1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * builtin-top.c
   4 *
   5 * Builtin top command: Display a continuously updated profile of
   6 * any workload, CPU or specific PID.
   7 *
   8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   9 *		 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  10 *
  11 * Improvements and fixes by:
  12 *
  13 *   Arjan van de Ven <arjan@linux.intel.com>
  14 *   Yanmin Zhang <yanmin.zhang@intel.com>
  15 *   Wu Fengguang <fengguang.wu@intel.com>
  16 *   Mike Galbraith <efault@gmx.de>
  17 *   Paul Mackerras <paulus@samba.org>
  18 */
  19#include "builtin.h"
  20
  21#include "perf.h"
  22
  23#include "util/annotate.h"
  24#include "util/bpf-event.h"
  25#include "util/cgroup.h"
  26#include "util/config.h"
  27#include "util/color.h"
  28#include "util/dso.h"
  29#include "util/evlist.h"
  30#include "util/evsel.h"
  31#include "util/evsel_config.h"
  32#include "util/event.h"
  33#include "util/machine.h"
  34#include "util/map.h"
  35#include "util/mmap.h"
  36#include "util/session.h"
  37#include "util/thread.h"
  38#include "util/symbol.h"
  39#include "util/synthetic-events.h"
  40#include "util/top.h"
  41#include "util/util.h"
  42#include <linux/rbtree.h>
  43#include <subcmd/parse-options.h>
  44#include "util/parse-events.h"
  45#include "util/callchain.h"
  46#include "util/cpumap.h"
  47#include "util/sort.h"
  48#include "util/string2.h"
  49#include "util/term.h"
  50#include "util/intlist.h"
  51#include "util/parse-branch-options.h"
  52#include "arch/common.h"
  53#include "ui/ui.h"
  54
  55#include "util/debug.h"
  56#include "util/ordered-events.h"
  57#include "util/pfm.h"
  58
  59#include <assert.h>
  60#include <elf.h>
  61#include <fcntl.h>
  62
  63#include <stdio.h>
  64#include <termios.h>
  65#include <unistd.h>
  66#include <inttypes.h>
  67
  68#include <errno.h>
  69#include <time.h>
  70#include <sched.h>
  71#include <signal.h>
  72
  73#include <sys/syscall.h>
  74#include <sys/ioctl.h>
  75#include <poll.h>
  76#include <sys/prctl.h>
  77#include <sys/wait.h>
  78#include <sys/uio.h>
  79#include <sys/utsname.h>
  80#include <sys/mman.h>
  81
  82#include <linux/stringify.h>
  83#include <linux/time64.h>
  84#include <linux/types.h>
  85#include <linux/err.h>
  86
  87#include <linux/ctype.h>
  88#include <perf/mmap.h>
  89
  90static volatile sig_atomic_t done;
  91static volatile sig_atomic_t resize;
  92
  93#define HEADER_LINE_NR  5
  94
  95static void perf_top__update_print_entries(struct perf_top *top)
  96{
  97	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  98}
  99
 100static void winch_sig(int sig __maybe_unused)
 101{
 102	resize = 1;
 103}
 104
 105static void perf_top__resize(struct perf_top *top)
 106{
 107	get_term_dimensions(&top->winsize);
 108	perf_top__update_print_entries(top);
 109}
 110
 111static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 112{
 113	struct evsel *evsel;
 114	struct symbol *sym;
 115	struct annotation *notes;
 116	struct map *map;
 117	int err = -1;
 118
 119	if (!he || !he->ms.sym)
 120		return -1;
 121
 122	evsel = hists_to_evsel(he->hists);
 123
 124	sym = he->ms.sym;
 125	map = he->ms.map;
 126
 127	/*
 128	 * We can't annotate with just /proc/kallsyms
 129	 */
 130	if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 131	    !dso__is_kcore(map->dso)) {
 132		pr_err("Can't annotate %s: No vmlinux file was found in the "
 133		       "path\n", sym->name);
 134		sleep(1);
 135		return -1;
 136	}
 137
 138	notes = symbol__annotation(sym);
 139	mutex_lock(&notes->lock);
 140
 141	if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 142		mutex_unlock(&notes->lock);
 143		pr_err("Not enough memory for annotating '%s' symbol!\n",
 144		       sym->name);
 145		sleep(1);
 146		return err;
 147	}
 148
 149	err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
 150	if (err == 0) {
 151		top->sym_filter_entry = he;
 152	} else {
 153		char msg[BUFSIZ];
 154		symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
 155		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 156	}
 157
 158	mutex_unlock(&notes->lock);
 159	return err;
 160}
 161
 162static void __zero_source_counters(struct hist_entry *he)
 163{
 164	struct symbol *sym = he->ms.sym;
 165	symbol__annotate_zero_histograms(sym);
 166}
 167
 168static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 169{
 170	struct utsname uts;
 171	int err = uname(&uts);
 172
 173	ui__warning("Out of bounds address found:\n\n"
 174		    "Addr:   %" PRIx64 "\n"
 175		    "DSO:    %s %c\n"
 176		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
 177		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 178		    "Arch:   %s\n"
 179		    "Kernel: %s\n"
 180		    "Tools:  %s\n\n"
 181		    "Not all samples will be on the annotation output.\n\n"
 182		    "Please report to linux-kernel@vger.kernel.org\n",
 183		    ip, map->dso->long_name, dso__symtab_origin(map->dso),
 184		    map->start, map->end, sym->start, sym->end,
 185		    sym->binding == STB_GLOBAL ? 'g' :
 186		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 187		    err ? "[unknown]" : uts.machine,
 188		    err ? "[unknown]" : uts.release, perf_version_string);
 189	if (use_browser <= 0)
 190		sleep(5);
 191
 192	map->erange_warned = true;
 193}
 194
 195static void perf_top__record_precise_ip(struct perf_top *top,
 196					struct hist_entry *he,
 197					struct perf_sample *sample,
 198					struct evsel *evsel, u64 ip)
 199	EXCLUSIVE_LOCKS_REQUIRED(he->hists->lock)
 200{
 201	struct annotation *notes;
 202	struct symbol *sym = he->ms.sym;
 203	int err = 0;
 204
 205	if (sym == NULL || (use_browser == 0 &&
 206			    (top->sym_filter_entry == NULL ||
 207			     top->sym_filter_entry->ms.sym != sym)))
 208		return;
 209
 210	notes = symbol__annotation(sym);
 211
 212	if (!mutex_trylock(&notes->lock))
 213		return;
 214
 215	err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 216
 217	mutex_unlock(&notes->lock);
 218
 219	if (unlikely(err)) {
 220		/*
 221		 * This function is now called with he->hists->lock held.
 222		 * Release it before going to sleep.
