evlist.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4 *
   5 * Parts came from builtin-{top,stat,record}.c, see those files for further
   6 * copyright notes.
   7 */
   8#include <api/fs/fs.h>
   9#include <errno.h>
  10#include <inttypes.h>
  11#include <poll.h>
  12#include "cpumap.h"
  13#include "util/mmap.h"
  14#include "thread_map.h"
  15#include "target.h"
  16#include "evlist.h"
  17#include "evsel.h"
  18#include "record.h"
  19#include "debug.h"
  20#include "units.h"
  21#include "bpf_counter.h"
  22#include <internal/lib.h> // page_size
  23#include "affinity.h"
  24#include "../perf.h"
  25#include "asm/bug.h"
  26#include "bpf-event.h"
  27#include "util/event.h"
  28#include "util/string2.h"
  29#include "util/perf_api_probe.h"
  30#include "util/evsel_fprintf.h"
  31#include "util/pmu.h"
  32#include "util/sample.h"
  33#include "util/bpf-filter.h"
  34#include "util/stat.h"
  35#include "util/util.h"
  36#include "util/env.h"
  37#include "util/intel-tpebs.h"
  38#include "util/metricgroup.h"
  39#include "util/strbuf.h"
  40#include <signal.h>
  41#include <unistd.h>
  42#include <sched.h>
  43#include <stdlib.h>
  44
  45#include "parse-events.h"
  46#include <subcmd/parse-options.h>
  47
  48#include <fcntl.h>
  49#include <sys/ioctl.h>
  50#include <sys/mman.h>
  51#include <sys/prctl.h>
  52#include <sys/timerfd.h>
  53#include <sys/wait.h>
  54
  55#include <linux/bitops.h>
  56#include <linux/hash.h>
  57#include <linux/log2.h>
  58#include <linux/err.h>
  59#include <linux/string.h>
  60#include <linux/time64.h>
  61#include <linux/zalloc.h>
  62#include <perf/evlist.h>
  63#include <perf/evsel.h>
  64#include <perf/cpumap.h>
  65#include <perf/mmap.h>
  66
  67#include <internal/xyarray.h>
  68
  69#ifdef LACKS_SIGQUEUE_PROTOTYPE
  70int sigqueue(pid_t pid, int sig, const union sigval value);
  71#endif
  72
  73#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  74#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  75
  76void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  77		  struct perf_thread_map *threads)
  78{
  79	perf_evlist__init(&evlist->core);
  80	perf_evlist__set_maps(&evlist->core, cpus, threads);
  81	evlist->workload.pid = -1;
  82	evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  83	evlist->ctl_fd.fd = -1;
  84	evlist->ctl_fd.ack = -1;
  85	evlist->ctl_fd.pos = -1;
  86	evlist->nr_br_cntr = -1;
  87	metricgroup__rblist_init(&evlist->metric_events);
  88	INIT_LIST_HEAD(&evlist->deferred_samples);
  89}
  90
  91struct evlist *evlist__new(void)
  92{
  93	struct evlist *evlist = zalloc(sizeof(*evlist));
  94
  95	if (evlist != NULL)
  96		evlist__init(evlist, NULL, NULL);
  97
  98	return evlist;
  99}
 100
 101struct evlist *evlist__new_default(void)
 102{
 103	struct evlist *evlist = evlist__new();
 104	bool can_profile_kernel;
 105	struct perf_pmu *pmu = NULL;
 106
 107	if (!evlist)
 108		return NULL;
 109
 110	can_profile_kernel = perf_event_paranoid_check(1);
 111
 112	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 113		char buf[256];
 114		int err;
 115
 116		snprintf(buf, sizeof(buf), "%s/cycles/%s", pmu->name,
 117			 can_profile_kernel ? "P" : "Pu");
 118		err = parse_event(evlist, buf);
 119		if (err) {
 120			evlist__delete(evlist);
 121			return NULL;
 122		}
 123	}
 124
 125	if (evlist->core.nr_entries > 1) {
 126		struct evsel *evsel;
 127
 128		evlist__for_each_entry(evlist, evsel)
 129			evsel__set_sample_id(evsel, /*can_sample_identifier=*/false);
 130	}
 131
 132	return evlist;
 133}
 134
 135struct evlist *evlist__new_dummy(void)
 136{
 137	struct evlist *evlist = evlist__new();
 138
 139	if (evlist && evlist__add_dummy(evlist)) {
 140		evlist__delete(evlist);
 141		evlist = NULL;
 142	}
 143
 144	return evlist;
 145}
 146
 147/**
 148 * evlist__set_id_pos - set the positions of event ids.
 149 * @evlist: selected event list
 150 *
 151 * Events with compatible sample types all have the same id_pos
 152 * and is_pos.  For convenience, put a copy on evlist.
 153 */
 154void evlist__set_id_pos(struct evlist *evlist)
 155{
 156	struct evsel *first = evlist__first(evlist);
 157
 158	evlist->id_pos = first->id_pos;
 159	evlist->is_pos = first->is_pos;
 160}
 161
 162static void evlist__update_id_pos(struct evlist *evlist)
 163{
 164	struct evsel *evsel;
 165
 166	evlist__for_each_entry(evlist, evsel)
 167		evsel__calc_id_pos(evsel);
 168
 169	evlist__set_id_pos(evlist);
 170}
 171
 172static void evlist__purge(struct evlist *evlist)
 173{
 174	struct evsel *pos, *n;
 175
 176	evlist__for_each_entry_safe(evlist, n, pos) {
 177		list_del_init(&pos->core.node);
 178		pos->evlist = NULL;
 179		evsel__delete(pos);
 180	}
 181
 182	evlist->core.nr_entries = 0;
 183}
 184
 185void evlist__exit(struct evlist *evlist)
 186{
 187	metricgroup__rblist_exit(&evlist->metric_events);
 188	event_enable_timer__exit(&evlist->eet);
 189	zfree(&evlist->mmap);
 190	zfree(&evlist->overwrite_mmap);
 191	perf_evlist__exit(&evlist->core);
 192}
 193
 194void evlist__delete(struct evlist *evlist)
 195{
 196	if (evlist == NULL)
 197		return;
 198
 199	evlist__free_stats(evlist);
 200	evlist__munmap(evlist);
 201	evlist__close(evlist);
 202	evlist__purge(evlist);
 203	evlist__exit(evlist);
 204	free(evlist);
 205}
 206
 207void evlist__add(struct evlist *evlist, struct evsel *entry)
 208{
 209	perf_evlist__add(&evlist->core, &entry->core);
 210	entry->evlist = evlist;
 211	entry->tracking = !entry->core.idx;
 212
 213	if (evlist->core.nr_entries == 1)
 214		evlist__set_id_pos(evlist);
 215}
 216
 217void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 218{
 219	evsel->evlist = NULL;
 220	perf_evlist__remove(&evlist->core, &evsel->core);
 221}
 222
 223void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
 224{
 225	while (!list_empty(list)) {
 226		struct evsel *evsel, *temp, *leader = NULL;
 227
 228		__evlist__for_each_entry_safe(list, temp, evsel) {
 229			list_del_init(&evsel->core.node);
 230			evlist__add(evlist, evsel);
 231			leader = evsel;
 232			break;
 233		}
 234
 235		__evlist__for_each_entry_safe(list, temp, evsel) {
 236			if (evsel__has_leader(evsel, leader)) {
 237				list_del_init(&evsel->core.node);
 238				evlist__add(evlist, evsel);
 239			}
 240		}
 241	}
 242}
 243
 244int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 245				       const struct evsel_str_handler *assocs, size_t nr_assocs)
 246{
 247	size_t i;
 248	int err;
 249
 250	for (i = 0; i < nr_assocs; i++) {
 251		// Adding a handler for an event not in this evlist, just ignore it.
