Perf stat metric grouping with hardware information
Subsystems (2)
v4 cover Perf stat metric grouping with hardware information
From: Weilin Wang <[email protected]>
Changes in v4:
- Update bitmap operation functions [Kan, Ian]
- Move data struct definitions to metricgroup.c for smaller scope [Ian]
- Update variable names and comments for better readbility [Ian]
- Exclude JSON files from this patch. A seperate patch set with complete set of JSON files will be submitted soon.
v3: https://lore.kernel.org/all/[email protected]/
Weilin Wang (15):
perf stat: Add new field in stat_config to enable hardware aware
grouping.
perf stat: Add basic functions for the hardware aware grouping perf pmu-events: Add functions in jevent.py to parse counter and event
info for hardware aware grouping
find_bit: add _find_last_and_bit() to support finding the most
significant set bit
perf stat: Add functions to set counter bitmaps for hardware-grouping
method
perf stat: Add functions to get counter info
perf stat: Add functions to create new group and assign events into
groups
perf stat: Add build string function and topdown events handling in
hardware-grouping
perf stat: Add function to handle special events in hardware-grouping perf stat: Add function to combine metrics for hardware-grouping perf stat: Handle taken alone in hardware-grouping
perf stat: Handle NMI in hardware-grouping
perf stat: Code refactoring in hardware-grouping
perf stat: Add tool events support in hardware-grouping perf stat: Add hardware-grouping cmd option to perf stat
tools/include/linux/find.h | 18 + tools/lib/find_bit.c | 33 + tools/perf/builtin-stat.c | 7 + tools/perf/pmu-events/jevents.py | 192 +++- tools/perf/pmu-events/pmu-events.h | 35 +- tools/perf/util/metricgroup.c | 1338 +++++++++++++++++++++++++--- tools/perf/util/metricgroup.h | 1 + tools/perf/util/stat.h | 1 + 8 files changed, 1464 insertions(+), 161 deletions(-)
--
2.42.0
RFC PATCH v4 1/15 perf stat: Add new field in stat_config to enable hardware aware grouping.
From: Weilin Wang <[email protected]>
Hardware counter and event information could be used to help creating event groups that better utilize hardware counters and improve multiplexing.
--- tools/perf/builtin-stat.c | 5 +++++ tools/perf/util/metricgroup.c | 5 +++++ tools/perf/util/metricgroup.h | 1 + tools/perf/util/stat.h | 1 + 4 files changed, 12 insertions(+)
tools/perf/builtin-stat.c +5 -0
@@ -2062,6 +2062,7 @@ static int add_default_attributes(void) stat_config.metric_no_threshold, stat_config.user_requested_cpu_list, stat_config.system_wide, + stat_config.hardware_aware_grouping, &stat_config.metric_events); }
@@ -2095,6 +2096,7 @@ static int add_default_attributes(void) stat_config.metric_no_threshold, stat_config.user_requested_cpu_list, stat_config.system_wide, + stat_config.hardware_aware_grouping, &stat_config.metric_events); }
@@ -2129,6 +2131,7 @@ static int add_default_attributes(void)/*metric_no_threshold=*/ true, stat_config.user_requested_cpu_list, stat_config.system_wide, + stat_config.hardware_aware_grouping, &stat_config.metric_events) < 0) return -1; }
@@ -2170,6 +2173,7 @@ static int add_default_attributes(void)/*metric_no_threshold=*/ true, stat_config.user_requested_cpu_list, stat_config.system_wide, + stat_config.hardware_aware_grouping, &stat_config.metric_events) < 0) return -1;
@@ -2702,6 +2706,7 @@ int cmd_stat(int argc, const char **argv) stat_config.metric_no_threshold, stat_config.user_requested_cpu_list, stat_config.system_wide, + stat_config.hardware_aware_grouping, &stat_config.metric_events);zfree (&metrics);
tools/perf/util/metricgroup.c +5 -0
@@ -1690,12 +1690,17 @@ int metricgroup__parse_groups(struct evlist *perf_evlist,bool metric_no_threshold, const char *user_requested_cpu_list, bool system_wide, + bool hardware_aware_grouping, struct rblist *metric_events) { const struct pmu_metrics_table *table = pmu_metrics_table__find();if (!table)return -EINVAL ; +if (hardware_aware_grouping) { +pr_debug ("Use hardware aware grouping instead of traditional metric grouping method\n" ); + } + return parse_groups(perf_evlist, pmu, str, metric_no_group, metric_no_merge, metric_no_threshold, user_requested_cpu_list, system_wide,
tools/perf/util/metricgroup.h +1 -0
@@ -77,6 +77,7 @@ int metricgroup__parse_groups(struct evlist *perf_evlist,bool metric_no_threshold, const char *user_requested_cpu_list, bool system_wide, + bool hardware_aware_grouping, struct rblist *metric_events); int metricgroup__parse_groups_test(struct evlist *evlist, const struct pmu_metrics_table *table,
tools/perf/util/stat.h +1 -0
@@ -86,6 +86,7 @@ struct perf_stat_config {bool metric_no_group;bool metric_no_merge;bool metric_no_threshold; +bool hardware_aware_grouping;bool stop_read_counter;bool iostat_run;char *user_requested_cpu_list;
Signature · 1 line
2.42.0
RFC PATCH v4 2/15 perf stat: Add basic functions for the hardware aware grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Add the first set of functions for the hardware aware grouping method. Function hw_aware_parse_groups() is the entry point of this metric grouping method. It does metric grouping on a combined list of events and will create a list of grouping strings as final results of the grouping method. These grouping strings will be used in the same manner as existing metric grouping process.
This method will fall back to normal grouping when hardware aware grouping return with err so that perf stat still executes and returns with correct result.
--- tools/perf/util/metricgroup.c | 218 +++++++++++++++++++++++++++++++++- 1 file changed, 217 insertions(+), 1 deletion(-)
tools/perf/util/metricgroup.c +217 -1
@@ -157,6 +157,14 @@ struct metric {struct evlist *evlist; }; +/** +* Each group is one node in the group string list. +*/ +struct metricgroup__group_strs { +struct list_head nd; +struct strbuf grouping_str; +}; +static void metric__watchdog_constraint_hint (const char *name,bool foot) {static bool violate_nmi_constraint;
@@ -1432,6 +1440,101 @@ static int build_combined_expr_ctx(const struct list_head *metric_list,return ret; } +/** +* hw_aware_build_grouping - Build event groupings by reading counter +* requirement of the events and counter available on the system from +* pmu-events. +* @ctx: the event identifiers parsed from metrics. +* @groupings: header to the list of final event grouping. +* @modifier: any modifiers added to the events. +*/ +static int hw_aware_build_grouping (struct expr_parse_ctx *ctx__maybe_unused , +struct list_head *groupings__maybe_unused , +const char *modifier__maybe_unused ) +{ +int ret =0 ; + +pr_debug ("This is a placeholder\n" ); +return ret; +} + +static void group_str_free (struct metricgroup__group_strs *g) +{ +if (!g) +return ; + +strbuf_release (&g->grouping_str); +free (g); +} + +static void metricgroup__free_grouping_strs (struct list_head + *grouping_strs) +{ +struct metricgroup__group_strs *g, *tmp; + +list_for_each_entry_safe (g, tmp, grouping_strs, nd) { +list_del_init (&g->nd); +group_str_free (g); + } +} + +/** +* hw_aware_parse_ids - Build the event string for the ids and parse them +* creating an evlist. The encoded metric_ids are decoded. Events are placed +* into groups based on event counter requirements and counter availabilities of +* the system. +* @metric_no_merge: is metric sharing explicitly disabled. +* @fake_pmu: used when testing metrics not supported by the current CPU. +* @ids: the event identifiers parsed from a metric. +* @modifier: any modifiers added to the events. +* @out_evlist: the created list of events. +*/ +static int hw_aware_parse_ids (struct perf_pmu *fake_pmu, +struct expr_parse_ctx *ids,const char *modifier, +struct evlist **out_evlist) +{ +struct parse_events_error parse_error; +struct evlist *parsed_evlist; +LIST_HEAD (groupings); +struct metricgroup__group_strs *group; +int ret; + + *out_evlist =NULL ; + ret =hw_aware_build_grouping (ids, &groupings, modifier); +if (ret) { +metricgroup__free_grouping_strs (&groupings); +return ret; + } + + parsed_evlist =evlist__new (); +if (!parsed_evlist) { + ret = -ENOMEM ; + goto err_out; + } +list_for_each_entry (group, &groupings, nd) { +struct strbuf *events = &group->grouping_str; + +pr_debug ("Parsing metric events '%s'\n" , events->buf); +parse_events_error__init (&parse_error); + ret =__parse_events (parsed_evlist, events->buf,/*pmu_filter=*/ NULL , + &parse_error, fake_pmu,/*warn_if_reordered=*/ false); +if (ret) { +parse_events_error__print (&parse_error, events->buf); + goto err_out; + } + ret =decode_all_metric_ids (parsed_evlist, modifier); +if (ret) + goto err_out; + } + *out_evlist = parsed_evlist; + parsed_evlist =NULL ; +err_out: +parse_events_error__exit (&parse_error); +evlist__delete (parsed_evlist); +metricgroup__free_grouping_strs (&groupings); +return ret; +} +/** * parse_ids - Build the event string for the ids and parse them creating an * evlist. The encoded metric_ids are decoded.
@@ -1520,6 +1623,114 @@ static int parse_ids(bool metric_no_merge, struct perf_pmu *fake_pmu,return ret; } +static int hw_aware_parse_groups (struct evlist *perf_evlist, +const char *pmu,const char *str, +bool metric_no_threshold, +const char *user_requested_cpu_list, +bool system_wide, +struct perf_pmu *fake_pmu, +struct rblist *metric_events_list, +const struct pmu_metrics_table *table) +{ +struct evlist *combined_evlist =NULL ; +LIST_HEAD (metric_list); +struct metric *m; +int ret; +bool metric_no_group = false; +bool metric_no_merge = false; + +if (metric_events_list->nr_entries ==0 ) +metricgroup__rblist_init (metric_events_list); + ret =metricgroup__add_metric_list (pmu, str, metric_no_group, metric_no_threshold, + user_requested_cpu_list, + system_wide, &metric_list, table); +if (ret) + goto out; + +/* Sort metrics from largest to smallest. */ +list_sort (NULL , &metric_list, metric_list_cmp); + +if (!metric_no_merge) { +struct expr_parse_ctx *combined =NULL ; + + ret =build_combined_expr_ctx (&metric_list, &combined); + +if (!ret && combined &&hashmap__size (combined->ids)) { + ret =hw_aware_parse_ids (fake_pmu, combined, +/*modifier=*/ NULL , + &combined_evlist); + } + +if (ret) + goto out; + +if (combined) +expr__ctx_free (combined); + } + +list_for_each_entry (m, &metric_list, nd) { +struct metric_expr *expr; +struct metric_event *me; +struct evsel **metric_events; + + ret =setup_metric_events (fake_pmu ?"all" : m->pmu, m->pctx->ids, + combined_evlist, &metric_events); +if (ret) { +pr_debug ("Cannot resolve IDs for %s: %s\n" , + m->metric_name, m->metric_expr); + goto out; + } + + me =metricgroup__lookup (metric_events_list, metric_events[0 ], true); + + expr =malloc (sizeof (struct metric_expr )); +if (!expr) { + ret = -ENOMEM ; +free (metric_events); + goto out; + } + + expr->metric_refs = m->metric_refs; + m->metric_refs =NULL ; + expr->metric_expr = m->metric_expr; +if (m->modifier) { +char *tmp; + +if (asprintf (&tmp,"%s:%s" , m->metric_name, m->modifier) <0 ) + expr->metric_name =NULL ; +else + expr->metric_name = tmp; + }else { + expr->metric_name =strdup (m->metric_name); + } + +if (!expr->metric_name) { + ret = -ENOMEM ; +free (metric_events); + goto out; + } + expr->metric_threshold = m->metric_threshold; + expr->metric_unit = m->metric_unit; + expr->metric_events = metric_events; + expr->runtime = m->pctx->sctx.runtime; +list_add (&expr->nd, &me->head); + } + +if (combined_evlist) { +evlist__splice_list_tail (perf_evlist, &combined_evlist->core.entries); +evlist__delete (combined_evlist); + } + +list_for_each_entry (m, &metric_list, nd) { +if (m->evlist) +evlist__splice_list_tail (perf_evlist, &m->evlist->core.entries); + } + +out: +metricgroup__free_metrics (&metric_list); +return ret; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,
@@ -1698,10 +1909,15 @@ int metricgroup__parse_groups(struct evlist *perf_evlist,if (!table)return -EINVAL ;if (hardware_aware_grouping) { +int ret;pr_debug ("Use hardware aware grouping instead of traditional metric grouping method\n" ); + ret =hw_aware_parse_groups (perf_evlist, pmu, str, + metric_no_threshold, user_requested_cpu_list, system_wide, +/*fake_pmu=*/ NULL , metric_events, table); +if (!ret) +return 0 ; } -return parse_groups (perf_evlist, pmu, str, metric_no_group, metric_no_merge, metric_no_threshold, user_requested_cpu_list, system_wide,/*fake_pmu=*/ NULL , metric_events, table);
Signature · 1 line
2.42.0
RFC PATCH v4 3/15 perf pmu-events: Add functions in jevent.py to parse counter and event info for hardware aware grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
These functions are added to parse event counter restrictions and counter availability info from json files so that the metric grouping method could do grouping based on the counter restriction of events and the counters that are available on the system.
--- tools/perf/pmu-events/jevents.py | 185 +++++++++++++++++++++++++++-- tools/perf/pmu-events/pmu-events.h | 34 +++++- 2 files changed, 207 insertions(+), 12 deletions(-)
tools/perf/pmu-events/jevents.py +177 -8
@@ -23,6 +23,8 @@ _metric_tables = [] _sys_metric_tables = [] # Mapping between sys event table names and sys metric table names. _sys_event_table_to_metric_table_mapping = {} +# List of regular PMU counter layout tables. +_pmu_layouts_tables = [] # Map from an event name to an architecture standard # JsonEvent. Architecture standard events are in json files in the top # f'{_args.starting_dir}/{_args.arch}' directory.
@@ -31,6 +33,10 @@ _arch_std_events = {} _pending_events = [] # Name of events table to be written out _pending_events_tblname = None +# PMU counter layout to write out when the layout table is closed +_pending_pmu_counts = [] +# Name of PMU counter layout table to be written out +_pending_pmu_counts_tblname = None # Metrics to write out when the table is closed _pending_metrics = [] # Name of metrics table to be written out
@@ -47,10 +53,17 @@ _json_event_attributes = [ 'event', # Short things in alphabetical order. 'compat', 'deprecated', 'perpkg', 'unit', + # The list of counter(s) this event could use + 'counters', # Longer things (the last won't be iterated over during decompress). 'long_desc' ] +# Attributes that are in pmu_unit_layout. +_json_layout_attributes = [ + 'pmu', 'desc', 'num_counters', 'num_fixed_counters' +] + # Attributes that are in pmu_metric rather than pmu_event. _json_metric_attributes = [ 'metric_name', 'metric_group', 'metric_expr', 'metric_threshold',
@@ -58,7 +71,9 @@ _json_metric_attributes = [ 'default_metricgroup_name', 'aggr_mode', 'event_grouping' ] # Attributes that are bools or enum int values, encoded as '0', '1',... -_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg'] +_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg', + 'num_counters', 'num_fixed_counters' +] def removesuffix(s: str, suffix: str) -> str: """Remove the suffix from a string
@@ -334,6 +349,12 @@ class JsonEvent: if 'Errata' in jd: extra_desc += ' Spec update: ' + jd['Errata'] self.pmu = unit_to_pmu(jd.get('Unit')) + # The list of counter(s) the event could be collected with + self.counters = jd.get('Counter') + # Number of generic counter + self.num_counters = jd.get('NumCounters') + # Number of fixed counter + self.num_fixed_counters = jd.get('NumFixedCounters') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')
@@ -405,8 +426,16 @@ class JsonEvent: s += f'\t{attr} = {value},\n' return s + '}' - def build_c_string(self, metric: bool) -> str: + def build_c_string(self, metric: bool, layout: bool) -> str: s = '' + if layout: + for attr in _json_layout_attributes: + x = getattr(self, attr) + if attr in _json_enum_attributes: + s += x if x else '0' + else: + s += f'{x}\\000' if x else '\\000' + return s for attr in _json_metric_attributes if metric else _json_event_attributes: x = getattr(self, attr) if metric and x and attr == 'metric_expr':
@@ -421,10 +450,10 @@ class JsonEvent: s += f'{x}\\000' if x else '\\000' return s - def to_c_string(self, metric: bool) -> str: + def to_c_string(self, metric: bool, layout: bool) -> str: """Representation of the event as a C struct initializer.""" - s = self.build_c_string(metric) + s = self.build_c_string(metric, layout) return f'{{ { _bcs.offsets[s] } }}, /* {s} */\n'
@@ -461,6 +490,8 @@ def preprocess_arch_std_files(archpath: str) -> None: _arch_std_events[event.name.lower()] = event if event.metric_name: _arch_std_events[event.metric_name.lower()] = event + if event.num_counters: + _arch_std_events[event.pmu.lower()] = event def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -470,6 +501,8 @@ def add_events_table_entries(item: os.DirEntry, topic: str) -> None: _pending_events.append(e) if e.metric_name: _pending_metrics.append(e) + if e.num_counters: + _pending_pmu_counts.append(e) def print_pending_events() -> None:
@@ -510,7 +543,7 @@ def print_pending_events() -> None: last_pmu = event.pmu pmus.add((event.pmu, pmu_name)) - _args.output_file.write(event.to_c_string(metric=False)) + _args.output_file.write(event.to_c_string(metric=False, layout=False)) _pending_events = [] _args.output_file.write(f"""
@@ -565,7 +598,7 @@ def print_pending_metrics() -> None: last_pmu = metric.pmu pmus.add((metric.pmu, pmu_name)) - _args.output_file.write(metric.to_c_string(metric=True)) + _args.output_file.write(metric.to_c_string(metric=True, layout=False)) _pending_metrics = [] _args.output_file.write(f"""
@@ -583,6 +616,35 @@ const struct pmu_table_entry {_pending_metrics_tblname}[] = {{ """) _args.output_file.write('};\n\n') +def print_pending_pmu_counter_layout_table() -> None: + '''Print counter layout data from counter.json file to counter layout table in + c-string''' + + def pmu_counts_cmp_key(j: JsonEvent) -> Tuple[bool, str, str]: + def fix_none(s: Optional[str]) -> str: + if s is None: + return '' + return s + + return (j.desc is not None, fix_none(j.pmu), fix_none(j.num_counters)) + + global _pending_pmu_counts + if not _pending_pmu_counts: + return + + global _pending_pmu_counts_tblname + global pmu_layouts_tables + _pmu_layouts_tables.append(_pending_pmu_counts_tblname) + + _args.output_file.write( + f'static const struct compact_pmu_event {_pending_pmu_counts_tblname}[] = {{\n') + + for pmu_layout in sorted(_pending_pmu_counts, key=pmu_counts_cmp_key): + _args.output_file.write(pmu_layout.to_c_string(metric=False, layout=True)) + _pending_pmu_counts = [] + + _args.output_file.write('};\n\n') + def get_topic(topic: str) -> str: if topic.endswith('metrics.json'): return 'metrics'
@@ -619,10 +681,12 @@ def preprocess_one_file(parents: Sequence[str], item: os.DirEntry) -> None: pmu_name = f"{event.pmu}\\000" if event.name: _bcs.add(pmu_name, metric=False) - _bcs.add(event.build_c_string(metric=False), metric=False) + _bcs.add(event.build_c_string(metric=False, layout=False), metric=False) if event.metric_name: _bcs.add(pmu_name, metric=True) - _bcs.add(event.build_c_string(metric=True), metric=True) + _bcs.add(event.build_c_string(metric=True, layout=False), metric=True) + if event.num_counters: + _bcs.add(event.build_c_string(metric=False, layout=True), metric=False) def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: """Process a JSON file during the main walk."""
@@ -636,11 +700,14 @@ def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: if item.is_dir() and is_leaf_dir(item.path): print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() global _pending_events_tblname _pending_events_tblname = file_name_to_table_name('pmu_events_', parents, item.name) global _pending_metrics_tblname _pending_metrics_tblname = file_name_to_table_name('pmu_metrics_', parents, item.name) + global _pending_pmu_counts_tblname + _pending_pmu_counts_tblname = file_name_to_table_name('pmu_layouts_', parents, item.name) if item.name == 'sys': _sys_event_table_to_metric_table_mapping[_pending_events_tblname] = _pending_metrics_tblname
@@ -674,6 +741,12 @@ struct pmu_metrics_table { uint32_t num_pmus; }; +/* Struct used to make the PMU counter layout table implementation opaque to callers. */ +struct pmu_layouts_table { + const struct compact_pmu_event *entries; + size_t length; +}; + /* * Map a CPU to its table of PMU events. The CPU is identified by the * cpuid field, which is an arch-specific identifier for the CPU.
@@ -687,6 +760,7 @@ struct pmu_events_map { const char *cpuid; struct pmu_events_table event_table; struct pmu_metrics_table metric_table; + struct pmu_layouts_table layout_table; }; /*
@@ -731,6 +805,12 @@ const struct pmu_events_map pmu_events_map[] = { metric_size = '0' if event_size == '0' and metric_size == '0': continue + layout_tblname = file_name_to_table_name('pmu_layouts_', [], row[2].replace('/', '_')) + if layout_tblname in _pmu_layouts_tables: + layout_size = f'ARRAY_SIZE({layout_tblname})' + else: + layout_tblname = 'NULL' + layout_size = '0' cpuid = row[0].replace('\\', '\\\\') _args.output_file.write(f"""{{ \t.arch = "{arch}",
@@ -742,6 +822,10 @@ const struct pmu_events_map pmu_events_map[] = { \t.metric_table = {{ \t\t.pmus = {metric_tblname}, \t\t.num_pmus = {metric_size} +\t}}, +\t.layout_table = {{ +\t\t.entries = {layout_tblname}, +\t\t.length = {layout_size} \t}} }}, """)
@@ -752,6 +836,7 @@ const struct pmu_events_map pmu_events_map[] = { \t.cpuid = 0, \t.event_table = { 0, 0 }, \t.metric_table = { 0, 0 }, +\t.layout_table = { 0, 0 }, } }; """)
@@ -840,6 +925,24 @@ static void decompress_metric(int offset, struct pmu_metric *pm) _args.output_file.write('\twhile (*p++);') _args.output_file.write("""} +static void decompress_layout(int offset, struct pmu_layout *pm) +{ +\tconst char *p = &big_c_string[offset]; +""") + for attr in _json_layout_attributes: + _args.output_file.write(f'\n\tpm->{attr} = ') + if attr in _json_enum_attributes: + _args.output_file.write("*p - '0';\n") + else: + _args.output_file.write("(*p == '\\0' ? NULL : p);\n") + if attr == _json_layout_attributes[-1]: + continue + if attr in _json_enum_attributes: + _args.output_file.write('\tp++;') + else: + _args.output_file.write('\twhile (*p++);') + _args.output_file.write("""} + static int pmu_events_table__for_each_event_pmu(const struct pmu_events_table *table, const struct pmu_table_entry *pmu, pmu_event_iter_fn fn,
@@ -995,6 +1098,21 @@ int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, return 0; } +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table *table, + pmu_layout_iter_fn fn, + void *data) { + for (size_t i = 0; i < table->length; i++) { + struct pmu_layout pm; + int ret; + + decompress_layout(table->entries[i].offset, &pm); + ret = fn(&pm, data); + if (ret) + return ret; + } + return 0; +} + static const struct pmu_events_map *map_for_pmu(struct perf_pmu *pmu) { static struct {
@@ -1090,6 +1208,33 @@ const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm return NULL; } +const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct perf_pmu *pmu) +{ + const struct pmu_layouts_table *table = NULL; + char *cpuid = perf_pmu__getcpuid(pmu); + int i; + + /* on some platforms which uses cpus map, cpuid can be NULL for + * PMUs other than CORE PMUs. + */ + if (!cpuid) + return NULL; + + i = 0; + for (;;) { + const struct pmu_events_map *map = &pmu_events_map[i++]; + if (!map->arch) + break; + + if (!strcmp_cpuid_str(map->cpuid, cpuid)) { + table = &map->layout_table; + break; + } + } + free(cpuid); + return table; +} + const struct pmu_events_table *find_core_events_table(const char *arch, const char *cpuid) { for (const struct pmu_events_map *tables = &pmu_events_map[0];
@@ -1111,6 +1256,16 @@ const struct pmu_metrics_table *find_core_metrics_table(const char *arch, const } return NULL; } +const struct pmu_layouts_table *find_core_layouts_table(const char *arch, const char *cpuid) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + if (!strcmp(tables->arch, arch) && !strcmp_cpuid_str(tables->cpuid, cpuid)) + return &tables->layout_table; + } + return NULL; +} int pmu_for_each_core_event(pmu_event_iter_fn fn, void *data) {
@@ -1139,6 +1294,19 @@ int pmu_for_each_core_metric(pmu_metric_iter_fn fn, void *data) return 0; } +int pmu_for_each_core_layout(pmu_layout_iter_fn fn, void *data) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + int ret = pmu_layouts_table__for_each_layout(&tables->layout_table, fn, data); + + if (ret) + return ret; + } + return 0; +} + const struct pmu_events_table *find_sys_events_table(const char *name) { for (const struct pmu_sys_events *tables = &pmu_sys_event_tables[0];
@@ -1295,6 +1463,7 @@ struct pmu_table_entry { ftw(arch_path, [], process_one_file) print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() print_mapping_table(archs) print_system_mapping_table()
tools/perf/pmu-events/pmu-events.h +30 -4
@@ -45,6 +45,11 @@ struct pmu_event {const char *desc;const char *topic;const char *long_desc; +/** +* The list of counter(s) the event could be collected on. +* eg., "0,1,2,3,4,5,6,7". +*/ +const char *counters;const char *pmu;const char *unit;bool perpkg;
@@ -67,8 +72,20 @@ struct pmu_metric {enum metric_event_groups event_grouping; }; +struct pmu_layout { +const char *pmu; +const char *desc; +/** Total number of generic counters*/ +int num_counters; +/** +* Total number of fixed counters. +* Set to zero if no fixed counter on the unit.*/ +int num_fixed_counters; +}; +struct pmu_events_table ;struct pmu_metrics_table ; +struct pmu_layouts_table ; typedef int (*pmu_event_iter_fn)(const struct pmu_event *pe, const struct pmu_events_table *table,
@@ -78,15 +95,21 @@ typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,const struct pmu_metrics_table *table, void *data); +typedef int (*pmu_layout_iter_fn )(const struct pmu_layout *pm, +void *data); +int pmu_events_table__for_each_event (const struct pmu_events_table *table,struct perf_pmu *pmu,pmu_event_iter_fn fn,void *data);int pmu_events_table__find_event (const struct pmu_events_table *table, -struct perf_pmu *pmu, -const char *name, -pmu_event_iter_fn fn, -void *data); +struct perf_pmu *pmu, +const char *name, +pmu_event_iter_fn fn, +void *data); +int pmu_layouts_table__for_each_layout (const struct pmu_layouts_table *table, +pmu_layout_iter_fn fn, +void *data);size_t pmu_events_table__num_events (const struct pmu_events_table *table,struct perf_pmu *pmu);
@@ -95,10 +118,13 @@ int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, pmconst struct pmu_events_table *perf_pmu__find_events_table (struct perf_pmu *pmu);const struct pmu_metrics_table *perf_pmu__find_metrics_table (struct perf_pmu *pmu); +const struct pmu_layouts_table *perf_pmu__find_layouts_table (struct perf_pmu *pmu);const struct pmu_events_table *find_core_events_table (const char *arch,const char *cpuid);const struct pmu_metrics_table *find_core_metrics_table (const char *arch,const char *cpuid); +const struct pmu_layouts_table *find_core_layouts_table (const char *arch,const char *cpuid);int pmu_for_each_core_event (pmu_event_iter_fn fn,void *data);int pmu_for_each_core_metric (pmu_metric_iter_fn fn,void *data); +int pmu_for_each_core_layout (pmu_layout_iter_fn fn,void *data);const struct pmu_events_table *find_sys_events_table (const char *name);const struct pmu_metrics_table *find_sys_metrics_table (const char *name);
Signature · 1 line
2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 03/15] perf pmu-events: Add functions in jevent.py to parse counter and event info for hardware aware grouping
Quoted 68 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
These functions are added to parse event counter restrictions and counter availability info from json files so that the metric grouping method could do grouping based on the counter restriction of events and the counters that are available on the system.