 223		 */
 224		mutex_unlock(&he->hists->lock);
 225
 226		if (err == -ERANGE && !he->ms.map->erange_warned)
 227			ui__warn_map_erange(he->ms.map, sym, ip);
 228		else if (err == -ENOMEM) {
 229			pr_err("Not enough memory for annotating '%s' symbol!\n",
 230			       sym->name);
 231			sleep(1);
 232		}
 233
 234		mutex_lock(&he->hists->lock);
 235	}
 236}
 237
 238static void perf_top__show_details(struct perf_top *top)
 239{
 240	struct hist_entry *he = top->sym_filter_entry;
 241	struct evsel *evsel;
 242	struct annotation *notes;
 243	struct symbol *symbol;
 244	int more;
 245
 246	if (!he)
 247		return;
 248
 249	evsel = hists_to_evsel(he->hists);
 250
 251	symbol = he->ms.sym;
 252	notes = symbol__annotation(symbol);
 253
 254	mutex_lock(&notes->lock);
 255
 256	symbol__calc_percent(symbol, evsel);
 257
 258	if (notes->src == NULL)
 259		goto out_unlock;
 260
 261	printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
 262	printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 263
 264	more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
 265
 266	if (top->evlist->enabled) {
 267		if (top->zero)
 268			symbol__annotate_zero_histogram(symbol, top->sym_evsel->core.idx);
 269		else
 270			symbol__annotate_decay_histogram(symbol, top->sym_evsel->core.idx);
 271	}
 272	if (more != 0)
 273		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 274out_unlock:
 275	mutex_unlock(&notes->lock);
 276}
 277
 278static void perf_top__resort_hists(struct perf_top *t)
 279{
 280	struct evlist *evlist = t->evlist;
 281	struct evsel *pos;
 282
 283	evlist__for_each_entry(evlist, pos) {
 284		struct hists *hists = evsel__hists(pos);
 285
 286		/*
 287		 * unlink existing entries so that they can be linked
 288		 * in a correct order in hists__match() below.
 289		 */
 290		hists__unlink(hists);
 291
 292		if (evlist->enabled) {
 293			if (t->zero) {
 294				hists__delete_entries(hists);
 295			} else {
 296				hists__decay_entries(hists, t->hide_user_symbols,
 297						     t->hide_kernel_symbols);
 298			}
 299		}
 300
 301		hists__collapse_resort(hists, NULL);
 302
 303		/* Non-group events are considered as leader */
 304		if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
 305			struct hists *leader_hists = evsel__hists(evsel__leader(pos));
 306
 307			hists__match(leader_hists, hists);
 308			hists__link(leader_hists, hists);
 309		}
 310	}
 311
 312	evlist__for_each_entry(evlist, pos) {
 313		evsel__output_resort(pos, NULL);
 314	}
 315}
 316
 317static void perf_top__print_sym_table(struct perf_top *top)
 318{
 319	char bf[160];
 320	int printed = 0;
 321	const int win_width = top->winsize.ws_col - 1;
 322	struct evsel *evsel = top->sym_evsel;
 323	struct hists *hists = evsel__hists(evsel);
 324
 325	puts(CONSOLE_CLEAR);
 326
 327	perf_top__header_snprintf(top, bf, sizeof(bf));
 328	printf("%s\n", bf);
 329
 330	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 331
 332	if (!top->record_opts.overwrite &&
 333	    (top->evlist->stats.nr_lost_warned !=
 334	     top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
 335		top->evlist->stats.nr_lost_warned =
 336			      top->evlist->stats.nr_events[PERF_RECORD_LOST];
 337		color_fprintf(stdout, PERF_COLOR_RED,
 338			      "WARNING: LOST %d chunks, Check IO/CPU overload",
 339			      top->evlist->stats.nr_lost_warned);
 340		++printed;
 341	}
 342
 343	if (top->sym_filter_entry) {
 344		perf_top__show_details(top);
 345		return;
 346	}
 347
 348	perf_top__resort_hists(top);
 349
 350	hists__output_recalc_col_len(hists, top->print_entries - printed);
 351	putchar('\n');
 352	hists__fprintf(hists, false, top->print_entries - printed, win_width,
 353		       top->min_percent, stdout, !symbol_conf.use_callchain);
 354}
 355
 356static void prompt_integer(int *target, const char *msg)
 357{
 358	char *buf = malloc(0), *p;
 359	size_t dummy = 0;
 360	int tmp;
 361
 362	fprintf(stdout, "\n%s: ", msg);
 363	if (getline(&buf, &dummy, stdin) < 0)
 364		return;
 365
 366	p = strchr(buf, '\n');
 367	if (p)
 368		*p = 0;
 369
 370	p = buf;
 371	while(*p) {
 372		if (!isdigit(*p))
 373			goto out_free;
 374		p++;
 375	}
 376	tmp = strtoul(buf, NULL, 10);
 377	*target = tmp;
 378out_free:
 379	free(buf);
 380}
 381
 382static void prompt_percent(int *target, const char *msg)
 383{
 384	int tmp = 0;
 385
 386	prompt_integer(&tmp, msg);
 387	if (tmp >= 0 && tmp <= 100)
 388		*target = tmp;
 389}
 390
 391static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 392{
 393	char *buf = malloc(0), *p;
 394	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 395	struct hists *hists = evsel__hists(top->sym_evsel);
 396	struct rb_node *next;
 397	size_t dummy = 0;
 398
 399	/* zero counters of active symbol */
 400	if (syme) {
 401		__zero_source_counters(syme);
 402		top->sym_filter_entry = NULL;
 403	}
 404
 405	fprintf(stdout, "\n%s: ", msg);
 406	if (getline(&buf, &dummy, stdin) < 0)
 407		goto out_free;
 408
 409	p = strchr(buf, '\n');
 410	if (p)
 411		*p = 0;
 412
 413	next = rb_first_cached(&hists->entries);
 414	while (next) {
 415		n = rb_entry(next, struct hist_entry, rb_node);
 416		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 417			found = n;
 418			break;
 419		}
 420		next = rb_next(&n->rb_node);
 421	}
 422
 423	if (!found) {
 424		fprintf(stderr, "Sorry, %s is not active.\n", buf);
 425		sleep(1);
 426	} else
 427		perf_top__parse_source(top, found);
 428
 429out_free:
 430	free(buf);
 431}
 432
 433static void perf_top__print_mapped_keys(struct perf_top *top)
 434{
 435	char *name = NULL;
 436
 437	if (top->sym_filter_entry) {
 438		struct symbol *sym = top->sym_filter_entry->ms.sym;
 439		name = sym->name;
 440	}
 441
 442	fprintf(stdout, "\nMapped keys:\n");
 443	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 444	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 445
 446	if (top->evlist->core.nr_entries > 1)
 447		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
 448
 449	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 450
 451	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 452	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 453	fprintf(stdout, "\t[S]     stop annotation.\n");
 454
 455	fprintf(stdout,
