Event parsing fixes
Ian Rogers
reviewerv3 cover Event parsing fixes
For the default events, parse from strings rather than use pre-cooked perf_event_attr. This fixes non-x86 heterogeneous CPUs where legacy hardware events wouldn't be opened for all PMUs. v2 was previously just patch 3 but it is extended in v3 to incorporate related fixes.
When a sysfs/json event is used in preference to a legacy event, allow evsel__match to still function using a saved hardware config number. This fixes hard coded metrics in stat-shadow for events like "instructions" on Intel that have a sysfs file.
Fix/improve uniquifying event names fixing the test "102: perf stat metrics (shadow stat) test:" that was broken by a formatting issue when the sysfs instructions event was used.
Having evsel->pmu_name and evsel->pmu->name is confusing, get rid of the former. Fix/improve evsel__sys_has_perf_metrics in the process.
Ian Rogers (5):
perf evsel: Add alternate_hw_config and use in evsel__match perf stat: Uniquify event name improvements
perf stat: Remove evlist__add_default_attrs use strings perf evsel x86: Make evsel__has_perf_metrics work for legacy events perf evsel: Remove pmu_name
tools/perf/arch/x86/util/evlist.c | 74 +------- tools/perf/arch/x86/util/evsel.c | 35 +++- tools/perf/builtin-diff.c | 6 +- tools/perf/builtin-stat.c | 291 ++++++++++++------------------ tools/perf/tests/parse-events.c | 2 +- tools/perf/util/evlist.c | 46 +---- tools/perf/util/evlist.h | 12 -- tools/perf/util/evsel.c | 28 ++- tools/perf/util/evsel.h | 22 +-- tools/perf/util/metricgroup.c | 4 +- tools/perf/util/parse-events.c | 52 ++++-- tools/perf/util/parse-events.h | 6 + tools/perf/util/pmu.c | 6 +- tools/perf/util/pmu.h | 2 +- tools/perf/util/stat-display.c | 101 ++++++++--- tools/perf/util/stat-shadow.c | 14 +- tools/perf/util/stat.c | 2 +- 17 files changed, 305 insertions(+), 398 deletions(-)
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerPATCH v3 1/5 perf evsel: Add alternate_hw_config and use in evsel__match
There are cases where we want to match events like instructions and cycles with legacy hardware values, in particular in stat-shadow's hard coded metrics. An evsel's name isn't a good point of reference as it gets altered, strstr would be too imprecise and re-parsing the event from its name is silly. Instead, hold the legacy hardware event name, determined during parsing, in the evsel for this matching case.
Inline evsel__match2 that is only used in builtin-diff.
--- tools/perf/builtin-diff.c | 6 ++-- tools/perf/util/evsel.c | 21 ++++++++++++++ tools/perf/util/evsel.h | 19 ++----------- tools/perf/util/parse-events.c | 51 +++++++++++++++++++++------------- tools/perf/util/parse-events.h | 6 ++++ tools/perf/util/pmu.c | 6 +++- tools/perf/util/pmu.h | 2 +- 7 files changed, 70 insertions(+), 41 deletions(-)
tools/perf/builtin-diff.c +3 -3
@@ -483,13 +483,13 @@ 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__match2 (evsel, e)) +if ((evsel->core.attr.type == e->core.attr.type) && + (evsel->core.attr.config == e->core.attr.config))return e; }
tools/perf/util/evsel.c +21 -0
@@ -295,6 +295,7 @@ void evsel__init(struct evsel *evsel, evsel->pmu_name =NULL ; evsel->group_pmu_name =NULL ; evsel->skippable = false; + evsel->alternate_hw_config =PERF_COUNT_HW_MAX ; }struct evsel *evsel__new_idx (struct perf_event_attr *attr,int idx)
@@ -441,6 +442,8 @@ struct evsel *evsel__clone(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:
@@ -1600,6 +1603,24 @@ 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) {u64 read_format = evsel->core.attr.read_format;
tools/perf/util/evsel.h +2 -17
@@ -101,6 +101,7 @@ struct evsel {int bpf_fd;struct bpf_object *bpf_obj;struct list_head config_terms; +u64 alternate_hw_config; };/*
@@ -354,26 +355,10 @@ 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); -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; -} +bool __evsel__match (const struct evsel *evsel,u32 type,u64 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 +32 -19
166 lines not on this page · raw message
tools/perf/util/parse-events.h +6 -0
@@ -127,6 +127,12 @@ 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 {
tools/perf/util/pmu.c +5 -1
@@ -1547,7 +1547,7 @@ 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, -struct parse_events_error *err) +u64 *alternate_hw_config,struct parse_events_error *err) {struct parse_events_term *term, *h;struct perf_pmu_alias *alias;
@@ -1575,6 +1575,7 @@ 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)
@@ -1583,6 +1584,9 @@ 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
@@ -211,7 +211,7 @@ __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, -struct parse_events_error *err); +u64 *alternate_hw_config,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);
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerPATCH v3 2/5 perf stat: Uniquify event name improvements
In reply to Ian RogersWithout aggregation on Intel:
``` $ perf stat -e instructions,cycles ... ```
Will use "cycles" for the name of the legacy cycles event but as "instructions" has a sysfs name it will and a "[cpu]" PMU suffix. This often breaks things as the space between the event and the PMU name look like an extra column. The existing uniquify logic was also uniquifying in cases when all events are core and not with uncore events, it was not correctly handling modifiers, etc.