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 185 +++++++++++++++++++++++++++-- tools/perf/pmu-events/pmu-events.h | 34 +++++- 2 files changed, 207 insertions(+), 12 deletions(-)tools/perf/pmu-events/jevents.py +16
-1@@ -23,6 +23,8 @@ _metric_tables = [] _sys_metric_tables = [] # Mapping between sys event table names and sys metric table names. _sys_event_table_to_metric_table_mapping = {} +# List of regular PMU counter layout tables. +_pmu_layouts_tables = [] # Map from an event name to an architecture standard # JsonEvent. Architecture standard events are in json files in the top # f'{_args.starting_dir}/{_args.arch}' directory.@@ -31,6 +33,10 @@ _arch_std_events = {} _pending_events = [] # Name of events table to be written out _pending_events_tblname = None +# PMU counter layout to write out when the layout table is closed +_pending_pmu_counts = [] +# Name of PMU counter layout table to be written out +_pending_pmu_counts_tblname = None # Metrics to write out when the table is closed _pending_metrics = [] # Name of metrics table to be written out@@ -47,10 +53,17 @@ _json_event_attributes = [ 'event', # Short things in alphabetical order. 'compat', 'deprecated', 'perpkg', 'unit', + # The list of counter(s) this event could use + 'counters', # Longer things (the last won't be iterated over during decompress). 'long_desc' ] +# Attributes that are in pmu_unit_layout. +_json_layout_attributes = [ + 'pmu', 'desc', 'num_counters', 'num_fixed_counters' +] + # Attributes that are in pmu_metric rather than pmu_event. _json_metric_attributes = [ 'metric_name', 'metric_group', 'metric_expr', 'metric_threshold',@@ -58,7 +71,9 @@ _json_metric_attributes = [ 'default_metricgroup_name', 'aggr_mode', 'event_grouping' ] # Attributes that are bools or enum int values, encoded as '0', '1',... -_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg'] +_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg', + 'num_counters', 'num_fixed_counters' +] def removesuffix(s: str, suffix: str) -> str: """Remove the suffix from a string
tools/perf/pmu-events/jevents.py class JsonEvent: · line 357
Quoted 9 lines
tools/perf/pmu-events/jevents.py +6
-0@@ -334,6 +349,12 @@ class JsonEvent: if 'Errata' in jd: extra_desc += ' Spec update: ' + jd['Errata'] self.pmu = unit_to_pmu(jd.get('Unit')) + # The list of counter(s) the event could be collected with + self.counters = jd.get('Counter') + # Number of generic counter + self.num_counters = jd.get('NumCounters') + # Number of fixed counter + self.num_fixed_counters = jd.get('NumFixedCounters')
I'm not a fan of counters vs fixed_counters, perhaps in the python:
self.counters['generic'].list = jd.get('Counter') self.counters['generic'].num = jd.get('NumCounters') self.counters['fixed'].num = jd.get('NumFixedCounters')
Would make things a bit more intention revealing. Fwiw, the json name is worse as Counter is really for a plural of things. I don't think we can update this in the Intel perfmon events, but maybe we can do it in the create_perf_json.py converter script and its generated json.
This is pretty trivial as clean up so I'd prefer it before my Reviewed-by unless there's something I'm overlooking.
Thanks,
Ian
Quoted 395 lines
tools/perf/pmu-events/jevents.py +155
-7filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')@@ -405,8 +426,16 @@ class JsonEvent: s += f'\t{attr} = {value},\n' return s + '}' - def build_c_string(self, metric: bool) -> str: + def build_c_string(self, metric: bool, layout: bool) -> str: s = '' + if layout: + for attr in _json_layout_attributes: + x = getattr(self, attr) + if attr in _json_enum_attributes: + s += x if x else '0' + else: + s += f'{x}\\000' if x else '\\000' + return s for attr in _json_metric_attributes if metric else _json_event_attributes: x = getattr(self, attr) if metric and x and attr == 'metric_expr':@@ -421,10 +450,10 @@ class JsonEvent: s += f'{x}\\000' if x else '\\000' return s - def to_c_string(self, metric: bool) -> str: + def to_c_string(self, metric: bool, layout: bool) -> str: """Representation of the event as a C struct initializer.""" - s = self.build_c_string(metric) + s = self.build_c_string(metric, layout) return f'{{ { _bcs.offsets[s] } }}, /* {s} */\n'@@ -461,6 +490,8 @@ def preprocess_arch_std_files(archpath: str) -> None: _arch_std_events[event.name.lower()] = event if event.metric_name: _arch_std_events[event.metric_name.lower()] = event + if event.num_counters: + _arch_std_events[event.pmu.lower()] = event def add_events_table_entries(item: os.DirEntry, topic: str) -> None:@@ -470,6 +501,8 @@ def add_events_table_entries(item: os.DirEntry, topic: str) -> None: _pending_events.append(e) if e.metric_name: _pending_metrics.append(e) + if e.num_counters: + _pending_pmu_counts.append(e) def print_pending_events() -> None:@@ -510,7 +543,7 @@ def print_pending_events() -> None: last_pmu = event.pmu pmus.add((event.pmu, pmu_name)) - _args.output_file.write(event.to_c_string(metric=False)) + _args.output_file.write(event.to_c_string(metric=False, layout=False)) _pending_events = [] _args.output_file.write(f"""@@ -565,7 +598,7 @@ def print_pending_metrics() -> None: last_pmu = metric.pmu pmus.add((metric.pmu, pmu_name)) - _args.output_file.write(metric.to_c_string(metric=True)) + _args.output_file.write(metric.to_c_string(metric=True, layout=False)) _pending_metrics = [] _args.output_file.write(f"""@@ -583,6 +616,35 @@ const struct pmu_table_entry {_pending_metrics_tblname}[] = {{ """) _args.output_file.write('};\n\n') +def print_pending_pmu_counter_layout_table() -> None: + '''Print counter layout data from counter.json file to counter layout table in + c-string''' + + def pmu_counts_cmp_key(j: JsonEvent) -> Tuple[bool, str, str]: + def fix_none(s: Optional[str]) -> str: + if s is None: + return '' + return s + + return (j.desc is not None, fix_none(j.pmu), fix_none(j.num_counters)) + + global _pending_pmu_counts + if not _pending_pmu_counts: + return + + global _pending_pmu_counts_tblname + global pmu_layouts_tables + _pmu_layouts_tables.append(_pending_pmu_counts_tblname) + + _args.output_file.write( + f'static const struct compact_pmu_event {_pending_pmu_counts_tblname}[] = {{\n') + + for pmu_layout in sorted(_pending_pmu_counts, key=pmu_counts_cmp_key): + _args.output_file.write(pmu_layout.to_c_string(metric=False, layout=True)) + _pending_pmu_counts = [] + + _args.output_file.write('};\n\n') + def get_topic(topic: str) -> str: if topic.endswith('metrics.json'): return 'metrics'@@ -619,10 +681,12 @@ def preprocess_one_file(parents: Sequence[str], item: os.DirEntry) -> None: pmu_name = f"{event.pmu}\\000" if event.name: _bcs.add(pmu_name, metric=False) - _bcs.add(event.build_c_string(metric=False), metric=False) + _bcs.add(event.build_c_string(metric=False, layout=False), metric=False) if event.metric_name: _bcs.add(pmu_name, metric=True) - _bcs.add(event.build_c_string(metric=True), metric=True) + _bcs.add(event.build_c_string(metric=True, layout=False), metric=True) + if event.num_counters: + _bcs.add(event.build_c_string(metric=False, layout=True), metric=False) def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: """Process a JSON file during the main walk."""@@ -636,11 +700,14 @@ def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: if item.is_dir() and is_leaf_dir(item.path): print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() global _pending_events_tblname _pending_events_tblname = file_name_to_table_name('pmu_events_', parents, item.name) global _pending_metrics_tblname _pending_metrics_tblname = file_name_to_table_name('pmu_metrics_', parents, item.name) + global _pending_pmu_counts_tblname + _pending_pmu_counts_tblname = file_name_to_table_name('pmu_layouts_', parents, item.name) if item.name == 'sys': _sys_event_table_to_metric_table_mapping[_pending_events_tblname] = _pending_metrics_tblname@@ -674,6 +741,12 @@ struct pmu_metrics_table { uint32_t num_pmus; }; +/* Struct used to make the PMU counter layout table implementation opaque to callers. */ +struct pmu_layouts_table { + const struct compact_pmu_event *entries; + size_t length; +}; + /* * Map a CPU to its table of PMU events. The CPU is identified by the * cpuid field, which is an arch-specific identifier for the CPU.@@ -687,6 +760,7 @@ struct pmu_events_map { const char *cpuid; struct pmu_events_table event_table; struct pmu_metrics_table metric_table; + struct pmu_layouts_table layout_table; }; /*@@ -731,6 +805,12 @@ const struct pmu_events_map pmu_events_map[] = { metric_size = '0' if event_size == '0' and metric_size == '0': continue + layout_tblname = file_name_to_table_name('pmu_layouts_', [], row[2].replace('/', '_')) + if layout_tblname in _pmu_layouts_tables: + layout_size = f'ARRAY_SIZE({layout_tblname})' + else: + layout_tblname = 'NULL' + layout_size = '0' cpuid = row[0].replace('\\', '\\\\') _args.output_file.write(f"""{{ \t.arch = "{arch}",@@ -742,6 +822,10 @@ const struct pmu_events_map pmu_events_map[] = { \t.metric_table = {{ \t\t.pmus = {metric_tblname}, \t\t.num_pmus = {metric_size} +\t}}, +\t.layout_table = {{ +\t\t.entries = {layout_tblname}, +\t\t.length = {layout_size} \t}} }}, """)@@ -752,6 +836,7 @@ const struct pmu_events_map pmu_events_map[] = { \t.cpuid = 0, \t.event_table = { 0, 0 }, \t.metric_table = { 0, 0 }, +\t.layout_table = { 0, 0 }, } }; """)@@ -840,6 +925,24 @@ static void decompress_metric(int offset, struct pmu_metric *pm) _args.output_file.write('\twhile (*p++);') _args.output_file.write("""} +static void decompress_layout(int offset, struct pmu_layout *pm) +{ +\tconst char *p = &big_c_string[offset]; +""") + for attr in _json_layout_attributes: + _args.output_file.write(f'\n\tpm->{attr} = ') + if attr in _json_enum_attributes: + _args.output_file.write("*p - '0';\n") + else: + _args.output_file.write("(*p == '\\0' ? NULL : p);\n") + if attr == _json_layout_attributes[-1]: + continue + if attr in _json_enum_attributes: + _args.output_file.write('\tp++;') + else: + _args.output_file.write('\twhile (*p++);') + _args.output_file.write("""} + static int pmu_events_table__for_each_event_pmu(const struct pmu_events_table *table, const struct pmu_table_entry *pmu, pmu_event_iter_fn fn,@@ -995,6 +1098,21 @@ int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, return 0; } +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table *table, + pmu_layout_iter_fn fn, + void *data) { + for (size_t i = 0; i < table->length; i++) { + struct pmu_layout pm; + int ret; + + decompress_layout(table->entries[i].offset, &pm); + ret = fn(&pm, data); + if (ret) + return ret; + } + return 0; +} + static const struct pmu_events_map *map_for_pmu(struct perf_pmu *pmu) { static struct {@@ -1090,6 +1208,33 @@ const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm return NULL; } +const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct perf_pmu *pmu) +{ + const struct pmu_layouts_table *table = NULL; + char *cpuid = perf_pmu__getcpuid(pmu); + int i; + + /* on some platforms which uses cpus map, cpuid can be NULL for + * PMUs other than CORE PMUs. + */ + if (!cpuid) + return NULL; + + i = 0; + for (;;) { + const struct pmu_events_map *map = &pmu_events_map[i++]; + if (!map->arch) + break; + + if (!strcmp_cpuid_str(map->cpuid, cpuid)) { + table = &map->layout_table; + break; + } + } + free(cpuid); + return table; +} + const struct pmu_events_table *find_core_events_table(const char *arch, const char *cpuid) { for (const struct pmu_events_map *tables = &pmu_events_map[0];@@ -1111,6 +1256,16 @@ const struct pmu_metrics_table *find_core_metrics_table(const char *arch, const } return NULL; } +const struct pmu_layouts_table *find_core_layouts_table(const char *arch, const char *cpuid) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + if (!strcmp(tables->arch, arch) && !strcmp_cpuid_str(tables->cpuid, cpuid)) + return &tables->layout_table; + } + return NULL; +} int pmu_for_each_core_event(pmu_event_iter_fn fn, void *data) {@@ -1139,6 +1294,19 @@ int pmu_for_each_core_metric(pmu_metric_iter_fn fn, void *data) return 0; } +int pmu_for_each_core_layout(pmu_layout_iter_fn fn, void *data) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + int ret = pmu_layouts_table__for_each_layout(&tables->layout_table, fn, data); + + if (ret) + return ret; + } + return 0; +} + const struct pmu_events_table *find_sys_events_table(const char *name) { for (const struct pmu_sys_events *tables = &pmu_sys_event_tables[0];@@ -1295,6 +1463,7 @@ struct pmu_table_entry { ftw(arch_path, [], process_one_file) print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() print_mapping_table(archs) print_system_mapping_table()tools/perf/pmu-events/pmu-events.h +30
-5@@ -45,6 +45,11 @@ struct pmu_event {const char *desc;const char *topic;const char *long_desc; +/** +* The list of counter(s) the event could be collected on. +* eg., "0,1,2,3,4,5,6,7". +*/ +const char *counters;const char *pmu;const char *unit;bool perpkg;@@ -67,8 +72,20 @@ struct pmu_metric {enum metric_event_groups event_grouping; }; +struct pmu_layout { +const char *pmu; +const char *desc; +/** Total number of generic counters*/ +int num_counters; +/** +* Total number of fixed counters. +* Set to zero if no fixed counter on the unit.*/ +int num_fixed_counters; +}; +struct pmu_events_table ;struct pmu_metrics_table ; +struct pmu_layouts_table ; typedef int (*pmu_event_iter_fn)(const struct pmu_event *pe, const struct pmu_events_table *table,@@ -78,15 +95,21 @@ typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,const struct pmu_metrics_table *table, void *data); +typedef int (*pmu_layout_iter_fn )(const struct pmu_layout *pm, +void *data); +int pmu_events_table__for_each_event (const struct pmu_events_table *table,struct perf_pmu *pmu,pmu_event_iter_fn fn,void *data);int pmu_events_table__find_event (const struct pmu_events_table *table, -struct perf_pmu *pmu, -const char *name, -pmu_event_iter_fn fn, -void *data); +struct perf_pmu *pmu, +const char *name, +pmu_event_iter_fn fn, +void *data); +int pmu_layouts_table__for_each_layout (const struct pmu_layouts_table *table, +pmu_layout_iter_fn fn, +void *data);size_t pmu_events_table__num_events (const struct pmu_events_table *table,struct perf_pmu *pmu);@@ -95,10 +118,13 @@ int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, pmconst struct pmu_events_table *perf_pmu__find_events_table (struct perf_pmu *pmu);const struct pmu_metrics_table *perf_pmu__find_metrics_table (struct perf_pmu *pmu); +const struct pmu_layouts_table *perf_pmu__find_layouts_table (struct perf_pmu *pmu);const struct pmu_events_table *find_core_events_table (const char *arch,const char *cpuid);const struct pmu_metrics_table *find_core_metrics_table (const char *arch,const char *cpuid); +const struct pmu_layouts_table *find_core_layouts_table (const char *arch,const char *cpuid);int pmu_for_each_core_event (pmu_event_iter_fn fn,void *data);int pmu_for_each_core_metric (pmu_metric_iter_fn fn,void *data); +int pmu_for_each_core_layout (pmu_layout_iter_fn fn,void *data);const struct pmu_events_table *find_sys_events_table (const char *name);const struct pmu_metrics_table *find_sys_metrics_table (const char *name); --2.42.0
RE: [RFC PATCH v4 03/15] perf pmu-events: Add functions in jevent.py to parse counter and event info for hardware aware grouping
Quoted 102 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 9:49 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 03/15] perf pmu-events: Add functions injevent.pyto parse counter and event info for hardware aware groupingQuoted 11 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
These functions are added to parse event counter restrictions and counter availability info from json files so that the metric grouping method could do grouping based on the counter restriction of events and the counters that are available on the system.
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 185+++++++++++++++++++++++++++--
tools/perf/pmu-
Quoted 4 lines
tools/perf/pmu-events/pmu-events.h | 34 +++++- 2 files changed, 207 insertions(+), 12 deletions(-)tools/perf/pmu- renamed from tools/perf/pmu-events/jevents.py +0
-0
events/jevents.pyQuoted 42 lines
index
2c7e5d61ce92..bc91b7efa49a100755tools/perf/pmu-events/jevents.py +13
-0@@ -23,6 +23,8 @@ _metric_tables = [] _sys_metric_tables = [] # Mapping between sys event table names and sys metric table names. _sys_event_table_to_metric_table_mapping = {} +# List of regular PMU counter layout tables. +_pmu_layouts_tables = [] # Map from an event name to an architecture standard # JsonEvent. Architecture standard events are in json files in the top # f'{_args.starting_dir}/{_args.arch}' directory.@@ -31,6 +33,10 @@ _arch_std_events = {} _pending_events = [] # Name of events table to be written out _pending_events_tblname = None +# PMU counter layout to write out when the layout table is closed +_pending_pmu_counts = [] +# Name of PMU counter layout table to be written out +_pending_pmu_counts_tblname = None # Metrics to write out when the table is closed _pending_metrics = [] # Name of metrics table to be written out@@ -47,10 +53,17 @@ _json_event_attributes = [ 'event', # Short things in alphabetical order. 'compat', 'deprecated', 'perpkg', 'unit', + # The list of counter(s) this event could use + 'counters', # Longer things (the last won't be iterated over during decompress). 'long_desc' ] +# Attributes that are in pmu_unit_layout. +_json_layout_attributes = [ + 'pmu', 'desc', 'num_counters', 'num_fixed_counters' +] + # Attributes that are in pmu_metric rather than pmu_event. _json_metric_attributes = [ 'metric_name', 'metric_group', 'metric_expr', 'metric_threshold',tools/perf/pmu-events/jevents.py _json_metric_attributes = [ · line 61
Quoted 4 lines
tools/perf/pmu-events/jevents.py +0
-1@@ -58,7 +71,9 @@ _json_metric_attributes = [ 'default_metricgroup_name', 'aggr_mode', 'event_grouping' ] # Attributes that are bools or enum int values, encoded as '0', '1',... -_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping','perpkg']
tools/perf/pmu-events/jevents.py _json_metric_attributes = [ · line 74
Quoted 1 line
tools/perf/pmu-events/jevents.py +1
-0+_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping','perpkg',
Quoted 6 lines
tools/perf/pmu-events/jevents.py +2
-0+ 'num_counters', 'num_fixed_counters' +] def removesuffix(s: str, suffix: str) -> str: """Remove the suffix from a stringtools/perf/pmu-events/jevents.py class JsonEvent: · line 357
Quoted 9 lines
tools/perf/pmu-events/jevents.py +6
-0@@ -334,6 +349,12 @@ class JsonEvent: if 'Errata' in jd: extra_desc += ' Spec update: ' + jd['Errata'] self.pmu = unit_to_pmu(jd.get('Unit')) + # The list of counter(s) the event could be collected with + self.counters = jd.get('Counter') + # Number of generic counter + self.num_counters = jd.get('NumCounters') + # Number of fixed counter + self.num_fixed_counters = jd.get('NumFixedCounters')I'm not a fan of counters vs
fixed_counters, perhaps in the python:
Quoted 4 lines
self.counters['generic'].list =
jd.get('Counter')self.counters['generic'].num =jd.get('NumCounters')self.counters['fixed'].num =jd.get('NumFixedCounters')
At this point of time, we cannot tell the content in 'Counter' would be for generic counter or fixed counter. Maybe we could do something like:
self.counters['list'] = jd.get('Counter') self.counters['num'].gp = jd.get('NumCounters') self.counters['num'].fixed = jd.get('NumFixedCounters')
One difficulty I still have is to get getattr() work with this type of attribute. Do you have some suggestion?
Thanks,
Weilin
Quoted 449 lines
Would make things a bit more intention revealing. Fwiw, the json name is worse as Counter is really for a plural of things. I don't think we can update this in the Intel perfmon events, but maybe we can do it in the
create_perf_json.pyconverter script and its generated json.This is pretty trivial as clean up so I'd prefer it before my Reviewed-by unless there's something I'm overlooking.
