dec3bb5739 100644 --- a/tools/perf/util/evlist.h +++ b/tools/perf/util/evlist.h @@ -369,4 +369,6 @@ struct evsel *evlist__find_evsel(struct evlist *evlist, int idx); void evlist__invalidate_all_cpus(struct evlist *evlist); bool evlist__has_hybrid_events(struct evlist *evlist); +void evlist__warn_hybrid_group(struct evlist *evlist); +bool evlist__hybrid_exist(struct evlist *evlist); #endif /* __PERF_EVLIST_H */ -- 2.17.1
[PATCH v2 17/27] perf evsel: Adjust hybrid event and global event mixed group
: 6: 800568167 1002104339 1002104339 cycles: 7: 800566760 1002102953 1002102953 WARNING: for cpu-clock, some CPU counts not read cpu-clock: 0: 0 0 0 cpu-clock: 1: 0 0 0 cpu-clock: 2: 0 0 0 cpu-clock: 3: 0 0 0 cpu-clock: 4: 0 0 0 cpu-clock: 5: 0 0 0 cpu-clock: 6: 0 0 0 cpu-clock: 7: 0 0 0 cpu-clock: 8: 0 0 0 cpu-clock: 9: 0 0 0 cpu-clock: 10: 0 0 0 cpu-clock: 11: 0 0 0 cpu-clock: 12: 0 0 0 cpu-clock: 13: 0 0 0 cpu-clock: 14: 0 0 0 cpu-clock: 15: 0 0 0 cpu-clock: 16: 1002125111 1002124999 1002124999 cpu-clock: 17: 1002118626 1002118442 1002118442 cpu-clock: 18: 1002115058 1002114853 1002114853 cpu-clock: 19: 1002111740 1002111730 1002111730 cpu-clock: 20: 1002109031 1002108582 1002108582 cpu-clock: 21: 1002105927 1002106441 1002106441 cpu-clock: 22: 1002104010 1002104339 1002104339 cpu-clock: 23: 1002102730 1002102953 1002102953 cycles: 6446120799 8016892339 8016892339 cpu-clock: 8016892233 8016892339 8016892339 Performance counter stats for 'system wide': 6,446,120,799 cpu_atom/cycles/ # 804.067 M/sec 8,016.89 msec cpu-clock #7.999 CPUs utilized 1.002212870 seconds time elapsed For cpu-clock, cpu16-cpu23 are set with valid group fd (cpu_atom/cycles/'s fd on that cpu). For counting results, cpu_atom/cycles/ has 8 cpus aggregation , that's correct. But for cpu-clock, it also has 8 cpus aggregation (cpu16-cpu23, but not all cpus), the code should be improved. Now one warning is displayed: "WARNING: for cpu-clock, some CPU counts not read". Signed-off-by: Jin Yao --- tools/perf/util/evsel.c | 105 ++-- tools/perf/util/stat.h | 1 + 2 files changed, 101 insertions(+), 5 deletions(-) diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c index e0b6227d263f..862fdc145f05 100644 --- a/tools/perf/util/evsel.c +++ b/tools/perf/util/evsel.c @@ -1464,15 +1464,26 @@ static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val perf_counts__set_loaded(counter->counts, cpu, thread, true); } -static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data) +static int evsel_cpuid_match(struct evsel *evsel1, struct evsel *evsel2, +int cpu) +{ + int cpuid; + + cpuid = perf_cpu_map__cpu(evsel1->core.cpus, cpu); + return perf_cpu_map__idx(evsel2->core.cpus, cpuid); +} + +static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, +u64 *data, int nr_members) { u64 read_format = leader->core.attr.read_format; struct sample_read_value *v; u64 nr, ena = 0, run = 0, i; + int idx; nr = *data++; - if (nr != (u64) leader->core.nr_members) + if (nr != (u64) nr_members) return -EINVAL; if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) @@ -1492,24 +1503,85 @@ static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, if (!counter) return -EINVAL; - evsel__set_count(counter, cpu, thread, v[i].value, ena, run); + if (evsel__is_hybrid_event(counter) || + evsel__is_hybrid_event(leader)) { + idx = evsel_cpuid_match(leader, counter, cpu); + if (idx == -1) + return -EINVAL; + } else + idx = cpu; + + evsel__set_count(counter, idx, thread, v[i].value, ena, run); } return 0; } +static int hybrid_read_size(struct evsel *leader, int cpu, int *nr_members) +{ + struct evsel *pos; + int nr = 1, back, new_size = 0, idx; + + for_each_group_member(pos, leader) { + idx = evsel_cpuid_match(leader, pos, cpu); + if (idx != -1) + nr++; + } + + if (nr != leader->core.nr_members) { + back = leader->core.nr_members; + leader->core.nr_members = nr; + new_size = perf_evsel__read_size(>core); + leader->core.nr_members = back; + } + + *nr_members = nr; + return new_size; +} + static int evsel__read_group(struct evsel *leader, int cpu, int thread) { struct perf_stat_evsel *ps = leader->stats; u64 read_format = leader->core.attr.read_format; int size = perf_evsel__read_size(>core); + int new_size, nr_members; u64 *data = ps->group_data; if (!(read_format & PERF_FORMAT_ID)) return -EINVAL; - if (!evsel__is_group_leader(leader)) + if (!evsel__is_group_leader(leader)) { + if (evsel__is_hybrid_event(leader->leader) && + !evsel__is_hybrid_event(leader)) { +
[PATCH v2 18/27] perf script: Support PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU
For a hybrid system, the perf subsystem doesn't know which PMU the events belong to. So the PMU aware version PERF_TYPE_HARDWARE_PMU and PERF_TYPE_HW_CACHE_PMU are introduced. Now define the new output[] entries for these two types. Signed-off-by: Jin Yao --- tools/perf/builtin-script.c | 24 1 file changed, 24 insertions(+) diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c index 5915f19cee55..d0e889e636d5 100644 --- a/tools/perf/builtin-script.c +++ b/tools/perf/builtin-script.c @@ -275,6 +275,30 @@ static struct { .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, }, + [PERF_TYPE_HARDWARE_PMU] = { + .user_set = false, + + .fields = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | + PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | + PERF_OUTPUT_EVNAME | PERF_OUTPUT_IP | + PERF_OUTPUT_SYM | PERF_OUTPUT_SYMOFFSET | + PERF_OUTPUT_DSO | PERF_OUTPUT_PERIOD, + + .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, + }, + + [PERF_TYPE_HW_CACHE_PMU] = { + .user_set = false, + + .fields = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | + PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | + PERF_OUTPUT_EVNAME | PERF_OUTPUT_IP | + PERF_OUTPUT_SYM | PERF_OUTPUT_SYMOFFSET | + PERF_OUTPUT_DSO | PERF_OUTPUT_PERIOD, + + .invalid_fields = PERF_OUTPUT_TRACE | PERF_OUTPUT_BPF_OUTPUT, + }, + [OUTPUT_TYPE_SYNTH] = { .user_set = false, -- 2.17.1
[PATCH v2 15/27] perf stat: Filter out unmatched aggregation for hybrid event
perf-stat has supported some aggregation modes, such as --per-core, --per-socket and etc. While for hybrid event, it may only available on part of cpus. So for --per-core, we need to filter out the unavailable cores, for --per-socket, filter out the unavailable sockets, and so on. Before: root@ssp-pwrt-002:~# ./perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2 1,604,426,524 cpu_core/cycles/ S0-D0-C4 2 1,604,408,224 cpu_core/cycles/ S0-D0-C8 2 1,605,995,644 cpu_core/cycles/ S0-D0-C12 2 1,628,056,554 cpu_core/cycles/ S0-D0-C16 2 1,611,488,734 cpu_core/cycles/ S0-D0-C20 2 1,616,314,761 cpu_core/cycles/ S0-D0-C24 2 1,603,558,295 cpu_core/cycles/ S0-D0-C28 2 1,603,541,128 cpu_core/cycles/ S0-D0-C32 0cpu_core/cycles/ S0-D0-C33 0cpu_core/cycles/ S0-D0-C34 0cpu_core/cycles/ S0-D0-C35 0cpu_core/cycles/ S0-D0-C36 0cpu_core/cycles/ S0-D0-C37 0cpu_core/cycles/ S0-D0-C38 0cpu_core/cycles/ S0-D0-C39 0cpu_core/cycles/ After: root@ssp-pwrt-002:~# ./perf stat --per-core -e cpu_core/cycles/ -a -- sleep 1 Performance counter stats for 'system wide': S0-D0-C0 2 1,621,781,943 cpu_core/cycles/ S0-D0-C4 2 1,621,755,088 cpu_core/cycles/ S0-D0-C8 2 1,604,276,920 cpu_core/cycles/ S0-D0-C12 2 1,603,446,963 cpu_core/cycles/ S0-D0-C16 2 1,604,231,725 cpu_core/cycles/ S0-D0-C20 2 1,603,435,286 cpu_core/cycles/ S0-D0-C24 2 1,603,387,250 cpu_core/cycles/ S0-D0-C28 2 1,604,173,183 cpu_core/cycles/ Signed-off-by: Jin Yao --- tools/perf/util/stat-display.c | 20 1 file changed, 20 insertions(+) diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c index ed37d8e7ea1a..2db7c36a03ad 100644 --- a/tools/perf/util/stat-display.c +++ b/tools/perf/util/stat-display.c @@ -634,6 +634,20 @@ static void aggr_cb(struct perf_stat_config *config, } } +static bool aggr_id_hybrid_matched(struct perf_stat_config *config, + struct evsel *counter, struct aggr_cpu_id id) +{ + struct aggr_cpu_id s; + + for (int i = 0; i < evsel__nr_cpus(counter); i++) { + s = config->aggr_get_id(config, evsel__cpus(counter), i); + if (cpu_map__compare_aggr_cpu_id(s, id)) + return true; + } + + return false; +} + static void print_counter_aggrdata(struct perf_stat_config *config, struct evsel *counter, int s, char *prefix, bool metric_only, @@ -647,6 +661,12 @@ static void print_counter_aggrdata(struct perf_stat_config *config, double uval; ad.id = id = config->aggr_map->map[s]; + + if (perf_pmu__hybrid_exist() && + !aggr_id_hybrid_matched(config, counter, id)) { + return; + } + ad.val = ad.ena = ad.run = 0; ad.nr = 0; if (!collect_data(config, counter, aggr_cb, )) -- 2.17.1
[PATCH v2 14/27] perf stat: Add default hybrid events
Previously if '-e' is not specified in perf stat, some software events and hardware events are added to evlist by default. root@otcpl-adl-s-2:~# ./perf stat -- ./triad_loop Performance counter stats for './triad_loop': 109.43 msec task-clock#0.993 CPUs utilized 1 context-switches #0.009 K/sec 0 cpu-migrations#0.000 K/sec 105 page-faults #0.960 K/sec 401,161,982 cycles#3.666 GHz 1,601,216,357 instructions #3.99 insn per cycle 200,217,751 branches # 1829.686 M/sec 14,555 branch-misses #0.01% of all branches 0.110176860 seconds time elapsed Among the events, cycles, instructions, branches and branch-misses are hardware events. One hybrid platform, two events are created for one hardware event. core cycles, atom cycles, core instructions, atom instructions, core branches, atom branches, core branch-misses, atom branch-misses These events will be added to evlist in order on hybrid platform if '-e' is not set. Since parse_events() has been supported to create two hardware events for one event on hybrid platform, so we just use parse_events(evlist, "cycles,instructions,branches,branch-misses") to create the default events and add them to evlist. After: root@ssp-pwrt-002:~# ./perf stat -- ./triad_loop Performance counter stats for './triad_loop': 290.77 msec task-clock#0.996 CPUs utilized 25 context-switches #0.086 K/sec 13 cpu-migrations#0.045 K/sec 107 page-faults #0.368 K/sec 449,620,957 cpu_core/cycles/ # 1546.334 M/sec cpu_atom/cycles/ (0.00%) 1,601,499,820 cpu_core/instructions/# 5507.870 M/sec cpu_atom/instructions/ (0.00%) 200,272,310 