Change the logic so that an initial pass that can disable uniquification is run. For individual counters, disable uniquification in more cases such as for consistency with legacy events or for libpfm4 events. Don't use the "[pmu]" style suffix in uniquification, always use "pmu/.../". Change how modifiers/terms are handled in the uniquification so that they look like parse-able events.
This fixes "102: perf stat metrics (shadow stat) test:" that has been failing due to "instructions [cpu]" breaking its column/awk logic when values aren't aggregated. This started happening when instructions could match a sysfs rather than a legacy event, so the fixes tag reflects this.
--- tools/perf/util/stat-display.c | 101 +++++++++++++++++++++++++-------- 1 file changed, 78 insertions(+), 23 deletions(-)
tools/perf/util/stat-display.c +78 -23
@@ -868,38 +868,66 @@ static void printout(struct perf_stat_config *config, struct outstate *os,static void uniquify_event_name (struct evsel *counter) { -char *new_name; -char *config; -int ret =0 ; +const char *name, *pmu_name; +char *new_name, *config; +int ret; -if (counter->uniquified_name || counter->use_config_name || - !counter->pmu_name || !strncmp (evsel__name (counter), counter->pmu_name, -strlen (counter->pmu_name))) +/* The evsel was already uniquified. */ +if (counter->uniquified_name)return ; - config =strchr (counter->name,'/' ); +/* Avoid checking to uniquify twice. */ + counter->uniquified_name = true; + +/* The evsel has a "name=" config term or is from libpfm. */ +if (counter->use_config_name || counter->is_libpfm_event) +return ; + +/* Legacy no PMU event, don't uniquify. */ +if (!counter->pmu || + (counter->pmu->type <PERF_TYPE_MAX && counter->pmu->type !=PERF_TYPE_RAW )) +return ; + +/* A sysfs or json event replacing a legacy event, don't uniquify. */ +if (counter->pmu->is_core && counter->alternate_hw_config !=PERF_COUNT_HW_MAX ) +return ; + + name =evsel__name (counter); + pmu_name = counter->pmu->name; +/* Already prefixed by the PMU name. */ +if (!strncmp (name, pmu_name,strlen (pmu_name))) +return ; + + config =strchr (name,'/' );if (config) { -if (asprintf (&new_name, -"%s%s" , counter->pmu_name, config) >0 ) { -free (counter->name); - counter->name = new_name; - } - }else { -if (evsel__is_hybrid (counter)) { - ret =asprintf (&new_name,"%s/%s/" , - counter->pmu_name, counter->name); +int len = config - name; + +if (config[1 ] =='/' ) { +/* case: event// */ + ret =asprintf (&new_name,"%s/%.*s/%s" , pmu_name, len, name, config +2 ); }else { - ret =asprintf (&new_name,"%s [%s]" , - counter->name, counter->pmu_name); +/* case: event/.../ */ + ret =asprintf (&new_name,"%s/%.*s,%s" , pmu_name, len, name, config +1 ); } + }else { + config =strchr (name,':' ); +if (config) { +/* case: event:.. */ +int len = config - name; -if (ret) { -free (counter->name); - counter->name = new_name; + ret =asprintf (&new_name,"%s/%.*s/%s" , pmu_name, len, name, config +1 ); + }else { +/* case: event */ + ret =asprintf (&new_name,"%s/%s/" , pmu_name, name); } } - - counter->uniquified_name = true; +if (ret >0 ) { +free (counter->name); + counter->name = new_name; + }else { +/* ENOMEM from asprintf. */ + counter->uniquified_name = false; + } }static bool hybrid_uniquify (struct evsel *evsel,struct perf_stat_config *config)
@@ -1541,6 +1569,31 @@ static void print_cgroup_counter(struct perf_stat_config *config, struct evlistprint_metric_end (config, os); } +static void disable_uniquify (struct evlist *evlist) +{ +struct evsel *counter; +struct perf_pmu *last_pmu =NULL ; +bool first = true; + +evlist__for_each_entry (evlist, counter) { +/* If PMUs vary then uniquify can be useful. */ +if (!first && counter->pmu != last_pmu) +return ; + first = false; +if (counter->pmu) { +/* Allow uniquify for uncore PMUs. */ +if (!counter->pmu->is_core) +return ; +/* Keep hybrid event names uniquified for clarity. */ +if (perf_pmus__num_core_pmus () >1 ) +return ; + } + } +evlist__for_each_entry_continue (evlist, counter) { + counter->uniquified_name = true; + } +} +void evlist__print_counters (struct evlist *evlist,struct perf_stat_config *config,struct target *_target,struct timespec *ts,int argc,const char **argv)