 456		"\t[K]     hide kernel symbols.             \t(%s)\n",
 457		top->hide_kernel_symbols ? "yes" : "no");
 458	fprintf(stdout,
 459		"\t[U]     hide user symbols.               \t(%s)\n",
 460		top->hide_user_symbols ? "yes" : "no");
 461	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 462	fprintf(stdout, "\t[qQ]    quit.\n");
 463}
 464
 465static int perf_top__key_mapped(struct perf_top *top, int c)
 466{
 467	switch (c) {
 468		case 'd':
 469		case 'e':
 470		case 'f':
 471		case 'z':
 472		case 'q':
 473		case 'Q':
 474		case 'K':
 475		case 'U':
 476		case 'F':
 477		case 's':
 478		case 'S':
 479			return 1;
 480		case 'E':
 481			return top->evlist->core.nr_entries > 1 ? 1 : 0;
 482		default:
 483			break;
 484	}
 485
 486	return 0;
 487}
 488
 489static bool perf_top__handle_keypress(struct perf_top *top, int c)
 490{
 491	bool ret = true;
 492
 493	if (!perf_top__key_mapped(top, c)) {
 494		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 495		struct termios save;
 496
 497		perf_top__print_mapped_keys(top);
 498		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 499		fflush(stdout);
 500
 501		set_term_quiet_input(&save);
 502
 503		poll(&stdin_poll, 1, -1);
 504		c = getc(stdin);
 505
 506		tcsetattr(0, TCSAFLUSH, &save);
 507		if (!perf_top__key_mapped(top, c))
 508			return ret;
 509	}
 510
 511	switch (c) {
 512		case 'd':
 513			prompt_integer(&top->delay_secs, "Enter display delay");
 514			if (top->delay_secs < 1)
 515				top->delay_secs = 1;
 516			break;
 517		case 'e':
 518			prompt_integer(&top->print_entries, "Enter display entries (lines)");
 519			if (top->print_entries == 0) {
 520				perf_top__resize(top);
 521				signal(SIGWINCH, winch_sig);
 522			} else {
 523				signal(SIGWINCH, SIG_DFL);
 524			}
 525			break;
 526		case 'E':
 527			if (top->evlist->core.nr_entries > 1) {
 528				/* Select 0 as the default event: */
 529				int counter = 0;
 530
 531				fprintf(stderr, "\nAvailable events:");
 532
 533				evlist__for_each_entry(top->evlist, top->sym_evsel)
 534					fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));
 535
 536				prompt_integer(&counter, "Enter details event counter");
 537
 538				if (counter >= top->evlist->core.nr_entries) {
 539					top->sym_evsel = evlist__first(top->evlist);
 540					fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
 541					sleep(1);
 542					break;
 543				}
 544				evlist__for_each_entry(top->evlist, top->sym_evsel)
 545					if (top->sym_evsel->core.idx == counter)
 546						break;
 547			} else
 548				top->sym_evsel = evlist__first(top->evlist);
 549			break;
 550		case 'f':
 551			prompt_integer(&top->count_filter, "Enter display event count filter");
 552			break;
 553		case 'F':
 554			prompt_percent(&top->annotation_opts.min_pcnt,
 555				       "Enter details display event filter (percent)");
 556			break;
 557		case 'K':
 558			top->hide_kernel_symbols = !top->hide_kernel_symbols;
 559			break;
 560		case 'q':
 561		case 'Q':
 562			printf("exiting.\n");
 563			if (top->dump_symtab)
 564				perf_session__fprintf_dsos(top->session, stderr);
 565			ret = false;
 566			break;
 567		case 's':
 568			perf_top__prompt_symbol(top, "Enter details symbol");
 569			break;
 570		case 'S':
 571			if (!top->sym_filter_entry)
 572				break;
 573			else {
 574				struct hist_entry *syme = top->sym_filter_entry;
 575
 576				top->sym_filter_entry = NULL;
 577				__zero_source_counters(syme);
 578			}
 579			break;
 580		case 'U':
 581			top->hide_user_symbols = !top->hide_user_symbols;
 582			break;
 583		case 'z':
 584			top->zero = !top->zero;
 585			break;
 586		default:
 587			break;
 588	}
 589
 590	return ret;
 591}
 592
 593static void perf_top__sort_new_samples(void *arg)
 594{
 595	struct perf_top *t = arg;
 596
 597	if (t->evlist->selected != NULL)
 598		t->sym_evsel = t->evlist->selected;
 599
 600	perf_top__resort_hists(t);
 601
 602	if (t->lost || t->drop)
 603		pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 604}
 605
 606static void stop_top(void)
 607{
 608	session_done = 1;
 609	done = 1;
 610}
 611
 612static void *display_thread_tui(void *arg)
 613{
 614	struct evsel *pos;
 615	struct perf_top *top = arg;
 616	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 617	struct hist_browser_timer hbt = {
 618		.timer		= perf_top__sort_new_samples,
 619		.arg		= top,
 620		.refresh	= top->delay_secs,
 621	};
 622	int ret;
 623
 624	/* In order to read symbols from other namespaces perf to  needs to call
 625	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 626	 * unshare(2) the fs so that we may continue to setns into namespaces
 627	 * that we're observing.
 628	 */
 629	unshare(CLONE_FS);
 630
 631	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 632
 633repeat:
 634	perf_top__sort_new_samples(top);
 635
 636	/*
 637	 * Initialize the uid_filter_str, in the future the TUI will allow
 638	 * Zooming in/out UIDs. For now just use whatever the user passed
 639	 * via --uid.
 640	 */
 641	evlist__for_each_entry(top->evlist, pos) {
 642		struct hists *hists = evsel__hists(pos);
 643		hists->uid_filter_str = top->record_opts.target.uid_str;
 644	}
 645
 646	ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 647				       &top->session->header.env, !top->record_opts.overwrite,
 648				       &top->annotation_opts);
 649	if (ret == K_RELOAD) {
 650		top->zero = true;
 651		goto repeat;
 652	} else
 653		stop_top();
 654
 655	return NULL;
 656}
 657
 658static void display_sig(int sig __maybe_unused)
 659{
 660	stop_top();
 661}
 662
 663static void display_setup_sig(void)
 664{
 665	signal(SIGSEGV, sighandler_dump_stack);
 666	signal(SIGFPE, sighandler_dump_stack);
 667	signal(SIGINT,  display_sig);
 668	signal(SIGQUIT, display_sig);
 669	signal(SIGTERM, display_sig);
 670}
 671
 672static void *display_thread(void *arg)
 673{
 674	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 675	struct termios save;
 676	struct perf_top *top = arg;
 677	int delay_msecs, c;
 678
 679	/* In order to read symbols from other namespaces perf to  needs to call
 680	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 681	 * unshare(2) the fs so that we may continue to setns into namespaces
 682	 * that we're observing.
