metricgroup.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (c) 2017, Intel Corporation.
   4 */
   5
   6/* Manage metrics and groups of metrics from JSON files */
   7
   8#include "metricgroup.h"
   9#include "debug.h"
  10#include "evlist.h"
  11#include "evsel.h"
  12#include "strbuf.h"
  13#include "pmu.h"
  14#include "pmus.h"
  15#include "print-events.h"
  16#include "smt.h"
  17#include "tool_pmu.h"
  18#include "expr.h"
  19#include "rblist.h"
  20#include <string.h>
  21#include <errno.h>
  22#include "strlist.h"
  23#include <assert.h>
  24#include <linux/ctype.h>
  25#include <linux/list_sort.h>
  26#include <linux/string.h>
  27#include <linux/zalloc.h>
  28#include <perf/cpumap.h>
  29#include <subcmd/parse-options.h>
  30#include <api/fs/fs.h>
  31#include "util.h"
  32#include <asm/bug.h>
  33#include "cgroup.h"
  34#include "util/hashmap.h"
  35
  36struct metric_event *metricgroup__lookup(struct rblist *metric_events,
  37					 struct evsel *evsel,
  38					 bool create)
  39{
  40	struct rb_node *nd;
  41	struct metric_event me = {
  42		.evsel = evsel
  43	};
  44
  45	if (!metric_events)
  46		return NULL;
  47
  48	if (evsel && evsel->metric_leader)
  49		me.evsel = evsel->metric_leader;
  50	nd = rblist__find(metric_events, &me);
  51	if (nd)
  52		return container_of(nd, struct metric_event, nd);
  53	if (create) {
  54		rblist__add_node(metric_events, &me);
  55		nd = rblist__find(metric_events, &me);
  56		if (nd)
  57			return container_of(nd, struct metric_event, nd);
  58	}
  59	return NULL;
  60}
  61
  62static int metric_event_cmp(struct rb_node *rb_node, const void *entry)
  63{
  64	struct metric_event *a = container_of(rb_node,
  65					      struct metric_event,
  66					      nd);
  67	const struct metric_event *b = entry;
  68
  69	if (a->evsel == b->evsel)
  70		return 0;
  71	if ((char *)a->evsel < (char *)b->evsel)
  72		return -1;
  73	return +1;
  74}
  75
  76static struct rb_node *metric_event_new(struct rblist *rblist __maybe_unused,
  77					const void *entry)
  78{
  79	struct metric_event *me = malloc(sizeof(struct metric_event));
  80
  81	if (!me)
  82		return NULL;
  83	memcpy(me, entry, sizeof(struct metric_event));
  84	me->evsel = ((struct metric_event *)entry)->evsel;
  85	me->is_default = false;
  86	INIT_LIST_HEAD(&me->head);
  87	return &me->nd;
  88}
  89
  90static void metric_event_delete(struct rblist *rblist __maybe_unused,
  91				struct rb_node *rb_node)
  92{
  93	struct metric_event *me = container_of(rb_node, struct metric_event, nd);
  94	struct metric_expr *expr, *tmp;
  95
  96	list_for_each_entry_safe(expr, tmp, &me->head, nd) {
  97		zfree(&expr->metric_name);
  98		zfree(&expr->metric_refs);
  99		zfree(&expr->metric_events);
 100		free(expr);
 101	}
 102
 103	free(me);
 104}
 105
 106void metricgroup__rblist_init(struct rblist *metric_events)
 107{
 108	rblist__init(metric_events);
 109	metric_events->node_cmp = metric_event_cmp;
 110	metric_events->node_new = metric_event_new;
 111	metric_events->node_delete = metric_event_delete;
 112}
 113
 114void metricgroup__rblist_exit(struct rblist *metric_events)
 115{
 116	rblist__exit(metric_events);
 117}
 118
 119/**
 120 * The metric under construction. The data held here will be placed in a
 121 * metric_expr.
 122 */
 123struct metric {
 124	struct list_head nd;
 125	/**
 126	 * The expression parse context importantly holding the IDs contained
 127	 * within the expression.
 128	 */
 129	struct expr_parse_ctx *pctx;
 130	const char *pmu;
 131	/** The name of the metric such as "IPC". */
 132	const char *metric_name;
 133	/** Modifier on the metric such as "u" or NULL for none. */
 134	const char *modifier;
 135	/** The expression to parse, for example, "instructions/cycles". */
 136	const char *metric_expr;
 137	/** Optional threshold expression where zero value is green, otherwise red. */
 138	const char *metric_threshold;
 139	/**
 140	 * The "ScaleUnit" that scales and adds a unit to the metric during
 141	 * output.
 142	 */
 143	const char *metric_unit;
 144	/**
 145	 * Optional name of the metric group reported
 146	 * if the Default metric group is being processed.
 147	 */
 148	const char *default_metricgroup_name;
 149	/** Optional null terminated array of referenced metrics. */
 150	struct metric_ref *metric_refs;
 151	/**
 152	 * Should events of the metric be grouped?
 153	 */
 154	bool group_events;
 155	/** Show events even if in the Default metric group. */
 156	bool default_show_events;
 157	/**
 158	 * Parsed events for the metric. Optional as events may be taken from a
 159	 * different metric whose group contains all the IDs necessary for this
 160	 * one.
 161	 */
 162	struct evlist *evlist;
 163};
 164
 165static void metric__watchdog_constraint_hint(const char *name, bool foot)
 166{
 167	static bool violate_nmi_constraint;
 168
 169	if (!foot) {
 170		pr_warning("Not grouping metric %s's events.\n", name);
 171		violate_nmi_constraint = true;
 172		return;
 173	}
 174
 175	if (!violate_nmi_constraint)
 176		return;
 177
 178	pr_warning("Try disabling the NMI watchdog to comply NO_NMI_WATCHDOG metric constraint:\n"
 179		   "    echo 0 > /proc/sys/kernel/nmi_watchdog\n"
 180		   "    perf stat ...\n"
 181		   "    echo 1 > /proc/sys/kernel/nmi_watchdog\n");
 182}
 183
 184static bool metric__group_events(const struct pmu_metric *pm, bool metric_no_threshold)
 185{
 186	switch (pm->event_grouping) {
 187	case MetricNoGroupEvents:
 188		return false;
 189	case MetricNoGroupEventsNmi:
 190		if (!sysctl__nmi_watchdog_enabled())
 191			return true;
 192		metric__watchdog_constraint_hint(pm->metric_name, /*foot=*/false);
 193		return false;
 194	case MetricNoGroupEventsSmt:
 195		return !smt_on();
 196	case MetricNoGroupEventsThresholdAndNmi:
 197		if (metric_no_threshold)
 198			return true;
 199		if (!sysctl__nmi_watchdog_enabled())
 200			return true;
 201		metric__watchdog_constraint_hint(pm->metric_name, /*foot=*/false);
 202		return false;
 203	case MetricGroupEvents:
 204	default:
 205		return true;
 206	}
 207}
 208
 209static void metric__free(struct metric *m)
 210{
 211	if (!m)
 212		return;
 213
 214	zfree(&m->metric_refs);
 215	expr__ctx_free(m->pctx);
 216	zfree(&m->modifier);
 217	evlist__delete(m->evlist);
 218	free(m);
 219}
 220
 221static struct metric *metric__new(const struct pmu_metric *pm,
 222				  const char *modifier,
 223				  bool metric_no_group,
 224				  bool metric_no_threshold,
 225				  int runtime,
 226				  const char *user_requested_cpu_list,
 227				  bool system_wide)
 228{
 229	struct metric *m;
 230
 231	m = zalloc(sizeof(*m));
 232	if (!m)
 233		return NULL;
 234
 235	m->pctx = expr__ctx_new();
 236	if (!m->pctx)
 237		goto out_err;
 238
 239	m->pmu = pm->pmu ?: "cpu";
 240	m->metric_name = pm->metric_name;
 241	m->default_metricgroup_name = pm->default_metricgroup_name ?: "";
 242	m->modifier = NULL;
 243	if (modifier) {
 244		m->modifier = strdup(modifier);
 245		if (!m->modifier)
 246			goto out_err;
 247	}
 248	m->metric_expr = pm->metric_expr;
 249	m->metric_threshold = pm->metric_threshold;
 250	m->metric_unit = pm->unit;
 251	m->pctx->sctx.user_requested_cpu_list = NULL;
 252	if (user_requested_cpu_list) {
 253		m->pctx->sctx.user_requested_cpu_list = strdup(user_requested_cpu_list);
 254		if (!m->pctx->sctx.user_requested_cpu_list)
 255			goto out_err;
 256	}
 257	m->pctx->sctx.runtime = runtime;
 258	m->pctx->sctx.system_wide = system_wide;
 259	m->group_events = !metric_no_group && metric__group_events(pm, metric_no_threshold);
 260	m->default_show_events = pm->default_show_events;
 261	m->metric_refs = NULL;
 262	m->evlist = NULL;
 263
 264	return m;
 265out_err:
 266	metric__free(m);
 267	return NULL;
 268}
 269
 270static bool contains_metric_id(struct evsel **metric_events, int num_events,
 271			       const char *metric_id)
 272{
 273	int i;
 274
 275	for (i = 0; i < num_events; i++) {
 276		if (!strcmp(evsel__metric_id(metric_events[i]), metric_id))
 277			return true;
 278	}
 279	return false;
 280}
 281
 282/**
 283 * setup_metric_events - Find a group of events in metric_evlist that correspond
 284 *                       to the IDs from a parsed metric expression.
