Consistently prioritize legacy events
Subsystems (2)
Ian Rogers
reviewerv2 cover Consistently prioritize legacy events
Legacy events were historically the priority over sysfs/json events. This behavior was changed in commit a24d9d9dc096 ("perf
parse-events: Make legacy events lower priority than sysfs/JSON") to fix issues on ARM Apple-M CPUs. Sysfs/json being the priority makes most sense, was preferred by RISC-V and gives abilities impossible with legacy events. Making this encoding consistent is proving impossible with endless bike shedding, people saying they see something as a fix then when that fix is provided continuing to object, etc. Given that in the RFC [1] legacy events are seen as being preferrable let's just follow up on that and make their use consistent.
If it isn't clear, I think this is the wrong decision and landing [2] should be the priority. When I try to have this conversation it just gets ignored and I want consistency more than my preferred prioritization.
[1] https://lore.kernel.org/lkml/[email protected]/
[2] https://lore.kernel.org/linux-perf-users/[email protected]/
v2: Missed revert of stat-display logic for json event prioritization.
Ian Rogers (2):
perf parse-events: Make legacy events always have priority over
sysfs/json
Revert "perf evsel: Add alternate_hw_config and use in evsel__match"
tools/perf/builtin-diff.c | 6 +-- tools/perf/tests/parse-events.c | 13 ++---- tools/perf/util/evsel.c | 21 --------- tools/perf/util/evsel.h | 19 +++++++- tools/perf/util/parse-events.c | 81 ++++++++++++--------------------- tools/perf/util/parse-events.h | 8 +--- tools/perf/util/parse-events.y | 2 +- tools/perf/util/pmu.c | 6 +-- tools/perf/util/pmu.h | 2 +- tools/perf/util/stat-display.c | 6 --- 10 files changed, 57 insertions(+), 107 deletions(-)
Signature · 1 line
2.49.0.rc0.332.g42c0ae87b1-goog
Ian Rogers
reviewerPATCH v2 1/2 perf parse-events: Make legacy events always have priority over sysfs/json
This reverts the behavior of commit a24d9d9dc096 ("perf parse-events:
Make legacy events lower priority than sysfs/JSON"). That commit states:
As part of that report Mark Rutland <[email protected]> requested that legacy events not be higher in priority when a PMU is specified reversing what has until this change been perf's default behavior.
with an Acked-by: Mark Rutland <[email protected]> tag. Mark expresses this preference here:
https://lore.kernel.org/lkml/[email protected]/
On an Alderlake this commit caused cpu_core/instructions/ to be encoded as:
------------------------------------------------------------
perf_event_attr:
type 4 (cpu_core)
size 136
config 0xc0 (instructions)
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
enable_on_exec 1
exclude_guest 1------------------------------------------------------------
While instructions would be encoded as:
------------------------------------------------------------
perf_event_attr:
type 0 (PERF_TYPE_HARDWARE)
size 136
config 0x400000001
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
enable_on_exec 1
exclude_guest 1------------------------------------------------------------
As the event parsing for legacy events uses parse_events_add_numeric, rather than the more generic PMU handling case, we end up with the same event name being encoded two ways. However, as instructions is also a cpu_atom event on Alderlake, or any hybrid system, the PMU prefixes are added in the stat output meaning instructions is reported as cpu_core/instructions/ and cpu_atom/instructions/ even though those events are encoded differently and potentially have different behaviors. We shouldn't be having the same event with two different encodings and I followed up the commit with changing the prioritization of legacy events also when no PMU is specified in commit 617824a7f0f7 ("perf parse-events: Prefer sysfs/JSON hardware events over legacy").