 683	 */
 684	unshare(CLONE_FS);
 685
 686	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 687
 688	display_setup_sig();
 689	pthread__unblock_sigwinch();
 690repeat:
 691	delay_msecs = top->delay_secs * MSEC_PER_SEC;
 692	set_term_quiet_input(&save);
 693	/* trash return*/
 694	clearerr(stdin);
 695	if (poll(&stdin_poll, 1, 0) > 0)
 696		getc(stdin);
 697
 698	while (!done) {
 699		perf_top__print_sym_table(top);
 700		/*
 701		 * Either timeout expired or we got an EINTR due to SIGWINCH,
 702		 * refresh screen in both cases.
 703		 */
 704		switch (poll(&stdin_poll, 1, delay_msecs)) {
 705		case 0:
 706			continue;
 707		case -1:
 708			if (errno == EINTR)
 709				continue;
 710			__fallthrough;
 711		default:
 712			c = getc(stdin);
 713			tcsetattr(0, TCSAFLUSH, &save);
 714
 715			if (perf_top__handle_keypress(top, c))
 716				goto repeat;
 717			stop_top();
 718		}
 719	}
 720
 721	tcsetattr(0, TCSAFLUSH, &save);
 722	return NULL;
 723}
 724
 725static int hist_iter__top_callback(struct hist_entry_iter *iter,
 726				   struct addr_location *al, bool single,
 727				   void *arg)
 728	EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
 729{
 730	struct perf_top *top = arg;
 
 731	struct evsel *evsel = iter->evsel;
 732
 733	if (perf_hpp_list.sym && single)
 734		perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);
 735
 736	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 737		     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 738		     NULL);
 739	return 0;
 740}
 741
 742static void perf_event__process_sample(struct perf_tool *tool,
 743				       const union perf_event *event,
 744				       struct evsel *evsel,
 745				       struct perf_sample *sample,
 746				       struct machine *machine)
 747{
 748	struct perf_top *top = container_of(tool, struct perf_top, tool);
 749	struct addr_location al;
 
 750
 751	if (!machine && perf_guest) {
 752		static struct intlist *seen;
 753
 754		if (!seen)
 755			seen = intlist__new(NULL);
 756
 757		if (!intlist__has_entry(seen, sample->pid)) {
 758			pr_err("Can't find guest [%d]'s kernel information\n",
 759				sample->pid);
 760			intlist__add(seen, sample->pid);
 761		}
 762		return;
 763	}
 764
 765	if (!machine) {
 766		pr_err("%u unprocessable samples recorded.\r",
 767		       top->session->evlist->stats.nr_unprocessable_samples++);
 768		return;
 769	}
 770
 771	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 772		top->exact_samples++;
 773
 774	if (machine__resolve(machine, &al, sample) < 0)
 775		return;
 776
 777	if (top->stitch_lbr)
 778		al.thread->lbr_stitch_enable = true;
 779
 780	if (!machine->kptr_restrict_warned &&
 781	    symbol_conf.kptr_restrict &&
 782	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
 783		if (!evlist__exclude_kernel(top->session->evlist)) {
 784			ui__warning(
 785"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 786"Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 787"Kernel%s samples will not be resolved.\n",
 788			  al.map && map__has_symbols(al.map) ?
 789			  " modules" : "");
 790			if (use_browser <= 0)
 791				sleep(5);
 792		}
 793		machine->kptr_restrict_warned = true;
 794	}
 795
 796	if (al.sym == NULL && al.map != NULL) {
 797		const char *msg = "Kernel samples will not be resolved.\n";
 798		/*
 799		 * As we do lazy loading of symtabs we only will know if the
 800		 * specified vmlinux file is invalid when we actually have a
 801		 * hit in kernel space and then try to load it. So if we get
 802		 * here and there are _no_ symbols in the DSO backing the
 803		 * kernel map, bail out.
 804		 *
 805		 * We may never get here, for instance, if we use -K/
 806		 * --hide-kernel-symbols, even if the user specifies an
 807		 * invalid --vmlinux ;-)
 808		 */
 809		if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 810		    __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 811			if (symbol_conf.vmlinux_name) {
 812				char serr[256];
 813				dso__strerror_load(al.map->dso, serr, sizeof(serr));
 814				ui__warning("The %s file can't be used: %s\n%s",
 815					    symbol_conf.vmlinux_name, serr, msg);
 816			} else {
 817				ui__warning("A vmlinux file was not found.\n%s",
 818					    msg);
 819			}
 820
 821			if (use_browser <= 0)
 822				sleep(5);
 823			top->vmlinux_warned = true;
 824		}
 825	}
 826
 827	if (al.sym == NULL || !al.sym->idle) {
 828		struct hists *hists = evsel__hists(evsel);
 829		struct hist_entry_iter iter = {
 830			.evsel		= evsel,
 831			.sample 	= sample,
 832			.add_entry_cb 	= hist_iter__top_callback,
 833		};
 834
 835		if (symbol_conf.cumulate_callchain)
 836			iter.ops = &hist_iter_cumulative;
 837		else
 838			iter.ops = &hist_iter_normal;
 839
 840		mutex_lock(&hists->lock);
 841
 842		if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
 
 843			pr_err("Problem incrementing symbol period, skipping event\n");
 844
 845		mutex_unlock(&hists->lock);
 846	}
 847
 848	addr_location__put(&al);
 849}
 850
 851static void
 852perf_top__process_lost(struct perf_top *top, union perf_event *event,
 853		       struct evsel *evsel)
 854{
 855	top->lost += event->lost.lost;
 856	top->lost_total += event->lost.lost;
 857	evsel->evlist->stats.total_lost += event->lost.lost;
 858}
 859
 860static void
 861perf_top__process_lost_samples(struct perf_top *top,
 862			       union perf_event *event,
 863			       struct evsel *evsel)
 864{
 865	top->lost += event->lost_samples.lost;
 866	top->lost_total += event->lost_samples.lost;
 867	evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
 868}
 869
 870static u64 last_timestamp;
 871
 872static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 873{
 874	struct record_opts *opts = &top->record_opts;
 875	struct evlist *evlist = top->evlist;
 876	struct mmap *md;
 877	union perf_event *event;
 878
 879	md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 880	if (perf_mmap__read_init(&md->core) < 0)
 881		return;
 882
 883	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 884		int ret;
 885
 886		ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 887		if (ret && ret != -1)
 888			break;
 889
 890		ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
 891		if (ret)
 892			break;
 893
 894		perf_mmap__consume(&md->core);
 895
 896		if (top->qe.rotate) {
 897			mutex_lock(&top->qe.mutex);
 898			top->qe.rotate = false;
 899			cond_signal(&top->qe.cond);
 900			mutex_unlock(&top->qe.mutex);
 901		}
 902	}
 903
 904	perf_mmap__read_done(&md->core);
 905}
 906
 907static void perf_top__mmap_read(struct perf_top *top)
 908{
 909	bool overwrite = top->record_opts.overwrite;
 910	struct evlist *evlist = top->evlist;
 911	int i;
 912
 913	if (overwrite)
 914		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 915
 916	for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 917		perf_top__mmap_read_idx(top, i);
 918
 919	if (overwrite) {
 920		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 921		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 922	}
 923}
 924
 925/*
 926 * Check per-event overwrite term.