 285 * @pmu: The PMU for the IDs.
 286 * @ids: the metric IDs to match.
 287 * @metric_evlist: the list of perf events.
 288 * @out_metric_events: holds the created metric events array.
 289 */
 290static int setup_metric_events(const char *pmu, struct hashmap *ids,
 291			       struct evlist *metric_evlist,
 292			       struct evsel ***out_metric_events)
 293{
 294	struct evsel **metric_events;
 295	const char *metric_id;
 296	struct evsel *ev;
 297	size_t ids_size, matched_events, i;
 298	bool all_pmus = !strcmp(pmu, "all") || perf_pmus__num_core_pmus() == 1 || !is_pmu_core(pmu);
 299
 300	*out_metric_events = NULL;
 301	ids_size = hashmap__size(ids);
 302
 303	metric_events = calloc(ids_size + 1, sizeof(void *));
 304	if (!metric_events)
 305		return -ENOMEM;
 306
 307	matched_events = 0;
 308	evlist__for_each_entry(metric_evlist, ev) {
 309		struct expr_id_data *val_ptr;
 310
 311		/* Don't match events for the wrong hybrid PMU. */
 312		if (!all_pmus && ev->pmu && evsel__is_hybrid(ev) &&
 313		    strcmp(ev->pmu->name, pmu))
 314			continue;
 315		/*
 316		 * Check for duplicate events with the same name. For
 317		 * example, uncore_imc/cas_count_read/ will turn into 6
 318		 * events per socket on skylakex. Only the first such
 319		 * event is placed in metric_events.
 320		 */
 321		metric_id = evsel__metric_id(ev);
 322		if (contains_metric_id(metric_events, matched_events, metric_id))
 323			continue;
 324		/*
 325		 * Does this event belong to the parse context? For
 326		 * combined or shared groups, this metric may not care
 327		 * about this event.
 328		 */
 329		if (hashmap__find(ids, metric_id, &val_ptr)) {
 330			pr_debug("Matched metric-id %s to %s\n", metric_id, evsel__name(ev));
 331			metric_events[matched_events++] = ev;
 332
 333			if (matched_events >= ids_size)
 334				break;
 335		}
 336	}
 337	if (matched_events < ids_size) {
 338		free(metric_events);
 339		return -EINVAL;
 340	}
 341	for (i = 0; i < ids_size; i++) {
 342		ev = metric_events[i];
 343		ev->collect_stat = true;
 344
 345		/*
 346		 * The metric leader points to the identically named
 347		 * event in metric_events.
 348		 */
 349		ev->metric_leader = ev;
 350		/*
 351		 * Mark two events with identical names in the same
 352		 * group (or globally) as being in use as uncore events
 353		 * may be duplicated for each pmu. Set the metric leader
 354		 * of such events to be the event that appears in
 355		 * metric_events.
 356		 */
 357		metric_id = evsel__metric_id(ev);
 358		evlist__for_each_entry_continue(metric_evlist, ev) {
 359			if (!strcmp(evsel__metric_id(ev), metric_id))
 360				ev->metric_leader = metric_events[i];
 361		}
 362	}
 363	*out_metric_events = metric_events;
 364	return 0;
 365}
 366
 367static bool match_metric_or_groups(const char *metric_or_groups, const char *sought)
 368{
 369	int len;
 370	char *m;
 371
 372	if (!sought)
 373		return false;
 374	if (!strcmp(sought, "all"))
 375		return true;
 376	if (!metric_or_groups)
 377		return !strcasecmp(sought, "No_group");
 378	len = strlen(sought);
 379	if (!strncasecmp(metric_or_groups, sought, len) &&
 380	    (metric_or_groups[len] == 0 || metric_or_groups[len] == ';'))
 381		return true;
 382	m = strchr(metric_or_groups, ';');
 383	return m && match_metric_or_groups(m + 1, sought);
 384}
 385
 386static bool match_pm_metric_or_groups(const struct pmu_metric *pm, const char *pmu,
 387				      const char *metric_or_groups)
 388{
 389	const char *pm_pmu = pm->pmu ?: "cpu";
 390
 391	if (strcmp(pmu, "all") && strcmp(pm_pmu, pmu))
 392		return false;
 393
 394	return match_metric_or_groups(pm->metric_group, metric_or_groups) ||
 395	       match_metric_or_groups(pm->metric_name, metric_or_groups);
 396}
 397
 398struct metricgroup_iter_data {
 399	pmu_metric_iter_fn fn;
 400	void *data;
 401};
 402
 403static int metricgroup__sys_event_iter(const struct pmu_metric *pm,
 404				       const struct pmu_metrics_table *table,
 405				       void *data)
 406{
 407	struct metricgroup_iter_data *d = data;
 408	struct perf_pmu *pmu = NULL;
 409
 410	if (!pm->metric_expr || !pm->compat)
 411		return 0;
 412
 413	while ((pmu = perf_pmus__scan(pmu))) {
 414
 415		if (!pmu->id || !pmu_uncore_identifier_match(pm->compat, pmu->id))
 416			continue;
 417
 418		return d->fn(pm, table, d->data);
 419	}
 420	return 0;
 421}
 422
 423int metricgroup__for_each_metric(const struct pmu_metrics_table *table, pmu_metric_iter_fn fn,
 424				 void *data)
 425{
 426	struct metricgroup_iter_data sys_data = {
 427		.fn = fn,
 428		.data = data,
 429	};
 430	const struct pmu_metrics_table *tables[2] = {
 431		table,
 432		pmu_metrics_table__default(),
 433	};
 434
 435	for (size_t i = 0; i < ARRAY_SIZE(tables); i++) {
 436		int ret;
 437
 438		if (!tables[i])
 439			continue;
 440
 441		ret = pmu_metrics_table__for_each_metric(tables[i], fn, data);
 442		if (ret)
 443			return ret;
 444	}
 445
 446	return pmu_for_each_sys_metric(metricgroup__sys_event_iter, &sys_data);
 447}
 448
 449static const char *code_characters = ",-=@";
 450
 451static int encode_metric_id(struct strbuf *sb, const char *x)
 452{
 453	char *c;
 454	int ret = 0;
 455
 456	for (; *x; x++) {
 457		c = strchr(code_characters, *x);
 458		if (c) {
 459			ret = strbuf_addch(sb, '!');
 460			if (ret)
 461				break;
 462
 463			ret = strbuf_addch(sb, '0' + (c - code_characters));
 464			if (ret)
 465				break;
 466		} else {
 467			ret = strbuf_addch(sb, *x);
 468			if (ret)
 469				break;
 470		}
 471	}
 472	return ret;
 473}
 474
 475static int decode_metric_id(struct strbuf *sb, const char *x)
 476{
 477	const char *orig = x;
 478	size_t i;
 479	char c;
 480	int ret;
 481
 482	for (; *x; x++) {
 483		c = *x;
 484		if (*x == '!') {
 485			x++;
 486			i = *x - '0';
 487			if (i > strlen(code_characters)) {
 488				pr_err("Bad metric-id encoding in: '%s'", orig);
 489				return -1;
 490			}
 491			c = code_characters[i];
 492		}
 493		ret = strbuf_addch(sb, c);
 494		if (ret)
 495			return ret;
 496	}
 497	return 0;
 498}
 499
 500static int decode_all_metric_ids(struct evlist *perf_evlist, const char *modifier)
 501{
 502	struct evsel *ev;
 503	struct strbuf sb = STRBUF_INIT;
 504	char *cur;
 505	int ret = 0;
 506
 507	evlist__for_each_entry(perf_evlist, ev) {
 508		if (!ev->metric_id)
 509			continue;
 510
 511		ret = strbuf_setlen(&sb, 0);
 512		if (ret)
 513			break;
 514
 515		ret = decode_metric_id(&sb, ev->metric_id);
 516		if (ret)
 517			break;
 518
 519		free((char *)ev->metric_id);
 520		ev->metric_id = strdup(sb.buf);
 521		if (!ev->metric_id) {
 522			ret = -ENOMEM;
 523			break;
 524		}
 525		/*
 526		 * If the name is just the parsed event, use the metric-id to
 527		 * give a more friendly display version.