Thanks,
Iantools/perf/pmu-events/jevents.py def preprocess_arch_std_files(archpath: str) ->
Quoted 35 lines
tools/perf/pmu-events/jevents.py +11
-3filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')@@ -405,8 +426,16 @@ class JsonEvent: s += f'\t{attr} = {value},\n' return s + '}' - def build_c_string(self, metric: bool) -> str: + def build_c_string(self, metric: bool, layout: bool) -> str: s = '' + if layout: + for attr in _json_layout_attributes: + x = getattr(self, attr) + if attr in _json_enum_attributes: + s += x if x else '0' + else: + s += f'{x}\\000' if x else '\\000' + return s for attr in _json_metric_attributes if metric else _json_event_attributes: x = getattr(self, attr) if metric and x and attr == 'metric_expr':@@ -421,10 +450,10 @@ class JsonEvent: s += f'{x}\\000' if x else '\\000' return s - def to_c_string(self, metric: bool) -> str: + def to_c_string(self, metric: bool, layout: bool) -> str: """Representation of the event as a C struct initializer.""" - s = self.build_c_string(metric) + s = self.build_c_string(metric, layout) return f'{{ { _bcs.offsets[s] } }}, /* {s} */\n'@@ -461,6 +490,8 @@ def preprocess_arch_std_files(archpath: str) ->None:
tools/perf/pmu-events/jevents.py def add_events_table_entries(item: os.DirEntry,
Quoted 9 lines
tools/perf/pmu-events/jevents.py +2
-0_arch_std_events[event.name.lower()] = event if event.metric_name: _arch_std_events[event.metric_name.lower()] = event + if event.num_counters: + _arch_std_events[event.pmu.lower()] = event def add_events_table_entries(item: os.DirEntry, topic: str) -> None:@@ -470,6 +501,8 @@ def add_events_table_entries(item: os.DirEntry,topic: str) -> None:
tools/perf/pmu-events/jevents.py const struct pmu_table_entry
Quoted 27 lines
tools/perf/pmu-events/jevents.py +4
-2_pending_events.append(e) if e.metric_name: _pending_metrics.append(e) + if e.num_counters: + _pending_pmu_counts.append(e) def print_pending_events() -> None:@@ -510,7 +543,7 @@ def print_pending_events() -> None: last_pmu = event.pmu pmus.add((event.pmu, pmu_name)) - _args.output_file.write(event.to_c_string(metric=False)) + _args.output_file.write(event.to_c_string(metric=False, layout=False)) _pending_events = [] _args.output_file.write(f"""@@ -565,7 +598,7 @@ def print_pending_metrics() -> None: last_pmu = metric.pmu pmus.add((metric.pmu, pmu_name)) - _args.output_file.write(metric.to_c_string(metric=True)) + _args.output_file.write(metric.to_c_string(metric=True, layout=False)) _pending_metrics = [] _args.output_file.write(f"""@@ -583,6 +616,35 @@ const struct pmu_table_entry{_pending_metrics_tblname}[] = {{
Quoted 15 lines
tools/perf/pmu-events/jevents.py +12
-0""") _args.output_file.write('};\n\n') +def print_pending_pmu_counter_layout_table() -> None: + '''Print counter layout data from counter.json file to counter layout table in + c-string''' + + def pmu_counts_cmp_key(j: JsonEvent) -> Tuple[bool, str, str]: + def fix_none(s: Optional[str]) -> str: + if s is None: + return '' + return s + + return (j.desc is not None, fix_none(j.pmu), fix_none(j.num_counters)) +tools/perf/pmu-events/jevents.py const struct pmu_table_entry {_pending_metrics_tblname}[] = {{
Quoted 10 lines
tools/perf/pmu-events/jevents.py +10
-0+ global _pending_pmu_counts + if not _pending_pmu_counts: + return + + global _pending_pmu_counts_tblname + global pmu_layouts_tables + _pmu_layouts_tables.append(_pending_pmu_counts_tblname) + + _args.output_file.write( + f'static const struct compact_pmu_event{_pending_pmu_counts_tblname}[] = {{\n')
tools/perf/pmu-events/jevents.py const struct pmu_table_entry · line 642
Quoted 2 lines
tools/perf/pmu-events/jevents.py +2
-0+ + for pmu_layout in sorted(_pending_pmu_counts,key=pmu_counts_cmp_key):
tools/perf/pmu-events/jevents.py const struct pmu_table_entry · line 643
Quoted 1 line
tools/perf/pmu-events/jevents.py +1
-0+ _args.output_file.write(pmu_layout.to_c_string(metric=False,layout=True))
tools/perf/pmu-events/jevents.py def preprocess_one_file(parents: Sequence[str],
Quoted 8 lines
tools/perf/pmu-events/jevents.py +4
-0+ _pending_pmu_counts = [] + + _args.output_file.write('};\n\n') + def get_topic(topic: str) -> str: if topic.endswith('metrics.json'): return 'metrics'@@ -619,10 +681,12 @@ def preprocess_one_file(parents: Sequence[str],item: os.DirEntry) -> None:
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str],
Quoted 15 lines
tools/perf/pmu-events/jevents.py +4
-2pmu_name = f"{event.pmu}\\000" if event.name: _bcs.add(pmu_name, metric=False) - _bcs.add(event.build_c_string(metric=False), metric=False) + _bcs.add(event.build_c_string(metric=False, layout=False), metric=False) if event.metric_name: _bcs.add(pmu_name, metric=True) - _bcs.add(event.build_c_string(metric=True), metric=True) + _bcs.add(event.build_c_string(metric=True, layout=False), metric=True) + if event.num_counters: + _bcs.add(event.build_c_string(metric=False, layout=True), metric=False) def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: """Process a JSON file during the main walk."""@@ -636,11 +700,14 @@ def process_one_file(parents: Sequence[str],item: os.DirEntry) -> None:
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None:
Quoted 7 lines
tools/perf/pmu-events/jevents.py +1
-0if item.is_dir() and is_leaf_dir(item.path): print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() global _pending_events_tblname _pending_events_tblname = file_name_to_table_name('pmu_events_',parents, item.name)
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 708
Quoted 2 lines
tools/perf/pmu-events/jevents.py +0
-0global _pending_metrics_tblname _pending_metrics_tblname = file_name_to_table_name('pmu_metrics_',parents, item.name)
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 710
Quoted 2 lines
tools/perf/pmu-events/jevents.py +2
-0+ global _pending_pmu_counts_tblname + _pending_pmu_counts_tblname =
file_name_to_table_name('pmu_layouts_', parents, item.name)tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 712
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0if item.name == 'sys': _sys_event_table_to_metric_table_mapping[_pending_events_tblname]=
_pending_metrics_tblnametools/perf/pmu-events/jevents.py struct pmu_metrics_table { · line 744
Quoted 5 lines
tools/perf/pmu-events/jevents.py +1
-0@@ -674,6 +741,12 @@ struct pmu_metrics_table { uint32_t num_pmus; }; +/* Struct used to make the PMU counter layout table implementationopaque to callers. */
tools/perf/pmu-events/jevents.py const struct pmu_events_map
Quoted 17 lines
tools/perf/pmu-events/jevents.py +6
-0+struct pmu_layouts_table { + const struct compact_pmu_event *entries; + size_t length; +}; + /* * Map a CPU to its table of PMU events. The CPU is identified by the * cpuid field, which is an arch-specific identifier for the CPU.@@ -687,6 +760,7 @@ struct pmu_events_map { const char *cpuid; struct pmu_events_table event_table; struct pmu_metrics_table metric_table; + struct pmu_layouts_table layout_table; }; /*@@ -731,6 +805,12 @@ const struct pmu_events_mappmu_events_map[] = {
tools/perf/pmu-events/jevents.py const struct pmu_events_map pmu_events_map[] = {
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0metric_size = '0' if event_size == '0' and metric_size == '0': continue + layout_tblname = file_name_to_table_name('pmu_layouts_', [],row[2].replace('/', '_'))
tools/perf/pmu-events/jevents.py const struct pmu_events_map
Quoted 9 lines
tools/perf/pmu-events/jevents.py +5
-0+ if layout_tblname in _pmu_layouts_tables: + layout_size = f'ARRAY_SIZE({layout_tblname})' + else: + layout_tblname = 'NULL' + layout_size = '0' cpuid = row[0].replace('\\', '\\\\') _args.output_file.write(f"""{{ \t.arch = "{arch}",@@ -742,6 +822,10 @@ const struct pmu_events_mappmu_events_map[] = {
tools/perf/pmu-events/jevents.py const struct pmu_events_map pmu_events_map[]
Quoted 11 lines
tools/perf/pmu-events/jevents.py +4
-0\t.metric_table = {{ \t\t.pmus = {metric_tblname}, \t\t.num_pmus = {metric_size} +\t}}, +\t.layout_table = {{ +\t\t.entries = {layout_tblname}, +\t\t.length = {layout_size} \t}} }}, """)@@ -752,6 +836,7 @@ const struct pmu_events_map pmu_events_map[]= {
tools/perf/pmu-events/jevents.py static void decompress_metric(int offset, struct
Quoted 8 lines
tools/perf/pmu-events/jevents.py +1
-0\t.cpuid = 0, \t.event_table = { 0, 0 }, \t.metric_table = { 0, 0 }, +\t.layout_table = { 0, 0 }, } }; """)@@ -840,6 +925,24 @@ static void decompress_metric(int offset, struct
pmu_metric*pm)Quoted 12 lines
tools/perf/pmu-events/jevents.py +9
-0_args.output_file.write('\twhile (*p++);') _args.output_file.write("""} +static void decompress_layout(int offset, struct pmu_layout *pm) +{ +\tconst char *p = &big_c_string[offset]; +""") + for attr in _json_layout_attributes: + _args.output_file.write(f'\n\tpm->{attr} = ') + if attr in _json_enum_attributes: + _args.output_file.write("*p - '0';\n") + else:tools/perf/pmu-events/jevents.py static void decompress_metric(int offset, struct pmu_metric *pm)
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ _args.output_file.write("(*p == '\\0' ? NULL : p);\n") + if attr == _json_layout_attributes[-1]: + continue + if attr in _json_enum_attributes: + _args.output_file.write('\tp++;') + else: + _args.output_file.write('\twhile (*p++);') + _args.output_file.write("""} + static int pmu_events_table__for_each_event_pmu(const struct
pmu_events_table*table,tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0const struct pmu_table_entry *pmu, pmu_event_iter_fn fn,@@ -995,6 +1098,21 @@ int pmu_metrics_table__for_each_metric(conststruct
pmu_metrics_table*table,tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table,
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0return 0; } +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table*table,
Quoted 5 lines
tools/perf/pmu-events/jevents.py +5
-0+ pmu_layout_iter_fn fn, + void *data) { + for (size_t i = 0; i < table->length; i++) { + struct pmu_layout pm; + int ret;tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table,
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ + decompress_layout(table->entries[i].offset, &pm); + ret = fn(&pm, data); + if (ret) + return ret; + } + return 0; +} + static const struct pmu_events_map *map_for_pmu(struct perf_pmu*pmu)
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0{ static struct {@@ -1090,6 +1208,33 @@ const struct pmu_metrics_table*perf_pmu__find_metrics_table(struct
perf_pmu*pmtools/perf/pmu-events/jevents.py const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0return NULL; } +const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct
perf_pmu*pmu)Quoted 17 lines
tools/perf/pmu-events/jevents.py +17
-0+{ + const struct pmu_layouts_table *table = NULL; + char *cpuid = perf_pmu__getcpuid(pmu); + int i; + + /* on some platforms which uses cpus map, cpuid can be NULL for + * PMUs other than CORE PMUs. + */ + if (!cpuid) + return NULL; + + i = 0; + for (;;) { + const struct pmu_events_map *map = &pmu_events_map[i++]; + if (!map->arch) + break; +tools/perf/pmu-events/jevents.py const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ if (!strcmp_cpuid_str(map->cpuid, cpuid)) { + table = &map->layout_table; + break; + } + } + free(cpuid); + return table; +} + const struct pmu_events_table *find_core_events_table(const char *arch,const char *cpuid)
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0{ for (const struct pmu_events_map *tables = &pmu_events_map[0];@@ -1111,6 +1256,16 @@ const struct pmu_metrics_table*find_core_metrics_table(const char *arch, const
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table *find_core_metrics_table(const char *arch, const
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0} return NULL; } +const struct pmu_layouts_table *find_core_layouts_table(const char *arch,const char *cpuid)
tools/perf/pmu-events/jevents.py int
Quoted 14 lines
tools/perf/pmu-events/jevents.py +9
-0+{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + if (!strcmp(tables->arch, arch) && !strcmp_cpuid_str(tables- cpuid, cpuid)) + return &tables->layout_table; + } + return NULL; +} int pmu_for_each_core_event(pmu_event_iter_fn fn, void *data){
tools/perf/pmu-events/jevents.py +0
-0@@ -1139,6 +1294,19 @@ int
pmu_for_each_core_metric(pmu_metric_iter_fn fn, void *data)tools/perf/pmu-
Quoted 29 lines
tools/perf/pmu-events/jevents.py +13
-0return 0; } +int pmu_for_each_core_layout(pmu_layout_iter_fn fn, void *data) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + int ret = pmu_layouts_table__for_each_layout(&tables- layout_table, fn, data); + + if (ret) + return ret; + } + return 0; +} + const struct pmu_events_table *find_sys_events_table(const char *name) {for (const struct pmu_sys_events *tables = &pmu_sys_event_tables[0];tools/perf/pmu-events/jevents.py +1
-0@@ -1295,6 +1463,7 @@ struct pmu_table_entry { ftw(arch_path, [], process_one_file) print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() print_mapping_table(archs) print_system_mapping_table()tools/perf/pmu- renamed from tools/perf/pmu-events/pmu-events.h +0
-0
events/pmu-events.htools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct
Quoted 37 lines
index
f5aa96f1685c..e245e4738970100644tools/perf/pmu-events/pmu-events.h +17
-0@@ -45,6 +45,11 @@ struct pmu_event {const char *desc;const char *topic;const char *long_desc; +/** +* The list of counter(s) the event could be collected on. +* eg., "0,1,2,3,4,5,6,7". +*/ +const char *counters;const char *pmu;const char *unit;bool perpkg;@@ -67,8 +72,20 @@ struct pmu_metric {enum metric_event_groups event_grouping; }; +struct pmu_layout { +const char *pmu; +const char *desc; +/** Total number of generic counters*/ +int num_counters; +/** +* Total number of fixed counters. +* Set to zero if no fixed counter on the unit.*/ +int num_fixed_counters; +}; +struct pmu_events_table ;struct pmu_metrics_table ; +struct pmu_layouts_table ; typedef int (*pmu_event_iter_fn)(const struct pmu_event *pe, const struct pmu_events_table *table,@@ -78,15 +95,21 @@ typedef int (*pmu_metric_iter_fn)(const struct
pmu_metric*pm,tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 7 lines
tools/perf/pmu-events/pmu-events.h +3
-0const struct pmu_metrics_table *table, void *data); +typedef int (*pmu_layout_iter_fn )(const struct pmu_layout *pm, +void *data); + int pmu_events_table__for_each_event(const struct pmu_events_table*table,
tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 13 lines
tools/perf/pmu-events/pmu-events.h +5
-4struct perf_pmu *pmu, pmu_event_iter_fn fn, void *data);int pmu_events_table__find_event (const struct pmu_events_table *table, -struct perf_pmu *pmu, -const char *name, -pmu_event_iter_fn fn, -void *data); +struct perf_pmu *pmu, +const char *name, +pmu_event_iter_fn fn, +void *data); +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table*table,
tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +2
-0+pmu_layout_iter_fn fn, + void *data); size_t pmu_events_table__num_events(const struct pmu_events_table*table,
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +0
-0struct perf_pmu *pmu);@@ -95,10 +118,13 @@ int pmu_metrics_table__for_each_metric(conststruct
pmu_metrics_table*table, pmtools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 118
Quoted 2 lines
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_events_table *perf_pmu__find_events_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 119
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 120
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +1
-0+const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 121
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_events_table *find_core_events_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 122
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_metrics_table *find_core_metrics_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 123
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +1
-0+const struct pmu_layouts_table *find_core_layouts_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, pm
Quoted 6 lines
tools/perf/pmu-events/pmu-events.h +1
-0int pmu_for_each_core_event (pmu_event_iter_fn fn,void *data);int pmu_for_each_core_metric (pmu_metric_iter_fn fn,void *data); +int pmu_for_each_core_layout (pmu_layout_iter_fn fn,void *data);const struct pmu_events_table *find_sys_events_table (const char *name); const struct pmu_metrics_table *find_sys_metrics_table(const char*name);
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +0
-1--2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 03/15] perf pmu-events: Add functions in jevent.py to parse counter and event info for hardware aware grouping
Quoted 108 lines · Weilin
Quoted 103 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 9:49 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 03/15] perf pmu-events: Add functions injevent.pyto parse counter and event info for hardware aware groupingQuoted 11 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
These functions are added to parse event counter restrictions and counter availability info from json files so that the metric grouping method could do grouping based on the counter restriction of events and the counters that are available on the system.
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 185+++++++++++++++++++++++++++--
tools/perf/pmu-
Quoted 4 lines
tools/perf/pmu-events/pmu-events.h | 34 +++++- 2 files changed, 207 insertions(+), 12 deletions(-)tools/perf/pmu- renamed from tools/perf/pmu-events/jevents.py +0
-0
events/jevents.pyQuoted 42 lines
index
2c7e5d61ce92..bc91b7efa49a100755tools/perf/pmu-events/jevents.py +13
-0@@ -23,6 +23,8 @@ _metric_tables = [] _sys_metric_tables = [] # Mapping between sys event table names and sys metric table names. _sys_event_table_to_metric_table_mapping = {} +# List of regular PMU counter layout tables. +_pmu_layouts_tables = [] # Map from an event name to an architecture standard # JsonEvent. Architecture standard events are in json files in the top # f'{_args.starting_dir}/{_args.arch}' directory.@@ -31,6 +33,10 @@ _arch_std_events = {} _pending_events = [] # Name of events table to be written out _pending_events_tblname = None +# PMU counter layout to write out when the layout table is closed +_pending_pmu_counts = [] +# Name of PMU counter layout table to be written out +_pending_pmu_counts_tblname = None # Metrics to write out when the table is closed _pending_metrics = [] # Name of metrics table to be written out@@ -47,10 +53,17 @@ _json_event_attributes = [ 'event', # Short things in alphabetical order. 'compat', 'deprecated', 'perpkg', 'unit', + # The list of counter(s) this event could use + 'counters', # Longer things (the last won't be iterated over during decompress). 'long_desc' ] +# Attributes that are in pmu_unit_layout. +_json_layout_attributes = [ + 'pmu', 'desc', 'num_counters', 'num_fixed_counters' +] + # Attributes that are in pmu_metric rather than pmu_event. _json_metric_attributes = [ 'metric_name', 'metric_group', 'metric_expr', 'metric_threshold',tools/perf/pmu-events/jevents.py _json_metric_attributes = [ · line 61
Quoted 4 lines
tools/perf/pmu-events/jevents.py +0
-1@@ -58,7 +71,9 @@ _json_metric_attributes = [ 'default_metricgroup_name', 'aggr_mode', 'event_grouping' ] # Attributes that are bools or enum int values, encoded as '0', '1',... -_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping','perpkg']
tools/perf/pmu-events/jevents.py _json_metric_attributes = [ · line 74
Quoted 1 line
tools/perf/pmu-events/jevents.py +1
-0+_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping','perpkg',
Quoted 6 lines
tools/perf/pmu-events/jevents.py +2
-0+ 'num_counters', 'num_fixed_counters' +] def removesuffix(s: str, suffix: str) -> str: """Remove the suffix from a stringtools/perf/pmu-events/jevents.py class JsonEvent: · line 357
Quoted 9 lines
tools/perf/pmu-events/jevents.py +6
-0@@ -334,6 +349,12 @@ class JsonEvent: if 'Errata' in jd: extra_desc += ' Spec update: ' + jd['Errata'] self.pmu = unit_to_pmu(jd.get('Unit')) + # The list of counter(s) the event could be collected with + self.counters = jd.get('Counter') + # Number of generic counter + self.num_counters = jd.get('NumCounters') + # Number of fixed counter + self.num_fixed_counters = jd.get('NumFixedCounters')I'm not a fan of counters vs
fixed_counters, perhaps in the python:Quoted 4 lines
self.counters['generic'].list =
jd.get('Counter')self.counters['generic'].num =jd.get('NumCounters')self.counters['fixed'].num =jd.get('NumFixedCounters')
Quoted 8 lines
At this point of time, we cannot tell the content in 'Counter' would be for generic counter or fixed counter. Maybe we could do something like:
self.counters['list'] =jd.get('Counter')self.counters['num'].gp =jd.get('NumCounters')self.counters['num'].fixed =jd.get('NumFixedCounters')One difficulty I still have is to get
getattr()work with this type of attribute. Do you have some suggestion?
Thanks Weilin, the getattr stuff in JsonEvent was a bit of a hack to avoid repetitively typing out each attribute. It would be possible in the case of a counters dict to just hard code its existence. But I'd play around with the code and see what's cleanest. I generally find your python style better than mine :-)
Thanks,
Ian
Quoted 452 lines
Thanks,
WeilinQuoted 449 lines
Would make things a bit more intention revealing. Fwiw, the json name is worse as Counter is really for a plural of things. I don't think we can update this in the Intel perfmon events, but maybe we can do it in the
create_perf_json.pyconverter script and its generated json.This is pretty trivial as clean up so I'd prefer it before my Reviewed-by unless there's something I'm overlooking.
Thanks,
Iantools/perf/pmu-events/jevents.py def preprocess_arch_std_files(archpath: str) ->
Quoted 35 lines
tools/perf/pmu-events/jevents.py +11
-3filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')@@ -405,8 +426,16 @@ class JsonEvent: s += f'\t{attr} = {value},\n' return s + '}' - def build_c_string(self, metric: bool) -> str: + def build_c_string(self, metric: bool, layout: bool) -> str: s = '' + if layout: + for attr in _json_layout_attributes: + x = getattr(self, attr) + if attr in _json_enum_attributes: + s += x if x else '0' + else: + s += f'{x}\\000' if x else '\\000' + return s for attr in _json_metric_attributes if metric else _json_event_attributes: x = getattr(self, attr) if metric and x and attr == 'metric_expr':@@ -421,10 +450,10 @@ class JsonEvent: s += f'{x}\\000' if x else '\\000' return s - def to_c_string(self, metric: bool) -> str: + def to_c_string(self, metric: bool, layout: bool) -> str: """Representation of the event as a C struct initializer.""" - s = self.build_c_string(metric) + s = self.build_c_string(metric, layout) return f'{{ { _bcs.offsets[s] } }}, /* {s} */\n'@@ -461,6 +490,8 @@ def preprocess_arch_std_files(archpath: str) ->None:
tools/perf/pmu-events/jevents.py def add_events_table_entries(item: os.DirEntry,
Quoted 9 lines
tools/perf/pmu-events/jevents.py +2
-0_arch_std_events[event.name.lower()] = event if event.metric_name: _arch_std_events[event.metric_name.lower()] = event + if event.num_counters: + _arch_std_events[event.pmu.lower()] = event def add_events_table_entries(item: os.DirEntry, topic: str) -> None:@@ -470,6 +501,8 @@ def add_events_table_entries(item: os.DirEntry,topic: str) -> None:
tools/perf/pmu-events/jevents.py const struct pmu_table_entry
Quoted 27 lines
tools/perf/pmu-events/jevents.py +4
-2_pending_events.append(e) if e.metric_name: _pending_metrics.append(e) + if e.num_counters: + _pending_pmu_counts.append(e) def print_pending_events() -> None:@@ -510,7 +543,7 @@ def print_pending_events() -> None: last_pmu = event.pmu pmus.add((event.pmu, pmu_name)) - _args.output_file.write(event.to_c_string(metric=False)) + _args.output_file.write(event.to_c_string(metric=False, layout=False)) _pending_events = [] _args.output_file.write(f"""@@ -565,7 +598,7 @@ def print_pending_metrics() -> None: last_pmu = metric.pmu pmus.add((metric.pmu, pmu_name)) - _args.output_file.write(metric.to_c_string(metric=True)) + _args.output_file.write(metric.to_c_string(metric=True, layout=False)) _pending_metrics = [] _args.output_file.write(f"""@@ -583,6 +616,35 @@ const struct pmu_table_entry{_pending_metrics_tblname}[] = {{
Quoted 15 lines
tools/perf/pmu-events/jevents.py +12
-0""") _args.output_file.write('};\n\n') +def print_pending_pmu_counter_layout_table() -> None: + '''Print counter layout data from counter.json file to counter layout table in + c-string''' + + def pmu_counts_cmp_key(j: JsonEvent) -> Tuple[bool, str, str]: + def fix_none(s: Optional[str]) -> str: + if s is None: + return '' + return s + + return (j.desc is not None, fix_none(j.pmu), fix_none(j.num_counters)) +tools/perf/pmu-events/jevents.py const struct pmu_table_entry {_pending_metrics_tblname}[] = {{
Quoted 10 lines
tools/perf/pmu-events/jevents.py +10
-0+ global _pending_pmu_counts + if not _pending_pmu_counts: + return + + global _pending_pmu_counts_tblname + global pmu_layouts_tables + _pmu_layouts_tables.append(_pending_pmu_counts_tblname) + + _args.output_file.write( + f'static const struct compact_pmu_event{_pending_pmu_counts_tblname}[] = {{\n')
tools/perf/pmu-events/jevents.py const struct pmu_table_entry · line 642
Quoted 2 lines
tools/perf/pmu-events/jevents.py +2
-0+ + for pmu_layout in sorted(_pending_pmu_counts,key=pmu_counts_cmp_key):
tools/perf/pmu-events/jevents.py const struct pmu_table_entry · line 643
Quoted 1 line
tools/perf/pmu-events/jevents.py +1
-0+ _args.output_file.write(pmu_layout.to_c_string(metric=False,layout=True))
tools/perf/pmu-events/jevents.py def preprocess_one_file(parents: Sequence[str],
Quoted 8 lines
tools/perf/pmu-events/jevents.py +4
-0+ _pending_pmu_counts = [] + + _args.output_file.write('};\n\n') + def get_topic(topic: str) -> str: if topic.endswith('metrics.json'): return 'metrics'@@ -619,10 +681,12 @@ def preprocess_one_file(parents: Sequence[str],item: os.DirEntry) -> None:
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str],
Quoted 15 lines
tools/perf/pmu-events/jevents.py +4
-2pmu_name = f"{event.pmu}\\000" if event.name: _bcs.add(pmu_name, metric=False) - _bcs.add(event.build_c_string(metric=False), metric=False) + _bcs.add(event.build_c_string(metric=False, layout=False), metric=False) if event.metric_name: _bcs.add(pmu_name, metric=True) - _bcs.add(event.build_c_string(metric=True), metric=True) + _bcs.add(event.build_c_string(metric=True, layout=False), metric=True) + if event.num_counters: + _bcs.add(event.build_c_string(metric=False, layout=True), metric=False) def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: """Process a JSON file during the main walk."""@@ -636,11 +700,14 @@ def process_one_file(parents: Sequence[str],item: os.DirEntry) -> None:
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None:
Quoted 7 lines
tools/perf/pmu-events/jevents.py +1
-0if item.is_dir() and is_leaf_dir(item.path): print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() global _pending_events_tblname _pending_events_tblname = file_name_to_table_name('pmu_events_',parents, item.name)
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 708
Quoted 2 lines
tools/perf/pmu-events/jevents.py +0
-0global _pending_metrics_tblname _pending_metrics_tblname = file_name_to_table_name('pmu_metrics_',parents, item.name)
tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 710
Quoted 2 lines
tools/perf/pmu-events/jevents.py +2
-0+ global _pending_pmu_counts_tblname + _pending_pmu_counts_tblname =
file_name_to_table_name('pmu_layouts_', parents, item.name)tools/perf/pmu-events/jevents.py def process_one_file(parents: Sequence[str], · line 712
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0if item.name == 'sys': _sys_event_table_to_metric_table_mapping[_pending_events_tblname]=
_pending_metrics_tblnametools/perf/pmu-events/jevents.py struct pmu_metrics_table { · line 744
Quoted 5 lines
tools/perf/pmu-events/jevents.py +1
-0@@ -674,6 +741,12 @@ struct pmu_metrics_table { uint32_t num_pmus; }; +/* Struct used to make the PMU counter layout table implementationopaque to callers. */
tools/perf/pmu-events/jevents.py const struct pmu_events_map
Quoted 17 lines
tools/perf/pmu-events/jevents.py +6
-0+struct pmu_layouts_table { + const struct compact_pmu_event *entries; + size_t length; +}; + /* * Map a CPU to its table of PMU events. The CPU is identified by the * cpuid field, which is an arch-specific identifier for the CPU.@@ -687,6 +760,7 @@ struct pmu_events_map { const char *cpuid; struct pmu_events_table event_table; struct pmu_metrics_table metric_table; + struct pmu_layouts_table layout_table; }; /*@@ -731,6 +805,12 @@ const struct pmu_events_mappmu_events_map[] = {
tools/perf/pmu-events/jevents.py const struct pmu_events_map pmu_events_map[] = {
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0metric_size = '0' if event_size == '0' and metric_size == '0': continue + layout_tblname = file_name_to_table_name('pmu_layouts_', [],row[2].replace('/', '_'))
tools/perf/pmu-events/jevents.py const struct pmu_events_map
Quoted 9 lines
tools/perf/pmu-events/jevents.py +5
-0+ if layout_tblname in _pmu_layouts_tables: + layout_size = f'ARRAY_SIZE({layout_tblname})' + else: + layout_tblname = 'NULL' + layout_size = '0' cpuid = row[0].replace('\\', '\\\\') _args.output_file.write(f"""{{ \t.arch = "{arch}",@@ -742,6 +822,10 @@ const struct pmu_events_mappmu_events_map[] = {
tools/perf/pmu-events/jevents.py const struct pmu_events_map pmu_events_map[]
Quoted 11 lines
tools/perf/pmu-events/jevents.py +4
-0\t.metric_table = {{ \t\t.pmus = {metric_tblname}, \t\t.num_pmus = {metric_size} +\t}}, +\t.layout_table = {{ +\t\t.entries = {layout_tblname}, +\t\t.length = {layout_size} \t}} }}, """)@@ -752,6 +836,7 @@ const struct pmu_events_map pmu_events_map[]= {
tools/perf/pmu-events/jevents.py static void decompress_metric(int offset, struct
Quoted 8 lines
tools/perf/pmu-events/jevents.py +1
-0\t.cpuid = 0, \t.event_table = { 0, 0 }, \t.metric_table = { 0, 0 }, +\t.layout_table = { 0, 0 }, } }; """)@@ -840,6 +925,24 @@ static void decompress_metric(int offset, struct
pmu_metric*pm)Quoted 12 lines
tools/perf/pmu-events/jevents.py +9
-0_args.output_file.write('\twhile (*p++);') _args.output_file.write("""} +static void decompress_layout(int offset, struct pmu_layout *pm) +{ +\tconst char *p = &big_c_string[offset]; +""") + for attr in _json_layout_attributes: + _args.output_file.write(f'\n\tpm->{attr} = ') + if attr in _json_enum_attributes: + _args.output_file.write("*p - '0';\n") + else:tools/perf/pmu-events/jevents.py static void decompress_metric(int offset, struct pmu_metric *pm)
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ _args.output_file.write("(*p == '\\0' ? NULL : p);\n") + if attr == _json_layout_attributes[-1]: + continue + if attr in _json_enum_attributes: + _args.output_file.write('\tp++;') + else: + _args.output_file.write('\twhile (*p++);') + _args.output_file.write("""} + static int pmu_events_table__for_each_event_pmu(const struct
pmu_events_table*table,tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0const struct pmu_table_entry *pmu, pmu_event_iter_fn fn,@@ -995,6 +1098,21 @@ int pmu_metrics_table__for_each_metric(conststruct
pmu_metrics_table*table,tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table,
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0return 0; } +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table*table,
Quoted 5 lines
tools/perf/pmu-events/jevents.py +5
-0+ pmu_layout_iter_fn fn, + void *data) { + for (size_t i = 0; i < table->length; i++) { + struct pmu_layout pm; + int ret;tools/perf/pmu-events/jevents.py int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table,
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ + decompress_layout(table->entries[i].offset, &pm); + ret = fn(&pm, data); + if (ret) + return ret; + } + return 0; +} + static const struct pmu_events_map *map_for_pmu(struct perf_pmu*pmu)
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0{ static struct {@@ -1090,6 +1208,33 @@ const struct pmu_metrics_table*perf_pmu__find_metrics_table(struct
perf_pmu*pmtools/perf/pmu-events/jevents.py const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0return NULL; } +const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct
perf_pmu*pmu)Quoted 17 lines
tools/perf/pmu-events/jevents.py +17
-0+{ + const struct pmu_layouts_table *table = NULL; + char *cpuid = perf_pmu__getcpuid(pmu); + int i; + + /* on some platforms which uses cpus map, cpuid can be NULL for + * PMUs other than CORE PMUs. + */ + if (!cpuid) + return NULL; + + i = 0; + for (;;) { + const struct pmu_events_map *map = &pmu_events_map[i++]; + if (!map->arch) + break; +tools/perf/pmu-events/jevents.py const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct perf_pmu *pm
Quoted 10 lines
tools/perf/pmu-events/jevents.py +9
-0+ if (!strcmp_cpuid_str(map->cpuid, cpuid)) { + table = &map->layout_table; + break; + } + } + free(cpuid); + return table; +} + const struct pmu_events_table *find_core_events_table(const char *arch,const char *cpuid)
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table
Quoted 3 lines
tools/perf/pmu-events/jevents.py +0
-0{ for (const struct pmu_events_map *tables = &pmu_events_map[0];@@ -1111,6 +1256,16 @@ const struct pmu_metrics_table*find_core_metrics_table(const char *arch, const
tools/perf/pmu-events/jevents.py const struct pmu_metrics_table *find_core_metrics_table(const char *arch, const
Quoted 4 lines
tools/perf/pmu-events/jevents.py +1
-0} return NULL; } +const struct pmu_layouts_table *find_core_layouts_table(const char *arch,const char *cpuid)
tools/perf/pmu-events/jevents.py int
Quoted 14 lines
tools/perf/pmu-events/jevents.py +9
-0+{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + if (!strcmp(tables->arch, arch) && !strcmp_cpuid_str(tables- cpuid, cpuid)) + return &tables->layout_table; + } + return NULL; +} int pmu_for_each_core_event(pmu_event_iter_fn fn, void *data){
tools/perf/pmu-events/jevents.py +0
-0@@ -1139,6 +1294,19 @@ int
pmu_for_each_core_metric(pmu_metric_iter_fn fn, void *data)tools/perf/pmu-
Quoted 29 lines
tools/perf/pmu-events/jevents.py +13
-0return 0; } +int pmu_for_each_core_layout(pmu_layout_iter_fn fn, void *data) +{ + for (const struct pmu_events_map *tables = &pmu_events_map[0]; + tables->arch; + tables++) { + int ret = pmu_layouts_table__for_each_layout(&tables- layout_table, fn, data); + + if (ret) + return ret; + } + return 0; +} + const struct pmu_events_table *find_sys_events_table(const char *name) {for (const struct pmu_sys_events *tables = &pmu_sys_event_tables[0];tools/perf/pmu-events/jevents.py +1
-0@@ -1295,6 +1463,7 @@ struct pmu_table_entry { ftw(arch_path, [], process_one_file) print_pending_events() print_pending_metrics() + print_pending_pmu_counter_layout_table() print_mapping_table(archs) print_system_mapping_table()tools/perf/pmu- renamed from tools/perf/pmu-events/pmu-events.h +0
-0
events/pmu-events.htools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct
Quoted 37 lines
index
f5aa96f1685c..e245e4738970100644tools/perf/pmu-events/pmu-events.h +17
-0@@ -45,6 +45,11 @@ struct pmu_event {const char *desc;const char *topic;const char *long_desc; +/** +* The list of counter(s) the event could be collected on. +* eg., "0,1,2,3,4,5,6,7". +*/ +const char *counters;const char *pmu;const char *unit;bool perpkg;@@ -67,8 +72,20 @@ struct pmu_metric {enum metric_event_groups event_grouping; }; +struct pmu_layout { +const char *pmu; +const char *desc; +/** Total number of generic counters*/ +int num_counters; +/** +* Total number of fixed counters. +* Set to zero if no fixed counter on the unit.*/ +int num_fixed_counters; +}; +struct pmu_events_table ;struct pmu_metrics_table ; +struct pmu_layouts_table ; typedef int (*pmu_event_iter_fn)(const struct pmu_event *pe, const struct pmu_events_table *table,@@ -78,15 +95,21 @@ typedef int (*pmu_metric_iter_fn)(const struct
pmu_metric*pm,tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 7 lines
tools/perf/pmu-events/pmu-events.h +3
-0const struct pmu_metrics_table *table, void *data); +typedef int (*pmu_layout_iter_fn )(const struct pmu_layout *pm, +void *data); + int pmu_events_table__for_each_event(const struct pmu_events_table*table,
tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 13 lines
tools/perf/pmu-events/pmu-events.h +5
-4struct perf_pmu *pmu, pmu_event_iter_fn fn, void *data);int pmu_events_table__find_event (const struct pmu_events_table *table, -struct perf_pmu *pmu, -const char *name, -pmu_event_iter_fn fn, -void *data); +struct perf_pmu *pmu, +const char *name, +pmu_event_iter_fn fn, +void *data); +int pmu_layouts_table__for_each_layout(const struct pmu_layouts_table*table,
tools/perf/pmu-events/pmu-events.h typedef int (*pmu_metric_iter_fn)(const struct pmu_metric *pm,
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +2
-0+pmu_layout_iter_fn fn, + void *data); size_t pmu_events_table__num_events(const struct pmu_events_table*table,
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +0
-0struct perf_pmu *pmu);@@ -95,10 +118,13 @@ int pmu_metrics_table__for_each_metric(conststruct
pmu_metrics_table*table, pmtools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 118
Quoted 2 lines
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_events_table *perf_pmu__find_events_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 119
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_metrics_table *perf_pmu__find_metrics_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 120
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +1
-0+const struct pmu_layouts_table *perf_pmu__find_layouts_table(struct
perf_pmu*pmu);tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 121
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_events_table *find_core_events_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 122
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +0
-0const struct pmu_metrics_table *find_core_metrics_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const · line 123
Quoted 1 line
tools/perf/pmu-events/pmu-events.h +1
-0+const struct pmu_layouts_table *find_core_layouts_table(const char *arch,const char *cpuid);
tools/perf/pmu-events/pmu-events.h int pmu_metrics_table__for_each_metric(const struct pmu_metrics_table *table, pm
Quoted 6 lines
tools/perf/pmu-events/pmu-events.h +1
-0int pmu_for_each_core_event (pmu_event_iter_fn fn,void *data);int pmu_for_each_core_metric (pmu_metric_iter_fn fn,void *data); +int pmu_for_each_core_layout (pmu_layout_iter_fn fn,void *data);const struct pmu_events_table *find_sys_events_table (const char *name); const struct pmu_metrics_table *find_sys_metrics_table(const char*name);
Quoted 3 lines
tools/perf/pmu-events/pmu-events.h +0
-1--2.42.0
RFC PATCH v4 4/15 find_bit: add _find_last_and_bit() to support finding the most significant set bit
In reply to [email protected]From: Weilin Wang <[email protected]>
This function is required for more efficient PMU counter assignment.