@@ -1554,6 +1607,8 @@ void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *conf .first = true, }; +disable_uniquify (evlist); +if (config->iostat_run) evlist->selected =evlist__first (evlist);
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerPATCH v3 3/5 perf stat: Remove evlist__add_default_attrs use strings
In reply to Ian Rogersadd_default_atttributes would add evsels by having pre-created perf_event_attr, however, this needed fixing for hybrid as the extended PMU type was necessary for each core PMU. The logic for this was in an arch specific x86 function and wasn't present for ARM, meaning that default events weren't being opened on all PMUs on ARM. Change the creation of the default events to use parse_events and strings as that will open the events on all PMUs.
Rather than try to detect events on PMUs before parsing, parse the event but skip its output in stat-display.
The previous order of hardware events was: cycles, stalled-cycles-frontend, stalled-cycles-backend, instructions. As instructions is a more fundamental concept the order is changed to:
instructions, cycles, stalled-cycles-frontend, stalled-cycles-backend.
--- tools/perf/arch/x86/util/evlist.c | 74 +------- tools/perf/builtin-stat.c | 291 ++++++++++++------------------ tools/perf/util/evlist.c | 43 ----- tools/perf/util/evlist.h | 12 -- tools/perf/util/stat-shadow.c | 4 +- 5 files changed, 117 insertions(+), 307 deletions(-)
tools/perf/arch/x86/util/evlist.c +3 -71
@@ -1,78 +1,10 @@// SPDX-License-Identifier: GPL-2.0 -#include <stdio.h> -#include "util/pmu.h" -#include "util/pmus.h" -#include "util/evlist.h" -#include "util/parse-events.h" -#include "util/event.h" +#include <string.h> +#include "../../../util/evlist.h" +#include "../../../util/evsel.h" #include "topdown.h" #include "evsel.h" -static int ___evlist__add_default_attrs (struct evlist *evlist, -struct perf_event_attr *attrs, -size_t nr_attrs) -{ -LIST_HEAD (head); -size_t i =0 ; - -for (i =0 ; i < nr_attrs; i++) -event_attr_init (attrs + i); - -if (perf_pmus__num_core_pmus () ==1 ) -return evlist__add_attrs (evlist, attrs, nr_attrs); - -for (i =0 ; i < nr_attrs; i++) { -struct perf_pmu *pmu =NULL ; - -if (attrs[i].type ==PERF_TYPE_SOFTWARE ) { -struct evsel *evsel =evsel__new (attrs + i); - -if (evsel ==NULL ) - goto out_delete_partial_list; -list_add_tail (&evsel->core.node, &head); -continue ; - } - -while ((pmu =perf_pmus__scan_core (pmu)) !=NULL ) { -struct perf_cpu_map *cpus; -struct evsel *evsel; - - evsel =evsel__new (attrs + i); -if (evsel ==NULL ) - goto out_delete_partial_list; - evsel->core.attr.config |= (__u64 )pmu->type <<PERF_PMU_TYPE_SHIFT ; - cpus =perf_cpu_map__get (pmu->cpus); - evsel->core.cpus = cpus; - evsel->core.own_cpus =perf_cpu_map__get (cpus); - evsel->pmu_name =strdup (pmu->name); -list_add_tail (&evsel->core.node, &head); - } - } - -evlist__splice_list_tail (evlist, &head); - -return 0 ; - -out_delete_partial_list: - { -struct evsel *evsel, *n; - -__evlist__for_each_entry_safe (&head, n, evsel) -evsel__delete (evsel); - } -return -1 ; -} - -int arch_evlist__add_default_attrs (struct evlist *evlist, -struct perf_event_attr *attrs, -size_t nr_attrs) -{ -if (!nr_attrs) -return 0 ; - -return ___evlist__add_default_attrs (evlist, attrs, nr_attrs); -} -int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs) { if (topdown_sys_has_perf_metrics() &&
tools/perf/builtin-stat.c +112 -179
401 lines not on this page · raw message
tools/perf/util/evlist.c +0 -43
49 lines not on this page · raw message
tools/perf/util/evlist.h +0 -12
@@ -100,18 +100,6 @@ void evlist__delete(struct evlist *evlist);void evlist__add (struct evlist *evlist,struct evsel *entry);void evlist__remove (struct evlist *evlist,struct evsel *evsel); -int evlist__add_attrs (struct evlist *evlist,struct perf_event_attr *attrs,size_t nr_attrs); - -int __evlist__add_default_attrs (struct evlist *evlist, -struct perf_event_attr *attrs,size_t nr_attrs); - -int arch_evlist__add_default_attrs (struct evlist *evlist, -struct perf_event_attr *attrs, -size_t nr_attrs); - -#define evlist__add_default_attrs (evlist, array) \ - arch_evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array)) -int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs);int evlist__add_dummy (struct evlist *evlist);