 252		struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 253		if (evsel == NULL)
 254			continue;
 255
 256		err = -EEXIST;
 257		if (evsel->handler != NULL)
 258			goto out;
 259		evsel->handler = assocs[i].handler;
 260	}
 261
 262	err = 0;
 263out:
 264	return err;
 265}
 266
 267static void evlist__set_leader(struct evlist *evlist)
 268{
 269	perf_evlist__set_leader(&evlist->core);
 270}
 271
 272static struct evsel *evlist__dummy_event(struct evlist *evlist)
 273{
 274	struct perf_event_attr attr = {
 275		.type	= PERF_TYPE_SOFTWARE,
 276		.config = PERF_COUNT_SW_DUMMY,
 277		.size	= sizeof(attr), /* to capture ABI version */
 278		/* Avoid frequency mode for dummy events to avoid associated timers. */
 279		.freq = 0,
 280		.sample_period = 1,
 281	};
 282
 283	return evsel__new_idx(&attr, evlist->core.nr_entries);
 284}
 285
 286int evlist__add_dummy(struct evlist *evlist)
 287{
 288	struct evsel *evsel = evlist__dummy_event(evlist);
 289
 290	if (evsel == NULL)
 291		return -ENOMEM;
 292
 293	evlist__add(evlist, evsel);
 294	return 0;
 295}
 296
 297struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
 298{
 299	struct evsel *evsel = evlist__dummy_event(evlist);
 300
 301	if (!evsel)
 302		return NULL;
 303
 304	evsel->core.attr.exclude_kernel = 1;
 305	evsel->core.attr.exclude_guest = 1;
 306	evsel->core.attr.exclude_hv = 1;
 307	evsel->core.system_wide = system_wide;
 308	evsel->no_aux_samples = true;
 309	evsel->name = strdup("dummy:u");
 310
 311	evlist__add(evlist, evsel);
 312	return evsel;
 313}
 314
 315#ifdef HAVE_LIBTRACEEVENT
 316struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
 317{
 318	struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0,
 319					       /*format=*/true);
 320
 321	if (IS_ERR(evsel))
 322		return evsel;
 323
 324	evsel__set_sample_bit(evsel, CPU);
 325	evsel__set_sample_bit(evsel, TIME);
 326
 327	evsel->core.system_wide = system_wide;
 328	evsel->no_aux_samples = true;
 329
 330	evlist__add(evlist, evsel);
 331	return evsel;
 332}
 333#endif
 334
 335struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
 336{
 337	struct evsel *evsel;
 338
 339	evlist__for_each_entry(evlist, evsel) {
 340		if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 341		    (strcmp(evsel->name, name) == 0))
 342			return evsel;
 343	}
 344
 345	return NULL;
 346}
 347
 348#ifdef HAVE_LIBTRACEEVENT
 349int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
 350{
 351	struct evsel *evsel = evsel__newtp(sys, name);
 352
 353	if (IS_ERR(evsel))
 354		return -1;
 355
 356	evsel->handler = handler;
 357	evlist__add(evlist, evsel);
 358	return 0;
 359}
 360#endif
 361
 362struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
 363{
 364	struct evlist_cpu_iterator itr = {
 365		.container = evlist,
 366		.evsel = NULL,
 367		.cpu_map_idx = 0,
 368		.evlist_cpu_map_idx = 0,
 369		.evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
 370		.cpu = (struct perf_cpu){ .cpu = -1},
 371		.affinity = affinity,
 372	};
 373
 374	if (evlist__empty(evlist)) {
 375		/* Ensure the empty list doesn't iterate. */
 376		itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
 377	} else {
 378		itr.evsel = evlist__first(evlist);
 379		if (itr.affinity) {
 380			itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
 381			affinity__set(itr.affinity, itr.cpu.cpu);
 382			itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
 383			/*
 384			 * If this CPU isn't in the evsel's cpu map then advance
 385			 * through the list.
 386			 */
 387			if (itr.cpu_map_idx == -1)
 388				evlist_cpu_iterator__next(&itr);
 389		}
 390	}
 391	return itr;
 392}
 393
 394void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
 395{
 396	while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
 397		evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
 398		evlist_cpu_itr->cpu_map_idx =
 399			perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 400					  evlist_cpu_itr->cpu);
 401		if (evlist_cpu_itr->cpu_map_idx != -1)
 402			return;
 403	}
 404	evlist_cpu_itr->evlist_cpu_map_idx++;
 405	if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
 406		evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
 407		evlist_cpu_itr->cpu =
 408			perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
 409					  evlist_cpu_itr->evlist_cpu_map_idx);
 410		if (evlist_cpu_itr->affinity)
 411			affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
 412		evlist_cpu_itr->cpu_map_idx =
 413			perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 414					  evlist_cpu_itr->cpu);
 415		/*
 416		 * If this CPU isn't in the evsel's cpu map then advance through
 417		 * the list.
 418		 */
 419		if (evlist_cpu_itr->cpu_map_idx == -1)
 420			evlist_cpu_iterator__next(evlist_cpu_itr);
 421	}
 422}
 423
 424bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
 425{
 426	return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
 427}
 428
 429static int evsel__strcmp(struct evsel *pos, char *evsel_name)
 430{
 431	if (!evsel_name)
 432		return 0;
 433	if (evsel__is_dummy_event(pos))
 434		return 1;
 435	return !evsel__name_is(pos, evsel_name);
 436}
 437
 438static int evlist__is_enabled(struct evlist *evlist)
 439{
 440	struct evsel *pos;
 441
 442	evlist__for_each_entry(evlist, pos) {
 443		if (!evsel__is_group_leader(pos) || !pos->core.fd)
 444			continue;
 445		/* If at least one event is enabled, evlist is enabled. */
 446		if (!pos->disabled)
 447			return true;
 448	}
 449	return false;
 450}
 451
 452static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 453{
 454	struct evsel *pos;
 455	struct evlist_cpu_iterator evlist_cpu_itr;
 456	struct affinity saved_affinity, *affinity = NULL;
 457	bool has_imm = false;
 458
 459	// See explanation in evlist__close()
 460	if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 461		if (affinity__setup(&saved_affinity) < 0)
 462			return;
 463		affinity = &saved_affinity;
 464	}
 465
 466	/* Disable 'immediate' events last */
 467	for (int imm = 0; imm <= 1; imm++) {
 468		evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 469			pos = evlist_cpu_itr.evsel;
 470			if (evsel__strcmp(pos, evsel_name))
 471				continue;
 472			if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
 473				continue;
 474			if (excl_dummy && evsel__is_dummy_event(pos))
 475				continue;
 476			if (pos->immediate)
 477				has_imm = true;
 478			if (pos->immediate != imm)
 479				continue;
 480			evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 481		}
 482		if (!has_imm)
 483			break;
 484	}
 485
 486	affinity__cleanup(affinity);
 487	evlist__for_each_entry(evlist, pos) {
 488		if (evsel__strcmp(pos, evsel_name))
 489			continue;
 490		if (!evsel__is_group_leader(pos) || !pos->core.fd)
 491			continue;
 492		if (excl_dummy && evsel__is_dummy_event(pos))
 493			continue;
 494		pos->disabled = true;
 495	}
 496
 497	/*
 498	 * If we disabled only single event, we need to check
 499	 * the enabled state of the evlist manually.
 500	 */
 501	if (evsel_name)
 502		evlist->enabled = evlist__is_enabled(evlist);
 503	else
 504		evlist->enabled = false;
 505}
 506
 507void evlist__disable(struct evlist *evlist)
 508{
 509	__evlist__disable(evlist, NULL, false);
 510}
 511
 512void evlist__disable_non_dummy(struct evlist *evlist)
 513{
 514	__evlist__disable(evlist, NULL, true);
 515}
 516
 517void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
 518{
 519	__evlist__disable(evlist, evsel_name, false);
 520}
 521
 522static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 523{
 524	struct evsel *pos;
 525	struct evlist_cpu_iterator evlist_cpu_itr;
 526	struct affinity saved_affinity, *affinity = NULL;
 527
 528	// See explanation in evlist__close()
 529	if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 530		if (affinity__setup(&saved_affinity) < 0)
 531			return;
 532		affinity = &saved_affinity;
 533	}
 534
 535	evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 536		pos = evlist_cpu_itr.evsel;
 537		if (evsel__strcmp(pos, evsel_name))
 538			continue;
 539		if (!evsel__is_group_leader(pos) || !pos->core.fd)
 540			continue;
 541		if (excl_dummy && evsel__is_dummy_event(pos))
 542			continue;
 543		evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 544	}
 545	affinity__cleanup(affinity);
 546	evlist__for_each_entry(evlist, pos) {
 547		if (evsel__strcmp(pos, evsel_name))
 548			continue;
 549		if (!evsel__is_group_leader(pos) || !pos->core.fd)
 550			continue;
 551		if (excl_dummy && evsel__is_dummy_event(pos))
 552			continue;
 553		pos->disabled = false;
 554	}
 555
 556	/*
 557	 * Even single event sets the 'enabled' for evlist,
 558	 * so the toggle can work properly and toggle to
 559	 * 'disabled' state.