 927 * perf top should support consistent term for all events.
 928 * - All events don't have per-event term
 929 *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 930 *   Nothing change, return 0.
 931 * - All events have same per-event term
 932 *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 933 *   Using the per-event setting to replace the opts->overwrite if
 934 *   they are different, then return 0.
 935 * - Events have different per-event term
 936 *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 937 *   Return -1
 938 * - Some of the event set per-event term, but some not.
 939 *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 940 *   Return -1
 941 */
 942static int perf_top__overwrite_check(struct perf_top *top)
 943{
 944	struct record_opts *opts = &top->record_opts;
 945	struct evlist *evlist = top->evlist;
 946	struct evsel_config_term *term;
 947	struct list_head *config_terms;
 948	struct evsel *evsel;
 949	int set, overwrite = -1;
 950
 951	evlist__for_each_entry(evlist, evsel) {
 952		set = -1;
 953		config_terms = &evsel->config_terms;
 954		list_for_each_entry(term, config_terms, list) {
 955			if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
 956				set = term->val.overwrite ? 1 : 0;
 957		}
 958
 959		/* no term for current and previous event (likely) */
 960		if ((overwrite < 0) && (set < 0))
 961			continue;
 962
 963		/* has term for both current and previous event, compare */
 964		if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 965			return -1;
 966
 967		/* no term for current event but has term for previous one */
 968		if ((overwrite >= 0) && (set < 0))
 969			return -1;
 970
 971		/* has term for current event */
 972		if ((overwrite < 0) && (set >= 0)) {
 973			/* if it's first event, set overwrite */
 974			if (evsel == evlist__first(evlist))
 975				overwrite = set;
 976			else
 977				return -1;
 978		}
 979	}
 980
 981	if ((overwrite >= 0) && (opts->overwrite != overwrite))
 982		opts->overwrite = overwrite;
 983
 984	return 0;
 985}
 986
 987static int perf_top_overwrite_fallback(struct perf_top *top,
 988				       struct evsel *evsel)
 989{
 990	struct record_opts *opts = &top->record_opts;
 991	struct evlist *evlist = top->evlist;
 992	struct evsel *counter;
 993
 994	if (!opts->overwrite)
 995		return 0;
 996
 997	/* only fall back when first event fails */
 998	if (evsel != evlist__first(evlist))
 999		return 0;
1000
1001	evlist__for_each_entry(evlist, counter)
1002		counter->core.attr.write_backward = false;
1003	opts->overwrite = false;
1004	pr_debug2("fall back to non-overwrite mode\n");
1005	return 1;
1006}
1007
1008static int perf_top__start_counters(struct perf_top *top)
1009{
1010	char msg[BUFSIZ];
1011	struct evsel *counter;
1012	struct evlist *evlist = top->evlist;
1013	struct record_opts *opts = &top->record_opts;
1014
1015	if (perf_top__overwrite_check(top)) {
1016		ui__error("perf top only support consistent per-event "
1017			  "overwrite setting for all events\n");
1018		goto out_err;
1019	}
1020
1021	evlist__config(evlist, opts, &callchain_param);
1022
1023	evlist__for_each_entry(evlist, counter) {
1024try_again:
1025		if (evsel__open(counter, top->evlist->core.user_requested_cpus,
1026				     top->evlist->core.threads) < 0) {
1027
1028			/*
1029			 * Specially handle overwrite fall back.
1030			 * Because perf top is the only tool which has
1031			 * overwrite mode by default, support
1032			 * both overwrite and non-overwrite mode, and
1033			 * require consistent mode for all events.
1034			 *
1035			 * May move it to generic code with more tools
1036			 * have similar attribute.
1037			 */
1038			if (perf_missing_features.write_backward &&
1039			    perf_top_overwrite_fallback(top, counter))
1040				goto try_again;
1041
1042			if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1043				if (verbose > 0)
1044					ui__warning("%s\n", msg);
1045				goto try_again;
1046			}
1047
1048			evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1049			ui__error("%s\n", msg);
1050			goto out_err;
1051		}
1052	}
1053
1054	if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1055		ui__error("Failed to mmap with %d (%s)\n",
1056			    errno, str_error_r(errno, msg, sizeof(msg)));
1057		goto out_err;
1058	}
1059
1060	return 0;
1061
1062out_err:
1063	return -1;
1064}
1065
1066static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1067{
1068	if (callchain->mode != CHAIN_NONE) {
1069		if (callchain_register_param(callchain) < 0) {
1070			ui__error("Can't register callchain params.\n");
1071			return -EINVAL;
1072		}
1073	}
1074
1075	return 0;
1076}
1077
1078static struct ordered_events *rotate_queues(struct perf_top *top)
1079{
1080	struct ordered_events *in = top->qe.in;
1081
1082	if (top->qe.in == &top->qe.data[1])
1083		top->qe.in = &top->qe.data[0];
1084	else
1085		top->qe.in = &top->qe.data[1];
1086
1087	return in;
1088}
1089
1090static void *process_thread(void *arg)
1091{
1092	struct perf_top *top = arg;
1093
1094	while (!done) {
1095		struct ordered_events *out, *in = top->qe.in;
1096
1097		if (!in->nr_events) {
1098			usleep(100);
1099			continue;
1100		}
1101
1102		out = rotate_queues(top);
1103
1104		mutex_lock(&top->qe.mutex);
1105		top->qe.rotate = true;
1106		cond_wait(&top->qe.cond, &top->qe.mutex);
1107		mutex_unlock(&top->qe.mutex);
1108
1109		if (ordered_events__flush(out, OE_FLUSH__TOP))
1110			pr_err("failed to process events\n");
1111	}
1112
1113	return NULL;
1114}
1115
1116/*
1117 * Allow only 'top->delay_secs' seconds behind samples.