tools/perf/util/stat-shadow.c +2 -2
@@ -76,7 +76,7 @@ void perf_stat__reset_shadow_stats(void)memset (&ru_stats,0 ,sizeof (ru_stats)); } -static enum stat_type evsel__stat_type (const struct evsel *evsel) +static enum stat_type evsel__stat_type (struct evsel *evsel) {/* Fake perf_hw_cache_op_id values for use with evsel__match. */ u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
@@ -152,7 +152,7 @@ static const char *get_ratio_color(const double ratios[3], double val)static double find_stat (const struct evsel *evsel,int aggr_idx,enum stat_type type) { -const struct evsel *cur; +struct evsel *cur;int evsel_ctx =evsel_context (evsel);evlist__for_each_entry (evsel->evlist, cur) {
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Re: [PATCH v3 3/5] perf stat: Remove evlist__add_default_attrs use strings
Quoted 11 lines · Ian Rogers
add_default_atttributeswould add evsels by having pre-createdperf_event_attr, however, this needed fixing for hybrid as the extended PMU type was necessary for each core PMU. The logic for this was in an arch specific x86 function and wasn't present for ARM, meaning that default events weren't being opened on all PMUs on ARM. Change the creation of the default events to useparse_eventsand strings as that will open the events on all PMUs.Rather than try to detect events on PMUs before parsing, parse the event but skip its output in stat-display.
Quoted 6 lines
The previous order of hardware events was: cycles, stalled-cycles-frontend, stalled-cycles-backend, instructions. As instructions is a more fundamental concept the order is changed to:
instructions, cycles, stalled-cycles-frontend, stalled-cycles-backend.
Taking a look at this one now. I think some example commands and outputs in the commit message would be helpful because there are quite a few different things mentioned in the closes link.
But I'm assuming this is just for the command without specifying an event:
$ perf stat
I didn't realise that wasn't working properly and I'd missed that Mark spotted it in that link.
Ian Rogers
reviewerRe: [PATCH v3 3/5] perf stat: Remove evlist__add_default_attrs use strings
Quoted 26 lines · James Clark
Quoted 11 lines · Ian Rogers
add_default_atttributeswould add evsels by having pre-createdperf_event_attr, however, this needed fixing for hybrid as the extended PMU type was necessary for each core PMU. The logic for this was in an arch specific x86 function and wasn't present for ARM, meaning that default events weren't being opened on all PMUs on ARM. Change the creation of the default events to useparse_eventsand strings as that will open the events on all PMUs.Rather than try to detect events on PMUs before parsing, parse the event but skip its output in stat-display.
Quoted 6 lines
The previous order of hardware events was: cycles, stalled-cycles-frontend, stalled-cycles-backend, instructions. As instructions is a more fundamental concept the order is changed to:
instructions, cycles, stalled-cycles-frontend, stalled-cycles-backend.Taking a look at this one now. I think some example commands and outputs in the commit message would be helpful because there are quite a few different things mentioned in the closes link.
Quoted 6 lines
But I'm assuming this is just for the command without specifying an event:
$ perf statI didn't realise that wasn't working properly and I'd missed that Mark spotted it in that link.
Hi James,
Do you want to take on owning fixing this? My concern is that by using event parsing:
+ "context-switches," + "cpu-migrations," + "page-faults," + "instructions," + "cycles," + "stalled-cycles-frontend," + "stalled-cycles-backend," + "branches," + "branch-misses",
any of the names could conflict with something advertised on an ARM PMU somewhere. Clearly cycles has already proved to be controversial and broken for perf record on Neoverse.