cpu_core/branches/# 688.776 M/sec cpu_atom/branches/ (0.00%) 15,255 cpu_core/branch-misses/ #0.052 M/sec cpu_atom/branch-misses/ (0.00%) 0.291897676 seconds time elapsed We can see two events are created for one hardware event. First one is core event the second one is atom event. One thing is, the shadow stats looks a bit different, now it's just 'M/sec'. The perf_stat__update_shadow_stats and perf_stat__print_shadow_stats need to be improved in future if we want to get the original shadow stats. Signed-off-by: Jin Yao --- tools/perf/builtin-stat.c | 22 ++ 1 file changed, 22 insertions(+) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 6c0a21323814..7a732508b2b4 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1162,6 +1162,13 @@ static int parse_stat_cgroups(const struct option *opt, return parse_cgroups(opt, str, unset); } +static int add_default_hybrid_events(struct evlist *evlist) +{ + struct parse_events_error err; + + return parse_events(evlist, "cycles,instructions,branches,branch-misses", ); +} + static struct option stat_options[] = { OPT_BOOLEAN('T', "transaction", _run, "hardware transaction statistics"), @@ -1637,6 +1644,12 @@ static int add_default_attributes(void) { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, +}; + struct perf_event_attr default_sw_attrs[] = { + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, + { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, }; /* @@ -1874,6 +1887,15 @@ static int add_default_attributes(void) } if (!evsel_list->core.nr_entries) { + perf_pmu__scan(NULL); + if (perf_pmu__hybrid_exist()) { + if (evlist__add_default_attrs(evsel_list, + default_sw_attrs) < 0) { + return -1; + } + return add_default_hybrid_events(evsel_list); + } + if (target__has_cpu()) default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; -- 2.17.1
[PATCH v2 13/27] perf evlist: Create two hybrid 'cycles' events by default
When evlist is empty, for example no '-e' specified in perf record, one default 'cycles' event is added to evlist. While on hybrid platform, it needs to create two default 'cycles' events. One is for core, the other is for atom. This patch actually calls evsel__new_cycles() two times to create two 'cycles' events. root@ssp-pwrt-002:~# ./perf record -vv -- sleep 1 ... perf_event_attr: type 6 size 120 config 0x4 { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|PERIOD read_format ID disabled 1 inherit 1 mmap 1 comm 1 freq 1 enable_on_exec 1 task 1 precise_ip 3 sample_id_all1 exclude_guest1 mmap21 comm_exec1 ksymbol 1 bpf_event1 sys_perf_event_open: pid 22300 cpu 0 group_fd -1 flags 0x8 = 5 sys_perf_event_open: pid 22300 cpu 1 group_fd -1 flags 0x8 = 6 sys_perf_event_open: pid 22300 cpu 2 group_fd -1 flags 0x8 = 7 sys_perf_event_open: pid 22300 cpu 3 group_fd -1 flags 0x8 = 9 sys_perf_event_open: pid 22300 cpu 4 group_fd -1 flags 0x8 = 10 sys_perf_event_open: pid 22300 cpu 5 group_fd -1 flags 0x8 = 11 sys_perf_event_open: pid 22300 cpu 6 group_fd -1 flags 0x8 = 12 sys_perf_event_open: pid 22300 cpu 7 group_fd -1 flags 0x8 = 13 sys_perf_event_open: pid 22300 cpu 8 group_fd -1 flags 0x8 = 14 sys_perf_event_open: pid 22300 cpu 9 group_fd -1 flags 0x8 = 15 sys_perf_event_open: pid 22300 cpu 10 group_fd -1 flags 0x8 = 16 sys_perf_event_open: pid 22300 cpu 11 group_fd -1 flags 0x8 = 17 sys_perf_event_open: pid 22300 cpu 12 group_fd -1 flags 0x8 = 18 sys_perf_event_open: pid 22300 cpu 13 