 560	 */
 561	evlist->enabled = true;
 562}
 563
 564void evlist__enable(struct evlist *evlist)
 565{
 566	__evlist__enable(evlist, NULL, false);
 567}
 568
 569void evlist__enable_non_dummy(struct evlist *evlist)
 570{
 571	__evlist__enable(evlist, NULL, true);
 572}
 573
 574void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
 575{
 576	__evlist__enable(evlist, evsel_name, false);
 577}
 578
 579void evlist__toggle_enable(struct evlist *evlist)
 580{
 581	(evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 582}
 583
 584int evlist__add_pollfd(struct evlist *evlist, int fd)
 585{
 586	return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
 587}
 588
 589int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 590{
 591	return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 592}
 593
 594#ifdef HAVE_EVENTFD_SUPPORT
 595int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
 596{
 597	return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
 598				       fdarray_flag__nonfilterable |
 599				       fdarray_flag__non_perf_event);
 600}
 601#endif
 602
 603int evlist__poll(struct evlist *evlist, int timeout)
 604{
 605	return perf_evlist__poll(&evlist->core, timeout);
 606}
 607
 608struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
 609{
 610	struct hlist_head *head;
 611	struct perf_sample_id *sid;
 612	int hash;
 613
 614	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 615	head = &evlist->core.heads[hash];
 616
 617	hlist_for_each_entry(sid, head, node)
 618		if (sid->id == id)
 619			return sid;
 620
 621	return NULL;
 622}
 623
 624struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
 625{
 626	struct perf_sample_id *sid;
 627
 628	if (evlist->core.nr_entries == 1 || !id)
 629		return evlist__first(evlist);
 630
 631	sid = evlist__id2sid(evlist, id);
 632	if (sid)
 633		return container_of(sid->evsel, struct evsel, core);
 634
 635	if (!evlist__sample_id_all(evlist))
 636		return evlist__first(evlist);
 637
 638	return NULL;
 639}
 640
 641struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
 642{
 643	struct perf_sample_id *sid;
 644
 645	if (!id)
 646		return NULL;
 647
 648	sid = evlist__id2sid(evlist, id);
 649	if (sid)
 650		return container_of(sid->evsel, struct evsel, core);
 651
 652	return NULL;
 653}
 654
 655static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
 656{
 657	const __u64 *array = event->sample.array;
 658	ssize_t n;
 659
 660	n = (event->header.size - sizeof(event->header)) >> 3;
 661
 662	if (event->header.type == PERF_RECORD_SAMPLE) {
 663		if (evlist->id_pos >= n)
 664			return -1;
 665		*id = array[evlist->id_pos];
 666	} else {
 667		if (evlist->is_pos > n)
 668			return -1;
 669		n -= evlist->is_pos;
 670		*id = array[n];
 671	}
 672	return 0;
 673}
 674
 675struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
 676{
 677	struct evsel *first = evlist__first(evlist);
 678	struct hlist_head *head;
 679	struct perf_sample_id *sid;
 680	int hash;
 681	u64 id;
 682
 683	if (evlist->core.nr_entries == 1)
 684		return first;
 685
 686	if (!first->core.attr.sample_id_all &&
 687	    event->header.type != PERF_RECORD_SAMPLE)
 688		return first;
 689
 690	if (evlist__event2id(evlist, event, &id))
 691		return NULL;
 692
 693	/* Synthesized events have an id of zero */
 694	if (!id)
 695		return first;
 696
 697	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 698	head = &evlist->core.heads[hash];
 699
 700	hlist_for_each_entry(sid, head, node) {
 701		if (sid->id == id)
 702			return container_of(sid->evsel, struct evsel, core);
 703	}
 704	return NULL;
 705}
 706
 707static int evlist__set_paused(struct evlist *evlist, bool value)
 708{
 709	int i;
 710
 711	if (!evlist->overwrite_mmap)
 712		return 0;
 713
 714	for (i = 0; i < evlist->core.nr_mmaps; i++) {
 715		int fd = evlist->overwrite_mmap[i].core.fd;
 716		int err;
 717
 718		if (fd < 0)
 719			continue;
 720		err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 721		if (err)
 722			return err;
 723	}
 724	return 0;
 725}
 726
 727static int evlist__pause(struct evlist *evlist)
 728{
 729	return evlist__set_paused(evlist, true);
 730}
 731
 732static int evlist__resume(struct evlist *evlist)
 733{
 734	return evlist__set_paused(evlist, false);
 735}
 736
 737static void evlist__munmap_nofree(struct evlist *evlist)
 738{
 739	int i;
 740
 741	if (evlist->mmap)
 742		for (i = 0; i < evlist->core.nr_mmaps; i++)
 743			perf_mmap__munmap(&evlist->mmap[i].core);
 744
 745	if (evlist->overwrite_mmap)
 746		for (i = 0; i < evlist->core.nr_mmaps; i++)
 747			perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 748}
 749
 750void evlist__munmap(struct evlist *evlist)
 751{
 752	evlist__munmap_nofree(evlist);
 753	zfree(&evlist->mmap);
 754	zfree(&evlist->overwrite_mmap);
 755}
 756
 757static void perf_mmap__unmap_cb(struct perf_mmap *map)
 758{
 759	struct mmap *m = container_of(map, struct mmap, core);
 760
 761	mmap__munmap(m);
 762}
 763
 764static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 765				       bool overwrite)
 766{
 767	int i;
 768	struct mmap *map;
 769
 770	map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 771	if (!map)
 772		return NULL;
 773
 774	for (i = 0; i < evlist->core.nr_mmaps; i++) {
 775		struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 776
 777		/*
 778		 * When the perf_mmap() call is made we grab one refcount, plus
 779		 * one extra to let perf_mmap__consume() get the last
 780		 * events after all real references (perf_mmap__get()) are
 781		 * dropped.
 782		 *
 783		 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 784		 * thus does perf_mmap__get() on it.
 785		 */
 786		perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 787	}
 788
 789	return map;
 790}
 791
 792static void
 793perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 794			 struct perf_evsel *_evsel,
 795			 struct perf_mmap_param *_mp,
 796			 int idx)
 797{
 798	struct evlist *evlist = container_of(_evlist, struct evlist, core);
 799	struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 800	struct evsel *evsel = container_of(_evsel, struct evsel, core);
 801
 802	auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
 803}
 804
 805static struct perf_mmap*
 806perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 807{
 808	struct evlist *evlist = container_of(_evlist, struct evlist, core);
 809	struct mmap *maps;
 810
 811	maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 812
 813	if (!maps) {
 814		maps = evlist__alloc_mmap(evlist, overwrite);
 815		if (!maps)
 816			return NULL;
 817
 818		if (overwrite) {
 819			evlist->overwrite_mmap = maps;
 820			if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 821				evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 822		} else {
 823			evlist->mmap = maps;
 824		}
 825	}
 826
 827	return &maps[idx].core;
 828}
 829
 830static int
 831perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 832			  int output, struct perf_cpu cpu)
 833{
 834	struct mmap *map = container_of(_map, struct mmap, core);
 835	struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 836
 837	return mmap__mmap(map, mp, output, cpu);
 838}
 839
 840unsigned long perf_event_mlock_kb_in_pages(void)
 841{
 842	unsigned long pages;
 843	int max;
 844
 845	if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 846		/*
 847		 * Pick a once upon a time good value, i.e. things look
 848		 * strange since we can't read a sysctl value, but lets not
 849		 * die yet...
 850		 */
 851		max = 512;
 852	} else {
 853		max -= (page_size / 1024);
 854	}
 855
 856	pages = (max * 1024) / page_size;
 857	if (!is_power_of_2(pages))
 858		pages = rounddown_pow_of_two(pages);
 859
 860	return pages;
 861}
 862
 863size_t evlist__mmap_size(unsigned long pages)
 864{
 865	if (pages == UINT_MAX)
 866		pages = perf_event_mlock_kb_in_pages();
 867	else if (!is_power_of_2(pages))
 868		return 0;
 869
 870	return (pages + 1) * page_size;
 871}
 872
 873static long parse_pages_arg(const char *str, unsigned long min,
 874			    unsigned long max)
 875{
 876	unsigned long pages, val;
 877	static struct parse_tag tags[] = {
 878		{ .tag  = 'B', .mult = 1       },
 879		{ .tag  = 'K', .mult = 1 << 10 },
 880		{ .tag  = 'M', .mult = 1 << 20 },
 881		{ .tag  = 'G', .mult = 1 << 30 },
 882		{ .tag  = 0 },
 883	};
 884
 885	if (str == NULL)
 886		return -EINVAL;
 887
 888	val = parse_tag_value(str, tags);
 889	if (val != (unsigned long) -1) {
 890		/* we got file size value */
 891		pages = PERF_ALIGN(val, page_size) / page_size;
 892	} else {
 893		/* we got pages count value */
 894		char *eptr;
 895		pages = strtoul(str, &eptr, 10);
 896		if (*eptr != '\0')
 897			return -EINVAL;
 898	}
 899
 900	if (pages == 0 && min == 0) {
 901		/* leave number of pages at 0 */
 902	} else if (!is_power_of_2(pages)) {
 903		char buf[100];
 904
 905		/* round pages up to next power of 2 */
 906		pages = roundup_pow_of_two(pages);
 907		if (!pages)
 908			return -EINVAL;
 909
 910		unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 911		pr_info("rounding mmap pages size to %s (%lu pages)\n",
 912			buf, pages);
 913	}
 914
 915	if (pages > max)
 916		return -EINVAL;
 917
 918	return pages;
 919}
 920
 921int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 922{
 923	unsigned long max = UINT_MAX;
 924	long pages;
 925
 926	if (max > SIZE_MAX / page_size)
 927		max = SIZE_MAX / page_size;
 928
 929	pages = parse_pages_arg(str, 1, max);
 930	if (pages < 0) {
 931		pr_err("Invalid argument for --mmap_pages/-m\n");
 932		return -1;
 933	}
 934
 935	*mmap_pages = pages;
 936	return 0;
 937}
 938
 939int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
 940{
 941	return __evlist__parse_mmap_pages(opt->value, str);
 942}
 943
 944/**
 945 * evlist__mmap_ex - Create mmaps to receive events.
 946 * @evlist: list of events
 947 * @pages: map length in pages
 948 * @overwrite: overwrite older events?
 949 * @auxtrace_pages - auxtrace map length in pages
 950 * @auxtrace_overwrite - overwrite older auxtrace data?
 951 *
 952 * If @overwrite is %false the user needs to signal event consumption using
 953 * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 954 * automatically.
 955 *
 956 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 957 * consumption using auxtrace_mmap__write_tail().
 958 *
 959 * Return: %0 on success, negative error code otherwise.
 960 */
 961int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 962			 unsigned int auxtrace_pages,
 963			 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 964			 int comp_level)
 965{
 966	/*
 967	 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 968	 * Its value is decided by evsel's write_backward.
 969	 * So &mp should not be passed through const pointer.