 528		 */
 529		if (strstr(ev->name, "metric-id=")) {
 530			bool has_slash = false;
 531
 532			zfree(&ev->name);
 533			for (cur = strchr(sb.buf, '@') ; cur; cur = strchr(++cur, '@')) {
 534				*cur = '/';
 535				has_slash = true;
 536			}
 537
 538			if (modifier) {
 539				if (!has_slash && !strchr(sb.buf, ':')) {
 540					ret = strbuf_addch(&sb, ':');
 541					if (ret)
 542						break;
 543				}
 544				ret = strbuf_addstr(&sb, modifier);
 545				if (ret)
 546					break;
 547			}
 548			ev->name = strdup(sb.buf);
 549			if (!ev->name) {
 550				ret = -ENOMEM;
 551				break;
 552			}
 553		}
 554	}
 555	strbuf_release(&sb);
 556	return ret;
 557}
 558
 559static int metricgroup__build_event_string(struct strbuf *events,
 560					   const struct expr_parse_ctx *ctx,
 561					   const char *modifier,
 562					   bool group_events)
 563{
 564	struct hashmap_entry *cur;
 565	size_t bkt;
 566	bool no_group = true, has_tool_events = false;
 567	bool tool_events[TOOL_PMU__EVENT_MAX] = {false};
 568	int ret = 0;
 569
 570#define RETURN_IF_NON_ZERO(x) do { if (x) return x; } while (0)
 571
 572	hashmap__for_each_entry(ctx->ids, cur, bkt) {
 573		const char *sep, *rsep, *id = cur->pkey;
 574		enum tool_pmu_event ev;
 575
 576		pr_debug("found event %s\n", id);
 577
 578		/* Always move tool events outside of the group. */
 579		ev = tool_pmu__str_to_event(id);
 580		if (ev != TOOL_PMU__EVENT_NONE) {
 581			has_tool_events = true;
 582			tool_events[ev] = true;
 583			continue;
 584		}
 585		/* Separate events with commas and open the group if necessary. */
 586		if (no_group) {
 587			if (group_events) {
 588				ret = strbuf_addch(events, '{');
 589				RETURN_IF_NON_ZERO(ret);
 590			}
 591
 592			no_group = false;
 593		} else {
 594			ret = strbuf_addch(events, ',');
 595			RETURN_IF_NON_ZERO(ret);
 596		}
 597		/*
 598		 * Encode the ID as an event string. Add a qualifier for
 599		 * metric_id that is the original name except with characters
 600		 * that parse-events can't parse replaced. For example,
 601		 * 'msr@tsc@' gets added as msr/tsc,metric-id=msr!3tsc!3/
 602		 */
 603		sep = strchr(id, '@');
 604		if (sep != NULL) {
 605			ret = strbuf_add(events, id, sep - id);
 606			RETURN_IF_NON_ZERO(ret);
 607			ret = strbuf_addch(events, '/');
 608			RETURN_IF_NON_ZERO(ret);
 609			rsep = strrchr(sep, '@');
 610			ret = strbuf_add(events, sep + 1, rsep - sep - 1);
 611			RETURN_IF_NON_ZERO(ret);
 612			ret = strbuf_addstr(events, ",metric-id=");
 613			RETURN_IF_NON_ZERO(ret);
 614			sep = rsep;
 615		} else {
 616			sep = strchr(id, ':');
 617			if (sep != NULL) {
 618				ret = strbuf_add(events, id, sep - id);
 619				RETURN_IF_NON_ZERO(ret);
 620			} else {
 621				ret = strbuf_addstr(events, id);
 622				RETURN_IF_NON_ZERO(ret);
 623			}
 624			ret = strbuf_addstr(events, "/metric-id=");
 625			RETURN_IF_NON_ZERO(ret);
 626		}
 627		ret = encode_metric_id(events, id);
 628		RETURN_IF_NON_ZERO(ret);
 629		ret = strbuf_addstr(events, "/");
 630		RETURN_IF_NON_ZERO(ret);
 631
 632		if (sep != NULL) {
 633			ret = strbuf_addstr(events, sep + 1);
 634			RETURN_IF_NON_ZERO(ret);
 635		}
 636		if (modifier) {
 637			ret = strbuf_addstr(events, modifier);
 638			RETURN_IF_NON_ZERO(ret);
 639		}
 640	}
 641	if (!no_group && group_events) {
 642		ret = strbuf_addf(events, "}:W");
 643		RETURN_IF_NON_ZERO(ret);
 644	}
 645	if (has_tool_events) {
 646		int i;
 647
 648		tool_pmu__for_each_event(i) {
 649			if (tool_events[i]) {
 650				if (!no_group) {
 651					ret = strbuf_addch(events, ',');
 652					RETURN_IF_NON_ZERO(ret);
 653				}
 654				no_group = false;
 655				ret = strbuf_addstr(events, tool_pmu__event_to_str(i));
 656				RETURN_IF_NON_ZERO(ret);
 657			}
 658		}
 659	}
 660
 661	return ret;
 662#undef RETURN_IF_NON_ZERO
 663}
 664
 665int __weak arch_get_runtimeparam(const struct pmu_metric *pm __maybe_unused)
 666{
 667	return 1;
 668}
 669
 670/*
 671 * A singly linked list on the stack of the names of metrics being
 672 * processed. Used to identify recursion.