1118 */
1119static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1120{
1121	union perf_event *event = qevent->event;
1122	u64 delay_timestamp;
1123
1124	if (event->header.type != PERF_RECORD_SAMPLE)
1125		return false;
1126
1127	delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1128	return delay_timestamp < last_timestamp;
1129}
1130
1131static int deliver_event(struct ordered_events *qe,
1132			 struct ordered_event *qevent)
1133{
1134	struct perf_top *top = qe->data;
1135	struct evlist *evlist = top->evlist;
1136	struct perf_session *session = top->session;
1137	union perf_event *event = qevent->event;
1138	struct perf_sample sample;
1139	struct evsel *evsel;
1140	struct machine *machine;
1141	int ret = -1;
1142
1143	if (should_drop(qevent, top)) {
1144		top->drop++;
1145		top->drop_total++;
1146		return 0;
1147	}
1148
1149	ret = evlist__parse_sample(evlist, event, &sample);
1150	if (ret) {
1151		pr_err("Can't parse sample, err = %d\n", ret);
1152		goto next_event;
1153	}
1154
1155	evsel = evlist__id2evsel(session->evlist, sample.id);
1156	assert(evsel != NULL);
1157
1158	if (event->header.type == PERF_RECORD_SAMPLE) {
1159		if (evswitch__discard(&top->evswitch, evsel))
1160			return 0;
1161		++top->samples;
1162	}
1163
1164	switch (sample.cpumode) {
1165	case PERF_RECORD_MISC_USER:
1166		++top->us_samples;
1167		if (top->hide_user_symbols)
1168			goto next_event;
1169		machine = &session->machines.host;
1170		break;
1171	case PERF_RECORD_MISC_KERNEL:
1172		++top->kernel_samples;
1173		if (top->hide_kernel_symbols)
1174			goto next_event;
1175		machine = &session->machines.host;
1176		break;
1177	case PERF_RECORD_MISC_GUEST_KERNEL:
1178		++top->guest_kernel_samples;
1179		machine = perf_session__find_machine(session,
1180						     sample.pid);
1181		break;
1182	case PERF_RECORD_MISC_GUEST_USER:
1183		++top->guest_us_samples;
1184		/*
1185		 * TODO: we don't process guest user from host side
1186		 * except simple counting.
1187		 */
1188		goto next_event;
1189	default:
1190		if (event->header.type == PERF_RECORD_SAMPLE)
1191			goto next_event;
1192		machine = &session->machines.host;
1193		break;
1194	}
1195
1196	if (event->header.type == PERF_RECORD_SAMPLE) {
1197		perf_event__process_sample(&top->tool, event, evsel,
1198					   &sample, machine);
1199	} else if (event->header.type == PERF_RECORD_LOST) {
1200		perf_top__process_lost(top, event, evsel);
1201	} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1202		perf_top__process_lost_samples(top, event, evsel);
1203	} else if (event->header.type < PERF_RECORD_MAX) {
1204		events_stats__inc(&session->evlist->stats, event->header.type);
1205		machine__process_event(machine, event, &sample);
1206	} else
1207		++session->evlist->stats.nr_unknown_events;
1208
1209	ret = 0;
1210next_event:
1211	return ret;
1212}
1213
1214static void init_process_thread(struct perf_top *top)
1215{
1216	ordered_events__init(&top->qe.data[0], deliver_event, top);
1217	ordered_events__init(&top->qe.data[1], deliver_event, top);
1218	ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1219	ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1220	top->qe.in = &top->qe.data[0];
1221	mutex_init(&top->qe.mutex);
1222	cond_init(&top->qe.cond);
1223}
1224
1225static int __cmd_top(struct perf_top *top)
1226{
1227	struct record_opts *opts = &top->record_opts;
1228	pthread_t thread, thread_process;
1229	int ret;
1230
1231	if (!top->annotation_opts.objdump_path) {
1232		ret = perf_env__lookup_objdump(&top->session->header.env,
1233					       &top->annotation_opts.objdump_path);
1234		if (ret)
1235			return ret;
1236	}
1237
1238	ret = callchain_param__setup_sample_type(&callchain_param);
1239	if (ret)
1240		return ret;
1241
1242	if (perf_session__register_idle_thread(top->session) < 0)
1243		return ret;
1244
1245	if (top->nr_threads_synthesize > 1)
1246		perf_set_multithreaded();
1247
1248	init_process_thread(top);
1249
1250	if (opts->record_namespaces)
1251		top->tool.namespace_events = true;
1252	if (opts->record_cgroup) {
1253#ifdef HAVE_FILE_HANDLE
1254		top->tool.cgroup_events = true;
1255#else
1256		pr_err("cgroup tracking is not supported.\n");
1257		return -1;
1258#endif
1259	}
1260
1261	ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1262						&top->session->machines.host,
1263						&top->record_opts);
1264	if (ret < 0)
1265		pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1266
1267	ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1268					     &top->session->machines.host);
1269	if (ret < 0)
1270		pr_debug("Couldn't synthesize cgroup events.\n");
1271
1272	machine__synthesize_threads(&top->session->machines.host, &opts->target,
1273				    top->evlist->core.threads, true, false,
1274				    top->nr_threads_synthesize);
1275
1276	if (top->nr_threads_synthesize > 1)
1277		perf_set_singlethreaded();
1278
1279	if (perf_hpp_list.socket) {
1280		ret = perf_env__read_cpu_topology_map(&perf_env);
1281		if (ret < 0) {
1282			char errbuf[BUFSIZ];
1283			const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1284
1285			ui__error("Could not read the CPU topology map: %s\n", err);
1286			return ret;
1287		}
1288	}
1289
1290	ret = perf_top__start_counters(top);
1291	if (ret)
1292		return ret;
1293
1294	top->session->evlist = top->evlist;
1295	perf_session__set_id_hdr_size(top->session);
1296
1297	/*
1298	 * When perf is starting the traced process, all the events (apart from
1299	 * group members) have enable_on_exec=1 set, so don't spoil it by
1300	 * prematurely enabling them.
1301	 *
1302	 * XXX 'top' still doesn't start workloads like record, trace, but should,
1303	 * so leave the check here.