 970	 */
 971	struct mmap_params mp = {
 972		.nr_cblocks	= nr_cblocks,
 973		.affinity	= affinity,
 974		.flush		= flush,
 975		.comp_level	= comp_level
 976	};
 977	struct perf_evlist_mmap_ops ops = {
 978		.idx  = perf_evlist__mmap_cb_idx,
 979		.get  = perf_evlist__mmap_cb_get,
 980		.mmap = perf_evlist__mmap_cb_mmap,
 981	};
 982
 983	evlist->core.mmap_len = evlist__mmap_size(pages);
 984	pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 985
 986	auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 987				   auxtrace_pages, auxtrace_overwrite);
 988
 989	return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 990}
 991
 992int evlist__mmap(struct evlist *evlist, unsigned int pages)
 993{
 994	return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 995}
 996
 997int evlist__create_maps(struct evlist *evlist, struct target *target)
 998{
 999	bool all_threads = (target->per_thread && target->system_wide);
1000	struct perf_cpu_map *cpus;
1001	struct perf_thread_map *threads;
1002
1003	/*
1004	 * If specify '-a' and '--per-thread' to perf record, perf record
1005	 * will override '--per-thread'. target->per_thread = false and
1006	 * target->system_wide = true.
1007	 *
1008	 * If specify '--per-thread' only to perf record,
1009	 * target->per_thread = true and target->system_wide = false.
1010	 *
1011	 * So target->per_thread && target->system_wide is false.
1012	 * For perf record, thread_map__new_str doesn't call
1013	 * thread_map__new_all_cpus. That will keep perf record's
1014	 * current behavior.
1015	 *
1016	 * For perf stat, it allows the case that target->per_thread and
1017	 * target->system_wide are all true. It means to collect system-wide
1018	 * per-thread data. thread_map__new_str will call
1019	 * thread_map__new_all_cpus to enumerate all threads.
1020	 */
1021	threads = thread_map__new_str(target->pid, target->tid, all_threads);
1022
1023	if (!threads)
1024		return -1;
1025
1026	if (target__uses_dummy_map(target) && !evlist__has_bpf_output(evlist))
1027		cpus = perf_cpu_map__new_any_cpu();
1028	else
1029		cpus = perf_cpu_map__new(target->cpu_list);
1030
1031	if (!cpus)
1032		goto out_delete_threads;
1033
1034	evlist->core.has_user_cpus = !!target->cpu_list;
1035
1036	perf_evlist__set_maps(&evlist->core, cpus, threads);
1037
1038	/* as evlist now has references, put count here */
1039	perf_cpu_map__put(cpus);
1040	perf_thread_map__put(threads);
1041
1042	return 0;
1043
1044out_delete_threads:
1045	perf_thread_map__put(threads);
1046	return -1;
1047}
1048
1049int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel,
1050			  struct target *target)
1051{
1052	struct evsel *evsel;
1053	int err = 0;
1054
1055	evlist__for_each_entry(evlist, evsel) {
1056		/*
1057		 * filters only work for tracepoint event, which doesn't have cpu limit.
1058		 * So evlist and evsel should always be same.
1059		 */
1060		if (evsel->filter) {
1061			err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1062			if (err) {
1063				*err_evsel = evsel;
1064				break;
1065			}
1066		}
1067
1068		/*
1069		 * non-tracepoint events can have BPF filters.
1070		 */
1071		if (!list_empty(&evsel->bpf_filters)) {
1072			err = perf_bpf_filter__prepare(evsel, target);
1073			if (err) {
1074				*err_evsel = evsel;
1075				break;
1076			}
1077		}
1078	}
1079
1080	return err;
1081}
1082
1083int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1084{
1085	struct evsel *evsel;
1086	int err = 0;
1087
1088	if (filter == NULL)
1089		return -1;
1090
1091	evlist__for_each_entry(evlist, evsel) {
1092		if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1093			continue;
1094
1095		err = evsel__set_filter(evsel, filter);
1096		if (err)
1097			break;
1098	}
1099
1100	return err;
1101}
1102
1103int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1104{
1105	struct evsel *evsel;
1106	int err = 0;
1107
1108	if (filter == NULL)
1109		return -1;
1110
1111	evlist__for_each_entry(evlist, evsel) {
1112		if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1113			continue;
1114
1115		err = evsel__append_tp_filter(evsel, filter);
1116		if (err)
1117			break;
1118	}
1119
1120	return err;
1121}
1122
1123char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1124{
1125	char *filter;
1126	size_t i;
1127
1128	for (i = 0; i < npids; ++i) {
1129		if (i == 0) {
1130			if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1131				return NULL;
1132		} else {
1133			char *tmp;
1134
1135			if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1136				goto out_free;
1137
1138			free(filter);
1139			filter = tmp;
1140		}
1141	}
1142
1143	return filter;
1144out_free:
1145	free(filter);
1146	return NULL;
1147}
1148
1149int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1150{
1151	char *filter = asprintf__tp_filter_pids(npids, pids);
1152	int ret = evlist__set_tp_filter(evlist, filter);
1153
1154	free(filter);
1155	return ret;
1156}
1157
1158int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1159{
1160	char *filter = asprintf__tp_filter_pids(npids, pids);
1161	int ret = evlist__append_tp_filter(evlist, filter);
1162
1163	free(filter);
1164	return ret;
1165}
1166
1167int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1168{
1169	return evlist__append_tp_filter_pids(evlist, 1, &pid);
1170}
1171
1172bool evlist__valid_sample_type(struct evlist *evlist)
1173{
1174	struct evsel *pos;
1175
1176	if (evlist->core.nr_entries == 1)
1177		return true;
1178
1179	if (evlist->id_pos < 0 || evlist->is_pos < 0)
1180		return false;
1181
1182	evlist__for_each_entry(evlist, pos) {
1183		if (pos->id_pos != evlist->id_pos ||
1184		    pos->is_pos != evlist->is_pos)
1185			return false;
1186	}
1187
1188	return true;
1189}
1190
1191u64 __evlist__combined_sample_type(struct evlist *evlist)
1192{
1193	struct evsel *evsel;
1194
1195	if (evlist->combined_sample_type)
1196		return evlist->combined_sample_type;
1197
1198	evlist__for_each_entry(evlist, evsel)
1199		evlist->combined_sample_type |= evsel->core.attr.sample_type;
1200
1201	return evlist->combined_sample_type;
1202}
1203
1204u64 evlist__combined_sample_type(struct evlist *evlist)
1205{
1206	evlist->combined_sample_type = 0;
1207	return __evlist__combined_sample_type(evlist);
1208}
1209
1210u64 evlist__combined_branch_type(struct evlist *evlist)
1211{
1212	struct evsel *evsel;
1213	u64 branch_type = 0;
1214
1215	evlist__for_each_entry(evlist, evsel)
1216		branch_type |= evsel->core.attr.branch_sample_type;
1217	return branch_type;
1218}
1219
1220static struct evsel *
1221evlist__find_dup_event_from_prev(struct evlist *evlist, struct evsel *event)
1222{
1223	struct evsel *pos;
1224
1225	evlist__for_each_entry(evlist, pos) {
1226		if (event == pos)
1227			break;
1228		if ((pos->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) &&
1229		    !strcmp(pos->name, event->name))
1230			return pos;
1231	}
1232	return NULL;
1233}
1234
1235#define MAX_NR_ABBR_NAME	(26 * 11)
1236
1237/*
1238 * The abbr name is from A to Z9. If the number of event
1239 * which requires the branch counter > MAX_NR_ABBR_NAME,
1240 * return NA.
1241 */
1242static void evlist__new_abbr_name(char *name)
1243{
1244	static int idx;
1245	int i = idx / 26;
1246
1247	if (idx >= MAX_NR_ABBR_NAME) {
1248		name[0] = 'N';
1249		name[1] = 'A';
1250		name[2] = '\0';
1251		return;
1252	}
1253
1254	name[0] = 'A' + (idx % 26);
1255
1256	if (!i)
1257		name[1] = '\0';
1258	else {
1259		name[1] = '0' + i - 1;
1260		name[2] = '\0';
1261	}
1262
1263	idx++;
1264}
1265
1266void evlist__update_br_cntr(struct evlist *evlist)
1267{
1268	struct evsel *evsel, *dup;
1269	int i = 0;
1270
1271	evlist__for_each_entry(evlist, evsel) {
1272		if (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) {
1273			evsel->br_cntr_idx = i++;
1274			evsel__leader(evsel)->br_cntr_nr++;
1275
1276			dup = evlist__find_dup_event_from_prev(evlist, evsel);
1277			if (dup)
1278				memcpy(evsel->abbr_name, dup->abbr_name, 3 * sizeof(char));
1279			else
1280				evlist__new_abbr_name(evsel->abbr_name);
1281		}
1282	}
1283	evlist->nr_br_cntr = i;
1284}
1285
1286bool evlist__valid_read_format(struct evlist *evlist)
1287{
1288	struct evsel *first = evlist__first(evlist), *pos = first;
1289	u64 read_format = first->core.attr.read_format;
1290	u64 sample_type = first->core.attr.sample_type;
1291
1292	evlist__for_each_entry(evlist, pos) {
1293		if (read_format != pos->core.attr.read_format) {
1294			pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1295				 read_format, (u64)pos->core.attr.read_format);
1296		}
1297	}
1298
1299	/* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1300	if ((sample_type & PERF_SAMPLE_READ) &&
1301	    !(read_format & PERF_FORMAT_ID)) {
1302		return false;
1303	}
1304
1305	return true;
1306}
1307
1308u16 evlist__id_hdr_size(struct evlist *evlist)
1309{
1310	struct evsel *first = evlist__first(evlist);
1311
1312	return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1313}
1314
1315bool evlist__valid_sample_id_all(struct evlist *evlist)
1316{
1317	struct evsel *first = evlist__first(evlist), *pos = first;
1318
1319	evlist__for_each_entry_continue(evlist, pos) {
1320		if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1321			return false;
1322	}
1323
1324	return true;
1325}
1326
1327bool evlist__sample_id_all(struct evlist *evlist)
1328{
1329	struct evsel *first = evlist__first(evlist);
1330	return first->core.attr.sample_id_all;
1331}
1332
1333void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1334{
1335	evlist->selected = evsel;
1336}
1337
1338void evlist__close(struct evlist *evlist)
1339{
1340	struct evsel *evsel;
1341	struct evlist_cpu_iterator evlist_cpu_itr;
1342	struct affinity affinity;
1343
1344	/*
1345	 * With perf record core.user_requested_cpus is usually NULL.