 673 */
 674struct visited_metric {
 675	const char *name;
 676	const struct visited_metric *parent;
 677};
 678
 679struct metricgroup_add_iter_data {
 680	struct list_head *metric_list;
 681	const char *pmu;
 682	const char *metric_name;
 683	const char *modifier;
 684	int *ret;
 685	bool *has_match;
 686	bool metric_no_group;
 687	bool metric_no_threshold;
 688	const char *user_requested_cpu_list;
 689	bool system_wide;
 690	struct metric *root_metric;
 691	const struct visited_metric *visited;
 692	const struct pmu_metrics_table *table;
 693};
 694
 695static int add_metric(struct list_head *metric_list,
 696		      const struct pmu_metric *pm,
 697		      const char *modifier,
 698		      bool metric_no_group,
 699		      bool metric_no_threshold,
 700		      const char *user_requested_cpu_list,
 701		      bool system_wide,
 702		      struct metric *root_metric,
 703		      const struct visited_metric *visited,
 704		      const struct pmu_metrics_table *table);
 705
 706static int metricgroup__find_metric_callback(const struct pmu_metric *pm,
 707					     const struct pmu_metrics_table *table  __maybe_unused,
 708					     void *vdata)
 709{
 710	struct pmu_metric *copied_pm = vdata;
 711
 712	memcpy(copied_pm, pm, sizeof(*pm));
 713	return 0;
 714}
 715
 716/**
 717 * resolve_metric - Locate metrics within the root metric and recursively add
 718 *                    references to them.
 719 * @metric_list: The list the metric is added to.
 720 * @pmu: The PMU name to resolve metrics on, or "all" for all PMUs.
 721 * @modifier: if non-null event modifiers like "u".
 722 * @metric_no_group: Should events written to events be grouped "{}" or
 723 *                   global. Grouping is the default but due to multiplexing the
 724 *                   user may override.
 725 * @user_requested_cpu_list: Command line specified CPUs to record on.
 726 * @system_wide: Are events for all processes recorded.
 727 * @root_metric: Metrics may reference other metrics to form a tree. In this
 728 *               case the root_metric holds all the IDs and a list of referenced
 729 *               metrics. When adding a root this argument is NULL.
 730 * @visited: A singly linked list of metric names being added that is used to
 731 *           detect recursion.
 732 * @table: The table that is searched for metrics, most commonly the table for the
 733 *       architecture perf is running upon.
 734 */
 735static int resolve_metric(struct list_head *metric_list,
 736			  struct perf_pmu *pmu,
 737			  const char *modifier,
 738			  bool metric_no_group,
 739			  bool metric_no_threshold,
 740			  const char *user_requested_cpu_list,
 741			  bool system_wide,
 742			  struct metric *root_metric,
 743			  const struct visited_metric *visited,
 744			  const struct pmu_metrics_table *table)
 745{
 746	struct hashmap_entry *cur;
 747	size_t bkt;
 748	struct to_resolve {
 749		/* The metric to resolve. */
 750		struct pmu_metric pm;
 751		/*
 752		 * The key in the IDs map, this may differ from in case,
 753		 * etc. from pm->metric_name.
 754		 */
 755		const char *key;
 756	} *pending = NULL;
 757	int i, ret = 0, pending_cnt = 0;
 758
 759	/*
 760	 * Iterate all the parsed IDs and if there's a matching metric and it to
 761	 * the pending array.
 762	 */
 763	hashmap__for_each_entry(root_metric->pctx->ids, cur, bkt) {
 764		struct pmu_metric pm;
 765
 766		if (pmu_metrics_table__find_metric(table, pmu, cur->pkey,
 767						   metricgroup__find_metric_callback,
 768						   &pm) != PMU_METRICS__NOT_FOUND) {
 769			pending = realloc(pending,
 770					(pending_cnt + 1) * sizeof(struct to_resolve));
 771			if (!pending)
 772				return -ENOMEM;
 773
 774			memcpy(&pending[pending_cnt].pm, &pm, sizeof(pm));
 775			pending[pending_cnt].key = cur->pkey;
 776			pending_cnt++;
 777		}
 778	}
 779
 780	/* Remove the metric IDs from the context. */
 781	for (i = 0; i < pending_cnt; i++)
 782		expr__del_id(root_metric->pctx, pending[i].key);
 783
 784	/*
 785	 * Recursively add all the metrics, IDs are added to the root metric's
 786	 * context.
 787	 */
 788	for (i = 0; i < pending_cnt; i++) {
 789		ret = add_metric(metric_list, &pending[i].pm, modifier, metric_no_group,
 790				 metric_no_threshold, user_requested_cpu_list, system_wide,
 791				 root_metric, visited, table);
 792		if (ret)
 793			break;
 794	}
 795
 796	free(pending);
 797	return ret;
 798}
 799
 800/**
 801 * __add_metric - Add a metric to metric_list.
 802 * @metric_list: The list the metric is added to.
 803 * @pm: The pmu_metric containing the metric to be added.
 804 * @modifier: if non-null event modifiers like "u".
 805 * @metric_no_group: Should events written to events be grouped "{}" or
 806 *                   global. Grouping is the default but due to multiplexing the
 807 *                   user may override.
 808 * @metric_no_threshold: Should threshold expressions be ignored?
 809 * @runtime: A special argument for the parser only known at runtime.
 810 * @user_requested_cpu_list: Command line specified CPUs to record on.
 811 * @system_wide: Are events for all processes recorded.
 812 * @root_metric: Metrics may reference other metrics to form a tree. In this
 813 *               case the root_metric holds all the IDs and a list of referenced
 814 *               metrics. When adding a root this argument is NULL.
 815 * @visited: A singly linked list of metric names being added that is used to
 816 *           detect recursion.
 817 * @table: The table that is searched for metrics, most commonly the table for the
 818 *       architecture perf is running upon.
 819 */
 820static int __add_metric(struct list_head *metric_list,
 821			const struct pmu_metric *pm,
 822			const char *modifier,
 823			bool metric_no_group,
 824			bool metric_no_threshold,
 825			int runtime,
 826			const char *user_requested_cpu_list,
 827			bool system_wide,
 828			struct metric *root_metric,
 829			const struct visited_metric *visited,
 830			const struct pmu_metrics_table *table)
 831{
 832	const struct visited_metric *vm;
 833	int ret;
 834	bool is_root = !root_metric;
 835	const char *expr;
 836	struct visited_metric visited_node = {
 837		.name = pm->metric_name,
 838		.parent = visited,
 839	};
 840
 841	for (vm = visited; vm; vm = vm->parent) {
 842		if (!strcmp(pm->metric_name, vm->name)) {
 843			pr_err("failed: recursion detected for %s\n", pm->metric_name);
 844			return -1;
 845		}
 846	}
 847
 848	if (is_root) {
 849		/*
 850		 * This metric is the root of a tree and may reference other
 851		 * metrics that are added recursively.
 852		 */
 853		root_metric = metric__new(pm, modifier, metric_no_group, metric_no_threshold,
 854					  runtime, user_requested_cpu_list, system_wide);
 855		if (!root_metric)
 856			return -ENOMEM;
 857
 858	} else {
 859		int cnt = 0;
 860
 861		/*
 862		 * This metric was referenced in a metric higher in the
 863		 * tree. Check if the same metric is already resolved in the
 864		 * metric_refs list.
 865		 */
 866		if (root_metric->metric_refs) {
 867			for (; root_metric->metric_refs[cnt].metric_name; cnt++) {
 868				if (!strcmp(pm->metric_name,
 869					    root_metric->metric_refs[cnt].metric_name))
 870					return 0;
 871			}
 872		}
 873
 874		/* Create reference. Need space for the entry and the terminator. */
 875		root_metric->metric_refs = realloc(root_metric->metric_refs,
 876						(cnt + 2) * sizeof(struct metric_ref));
 877		if (!root_metric->metric_refs)
 878			return -ENOMEM;
 879
 880		/*
 881		 * Intentionally passing just const char pointers,
 882		 * from 'pe' object, so they never go away. We don't
 883		 * need to change them, so there's no need to create
 884		 * our own copy.
 885		 */
 886		root_metric->metric_refs[cnt].metric_name = pm->metric_name;
 887		root_metric->metric_refs[cnt].metric_expr = pm->metric_expr;
 888
 889		/* Null terminate array. */
 890		root_metric->metric_refs[cnt+1].metric_name = NULL;
 891		root_metric->metric_refs[cnt+1].metric_expr = NULL;
 892	}
 893
 894	/*
 895	 * For both the parent and referenced metrics, we parse
 896	 * all the metric's IDs and add it to the root context.