When we use bitmap to log available PMU counters and counters that support a given event, we want to find a most significant set bit so that we could starting assigning counters with larger index first. This is helpful because counters with smaller indexes usually are more generic and support more events.
--- tools/include/linux/find.h | 18 ++++++++++++++++++ tools/lib/find_bit.c | 33 +++++++++++++++++++++++++++++++++ 2 files changed, 51 insertions(+)
tools/include/linux/find.h +18 -0
@@ -18,6 +18,8 @@ extern unsigned long _find_first_bit(const unsigned long *addr, unsigned long siextern unsigned long _find_first_and_bit (const unsigned long *addr1,const unsigned long *addr2,unsigned long size);extern unsigned long _find_first_zero_bit (const unsigned long *addr,unsigned long size); +extern unsigned long _find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2,unsigned long size);#ifndef find_next_bit/**
@@ -174,4 +176,20 @@ unsigned long find_first_zero_bit(const unsigned long *addr, unsigned long size) }#endif +#ifndef find_last_and_bit +static inline +unsigned long find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2, +unsigned long size) +{ +if (small_const_nbits (size)) { +unsigned long val = *addr1 & *addr2 &GENMASK (size -1 ,0 ); + +return val ?__fls (val) : size; + } + +return _find_last_and_bit (addr1, addr2, size); +} +#endif +#endif /*__LINUX_FIND_H_ */
tools/lib/find_bit.c +33 -0
@@ -67,6 +67,27 @@ out: \ sz; \ }) +/* +* Common helper for find_bit() function family +* @FETCH: The expression that fetches and pre-processes each word of bitmap(s) +* @MUNGE: The expression that post-processes a word containing found bit (may be empty) +* @size: The bitmap size in bits +*/ +#define FIND_LAST_BIT (FETCH ,MUNGE , size) \ +({ \ + unsigned long idx, val, sz = (size); \ + \ + for (idx = ((size - 1) / BITS_PER_LONG); idx >= 0; idx--) { \ + val = (FETCH); \ + if (val) { \ + sz = min(idx * BITS_PER_LONG + __fls(MUNGE(val)), sz); \ + break; \ + } \ + } \ + \ + sz; \ +}) +#ifndef find_first_bit/* * Find the first set bit in a memory region.
@@ -121,3 +142,15 @@ unsigned long _find_next_zero_bit(const unsigned long *addr, unsigned long nbits return FIND_NEXT_BIT(~addr[idx], /* nop */, nbits, start); }#endif + +#ifndef find_last_and_bit +/* +* Find the last set bit in two memory regions. +*/ +unsigned long _find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2, +unsigned long size) +{ + return FIND_LAST_BIT(addr1[idx] & addr2[idx], /* nop */, size); +} +#endif
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 04/15] find_bit: add _find_last_and_bit() to support finding the most significant set bit
Quoted 87 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
This function is required for more efficient PMU counter assignment.
When we use bitmap to log available PMU counters and counters that support a given event, we want to find a most significant set bit so that we could starting assigning counters with larger index first. This is helpful because counters with smaller indexes usually are more generic and support more events.
Signed-off-by Weilin Wang <[email protected]>--- tools/include/linux/find.h | 18 ++++++++++++++++++ tools/lib/find_bit.c | 33 +++++++++++++++++++++++++++++++++ 2 files changed, 51 insertions(+)tools/include/linux/find.h +18
-0@@ -18,6 +18,8 @@ extern unsigned long _find_first_bit(const unsigned long *addr, unsigned long siextern unsigned long _find_first_and_bit (const unsigned long *addr1,const unsigned long *addr2,unsigned long size);extern unsigned long _find_first_zero_bit (const unsigned long *addr,unsigned long size); +extern unsigned long _find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2,unsigned long size);#ifndef find_next_bit/** @@ -174,4 +176,20 @@ unsigned long find_first_zero_bit(const unsigned long *addr, unsigned long size) }#endif +#ifndef find_last_and_bit +static inline +unsigned long find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2, +unsigned long size) +{ +if (small_const_nbits (size)) { +unsigned long val = *addr1 & *addr2 &GENMASK (size -1 ,0 ); + +return val ?__fls (val) : size; + } + +return _find_last_and_bit (addr1, addr2, size); +} +#endif +#endif /*__LINUX_FIND_H_ */ tools/lib/find_bit.c +21
-0@@ -67,6 +67,27 @@ out: \ sz; \ }) +/* +* Common helper for find_bit() function family +* @FETCH: The expression that fetches and pre-processes each word of bitmap(s) +* @MUNGE: The expression that post-processes a word containing found bit (may be empty) +* @size: The bitmap size in bits +*/ +#define FIND_LAST_BIT (FETCH ,MUNGE , size) \ +({ \ + unsigned long idx, val, sz = (size); \ + \ + for (idx = ((size - 1) / BITS_PER_LONG); idx >= 0; idx--) { \ + val = (FETCH); \ + if (val) { \ + sz = min(idx * BITS_PER_LONG + __fls(MUNGE(val)), sz); \ + break; \ + } \ + } \ + \ + sz; \ +}) +#ifndef find_first_bit/* * Find the first set bit in a memory region. @@ -121,3 +142,15 @@ unsigned long _find_next_zero_bit(const unsigned long *addr, unsigned long nbits return FIND_NEXT_BIT(~addr[idx], /* nop */, nbits, start); }#endif
tools/lib/find_bit.c unsigned long _find_next_zero_bit(const unsigned long *addr, unsigned long nbits · line 154
Quoted 10 lines
tools/lib/find_bit.c +10
-0+ +#ifndef find_last_and_bit +/* +* Find the last set bit in two memory regions. +*/ +unsigned long _find_last_and_bit (const unsigned long *addr1, +const unsigned long *addr2, +unsigned long size) +{ + return FIND_LAST_BIT(addr1[idx] & addr2[idx], /* nop */, size);
The "/* nop */" argument is weird but existing style.
Thanks,
Ian
Quoted 6 lines
tools/lib/find_bit.c +2
-1+} +#endif --2.42.0
RFC PATCH v4 5/15 perf stat: Add functions to set counter bitmaps for hardware-grouping method
In reply to [email protected]From: Weilin Wang <[email protected]>
Add metricgroup__event_info data structure to represent an event in the metric grouping context; the list of counters and the PMU name an event should be collected with.
Add functions to parse event counter info from pmu-events and generate a list of metricgroup__event_info data to prepare grouping.
--- tools/perf/util/metricgroup.c | 211 +++++++++++++++++++++++++++++++++- 1 file changed, 208 insertions(+), 3 deletions(-)
tools/perf/util/metricgroup.c +208 -3
@@ -24,6 +24,7 @@#include <linux/list_sort.h> #include <linux/string.h> #include <linux/zalloc.h> +#include <linux/bitmap.h> #include <perf/cpumap.h> #include <subcmd/parse-options.h> #include <api/fs/fs.h>
@@ -157,6 +158,27 @@ struct metric {struct evlist *evlist; }; +/* Maximum number of counters per PMU*/ +#define NR_COUNTERS 16 + +/** +* An event used in a metric. This info is for metric grouping. +*/ +struct metricgroup__event_info { +struct list_head nd; +/** The name of the event. */ +const char *name; +/** The name of the pmu the event be collected on. */ +const char *pmu_name; +/** The event uses fixed counter*/ +bool fixed_counter; +/** The event uses special counters that we consider that as free counter +* during the event grouping*/ +bool free_counter; +/** The counters the event allowed to be collected on. */ + DECLARE_BITMAP(counters, NR_COUNTERS); +}; +/** * Each group is one node in the group string list. */
@@ -1440,6 +1462,175 @@ static int build_combined_expr_ctx(const struct list_head *metric_list,return ret; } +/** +* set_counter_bitmap - The counter bitmap: [0-15]. +*/ +static int set_counter_bitmap (int pos,unsigned long *bitmap) +{ +if (pos >=NR_COUNTERS || pos <0 ) +return -EINVAL ; +__set_bit (pos, bitmap); +return 0 ; +} + +static int parse_fixed_counter (const char *counter, +unsigned long *bitmap, +bool *fixed) +{ +int ret = -ENOENT ; +//TODO: this pattern is different on some other platforms +const char *pattern ="Fixed counter " ; +int pos =0 ; + +if (!strncmp (counter, pattern,strlen (pattern))) { + pos =atoi (counter +strlen (pattern)); + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + *fixed = true; +return 0 ; + } +return ret; +} + +/** +* parse_counter - Parse event counter info from pmu-events and set up bitmap +* accordingly. +* +* @counter: counter info string to be parsed. +* @bitmap: bitmap to set based on counter info parsed. +* @fixed: is set to true if the event uses fixed counter. +*/ +static int parse_counter (const char *counter, +unsigned long *bitmap, +bool *fixed) +{ +int ret =0 ; +char *p; +char *tok; +int pos =0 ; + + ret =parse_fixed_counter (counter, bitmap, fixed); +// ret==0 means matched with fixed counter +if (ret ==0 ) +return ret; + + p =strdup (counter); + tok =strtok (p,"," ); +if (!tok) +return -ENOENT ; + +while (tok) { + pos =atoi (tok); + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + tok =strtok (NULL ,"," ); + } +return 0 ; +} + +static struct metricgroup__event_info *event_info__new (const char *name, +const char *pmu_name, +const char *counter, +bool free_counter) +{ +int ret =0 ; +char *bit_buf =malloc (NR_COUNTERS ); +bool fixed_counter = false; +struct metricgroup__event_info *e; + + e =zalloc (sizeof (*e)); +if (!e) +return NULL ; +if (!pmu_name) + pmu_name ="core" ; + + e->name = name; + e->free_counter = free_counter; + e->pmu_name =strdup (pmu_name); +if (free_counter) { + ret =set_counter_bitmap (0 , e->counters); +if (ret) +return NULL ; + }else { + ret =parse_counter (counter, e->counters, &fixed_counter); +if (ret) +return NULL ; + e->fixed_counter = fixed_counter; + } + +bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); +pr_debug ("Event %s requires pmu %s counter: %s bitmap %s, [pmu=%s]\n" , + e->name, e->pmu_name, counter, bit_buf, pmu_name); + +return e; +} + +struct metricgroup__add_metric_event_data { +struct list_head *list; +/* pure event name, exclude umask and other info*/ +const char *event_name; +/* event name and umask if applicable*/ +const char *event_id; +}; + +static int metricgroup__add_metric_event_callback (const struct pmu_event *pe, +const struct pmu_events_table *table__maybe_unused , +void *data) +{ +struct metricgroup__event_info *event; +struct metricgroup__add_metric_event_data *d = data; + +if (!strcasecmp (pe->name, d->event_name)) { + event =event_info__new (d->event_id, pe->pmu, pe->counter,/*free_counter=*/ false); +if (!event) +return -ENOMEM ; +list_add (&event->nd, d->list); + } + +return 0 ; +} + +/** +* get_metricgroup_events - Find counter requirement of events from the +* pmu_events table +* @full_id: the full event identifiers. +* @table: pmu_events table that is searched for event data. +* @event_info_list: the list that the new event counter info added to. +*/ +static int get_metricgroup_events (const char *full_id, +const struct pmu_events_table *table, +struct list_head *event_info_list) +{ +LIST_HEAD (list); +int ret =0 ; +const char *id; +const char *rsep, *sep =strchr (full_id,'@' ); + +if (sep) { + rsep =strchr (full_id,',' ); + id =strndup (sep +1 , rsep - sep -1 ); +if (ret) + goto out; + }else { + id = full_id; + } + { +struct metricgroup__add_metric_event_data data = { + .list = &list, + .event_name = id, + .event_id = full_id, + }; + ret =pmu_events_table_for_each_event (table, + metricgroup__add_metric_event_callback, &data); + } + +out: +list_splice (&list, event_info_list); +return ret; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from
@@ -1453,9 +1644,25 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, const char *modifier __maybe_unused) {int ret =0 ; +struct hashmap_entry *cur; +LIST_HEAD (pmu_info_list); +LIST_HEAD (event_info_list); +size_t bkt; +const struct pmu_events_table *etable =perf_pmu__find_events_table (NULL ); + +#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) +hashmap__for_each_entry (ctx->ids, cur, bkt) { +const char *id = cur->pkey; + +pr_debug ("found event %s\n" , id); + + ret =get_metricgroup_events (id, etable, &event_info_list); +if (ret) +return ret; + } -pr_debug ("This is a placeholder\n" );return ret; +#undef RETURN_IF_NON_ZERO }static void group_str_free (struct metricgroup__group_strs *g)
@@ -1529,8 +1736,6 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, *out_evlist = parsed_evlist; parsed_evlist =NULL ; err_out: -parse_events_error__exit (&parse_error); -evlist__delete (parsed_evlist);metricgroup__free_grouping_strs (&groupings);return ret; }
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 05/15] perf stat: Add functions to set counter bitmaps for hardware-grouping method
Quoted 11 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add
metricgroup__event_infodata structure to represent an event in the metric grouping context; the list of counters and the PMU name an event should be collected with.Add functions to parse event counter info from pmu-events and generate a list of
metricgroup__event_infodata to prepare grouping.Signed-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 259 lines
--- tools/perf/util/metricgroup.c | 211 +++++++++++++++++++++++++++++++++- 1 file changed, 208 insertions(+), 3 deletions(-)tools/perf/util/metricgroup.c +208
-4@@ -24,6 +24,7 @@#include <linux/list_sort.h> #include <linux/string.h> #include <linux/zalloc.h> +#include <linux/bitmap.h> #include <perf/cpumap.h> #include <subcmd/parse-options.h> #include <api/fs/fs.h> @@ -157,6 +158,27 @@ struct metric {struct evlist *evlist; }; +/* Maximum number of counters per PMU*/ +#define NR_COUNTERS 16 + +/** +* An event used in a metric. This info is for metric grouping. +*/ +struct metricgroup__event_info { +struct list_head nd; +/** The name of the event. */ +const char *name; +/** The name of the pmu the event be collected on. */ +const char *pmu_name; +/** The event uses fixed counter*/ +bool fixed_counter; +/** The event uses special counters that we consider that as free counter +* during the event grouping*/ +bool free_counter; +/** The counters the event allowed to be collected on. */ + DECLARE_BITMAP(counters, NR_COUNTERS); +}; +/** * Each group is one node in the group string list. */ @@ -1440,6 +1462,175 @@ static int build_combined_expr_ctx(const struct list_head *metric_list,return ret; } +/** +* set_counter_bitmap - The counter bitmap: [0-15]. +*/ +static int set_counter_bitmap (int pos,unsigned long *bitmap) +{ +if (pos >=NR_COUNTERS || pos <0 ) +return -EINVAL ; +__set_bit (pos, bitmap); +return 0 ; +} + +static int parse_fixed_counter (const char *counter, +unsigned long *bitmap, +bool *fixed) +{ +int ret = -ENOENT ; +//TODO: this pattern is different on some other platforms +const char *pattern ="Fixed counter " ; +int pos =0 ; + +if (!strncmp (counter, pattern,strlen (pattern))) { + pos =atoi (counter +strlen (pattern)); + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + *fixed = true; +return 0 ; + } +return ret; +} + +/** +* parse_counter - Parse event counter info from pmu-events and set up bitmap +* accordingly. +* +* @counter: counter info string to be parsed. +* @bitmap: bitmap to set based on counter info parsed. +* @fixed: is set to true if the event uses fixed counter. +*/ +static int parse_counter (const char *counter, +unsigned long *bitmap, +bool *fixed) +{ +int ret =0 ; +char *p; +char *tok; +int pos =0 ; + + ret =parse_fixed_counter (counter, bitmap, fixed); +// ret==0 means matched with fixed counter +if (ret ==0 ) +return ret; + + p =strdup (counter); + tok =strtok (p,"," ); +if (!tok) +return -ENOENT ; + +while (tok) { + pos =atoi (tok); + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + tok =strtok (NULL ,"," ); + } +return 0 ; +} + +static struct metricgroup__event_info *event_info__new (const char *name, +const char *pmu_name, +const char *counter, +bool free_counter) +{ +int ret =0 ; +char *bit_buf =malloc (NR_COUNTERS ); +bool fixed_counter = false; +struct metricgroup__event_info *e; + + e =zalloc (sizeof (*e)); +if (!e) +return NULL ; +if (!pmu_name) + pmu_name ="core" ; + + e->name = name; + e->free_counter = free_counter; + e->pmu_name =strdup (pmu_name); +if (free_counter) { + ret =set_counter_bitmap (0 , e->counters); +if (ret) +return NULL ; + }else { + ret =parse_counter (counter, e->counters, &fixed_counter); +if (ret) +return NULL ; + e->fixed_counter = fixed_counter; + } + +bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); +pr_debug ("Event %s requires pmu %s counter: %s bitmap %s, [pmu=%s]\n" , + e->name, e->pmu_name, counter, bit_buf, pmu_name); + +return e; +} + +struct metricgroup__add_metric_event_data { +struct list_head *list; +/* pure event name, exclude umask and other info*/ +const char *event_name; +/* event name and umask if applicable*/ +const char *event_id; +}; + +static int metricgroup__add_metric_event_callback (const struct pmu_event *pe, +const struct pmu_events_table *table__maybe_unused , +void *data) +{ +struct metricgroup__event_info *event; +struct metricgroup__add_metric_event_data *d = data; + +if (!strcasecmp (pe->name, d->event_name)) { + event =event_info__new (d->event_id, pe->pmu, pe->counter,/*free_counter=*/ false); +if (!event) +return -ENOMEM ; +list_add (&event->nd, d->list); + } + +return 0 ; +} + +/** +* get_metricgroup_events - Find counter requirement of events from the +* pmu_events table +* @full_id: the full event identifiers. +* @table: pmu_events table that is searched for event data. +* @event_info_list: the list that the new event counter info added to. +*/ +static int get_metricgroup_events (const char *full_id, +const struct pmu_events_table *table, +struct list_head *event_info_list) +{ +LIST_HEAD (list); +int ret =0 ; +const char *id; +const char *rsep, *sep =strchr (full_id,'@' ); + +if (sep) { + rsep =strchr (full_id,',' ); + id =strndup (sep +1 , rsep - sep -1 ); +if (ret) + goto out; + }else { + id = full_id; + } + { +struct metricgroup__add_metric_event_data data = { + .list = &list, + .event_name = id, + .event_id = full_id, + }; + ret =pmu_events_table_for_each_event (table, + metricgroup__add_metric_event_callback, &data); + } + +out: +list_splice (&list, event_info_list); +return ret; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from @@ -1453,9 +1644,25 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, const char *modifier __maybe_unused) {int ret =0 ; +struct hashmap_entry *cur; +LIST_HEAD (pmu_info_list); +LIST_HEAD (event_info_list); +size_t bkt; +const struct pmu_events_table *etable =perf_pmu__find_events_table (NULL ); + +#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) +hashmap__for_each_entry (ctx->ids, cur, bkt) { +const char *id = cur->pkey; + +pr_debug ("found event %s\n" , id); + + ret =get_metricgroup_events (id, etable, &event_info_list); +if (ret) +return ret; + } -pr_debug ("This is a placeholder\n" );return ret; +#undef RETURN_IF_NON_ZERO }static void group_str_free (struct metricgroup__group_strs *g)@@ -1529,8 +1736,6 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, *out_evlist = parsed_evlist; parsed_evlist =NULL ; err_out: -parse_events_error__exit (&parse_error); -evlist__delete (parsed_evlist);metricgroup__free_grouping_strs (&groupings);return ret; } --2.42.0
RFC PATCH v4 6/15 perf stat: Add functions to get counter info
In reply to [email protected]From: Weilin Wang <[email protected]>
Add data structure metricgroup__pmu_counters to represent hardware counters available in the system.
Add functions to parse pmu-events and create the list of pmu_info_list to hold the counter information of the system.
Add functions to free pmu_info_list and event_info_list before exit grouping for hardware-grouping method
This method would fall back to normal grouping when event json files do not support hardware aware grouping.