1346	 * Use the old method to handle this for now.
1347	 */
1348	if (!evlist->core.user_requested_cpus ||
1349	    cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1350		evlist__for_each_entry_reverse(evlist, evsel)
1351			evsel__close(evsel);
1352		return;
1353	}
1354
1355	if (affinity__setup(&affinity) < 0)
1356		return;
1357
1358	evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1359		perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1360				      evlist_cpu_itr.cpu_map_idx);
1361	}
1362
1363	affinity__cleanup(&affinity);
1364	evlist__for_each_entry_reverse(evlist, evsel) {
1365		perf_evsel__free_fd(&evsel->core);
1366		perf_evsel__free_id(&evsel->core);
1367	}
1368	perf_evlist__reset_id_hash(&evlist->core);
1369}
1370
1371static int evlist__create_syswide_maps(struct evlist *evlist)
1372{
1373	struct perf_cpu_map *cpus;
1374	struct perf_thread_map *threads;
1375
1376	/*
1377	 * Try reading /sys/devices/system/cpu/online to get
1378	 * an all cpus map.
1379	 *
1380	 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1381	 * code needs an overhaul to properly forward the
1382	 * error, and we may not want to do that fallback to a
1383	 * default cpu identity map :-\
1384	 */
1385	cpus = perf_cpu_map__new_online_cpus();
1386	if (!cpus)
1387		return -ENOMEM;
1388
1389	threads = perf_thread_map__new_dummy();
1390	if (!threads) {
1391		perf_cpu_map__put(cpus);
1392		return -ENOMEM;
1393	}
1394
1395	perf_evlist__set_maps(&evlist->core, cpus, threads);
1396	perf_thread_map__put(threads);
1397	perf_cpu_map__put(cpus);
1398	return 0;
1399}
1400
1401int evlist__open(struct evlist *evlist)
1402{
1403	struct evsel *evsel;
1404	int err;
1405
1406	/*
1407	 * Default: one fd per CPU, all threads, aka systemwide
1408	 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1409	 */
1410	if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1411		err = evlist__create_syswide_maps(evlist);
1412		if (err < 0)
1413			goto out_err;
1414	}
1415
1416	evlist__update_id_pos(evlist);
1417
1418	evlist__for_each_entry(evlist, evsel) {
1419		err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1420		if (err < 0)
1421			goto out_err;
1422	}
1423
1424	return 0;
1425out_err:
1426	evlist__close(evlist);
1427	errno = -err;
1428	return err;
1429}
1430
1431int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1432			     bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1433{
1434	int child_ready_pipe[2], go_pipe[2];
1435	char bf;
1436
1437	evlist->workload.cork_fd = -1;
1438
1439	if (pipe(child_ready_pipe) < 0) {
1440		perror("failed to create 'ready' pipe");
1441		return -1;
1442	}
1443
1444	if (pipe(go_pipe) < 0) {
1445		perror("failed to create 'go' pipe");
1446		goto out_close_ready_pipe;
1447	}
1448
1449	evlist->workload.pid = fork();
1450	if (evlist->workload.pid < 0) {
1451		perror("failed to fork");
1452		goto out_close_pipes;
1453	}
1454
1455	if (!evlist->workload.pid) {
1456		int ret;
1457
1458		if (pipe_output)
1459			dup2(2, 1);
1460
1461		signal(SIGTERM, SIG_DFL);
1462
1463		close(child_ready_pipe[0]);
1464		close(go_pipe[1]);
1465		fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1466
1467		/*
1468		 * Change the name of this process not to confuse --exclude-perf users
1469		 * that sees 'perf' in the window up to the execvp() and thinks that
1470		 * perf samples are not being excluded.
1471		 */
1472		prctl(PR_SET_NAME, "perf-exec");
1473
1474		/*
1475		 * Tell the parent we're ready to go
1476		 */
1477		close(child_ready_pipe[1]);
1478
1479		/*
1480		 * Wait until the parent tells us to go.
1481		 */
1482		ret = read(go_pipe[0], &bf, 1);
1483		/*
1484		 * The parent will ask for the execvp() to be performed by
1485		 * writing exactly one byte, in workload.cork_fd, usually via
1486		 * evlist__start_workload().
1487		 *
1488		 * For cancelling the workload without actually running it,
1489		 * the parent will just close workload.cork_fd, without writing
1490		 * anything, i.e. read will return zero and we just exit()
1491		 * here (See evlist__cancel_workload()).
1492		 */
1493		if (ret != 1) {
1494			if (ret == -1)
1495				perror("unable to read pipe");
1496			exit(ret);
1497		}
1498
1499		execvp(argv[0], (char **)argv);
1500
1501		if (exec_error) {
1502			union sigval val;
1503
1504			val.sival_int = errno;
1505			if (sigqueue(getppid(), SIGUSR1, val))
1506				perror(argv[0]);
1507		} else
1508			perror(argv[0]);
1509		exit(-1);
1510	}
1511
1512	if (exec_error) {
1513		struct sigaction act = {
1514			.sa_flags     = SA_SIGINFO,
1515			.sa_sigaction = exec_error,
1516		};
1517		sigaction(SIGUSR1, &act, NULL);
1518	}
1519
1520	if (target__none(target)) {
1521		if (evlist->core.threads == NULL) {
1522			fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1523				__func__, __LINE__);
1524			goto out_close_pipes;
1525		}
1526		perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1527	}
1528
1529	close(child_ready_pipe[1]);
1530	close(go_pipe[0]);
1531	/*
1532	 * wait for child to settle
1533	 */
1534	if (read(child_ready_pipe[0], &bf, 1) == -1) {
1535		perror("unable to read pipe");
1536		goto out_close_pipes;
1537	}
1538
1539	fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1540	evlist->workload.cork_fd = go_pipe[1];
1541	close(child_ready_pipe[0]);
1542	return 0;
1543
1544out_close_pipes:
1545	close(go_pipe[0]);
1546	close(go_pipe[1]);
1547out_close_ready_pipe:
1548	close(child_ready_pipe[0]);
1549	close(child_ready_pipe[1]);
1550	return -1;
1551}
1552
1553int evlist__start_workload(struct evlist *evlist)
1554{
1555	if (evlist->workload.cork_fd >= 0) {
1556		char bf = 0;
1557		int ret;
1558		/*
1559		 * Remove the cork, let it rip!