 897	 */
 898	ret = 0;
 899	expr = pm->metric_expr;
 900	if (is_root && pm->metric_threshold) {
 901		/*
 902		 * Threshold expressions are built off the actual metric. Switch
 903		 * to use that in case of additional necessary events. Change
 904		 * the visited node name to avoid this being flagged as
 905		 * recursion. If the threshold events are disabled, just use the
 906		 * metric's name as a reference. This allows metric threshold
 907		 * computation if there are sufficient events.
 908		 */
 909		assert(strstr(pm->metric_threshold, pm->metric_name));
 910		expr = metric_no_threshold ? pm->metric_name : pm->metric_threshold;
 911		visited_node.name = "__threshold__";
 912	}
 913	if (expr__find_ids(expr, NULL, root_metric->pctx) < 0) {
 914		/* Broken metric. */
 915		ret = -EINVAL;
 916	}
 917	if (!ret) {
 918		/* Resolve referenced metrics. */
 919		struct perf_pmu *pmu;
 920
 921		if (pm->pmu && pm->pmu[0] != '\0')
 922			pmu = perf_pmus__find(pm->pmu);
 923		else
 924			pmu = perf_pmus__scan_core(/*pmu=*/ NULL);
 925
 926		ret = resolve_metric(metric_list, pmu, modifier, metric_no_group,
 927				     metric_no_threshold, user_requested_cpu_list,
 928				     system_wide, root_metric, &visited_node,
 929				     table);
 930	}
 931	if (ret) {
 932		if (is_root)
 933			metric__free(root_metric);
 934
 935	} else if (is_root)
 936		list_add(&root_metric->nd, metric_list);
 937
 938	return ret;
 939}
 940
 941static int add_metric(struct list_head *metric_list,
 942		      const struct pmu_metric *pm,
 943		      const char *modifier,
 944		      bool metric_no_group,
 945		      bool metric_no_threshold,
 946		      const char *user_requested_cpu_list,
 947		      bool system_wide,
 948		      struct metric *root_metric,
 949		      const struct visited_metric *visited,
 950		      const struct pmu_metrics_table *table)
 951{
 952	int ret = 0;
 953
 954	pr_debug("metric expr %s for %s\n", pm->metric_expr, pm->metric_name);
 955
 956	if (!strstr(pm->metric_expr, "?")) {
 957		ret = __add_metric(metric_list, pm, modifier, metric_no_group,
 958				   metric_no_threshold, 0, user_requested_cpu_list,
 959				   system_wide, root_metric, visited, table);
 960	} else {
 961		int j, count;
 962
 963		count = arch_get_runtimeparam(pm);
 964
 965		/* This loop is added to create multiple
 966		 * events depend on count value and add
 967		 * those events to metric_list.
 968		 */
 969
 970		for (j = 0; j < count && !ret; j++)
 971			ret = __add_metric(metric_list, pm, modifier, metric_no_group,
 972					   metric_no_threshold, j, user_requested_cpu_list,
 973					   system_wide, root_metric, visited, table);
 974	}
 975
 976	return ret;
 977}
 978
 979/**
 980 * metric_list_cmp - list_sort comparator that sorts metrics with more events to
 981 *                   the front. tool events are excluded from the count.
 982 */
 983static int metric_list_cmp(void *priv __maybe_unused, const struct list_head *l,
 984			   const struct list_head *r)
 985{
 986	const struct metric *left = container_of(l, struct metric, nd);
 987	const struct metric *right = container_of(r, struct metric, nd);
 988	struct expr_id_data *data;
 989	int i, left_count, right_count;
 990
 991	left_count = hashmap__size(left->pctx->ids);
 992	tool_pmu__for_each_event(i) {
 993		if (!expr__get_id(left->pctx, tool_pmu__event_to_str(i), &data))
 994			left_count--;
 995	}
 996
 997	right_count = hashmap__size(right->pctx->ids);
 998	tool_pmu__for_each_event(i) {
 999		if (!expr__get_id(right->pctx, tool_pmu__event_to_str(i), &data))
1000			right_count--;
1001	}
1002
1003	return right_count - left_count;
1004}
1005
1006/**
1007 * default_metricgroup_cmp - Implements complex key for the Default metricgroup
1008 *			     that first sorts by default_metricgroup_name, then
1009 *			     metric_name.
1010 */
1011static int default_metricgroup_cmp(void *priv __maybe_unused,
1012				   const struct list_head *l,
1013				   const struct list_head *r)
1014{
1015	const struct metric *left = container_of(l, struct metric, nd);
1016	const struct metric *right = container_of(r, struct metric, nd);
1017	int diff = strcmp(right->default_metricgroup_name, left->default_metricgroup_name);
1018
1019	if (diff)
1020		return diff;
1021
1022	return strcmp(right->metric_name, left->metric_name);
1023}
1024
1025struct metricgroup__add_metric_data {
1026	struct list_head *list;
1027	const char *pmu;
1028	const char *metric_name;
1029	const char *modifier;
1030	const char *user_requested_cpu_list;
1031	bool metric_no_group;
1032	bool metric_no_threshold;
1033	bool system_wide;
1034	bool has_match;
1035};
1036
1037static int metricgroup__add_metric_callback(const struct pmu_metric *pm,
1038					    const struct pmu_metrics_table *table,
1039					    void *vdata)
1040{
1041	struct metricgroup__add_metric_data *data = vdata;
1042	int ret = 0;
1043
1044	if (pm->metric_expr && match_pm_metric_or_groups(pm, data->pmu, data->metric_name)) {
1045		bool metric_no_group = data->metric_no_group ||
1046			match_metric_or_groups(pm->metricgroup_no_group, data->metric_name);
1047
1048		data->has_match = true;
1049		ret = add_metric(data->list, pm, data->modifier, metric_no_group,
1050				 data->metric_no_threshold, data->user_requested_cpu_list,
1051				 data->system_wide, /*root_metric=*/NULL,
1052				 /*visited_metrics=*/NULL, table);
1053	}
1054	return ret;
1055}
1056
1057/**
1058 * metricgroup__add_metric - Find and add a metric, or a metric group.
1059 * @pmu: The PMU name to search for metrics on, or "all" for all PMUs.
1060 * @metric_name: The name of the metric or metric group. For example, "IPC"
1061 *               could be the name of a metric and "TopDownL1" the name of a
1062 *               metric group.
1063 * @modifier: if non-null event modifiers like "u".
1064 * @metric_no_group: Should events written to events be grouped "{}" or
1065 *                   global. Grouping is the default but due to multiplexing the
1066 *                   user may override.
1067 * @user_requested_cpu_list: Command line specified CPUs to record on.
1068 * @system_wide: Are events for all processes recorded.
1069 * @metric_list: The list that the metric or metric group are added to.
1070 * @table: The table that is searched for metrics, most commonly the table for the
1071 *       architecture perf is running upon.
1072 */
1073static int metricgroup__add_metric(const char *pmu, const char *metric_name, const char *modifier,
1074				   bool metric_no_group, bool metric_no_threshold,
1075				   const char *user_requested_cpu_list,
1076				   bool system_wide,
1077				   struct list_head *metric_list,
1078				   const struct pmu_metrics_table *table)
1079{
1080	LIST_HEAD(list);
1081	int ret;
1082	struct metricgroup__add_metric_data data = {
1083		.list = &list,
1084		.pmu = pmu,
1085		.metric_name = metric_name,
1086		.modifier = modifier,
1087		.metric_no_group = metric_no_group,
1088		.metric_no_threshold = metric_no_threshold,
1089		.user_requested_cpu_list = user_requested_cpu_list,
1090		.system_wide = system_wide,
1091		.has_match = false,
1092	};
1093
1094	/*
1095	 * Iterate over all metrics seeing if metric matches either the
1096	 * name or group. When it does add the metric to the list.