Thanks,
Ian
Ian Rogers
reviewerRe: [PATCH v3 3/5] perf stat: Remove evlist__add_default_attrs use strings
Quoted 50 lines · Ian Rogers
Quoted 26 lines · James Clark
Quoted 11 lines · Ian Rogers
add_default_atttributeswould add evsels by having pre-createdperf_event_attr, however, this needed fixing for hybrid as the extended PMU type was necessary for each core PMU. The logic for this was in an arch specific x86 function and wasn't present for ARM, meaning that default events weren't being opened on all PMUs on ARM. Change the creation of the default events to useparse_eventsand strings as that will open the events on all PMUs.Rather than try to detect events on PMUs before parsing, parse the event but skip its output in stat-display.
Quoted 6 lines
The previous order of hardware events was: cycles, stalled-cycles-frontend, stalled-cycles-backend, instructions. As instructions is a more fundamental concept the order is changed to:
instructions, cycles, stalled-cycles-frontend, stalled-cycles-backend.Taking a look at this one now. I think some example commands and outputs in the commit message would be helpful because there are quite a few different things mentioned in the closes link.
Quoted 6 lines
But I'm assuming this is just for the command without specifying an event:
$ perf statI didn't realise that wasn't working properly and I'd missed that Mark spotted it in that link.
Hi James,
Do you want to take on owning fixing this? My concern is that by using event parsing:
tools/perf/builtin-stat.c +9
-0+"context-switches," +"cpu-migrations," +"page-faults," +"instructions," +"cycles," +"stalled-cycles-frontend," +"stalled-cycles-backend," +"branches," + "branch-misses",
Quoted 3 lines
any of the names could conflict with something advertised on an ARM PMU somewhere. Clearly cycles has already proved to be controversial and broken for perf record on Neoverse.
Oh and these events are also broken on ARM M? PMUs with the revert that's in v6.10, so I don't know how to go about fixing this.
Thanks,
Ian
Re: [PATCH v3 3/5] perf stat: Remove evlist__add_default_attrs use strings
Quoted 55 lines · Ian Rogers
Quoted 50 lines · Ian Rogers
Quoted 26 lines · James Clark
Quoted 11 lines · Ian Rogers
add_default_atttributeswould add evsels by having pre-createdperf_event_attr, however, this needed fixing for hybrid as the extended PMU type was necessary for each core PMU. The logic for this was in an arch specific x86 function and wasn't present for ARM, meaning that default events weren't being opened on all PMUs on ARM. Change the creation of the default events to useparse_eventsand strings as that will open the events on all PMUs.Rather than try to detect events on PMUs before parsing, parse the event but skip its output in stat-display.
Quoted 6 lines
The previous order of hardware events was: cycles, stalled-cycles-frontend, stalled-cycles-backend, instructions. As instructions is a more fundamental concept the order is changed to:
instructions, cycles, stalled-cycles-frontend, stalled-cycles-backend.Taking a look at this one now. I think some example commands and outputs in the commit message would be helpful because there are quite a few different things mentioned in the closes link.
Quoted 6 lines
But I'm assuming this is just for the command without specifying an event:
$ perf statI didn't realise that wasn't working properly and I'd missed that Mark spotted it in that link.
Hi James,
Do you want to take on owning fixing this? My concern is that by using event parsing:
tools/perf/builtin-stat.c +9
-0+"context-switches," +"cpu-migrations," +"page-faults," +"instructions," +"cycles," +"stalled-cycles-frontend," +"stalled-cycles-backend," +"branches," + "branch-misses",Quoted 3 lines
any of the names could conflict with something advertised on an ARM PMU somewhere. Clearly cycles has already proved to be controversial and broken for perf record on Neoverse.
Quoted 5 lines
Oh and these events are also broken on ARM M? PMUs with the revert that's in v6.10, so I don't know how to go about fixing this.
Thanks,
Ian
I'll take it yep. But the fix probably needs to take into context whatever else we decide with the other issue. Conflicting names might not actually be an issue if we work around that with the other change. But I'm still getting up to speed with the other one.
James
Ian Rogers
reviewerPATCH v3 4/5 perf evsel x86: Make evsel__has_perf_metrics work for legacy events
In reply to Ian RogersUse PMU interface to better detect core PMU for legacy events. Look for slots event on core PMU if it is appropriate for the event.