--- tools/perf/util/metricgroup.c | 101 ++++++++++++++++++++++++++++++++-- 1 file changed, 97 insertions(+), 4 deletions(-)
tools/perf/util/metricgroup.c +97 -4
@@ -179,6 +179,20 @@ struct metricgroup__event_info {DECLARE_BITMAP (counters,NR_COUNTERS ); }; +/** +* A node is the counter availability of a pmu. +* This info is built up at the beginning from JSON file and +* used as a reference in metric grouping process. +*/ +struct metricgroup__pmu_counters { +struct list_head nd; +/** The name of the pmu the event collected on. */ +const char *name; +/** The number of gp counters in the pmu. */ +size_t num_counters; +size_t num_fixed_counters; +}; +/** * Each group is one node in the group string list. */
@@ -1530,6 +1544,27 @@ static int parse_counter(const char *counter,return 0 ; } +static void metricgroup__free_event_info (struct list_head + *event_info_list) +{ +struct metricgroup__event_info *e, *tmp; + +list_for_each_entry_safe (e, tmp, event_info_list, nd) { +list_del_init (&e->nd); +free (e); + } +} + +static void metricgroup__free_pmu_info (struct list_head *pmu_info_list) +{ +struct metricgroup__pmu_counters *p, *tmp; + +list_for_each_entry_safe (p, tmp, pmu_info_list, nd) { +list_del_init (&p->nd); +free (p); + } +} + static struct metricgroup__event_info *event_info__new(const char *name, const char *pmu_name, const char *counter,
@@ -1548,7 +1583,7 @@ static struct metricgroup__event_info *event_info__new(const char *name, e->name = name; e->free_counter = free_counter; - e->pmu_name =strdup (pmu_name); + e->pmu_name = pmu_name;if (free_counter) { ret =set_counter_bitmap (0 , e->counters);if (ret)
@@ -1583,7 +1618,9 @@ static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,struct metricgroup__add_metric_event_data *d = data;if (!strcasecmp (pe->name, d->event_name)) { - event =event_info__new (d->event_id, pe->pmu, pe->counter,/*free_counter=*/ false); +if (!pe->counters) +return -EINVAL ; + event =event_info__new (d->event_id, pe->pmu, pe->counters,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);
@@ -1622,7 +1659,7 @@ static int get_metricgroup_events(const char *full_id, .event_name = id, .event_id = full_id, }; - ret =pmu_events_table_for_each_event (table, + ret =pmu_events_table__for_each_event (table,/*pmu=*/ NULL , metricgroup__add_metric_event_callback, &data); }
@@ -1631,6 +1668,57 @@ static int get_metricgroup_events(const char *full_id,return ret; } +static struct metricgroup__pmu_counters *pmu_layout__new (const struct pmu_layout *pl) +{ +struct metricgroup__pmu_counters *l; + + l =zalloc (sizeof (*l)); + +if (!l) +return NULL ; + + l->name = pl->pmu; + l->num_counters = pl->num_counters; + l->num_fixed_counters = pl->num_fixed_counters; +pr_debug ("create new pmu_layout: [pmu]=%s, [gp_size]=%ld, [fixed_size]=%ld\n" , + l->name, l->num_counters, l->num_fixed_counters); +return l; +} + +static int metricgroup__add_pmu_layout_callback (const struct pmu_layout *pl, +void *data) +{ +struct metricgroup__pmu_counters *pmu; +struct list_head *d = data; +int ret =0 ; + + pmu =pmu_layout__new (pl); +if (!pmu) +return -ENOMEM ; +list_add (&pmu->nd, d); +return ret; +} + +/** +* get_pmu_counter_layouts - Find counter info of the architecture from +* the pmu_layouts table +* @pmu_info_list: the list that the new counter info of a pmu is added to. +* @table: pmu_layouts table that is searched for counter info. +*/ +static int get_pmu_counter_layouts (struct list_head *pmu_info_list, +const struct pmu_layouts_table + *table) +{ +LIST_HEAD (list); +int ret; + + ret =pmu_layouts_table__for_each_layout (table, + metricgroup__add_pmu_layout_callback, &list); + +list_splice (&list, pmu_info_list); +return ret; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from
@@ -1649,6 +1737,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,LIST_HEAD (event_info_list);size_t bkt;const struct pmu_events_table *etable =perf_pmu__find_events_table (NULL ); +const struct pmu_layouts_table *ltable =perf_pmu__find_layouts_table (NULL );#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0)hashmap__for_each_entry (ctx->ids, cur, bkt) {
@@ -1658,9 +1747,13 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, ret =get_metricgroup_events (id, etable, &event_info_list);if (ret) -return ret; + goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable); +err_out: +metricgroup__free_event_info (&event_info_list); +metricgroup__free_pmu_info (&pmu_info_list);return ret;#undef RETURN_IF_NON_ZERO }
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 06/15] perf stat: Add functions to get counter info
Quoted 9 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add data structure
metricgroup__pmu_countersto represent hardware counters available in the system.Add functions to parse pmu-events and create the list of
pmu_info_listto hold the counter information of the system.
Quoted 7 lines
Add functions to free
pmu_info_listandevent_info_listbefore exit grouping for hardware-grouping methodThis method would fall back to normal grouping when event json files do not support hardware aware grouping.
Signed-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 170 lines
--- tools/perf/util/metricgroup.c | 101 ++++++++++++++++++++++++++++++++-- 1 file changed, 97 insertions(+), 4 deletions(-)tools/perf/util/metricgroup.c +97
-5@@ -179,6 +179,20 @@ struct metricgroup__event_info {DECLARE_BITMAP (counters,NR_COUNTERS ); }; +/** +* A node is the counter availability of a pmu. +* This info is built up at the beginning from JSON file and +* used as a reference in metric grouping process. +*/ +struct metricgroup__pmu_counters { +struct list_head nd; +/** The name of the pmu the event collected on. */ +const char *name; +/** The number of gp counters in the pmu. */ +size_t num_counters; +size_t num_fixed_counters; +}; +/** * Each group is one node in the group string list. */ @@ -1530,6 +1544,27 @@ static int parse_counter(const char *counter,return 0 ; } +static void metricgroup__free_event_info (struct list_head + *event_info_list) +{ +struct metricgroup__event_info *e, *tmp; + +list_for_each_entry_safe (e, tmp, event_info_list, nd) { +list_del_init (&e->nd); +free (e); + } +} + +static void metricgroup__free_pmu_info (struct list_head *pmu_info_list) +{ +struct metricgroup__pmu_counters *p, *tmp; + +list_for_each_entry_safe (p, tmp, pmu_info_list, nd) { +list_del_init (&p->nd); +free (p); + } +} + static struct metricgroup__event_info *event_info__new(const char *name, const char *pmu_name, const char *counter,@@ -1548,7 +1583,7 @@ static struct metricgroup__event_info *event_info__new(const char *name, e->name = name; e->free_counter = free_counter; - e->pmu_name =strdup (pmu_name); + e->pmu_name = pmu_name;if (free_counter) { ret =set_counter_bitmap (0 , e->counters);if (ret)@@ -1583,7 +1618,9 @@ static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,struct metricgroup__add_metric_event_data *d = data;if (!strcasecmp (pe->name, d->event_name)) { - event =event_info__new (d->event_id, pe->pmu, pe->counter,/*free_counter=*/ false); +if (!pe->counters) +return -EINVAL ; + event =event_info__new (d->event_id, pe->pmu, pe->counters,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);@@ -1622,7 +1659,7 @@ static int get_metricgroup_events(const char *full_id, .event_name = id, .event_id = full_id, }; - ret =pmu_events_table_for_each_event (table, + ret =pmu_events_table__for_each_event (table,/*pmu=*/ NULL , metricgroup__add_metric_event_callback, &data); }@@ -1631,6 +1668,57 @@ static int get_metricgroup_events(const char *full_id,return ret; } +static struct metricgroup__pmu_counters *pmu_layout__new (const struct pmu_layout *pl) +{ +struct metricgroup__pmu_counters *l; + + l =zalloc (sizeof (*l)); + +if (!l) +return NULL ; + + l->name = pl->pmu; + l->num_counters = pl->num_counters; + l->num_fixed_counters = pl->num_fixed_counters; +pr_debug ("create new pmu_layout: [pmu]=%s, [gp_size]=%ld, [fixed_size]=%ld\n" , + l->name, l->num_counters, l->num_fixed_counters); +return l; +} + +static int metricgroup__add_pmu_layout_callback (const struct pmu_layout *pl, +void *data) +{ +struct metricgroup__pmu_counters *pmu; +struct list_head *d = data; +int ret =0 ; + + pmu =pmu_layout__new (pl); +if (!pmu) +return -ENOMEM ; +list_add (&pmu->nd, d); +return ret; +} + +/** +* get_pmu_counter_layouts - Find counter info of the architecture from +* the pmu_layouts table +* @pmu_info_list: the list that the new counter info of a pmu is added to. +* @table: pmu_layouts table that is searched for counter info. +*/ +static int get_pmu_counter_layouts (struct list_head *pmu_info_list, +const struct pmu_layouts_table + *table) +{ +LIST_HEAD (list); +int ret; + + ret =pmu_layouts_table__for_each_layout (table, + metricgroup__add_pmu_layout_callback, &list); + +list_splice (&list, pmu_info_list); +return ret; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from @@ -1649,6 +1737,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,LIST_HEAD (event_info_list);size_t bkt;const struct pmu_events_table *etable =perf_pmu__find_events_table (NULL ); +const struct pmu_layouts_table *ltable =perf_pmu__find_layouts_table (NULL );#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0)hashmap__for_each_entry (ctx->ids, cur, bkt) {@@ -1658,9 +1747,13 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, ret =get_metricgroup_events (id, etable, &event_info_list);if (ret) -return ret; + goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable); +err_out: +metricgroup__free_event_info (&event_info_list); +metricgroup__free_pmu_info (&pmu_info_list);return ret;#undef RETURN_IF_NON_ZERO } --2.42.0
RFC PATCH v4 7/15 perf stat: Add functions to create new group and assign events into groups
In reply to [email protected]From: Weilin Wang <[email protected]>
Add struct metricgroup__pmu_group_list to hold the lists of groups from different PMUs. Each PMU has one separate list.
Add struct metricgroup__group as one node (one group in the grouping result) of the metricgroup__pmu_group_list. It uses two bitmaps to log counter availabilities(gp counters and fixed counters).
Add functions to create group and assign event into the groups based on the event restrictions (struct metricgroup__event_info) and counter availability (pmu_info_list and bitmaps). New group is inserted into the list of groups.
Add functions to handle counter bitmaps. Add functions do find and insert operations to handle inserting event into groups.
Add function to fill all bits of one counter bitmap. Add functions to create new groups when no counter is available in all the existing groups.
--- tools/perf/util/metricgroup.c | 296 ++++++++++++++++++++++++++++++++++ 1 file changed, 296 insertions(+)
tools/perf/util/metricgroup.c +296 -0
@@ -192,6 +192,41 @@ struct metricgroup__pmu_counters {size_t num_counters;size_t num_fixed_counters; }; +/** +* A list of event groups for this pmu. +* This is updated during the grouping. +*/ +struct metricgroup__pmu_group_list { +struct list_head nd; +/** The name of the pmu(/core) the events collected on. */ +const char *pmu_name; +/** The number of gp counters in the pmu(/core). */ +size_t num_counters; +/** The number of fixed counters in the pmu(/core) if applicable. */ +size_t num_fixed_counters; +/** Head to the list of groups using this pmu(/core)*/ +struct list_head group_head; +}; +/** +* This is one node in the metricgroup__pmu_group_list. +* It represents on group. +*/ +struct metricgroup__group { +struct list_head nd; +/** The bitmaps represent availability of the counters. +* They are updated once the corresponding counter is used by +* an event (event inserted into the group). +*/ + DECLARE_BITMAP(gp_counters, NR_COUNTERS); + DECLARE_BITMAP(fixed_counters, NR_COUNTERS); +/** Head to the list of event names in this group*/ +struct list_head event_head; +}; + +struct metricgroup__group_events { +struct list_head nd; +const char *event_name; +};/** * Each group is one node in the group string list.
@@ -1487,6 +1522,34 @@ static int set_counter_bitmap(int pos, unsigned long *bitmap)return 0 ; } +/** +* Returns 0 on success. Finds the last counter that is not used in pmu_counters +* and supports the event, included in event_counters. +*/ +static int find_counter_bitmap (const unsigned long *pmu_counters, +const unsigned long *event_counters, +unsigned long *bit) +{ +/*It is helpful to assign from the highest bit because some events can +*only be collected using GP0-3. +*/ +unsigned long find_bit =find_last_and_bit (pmu_counters, event_counters,NR_COUNTERS ); + +if (find_bit ==NR_COUNTERS ) +return -ERANGE ; + *bit = find_bit; +return 0 ; +} + +static int use_counter_bitmap (unsigned long *bitmap, +unsigned long find_bit) +{ +if (find_bit >=NR_COUNTERS ) +return -EINVAL ; +__clear_bit (find_bit, bitmap); +return 0 ; +} +static int parse_fixed_counter (const char *counter,unsigned long *bitmap,bool *fixed)
@@ -1544,6 +1607,38 @@ static int parse_counter(const char *counter,return 0 ; } +static void group_event_list_free (struct metricgroup__group *groups) +{ +struct metricgroup__group_events *e, *tmp; + +list_for_each_entry_safe (e, tmp, &groups->event_head, nd) { +list_del_init (&e->nd); +free (e); + } +} + +static void group_list_free (struct metricgroup__pmu_group_list *groups) +{ +struct metricgroup__group *g, *tmp; + +list_for_each_entry_safe (g, tmp, &groups->group_head, nd) { +list_del_init (&g->nd); +group_event_list_free (g); +free (g); + } +} + +static void metricgroup__free_group_list (struct list_head *groups) +{ +struct metricgroup__pmu_group_list *g, *tmp; + +list_for_each_entry_safe (g, tmp, groups, nd) { +list_del_init (&g->nd); +group_list_free (g); +free (g); + } +} +static void metricgroup__free_event_info (struct list_head *event_info_list) {
@@ -1719,6 +1814,203 @@ static int get_pmu_counter_layouts(struct list_head *pmu_info_list,return ret; } +static int fill_counter_bitmap (unsigned long *bitmap,int start,int size) +{ +int ret; + +bitmap_zero (bitmap,NR_COUNTERS ); + +for (int pos = start; pos < start + size; pos++) { + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + } +return 0 ; +} + +/** +* Find if there is a counter available for event e in current_group. If a +* counter is available, use this counter by filling the bit in the correct +* counter bitmap. Otherwise, return error (-ERANGE). +*/ +static int find_and_set_counters (struct metricgroup__event_info *e, +struct metricgroup__group *current_group) +{ +int ret; +unsigned long find_bit =0 ; + +if (e->free_counter) +return 0 ; +if (e->fixed_counter) { + ret =find_counter_bitmap (current_group->fixed_counters, e->counters, + &find_bit); +if (ret) +return ret; +pr_debug ("found counter for [event]=%s [e->fixed_counters]=%lu\n" , + e->name, *current_group->fixed_counters); + ret =use_counter_bitmap (current_group->fixed_counters, find_bit); + }else { + ret =find_counter_bitmap (current_group->gp_counters, e->counters, + &find_bit); +if (ret) +return ret; +pr_debug ("found counter for [event]=%s [e->gp_counters]=%lu\n" , + e->name, *current_group->gp_counters); + ret =use_counter_bitmap (current_group->gp_counters, find_bit); + } +return ret; +} + +static int _insert_event (struct metricgroup__event_info *e, +struct metricgroup__group *group) +{ +struct metricgroup__group_events *event =malloc (sizeof (struct metricgroup__group_events )); + +if (!event) +return -ENOMEM ; + event->event_name = e->name; +if (e->fixed_counter) +list_add (&event->nd, &group->event_head); +else +list_add_tail (&event->nd, &group->event_head); +return 0 ; +} + +/** +* Insert the new_group node at the end of the group list. +*/ +static int insert_new_group (struct list_head *head, +struct metricgroup__group *new_group, +size_t num_counters, +size_t num_fixed_counters) +{ +INIT_LIST_HEAD (&new_group->event_head); +fill_counter_bitmap (new_group->gp_counters,0 , num_counters); +fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); +list_add_tail (&new_group->nd, head); +return 0 ; +} + +/** +* Insert event e into a group capable to include it +* +*/ +static int insert_event_to_group (struct metricgroup__event_info *e, +struct metricgroup__pmu_group_list *pmu_group_head) +{ +struct metricgroup__group *g; +int ret; +struct list_head *head; + +list_for_each_entry (g, &pmu_group_head->group_head, nd) { + ret =find_and_set_counters (e, g); +if (!ret) {/* return if successfully find and set counter*/ + ret =_insert_event (e, g); +return ret; + } + } +/* +* We were not able to find an existing group to insert this event. +* Continue to create a new group and insert the event in it. +*/ + { +struct metricgroup__group *current_group = +malloc (sizeof (struct metricgroup__group )); + +if (!current_group) +return -ENOMEM ; +pr_debug ("create_new_group for [event] %s\n" , e->name); + + head = &pmu_group_head->group_head; + ret =insert_new_group (head, current_group, pmu_group_head->num_counters, + pmu_group_head->num_fixed_counters); +if (ret) +return ret; + ret =find_and_set_counters (e, current_group); +if (ret) +return ret; + ret =_insert_event (e, current_group); + } + +return ret; +} + +/** +* assign_event_grouping - Assign an event into a group. If existing group +* cannot include it, create a new group and insert the event to it. +*/ +static int assign_event_grouping (struct metricgroup__event_info *e, +struct list_head *pmu_info_list, +struct list_head *groups) +{ +int ret =0 ; + +struct metricgroup__pmu_group_list *g =NULL ; +struct metricgroup__pmu_group_list *pmu_group_head =NULL ; + +list_for_each_entry (g, groups, nd) { +if (!strcasecmp (g->pmu_name, e->pmu_name)) { +pr_debug ("found group for event %s in pmu %s\n" , e->name, g->pmu_name); + pmu_group_head = g; +break ; + } + } +if (!pmu_group_head) { +struct metricgroup__pmu_counters *p; + + pmu_group_head =malloc (sizeof (struct metricgroup__pmu_group_list )); +if (!pmu_group_head) +return -ENOMEM ; +INIT_LIST_HEAD (&pmu_group_head->group_head); +pr_debug ("create new group for event %s in pmu %s\n" , e->name, e->pmu_name); + pmu_group_head->pmu_name = e->pmu_name; +list_for_each_entry (p, pmu_info_list, nd) { +if (!strcasecmp (p->name, e->pmu_name)) { + pmu_group_head->num_counters = p->num_counters; + pmu_group_head->num_fixed_counters = p->num_fixed_counters; +break ; + } + } +list_add_tail (&pmu_group_head->nd, groups); + } + + ret =insert_event_to_group (e, pmu_group_head); +return ret; +} + +/** +* create_grouping - Create a list of groups and place all the events of +* event_info_list into these groups. +* @pmu_info_list: the list of PMU units info based on pmu-events data, used for +* creating new groups. +* @event_info_list: the list of events to be grouped. +* @groupings: the list of groups with events placed in. +* @modifier: any modifiers added to the events. +*/ +static int create_grouping (struct list_head *pmu_info_list, +struct list_head *event_info_list, +struct list_head *groupings__maybe_unused , +const char *modifier__maybe_unused ) +{ +int ret =0 ; +struct metricgroup__event_info *e; +LIST_HEAD (groups); +char *bit_buf =malloc (NR_COUNTERS ); + +//TODO: for each new core group, we should consider to add events that uses fixed counters +list_for_each_entry (e, event_info_list, nd) { +bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); +pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, + e->pmu_name, bit_buf); + ret =assign_event_grouping (e, pmu_info_list, &groups); +if (ret) + goto out; + } +out: +metricgroup__free_group_list (&groups); +return ret; +}; +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from
@@ -1750,6 +2042,10 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, goto err_out; } ret =get_pmu_counter_layouts (&pmu_info_list, ltable); +if (ret) + goto err_out; + ret =create_grouping (&pmu_info_list, &event_info_list, groupings, + modifier); err_out:metricgroup__free_event_info (&event_info_list);
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 07/15] perf stat: Add functions to create new group and assign events into groups
Quoted 15 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add struct
metricgroup__pmu_group_listto hold the lists of groups from different PMUs. Each PMU has one separate list.Add struct
metricgroup__groupas one node (one group in the grouping result) of themetricgroup__pmu_group_list. It uses two bitmaps to log counter availabilities(gp counters and fixed counters).Add functions to create group and assign event into the groups based on the event restrictions (struct
metricgroup__event_info) and counter availability (pmu_info_listand bitmaps). New group is inserted into the list of groups.
Quoted 7 lines
Add functions to handle counter bitmaps. Add functions do find and insert operations to handle inserting event into groups.
Add function to fill all bits of one counter bitmap. Add functions to create new groups when no counter is available in all the existing groups.
Signed-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 342 lines
--- tools/perf/util/metricgroup.c | 296 ++++++++++++++++++++++++++++++++++ 1 file changed, 296 insertions(+)tools/perf/util/metricgroup.c +296
-1@@ -192,6 +192,41 @@ struct metricgroup__pmu_counters {size_t num_counters;size_t num_fixed_counters; }; +/** +* A list of event groups for this pmu. +* This is updated during the grouping. +*/ +struct metricgroup__pmu_group_list { +struct list_head nd; +/** The name of the pmu(/core) the events collected on. */ +const char *pmu_name; +/** The number of gp counters in the pmu(/core). */ +size_t num_counters; +/** The number of fixed counters in the pmu(/core) if applicable. */ +size_t num_fixed_counters; +/** Head to the list of groups using this pmu(/core)*/ +struct list_head group_head; +}; +/** +* This is one node in the metricgroup__pmu_group_list. +* It represents on group. +*/ +struct metricgroup__group { +struct list_head nd; +/** The bitmaps represent availability of the counters. +* They are updated once the corresponding counter is used by +* an event (event inserted into the group). +*/ + DECLARE_BITMAP(gp_counters, NR_COUNTERS); + DECLARE_BITMAP(fixed_counters, NR_COUNTERS); +/** Head to the list of event names in this group*/ +struct list_head event_head; +}; + +struct metricgroup__group_events { +struct list_head nd; +const char *event_name; +};/** * Each group is one node in the group string list. @@ -1487,6 +1522,34 @@ static int set_counter_bitmap(int pos, unsigned long *bitmap)return 0 ; } +/** +* Returns 0 on success. Finds the last counter that is not used in pmu_counters +* and supports the event, included in event_counters. +*/ +static int find_counter_bitmap (const unsigned long *pmu_counters, +const unsigned long *event_counters, +unsigned long *bit) +{ +/*It is helpful to assign from the highest bit because some events can +*only be collected using GP0-3. +*/ +unsigned long find_bit =find_last_and_bit (pmu_counters, event_counters,NR_COUNTERS ); + +if (find_bit ==NR_COUNTERS ) +return -ERANGE ; + *bit = find_bit; +return 0 ; +} + +static int use_counter_bitmap (unsigned long *bitmap, +unsigned long find_bit) +{ +if (find_bit >=NR_COUNTERS ) +return -EINVAL ; +__clear_bit (find_bit, bitmap); +return 0 ; +} +static int parse_fixed_counter (const char *counter,unsigned long *bitmap,bool *fixed)@@ -1544,6 +1607,38 @@ static int parse_counter(const char *counter,return 0 ; } +static void group_event_list_free (struct metricgroup__group *groups) +{ +struct metricgroup__group_events *e, *tmp; + +list_for_each_entry_safe (e, tmp, &groups->event_head, nd) { +list_del_init (&e->nd); +free (e); + } +} + +static void group_list_free (struct metricgroup__pmu_group_list *groups) +{ +struct metricgroup__group *g, *tmp; + +list_for_each_entry_safe (g, tmp, &groups->group_head, nd) { +list_del_init (&g->nd); +group_event_list_free (g); +free (g); + } +} + +static void metricgroup__free_group_list (struct list_head *groups) +{ +struct metricgroup__pmu_group_list *g, *tmp; + +list_for_each_entry_safe (g, tmp, groups, nd) { +list_del_init (&g->nd); +group_list_free (g); +free (g); + } +} +static void metricgroup__free_event_info (struct list_head *event_info_list) {@@ -1719,6 +1814,203 @@ static int get_pmu_counter_layouts(struct list_head *pmu_info_list,return ret; } +static int fill_counter_bitmap (unsigned long *bitmap,int start,int size) +{ +int ret; + +bitmap_zero (bitmap,NR_COUNTERS ); + +for (int pos = start; pos < start + size; pos++) { + ret =set_counter_bitmap (pos, bitmap); +if (ret) +return ret; + } +return 0 ; +} + +/** +* Find if there is a counter available for event e in current_group. If a +* counter is available, use this counter by filling the bit in the correct +* counter bitmap. Otherwise, return error (-ERANGE). +*/ +static int find_and_set_counters (struct metricgroup__event_info *e, +struct metricgroup__group *current_group) +{ +int ret; +unsigned long find_bit =0 ; + +if (e->free_counter) +return 0 ; +if (e->fixed_counter) { + ret =find_counter_bitmap (current_group->fixed_counters, e->counters, + &find_bit); +if (ret) +return ret; +pr_debug ("found counter for [event]=%s [e->fixed_counters]=%lu\n" , + e->name, *current_group->fixed_counters); + ret =use_counter_bitmap (current_group->fixed_counters, find_bit); + }else { + ret =find_counter_bitmap (current_group->gp_counters, e->counters, + &find_bit); +if (ret) +return ret; +pr_debug ("found counter for [event]=%s [e->gp_counters]=%lu\n" , + e->name, *current_group->gp_counters); + ret =use_counter_bitmap (current_group->gp_counters, find_bit); + } +return ret; +} + +static int _insert_event (struct metricgroup__event_info *e, +struct metricgroup__group *group) +{ +struct metricgroup__group_events *event =malloc (sizeof (struct metricgroup__group_events )); + +if (!event) +return -ENOMEM ; + event->event_name = e->name; +if (e->fixed_counter) +list_add (&event->nd, &group->event_head); +else +list_add_tail (&event->nd, &group->event_head); +return 0 ; +} + +/** +* Insert the new_group node at the end of the group list. +*/ +static int insert_new_group (struct list_head *head, +struct metricgroup__group *new_group, +size_t num_counters, +size_t num_fixed_counters) +{ +INIT_LIST_HEAD (&new_group->event_head); +fill_counter_bitmap (new_group->gp_counters,0 , num_counters); +fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); +list_add_tail (&new_group->nd, head); +return 0 ; +} + +/** +* Insert event e into a group capable to include it +* +*/ +static int insert_event_to_group (struct metricgroup__event_info *e, +struct metricgroup__pmu_group_list *pmu_group_head) +{ +struct metricgroup__group *g; +int ret; +struct list_head *head; + +list_for_each_entry (g, &pmu_group_head->group_head, nd) { + ret =find_and_set_counters (e, g); +if (!ret) {/* return if successfully find and set counter*/ + ret =_insert_event (e, g); +return ret; + } + } +/* +* We were not able to find an existing group to insert this event. +* Continue to create a new group and insert the event in it. +*/ + { +struct metricgroup__group *current_group = +malloc (sizeof (struct metricgroup__group )); + +if (!current_group) +return -ENOMEM ; +pr_debug ("create_new_group for [event] %s\n" , e->name); + + head = &pmu_group_head->group_head; + ret =insert_new_group (head, current_group, pmu_group_head->num_counters, + pmu_group_head->num_fixed_counters); +if (ret) +return ret; + ret =find_and_set_counters (e, current_group); +if (ret) +return ret; + ret =_insert_event (e, current_group); + } + +return ret; +} + +/** +* assign_event_grouping - Assign an event into a group. If existing group +* cannot include it, create a new group and insert the event to it. +*/ +static int assign_event_grouping (struct metricgroup__event_info *e, +struct list_head *pmu_info_list, +struct list_head *groups) +{ +int ret =0 ; + +struct metricgroup__pmu_group_list *g =NULL ; +struct metricgroup__pmu_group_list *pmu_group_head =NULL ; + +list_for_each_entry (g, groups, nd) { +if (!strcasecmp (g->pmu_name, e->pmu_name)) { +pr_debug ("found group for event %s in pmu %s\n" , e->name, g->pmu_name); + pmu_group_head = g; +break ; + } + } +if (!pmu_group_head) { +struct metricgroup__pmu_counters *p; + + pmu_group_head =malloc (sizeof (struct metricgroup__pmu_group_list )); +if (!pmu_group_head) +return -ENOMEM ; +INIT_LIST_HEAD (&pmu_group_head->group_head); +pr_debug ("create new group for event %s in pmu %s\n" , e->name, e->pmu_name); + pmu_group_head->pmu_name = e->pmu_name; +list_for_each_entry (p, pmu_info_list, nd) { +if (!strcasecmp (p->name, e->pmu_name)) { + pmu_group_head->num_counters = p->num_counters; + pmu_group_head->num_fixed_counters = p->num_fixed_counters; +break ; + } + } +list_add_tail (&pmu_group_head->nd, groups); + } + + ret =insert_event_to_group (e, pmu_group_head); +return ret; +} + +/** +* create_grouping - Create a list of groups and place all the events of +* event_info_list into these groups. +* @pmu_info_list: the list of PMU units info based on pmu-events data, used for +* creating new groups. +* @event_info_list: the list of events to be grouped. +* @groupings: the list of groups with events placed in. +* @modifier: any modifiers added to the events. +*/ +static int create_grouping (struct list_head *pmu_info_list, +struct list_head *event_info_list, +struct list_head *groupings__maybe_unused , +const char *modifier__maybe_unused ) +{ +int ret =0 ; +struct metricgroup__event_info *e; +LIST_HEAD (groups); +char *bit_buf =malloc (NR_COUNTERS ); + +//TODO: for each new core group, we should consider to add events that uses fixed counters +list_for_each_entry (e, event_info_list, nd) { +bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); +pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, + e->pmu_name, bit_buf); + ret =assign_event_grouping (e, pmu_info_list, &groups); +if (ret) + goto out; + } +out: +metricgroup__free_group_list (&groups); +return ret; +}; +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from @@ -1750,6 +2042,10 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, goto err_out; } ret =get_pmu_counter_layouts (&pmu_info_list, ltable); +if (ret) + goto err_out; + ret =create_grouping (&pmu_info_list, &event_info_list, groupings, + modifier); err_out:metricgroup__free_event_info (&event_info_list); --2.42.0
RFC PATCH v4 8/15 perf stat: Add build string function and topdown events handling in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Add the function to generate final grouping strings. This function is very similar to the existing metricgroup__build_event_string() function. The difference is that the input data includes a list of grouping lists.