1097	 */
1098	ret = metricgroup__for_each_metric(table, metricgroup__add_metric_callback, &data);
1099	if (!ret && !data.has_match)
1100		ret = -EINVAL;
1101
1102	/*
1103	 * add to metric_list so that they can be released
1104	 * even if it's failed
1105	 */
1106	list_splice(&list, metric_list);
1107	return ret;
1108}
1109
1110/**
1111 * metricgroup__add_metric_list - Find and add metrics, or metric groups,
1112 *                                specified in a list.
1113 * @pmu: A pmu to restrict the metrics to, or "all" for all PMUS.
1114 * @list: the list of metrics or metric groups. For example, "IPC,CPI,TopDownL1"
1115 *        would match the IPC and CPI metrics, and TopDownL1 would match all
1116 *        the metrics in the TopDownL1 group.
1117 * @metric_no_group: Should events written to events be grouped "{}" or
1118 *                   global. Grouping is the default but due to multiplexing the
1119 *                   user may override.
1120 * @user_requested_cpu_list: Command line specified CPUs to record on.
1121 * @system_wide: Are events for all processes recorded.
1122 * @metric_list: The list that metrics are added to.
1123 * @table: The table that is searched for metrics, most commonly the table for the
1124 *       architecture perf is running upon.
1125 */
1126static int metricgroup__add_metric_list(const char *pmu, const char *list,
1127					bool metric_no_group,
1128					bool metric_no_threshold,
1129					const char *user_requested_cpu_list,
1130					bool system_wide, struct list_head *metric_list,
1131					const struct pmu_metrics_table *table)
1132{
1133	char *list_itr, *list_copy, *metric_name, *modifier;
1134	int ret, count = 0;
1135
1136	list_copy = strdup(list);
1137	if (!list_copy)
1138		return -ENOMEM;
1139	list_itr = list_copy;
1140
1141	while ((metric_name = strsep(&list_itr, ",")) != NULL) {
1142		modifier = strchr(metric_name, ':');
1143		if (modifier)
1144			*modifier++ = '\0';
1145
1146		ret = metricgroup__add_metric(pmu, metric_name, modifier,
1147					      metric_no_group, metric_no_threshold,
1148					      user_requested_cpu_list,
1149					      system_wide, metric_list, table);
1150		if (ret == -EINVAL)
1151			pr_err("Cannot find metric or group `%s'\n", metric_name);
1152
1153		if (ret)
1154			break;
1155
1156		count++;
1157	}
1158	free(list_copy);
1159
1160	if (!ret) {
1161		/*
1162		 * Warn about nmi_watchdog if any parsed metrics had the
1163		 * NO_NMI_WATCHDOG constraint.
1164		 */
1165		metric__watchdog_constraint_hint(NULL, /*foot=*/true);
1166		/* No metrics. */
1167		if (count == 0)
1168			return -EINVAL;
1169	}
1170	return ret;
1171}
1172
1173static void metricgroup__free_metrics(struct list_head *metric_list)
1174{
1175	struct metric *m, *tmp;
1176
1177	list_for_each_entry_safe (m, tmp, metric_list, nd) {
1178		list_del_init(&m->nd);
1179		metric__free(m);
1180	}
1181}
1182
1183/**
1184 * find_tool_events - Search for the pressence of tool events in metric_list.
1185 * @metric_list: List to take metrics from.
1186 * @tool_events: Array of false values, indices corresponding to tool events set
1187 *               to true if tool event is found.
1188 */
1189static void find_tool_events(const struct list_head *metric_list,
1190			     bool tool_events[TOOL_PMU__EVENT_MAX])
1191{
1192	struct metric *m;
1193
1194	list_for_each_entry(m, metric_list, nd) {
1195		int i;
1196
1197		tool_pmu__for_each_event(i) {
1198			struct expr_id_data *data;
1199
1200			if (!tool_events[i] &&
1201			    !expr__get_id(m->pctx, tool_pmu__event_to_str(i), &data))
1202				tool_events[i] = true;
1203		}
1204	}
1205}
1206
1207/**
1208 * build_combined_expr_ctx - Make an expr_parse_ctx with all !group_events
1209 *                           metric IDs, as the IDs are held in a set,
1210 *                           duplicates will be removed.
1211 * @metric_list: List to take metrics from.
1212 * @combined: Out argument for result.
1213 */
1214static int build_combined_expr_ctx(const struct list_head *metric_list,
1215				   struct expr_parse_ctx **combined)
1216{
1217	struct hashmap_entry *cur;
1218	size_t bkt;
1219	struct metric *m;
1220	char *dup;
1221	int ret;
1222
1223	*combined = expr__ctx_new();
1224	if (!*combined)
1225		return -ENOMEM;
1226
1227	list_for_each_entry(m, metric_list, nd) {
1228		if (!m->group_events && !m->modifier) {
1229			hashmap__for_each_entry(m->pctx->ids, cur, bkt) {
1230				dup = strdup(cur->pkey);
1231				if (!dup) {
1232					ret = -ENOMEM;
1233					goto err_out;
1234				}
1235				ret = expr__add_id(*combined, dup);
1236				if (ret)
1237					goto err_out;
1238			}
1239		}
1240	}
1241	return 0;
1242err_out:
1243	expr__ctx_free(*combined);
1244	*combined = NULL;
1245	return ret;
1246}
1247
1248/**
1249 * parse_ids - Build the event string for the ids and parse them creating an
1250 *             evlist. The encoded metric_ids are decoded.
1251 * @metric_no_merge: is metric sharing explicitly disabled.
1252 * @fake_pmu: use a fake PMU when testing metrics not supported by the current CPU.
1253 * @ids: the event identifiers parsed from a metric.
1254 * @modifier: any modifiers added to the events.
1255 * @group_events: should events be placed in a weak group.
1256 * @tool_events: entries set true if the tool event of index could be present in
1257 *               the overall list of metrics.
1258 * @out_evlist: the created list of events.
1259 */
1260static int parse_ids(bool metric_no_merge, bool fake_pmu,
1261		     struct expr_parse_ctx *ids, const char *modifier,
1262		     bool group_events, const bool tool_events[TOOL_PMU__EVENT_MAX],
1263		     struct evlist **out_evlist)
1264{
1265	struct parse_events_error parse_error;
1266	struct evlist *parsed_evlist;
1267	struct strbuf events = STRBUF_INIT;
1268	int ret;
1269
1270	*out_evlist = NULL;
1271	if (!metric_no_merge || hashmap__size(ids->ids) == 0) {
1272		bool added_event = false;
1273		int i;
1274		/*
1275		 * We may fail to share events between metrics because a tool
1276		 * event isn't present in one metric. For example, a ratio of
1277		 * cache misses doesn't need duration_time but the same events
1278		 * may be used for a misses per second. Events without sharing
1279		 * implies multiplexing, that is best avoided, so place
1280		 * all tool events in every group.
1281		 *
1282		 * Also, there may be no ids/events in the expression parsing
1283		 * context because of constant evaluation, e.g.:
1284		 *    event1 if #smt_on else 0
1285		 * Add a tool event to avoid a parse error on an empty string.