1560		 */
1561		ret = write(evlist->workload.cork_fd, &bf, 1);
1562		if (ret < 0)
1563			perror("unable to write to pipe");
1564
1565		close(evlist->workload.cork_fd);
1566		evlist->workload.cork_fd = -1;
1567		return ret;
1568	}
1569
1570	return 0;
1571}
1572
1573void evlist__cancel_workload(struct evlist *evlist)
1574{
1575	int status;
1576
1577	if (evlist->workload.cork_fd >= 0) {
1578		close(evlist->workload.cork_fd);
1579		evlist->workload.cork_fd = -1;
1580		waitpid(evlist->workload.pid, &status, WNOHANG);
1581	}
1582}
1583
1584int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1585{
1586	struct evsel *evsel = evlist__event2evsel(evlist, event);
1587	int ret;
1588
1589	if (!evsel)
1590		return -EFAULT;
1591	ret = evsel__parse_sample(evsel, event, sample);
1592	if (ret)
1593		return ret;
1594	if (perf_guest && sample->id) {
1595		struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1596
1597		if (sid) {
1598			sample->machine_pid = sid->machine_pid;
1599			sample->vcpu = sid->vcpu.cpu;
1600		}
1601	}
1602	return 0;
1603}
1604
1605int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1606{
1607	struct evsel *evsel = evlist__event2evsel(evlist, event);
1608
1609	if (!evsel)
1610		return -EFAULT;
1611	return evsel__parse_sample_timestamp(evsel, event, timestamp);
1612}
1613
1614int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1615{
1616	int printed, value;
1617	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1618
1619	switch (err) {
1620	case EACCES:
1621	case EPERM:
1622		printed = scnprintf(buf, size,
1623				    "Error:\t%s.\n"
1624				    "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1625
1626		value = perf_event_paranoid();
1627
1628		printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1629
1630		if (value >= 2) {
1631			printed += scnprintf(buf + printed, size - printed,
1632					     "For your workloads it needs to be <= 1\nHint:\t");
1633		}
1634		printed += scnprintf(buf + printed, size - printed,
1635				     "For system wide tracing it needs to be set to -1.\n");
1636
1637		printed += scnprintf(buf + printed, size - printed,
1638				    "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1639				    "Hint:\tThe current value is %d.", value);
1640		break;
1641	case EINVAL: {
1642		struct evsel *first = evlist__first(evlist);
1643		int max_freq;
1644
1645		if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1646			goto out_default;
1647
1648		if (first->core.attr.sample_freq < (u64)max_freq)
1649			goto out_default;
1650
1651		printed = scnprintf(buf, size,
1652				    "Error:\t%s.\n"
1653				    "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1654				    "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1655				    emsg, max_freq, first->core.attr.sample_freq);
1656		break;
1657	}
1658	default:
1659out_default:
1660		scnprintf(buf, size, "%s", emsg);
1661		break;
1662	}
1663
1664	return 0;
1665}
1666
1667int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1668{
1669	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1670	int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1671
1672	switch (err) {
1673	case EPERM:
1674		sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1675		printed += scnprintf(buf + printed, size - printed,
1676				     "Error:\t%s.\n"
1677				     "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1678				     "Hint:\tTried using %zd kB.\n",
1679				     emsg, pages_max_per_user, pages_attempted);
1680
1681		if (pages_attempted >= pages_max_per_user) {
1682			printed += scnprintf(buf + printed, size - printed,
1683					     "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1684					     pages_max_per_user + pages_attempted);
1685		}
1686
1687		printed += scnprintf(buf + printed, size - printed,
1688				     "Hint:\tTry using a smaller -m/--mmap-pages value.");
1689		break;
1690	default:
1691		scnprintf(buf, size, "%s", emsg);
1692		break;
1693	}
1694
1695	return 0;
1696}
1697
1698void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1699{
1700	struct evsel *evsel, *n;
1701	LIST_HEAD(move);
1702
1703	if (move_evsel == evlist__first(evlist))
1704		return;
1705
1706	evlist__for_each_entry_safe(evlist, n, evsel) {
1707		if (evsel__leader(evsel) == evsel__leader(move_evsel))
1708			list_move_tail(&evsel->core.node, &move);
1709	}
1710
1711	list_splice(&move, &evlist->core.entries);
1712}
1713
1714struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1715{
1716	struct evsel *evsel;
1717
1718	evlist__for_each_entry(evlist, evsel) {
1719		if (evsel->tracking)
1720			return evsel;
1721	}
1722
1723	return evlist__first(evlist);
1724}
1725
1726void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1727{
1728	struct evsel *evsel;
1729
1730	if (tracking_evsel->tracking)
1731		return;
1732
1733	evlist__for_each_entry(evlist, evsel) {
1734		if (evsel != tracking_evsel)
1735			evsel->tracking = false;
1736	}
1737
1738	tracking_evsel->tracking = true;
1739}
1740
1741struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide)
1742{
1743	struct evsel *evsel;
1744
1745	evsel = evlist__get_tracking_event(evlist);
1746	if (!evsel__is_dummy_event(evsel)) {
1747		evsel = evlist__add_aux_dummy(evlist, system_wide);
1748		if (!evsel)
1749			return NULL;
1750
1751		evlist__set_tracking_event(evlist, evsel);
1752	} else if (system_wide) {
1753		perf_evlist__go_system_wide(&evlist->core, &evsel->core);
1754	}
1755
1756	return evsel;
1757}
1758
1759struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1760{
1761	struct evsel *evsel;
1762
1763	evlist__for_each_entry(evlist, evsel) {
1764		if (!evsel->name)
1765			continue;
1766		if (evsel__name_is(evsel, str))
1767			return evsel;
1768	}
1769
1770	return NULL;
1771}
1772
1773void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1774{
1775	enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1776	enum action {
1777		NONE,
1778		PAUSE,
1779		RESUME,
1780	} action = NONE;
1781
1782	if (!evlist->overwrite_mmap)
1783		return;
1784
1785	switch (old_state) {
1786	case BKW_MMAP_NOTREADY: {
1787		if (state != BKW_MMAP_RUNNING)
1788			goto state_err;
1789		break;
1790	}
1791	case BKW_MMAP_RUNNING: {
1792		if (state != BKW_MMAP_DATA_PENDING)
1793			goto state_err;
1794		action = PAUSE;
1795		break;
1796	}
1797	case BKW_MMAP_DATA_PENDING: {
1798		if (state != BKW_MMAP_EMPTY)
1799			goto state_err;
1800		break;
1801	}
1802	case BKW_MMAP_EMPTY: {
1803		if (state != BKW_MMAP_RUNNING)
1804			goto state_err;
1805		action = RESUME;
1806		break;
1807	}
1808	default:
1809		WARN_ONCE(1, "Shouldn't get there\n");
1810	}
1811
1812	evlist->bkw_mmap_state = state;
1813
1814	switch (action) {
1815	case PAUSE:
1816		evlist__pause(evlist);
1817		break;
1818	case RESUME:
1819		evlist__resume(evlist);
1820		break;
1821	case NONE:
1822	default:
1823		break;
1824	}
1825
1826state_err:
1827	return;
1828}
1829
1830bool evlist__exclude_kernel(struct evlist *evlist)
1831{
1832	struct evsel *evsel;
1833
1834	evlist__for_each_entry(evlist, evsel) {
1835		if (!evsel->core.attr.exclude_kernel)
1836			return false;
1837	}
1838
1839	return true;
1840}
1841
1842/*
1843 * Events in data file are not collect in groups, but we still want
1844 * the group display. Set the artificial group and set the leader's
1845 * forced_leader flag to notify the display code.
1846 */
1847void evlist__force_leader(struct evlist *evlist)
1848{
1849	if (evlist__nr_groups(evlist) == 0) {
1850		struct evsel *leader = evlist__first(evlist);
1851
1852		evlist__set_leader(evlist);
1853		leader->forced_leader = true;
1854	}
1855}
1856
1857struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1858{
1859	struct evsel *c2, *leader;
1860	bool is_open = true;
1861
1862	leader = evsel__leader(evsel);
1863
1864	pr_debug("Weak group for %s/%d failed\n",
1865			leader->name, leader->core.nr_members);
1866
1867	/*
1868	 * for_each_group_member doesn't work here because it doesn't
1869	 * include the first entry.
1870	 */
1871	evlist__for_each_entry(evsel_list, c2) {
1872		if (c2 == evsel)
1873			is_open = false;
1874		if (evsel__has_leader(c2, leader)) {
1875			if (is_open && close)
1876				perf_evsel__close(&c2->core);
1877			/*
1878			 * We want to close all members of the group and reopen
1879			 * them. Some events, like Intel topdown, require being
1880			 * in a group and so keep these in the group.
1881			 */
1882			evsel__remove_from_group(c2, leader);
1883
1884			/*
1885			 * Set this for all former members of the group
1886			 * to indicate they get reopened.
1887			 */
1888			c2->reset_group = true;
1889		}
1890	}
1891	/* Reset the leader count if all entries were removed. */
1892	if (leader->core.nr_members == 1)
1893		leader->core.nr_members = 0;
1894	return leader;
1895}
1896
1897static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1898{
1899	char *s, *p;
1900	int ret = 0, fd;
1901
1902	if (strncmp(str, "fifo:", 5))
1903		return -EINVAL;
1904
1905	str += 5;
1906	if (!*str || *str == ',')
1907		return -EINVAL;
1908
1909	s = strdup(str);
1910	if (!s)
1911		return -ENOMEM;
1912
1913	p = strchr(s, ',');
1914	if (p)
1915		*p = '\0';
1916
1917	/*
1918	 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1919	 * end of a FIFO to be repeatedly opened and closed.