--- tools/perf/util/metricgroup.c | 97 +++++++++++++++++++++++++++++++++-- 1 file changed, 93 insertions(+), 4 deletions(-)
tools/perf/util/metricgroup.c +93 -4
@@ -1978,6 +1978,96 @@ static int assign_event_grouping(struct metricgroup__event_info *e,return ret; } +static int hw_aware_metricgroup__build_event_string (struct list_head *group_strs, +const char *modifier, +struct list_head *groups) +{ +struct metricgroup__pmu_group_list *p; +struct metricgroup__group *g; +struct metricgroup__group_events *ge; +bool no_group = true; +int ret =0 ; + +#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) + +list_for_each_entry (p, groups, nd) { +list_for_each_entry (g, &p->group_head, nd) { +struct strbuf *events; +struct metricgroup__group_strs *new_group_str = +malloc (sizeof (struct metricgroup__group_strs )); + +if (!new_group_str) +return -ENOMEM ; +strbuf_init (&new_group_str->grouping_str,0 ); + events = &new_group_str->grouping_str; + ret =strbuf_addch (events,'{' ); +RETURN_IF_NON_ZERO (ret); + no_group = true; +list_for_each_entry (ge, &g->event_head, nd) { +const char *sep, *rsep, *id = ge->event_name; + +pr_debug ("found event %s\n" , id); + +/* Separate events with commas and open the group if necessary. */ +if (!no_group) { + ret =strbuf_addch (events,',' ); +RETURN_IF_NON_ZERO (ret); + } +/* +* Encode the ID as an event string. Add a qualifier for +* metric_id that is the original name except with characters +* that parse-events can't parse replaced. For example, +* 'msr@tsc@' gets added as msr/tsc,metric-id=msr!3tsc!3/ +*/ + sep =strchr (id,'@' ); +if (sep) { + ret =strbuf_add (events, id, sep - id); +RETURN_IF_NON_ZERO (ret); + ret =strbuf_addch (events,'/' ); +RETURN_IF_NON_ZERO (ret); + rsep =strrchr (sep,'@' ); + ret =strbuf_add (events, sep +1 , rsep - sep -1 ); +RETURN_IF_NON_ZERO (ret); + ret =strbuf_addstr (events,",metric-id=" ); +RETURN_IF_NON_ZERO (ret); + sep = rsep; + }else { + sep =strchr (id,':' ); +if (sep) { + ret =strbuf_add (events, id, sep - id); +RETURN_IF_NON_ZERO (ret); + }else { + ret =strbuf_addstr (events, id); +RETURN_IF_NON_ZERO (ret); + } + ret =strbuf_addstr (events,"/metric-id=" ); +RETURN_IF_NON_ZERO (ret); + } + ret =encode_metric_id (events, id); +RETURN_IF_NON_ZERO (ret); + ret =strbuf_addstr (events,"/" ); +RETURN_IF_NON_ZERO (ret); + +if (sep) { + ret =strbuf_addstr (events, sep +1 ); +RETURN_IF_NON_ZERO (ret); + } +if (modifier) { + ret =strbuf_addstr (events, modifier); +RETURN_IF_NON_ZERO (ret); + } + no_group = false; + } + ret =strbuf_addf (events,"}:W" ); +RETURN_IF_NON_ZERO (ret); +pr_debug ("events-buf: %s\n" , events->buf); +list_add_tail (&new_group_str->nd, group_strs); + } + } +return ret; +#undef RETURN_IF_NON_ZERO +} +/** * create_grouping - Create a list of groups and place all the events of * event_info_list into these groups.
@@ -1989,8 +2079,8 @@ static int assign_event_grouping(struct metricgroup__event_info *e,*/ static int create_grouping (struct list_head *pmu_info_list,struct list_head *event_info_list, -struct list_head *groupings__maybe_unused , -const char *modifier__maybe_unused ) +struct list_head *groupings, +const char *modifier) {int ret =0 ;struct metricgroup__event_info *e;
@@ -2006,6 +2096,7 @@ static int create_grouping(struct list_head *pmu_info_list,if (ret) goto out; } + ret =hw_aware_metricgroup__build_event_string (groupings, modifier, &groups); out:metricgroup__free_group_list (&groups);return ret;
@@ -2035,8 +2126,6 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,hashmap__for_each_entry (ctx->ids, cur, bkt) {const char *id = cur->pkey; -pr_debug ("found event %s\n" , id); - ret =get_metricgroup_events (id, etable, &event_info_list);if (ret) goto err_out;
Signature · 1 line
2.42.0
RFC PATCH v4 9/15 perf stat: Add function to handle special events in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
There are some special events like topdown events and TSC that are not described in pmu-event JSON files. Add support to handle this type of events. This should be considered as a temporary solution because including these events in JSON files would be a better solution.
--- tools/perf/util/metricgroup.c | 38 ++++++++++++++++++++++++++++++++++- 1 file changed, 37 insertions(+), 1 deletion(-)
tools/perf/util/metricgroup.c +37 -1
@@ -160,6 +160,20 @@ struct metric {/* Maximum number of counters per PMU*/ #define NR_COUNTERS 16 +/* Special events that are not described in pmu-event JSON files. +* topdown-* and TSC use dedicated registers, set as free +* counter for grouping purpose +*/ +enum special_events { +TOPDOWN =0 , +TSC =1 , +SPECIAL_EVENT_MAX , +}; + +static const char *const special_event_names[SPECIAL_EVENT_MAX ] = { +"topdown-" , +"TSC" , +};/** * An event used in a metric. This info is for metric grouping.
@@ -2102,6 +2116,15 @@ static int create_grouping(struct list_head *pmu_info_list,return ret; }; +static bool is_special_event (const char *id) +{ +for (int i =0 ; i <SPECIAL_EVENT_MAX ; i++) { +if (!strncmp (id, special_event_names[i],strlen (special_event_names[i]))) +return true; + } +return false; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from
@@ -2126,6 +2149,17 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,hashmap__for_each_entry (ctx->ids, cur, bkt) {const char *id = cur->pkey; +if (is_special_event (id)) { +struct metricgroup__event_info *event; + + event =event_info__new (id,"default_core" ,"0" , +/*free_counter=*/ true); +if (!event) + goto err_out; + +list_add (&event->nd, &event_info_list); +continue ; + } ret =get_metricgroup_events (id, etable, &event_info_list);if (ret) goto err_out;
@@ -2597,8 +2631,10 @@ int metricgroup__parse_groups(struct evlist *perf_evlist, ret =hw_aware_parse_groups (perf_evlist, pmu, str, metric_no_threshold, user_requested_cpu_list, system_wide,/*fake_pmu=*/ NULL , metric_events, table); -if (!ret) +if (!ret) { +pr_info ("Hardware aware grouping completed\n" );return 0 ; + } } return parse_groups(perf_evlist, pmu, str, metric_no_group, metric_no_merge,
Signature · 1 line
2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 09/15] perf stat: Add function to handle special events in hardware-grouping
Quoted 18 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
There are some special events like topdown events and TSC that are not described in pmu-event JSON files. Add support to handle this type of events. This should be considered as a temporary solution because including these events in JSON files would be a better solution.
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/util/metricgroup.c | 38 ++++++++++++++++++++++++++++++++++- 1 file changed, 37 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +0
-0
tools/perf/util/metricgroup.c struct metric { · line 165
Quoted 6 lines
tools/perf/util/metricgroup.c +3
-0@@ -160,6 +160,20 @@ struct metric {/* Maximum number of counters per PMU*/ #define NR_COUNTERS 16 +/* Special events that are not described in pmu-event JSON files. +* topdown-* and TSC use dedicated registers, set as free +* counter for grouping purpose
msr/tsc/ is a software event where reading the value is done by rdtsc. Unlike tool events like duration_time we want msr/tsc/ in the group with the other hardware events so its running/enabled time scaling matches.
To some extent the topdown- events do exist in the json as "TOPDOWN.*". Looking at tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json I see just TOPDOWN.BACKEND_BOUND_SLOTS. Perhaps we can add the rest rather than have a workaround here?
If topdown events are in json and msr/tsc/ is treated like a software event, as we do here:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/parse-events.c?h=perf-tools-next#n1920
perhaps we don't need the special events category?
Thanks,
Ian
Quoted 63 lines
tools/perf/util/metricgroup.c +34
-2+*/ +enum special_events { +TOPDOWN =0 , +TSC =1 , +SPECIAL_EVENT_MAX , +}; + +static const char *const special_event_names[SPECIAL_EVENT_MAX ] = { +"topdown-" , +"TSC" , +};/** * An event used in a metric. This info is for metric grouping. @@ -2102,6 +2116,15 @@ static int create_grouping(struct list_head *pmu_info_list,return ret; }; +static bool is_special_event (const char *id) +{ +for (int i =0 ; i <SPECIAL_EVENT_MAX ; i++) { +if (!strncmp (id, special_event_names[i],strlen (special_event_names[i]))) +return true; + } +return false; +} +/** * hw_aware_build_grouping - Build event groupings by reading counter * requirement of the events and counter available on the system from @@ -2126,6 +2149,17 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,hashmap__for_each_entry (ctx->ids, cur, bkt) {const char *id = cur->pkey; +if (is_special_event (id)) { +struct metricgroup__event_info *event; + + event =event_info__new (id,"default_core" ,"0" , +/*free_counter=*/ true); +if (!event) + goto err_out; + +list_add (&event->nd, &event_info_list); +continue ; + } ret =get_metricgroup_events (id, etable, &event_info_list);if (ret) goto err_out;@@ -2597,8 +2631,10 @@ int metricgroup__parse_groups(struct evlist *perf_evlist, ret =hw_aware_parse_groups (perf_evlist, pmu, str, metric_no_threshold, user_requested_cpu_list, system_wide,/*fake_pmu=*/ NULL , metric_events, table); -if (!ret) +if (!ret) { +pr_info ("Hardware aware grouping completed\n" );return 0 ; + } } return parse_groups(perf_evlist, pmu, str, metric_no_group, metric_no_merge, --2.42.0
RFC PATCH v4 10/15 perf stat: Add function to combine metrics for hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
This function is very similar to the existing build_combined_expr_ctx(). Should be able to reuse current function instead of adding a new one. Will fix this later.
--- tools/perf/util/metricgroup.c | 47 ++++++++++++++++++++++++++++++++--- 1 file changed, 44 insertions(+), 3 deletions(-)
tools/perf/util/metricgroup.c +44 -3
@@ -1525,6 +1525,46 @@ static int build_combined_expr_ctx(const struct list_head *metric_list,return ret; } +/** +* hw_aware_build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events +* metric IDs, as the IDs are held in a set, +* duplicates will be removed. +* @metric_list: List to take metrics from. +* @combined: Out argument for result. +*/ +static int hw_aware_build_combined_expr_ctx (const struct list_head *metric_list, +struct expr_parse_ctx **combined) +{ +struct hashmap_entry *cur; +size_t bkt; +struct metric *m; +char *dup; +int ret; + + *combined =expr__ctx_new (); +if (!*combined) +return -ENOMEM ; + +list_for_each_entry (m, metric_list, nd) { +hashmap__for_each_entry (m->pctx->ids, cur, bkt) { +pr_debug2 ("metric: %s\n" , m->metric_name); + dup =strdup (cur->pkey); +if (!dup) { + ret = -ENOMEM ; + goto err_out; + } + ret =expr__add_id (*combined, dup); +if (ret) + goto err_out; + } + } +return 0 ; +err_out: +expr__ctx_free (*combined); + *combined =NULL ; +return ret; +} +/** * set_counter_bitmap - The counter bitmap: [0-15]. */
@@ -1815,8 +1855,7 @@ static int metricgroup__add_pmu_layout_callback(const struct pmu_layout *pl,* @table: pmu_layouts table that is searched for counter info. */ static int get_pmu_counter_layouts (struct list_head *pmu_info_list, -const struct pmu_layouts_table - *table) +const struct pmu_layouts_table *table) {LIST_HEAD (list);int ret;
@@ -2248,6 +2287,8 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, *out_evlist = parsed_evlist; parsed_evlist =NULL ; err_out: +parse_events_error__exit (&parse_error); +evlist__delete (parsed_evlist);metricgroup__free_grouping_strs (&groupings);return ret; }
@@ -2370,7 +2411,7 @@ static int hw_aware_parse_groups(struct evlist *perf_evlist,if (!metric_no_merge) {struct expr_parse_ctx *combined =NULL ; - ret =build_combined_expr_ctx (&metric_list, &combined); + ret =hw_aware_build_combined_expr_ctx (&metric_list, &combined);if (!ret && combined &&hashmap__size (combined->ids)) { ret = hw_aware_parse_ids(fake_pmu, combined,
Signature · 1 line
2.42.0
RFC PATCH v4 11/15 perf stat: Handle taken alone in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Add taken alone into consideration when grouping. Only one taken alone event is supported per group.
--- tools/perf/pmu-events/jevents.py | 7 ++++++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 20 +++++++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-)
tools/perf/pmu-events/jevents.py +6 -1
@@ -56,7 +56,9 @@ _json_event_attributes = [ # The list of counter(s) this event could use 'counters', # Longer things (the last won't be iterated over during decompress). - 'long_desc' + 'long_desc', + # Taken alone event could not be collected in the same group with other taken alone event + 'taken_alone' ] # Attributes that are in pmu_unit_layout.
@@ -355,6 +357,9 @@ class JsonEvent: self.num_counters = jd.get('NumCounters') # Number of fixed counter self.num_fixed_counters = jd.get('NumFixedCounters') + # If the event is taken alone event, which cannot be grouped with any other + # taken alone event. + self.taken_alone = jd.get('TakenAlone') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')
tools/perf/pmu-events/pmu-events.h +1 -0
@@ -54,6 +54,7 @@ struct pmu_event {const char *unit;bool perpkg;bool deprecated; +bool taken_alone; };struct pmu_metric {
tools/perf/util/metricgroup.c +15 -5
@@ -189,6 +189,7 @@ struct metricgroup__event_info {/** The event uses special counters that we consider that as free counter * during the event grouping*/ bool free_counter; +bool taken_alone;/** The counters the event allowed to be collected on. */ DECLARE_BITMAP (counters,NR_COUNTERS ); };
@@ -235,6 +236,7 @@ struct metricgroup__group {DECLARE_BITMAP (fixed_counters,NR_COUNTERS );/** Head to the list of event names in this group*/ struct list_head event_head; +bool taken_alone; };struct metricgroup__group_events {
@@ -1717,6 +1719,7 @@ static void metricgroup__free_pmu_info(struct list_head *pmu_info_list)static struct metricgroup__event_info *event_info__new (const char *name,const char *pmu_name,const char *counter, +bool taken_alone,bool free_counter) {int ret =0 ;
@@ -1731,6 +1734,7 @@ static struct metricgroup__event_info *event_info__new(const char *name, pmu_name ="core" ; e->name = name; + e->taken_alone = taken_alone; e->free_counter = free_counter; e->pmu_name = pmu_name;if (free_counter) {
@@ -1769,7 +1773,8 @@ static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,if (!strcasecmp (pe->name, d->event_name)) {if (!pe->counters)return -EINVAL ; - event =event_info__new (d->event_id, pe->pmu, pe->counters,/*free_counter=*/ false); + event =event_info__new (d->event_id, pe->pmu, pe->counters, + pe->taken_alone,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);
@@ -1892,6 +1897,8 @@ static int find_and_set_counters(struct metricgroup__event_info *e,int ret;unsigned long find_bit =0 ; +if (e->taken_alone && current_group->taken_alone) +return -ENOSPC ;if (e->free_counter)return 0 ;if (e->fixed_counter) {
@@ -1926,11 +1933,13 @@ static int _insert_event(struct metricgroup__event_info *e,list_add (&event->nd, &group->event_head); else list_add_tail(&event->nd, &group->event_head); +if (e->taken_alone) + group->taken_alone = true;return 0 ; }/** -* Insert the new_group node at the end of the group list. +* Initialize the new group and insert it to the end of the group list. */ static int insert_new_group(struct list_head *head, struct metricgroup__group *new_group,
@@ -1940,6 +1949,7 @@ static int insert_new_group(struct list_head *head,INIT_LIST_HEAD (&new_group->event_head);fill_counter_bitmap (new_group->gp_counters,0 , num_counters);fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); + new_group->taken_alone = false;list_add_tail (&new_group->nd, head);return 0 ; }
@@ -2143,8 +2153,8 @@ static int create_grouping(struct list_head *pmu_info_list,//TODO: for each new core group, we should consider to add events that uses fixed counters list_for_each_entry (e, event_info_list, nd) {bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); -pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, - e->pmu_name, bit_buf); +pr_debug ("Event name %s, [pmu]=%s, [counters]=%s, [taken_alone]=%d\n" , + e->name, e->pmu_name, bit_buf, e->taken_alone); ret =assign_event_grouping (e, pmu_info_list, &groups);if (ret) goto out;
@@ -2191,7 +2201,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,if (is_special_event (id)) {struct metricgroup__event_info *event; - event =event_info__new (id,"default_core" ,"0" , + event =event_info__new (id,"default_core" ,"0" , false,/*free_counter=*/ true);if (!event) goto err_out;
Signature · 1 line
2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware-grouping
Quoted 5 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add taken alone into consideration when grouping. Only one taken alone event is supported per group.
Is there an example of this?
Quoted 134 lines
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 7 ++++++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 20 +++++++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-)tools/perf/pmu-events/jevents.py +6
-1@@ -56,7 +56,9 @@ _json_event_attributes = [ # The list of counter(s) this event could use 'counters', # Longer things (the last won't be iterated over during decompress). - 'long_desc' + 'long_desc', + # Taken alone event could not be collected in the same group with other taken alone event + 'taken_alone' ] # Attributes that are in pmu_unit_layout.@@ -355,6 +357,9 @@ class JsonEvent: self.num_counters = jd.get('NumCounters') # Number of fixed counter self.num_fixed_counters = jd.get('NumFixedCounters') + # If the event is taken alone event, which cannot be grouped with any other + # taken alone event. + self.taken_alone = jd.get('TakenAlone') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')tools/perf/pmu-events/pmu-events.h +1
-0@@ -54,6 +54,7 @@ struct pmu_event {const char *unit;bool perpkg;bool deprecated; +bool taken_alone; };struct pmu_metric {tools/perf/util/metricgroup.c +14
-4@@ -189,6 +189,7 @@ struct metricgroup__event_info {/** The event uses special counters that we consider that as free counter * during the event grouping*/ bool free_counter; +bool taken_alone;/** The counters the event allowed to be collected on. */ DECLARE_BITMAP (counters,NR_COUNTERS ); };@@ -235,6 +236,7 @@ struct metricgroup__group {DECLARE_BITMAP (fixed_counters,NR_COUNTERS );/** Head to the list of event names in this group*/ struct list_head event_head; +bool taken_alone; };struct metricgroup__group_events {@@ -1717,6 +1719,7 @@ static void metricgroup__free_pmu_info(struct list_head *pmu_info_list)static struct metricgroup__event_info *event_info__new (const char *name,const char *pmu_name,const char *counter, +bool taken_alone,bool free_counter) {int ret =0 ;@@ -1731,6 +1734,7 @@ static struct metricgroup__event_info *event_info__new(const char *name, pmu_name ="core" ; e->name = name; + e->taken_alone = taken_alone; e->free_counter = free_counter; e->pmu_name = pmu_name;if (free_counter) {@@ -1769,7 +1773,8 @@ static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,if (!strcasecmp (pe->name, d->event_name)) {if (!pe->counters)return -EINVAL ; - event =event_info__new (d->event_id, pe->pmu, pe->counters,/*free_counter=*/ false); + event =event_info__new (d->event_id, pe->pmu, pe->counters, + pe->taken_alone,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);@@ -1892,6 +1897,8 @@ static int find_and_set_counters(struct metricgroup__event_info *e,int ret;unsigned long find_bit =0 ; +if (e->taken_alone && current_group->taken_alone) +return -ENOSPC ;if (e->free_counter)return 0 ;if (e->fixed_counter) {@@ -1926,11 +1933,13 @@ static int _insert_event(struct metricgroup__event_info *e,list_add (&event->nd, &group->event_head); else list_add_tail(&event->nd, &group->event_head); +if (e->taken_alone) + group->taken_alone = true;return 0 ; }/** -* Insert the new_group node at the end of the group list. +* Initialize the new group and insert it to the end of the group list. */ static int insert_new_group(struct list_head *head, struct metricgroup__group *new_group,@@ -1940,6 +1949,7 @@ static int insert_new_group(struct list_head *head,INIT_LIST_HEAD (&new_group->event_head);fill_counter_bitmap (new_group->gp_counters,0 , num_counters);fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); + new_group->taken_alone = false;list_add_tail (&new_group->nd, head);return 0 ; }@@ -2143,8 +2153,8 @@ static int create_grouping(struct list_head *pmu_info_list,//TODO: for each new core group, we should consider to add events that uses fixed counters list_for_each_entry (e, event_info_list, nd) {bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS ); -pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, - e->pmu_name, bit_buf); +pr_debug ("Event name %s, [pmu]=%s, [counters]=%s, [taken_alone]=%d\n" , + e->name, e->pmu_name, bit_buf, e->taken_alone); ret =assign_event_grouping (e, pmu_info_list, &groups);if (ret) goto out;
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, · line 2204
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-1@@ -2191,7 +2201,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused,if (is_special_event (id)) {struct metricgroup__event_info *event; - event = event_info__new(id, "default_core", "0", + event = event_info__new(id, "default_core", "0", false,
nit: document constant arguments, so "/*taken_alone=*/false,"
Thanks,
Ian
Quoted 6 lines
tools/perf/util/metricgroup.c +0
-1/*free_counter=*/true);if (!event) goto err_out; --2.42.0
RE: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware-grouping
Quoted 17 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 10:25 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware- grouping
Quoted 8 lines
Quoted 5 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add taken alone into consideration when grouping. Only one taken alone event is supported per group.
Is there an example of this?
Yes, there are about 20+ taken alone core events in SPR. Events like UOPS_RETIRED.MS, FRONTEND_RETIRED.DSB_MISS, INT_MISC.UNKNOWN_BRANCH_CYCLES, and etc.
If we run perf stat -e with two such kind of events, we will see multiplexing. ./perf stat -e "INT_MISC.UNKNOWN_BRANCH_CYCLES,UOPS_RETIRED.MS" -a sleep 1
Performance counter stats for 'system wide':
3,257,425 INT_MISC.UNKNOWN_BRANCH_CYCLES (50.04%)
134,098,908 UOPS_RETIRED.MS (49.96%)1.008902540 seconds time elapsed
This multiplexing should not happen if we run only one such type of event at a time. ./perf stat -e "INT_MISC.CLEAR_RESTEER_CYCLES,UOPS_RETIRED.MS" -a sleep 1
Performance counter stats for 'system wide':
5,107,299 INT_MISC.CLEAR_RESTEER_CYCLES
21,929,170 UOPS_RETIRED.MS
1.009937991 seconds time elapsedShould I add this example in the commit message?
Thanks,
Weilin
Quoted 169 lines
tools/perf/pmu-
Quoted 9 lines
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 7 ++++++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 20 +++++++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-)tools/perf/pmu- renamed from tools/perf/pmu-events/jevents.py +0
-0
events/jevents.pytools/perf/pmu-events/jevents.py _json_event_attributes = [ · line 60
Quoted 10 lines
index
bc91b7efa49a..4fbb367a3228100755tools/perf/pmu-events/jevents.py +2
-1@@ -56,7 +56,9 @@ _json_event_attributes = [ # The list of counter(s) this event could use 'counters', # Longer things (the last won't be iterated over during decompress). - 'long_desc' + 'long_desc', + # Taken alone event could not be collected in the same group with othertaken alone event
tools/perf/pmu-events/jevents.py class JsonEvent: · line 360
Quoted 9 lines
tools/perf/pmu-events/jevents.py +2
-0+ 'taken_alone' ] # Attributes that are in pmu_unit_layout.@@ -355,6 +357,9 @@ class JsonEvent: self.num_counters = jd.get('NumCounters') # Number of fixed counter self.num_fixed_counters = jd.get('NumFixedCounters') + # If the event is taken alone event, which cannot be grouped with anyother
tools/perf/pmu-
Quoted 6 lines
tools/perf/pmu-events/jevents.py +2
-0+ # taken alone event. + self.taken_alone = jd.get('TakenAlone') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')tools/perf/pmu- renamed from tools/perf/pmu-events/pmu-events.h +0
-0
events/pmu-events.htools/perf/util/metricgroup.c static void metricgroup__free_pmu_info(struct
Quoted 32 lines
index
e245e4738970..837edfeb676a100644tools/perf/pmu-events/pmu-events.h +1
-0@@ -54,6 +54,7 @@ struct pmu_event {const char *unit;bool perpkg;bool deprecated; +bool taken_alone; };struct pmu_metric {tools/perf/util/metricgroup.c +2
-0@@ -189,6 +189,7 @@ struct metricgroup__event_info {/** The event uses special counters that we consider that as free counter * during the event grouping*/ bool free_counter; +bool taken_alone;/** The counters the event allowed to be collected on. */ DECLARE_BITMAP (counters,NR_COUNTERS ); };@@ -235,6 +236,7 @@ struct metricgroup__group {DECLARE_BITMAP (fixed_counters,NR_COUNTERS );/** Head to the list of event names in this group*/ struct list_head event_head; +bool taken_alone; };struct metricgroup__group_events {@@ -1717,6 +1719,7 @@ static void metricgroup__free_pmu_info(struct
list_head*pmu_info_list)tools/perf/util/metricgroup.c static struct metricgroup__event_info
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0static struct metricgroup__event_info *event_info__new (const char *name,const char *pmu_name,const char *counter, +bool taken_alone,bool free_counter) {int ret =0 ;@@ -1731,6 +1734,7 @@ static struct metricgroup__event_info*event_info__new(const char *name,
tools/perf/util/metricgroup.c static int
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0pmu_name ="core" ; e->name = name; + e->taken_alone = taken_alone; e->free_counter = free_counter; e->pmu_name = pmu_name;if (free_counter) {@@ -1769,7 +1773,8 @@ static intmetricgroup__add_metric_event_callback(const struct
pmu_event*pe,tools/perf/util/metricgroup.c static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,
Quoted 4 lines
tools/perf/util/metricgroup.c +0
-1if (!strcasecmp (pe->name, d->event_name)) {if (!pe->counters)return -EINVAL ; - event = event_info__new(d->event_id, pe->pmu, pe->counters,/*free_counter=*/false);
tools/perf/util/metricgroup.c static int find_and_set_counters(struct
Quoted 6 lines
tools/perf/util/metricgroup.c +2
-0+ event =event_info__new (d->event_id, pe->pmu, pe->counters, + pe->taken_alone,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);@@ -1892,6 +1897,8 @@ static int find_and_set_counters(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int _insert_event(struct
Quoted 9 lines
tools/perf/util/metricgroup.c +2
-0int ret;unsigned long find_bit =0 ; +if (e->taken_alone && current_group->taken_alone) +return -ENOSPC ;if (e->free_counter)return 0 ;if (e->fixed_counter) {@@ -1926,11 +1933,13 @@ static int _insert_event(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int insert_new_group(struct list_head
Quoted 15 lines
tools/perf/util/metricgroup.c +3
-1list_add (&event->nd, &group->event_head); else list_add_tail(&event->nd, &group->event_head); +if (e->taken_alone) + group->taken_alone = true;return 0 ; }/** -* Insert the new_group node at the end of the group list. +* Initialize the new group and insert it to the end of the group list. */ static int insert_new_group(struct list_head *head, struct metricgroup__group *new_group,@@ -1940,6 +1949,7 @@ static int insert_new_group(struct list_head*head,
tools/perf/util/metricgroup.c static int insert_new_group(struct list_head *head,
Quoted 3 lines
tools/perf/util/metricgroup.c +0
-0INIT_LIST_HEAD (&new_group->event_head);fill_counter_bitmap (new_group->gp_counters,0 , num_counters); fill_counter_bitmap(new_group->fixed_counters, 0,
num_fixed_counters);tools/perf/util/metricgroup.c static int create_grouping(struct list_head
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ new_group->taken_alone = false;list_add_tail (&new_group->nd, head);return 0 ; }@@ -2143,8 +2153,8 @@ static int create_grouping(struct list_head
*pmu_info_list,tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2153
Quoted 1 line
tools/perf/util/metricgroup.c +0
-0//TODO: for each new core group, we should consider to add events that uses fixed counters
tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2155
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0list_for_each_entry (e, event_info_list, nd) { bitmap_scnprintf(e->counters, NR_COUNTERS, bit_buf,
NR_COUNTERS);tools/perf/util/metricgroup.c static int create_grouping(struct list_head *pmu_info_list,
Quoted 3 lines
tools/perf/util/metricgroup.c +1
-2-pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, - e->pmu_name, bit_buf); + pr_debug("Event name %s, [pmu]=%s, [counters]=%s,[taken_alone]=%d\n",
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ e->name, e->pmu_name, bit_buf, e->taken_alone); ret =assign_event_grouping (e, pmu_info_list, &groups);if (ret) goto out;@@ -2191,7 +2201,7 @@ static int hw_aware_build_grouping(struct
expr_parse_ctx*ctx__maybe_unused,tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, · line 2204
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-1if (is_special_event (id)) {struct metricgroup__event_info *event; - event = event_info__new(id, "default_core", "0", + event = event_info__new(id, "default_core", "0", false,nit: document constant arguments, so "/*taken_alone=*/false,"
Thanks,
IanQuoted 6 lines
tools/perf/util/metricgroup.c +0
-1/*free_counter=*/true);if (!event) goto err_out; --2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware-grouping
Quoted 25 lines · Weilin
Quoted 25 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 10:25 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware- groupingQuoted 5 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add taken alone into consideration when grouping. Only one taken alone event is supported per group.