1286		 */
1287		tool_pmu__for_each_event(i) {
1288			if (tool_events[i]) {
1289				char *tmp = strdup(tool_pmu__event_to_str(i));
1290
1291				if (!tmp)
1292					return -ENOMEM;
1293				ids__insert(ids->ids, tmp);
1294				added_event = true;
1295			}
1296		}
1297		if (!added_event && hashmap__size(ids->ids) == 0) {
1298			char *tmp = strdup("duration_time");
1299
1300			if (!tmp)
1301				return -ENOMEM;
1302			ids__insert(ids->ids, tmp);
1303		}
1304	}
1305	ret = metricgroup__build_event_string(&events, ids, modifier,
1306					      group_events);
1307	if (ret)
1308		return ret;
1309
1310	parsed_evlist = evlist__new();
1311	if (!parsed_evlist) {
1312		ret = -ENOMEM;
1313		goto err_out;
1314	}
1315	pr_debug("Parsing metric events '%s'\n", events.buf);
1316	parse_events_error__init(&parse_error);
1317	ret = __parse_events(parsed_evlist, events.buf, /*pmu_filter=*/NULL,
1318			     &parse_error, fake_pmu, /*warn_if_reordered=*/false,
1319			     /*fake_tp=*/false);
1320	if (ret) {
1321		parse_events_error__print(&parse_error, events.buf);
1322		goto err_out;
1323	}
1324	ret = decode_all_metric_ids(parsed_evlist, modifier);
1325	if (ret)
1326		goto err_out;
1327
1328	*out_evlist = parsed_evlist;
1329	parsed_evlist = NULL;
1330err_out:
1331	parse_events_error__exit(&parse_error);
1332	evlist__delete(parsed_evlist);
1333	strbuf_release(&events);
1334	return ret;
1335}
1336
1337/* How many times will a given evsel be used in a set of metrics? */
1338static int count_uses(struct list_head *metric_list, struct evsel *evsel)
1339{
1340	const char *metric_id = evsel__metric_id(evsel);
1341	struct metric *m;
1342	int uses = 0;
1343
1344	list_for_each_entry(m, metric_list, nd) {
1345		if (hashmap__find(m->pctx->ids, metric_id, NULL))
1346			uses++;
1347	}
1348	return uses;
1349}
1350
1351/*
1352 * Select the evsel that stat-display will use to trigger shadow/metric
1353 * printing. Pick the least shared non-tool evsel, encouraging metrics to be
1354 * with a hardware counter that is specific to them.
1355 */
1356static struct evsel *pick_display_evsel(struct list_head *metric_list,
1357					struct evsel **metric_events)
1358{
1359	struct evsel *selected = metric_events[0];
1360	size_t selected_uses;
1361	bool selected_is_tool;
1362
1363	if (!selected)
1364		return NULL;
1365
1366	selected_uses = count_uses(metric_list, selected);
1367	selected_is_tool = evsel__is_tool(selected);
1368	for (int i = 1; metric_events[i]; i++) {
1369		struct evsel *candidate = metric_events[i];
1370		size_t candidate_uses = count_uses(metric_list, candidate);
1371
1372		if ((selected_is_tool && !evsel__is_tool(candidate)) ||
1373		    (candidate_uses < selected_uses)) {
1374			selected = candidate;
1375			selected_uses = candidate_uses;
1376			selected_is_tool = evsel__is_tool(selected);
1377		}
1378	}
1379	return selected;
1380}
1381
1382static int parse_groups(struct evlist *perf_evlist,
1383			const char *pmu, const char *str,
1384			bool metric_no_group,
1385			bool metric_no_merge,
1386			bool metric_no_threshold,
1387			const char *user_requested_cpu_list,
1388			bool system_wide,
1389			bool fake_pmu,
1390			const struct pmu_metrics_table *table)
1391{
1392	struct evlist *combined_evlist = NULL;
1393	LIST_HEAD(metric_list);
1394	struct metric *m;
1395	bool tool_events[TOOL_PMU__EVENT_MAX] = {false};
1396	bool is_default = !strcmp(str, "Default");
1397	int ret;
1398
1399	ret = metricgroup__add_metric_list(pmu, str, metric_no_group, metric_no_threshold,
1400					   user_requested_cpu_list,
1401					   system_wide, &metric_list, table);
1402	if (ret)
1403		goto out;
1404
1405	/* Sort metrics from largest to smallest. */
1406	list_sort(NULL, &metric_list, metric_list_cmp);
1407
1408	if (!metric_no_merge) {
1409		struct expr_parse_ctx *combined = NULL;
1410
1411		find_tool_events(&metric_list, tool_events);
1412
1413		ret = build_combined_expr_ctx(&metric_list, &combined);
1414
1415		if (!ret && combined && hashmap__size(combined->ids)) {
1416			ret = parse_ids(metric_no_merge, fake_pmu, combined,
1417					/*modifier=*/NULL,
1418					/*group_events=*/false,
1419					tool_events,
1420					&combined_evlist);
1421		}
1422		if (combined)
1423			expr__ctx_free(combined);
1424
1425		if (ret)
1426			goto out;
1427	}
1428
1429	if (is_default)
1430		list_sort(NULL, &metric_list, default_metricgroup_cmp);
1431
1432	list_for_each_entry(m, &metric_list, nd) {
1433		struct metric_event *me;
1434		struct evsel **metric_events;
1435		struct evlist *metric_evlist = NULL;
1436		struct metric *n;
1437		struct metric_expr *expr;
1438
1439		if (combined_evlist && !m->group_events) {
1440			metric_evlist = combined_evlist;
1441		} else if (!metric_no_merge) {
1442			/*
1443			 * See if the IDs for this metric are a subset of an
1444			 * earlier metric.
1445			 */
1446			list_for_each_entry(n, &metric_list, nd) {
1447				if (m == n)
1448					break;
1449
1450				if (n->evlist == NULL)
1451					continue;
1452
1453				if ((!m->modifier && n->modifier) ||
1454				    (m->modifier && !n->modifier) ||
1455				    (m->modifier && n->modifier &&
1456					    strcmp(m->modifier, n->modifier)))
1457					continue;
1458
1459				if ((!m->pmu && n->pmu) ||
1460				    (m->pmu && !n->pmu) ||
1461				    (m->pmu && n->pmu && strcmp(m->pmu, n->pmu)))
1462					continue;
1463
1464				if (expr__subset_of_ids(n->pctx, m->pctx)) {
1465					pr_debug("Events in '%s' fully contained within '%s'\n",
1466						 m->metric_name, n->metric_name);
1467					metric_evlist = n->evlist;
1468					break;
1469				}
1470
1471			}
1472		}
1473		if (!metric_evlist) {
1474			ret = parse_ids(metric_no_merge, fake_pmu, m->pctx, m->modifier,
1475					m->group_events, tool_events, &m->evlist);
1476			if (ret)
1477				goto out;
1478
1479			metric_evlist = m->evlist;
1480		}
1481		ret = setup_metric_events(fake_pmu ? "all" : m->pmu, m->pctx->ids,
1482					  metric_evlist, &metric_events);
1483		if (ret) {
1484			pr_err("Cannot resolve IDs for %s: %s\n",
1485				m->metric_name, m->metric_expr);
1486			goto out;
1487		}
1488
1489		me = metricgroup__lookup(&perf_evlist->metric_events,
1490					 pick_display_evsel(&metric_list, metric_events),
1491					 /*create=*/true);
1492
1493		expr = malloc(sizeof(struct metric_expr));
1494		if (!expr) {
1495			ret = -ENOMEM;
1496			free(metric_events);
1497			goto out;
1498		}
1499
1500		expr->metric_refs = m->metric_refs;
1501		m->metric_refs = NULL;
1502		expr->metric_expr = m->metric_expr;
1503		if (m->modifier) {
1504			char *tmp;
1505
1506			if (asprintf(&tmp, "%s:%s", m->metric_name, m->modifier) < 0)
1507				expr->metric_name = NULL;
1508			else
1509				expr->metric_name = tmp;
1510		} else
1511			expr->metric_name = strdup(m->metric_name);
1512
1513		if (!expr->metric_name) {
1514			ret = -ENOMEM;
1515			free(expr);
1516			free(metric_events);
1517			goto out;
1518		}
1519		if (m->default_show_events) {
1520			struct evsel *pos;
1521
1522			for (int i = 0; metric_events[i]; i++)