--- tools/perf/arch/x86/util/evsel.c | 31 ++++++++++++++++++++++++++----- 1 file changed, 26 insertions(+), 5 deletions(-)
tools/perf/arch/x86/util/evsel.c +26 -5
@@ -21,7 +21,8 @@ void arch_evsel__set_sample_weight(struct evsel *evsel)/* Check whether the evsel's PMU supports the perf metrics */ bool evsel__sys_has_perf_metrics (const struct evsel *evsel) { -const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; +struct perf_pmu *pmu; +u32 type = evsel->core.attr.type;/* * The PERF_TYPE_RAW type is the core PMU type, e.g., "cpu" PMU
@@ -31,11 +32,31 @@ bool evsel__sys_has_perf_metrics(const struct evsel *evsel)* Checking both the PERF_TYPE_RAW type and the slots event * should be good enough to detect the perf metrics feature. */ -if ((evsel->core.attr.type ==PERF_TYPE_RAW ) && -perf_pmus__have_event (pmu_name,"slots" )) -return true; +again: +switch (type) { +case PERF_TYPE_HARDWARE : +case PERF_TYPE_HW_CACHE : + type = evsel->core.attr.config >>PERF_PMU_TYPE_SHIFT ; +if (type) + goto again; +break ; +case PERF_TYPE_RAW : +break ; +default : +return false; + } + + pmu = evsel->pmu; +if (pmu == &perf_pmu__fake) + pmu =NULL ; -return false; +if (!pmu) { +while ((pmu =perf_pmus__scan_core (pmu)) !=NULL ) { +if (pmu->type ==PERF_TYPE_RAW ) +break ; + } + } +return pmu &&perf_pmu__have_event (pmu,"slots" ); }bool arch_evsel__must_be_in_group (const struct evsel *evsel)
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerPATCH v3 5/5 perf evsel: Remove pmu_name
In reply to Ian Rogers"evsel->pmu_name" is only ever assigned a strdup of "pmu->name", a strdup of "evsel->pmu_name" or NULL. As such, prefer to use "pmu->name" directly and even to directly compare PMUs than PMU names. For safety, add some additional NULL tests.
--- tools/perf/arch/x86/util/evsel.c | 4 ++-- tools/perf/tests/parse-events.c | 2 +- tools/perf/util/evlist.c | 3 ++- tools/perf/util/evsel.c | 7 ------- tools/perf/util/evsel.h | 3 +-- tools/perf/util/metricgroup.c | 4 ++-- tools/perf/util/parse-events.c | 1 - tools/perf/util/stat-shadow.c | 10 +++++----- tools/perf/util/stat.c | 2 +- 9 files changed, 14 insertions(+), 22 deletions(-)
tools/perf/arch/x86/util/evsel.c +2 -2
@@ -84,7 +84,7 @@ int arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)return scnprintf (bf, size,"%s" , event_name);return scnprintf (bf, size,"%s/%s/" , - evsel->pmu_name ? evsel->pmu_name :"cpu" , + evsel->pmu ? evsel->pmu->name :"cpu" , event_name); }
@@ -129,7 +129,7 @@ int arch_evsel__open_strerror(struct evsel *evsel, char *msg, size_t size)return 0 ;if (!evsel->core.attr.precise_ip && - !(evsel->pmu_name && !strncmp (evsel->pmu_name,"ibs" ,3 ))) + !(evsel->pmu && !strncmp (evsel->pmu->name,"ibs" ,3 )))return 0 ;/* More verbose IBS errors. */
tools/perf/tests/parse-events.c +1 -1
@@ -730,7 +730,7 @@ static int test__checkevent_pmu_events(struct evlist *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 || -strcmp (evsel->pmu_name,"cpu" )); +strcmp (evsel->pmu->name,"cpu" ));TEST_ASSERT_VAL ("wrong exclude_user" , !evsel->core.attr.exclude_user); TEST_ASSERT_VAL("wrong exclude_kernel",
tools/perf/util/evlist.c +2 -1
@@ -2504,7 +2504,8 @@ void evlist__uniquify_name(struct evlist *evlist)else attributes = empty_attributes; -if (asprintf (&new_name,"%s/%s/%s" , pos->pmu_name, pos->name, attributes +1 )) { +if (asprintf (&new_name,"%s/%s/%s" , pos->pmu ?"" : pos->pmu->name, + pos->name, attributes +1 )) {free (pos->name); pos->name = new_name; }else {
tools/perf/util/evsel.c +0 -7
@@ -292,7 +292,6 @@ void evsel__init(struct evsel *evsel, evsel->metric_events =NULL ; evsel->per_pkg_mask =NULL ; evsel->collect_stat = false; - evsel->pmu_name =NULL ; evsel->group_pmu_name =NULL ; evsel->skippable = false; evsel->alternate_hw_config =PERF_COUNT_HW_MAX ;