1920	 */
1921	fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1922	if (fd < 0) {
1923		pr_err("Failed to open '%s'\n", s);
1924		ret = -errno;
1925		goto out_free;
1926	}
1927	*ctl_fd = fd;
1928	*ctl_fd_close = true;
1929
1930	if (p && *++p) {
1931		/* O_RDWR | O_NONBLOCK means the other end need not be open */
1932		fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1933		if (fd < 0) {
1934			pr_err("Failed to open '%s'\n", p);
1935			ret = -errno;
1936			goto out_free;
1937		}
1938		*ctl_fd_ack = fd;
1939	}
1940
1941out_free:
1942	free(s);
1943	return ret;
1944}
1945
1946int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1947{
1948	char *comma = NULL, *endptr = NULL;
1949
1950	*ctl_fd_close = false;
1951
1952	if (strncmp(str, "fd:", 3))
1953		return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1954
1955	*ctl_fd = strtoul(&str[3], &endptr, 0);
1956	if (endptr == &str[3])
1957		return -EINVAL;
1958
1959	comma = strchr(str, ',');
1960	if (comma) {
1961		if (endptr != comma)
1962			return -EINVAL;
1963
1964		*ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1965		if (endptr == comma + 1 || *endptr != '\0')
1966			return -EINVAL;
1967	}
1968
1969	return 0;
1970}
1971
1972void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1973{
1974	if (*ctl_fd_close) {
1975		*ctl_fd_close = false;
1976		close(ctl_fd);
1977		if (ctl_fd_ack >= 0)
1978			close(ctl_fd_ack);
1979	}
1980}
1981
1982int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1983{
1984	if (fd == -1) {
1985		pr_debug("Control descriptor is not initialized\n");
1986		return 0;
1987	}
1988
1989	evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1990						     fdarray_flag__nonfilterable |
1991						     fdarray_flag__non_perf_event);
1992	if (evlist->ctl_fd.pos < 0) {
1993		evlist->ctl_fd.pos = -1;
1994		pr_err("Failed to add ctl fd entry: %m\n");
1995		return -1;
1996	}
1997
1998	evlist->ctl_fd.fd = fd;
1999	evlist->ctl_fd.ack = ack;
2000
2001	return 0;
2002}
2003
2004bool evlist__ctlfd_initialized(struct evlist *evlist)
2005{
2006	return evlist->ctl_fd.pos >= 0;
2007}
2008
2009int evlist__finalize_ctlfd(struct evlist *evlist)
2010{
2011	struct pollfd *entries = evlist->core.pollfd.entries;
2012
2013	if (!evlist__ctlfd_initialized(evlist))
2014		return 0;
2015
2016	entries[evlist->ctl_fd.pos].fd = -1;
2017	entries[evlist->ctl_fd.pos].events = 0;
2018	entries[evlist->ctl_fd.pos].revents = 0;
2019
2020	evlist->ctl_fd.pos = -1;
2021	evlist->ctl_fd.ack = -1;
2022	evlist->ctl_fd.fd = -1;
2023
2024	return 0;
2025}
2026
2027static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
2028			      char *cmd_data, size_t data_size)
2029{
2030	int err;
2031	char c;
2032	size_t bytes_read = 0;
2033
2034	*cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2035	memset(cmd_data, 0, data_size);
2036	data_size--;
2037
2038	do {
2039		err = read(evlist->ctl_fd.fd, &c, 1);
2040		if (err > 0) {
2041			if (c == '\n' || c == '\0')
2042				break;
2043			cmd_data[bytes_read++] = c;
2044			if (bytes_read == data_size)
2045				break;
2046			continue;
2047		} else if (err == -1) {
2048			if (errno == EINTR)
2049				continue;
2050			if (errno == EAGAIN || errno == EWOULDBLOCK)
2051				err = 0;
2052			else
2053				pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
2054		}
2055		break;
2056	} while (1);
2057
2058	pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
2059		 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
2060
2061	if (bytes_read > 0) {
2062		if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
2063			     (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
2064			*cmd = EVLIST_CTL_CMD_ENABLE;
2065		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
2066				    (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
2067			*cmd = EVLIST_CTL_CMD_DISABLE;
2068		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2069				    (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2070			*cmd = EVLIST_CTL_CMD_SNAPSHOT;
2071			pr_debug("is snapshot\n");
2072		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2073				    (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2074			*cmd = EVLIST_CTL_CMD_EVLIST;
2075		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2076				    (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2077			*cmd = EVLIST_CTL_CMD_STOP;
2078		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2079				    (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2080			*cmd = EVLIST_CTL_CMD_PING;
2081		}
2082	}
2083
2084	return bytes_read ? (int)bytes_read : err;
2085}
2086
2087int evlist__ctlfd_ack(struct evlist *evlist)
2088{
2089	int err;
2090
2091	if (evlist->ctl_fd.ack == -1)
2092		return 0;
2093
2094	err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2095		    sizeof(EVLIST_CTL_CMD_ACK_TAG));
2096	if (err == -1)
2097		pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2098
2099	return err;
2100}
2101
2102static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2103{
2104	char *data = cmd_data + cmd_size;
2105
2106	/* no argument */
2107	if (!*data)
2108		return 0;
2109
2110	/* there's argument */
2111	if (*data == ' ') {
2112		*arg = data + 1;
2113		return 1;
2114	}
2115
2116	/* malformed */
2117	return -1;
2118}
2119
2120static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2121{
2122	struct evsel *evsel;
2123	char *name;
2124	int err;
2125
2126	err = get_cmd_arg(cmd_data,
2127			  enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2128				   sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2129			  &name);
2130	if (err < 0) {
2131		pr_info("failed: wrong command\n");
2132		return -1;
2133	}
2134
2135	if (err) {
2136		evsel = evlist__find_evsel_by_str(evlist, name);
2137		if (evsel) {
2138			if (enable)
2139				evlist__enable_evsel(evlist, name);
2140			else
2141				evlist__disable_evsel(evlist, name);
2142			pr_info("Event %s %s\n", evsel->name,
2143				enable ? "enabled" : "disabled");
2144		} else {
2145			pr_info("failed: can't find '%s' event\n", name);
2146		}
2147	} else {
2148		if (enable) {
2149			evlist__enable(evlist);
2150			pr_info(EVLIST_ENABLED_MSG);
2151		} else {
2152			evlist__disable(evlist);
2153			pr_info(EVLIST_DISABLED_MSG);
2154		}
2155	}
2156
2157	return 0;
2158}
2159
2160static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2161{
2162	struct perf_attr_details details = { .verbose = false, };
2163	struct evsel *evsel;
2164	char *arg;
2165	int err;
2166
2167	err = get_cmd_arg(cmd_data,
2168			  sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2169			  &arg);
2170	if (err < 0) {
2171		pr_info("failed: wrong command\n");
2172		return -1;
2173	}
2174
2175	if (err) {
2176		if (!strcmp(arg, "-v")) {
2177			details.verbose = true;
2178		} else if (!strcmp(arg, "-g")) {
2179			details.event_group = true;
2180		} else if (!strcmp(arg, "-F")) {
2181			details.freq = true;
2182		} else {
2183			pr_info("failed: wrong command\n");
2184			return -1;
2185		}
2186	}
2187
2188	evlist__for_each_entry(evlist, evsel)
2189		evsel__fprintf(evsel, &details, stderr);
2190
2191	return 0;
2192}
2193
2194int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2195{
2196	int err = 0;
2197	char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2198	int ctlfd_pos = evlist->ctl_fd.pos;
2199	struct pollfd *entries = evlist->core.pollfd.entries;
2200
2201	if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2202		return 0;
2203
2204	if (entries[ctlfd_pos].revents & POLLIN) {
2205		err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2206					 EVLIST_CTL_CMD_MAX_LEN);
2207		if (err > 0) {
2208			switch (*cmd) {
2209			case EVLIST_CTL_CMD_ENABLE:
2210			case EVLIST_CTL_CMD_DISABLE:
2211				err = evlist__ctlfd_enable(evlist, cmd_data,
2212							   *cmd == EVLIST_CTL_CMD_ENABLE);
2213				break;
2214			case EVLIST_CTL_CMD_EVLIST:
2215				err = evlist__ctlfd_list(evlist, cmd_data);
2216				break;
2217			case EVLIST_CTL_CMD_SNAPSHOT:
2218			case EVLIST_CTL_CMD_STOP:
2219			case EVLIST_CTL_CMD_PING:
2220				break;
2221			case EVLIST_CTL_CMD_ACK:
2222			case EVLIST_CTL_CMD_UNSUPPORTED:
2223			default:
2224				pr_debug("ctlfd: unsupported %d\n", *cmd);
2225				break;
2226			}
2227			if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2228			      *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2229				evlist__ctlfd_ack(evlist);
2230		}
2231	}
2232
2233	if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2234		evlist__finalize_ctlfd(evlist);
2235	else
2236		entries[ctlfd_pos].revents = 0;
2237
2238	return err;
2239}
2240
2241/**
2242 * struct event_enable_time - perf record -D/--delay single time range.
2243 * @start: start of time range to enable events in milliseconds
2244 * @end: end of time range to enable events in milliseconds
2245 *
2246 * N.B. this structure is also accessed as an array of int.
2247 */
2248struct event_enable_time {
2249	int	start;
2250	int	end;
2251};
2252
2253static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2254{
2255	const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2256	int ret, start, end, n;
2257
2258	ret = sscanf(str, fmt, &start, &end, &n);
2259	if (ret != 2 || end <= start)
2260		return -EINVAL;
2261	if (range) {
2262		range->start = start;
2263		range->end = end;
2264	}
2265	return n;
2266}
2267
2268static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2269{
2270	int incr = !!range;
2271	bool first = true;
2272	ssize_t ret, cnt;
2273
2274	for (cnt = 0; *str; cnt++) {
2275		ret = parse_event_enable_time(str, range, first);
2276		if (ret < 0)
2277			return ret;
2278		/* Check no overlap */
2279		if (!first && range && range->start <= range[-1].end)
2280			return -EINVAL;
2281		str += ret;
2282		range += incr;
2283		first = false;
2284	}
2285	return cnt;
2286}
2287
2288/**
2289 * struct event_enable_timer - control structure for perf record -D/--delay.
2290 * @evlist: event list
2291 * @times: time ranges that events are enabled (N.B. this is also accessed as an
2292 *         array of int)
2293 * @times_cnt: number of time ranges
2294 * @timerfd: timer file descriptor
2295 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2296 * @times_step: current position in (int *)@times)[],
2297 *              refer event_enable_timer__process()
2298 *
2299 * Note, this structure is only used when there are time ranges, not when there
2300 * is only an initial delay.