group_fd -1 flags 0x8 = 19 sys_perf_event_open: pid 22300 cpu 14 group_fd -1 flags 0x8 = 20 sys_perf_event_open: pid 22300 cpu 15 group_fd -1 flags 0x8 = 21 perf_event_attr: type 6 size 120 config 0xa { sample_period, sample_freq } 4000 sample_type IP|TID|TIME|ID|PERIOD read_format ID disabled 1 inherit 1 freq 1 enable_on_exec 1 precise_ip 3 sample_id_all1 exclude_guest1 sys_perf_event_open: pid 22300 cpu 16 group_fd -1 flags 0x8 = 22 sys_perf_event_open: pid 22300 cpu 17 group_fd -1 flags 0x8 = 23 sys_perf_event_open: pid 22300 cpu 18 group_fd -1 flags 0x8 = 24 sys_perf_event_open: pid 22300 cpu 19 group_fd -1 flags 0x8 = 25 sys_perf_event_open: pid 22300 cpu 20 group_fd -1 flags 0x8 = 26 sys_perf_event_open: pid 22300 cpu 21 group_fd -1 flags 0x8 = 27 sys_perf_event_open: pid 22300 cpu 22 group_fd -1 flags 0x8 = 28 sys_perf_event_open: pid 22300 cpu 23 group_fd -1 flags 0x8 = 29 ... We can see one core 'cycles' (0x4) is enabled on cpu0-cpu15 and atom 'cycles' (0xa) is enabled on cpu16-cpu23. Signed-off-by: Jin Yao --- tools/perf/builtin-record.c | 10 ++ tools/perf/util/evlist.c| 32 +++- tools/perf/util/evsel.c | 6 +++--- tools/perf/util/evsel.h | 2 +- 4 files changed, 41 insertions(+), 9 deletions(-) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 35465d1db6dd..363ea1047148 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -2786,10 +2786,12 @@ int cmd_record(int argc, const char **argv) if (record.opts.overwrite) record.opts.tail_synthesize = true; - if (rec->evlist->core.nr_entries == 0 && - __evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) { - pr_err("Not enough memory for event selector list\n"); - goto out; + if (rec->evlist->core.nr_entries == 0) { + perf_pmu__scan(NULL); + if (__evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) { + pr_err("Not enough memory for event selector list\n"); + got
[PATCH v2 11/27] perf parse-events: Support hardware events inside PMU
On hybrid platform, some hardware events are only available on a specific pmu. For example, 'L1-dcache-load-misses' is only available on 'cpu_core' pmu. And even for the event which can be available on both pmus, the user also may want to just enable one event. So now following syntax is supported: cpu_core// cpu_core// cpu_core// cpu_atom// cpu_atom// cpu_atom// It limits the event to be enabled only on a specified pmu. The patch uses this idea, for example, if we use "cpu_core/LLC-loads/", in parse_events_add_pmu(), term->config is "LLC-loads". We create a new "parse_events_state" with the pmu_name and use parse_events__scanner to scan the term->config (the string "LLC-loads" in this example). The parse_events_add_cache() will be called during parsing. The parse_state->pmu_name is used to identify the pmu where the event is enabled. Let's see examples: root@ssp-pwrt-002:~# ./perf stat -e cpu_core/cycles/,cpu_core/LLC-loads/ -vv -- ./triad_loop Control descriptor is not initialized perf_event_attr: type 6 size 120 config 0x4 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7267 cpu -1 group_fd -1 flags 0x8 = 3 perf_event_attr: type 7 size 120 config 0x40002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7267 cpu -1 group_fd -1 flags 0x8 = 4 cycles: 0: 449252097 29724 29724 LLC-loads: 0: 1857 29724 29724 cycles: 449252097 29724 29724 LLC-loads: 1857 29724 29724 Performance counter stats for './triad_loop': 449,252,097 cpu_core/cycles/ 1,857 