1304	 */
1305        if (!target__none(&opts->target))
1306		evlist__enable(top->evlist);
1307
1308	ret = -1;
1309	if (pthread_create(&thread_process, NULL, process_thread, top)) {
1310		ui__error("Could not create process thread.\n");
1311		return ret;
1312	}
1313
1314	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1315							    display_thread), top)) {
1316		ui__error("Could not create display thread.\n");
1317		goto out_join_thread;
1318	}
1319
1320	if (top->realtime_prio) {
1321		struct sched_param param;
1322
1323		param.sched_priority = top->realtime_prio;
1324		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1325			ui__error("Could not set realtime priority.\n");
1326			goto out_join;
1327		}
1328	}
1329
1330	/* Wait for a minimal set of events before starting the snapshot */
1331	evlist__poll(top->evlist, 100);
1332
1333	perf_top__mmap_read(top);
1334
1335	while (!done) {
1336		u64 hits = top->samples;
1337
1338		perf_top__mmap_read(top);
1339
1340		if (opts->overwrite || (hits == top->samples))
1341			ret = evlist__poll(top->evlist, 100);
1342
1343		if (resize) {
1344			perf_top__resize(top);
1345			resize = 0;
1346		}
1347	}
1348
1349	ret = 0;
1350out_join:
1351	pthread_join(thread, NULL);
1352out_join_thread:
1353	cond_signal(&top->qe.cond);
1354	pthread_join(thread_process, NULL);
1355	return ret;
1356}
1357
1358static int
1359callchain_opt(const struct option *opt, const char *arg, int unset)
1360{
1361	symbol_conf.use_callchain = true;
1362	return record_callchain_opt(opt, arg, unset);
1363}
1364
1365static int
1366parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1367{
1368	struct callchain_param *callchain = opt->value;
1369
1370	callchain->enabled = !unset;
1371	callchain->record_mode = CALLCHAIN_FP;
1372
1373	/*
1374	 * --no-call-graph
1375	 */
1376	if (unset) {
1377		symbol_conf.use_callchain = false;
1378		callchain->record_mode = CALLCHAIN_NONE;
1379		return 0;
1380	}
1381
1382	return parse_callchain_top_opt(arg);
1383}
1384
1385static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1386{
1387	if (!strcmp(var, "top.call-graph")) {
1388		var = "call-graph.record-mode";
1389		return perf_default_config(var, value, cb);
1390	}
1391	if (!strcmp(var, "top.children")) {
1392		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1393		return 0;
1394	}
1395
1396	return 0;
1397}
1398
1399static int
1400parse_percent_limit(const struct option *opt, const char *arg,
1401		    int unset __maybe_unused)
1402{
1403	struct perf_top *top = opt->value;
1404
1405	top->min_percent = strtof(arg, NULL);
1406	return 0;
1407}
1408
1409const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1410	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1411
1412int cmd_top(int argc, const char **argv)
1413{
1414	char errbuf[BUFSIZ];
1415	struct perf_top top = {
1416		.count_filter	     = 5,
1417		.delay_secs	     = 2,
1418		.record_opts = {
1419			.mmap_pages	= UINT_MAX,
1420			.user_freq	= UINT_MAX,
1421			.user_interval	= ULLONG_MAX,
1422			.freq		= 4000, /* 4 KHz */
1423			.target		= {
1424				.uses_mmap   = true,
1425			},
1426			/*
1427			 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1428			 * when we pause, fix that and reenable. Probably using a
1429			 * separate evlist with a dummy event, i.e. a non-overwrite
1430			 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1431			 * stays in overwrite mode. -acme
1432			 * */
1433			.overwrite	= 0,
1434			.sample_time	= true,
1435			.sample_time_set = true,
1436		},
1437		.max_stack	     = sysctl__max_stack(),
1438		.annotation_opts     = annotation__default_options,
1439		.nr_threads_synthesize = UINT_MAX,
1440	};
1441	struct record_opts *opts = &top.record_opts;
1442	struct target *target = &opts->target;
1443	const struct option options[] = {
1444	OPT_CALLBACK('e', "event", &top.evlist, "event",
1445		     "event selector. use 'perf list' to list available events",
1446		     parse_events_option),
1447	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1448	OPT_STRING('p', "pid", &target->pid, "pid",
1449		    "profile events on existing process id"),
1450	OPT_STRING('t', "tid", &target->tid, "tid",
1451		    "profile events on existing thread id"),
1452	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1453			    "system-wide collection from all CPUs"),
1454	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1455		    "list of cpus to monitor"),
1456	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1457		   "file", "vmlinux pathname"),
1458	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1459		    "don't load vmlinux even if found"),
1460	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1461		   "file", "kallsyms pathname"),
1462	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1463		    "hide kernel symbols"),
1464	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1465		     "number of mmap data pages", evlist__parse_mmap_pages),
1466	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1467		    "collect data with this RT SCHED_FIFO priority"),
1468	OPT_INTEGER('d', "delay", &top.delay_secs,
1469		    "number of seconds to delay between refreshes"),
1470	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1471			    "dump the symbol table used for profiling"),
1472	OPT_INTEGER('f', "count-filter", &top.count_filter,
1473		    "only display functions with more events than this"),
 
 
1474	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1475		    "child tasks do not inherit counters"),
1476	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1477		    "symbol to annotate"),
1478	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1479	OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1480		     "profile at this frequency",
1481		      record__parse_freq),
1482	OPT_INTEGER('E', "entries", &top.print_entries,
1483		    "display this many functions"),
1484	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1485		    "hide user symbols"),
1486#ifdef HAVE_SLANG_SUPPORT
1487	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1488#endif
1489	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1490	OPT_INCR('v', "verbose", &verbose,
1491		    "be more verbose (show counter open errors, etc)"),
1492	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1493		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1494		   " Please refer the man page for the complete list."),
1495	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1496		   "output field(s): overhead, period, sample plus all of sort keys"),
1497	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1498		    "Show a column with the number of samples"),
1499	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1500			   NULL, "enables call-graph recording and display",
1501			   &callchain_opt),
1502	OPT_CALLBACK(0, "call-graph", &callchain_param,
1503		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1504		     top_callchain_help, &parse_callchain_opt),
1505	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1506		    "Accumulate callchains of children and show total overhead as well"),
1507	OPT_INTEGER(0, "max-stack", &top.max_stack,
1508		    "Set the maximum stack depth when parsing the callchain. "
1509		    "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1510	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1511		   "ignore callees of these functions in call graphs",
1512		   report_parse_ignore_callees_opt),
1513	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1514		    "Show a column with the sum of periods"),
1515	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1516		   "only consider symbols in these dsos"),
1517	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1518		   "only consider symbols in these comms"),
1519	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1520		   "only consider these symbols"),
1521	OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1522		    "Interleave source code with assembly code (default)"),
1523	OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1524		    "Display raw encoding of assembly instructions (default)"),
1525	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1526		    "Enable kernel symbol demangling"),
1527	OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1528	OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1529		    "objdump binary to use for disassembly and annotations"),
1530	OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1531		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1532	OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1533		    "Add prefix to source file path names in programs (with --prefix-strip)"),
1534	OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1535		    "Strip first N entries of source file path name in programs (with --prefix)"),
1536	OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1537	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1538		     "Don't show entries under that percent", parse_percent_limit),
1539	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1540		     "How to display percentage of filtered entries", parse_filter_percentage),
1541	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1542		   "width[,width...]",
1543		   "don't try to adjust column width, use these fixed values"),
1544	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1545			"per thread proc mmap processing timeout in ms"),
1546	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1547		     "branch any", "sample any taken branches",
1548		     parse_branch_stack),
1549	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1550		     "branch filter mask", "branch stack filter modes",
1551		     parse_branch_stack),
1552	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1553		    "Show raw trace event output (do not use print fmt or plugins)"),
1554	OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1555		    "Show entries in a hierarchy"),
1556	OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1557		    "Use a backward ring buffer, default: no"),
1558	OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1559	OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1560			"number of thread to run event synthesize"),
1561	OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1562		     "monitor event in cgroup name only", parse_cgroups),
1563	OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1564		    "Record namespaces events"),
1565	OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1566		    "Record cgroup events"),
1567	OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1568		    "Sort the output by the event at the index n in group. "
1569		    "If n is invalid, sort by the first event. "
1570		    "WARNING: should be used on grouped events."),
1571	OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1572		    "Enable LBR callgraph stitching approach"),
1573#ifdef HAVE_LIBPFM
1574	OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1575		"libpfm4 event selector. use 'perf list' to list available events",
1576		parse_libpfm_events_option),
1577#endif
1578	OPTS_EVSWITCH(&top.evswitch),
1579	OPT_END()
1580	};
1581	const char * const top_usage[] = {
1582		"perf top [<options>]",
1583		NULL
1584	};
1585	int status = hists__init();
1586
1587	if (status < 0)
1588		return status;
1589
1590	top.annotation_opts.min_pcnt = 5;
1591	top.annotation_opts.context  = 4;
1592
1593	top.evlist = evlist__new();
1594	if (top.evlist == NULL)
1595		return -ENOMEM;
1596
1597	status = perf_config(perf_top_config, &top);
1598	if (status)
1599		return status;
1600	/*
1601	 * Since the per arch annotation init routine may need the cpuid, read
1602	 * it here, since we are not getting this from the perf.data header.