1523				metric_events[i]->default_show_events = true;
1524			evlist__for_each_entry(metric_evlist, pos) {
1525				if (pos->metric_leader && pos->metric_leader->default_show_events)
1526					pos->default_show_events = true;
1527			}
1528		}
1529		expr->metric_threshold = m->metric_threshold;
1530		expr->metric_unit = m->metric_unit;
1531		expr->metric_events = metric_events;
1532		expr->runtime = m->pctx->sctx.runtime;
1533		expr->default_metricgroup_name = m->default_metricgroup_name;
1534		me->is_default = is_default;
1535		list_add(&expr->nd, &me->head);
1536	}
1537
1538
1539	if (combined_evlist) {
1540		evlist__splice_list_tail(perf_evlist, &combined_evlist->core.entries);
1541		evlist__delete(combined_evlist);
1542	}
1543
1544	list_for_each_entry(m, &metric_list, nd) {
1545		if (m->evlist)
1546			evlist__splice_list_tail(perf_evlist, &m->evlist->core.entries);
1547	}
1548
1549out:
1550	metricgroup__free_metrics(&metric_list);
1551	return ret;
1552}
1553
1554int metricgroup__parse_groups(struct evlist *perf_evlist,
1555			      const char *pmu,
1556			      const char *str,
1557			      bool metric_no_group,
1558			      bool metric_no_merge,
1559			      bool metric_no_threshold,
1560			      const char *user_requested_cpu_list,
1561			      bool system_wide,
1562			      bool hardware_aware_grouping)
1563{
1564	const struct pmu_metrics_table *table = pmu_metrics_table__find();
1565
1566	if (hardware_aware_grouping)
1567		pr_debug("Use hardware aware grouping instead of traditional metric grouping method\n");
1568
1569	return parse_groups(perf_evlist, pmu, str, metric_no_group, metric_no_merge,
1570			    metric_no_threshold, user_requested_cpu_list, system_wide,
1571			    /*fake_pmu=*/false, table);
1572}
1573
1574int metricgroup__parse_groups_test(struct evlist *evlist,
1575				   const struct pmu_metrics_table *table,
1576				   const char *str)
1577{
1578	return parse_groups(evlist, "all", str,
1579			    /*metric_no_group=*/false,
1580			    /*metric_no_merge=*/false,
1581			    /*metric_no_threshold=*/false,
1582			    /*user_requested_cpu_list=*/NULL,
1583			    /*system_wide=*/false,
1584			    /*fake_pmu=*/true, table);
1585}
1586
1587struct metricgroup__has_metric_data {
1588	const char *pmu;
1589	const char *metric_or_groups;
1590};
1591static int metricgroup__has_metric_or_groups_callback(const struct pmu_metric *pm,
1592						      const struct pmu_metrics_table *table
1593							__maybe_unused,
1594						      void *vdata)
1595{
1596	struct metricgroup__has_metric_data *data = vdata;
1597
1598	return match_pm_metric_or_groups(pm, data->pmu, data->metric_or_groups) ? 1 : 0;
1599}
1600
1601bool metricgroup__has_metric_or_groups(const char *pmu, const char *metric_or_groups)
1602{
1603	const struct pmu_metrics_table *table = pmu_metrics_table__find();
1604	struct metricgroup__has_metric_data data = {
1605		.pmu = pmu,
1606		.metric_or_groups = metric_or_groups,
1607	};
1608
1609	return metricgroup__for_each_metric(table,
1610					    metricgroup__has_metric_or_groups_callback,
1611					    &data)
1612		? true : false;
1613}
1614
1615static int metricgroup__topdown_max_level_callback(const struct pmu_metric *pm,
1616					    const struct pmu_metrics_table *table __maybe_unused,
1617					    void *data)
1618{
1619	unsigned int *max_level = data;
1620	unsigned int level;
1621	const char *p = strstr(pm->metric_group ?: "", "TopdownL");
1622
1623	if (!p || p[8] == '\0')
1624		return 0;
1625
1626	level = p[8] - '0';
1627	if (level > *max_level)
1628		*max_level = level;
1629
1630	return 0;
1631}
1632
1633unsigned int metricgroups__topdown_max_level(void)
1634{
1635	unsigned int max_level = 0;
1636	const struct pmu_metrics_table *table = pmu_metrics_table__find();
1637
1638	if (!table)
1639		return false;
1640
1641	pmu_metrics_table__for_each_metric(table, metricgroup__topdown_max_level_callback,
1642					  &max_level);
1643	return max_level;
1644}
1645
1646int metricgroup__copy_metric_events(struct evlist *evlist, struct cgroup *cgrp,
1647				    struct rblist *new_metric_events,
1648				    struct rblist *old_metric_events)
1649{
1650	unsigned int i;
1651
1652	for (i = 0; i < rblist__nr_entries(old_metric_events); i++) {
1653		struct rb_node *nd;
1654		struct metric_event *old_me, *new_me;
1655		struct metric_expr *old_expr, *new_expr;
1656		struct evsel *evsel;
1657		size_t alloc_size;
1658		int idx, nr;
1659
1660		nd = rblist__entry(old_metric_events, i);
1661		old_me = container_of(nd, struct metric_event, nd);
1662
1663		evsel = evlist__find_evsel(evlist, old_me->evsel->core.idx);
1664		if (!evsel)
1665			return -EINVAL;
1666		new_me = metricgroup__lookup(new_metric_events, evsel, /*create=*/true);
1667		if (!new_me)
1668			return -ENOMEM;
1669
1670		pr_debug("copying metric event for cgroup '%s': %s (idx=%d)\n",
1671			 cgrp ? cgrp->name : "root", evsel->name, evsel->core.idx);
1672
1673		new_me->is_default = old_me->is_default;
1674		list_for_each_entry(old_expr, &old_me->head, nd) {
1675			new_expr = malloc(sizeof(*new_expr));
1676			if (!new_expr)
1677				return -ENOMEM;
1678
1679			new_expr->metric_expr = old_expr->metric_expr;
1680			new_expr->metric_threshold = old_expr->metric_threshold;
1681			new_expr->metric_name = strdup(old_expr->metric_name);
1682			if (!new_expr->metric_name)
1683				return -ENOMEM;
1684
1685			new_expr->metric_unit = old_expr->metric_unit;
1686			new_expr->runtime = old_expr->runtime;
1687			new_expr->default_metricgroup_name = old_expr->default_metricgroup_name;
1688
1689			if (old_expr->metric_refs) {
1690				/* calculate number of metric_events */
1691				for (nr = 0; old_expr->metric_refs[nr].metric_name; nr++)
1692					continue;
1693				alloc_size = sizeof(*new_expr->metric_refs);
1694				new_expr->metric_refs = calloc(nr + 1, alloc_size);
1695				if (!new_expr->metric_refs) {
1696					free(new_expr);
1697					return -ENOMEM;
1698				}
1699
1700				memcpy(new_expr->metric_refs, old_expr->metric_refs,
1701				       nr * alloc_size);
1702			} else {
1703				new_expr->metric_refs = NULL;
1704			}
1705
1706			/* calculate number of metric_events */
1707			for (nr = 0; old_expr->metric_events[nr]; nr++)
1708				continue;
1709			alloc_size = sizeof(*new_expr->metric_events);
1710			new_expr->metric_events = calloc(nr + 1, alloc_size);
1711			if (!new_expr->metric_events) {
1712				zfree(&new_expr->metric_refs);
1713				free(new_expr);
1714				return -ENOMEM;
1715			}
1716
1717			/* copy evsel in the same position */
1718			for (idx = 0; idx < nr; idx++) {
1719				evsel = old_expr->metric_events[idx];
1720				evsel = evlist__find_evsel(evlist, evsel->core.idx);
1721				if (evsel == NULL) {
1722					zfree(&new_expr->metric_events);
1723					zfree(&new_expr->metric_refs);
1724					free(new_expr);
1725					return -EINVAL;
1726				}
1727				new_expr->metric_events[idx] = evsel;
1728			}
1729
1730			list_add(&new_expr->nd, &new_me->head);
1731		}
1732	}
1733	return 0;
1734}