cpu_core/LLC-loads/ 0.298898415 seconds time elapsed root@ssp-pwrt-002:~# ./perf stat -e cpu_atom/cycles/,cpu_atom/LLC-loads/ -vv -- taskset -c 16 ./triad_loop Control descriptor is not initialized perf_event_attr: type 6 size 120 config 0xa sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7339 cpu -1 group_fd -1 flags 0x8 = 3 perf_event_attr: type 7 size 120 config 0xa0002 sample_type IDENTIFIER read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING disabled 1 inherit 1 enable_on_exec 1 exclude_guest1 sys_perf_event_open: pid 7339 cpu -1 group_fd -1 flags 0x8 = 4 cycles: 0: 602020010 343657939 342553275 LLC-loads: 0: 3537 343657939 342553275 cycles: 603961400 343657939 342553275 LLC-loads: 3548 343657939 342553275 Performance counter stats for 'taskset -c 16 ./triad_loop': 603,961,400 cpu_atom/cycles/ (99.68%) 3,548 cpu_atom/LLC-loads/ (99.68%) 0.344904585 seconds time elapsed Signed-off-by: Jin Yao --- tools/perf/util/parse-events.c | 100 +++-- tools/perf/util/parse-events.h | 6 +- tools/perf/util/parse-events.y | 21 ++- 3 files changed, 105 insertions(+), 22 deletions(-) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 09e42245f71a..30435adc7a7b 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -489,7 +489,8 @@ static int create_hybrid_cache_eve
[PATCH v2 12/27] perf parse-events: Support hybrid raw events
293915211 293915211 cpu_core/r3c/: 449000613 293915211 293915211 Performance counter stats for './triad_loop': 449,000,613 cpu_core/r3c/ 0.294859229 seconds time elapsed Signed-off-by: Jin Yao --- tools/perf/util/parse-events.c | 56 +- 1 file changed, 55 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 30435adc7a7b..9b2a33103a57 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -1532,6 +1532,55 @@ static int add_hybrid_numeric(struct parse_events_state *parse_state, return 0; } +static int create_hybrid_raw_event(struct parse_events_state *parse_state, + struct list_head *list, + struct perf_event_attr *attr, + struct list_head *head_config, + struct list_head *config_terms, + struct perf_pmu *pmu) +{ + struct evsel *evsel; + + attr->type = pmu->type; + evsel = __add_event(list, _state->idx, attr, true, + get_config_name(head_config), + pmu, config_terms, false, NULL); + if (evsel) + evsel->pmu_name = strdup(pmu->name); + else + return -ENOMEM; + + return 0; +} + +static int add_hybrid_raw(struct parse_events_state *parse_state, + struct list_head *list, + struct perf_event_attr *attr, + struct list_head *head_config, + struct list_head *config_terms, + bool *hybrid) +{ + struct perf_pmu *pmu; + int ret; + + *hybrid = false; + perf_pmu__for_each_hybrid_pmu(pmu) { + *hybrid = true; + if (parse_state->pmu_name && + strcmp(parse_state->pmu_name, pmu->name)) { + continue; + } + + ret = create_hybrid_raw_event(parse_state, list, attr, + head_config, config_terms, + pmu); + if (ret) + return ret; + } + + return 0; +} + int parse_events_add_numeric(struct parse_events_state *parse_state, struct list_head *list, u32 type, u64 config, @@ -1558,7 +1607,12 @@ int parse_events_add_numeric(struct parse_events_state *parse_state, /* * Skip the software dummy event. */ - if (type != PERF_TYPE_SOFTWARE) { + if (type == PERF_TYPE_RAW) { + ret = add_hybrid_raw(parse_state, list, , head_config, +_terms, ); + if (hybrid) + return ret; + } else if (type != PERF_TYPE_SOFTWARE) { if (!perf_pmu__hybrid_exist()) perf_pmu__scan(NULL); -- 2.17.1