2301 */
2302struct event_enable_timer {
2303	struct evlist *evlist;
2304	struct event_enable_time *times;
2305	size_t	times_cnt;
2306	int	timerfd;
2307	int	pollfd_pos;
2308	size_t	times_step;
2309};
2310
2311static int str_to_delay(const char *str)
2312{
2313	char *endptr;
2314	long d;
2315
2316	d = strtol(str, &endptr, 10);
2317	if (*endptr || d > INT_MAX || d < -1)
2318		return 0;
2319	return d;
2320}
2321
2322int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2323				    const char *str, int unset)
2324{
2325	enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2326	struct event_enable_timer *eet;
2327	ssize_t times_cnt;
2328	ssize_t ret;
2329	int err;
2330
2331	if (unset)
2332		return 0;
2333
2334	opts->target.initial_delay = str_to_delay(str);
2335	if (opts->target.initial_delay)
2336		return 0;
2337
2338	ret = parse_event_enable_times(str, NULL);
2339	if (ret < 0)
2340		return ret;
2341
2342	times_cnt = ret;
2343	if (times_cnt == 0)
2344		return -EINVAL;
2345
2346	eet = zalloc(sizeof(*eet));
2347	if (!eet)
2348		return -ENOMEM;
2349
2350	eet->times = calloc(times_cnt, sizeof(*eet->times));
2351	if (!eet->times) {
2352		err = -ENOMEM;
2353		goto free_eet;
2354	}
2355
2356	if (parse_event_enable_times(str, eet->times) != times_cnt) {
2357		err = -EINVAL;
2358		goto free_eet_times;
2359	}
2360
2361	eet->times_cnt = times_cnt;
2362
2363	eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2364	if (eet->timerfd == -1) {
2365		err = -errno;
2366		pr_err("timerfd_create failed: %s\n", strerror(errno));
2367		goto free_eet_times;
2368	}
2369
2370	eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2371	if (eet->pollfd_pos < 0) {
2372		err = eet->pollfd_pos;
2373		goto close_timerfd;
2374	}
2375
2376	eet->evlist = evlist;
2377	evlist->eet = eet;
2378	opts->target.initial_delay = eet->times[0].start;
2379
2380	return 0;
2381
2382close_timerfd:
2383	close(eet->timerfd);
2384free_eet_times:
2385	zfree(&eet->times);
2386free_eet:
2387	free(eet);
2388	return err;
2389}
2390
2391static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2392{
2393	struct itimerspec its = {
2394		.it_value.tv_sec = ms / MSEC_PER_SEC,
2395		.it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2396	};
2397	int err = 0;
2398
2399	if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2400		err = -errno;
2401		pr_err("timerfd_settime failed: %s\n", strerror(errno));
2402	}
2403	return err;
2404}
2405
2406int event_enable_timer__start(struct event_enable_timer *eet)
2407{
2408	int ms;
2409
2410	if (!eet)
2411		return 0;
2412
2413	ms = eet->times[0].end - eet->times[0].start;
2414	eet->times_step = 1;
2415
2416	return event_enable_timer__set_timer(eet, ms);
2417}
2418
2419int event_enable_timer__process(struct event_enable_timer *eet)
2420{
2421	struct pollfd *entries;
2422	short revents;
2423
2424	if (!eet)
2425		return 0;
2426
2427	entries = eet->evlist->core.pollfd.entries;
2428	revents = entries[eet->pollfd_pos].revents;
2429	entries[eet->pollfd_pos].revents = 0;
2430
2431	if (revents & POLLIN) {
2432		size_t step = eet->times_step;
2433		size_t pos = step / 2;
2434
2435		if (step & 1) {
2436			evlist__disable_non_dummy(eet->evlist);
2437			pr_info(EVLIST_DISABLED_MSG);
2438			if (pos >= eet->times_cnt - 1) {
2439				/* Disarm timer */
2440				event_enable_timer__set_timer(eet, 0);
2441				return 1; /* Stop */
2442			}
2443		} else {
2444			evlist__enable_non_dummy(eet->evlist);
2445			pr_info(EVLIST_ENABLED_MSG);
2446		}
2447
2448		step += 1;
2449		pos = step / 2;
2450
2451		if (pos < eet->times_cnt) {
2452			int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2453			int ms = times[step] - times[step - 1];
2454
2455			eet->times_step = step;
2456			return event_enable_timer__set_timer(eet, ms);
2457		}
2458	}
2459
2460	return 0;
2461}
2462
2463void event_enable_timer__exit(struct event_enable_timer **ep)
2464{
2465	if (!ep || !*ep)
2466		return;
2467	zfree(&(*ep)->times);
2468	zfree(ep);
2469}
2470
2471struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2472{
2473	struct evsel *evsel;
2474
2475	evlist__for_each_entry(evlist, evsel) {
2476		if (evsel->core.idx == idx)
2477			return evsel;
2478	}
2479	return NULL;
2480}
2481
2482void evlist__format_evsels(struct evlist *evlist, struct strbuf *sb, size_t max_length)
2483{
2484	struct evsel *evsel, *leader = NULL;
2485	bool first = true;
2486
2487	evlist__for_each_entry(evlist, evsel) {
2488		struct evsel *new_leader = evsel__leader(evsel);
2489
2490		if (evsel__is_dummy_event(evsel))
2491			continue;
2492
2493		if (leader != new_leader && leader && leader->core.nr_members > 1)
2494			strbuf_addch(sb, '}');
2495
2496		if (!first)
2497			strbuf_addch(sb, ',');
2498
2499		if (sb->len > max_length) {
2500			strbuf_addstr(sb, "...");
2501			return;
2502		}
2503		if (leader != new_leader && new_leader->core.nr_members > 1)
2504			strbuf_addch(sb, '{');
2505
2506		strbuf_addstr(sb, evsel__name(evsel));
2507		first = false;
2508		leader = new_leader;
2509	}
2510	if (leader && leader->core.nr_members > 1)
2511		strbuf_addch(sb, '}');
2512}
2513
2514void evlist__check_mem_load_aux(struct evlist *evlist)
2515{
2516	struct evsel *leader, *evsel, *pos;
2517
2518	/*
2519	 * For some platforms, the 'mem-loads' event is required to use
2520	 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2521	 * must be the group leader. Now we disable this group before reporting
2522	 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2523	 * any valid memory load information.
2524	 */
2525	evlist__for_each_entry(evlist, evsel) {
2526		leader = evsel__leader(evsel);
2527		if (leader == evsel)
2528			continue;
2529
2530		if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2531			for_each_group_evsel(pos, leader) {
2532				evsel__set_leader(pos, pos);
2533				pos->core.nr_members = 0;
2534			}
2535		}
2536	}
2537}
2538
2539/**
2540 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs
2541 *     and warn if the user CPU list is inapplicable for the event's PMU's
2542 *     CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a
2543 *     user requested CPU and so any online CPU is applicable. Core PMUs handle
2544 *     events on the CPUs in their list and otherwise the event isn't supported.
2545 * @evlist: The list of events being checked.
2546 * @cpu_list: The user provided list of CPUs.
2547 */
2548void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list)
2549{
2550	struct perf_cpu_map *user_requested_cpus;
2551	struct evsel *pos;
2552
2553	if (!cpu_list)
2554		return;
2555
2556	user_requested_cpus = perf_cpu_map__new(cpu_list);
2557	if (!user_requested_cpus)
2558		return;
2559
2560	evlist__for_each_entry(evlist, pos) {
2561		evsel__warn_user_requested_cpus(pos, user_requested_cpus);
2562	}
2563	perf_cpu_map__put(user_requested_cpus);
2564}
2565
2566/* Should uniquify be disabled for the evlist? */
2567static bool evlist__disable_uniquify(const struct evlist *evlist)
2568{
2569	struct evsel *counter;
2570	struct perf_pmu *last_pmu = NULL;
2571	bool first = true;
2572
2573	evlist__for_each_entry(evlist, counter) {
2574		/* If PMUs vary then uniquify can be useful. */
2575		if (!first && counter->pmu != last_pmu)
2576			return false;
2577		first = false;
2578		if (counter->pmu) {
2579			/* Allow uniquify for uncore PMUs. */
2580			if (!counter->pmu->is_core)
2581				return false;
2582			/* Keep hybrid event names uniquified for clarity. */
2583			if (perf_pmus__num_core_pmus() > 1)
2584				return false;
2585		}
2586		last_pmu = counter->pmu;
2587	}
2588	return true;
2589}
2590
2591static bool evlist__set_needs_uniquify(struct evlist *evlist, const struct perf_stat_config *config)
2592{
2593	struct evsel *counter;
2594	bool needs_uniquify = false;
2595
2596	if (evlist__disable_uniquify(evlist)) {
2597		evlist__for_each_entry(evlist, counter)
2598			counter->uniquified_name = true;
2599		return false;
2600	}
2601
2602	evlist__for_each_entry(evlist, counter) {
2603		if (evsel__set_needs_uniquify(counter, config))
2604			needs_uniquify = true;
2605	}
2606	return needs_uniquify;
2607}
2608
2609void evlist__uniquify_evsel_names(struct evlist *evlist, const struct perf_stat_config *config)
2610{
2611	if (evlist__set_needs_uniquify(evlist, config)) {
2612		struct evsel *pos;
2613
2614		evlist__for_each_entry(evlist, pos)
2615			evsel__uniquify_counter(pos);
2616	}
2617}
2618
2619bool evlist__has_bpf_output(struct evlist *evlist)
2620{
2621	struct evsel *evsel;
2622
2623	evlist__for_each_entry(evlist, evsel) {
2624		if (evsel__is_bpf_output(evsel))
2625			return true;
2626	}
2627
2628	return false;
2629}
2630
2631bool evlist__needs_bpf_sb_event(struct evlist *evlist)
2632{
2633	struct evsel *evsel;
2634
2635	evlist__for_each_entry(evlist, evsel) {
2636		if (evsel__is_dummy_event(evsel))
2637			continue;
2638		if (!evsel->core.attr.exclude_kernel)
2639			return true;
2640	}
2641
2642	return false;
2643}