@@ -390,11 +389,6 @@ struct evsel *evsel__clone(struct evsel *orig)if (evsel->group_name ==NULL ) goto out_err; } -if (orig->pmu_name) { - evsel->pmu_name =strdup (orig->pmu_name); -if (evsel->pmu_name ==NULL ) - goto out_err; - }if (orig->group_pmu_name) { evsel->group_pmu_name =strdup (orig->group_pmu_name);if (evsel->group_pmu_name ==NULL )
@@ -1481,7 +1475,6 @@ void evsel__exit(struct evsel *evsel)zfree (&evsel->group_name);zfree (&evsel->name);zfree (&evsel->filter); -zfree (&evsel->pmu_name);zfree (&evsel->group_pmu_name);zfree (&evsel->unit);zfree (&evsel->metric_id);
tools/perf/util/evsel.h +1 -2
@@ -72,7 +72,6 @@ struct evsel {struct {char *name;char *group_name; -const char *pmu_name;const char *group_pmu_name;#ifdef HAVE_LIBTRACEEVENT struct tep_event *tp_format;
@@ -169,7 +168,7 @@ struct evsel {unsigned long open_flags;int precise_ip_original; -/* for missing_features */ +/* The PMU the event is from. Used for missing_features, PMU name, etc. */ struct perf_pmu *pmu; };
tools/perf/util/metricgroup.c +2 -2
@@ -297,8 +297,8 @@ static int setup_metric_events(const char *pmu, struct hashmap *ids,struct expr_id_data *val_ptr;/* Don't match events for the wrong hybrid PMU. */ -if (!all_pmus && ev->pmu_name &&evsel__is_hybrid (ev) && -strcmp (ev->pmu_name, pmu)) +if (!all_pmus && ev->pmu &&evsel__is_hybrid (ev) && +strcmp (ev->pmu->name, pmu))continue ;/* * Check for duplicate events with the same name. For
tools/perf/util/parse-events.c +0 -1
@@ -265,7 +265,6 @@ __add_event(struct list_head *list, int *idx, evsel->core.is_pmu_core = pmu ? pmu->is_core : false; evsel->auto_merge_stats = auto_merge_stats; evsel->pmu = pmu; - evsel->pmu_name = pmu ?strdup (pmu->name) :NULL ; evsel->alternate_hw_config = alternate_hw_config;if (name)
tools/perf/util/stat-shadow.c +5 -5
@@ -566,7 +566,7 @@ static void perf_stat__print_metricgroup_header(struct perf_stat_config *config, {bool need_full_name =perf_pmus__num_core_pmus () >1 ;static const char *last_name; -static const char *last_pmu; +static const struct perf_pmu *last_pmu;char full_name[64 ];/*
@@ -577,21 +577,21 @@ static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,* different metric events. */ if (last_name && !strcmp (last_name, name)) { -if (!need_full_name || !strcmp (last_pmu, evsel->pmu_name)) { +if (!need_full_name || last_pmu != evsel->pmu) { out->print_metricgroup_header (config, ctxp,NULL );return ; } } -if (need_full_name) -scnprintf (full_name,sizeof (full_name),"%s (%s)" , name, evsel->pmu_name); +if (need_full_name && evsel->pmu) +scnprintf (full_name,sizeof (full_name),"%s (%s)" , name, evsel->pmu->name);else scnprintf (full_name,sizeof (full_name),"%s" , name); out->print_metricgroup_header (config, ctxp, full_name); last_name = name; - last_pmu = evsel->pmu_name; + last_pmu = evsel->pmu; }/**
tools/perf/util/stat.c +1 -1
@@ -553,7 +553,7 @@ static bool evsel__is_alias(struct evsel *evsel_a, struct evsel *evsel_b)if (evsel__is_clock (evsel_a) !=evsel__is_clock (evsel_b))return false; -return !!strcmp (evsel_a->pmu_name, evsel_b->pmu_name); +return evsel_a->pmu != evsel_b->pmu; }static void evsel__merge_aliases (struct evsel *evsel)
Signature · 1 line
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerQuoted 18 lines · Ian Rogers
For the default events, parse from strings rather than use pre-cooked
perf_event_attr. This fixes non-x86 heterogeneous CPUs where legacy hardware events wouldn't be opened for all PMUs. v2 was previously just patch 3 but it is extended in v3 to incorporate related fixes.When a sysfs/json event is used in preference to a legacy event, allow
evsel__matchto still function using a saved hardware config number. This fixes hard coded metrics in stat-shadow for events like "instructions" on Intel that have a sysfs file.Fix/improve uniquifying event names fixing the test "102: perf stat metrics (shadow stat) test:" that was broken by a formatting issue when the sysfs instructions event was used.
