Message-ID: <b9599b44-79a6-4a09-9424-3add58ad53f0@linux.intel.com>
Date:   Mon, 30 Oct 2023 11:07:21 -0400
From:   "Liang, Kan" <kan.liang@...ux.intel.com>
To:     weilin.wang@...el.com, Ian Rogers <irogers@...gle.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Arnaldo Carvalho de Melo <acme@...nel.org>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        Adrian Hunter <adrian.hunter@...el.com>
Cc:     linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org,
        Perry Taylor <perry.taylor@...el.com>,
        Samantha Alt <samantha.alt@...el.com>,
        Caleb Biggers <caleb.biggers@...el.com>,
        Mark Rutland <mark.rutland@....com>,
        Yang Jihong <yangjihong1@...wei.com>
Subject: Re: [RFC PATCH v2 02/17] perf stat: Add basic functions for the
 hardware-grouping stat cmd option



On 2023-10-13 9:51 p.m., weilin.wang@...el.com wrote:
> From: Weilin Wang <weilin.wang@...el.com>
> 
> Add the first set of functions for the hardware-grouping method. Function
> hw_awre_parse_groups() is the entry point of this metric grouping method.

hw_aware_parse_groups()

> 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 mannor as existing metric grouping

manner

> 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.
> 
> Signed-off-by: Weilin Wang <weilin.wang@...el.com>
> ---
>  tools/perf/util/metricgroup.c | 210 +++++++++++++++++++++++++++++++++-
>  tools/perf/util/metricgroup.h |   9 ++
>  2 files changed, 218 insertions(+), 1 deletion(-)
> 
> diff --git a/tools/perf/util/metricgroup.c b/tools/perf/util/metricgroup.c
> index b08af6860..8d4e29eb1 100644
> --- a/tools/perf/util/metricgroup.c
> +++ b/tools/perf/util/metricgroup.c
> @@ -1432,6 +1432,101 @@ 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");

I think the patch series still use a reversed way to organize the
patches. So you have to insert such the useless place holder here.

I think a better way is to start the series from the essential blocks,
then build more complex functions on top of the blocks, and finally
add/enable the new feature.

> +	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)

The hw_aware_parse_ids() is very similar to parse_ids(). Can we factor a
common function to avoid the duplication?


> +{
> +	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 +1615,114 @@ 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)

The hw_aware_parse_groups() is very similar to parse_groups(). Please
factor out a common function to avoid the duplication as well.

Thanks,
Kan
> +{
> +	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 +1901,15 @@ 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);
> diff --git a/tools/perf/util/metricgroup.h b/tools/perf/util/metricgroup.h
> index 779f6ede1..89809df85 100644
> --- a/tools/perf/util/metricgroup.h
> +++ b/tools/perf/util/metricgroup.h
> @@ -6,6 +6,7 @@
>  #include <linux/rbtree.h>
>  #include <stdbool.h>
>  #include "pmu-events/pmu-events.h"
> +#include "strbuf.h"
>  
>  struct evlist;
>  struct evsel;
> @@ -66,6 +67,14 @@ struct metric_expr {
>  	int runtime;
>  };
>  
> +/**
> + * Each group is one node in the group string list.
> + */
> +struct metricgroup__group_strs {
> +	struct list_head nd;
> +	struct strbuf grouping_str;
> +};
> +
>  struct metric_event *metricgroup__lookup(struct rblist *metric_events,
>  					 struct evsel *evsel,
>  					 bool create);

Powered by blists - more mailing lists