RISC-V have been asking that legacy events also be a lower priority than sysfs/json to avoid driver complexity. They've since had to work around this issue meaning it isn't a necessity:
If the overriding legacy with JSON is available, each future vendor may just provide the json file instead of modifying the driver. However, it will be a matter of convenience and clutter free future rather than a necessity at this point.
https://lists.riscv.org/g/sig-perf-analysis/topic/110906276#msg458
Landing the commit making encoding consistent and having sysfs/json a priority has been pushed back against by Namhyung Kim, Arnaldo Carvalho de Melo and James Clark. James assert the ARM drivers are now fixed and don't need the prioritization. RISC-V's interests have been ignored. The fact that legacy events are antiquated, imprecise in meaning, brittle, misleading and don't give us a way to fix kernel issues in software, have also been ignored:
https://lore.kernel.org/lkml/Z8sMcta0zTWeOso4@x1/
Given the push back against sysfs/json being a priority, contrary to Mark Rutland's insistence this was the sensible way to do things, contrary to RISC-V's wishes, and contrary to my own preference that legacy events do become legacy, this patch changes the priority back to legacy events being the priority consistently - that is instructions and cpu_core/instructions/ are both encoded using legacy events. It isn't a straight revert as clean-ups/refactorings are kept in place.
--- tools/perf/tests/parse-events.c | 13 +++---------- tools/perf/util/parse-events.c | 26 +++++++++----------------- 2 files changed, 12 insertions(+), 27 deletions(-)
tools/perf/tests/parse-events.c +3 -10
@@ -162,16 +162,9 @@ static int test__checkevent_numeric(struct evlist *evlist)static int assert_hw (struct perf_evsel *evsel,enum perf_hw_id id,const char *name) { -struct perf_pmu *pmu; - -if (evsel->attr.type ==PERF_TYPE_HARDWARE ) { -TEST_ASSERT_VAL ("wrong config" ,test_perf_config (evsel, id)); -return 0 ; - } - pmu =perf_pmus__find_by_type (evsel->attr.type); - -TEST_ASSERT_VAL ("unexpected PMU type" , pmu); -TEST_ASSERT_VAL ("PMU missing event" ,perf_pmu__have_event (pmu, name)); +TEST_ASSERT_VAL ("wrong type" ,PERF_TYPE_HARDWARE == evsel->attr.type); +TEST_ASSERT_VAL ("wrong config" ,test_perf_config (evsel, id)); +TEST_ASSERT_VAL ("wrong name" , !strcmp (evsel__hw_names[id], name));return 0 ; }
tools/perf/util/parse-events.c +9 -17
@@ -1023,12 +1023,10 @@ static int config_term_pmu(struct perf_event_attr *attr,return -EINVAL ; }/* -* Rewrite the PMU event to a legacy cache one unless the PMU -* doesn't support legacy cache events or the event is present -* within the PMU. +* Rewrite the PMU event to a legacy cache one as legacy events +* take priority over sysfs/json. */ -if (perf_pmu__supports_legacy_cache (pmu) && - !perf_pmu__have_event (pmu, term->config)) { +if (perf_pmu__supports_legacy_cache (pmu)) { attr->type =PERF_TYPE_HW_CACHE ;return parse_events__decode_legacy_cache (term->config, pmu->type, &attr->config);
@@ -1049,19 +1047,13 @@ static int config_term_pmu(struct perf_event_attr *attr,return -EINVAL ; }/* -* If the PMU has a sysfs or json event prefer it over -* legacy. ARM requires this. +* Rewrite the PMU event to a legacy cache one as legacy events +* take priority over sysfs/json. */ -if (perf_pmu__have_event (pmu, term->config)) { - term->type_term =PARSE_EVENTS__TERM_TYPE_USER ; - term->no_value = true; - term->alternate_hw_config = true; - }else { - attr->type =PERF_TYPE_HARDWARE ; - attr->config = term->val.num; -if (perf_pmus__supports_extended_type ()) - attr->config |= (__u64 )pmu->type <<PERF_PMU_TYPE_SHIFT ; - } + attr->type =PERF_TYPE_HARDWARE ; + attr->config = term->val.num; +if (perf_pmus__supports_extended_type ()) + attr->config |= (__u64 )pmu->type <<PERF_PMU_TYPE_SHIFT ;return 0 ; } if (term->type_term == PARSE_EVENTS__TERM_TYPE_USER ||
Signature · 1 line
2.49.0.rc0.332.g42c0ae87b1-goog
Ian Rogers
reviewerPATCH v2 2/2 Revert "perf evsel: Add alternate_hw_config and use in evsel__match"
In reply to Ian RogersThis reverts commit 22a4db3c36034e2b034c5b88414680857fc59cf4.