1603	 */
1604	status = perf_env__read_cpuid(&perf_env);
1605	if (status) {
1606		/*
1607		 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1608		 * warn the user explicitly.
1609		 */
1610		eprintf(status == ENOSYS ? 1 : 0, verbose,
1611			"Couldn't read the cpuid for this machine: %s\n",
1612			str_error_r(errno, errbuf, sizeof(errbuf)));
1613	}
1614	top.evlist->env = &perf_env;
1615
1616	argc = parse_options(argc, argv, options, top_usage, 0);
1617	if (argc)
1618		usage_with_options(top_usage, options);
1619
1620	status = symbol__validate_sym_arguments();
1621	if (status)
1622		goto out_delete_evlist;
1623
1624	if (annotate_check_args(&top.annotation_opts) < 0)
1625		goto out_delete_evlist;
1626
1627	if (!top.evlist->core.nr_entries &&
1628	    evlist__add_default(top.evlist) < 0) {
1629		pr_err("Not enough memory for event selector list\n");
1630		goto out_delete_evlist;
1631	}
1632
1633	status = evswitch__init(&top.evswitch, top.evlist, stderr);
1634	if (status)
1635		goto out_delete_evlist;
1636
1637	if (symbol_conf.report_hierarchy) {
1638		/* disable incompatible options */
1639		symbol_conf.event_group = false;
1640		symbol_conf.cumulate_callchain = false;
1641
1642		if (field_order) {
1643			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1644			parse_options_usage(top_usage, options, "fields", 0);
1645			parse_options_usage(NULL, options, "hierarchy", 0);
1646			goto out_delete_evlist;
1647		}
1648	}
1649
1650	if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1651		pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1652		goto out_delete_evlist;
1653	}
1654
1655	if (nr_cgroups > 0 && opts->record_cgroup) {
1656		pr_err("--cgroup and --all-cgroups cannot be used together\n");
1657		goto out_delete_evlist;
1658	}
1659
1660	if (opts->branch_stack && callchain_param.enabled)
1661		symbol_conf.show_branchflag_count = true;
1662
1663	sort__mode = SORT_MODE__TOP;
1664	/* display thread wants entries to be collapsed in a different tree */
1665	perf_hpp_list.need_collapse = 1;
1666
1667	if (top.use_stdio)
1668		use_browser = 0;
1669#ifdef HAVE_SLANG_SUPPORT
1670	else if (top.use_tui)
1671		use_browser = 1;
1672#endif
1673
1674	setup_browser(false);
1675
1676	if (setup_sorting(top.evlist) < 0) {
1677		if (sort_order)
1678			parse_options_usage(top_usage, options, "s", 1);
1679		if (field_order)
1680			parse_options_usage(sort_order ? NULL : top_usage,
1681					    options, "fields", 0);
1682		goto out_delete_evlist;
1683	}
1684
1685	status = target__validate(target);
1686	if (status) {
1687		target__strerror(target, status, errbuf, BUFSIZ);
1688		ui__warning("%s\n", errbuf);
1689	}
1690
1691	status = target__parse_uid(target);
1692	if (status) {
1693		int saved_errno = errno;
1694
1695		target__strerror(target, status, errbuf, BUFSIZ);
1696		ui__error("%s\n", errbuf);
1697
1698		status = -saved_errno;
1699		goto out_delete_evlist;
1700	}
1701
1702	if (target__none(target))
1703		target->system_wide = true;
1704
1705	if (evlist__create_maps(top.evlist, target) < 0) {
1706		ui__error("Couldn't create thread/CPU maps: %s\n",
1707			  errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1708		status = -errno;
1709		goto out_delete_evlist;
1710	}
1711
1712	if (top.delay_secs < 1)
1713		top.delay_secs = 1;
1714
1715	if (record_opts__config(opts)) {
1716		status = -EINVAL;
1717		goto out_delete_evlist;
1718	}
1719
1720	top.sym_evsel = evlist__first(top.evlist);
1721
1722	if (!callchain_param.enabled) {
1723		symbol_conf.cumulate_callchain = false;
1724		perf_hpp__cancel_cumulate();
1725	}
1726
1727	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1728		callchain_param.order = ORDER_CALLER;
1729
1730	status = symbol__annotation_init();
1731	if (status < 0)
1732		goto out_delete_evlist;
1733
1734	annotation_config__init(&top.annotation_opts);
1735
1736	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1737	status = symbol__init(NULL);
1738	if (status < 0)
1739		goto out_delete_evlist;
1740
1741	sort__setup_elide(stdout);
1742
1743	get_term_dimensions(&top.winsize);
1744	if (top.print_entries == 0) {
1745		perf_top__update_print_entries(&top);
1746		signal(SIGWINCH, winch_sig);
1747	}
1748
1749	top.session = perf_session__new(NULL, NULL);
1750	if (IS_ERR(top.session)) {
1751		status = PTR_ERR(top.session);
1752		goto out_delete_evlist;
1753	}
1754
1755#ifdef HAVE_LIBBPF_SUPPORT
1756	if (!top.record_opts.no_bpf_event) {
1757		top.sb_evlist = evlist__new();
1758
1759		if (top.sb_evlist == NULL) {
1760			pr_err("Couldn't create side band evlist.\n.");
1761			status = -EINVAL;
1762			goto out_delete_evlist;
1763		}
1764
1765		if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1766			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1767			status = -EINVAL;
1768			goto out_delete_evlist;
1769		}
1770	}
1771#endif
1772
1773	if (evlist__start_sb_thread(top.sb_evlist, target)) {
1774		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1775		opts->no_bpf_event = true;
1776	}
1777
1778	status = __cmd_top(&top);
1779
1780	if (!opts->no_bpf_event)
1781		evlist__stop_sb_thread(top.sb_evlist);
1782
1783out_delete_evlist:
1784	evlist__delete(top.evlist);
1785	perf_session__delete(top.session);
1786
1787	return status;
1788}