Having
evsel->pmu_nameandevsel->pmu->nameis confusing, get rid of the former. Fix/improveevsel__sys_has_perf_metricsin the process.
Quoted 6 lines
Ian Rogers (5):
perf evsel: Addalternate_hw_configand use inevsel__matchperf stat: Uniquify event name improvements
perf stat: Removeevlist__add_default_attrsuse strings perf evsel x86: Makeevsel__has_perf_metricswork for legacy events perf evsel: Removepmu_name
Hopefully the first 3 patches here can be a priority given the fixes.
Thanks,
Ian
Quoted 22 lines
tools/perf/arch/x86/util/evlist.c | 74 +------- tools/perf/arch/x86/util/evsel.c | 35 +++- tools/perf/builtin-diff.c | 6 +- tools/perf/builtin-stat.c | 291 ++++++++++++------------------ tools/perf/tests/parse-events.c | 2 +- tools/perf/util/evlist.c | 46 +---- tools/perf/util/evlist.h | 12 -- tools/perf/util/evsel.c | 28 ++- tools/perf/util/evsel.h | 22 +-- tools/perf/util/metricgroup.c | 4 +- tools/perf/util/parse-events.c | 52 ++++-- tools/perf/util/parse-events.h | 6 + tools/perf/util/pmu.c | 6 +- tools/perf/util/pmu.h | 2 +- tools/perf/util/stat-display.c | 101 ++++++++--- tools/perf/util/stat-shadow.c | 14 +- tools/perf/util/stat.c | 2 +- 17 files changed, 305 insertions(+), 398 deletions(-)--
2.45.0.118.g7fe29c98d7-goog
Ian Rogers
reviewerQuoted 20 lines · Ian Rogers
Quoted 18 lines · Ian Rogers
For the default events, parse from strings rather than use pre-cooked
perf_event_attr. This fixes non-x86 heterogeneous CPUs where legacy hardware events wouldn't be opened for all PMUs. v2 was previously just patch 3 but it is extended in v3 to incorporate related fixes.When a sysfs/json event is used in preference to a legacy event, allow
evsel__matchto still function using a saved hardware config number. This fixes hard coded metrics in stat-shadow for events like "instructions" on Intel that have a sysfs file.Fix/improve uniquifying event names fixing the test "102: perf stat metrics (shadow stat) test:" that was broken by a formatting issue when the sysfs instructions event was used.
Having
evsel->pmu_nameandevsel->pmu->nameis confusing, get rid of the former. Fix/improveevsel__sys_has_perf_metricsin the process.
Quoted 8 lines
Quoted 6 lines
Ian Rogers (5):
perf evsel: Addalternate_hw_configand use inevsel__matchperf stat: Uniquify event name improvements
perf stat: Removeevlist__add_default_attrsuse strings perf evsel x86: Makeevsel__has_perf_metricswork for legacy events perf evsel: Removepmu_nameHopefully the first 3 patches here can be a priority given the fixes.
Testing:
https://lore.kernel.org/lkml/[email protected]/
git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools.git
tags/perf-tools-for-v6.10-1-2024-05-21
I see this failure:
``` 102: perf stat metrics (shadow stat) test: --- start --- test child forked, pid 785992 IPC is different: 1.31 != (2725291544 / 2085879817) ---- end(-1) ---- ```
Which is addressed in the first 3 patches here.
Thanks,
Ian
Quoted 25 lines
Thanks,
IanQuoted 22 lines
tools/perf/arch/x86/util/evlist.c | 74 +------- tools/perf/arch/x86/util/evsel.c | 35 +++- tools/perf/builtin-diff.c | 6 +- tools/perf/builtin-stat.c | 291 ++++++++++++------------------ tools/perf/tests/parse-events.c | 2 +- tools/perf/util/evlist.c | 46 +---- tools/perf/util/evlist.h | 12 -- tools/perf/util/evsel.c | 28 ++- tools/perf/util/evsel.h | 22 +-- tools/perf/util/metricgroup.c | 4 +- tools/perf/util/parse-events.c | 52 ++++-- tools/perf/util/parse-events.h | 6 + tools/perf/util/pmu.c | 6 +- tools/perf/util/pmu.h | 2 +- tools/perf/util/stat-display.c | 101 ++++++++--- tools/perf/util/stat-shadow.c | 14 +- tools/perf/util/stat.c | 2 +- 17 files changed, 305 insertions(+), 398 deletions(-)--
2.45.0.118.g7fe29c98d7-goog