Now legacy events are prioritized there is no need to worry about a sysfs/json event being used when trying to pattern match a legacy event.
--- tools/perf/builtin-diff.c | 6 ++-- tools/perf/util/evsel.c | 21 ------------- tools/perf/util/evsel.h | 19 ++++++++++-- tools/perf/util/parse-events.c | 55 +++++++++++++--------------------- tools/perf/util/parse-events.h | 8 +---- tools/perf/util/parse-events.y | 2 +- tools/perf/util/pmu.c | 6 +--- tools/perf/util/pmu.h | 2 +- tools/perf/util/stat-display.c | 6 ---- 9 files changed, 45 insertions(+), 80 deletions(-)
tools/perf/builtin-diff.c +3 -3
@@ -470,13 +470,13 @@ static int diff__process_sample_event(const struct perf_tool *tool,static struct perf_diff pdiff; -static struct evsel *evsel_match (struct evsel *evsel,struct evlist *evlist) +static struct evsel *evsel_match (struct evsel *evsel, +struct evlist *evlist) {struct evsel *e;evlist__for_each_entry (evlist, e) { -if ((evsel->core.attr.type == e->core.attr.type) && - (evsel->core.attr.config == e->core.attr.config)) +if (evsel__match2 (evsel, e))return e; }
tools/perf/util/evsel.c +0 -21
@@ -404,7 +404,6 @@ void evsel__init(struct evsel *evsel, evsel->collect_stat = false; evsel->group_pmu_name =NULL ; evsel->skippable = false; - evsel->alternate_hw_config =PERF_COUNT_HW_MAX ; evsel->script_output_type =-1 ;// FIXME: OUTPUT_TYPE_UNSET, see builtin-script.c }
@@ -551,8 +550,6 @@ struct evsel *evsel__clone(struct evsel *dest, struct evsel *orig)if (evsel__copy_config_terms (evsel, orig) <0 ) goto out_err; - evsel->alternate_hw_config = orig->alternate_hw_config; -return evsel; out_err:
@@ -1848,24 +1845,6 @@ static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)return evsel__process_group_data (leader, cpu_map_idx, thread, data); } -bool __evsel__match (const struct evsel *evsel,u32 type,u64 config) -{ - -u32 e_type = evsel->core.attr.type; -u64 e_config = evsel->core.attr.config; - -if (e_type != type) { -return type ==PERF_TYPE_HARDWARE && evsel->pmu && evsel->pmu->is_core && - evsel->alternate_hw_config == config; - } - -if ((type ==PERF_TYPE_HARDWARE || type ==PERF_TYPE_HW_CACHE ) && -perf_pmus__supports_extended_type ()) - e_config &=PERF_HW_EVENT_MASK ; - -return e_config == config; -} -int evsel__read_counter (struct evsel *evsel,int cpu_map_idx,int thread) {if (evsel__is_tool (evsel))
tools/perf/util/evsel.h +17 -2
@@ -87,7 +87,6 @@ struct evsel {int bpf_fd;struct bpf_object *bpf_obj;struct list_head config_terms; -u64 alternate_hw_config; };/*
@@ -370,10 +369,26 @@ u64 format_field__intval(struct tep_format_field *field, struct perf_sample *samstruct tep_format_field *evsel__field (struct evsel *evsel,const char *name);struct tep_format_field *evsel__common_field (struct evsel *evsel,const char *name); -bool __evsel__match (const struct evsel *evsel,u32 type,u64 config); +static inline bool __evsel__match (const struct evsel *evsel,u32 type,u64 config) +{ +if (evsel->core.attr.type != type) +return false; + +if ((type ==PERF_TYPE_HARDWARE || type ==PERF_TYPE_HW_CACHE ) && +perf_pmus__supports_extended_type ()) +return (evsel->core.attr.config &PERF_HW_EVENT_MASK ) == config; + +return evsel->core.attr.config == config; +}#define evsel__match (evsel, t, c) __evsel__match(evsel, PERF_TYPE_##t, PERF_COUNT_##c) +static inline bool evsel__match2 (struct evsel *e1,struct evsel *e2) +{ +return (e1->core.attr.type == e2->core.attr.type) && + (e1->core.attr.config == e2->core.attr.config); +} +int evsel__read_counter (struct evsel *evsel,int cpu_map_idx,int thread);int __evsel__read_on_cpu (struct evsel *evsel,int cpu_map_idx,int thread,bool scale);