Quoted 8 lines
Quoted 1 line
Is there an example of this?
Yes, there are about 20+ taken alone core events in SPR. Events like UOPS_RETIRED.MS, FRONTEND_RETIRED.DSB_MISS, INT_MISC.UNKNOWN_BRANCH_CYCLES, and etc.
If we run
perf stat -ewith two such kind of events, we will see multiplexing. ./perf stat -e "INT_MISC.UNKNOWN_BRANCH_CYCLES,UOPS_RETIRED.MS" -a sleep 1
Should these events be in a group?
Quoted 10 lines
Performance counter stats for 'system wide':
3,257,425 INT_MISC.UNKNOWN_BRANCH_CYCLES (50.04%) 134,098,908 UOPS_RETIRED.MS (49.96%)1.008902540 seconds time elapsedThis multiplexing should not happen if we run only one such type of event at a time. ./perf stat -e "INT_MISC.CLEAR_RESTEER_CYCLES,UOPS_RETIRED.MS" -a sleep 1
Quoted 8 lines
Performance counter stats for 'system wide':
5,107,299 INT_MISC.CLEAR_RESTEER_CYCLES 21,929,170 UOPS_RETIRED.MS 1.009937991 seconds time elapsedShould I add this example in the commit message?
So taken alone means that the events share a counter? What's going wrong in the first case and causing the multiplexing?
Thanks,
Ian
Quoted 172 lines
Thanks,
WeilinQuoted 169 lines
tools/perf/pmu-
Quoted 9 lines
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 7 ++++++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 20 +++++++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-)tools/perf/pmu- renamed from tools/perf/pmu-events/jevents.py +0
-0
events/jevents.pytools/perf/pmu-events/jevents.py _json_event_attributes = [ · line 60
Quoted 10 lines
index
bc91b7efa49a..4fbb367a3228100755tools/perf/pmu-events/jevents.py +2
-1@@ -56,7 +56,9 @@ _json_event_attributes = [ # The list of counter(s) this event could use 'counters', # Longer things (the last won't be iterated over during decompress). - 'long_desc' + 'long_desc', + # Taken alone event could not be collected in the same group with othertaken alone event
tools/perf/pmu-events/jevents.py class JsonEvent: · line 360
Quoted 9 lines
tools/perf/pmu-events/jevents.py +2
-0+ 'taken_alone' ] # Attributes that are in pmu_unit_layout.@@ -355,6 +357,9 @@ class JsonEvent: self.num_counters = jd.get('NumCounters') # Number of fixed counter self.num_fixed_counters = jd.get('NumFixedCounters') + # If the event is taken alone event, which cannot be grouped with anyother
tools/perf/pmu-
Quoted 6 lines
tools/perf/pmu-events/jevents.py +2
-0+ # taken alone event. + self.taken_alone = jd.get('TakenAlone') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')tools/perf/pmu- renamed from tools/perf/pmu-events/pmu-events.h +0
-0
events/pmu-events.htools/perf/util/metricgroup.c static void metricgroup__free_pmu_info(struct
Quoted 32 lines
index
e245e4738970..837edfeb676a100644tools/perf/pmu-events/pmu-events.h +1
-0@@ -54,6 +54,7 @@ struct pmu_event {const char *unit;bool perpkg;bool deprecated; +bool taken_alone; };struct pmu_metric {tools/perf/util/metricgroup.c +2
-0@@ -189,6 +189,7 @@ struct metricgroup__event_info {/** The event uses special counters that we consider that as free counter * during the event grouping*/ bool free_counter; +bool taken_alone;/** The counters the event allowed to be collected on. */ DECLARE_BITMAP (counters,NR_COUNTERS ); };@@ -235,6 +236,7 @@ struct metricgroup__group {DECLARE_BITMAP (fixed_counters,NR_COUNTERS );/** Head to the list of event names in this group*/ struct list_head event_head; +bool taken_alone; };struct metricgroup__group_events {@@ -1717,6 +1719,7 @@ static void metricgroup__free_pmu_info(struct
list_head*pmu_info_list)tools/perf/util/metricgroup.c static struct metricgroup__event_info
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0static struct metricgroup__event_info *event_info__new (const char *name,const char *pmu_name,const char *counter, +bool taken_alone,bool free_counter) {int ret =0 ;@@ -1731,6 +1734,7 @@ static struct metricgroup__event_info*event_info__new(const char *name,
tools/perf/util/metricgroup.c static int
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0pmu_name ="core" ; e->name = name; + e->taken_alone = taken_alone; e->free_counter = free_counter; e->pmu_name = pmu_name;if (free_counter) {@@ -1769,7 +1773,8 @@ static intmetricgroup__add_metric_event_callback(const struct
pmu_event*pe,tools/perf/util/metricgroup.c static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,
Quoted 4 lines
tools/perf/util/metricgroup.c +0
-1if (!strcasecmp (pe->name, d->event_name)) {if (!pe->counters)return -EINVAL ; - event = event_info__new(d->event_id, pe->pmu, pe->counters,/*free_counter=*/false);
tools/perf/util/metricgroup.c static int find_and_set_counters(struct
Quoted 6 lines
tools/perf/util/metricgroup.c +2
-0+ event =event_info__new (d->event_id, pe->pmu, pe->counters, + pe->taken_alone,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);@@ -1892,6 +1897,8 @@ static int find_and_set_counters(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int _insert_event(struct
Quoted 9 lines
tools/perf/util/metricgroup.c +2
-0int ret;unsigned long find_bit =0 ; +if (e->taken_alone && current_group->taken_alone) +return -ENOSPC ;if (e->free_counter)return 0 ;if (e->fixed_counter) {@@ -1926,11 +1933,13 @@ static int _insert_event(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int insert_new_group(struct list_head
Quoted 15 lines
tools/perf/util/metricgroup.c +3
-1list_add (&event->nd, &group->event_head); else list_add_tail(&event->nd, &group->event_head); +if (e->taken_alone) + group->taken_alone = true;return 0 ; }/** -* Insert the new_group node at the end of the group list. +* Initialize the new group and insert it to the end of the group list. */ static int insert_new_group(struct list_head *head, struct metricgroup__group *new_group,@@ -1940,6 +1949,7 @@ static int insert_new_group(struct list_head*head,
tools/perf/util/metricgroup.c static int insert_new_group(struct list_head *head,
Quoted 3 lines
tools/perf/util/metricgroup.c +0
-0INIT_LIST_HEAD (&new_group->event_head);fill_counter_bitmap (new_group->gp_counters,0 , num_counters); fill_counter_bitmap(new_group->fixed_counters, 0,
num_fixed_counters);tools/perf/util/metricgroup.c static int create_grouping(struct list_head
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ new_group->taken_alone = false;list_add_tail (&new_group->nd, head);return 0 ; }@@ -2143,8 +2153,8 @@ static int create_grouping(struct list_head
*pmu_info_list,tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2153
Quoted 1 line
tools/perf/util/metricgroup.c +0
-0//TODO: for each new core group, we should consider to add events that uses fixed counters
tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2155
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0list_for_each_entry (e, event_info_list, nd) { bitmap_scnprintf(e->counters, NR_COUNTERS, bit_buf,
NR_COUNTERS);tools/perf/util/metricgroup.c static int create_grouping(struct list_head *pmu_info_list,
Quoted 3 lines
tools/perf/util/metricgroup.c +1
-2-pr_debug ("Event name %s, [pmu]=%s, [counters]=%s\n" , e->name, - e->pmu_name, bit_buf); + pr_debug("Event name %s, [pmu]=%s, [counters]=%s,[taken_alone]=%d\n",
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ e->name, e->pmu_name, bit_buf, e->taken_alone); ret =assign_event_grouping (e, pmu_info_list, &groups);if (ret) goto out;@@ -2191,7 +2201,7 @@ static int hw_aware_build_grouping(struct
expr_parse_ctx*ctx__maybe_unused,tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, · line 2204
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-1if (is_special_event (id)) {struct metricgroup__event_info *event; - event = event_info__new(id, "default_core", "0", + event = event_info__new(id, "default_core", "0", false,nit: document constant arguments, so "/*taken_alone=*/false,"
Thanks,
IanQuoted 6 lines
tools/perf/util/metricgroup.c +0
-1/*free_counter=*/true);if (!event) goto err_out; --2.42.0
RE: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware-grouping
Quoted 77 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Tuesday, March 26, 2024 5:05 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 11/15] perf stat: Handle taken alone in hardware- groupingQuoted 14 lines · Weilin
Quoted 14 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 10:25 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>;Biggers,
Quoted 2 lines
Quoted 2 lines
Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 11/15] perf stat: Handle taken alone inhardware-
Quoted 9 lines
Quoted 9 lines
grouping
Quoted 5 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add taken alone into consideration when grouping. Only one taken alone event is supported per group.
Quoted 8 lines
Quoted 1 line
Is there an example of this?
Yes, there are about 20+ taken alone core events in SPR. Events like UOPS_RETIRED.MS, FRONTEND_RETIRED.DSB_MISS, INT_MISC.UNKNOWN_BRANCH_CYCLES, and etc.
If we run
perf stat -ewith two such kind of events, we will see multiplexing. ./perf stat -e "INT_MISC.UNKNOWN_BRANCH_CYCLES,UOPS_RETIRED.MS"-a sleep 1
Should these events be in a group?
Quoted 3 lines
Performance counter stats for 'system wide':
3,257,425 INT_MISC.UNKNOWN_BRANCH_CYCLES(50.04%)
Quoted 1 line
134,098,908 UOPS_RETIRED.MS(49.96%)
Quoted 4 lines
1.008902540 seconds time elapsedThis multiplexing should not happen if we run only one such type of event at
a time.
Quoted 1 line
./perf stat -e "INT_MISC.CLEAR_RESTEER_CYCLES,UOPS_RETIRED.MS" -a
sleep 1
Quoted 6 lines
Performance counter stats for 'system wide':
5,107,299 INT_MISC.CLEAR_RESTEER_CYCLES 21,929,170 UOPS_RETIRED.MS
Quoted 6 lines
Quoted 3 lines
1.009937991 seconds time elapsedShould I add this example in the commit message?
So taken alone means that the events share a counter? What's going wrong in the first case and causing the multiplexing?
I think multiple taken alone events cannot get collected in one group. That's why we see the multiplexing. But I'm not clear how that works in hardware. It's probably not caused by they share a counter though, because I could still see these events could be collected in a list of different gp counters.
Quoted 182 lines
Thanks,
IanQuoted 13 lines
Thanks,
WeilinQuoted 10 lines
tools/perf/pmu-
Quoted 9 lines
Signed-off-by Weilin Wang <[email protected]>--- tools/perf/pmu-events/jevents.py | 7 ++++++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 20 +++++++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-)tools/perf/pmu- renamed from tools/perf/pmu-events/jevents.py +0
-0
events/jevents.pyQuoted 10 lines
Quoted 10 lines
tools/perf/pmu-events/jevents.py _json_event_attributes = [ · line 60
Quoted 10 lines
index
bc91b7efa49a..4fbb367a3228100755tools/perf/pmu-events/jevents.py +2
-1@@ -56,7 +56,9 @@ _json_event_attributes = [ # The list of counter(s) this event could use 'counters', # Longer things (the last won't be iterated over during decompress). - 'long_desc' + 'long_desc', + # Taken alone event could not be collected in the same group withother
Quoted 28 lines
Quoted 28 lines
taken alone event
tools/perf/pmu-events/jevents.py class JsonEvent: · line 360
Quoted 9 lines
tools/perf/pmu-events/jevents.py +2
-0+ 'taken_alone' ] # Attributes that are in pmu_unit_layout.@@ -355,6 +357,9 @@ class JsonEvent: self.num_counters = jd.get('NumCounters') # Number of fixed counter self.num_fixed_counters = jd.get('NumFixedCounters') + # If the event is taken alone event, which cannot be grouped with anyother
tools/perf/pmu-
Quoted 6 lines
tools/perf/pmu-events/jevents.py +2
-0+ # taken alone event. + self.taken_alone = jd.get('TakenAlone') filter = jd.get('Filter') self.unit = jd.get('ScaleUnit') self.perpkg = jd.get('PerPkg')tools/perf/pmu- renamed from tools/perf/pmu-events/pmu-events.h +0
-0
events/pmu-events.hQuoted 10 lines
index
e245e4738970..837edfeb676a100644tools/perf/pmu-events/pmu-events.h +1
-0@@ -54,6 +54,7 @@ struct pmu_event {const char *unit;bool perpkg;bool deprecated; +bool taken_alone; };Quoted 7 lines
Quoted 7 lines
tools/perf/util/metricgroup.c struct metricgroup__event_info { · line 189
Quoted 7 lines
tools/perf/pmu-events/pmu-events.h +0
-0struct pmu_metric {tools/perf/util/metricgroup.c +0
-0@@ -189,6 +189,7 @@ struct metricgroup__event_info {/** The event uses special counters that we consider that as free counter
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
tools/perf/util/metricgroup.c +2
-0* during the event grouping*/ bool free_counter; +bool taken_alone;/** The counters the event allowed to be collected on. */ DECLARE_BITMAP (counters,NR_COUNTERS ); };@@ -235,6 +236,7 @@ struct metricgroup__group {DECLARE_BITMAP (fixed_counters,NR_COUNTERS );/** Head to the list of event names in this group*/ struct list_head event_head; +bool taken_alone; };Quoted 2 lines
Quoted 2 lines
tools/perf/util/metricgroup.c static void
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0struct metricgroup__group_events {@@ -1717,6 +1719,7 @@ static voidmetricgroup__free_pmu_info(struct
Quoted 2 lines
Quoted 2 lines
list_head*pmu_info_list)Quoted 1 line
tools/perf/util/metricgroup.c +0
-0static struct metricgroup__event_info *event_info__new(const char*name,
Quoted 58 lines
Quoted 58 lines
tools/perf/util/metricgroup.c static struct metricgroup__event_info
Quoted 7 lines
tools/perf/util/metricgroup.c +1
-0const char *pmu_name, const char *counter, + bool taken_alone, bool free_counter) { int ret = 0;@@ -1731,6 +1734,7 @@ static struct metricgroup__event_info*event_info__new(const char *name,
tools/perf/util/metricgroup.c static int
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0pmu_name ="core" ; e->name = name; + e->taken_alone = taken_alone; e->free_counter = free_counter; e->pmu_name = pmu_name;if (free_counter) {@@ -1769,7 +1773,8 @@ static intmetricgroup__add_metric_event_callback(const struct
pmu_event*pe,tools/perf/util/metricgroup.c static int metricgroup__add_metric_event_callback(const struct pmu_event *pe,
Quoted 4 lines
tools/perf/util/metricgroup.c +0
-1if (!strcasecmp (pe->name, d->event_name)) {if (!pe->counters)return -EINVAL ; - event = event_info__new(d->event_id, pe->pmu, pe->counters,/*free_counter=*/false);
tools/perf/util/metricgroup.c static int find_and_set_counters(struct
Quoted 6 lines
tools/perf/util/metricgroup.c +2
-0+ event =event_info__new (d->event_id, pe->pmu, pe->counters, + pe->taken_alone,/*free_counter=*/ false);if (!event)return -ENOMEM ;list_add (&event->nd, d->list);@@ -1892,6 +1897,8 @@ static int find_and_set_counters(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int _insert_event(struct
Quoted 9 lines
tools/perf/util/metricgroup.c +2
-0int ret;unsigned long find_bit =0 ; +if (e->taken_alone && current_group->taken_alone) +return -ENOSPC ;if (e->free_counter)return 0 ;if (e->fixed_counter) {@@ -1926,11 +1933,13 @@ static int _insert_event(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int insert_new_group(struct list_head
Quoted 15 lines
tools/perf/util/metricgroup.c +3
-1list_add (&event->nd, &group->event_head); else list_add_tail(&event->nd, &group->event_head); +if (e->taken_alone) + group->taken_alone = true;return 0 ; }/** -* Insert the new_group node at the end of the group list. +* Initialize the new group and insert it to the end of the group list. */ static int insert_new_group(struct list_head *head, struct metricgroup__group *new_group,@@ -1940,6 +1949,7 @@ static int insert_new_group(struct list_head*head,
Quoted 3 lines
tools/perf/util/metricgroup.c +0
-0INIT_LIST_HEAD (&new_group->event_head);fill_counter_bitmap (new_group->gp_counters,0 , num_counters); fill_counter_bitmap(new_group->fixed_counters, 0,Quoted 8 lines
Quoted 8 lines
num_fixed_counters);tools/perf/util/metricgroup.c static int create_grouping(struct list_head
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ new_group->taken_alone = false;list_add_tail (&new_group->nd, head);return 0 ; }@@ -2143,8 +2153,8 @@ static int create_grouping(struct list_head
*pmu_info_list,Quoted 1 line
tools/perf/util/metricgroup.c +0
-0//TODO: for each new core group, we should consider to add events that
Quoted 32 lines
Quoted 5 lines
uses fixed counters
tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2155
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0list_for_each_entry (e, event_info_list, nd) { bitmap_scnprintf(e->counters, NR_COUNTERS, bit_buf,
NR_COUNTERS);Quoted 1 line
tools/perf/util/metricgroup.c +0
-1- pr_debug("Event name %s, [pmu]=%s, [counters]=%s\n", e-name,
Quoted 26 lines
tools/perf/util/metricgroup.c static int create_grouping(struct list_head · line 2156
Quoted 2 lines
tools/perf/util/metricgroup.c +1
-1- e->pmu_name, bit_buf); + pr_debug("Event name %s, [pmu]=%s, [counters]=%s,[taken_alone]=%d\n",
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ e->name, e->pmu_name, bit_buf, e->taken_alone); ret =assign_event_grouping (e, pmu_info_list, &groups);if (ret) goto out;@@ -2191,7 +2201,7 @@ static int hw_aware_build_grouping(struct
expr_parse_ctx*ctx__maybe_unused,tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, · line 2204
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-1if (is_special_event (id)) {struct metricgroup__event_info *event; - event = event_info__new(id, "default_core", "0", + event = event_info__new(id, "default_core", "0", false,nit: document constant arguments, so "/*taken_alone=*/false,"
Thanks,
IanQuoted 6 lines
tools/perf/util/metricgroup.c +0
-1/*free_counter=*/true);if (!event) goto err_out; --2.42.0
RFC PATCH v4 12/15 perf stat: Handle NMI in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Add an easy nmi watchdog support in grouping. When nmi watchdog is enabled, we reduce the total num of events could be assigned to one group by 1. A more efficient solution will be added in later.
--- tools/perf/util/metricgroup.c | 4 ++++ 1 file changed, 4 insertions(+)
tools/perf/util/metricgroup.c +4 -0
@@ -1947,6 +1947,10 @@ static int insert_new_group(struct list_head *head, size_t num_fixed_counters) {INIT_LIST_HEAD (&new_group->event_head); +if (sysctl__nmi_watchdog_enabled ()) { +pr_debug ("NMI watchdog is enabled. Reduce num of counters by 1\n" ); + num_counters -=1 ; + }fill_counter_bitmap (new_group->gp_counters,0 , num_counters);fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); new_group->taken_alone = false;
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 12/15] perf stat: Handle NMI in hardware-grouping
Quoted 8 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add an easy nmi watchdog support in grouping. When nmi watchdog is enabled, we reduce the total num of events could be assigned to one group by 1. A more efficient solution will be added in later.
Signed-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 22 lines
--- tools/perf/util/metricgroup.c | 4 ++++ 1 file changed, 4 insertions(+)tools/perf/util/metricgroup.c +4
-1@@ -1947,6 +1947,10 @@ static int insert_new_group(struct list_head *head, size_t num_fixed_counters) {INIT_LIST_HEAD (&new_group->event_head); +if (sysctl__nmi_watchdog_enabled ()) { +pr_debug ("NMI watchdog is enabled. Reduce num of counters by 1\n" ); + num_counters -=1 ; + }fill_counter_bitmap (new_group->gp_counters,0 , num_counters);fill_counter_bitmap (new_group->fixed_counters,0 , num_fixed_counters); new_group->taken_alone = false; --2.42.0
RFC PATCH v4 13/15 perf stat: Code refactoring in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Decouple the step to generate final grouping strings out from the build_grouping step so that we could do single metric grouping and then merge groups if needed later.
--- tools/perf/util/metricgroup.c | 50 +++++++++++++++++------------------ 1 file changed, 25 insertions(+), 25 deletions(-)
tools/perf/util/metricgroup.c +25 -25
@@ -1896,9 +1896,10 @@ static int find_and_set_counters(struct metricgroup__event_info *e, {int ret;unsigned long find_bit =0 ; - -if (e->taken_alone && current_group->taken_alone) +if (e->taken_alone && current_group->taken_alone) { +pr_debug ("current group with taken alone event already\n" );return -ENOSPC ; + }if (e->free_counter)return 0 ;if (e->fixed_counter) {
@@ -2017,7 +2018,8 @@ static int assign_event_grouping(struct metricgroup__event_info *e,list_for_each_entry (g, groups, nd) {if (!strcasecmp (g->pmu_name, e->pmu_name)) { -pr_debug ("found group for event %s in pmu %s\n" , e->name, g->pmu_name); +pr_debug ("found group header for event %s in pmu %s\n" , + e->name, g->pmu_name); pmu_group_head = g;break ; }
@@ -2146,26 +2148,22 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs*/ static int create_grouping (struct list_head *pmu_info_list,struct list_head *event_info_list, -struct list_head *groupings, -const char *modifier) +struct list_head *grouping) {int ret =0 ;struct metricgroup__event_info *e; -LIST_HEAD (groups);char *bit_buf =malloc (NR_COUNTERS ); -//TODO: for each new core group, we should consider to add events that uses fixed counters +//TODO: for each new core group, we could consider to add events that +//uses fixed counters list_for_each_entry (e, event_info_list, nd) {bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS );pr_debug ("Event name %s, [pmu]=%s, [counters]=%s, [taken_alone]=%d\n" , e->name, e->pmu_name, bit_buf, e->taken_alone); - ret =assign_event_grouping (e, pmu_info_list, &groups); + ret =assign_event_grouping (e, pmu_info_list, grouping);if (ret) - goto out; +return ret; } - ret =hw_aware_metricgroup__build_event_string (groupings, modifier, &groups); -out: -metricgroup__free_group_list (&groups);return ret; };
@@ -2186,9 +2184,8 @@ static bool is_special_event(const char *id)* @groupings: header to the list of final event grouping. * @modifier: any modifiers added to the events. */ -static int hw_aware_build_grouping (struct expr_parse_ctx *ctx__maybe_unused , -struct list_head *groupings__maybe_unused , -const char *modifier__maybe_unused ) +static int hw_aware_build_grouping (struct expr_parse_ctx *ctx, +struct list_head *grouping) {int ret =0 ;struct hashmap_entry *cur;
@@ -2220,8 +2217,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out; - ret =create_grouping (&pmu_info_list, &event_info_list, groupings, - modifier); + ret =create_grouping (&pmu_info_list, &event_info_list, grouping); err_out:metricgroup__free_event_info (&event_info_list);
@@ -2267,23 +2263,25 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, {struct parse_events_error parse_error;struct evlist *parsed_evlist; -LIST_HEAD (groupings); +LIST_HEAD (grouping_str); +LIST_HEAD (grouping);struct metricgroup__group_strs *group;int ret; *out_evlist =NULL ; - ret =hw_aware_build_grouping (ids, &groupings, modifier); -if (ret) { -metricgroup__free_grouping_strs (&groupings); -return ret; - } + ret =hw_aware_build_grouping (ids, &grouping); +if (ret) + goto out; + ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, &grouping); +if (ret) + goto out; parsed_evlist =evlist__new ();if (!parsed_evlist) { ret = -ENOMEM ; goto err_out; } -list_for_each_entry (group, &groupings, nd) { +list_for_each_entry (group, &grouping_str, nd) {struct strbuf *events = &group->grouping_str;pr_debug ("Parsing metric events '%s'\n" , events->buf);
@@ -2303,7 +2301,9 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, err_out:parse_events_error__exit (&parse_error);evlist__delete (parsed_evlist); -metricgroup__free_grouping_strs (&groupings); +out: +metricgroup__free_group_list (&grouping); +metricgroup__free_grouping_strs (&grouping_str);return ret; }
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 13/15] perf stat: Code refactoring in hardware-grouping
Quoted 8 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Decouple the step to generate final grouping strings out from the
build_groupingstep so that we could do single metric grouping and then merge groups if needed later.Signed-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 132 lines
--- tools/perf/util/metricgroup.c | 50 +++++++++++++++++------------------ 1 file changed, 25 insertions(+), 25 deletions(-)tools/perf/util/metricgroup.c +25
-26@@ -1896,9 +1896,10 @@ static int find_and_set_counters(struct metricgroup__event_info *e, {int ret;unsigned long find_bit =0 ; - -if (e->taken_alone && current_group->taken_alone) +if (e->taken_alone && current_group->taken_alone) { +pr_debug ("current group with taken alone event already\n" );return -ENOSPC ; + }if (e->free_counter)return 0 ;if (e->fixed_counter) {@@ -2017,7 +2018,8 @@ static int assign_event_grouping(struct metricgroup__event_info *e,list_for_each_entry (g, groups, nd) {if (!strcasecmp (g->pmu_name, e->pmu_name)) { -pr_debug ("found group for event %s in pmu %s\n" , e->name, g->pmu_name); +pr_debug ("found group header for event %s in pmu %s\n" , + e->name, g->pmu_name); pmu_group_head = g;break ; }@@ -2146,26 +2148,22 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs*/ static int create_grouping (struct list_head *pmu_info_list,struct list_head *event_info_list, -struct list_head *groupings, -const char *modifier) +struct list_head *grouping) {int ret =0 ;struct metricgroup__event_info *e; -LIST_HEAD (groups);char *bit_buf =malloc (NR_COUNTERS ); -//TODO: for each new core group, we should consider to add events that uses fixed counters +//TODO: for each new core group, we could consider to add events that +//uses fixed counters list_for_each_entry (e, event_info_list, nd) {bitmap_scnprintf (e->counters,NR_COUNTERS , bit_buf,NR_COUNTERS );pr_debug ("Event name %s, [pmu]=%s, [counters]=%s, [taken_alone]=%d\n" , e->name, e->pmu_name, bit_buf, e->taken_alone); - ret =assign_event_grouping (e, pmu_info_list, &groups); + ret =assign_event_grouping (e, pmu_info_list, grouping);if (ret) - goto out; +return ret; } - ret =hw_aware_metricgroup__build_event_string (groupings, modifier, &groups); -out: -metricgroup__free_group_list (&groups);return ret; };@@ -2186,9 +2184,8 @@ static bool is_special_event(const char *id)* @groupings: header to the list of final event grouping. * @modifier: any modifiers added to the events. */ -static int hw_aware_build_grouping (struct expr_parse_ctx *ctx__maybe_unused , -struct list_head *groupings__maybe_unused , -const char *modifier__maybe_unused ) +static int hw_aware_build_grouping (struct expr_parse_ctx *ctx, +struct list_head *grouping) {int ret =0 ;struct hashmap_entry *cur;@@ -2220,8 +2217,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx __maybe_unused, ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out; - ret =create_grouping (&pmu_info_list, &event_info_list, groupings, - modifier); + ret =create_grouping (&pmu_info_list, &event_info_list, grouping); err_out:metricgroup__free_event_info (&event_info_list);@@ -2267,23 +2263,25 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, {struct parse_events_error parse_error;struct evlist *parsed_evlist; -LIST_HEAD (groupings); +LIST_HEAD (grouping_str); +LIST_HEAD (grouping);struct metricgroup__group_strs *group;int ret; *out_evlist =NULL ; - ret =hw_aware_build_grouping (ids, &groupings, modifier); -if (ret) { -metricgroup__free_grouping_strs (&groupings); -return ret; - } + ret =hw_aware_build_grouping (ids, &grouping); +if (ret) + goto out; + ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, &grouping); +if (ret) + goto out; parsed_evlist =evlist__new ();if (!parsed_evlist) { ret = -ENOMEM ; goto err_out; } -list_for_each_entry (group, &groupings, nd) { +list_for_each_entry (group, &grouping_str, nd) {struct strbuf *events = &group->grouping_str;pr_debug ("Parsing metric events '%s'\n" , events->buf);@@ -2303,7 +2301,9 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, err_out:parse_events_error__exit (&parse_error);evlist__delete (parsed_evlist); -metricgroup__free_grouping_strs (&groupings); +out: +metricgroup__free_group_list (&grouping); +metricgroup__free_grouping_strs (&grouping_str);return ret; } --2.42.0
RFC PATCH v4 14/15 perf stat: Add tool events support in hardware-grouping
In reply to [email protected]From: Weilin Wang <[email protected]>
Add tool events into default_core grouping strings if find tool events so that metrics use tool events could be correctly calculated. Need this step to support TopdownL4-L5.