tools/perf/util/parse-events.c +21 -34
175 lines not on this page · raw message
tools/perf/util/parse-events.h +1 -7
@@ -128,12 +128,6 @@ struct parse_events_term {* value is assumed to be 1. An event name also has no value. */ bool no_value; -/** -* @alternate_hw_config: config is the event name but num is an -* alternate PERF_TYPE_HARDWARE config value which is often nice for the -* sake of quick matching. -*/ -bool alternate_hw_config; };struct parse_events_error {
@@ -244,7 +238,7 @@ struct evsel *parse_events__add_event(int idx, struct perf_event_attr *attr, struct perf_pmu *pmu);int parse_events_multi_pmu_add (struct parse_events_state *parse_state, -const char *event_name,u64 hw_config, +const char *event_name,const struct parse_events_terms *const_parsed_terms,struct list_head **listp,void *loc);
tools/perf/util/parse-events.y +1 -1
@@ -290,7 +290,7 @@ PE_NAME sep_dc struct list_head *list; int err; - err = parse_events_multi_pmu_add(_parse_state, $1, PERF_COUNT_HW_MAX, NULL, &list, &@1); + err = parse_events_multi_pmu_add(_parse_state, $1, NULL, &list, &@1); if (err < 0) { struct parse_events_state *parse_state = _parse_state; struct parse_events_error *error = parse_state->error;
tools/perf/util/pmu.c +1 -5
@@ -1662,7 +1662,7 @@ static int check_info_data(struct perf_pmu *pmu,*/ int perf_pmu__check_alias (struct perf_pmu *pmu,struct parse_events_terms *head_terms,struct perf_pmu_info *info,bool *rewrote_terms, -u64 *alternate_hw_config,struct parse_events_error *err) +struct parse_events_error *err) {struct parse_events_term *term, *h;struct perf_pmu_alias *alias;
@@ -1699,7 +1699,6 @@ int perf_pmu__check_alias(struct perf_pmu *pmu, struct parse_events_terms *head_ NULL);return ret; } - *rewrote_terms = true; ret =check_info_data (pmu, alias, info, err, term->err_term);if (ret)
@@ -1708,9 +1707,6 @@ int perf_pmu__check_alias(struct perf_pmu *pmu, struct parse_events_terms *head_if (alias->per_pkg) info->per_pkg = true; -if (term->alternate_hw_config) - *alternate_hw_config = term->val.num; -list_del_init (&term->list);parse_events_term__delete (term); }
tools/perf/util/pmu.h +1 -1
@@ -231,7 +231,7 @@ __u64 perf_pmu__format_bits(struct perf_pmu *pmu, const char *name);int perf_pmu__format_type (struct perf_pmu *pmu,const char *name);int perf_pmu__check_alias (struct perf_pmu *pmu,struct parse_events_terms *head_terms,struct perf_pmu_info *info,bool *rewrote_terms, -u64 *alternate_hw_config,struct parse_events_error *err); +struct parse_events_error *err);int perf_pmu__find_event (struct perf_pmu *pmu,const char *event,void *state,pmu_event_callback cb);void perf_pmu_format__set_value (void *format,int config,unsigned long *bits);
tools/perf/util/stat-display.c +0 -6
@@ -1699,12 +1699,6 @@ static void evsel__set_needs_uniquify(struct evsel *counter, const struct perf_sreturn ; } -if (counter->pmu && counter->pmu->is_core && - counter->alternate_hw_config !=PERF_COUNT_HW_MAX ) { -/* A sysfs or json event replacing a legacy event, don't uniquify. */ -return ; - } -if (config->aggr_mode ==AGGR_NONE ) {/* Always unique with no aggregation. */ counter->needs_uniquify = true;
Signature · 1 line
2.49.0.rc0.332.g42c0ae87b1-goog