--- tools/perf/util/metricgroup.c | 49 ++++++++++++++++++++++++++++++++++- 1 file changed, 48 insertions(+), 1 deletion(-)
tools/perf/util/metricgroup.c +48 -1
@@ -1486,6 +1486,35 @@ static void find_tool_events(const struct list_head *metric_list, } } +/** +* get_tool_event_str - Generate and return a string with all the used tool +* event names. +*/ +static int get_tool_event_str (struct strbuf *events, +const bool tool_events[PERF_TOOL_MAX ], +bool *has_tool_event) +{ +int i =0 ; +int ret; + +perf_tool_event__for_each_event (i) { +if (tool_events[i]) { +const char *tmp =strdup (perf_tool_event__to_str (i)); + +if (!tmp) +return -ENOMEM ; + *has_tool_event = true; + ret =strbuf_addstr (events,"," ); +if (ret) +return ret; + ret =strbuf_addstr (events, tmp); +if (ret) +return ret; + } + } +return 0 ; +} +/** * build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events * metric IDs, as the IDs are held in a set,
@@ -2049,6 +2078,7 @@ static int assign_event_grouping(struct metricgroup__event_info *e,static int hw_aware_metricgroup__build_event_string (struct list_head *group_strs,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct list_head *groups) {struct metricgroup__pmu_group_list *p;
@@ -2056,8 +2086,12 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strsstruct metricgroup__group_events *ge;bool no_group = true;int ret =0 ; +struct strbuf tool_event_str =STRBUF_INIT ; +bool has_tool_event = false;#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) + ret =get_tool_event_str (&tool_event_str, tool_events, &has_tool_event); +RETURN_IF_NON_ZERO (ret);list_for_each_entry (p, groups, nd) {list_for_each_entry (g, &p->group_head, nd) {
@@ -2129,6 +2163,12 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs } ret =strbuf_addf (events,"}:W" );RETURN_IF_NON_ZERO (ret); + +if (!strcmp (p->pmu_name,"default_core" ) && has_tool_event) { + ret =strbuf_addstr (events, tool_event_str.buf); +RETURN_IF_NON_ZERO (ret); + } +pr_debug ("events-buf: %s\n" , events->buf);list_add_tail (&new_group_str->nd, group_strs); }
@@ -2214,6 +2254,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx,if (ret) goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out;
@@ -2259,6 +2300,7 @@ static void metricgroup__free_grouping_strs(struct list_head*/ static int hw_aware_parse_ids (struct perf_pmu *fake_pmu,struct expr_parse_ctx *ids,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct evlist **out_evlist) {struct parse_events_error parse_error;
@@ -2272,7 +2314,8 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, ret =hw_aware_build_grouping (ids, &grouping);if (ret) goto out; - ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, &grouping); + ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, + tool_events, &grouping);if (ret) goto out;
@@ -2407,6 +2450,7 @@ static int hw_aware_parse_groups(struct evlist *perf_evlist,struct evlist *combined_evlist =NULL ;LIST_HEAD (metric_list);struct metric *m; +bool tool_events[PERF_TOOL_MAX ] = {false};int ret;bool metric_no_group = false;bool metric_no_merge = false;
@@ -2425,11 +2469,14 @@ static int hw_aware_parse_groups(struct evlist *perf_evlist,if (!metric_no_merge) {struct expr_parse_ctx *combined =NULL ; +find_tool_events (&metric_list, tool_events); + ret =hw_aware_build_combined_expr_ctx (&metric_list, &combined);if (!ret && combined &&hashmap__size (combined->ids)) { ret =hw_aware_parse_ids (fake_pmu, combined,/*modifier=*/ NULL , + tool_events, &combined_evlist); }
Signature · 1 line
2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 14/15] perf stat: Add tool events support in hardware-grouping
Quoted 61 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add tool events into
default_coregrouping strings if find tool events so that metrics use tool events could be correctly calculated. Need this step to support TopdownL4-L5.Signed-off-by Weilin Wang <[email protected]>--- tools/perf/util/metricgroup.c | 49 ++++++++++++++++++++++++++++++++++- 1 file changed, 48 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +30
-0@@ -1486,6 +1486,35 @@ static void find_tool_events(const struct list_head *metric_list, } } +/** +* get_tool_event_str - Generate and return a string with all the used tool +* event names. +*/ +static int get_tool_event_str (struct strbuf *events, +const bool tool_events[PERF_TOOL_MAX ], +bool *has_tool_event) +{ +int i =0 ; +int ret; + +perf_tool_event__for_each_event (i) { +if (tool_events[i]) { +const char *tmp =strdup (perf_tool_event__to_str (i)); + +if (!tmp) +return -ENOMEM ; + *has_tool_event = true; + ret =strbuf_addstr (events,"," ); +if (ret) +return ret; + ret =strbuf_addstr (events, tmp); +if (ret) +return ret; + } + } +return 0 ; +} +/** * build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events * metric IDs, as the IDs are held in a set, @@ -2049,6 +2078,7 @@ static int assign_event_grouping(struct metricgroup__event_info *e,static int hw_aware_metricgroup__build_event_string (struct list_head *group_strs,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct list_head *groups) {struct metricgroup__pmu_group_list *p;
tools/perf/util/metricgroup.c static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs · line 2093
Quoted 8 lines
tools/perf/util/metricgroup.c +3
-0@@ -2056,8 +2086,12 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strsstruct metricgroup__group_events *ge;bool no_group = true;int ret =0 ; +struct strbuf tool_event_str =STRBUF_INIT ; +bool has_tool_event = false;#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) + ret =get_tool_event_str (&tool_event_str, tool_events, &has_tool_event);
Does metricgroup__build_event_string need updating to use this helper function too?
Thanks,
Ian
Quoted 69 lines
tools/perf/util/metricgroup.c +15
-2+RETURN_IF_NON_ZERO (ret);list_for_each_entry (p, groups, nd) {list_for_each_entry (g, &p->group_head, nd) {@@ -2129,6 +2163,12 @@ static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs } ret =strbuf_addf (events,"}:W" );RETURN_IF_NON_ZERO (ret); + +if (!strcmp (p->pmu_name,"default_core" ) && has_tool_event) { + ret =strbuf_addstr (events, tool_event_str.buf); +RETURN_IF_NON_ZERO (ret); + } +pr_debug ("events-buf: %s\n" , events->buf);list_add_tail (&new_group_str->nd, group_strs); }@@ -2214,6 +2254,7 @@ static int hw_aware_build_grouping(struct expr_parse_ctx *ctx,if (ret) goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out;@@ -2259,6 +2300,7 @@ static void metricgroup__free_grouping_strs(struct list_head*/ static int hw_aware_parse_ids (struct perf_pmu *fake_pmu,struct expr_parse_ctx *ids,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct evlist **out_evlist) {struct parse_events_error parse_error;@@ -2272,7 +2314,8 @@ static int hw_aware_parse_ids(struct perf_pmu *fake_pmu, ret =hw_aware_build_grouping (ids, &grouping);if (ret) goto out; - ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, &grouping); + ret =hw_aware_metricgroup__build_event_string (&grouping_str, modifier, + tool_events, &grouping);if (ret) goto out;@@ -2407,6 +2450,7 @@ static int hw_aware_parse_groups(struct evlist *perf_evlist,struct evlist *combined_evlist =NULL ;LIST_HEAD (metric_list);struct metric *m; +bool tool_events[PERF_TOOL_MAX ] = {false};int ret;bool metric_no_group = false;bool metric_no_merge = false;@@ -2425,11 +2469,14 @@ static int hw_aware_parse_groups(struct evlist *perf_evlist,if (!metric_no_merge) {struct expr_parse_ctx *combined =NULL ; +find_tool_events (&metric_list, tool_events); + ret =hw_aware_build_combined_expr_ctx (&metric_list, &combined);if (!ret && combined &&hashmap__size (combined->ids)) { ret =hw_aware_parse_ids (fake_pmu, combined,/*modifier=*/ NULL , + tool_events, &combined_evlist); } --2.42.0
RE: [RFC PATCH v4 14/15] perf stat: Add tool events support in hardware-grouping
Quoted 91 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 10:56 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 14/15] perf stat: Add tool events support in hardware-groupingQuoted 10 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add tool events into
default_coregrouping strings if find tool events so that metrics use tool events could be correctly calculated. Need this step to support TopdownL4-L5.Signed-off-by Weilin Wang <[email protected]>--- tools/perf/util/metricgroup.c | 49++++++++++++++++++++++++++++++++++-
tools/perf/util/metricgroup.c static void find_tool_events(const struct
Quoted 7 lines
1 file changed, 48 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +0
-0@@ -1486,6 +1486,35 @@ static void find_tool_events(const struct
list_head*metric_list,Quoted 24 lines
tools/perf/util/metricgroup.c +21
-0} } +/** +* get_tool_event_str - Generate and return a string with all the used tool +* event names. +*/ +static int get_tool_event_str (struct strbuf *events, +const bool tool_events[PERF_TOOL_MAX ], +bool *has_tool_event) +{ +int i =0 ; +int ret; + +perf_tool_event__for_each_event (i) { +if (tool_events[i]) { +const char *tmp =strdup (perf_tool_event__to_str (i)); + +if (!tmp) +return -ENOMEM ; + *has_tool_event = true; + ret =strbuf_addstr (events,"," ); +if (ret) +return ret;tools/perf/util/metricgroup.c static void find_tool_events(const struct list_head *metric_list,
Quoted 10 lines
tools/perf/util/metricgroup.c +8
-0+ ret =strbuf_addstr (events, tmp); +if (ret) +return ret; + } + } +return 0 ; +} +/** * build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events
tools/perf/util/metricgroup.c static int assign_event_grouping(struct
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0* metric IDs, as the IDs are held in a set, @@ -2049,6 +2078,7 @@ static int assign_event_grouping(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int assign_event_grouping(struct · line 2078
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0static int hw_aware_metricgroup__build_event_string(struct list_head
*group_strs,tools/perf/util/metricgroup.c static int
Quoted 6 lines
tools/perf/util/metricgroup.c +1
-0const char *modifier, + const bool tool_events[PERF_TOOL_MAX], struct list_head *groups) { struct metricgroup__pmu_group_list *p;@@ -2056,8 +2086,12 @@ static inthw_aware_metricgroup__build_event_string(struct
list_head*group_strsQuoted 6 lines
tools/perf/util/metricgroup.c +2
-0struct metricgroup__group_events *ge;bool no_group = true;int ret =0 ; +struct strbuf tool_event_str =STRBUF_INIT ; +bool has_tool_event = false;
Quoted 6 lines
tools/perf/util/metricgroup.c static int · line 2093
Quoted 2 lines
tools/perf/util/metricgroup.c +1
-0#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) + ret = get_tool_event_str(&tool_event_str, tool_events,
&has_tool_event);Does
metricgroup__build_event_stringneed updating to use this helper function too?
Hi Ian,
In parse_ids, tool events are inserted to the ids that passed to metricgroup__build_event_string. If we want to use this function here, I think we also need to update the code in parse_ids to not insert tool events.
I could add this change if you think we should do this update.
Thanks,
Weilin
Quoted 83 lines
Thanks,
Iantools/perf/util/metricgroup.c static int
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+RETURN_IF_NON_ZERO (ret);list_for_each_entry (p, groups, nd) {list_for_each_entry (g, &p->group_head, nd) {@@ -2129,6 +2163,12 @@ static inthw_aware_metricgroup__build_event_string(struct
list_head*group_strstools/perf/util/metricgroup.c static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs
Quoted 5 lines
tools/perf/util/metricgroup.c +2
-0} ret =strbuf_addf (events,"}:W" );RETURN_IF_NON_ZERO (ret); + +if (!strcmp (p->pmu_name,"default_core" ) && has_tool_event){
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct
Quoted 8 lines
tools/perf/util/metricgroup.c +4
-0+ ret =strbuf_addstr (events, tool_event_str.buf); +RETURN_IF_NON_ZERO (ret); + } +pr_debug ("events-buf: %s\n" , events->buf);list_add_tail (&new_group_str->nd, group_strs); }@@ -2214,6 +2254,7 @@ static int hw_aware_build_grouping(struct
expr_parse_ctx*ctx,tools/perf/util/metricgroup.c static void
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0if (ret) goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out;@@ -2259,6 +2300,7 @@ static voidmetricgroup__free_grouping_strs(struct
list_headtools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0*/ static int hw_aware_parse_ids (struct perf_pmu *fake_pmu,struct expr_parse_ctx *ids,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct evlist **out_evlist) {struct parse_events_error parse_error;@@ -2272,7 +2314,8 @@ static int hw_aware_parse_ids(struct perf_pmu
*fake_pmu,tools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu *fake_pmu,
Quoted 4 lines
tools/perf/util/metricgroup.c +0
-1ret =hw_aware_build_grouping (ids, &grouping);if (ret) goto out; - ret = hw_aware_metricgroup__build_event_string(&grouping_str,modifier, &grouping);
tools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu · line 2317
Quoted 1 line
tools/perf/util/metricgroup.c +1
-0+ ret = hw_aware_metricgroup__build_event_string(&grouping_str,modifier,
tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ tool_events, &grouping);if (ret) goto out;@@ -2407,6 +2450,7 @@ static int hw_aware_parse_groups(struct evlist
*perf_evlist,tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0struct evlist *combined_evlist =NULL ;LIST_HEAD (metric_list);struct metric *m; +bool tool_events[PERF_TOOL_MAX ] = {false};int ret;bool metric_no_group = false;bool metric_no_merge = false;@@ -2425,11 +2469,14 @@ static int hw_aware_parse_groups(struct evlist
*perf_evlist,tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist *perf_evlist,
Quoted 6 lines
tools/perf/util/metricgroup.c +2
-0if (!metric_no_merge) {struct expr_parse_ctx *combined =NULL ; +find_tool_events (&metric_list, tool_events); + ret = hw_aware_build_combined_expr_ctx(&metric_list,&combined);
Quoted 11 lines
tools/perf/util/metricgroup.c +1
-1if (!ret && combined &&hashmap__size (combined->ids)) { ret =hw_aware_parse_ids (fake_pmu, combined,/*modifier=*/ NULL , + tool_events, &combined_evlist); } --2.42.0
Ian Rogers
reviewerRe: [RFC PATCH v4 14/15] perf stat: Add tool events support in hardware-grouping
Quoted 99 lines · Weilin
Quoted 92 lines
-----Original Message-----
From: Ian Rogers <[email protected]>
Sent: Saturday, March 23, 2024 10:56 PM
To: Wang, Weilin <[email protected]>
Cc: Kan Liang <[email protected]>; Namhyung Kim <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Alexander Shishkin <[email protected]>; Jiri Olsa <[email protected]>; Hunter, Adrian <[email protected]>; linux-perf- [email protected]; [email protected]; Taylor, Perry <[email protected]>; Alt, Samantha <[email protected]>; Biggers, Caleb <[email protected]>; Mark Rutland <[email protected]>
Subject: Re: [RFC PATCH v4 14/15] perf stat: Add tool events support in hardware-groupingQuoted 10 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add tool events into
default_coregrouping strings if find tool events so that metrics use tool events could be correctly calculated. Need this step to support TopdownL4-L5.Signed-off-by Weilin Wang <[email protected]>--- tools/perf/util/metricgroup.c | 49++++++++++++++++++++++++++++++++++-
tools/perf/util/metricgroup.c static void find_tool_events(const struct
Quoted 7 lines
1 file changed, 48 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +0
-0@@ -1486,6 +1486,35 @@ static void find_tool_events(const struct
list_head*metric_list,Quoted 24 lines
tools/perf/util/metricgroup.c +21
-0} } +/** +* get_tool_event_str - Generate and return a string with all the used tool +* event names. +*/ +static int get_tool_event_str (struct strbuf *events, +const bool tool_events[PERF_TOOL_MAX ], +bool *has_tool_event) +{ +int i =0 ; +int ret; + +perf_tool_event__for_each_event (i) { +if (tool_events[i]) { +const char *tmp =strdup (perf_tool_event__to_str (i)); + +if (!tmp) +return -ENOMEM ; + *has_tool_event = true; + ret =strbuf_addstr (events,"," ); +if (ret) +return ret;tools/perf/util/metricgroup.c static void find_tool_events(const struct list_head *metric_list,
Quoted 10 lines
tools/perf/util/metricgroup.c +8
-0+ ret =strbuf_addstr (events, tmp); +if (ret) +return ret; + } + } +return 0 ; +} +/** * build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events
tools/perf/util/metricgroup.c static int assign_event_grouping(struct
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0* metric IDs, as the IDs are held in a set, @@ -2049,6 +2078,7 @@ static int assign_event_grouping(struct
metricgroup__event_info*e,tools/perf/util/metricgroup.c static int assign_event_grouping(struct · line 2078
Quoted 2 lines
tools/perf/util/metricgroup.c +0
-0static int hw_aware_metricgroup__build_event_string(struct list_head
*group_strs,tools/perf/util/metricgroup.c static int
Quoted 6 lines
tools/perf/util/metricgroup.c +1
-0const char *modifier, + const bool tool_events[PERF_TOOL_MAX], struct list_head *groups) { struct metricgroup__pmu_group_list *p;@@ -2056,8 +2086,12 @@ static inthw_aware_metricgroup__build_event_string(struct
list_head*group_strsQuoted 6 lines
tools/perf/util/metricgroup.c +2
-0struct metricgroup__group_events *ge;bool no_group = true;int ret =0 ; +struct strbuf tool_event_str =STRBUF_INIT ; +bool has_tool_event = false;Quoted 6 lines
tools/perf/util/metricgroup.c static int · line 2093
Quoted 2 lines
tools/perf/util/metricgroup.c +1
-0#define RETURN_IF_NON_ZERO (x) do { if (x) return x; } while (0) + ret = get_tool_event_str(&tool_event_str, tool_events,
&has_tool_event);Does
metricgroup__build_event_stringneed updating to use this helper function too?
Quoted 7 lines
Hi Ian,
In
parse_ids, tool events are inserted to the ids that passed tometricgroup__build_event_string. If we want to use this function here, I think we also need to update the code inparse_idsto not insert tool events.I could add this change if you think we should do this update.
I like code-reuse and sharing logic whenever possible, it seems potentially messy though. I'll leave it up to your judgement as to whether it is worth cleaning up. Perhaps there should be a TODO if not.
Thanks,
Ian
Quoted 86 lines
Thanks,
WeilinQuoted 83 lines
Thanks,
Iantools/perf/util/metricgroup.c static int
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+RETURN_IF_NON_ZERO (ret);list_for_each_entry (p, groups, nd) {list_for_each_entry (g, &p->group_head, nd) {@@ -2129,6 +2163,12 @@ static inthw_aware_metricgroup__build_event_string(struct
list_head*group_strstools/perf/util/metricgroup.c static int hw_aware_metricgroup__build_event_string(struct list_head *group_strs
Quoted 5 lines
tools/perf/util/metricgroup.c +2
-0} ret =strbuf_addf (events,"}:W" );RETURN_IF_NON_ZERO (ret); + +if (!strcmp (p->pmu_name,"default_core" ) && has_tool_event){
tools/perf/util/metricgroup.c static int hw_aware_build_grouping(struct
Quoted 8 lines
tools/perf/util/metricgroup.c +4
-0+ ret =strbuf_addstr (events, tool_event_str.buf); +RETURN_IF_NON_ZERO (ret); + } +pr_debug ("events-buf: %s\n" , events->buf);list_add_tail (&new_group_str->nd, group_strs); }@@ -2214,6 +2254,7 @@ static int hw_aware_build_grouping(struct
expr_parse_ctx*ctx,tools/perf/util/metricgroup.c static void
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0if (ret) goto err_out; } + ret =get_pmu_counter_layouts (&pmu_info_list, ltable);if (ret) goto err_out;@@ -2259,6 +2300,7 @@ static voidmetricgroup__free_grouping_strs(struct
list_headtools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0*/ static int hw_aware_parse_ids (struct perf_pmu *fake_pmu,struct expr_parse_ctx *ids,const char *modifier, +const bool tool_events[PERF_TOOL_MAX ],struct evlist **out_evlist) {struct parse_events_error parse_error;@@ -2272,7 +2314,8 @@ static int hw_aware_parse_ids(struct perf_pmu
*fake_pmu,tools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu *fake_pmu,
Quoted 4 lines
tools/perf/util/metricgroup.c +0
-1ret =hw_aware_build_grouping (ids, &grouping);if (ret) goto out; - ret = hw_aware_metricgroup__build_event_string(&grouping_str,modifier, &grouping);
tools/perf/util/metricgroup.c static int hw_aware_parse_ids(struct perf_pmu · line 2317
Quoted 1 line
tools/perf/util/metricgroup.c +1
-0+ ret = hw_aware_metricgroup__build_event_string(&grouping_str,modifier,
tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist
Quoted 5 lines
tools/perf/util/metricgroup.c +1
-0+ tool_events, &grouping);if (ret) goto out;@@ -2407,6 +2450,7 @@ static int hw_aware_parse_groups(struct evlist
*perf_evlist,tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist
Quoted 8 lines
tools/perf/util/metricgroup.c +1
-0struct evlist *combined_evlist =NULL ;LIST_HEAD (metric_list);struct metric *m; +bool tool_events[PERF_TOOL_MAX ] = {false};int ret;bool metric_no_group = false;bool metric_no_merge = false;@@ -2425,11 +2469,14 @@ static int hw_aware_parse_groups(struct evlist
*perf_evlist,tools/perf/util/metricgroup.c static int hw_aware_parse_groups(struct evlist *perf_evlist,
Quoted 6 lines
tools/perf/util/metricgroup.c +2
-0if (!metric_no_merge) {struct expr_parse_ctx *combined =NULL ; +find_tool_events (&metric_list, tool_events); + ret = hw_aware_build_combined_expr_ctx(&metric_list,&combined);
Quoted 11 lines
tools/perf/util/metricgroup.c +1
-1if (!ret && combined &&hashmap__size (combined->ids)) { ret =hw_aware_parse_ids (fake_pmu, combined,/*modifier=*/ NULL , + tool_events, &combined_evlist); } --2.42.0
RFC PATCH v4 15/15 perf stat: Add hardware-grouping cmd option to perf stat
In reply to [email protected]From: Weilin Wang <[email protected]>
Add a cmd option to allow user to choose this new metric grouping method.
$ perf stat -M TopdownL1 -a --hardware-grouping
--- tools/perf/builtin-stat.c | 2 ++ 1 file changed, 2 insertions(+)
tools/perf/builtin-stat.c +2 -0
@@ -1257,6 +1257,8 @@ static struct option stat_options[] = { "don't try to share events between metrics in a group"),OPT_BOOLEAN (0 ,"metric-no-threshold" , &stat_config.metric_no_threshold,"disable adding events for the metric threshold calculation" ), +OPT_BOOLEAN (0 ,"hardware-grouping" , &stat_config.hardware_aware_grouping, +"Use hardware aware metric grouping method" ),OPT_BOOLEAN (0 ,"topdown" , &topdown_run,"measure top-down statistics" ), OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
Signature · 1 line
2.42.0
Re: [RFC PATCH v4 15/15] perf stat: Add hardware-grouping cmd option to perf stat
Quoted 8 lines · On Thu, Feb 8, 2024 at 7:14 PM <[email protected]> wrote:
From: Weilin Wang <[email protected]>
Add a cmd option to allow user to choose this new metric grouping method.
$ perf stat -M TopdownL1 -a --hardware-groupingSigned-off-by Weilin Wang <[email protected]>
Thanks,
Ian
Quoted 20 lines
--- tools/perf/builtin-stat.c | 2 ++ 1 file changed, 2 insertions(+)tools/perf/builtin-stat.c +2
-1@@ -1257,6 +1257,8 @@ static struct option stat_options[] = { "don't try to share events between metrics in a group"),OPT_BOOLEAN (0 ,"metric-no-threshold" , &stat_config.metric_no_threshold,"disable adding events for the metric threshold calculation" ), +OPT_BOOLEAN (0 ,"hardware-grouping" , &stat_config.hardware_aware_grouping, +"Use hardware aware metric grouping method" ),OPT_BOOLEAN (0 ,"topdown" , &topdown_run,"measure top-down statistics" ), OPT_UINTEGER(0, "td-level", &stat_config.topdown_level, --2.42.0