builtin-top.c

   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/stat.h"
  39#include "util/symbol.h"
  40#include "util/synthetic-events.h"
  41#include "util/top.h"
  42#include "util/util.h"
  43#include <linux/rbtree.h>
  44#include <subcmd/parse-options.h>
  45#include "util/parse-events.h"
  46#include "util/callchain.h"
  47#include "util/cpumap.h"
  48#include "util/sort.h"
  49#include "util/string2.h"
  50#include "util/term.h"
  51#include "util/intlist.h"
  52#include "util/parse-branch-options.h"
  53#include "arch/common.h"
  54#include "ui/ui.h"
  55
  56#include "util/debug.h"
  57#include "util/ordered-events.h"
  58#include "util/pfm.h"
  59
  60#include <assert.h>
  61#include <elf.h>
  62#include <fcntl.h>
  63
  64#include <stdio.h>
  65#include <termios.h>
  66#include <unistd.h>
  67#include <inttypes.h>
  68
  69#include <errno.h>
  70#include <time.h>
  71#include <sched.h>
  72#include <signal.h>
  73
  74#include <sys/syscall.h>
  75#include <sys/ioctl.h>
  76#include <poll.h>
  77#include <sys/prctl.h>
  78#include <sys/wait.h>
  79#include <sys/uio.h>
  80#include <sys/utsname.h>
  81#include <sys/mman.h>
  82
  83#include <linux/stringify.h>
  84#include <linux/time64.h>
  85#include <linux/types.h>
  86#include <linux/err.h>
  87
  88#include <linux/ctype.h>
  89#include <perf/mmap.h>
  90
  91static volatile sig_atomic_t done;
  92static volatile sig_atomic_t resize;
  93
  94#define HEADER_LINE_NR  5
  95
  96static void perf_top__update_print_entries(struct perf_top *top)
  97{
  98	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  99}
 100
 101static void winch_sig(int sig __maybe_unused)
 102{
 103	resize = 1;
 104}
 105
 106static void perf_top__resize(struct perf_top *top)
 107{
 108	get_term_dimensions(&top->winsize);
 109	perf_top__update_print_entries(top);
 110}
 111
 112static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 113{
 114	struct evsel *evsel;
 115	struct symbol *sym;
 116	struct annotation *notes;
 117	struct map *map;
 118	struct dso *dso;
 119	int err = -1;
 120
 121	if (!he || !he->ms.sym)
 122		return -1;
 123
 124	evsel = hists_to_evsel(he->hists);
 125
 126	sym = he->ms.sym;
 127	map = he->ms.map;
 128	dso = map__dso(map);
 129
 130	/*
 131	 * We can't annotate with just /proc/kallsyms
 132	 */
 133	if (dso__symtab_type(dso) == DSO_BINARY_TYPE__KALLSYMS && !dso__is_kcore(dso)) {
 134		pr_err("Can't annotate %s: No vmlinux file was found in the "
 135		       "path\n", sym->name);
 136		sleep(1);
 137		return -1;
 138	}
 139
 140	notes = symbol__annotation(sym);
 141	annotation__lock(notes);
 142
 143	if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 144		annotation__unlock(notes);
 145		pr_err("Not enough memory for annotating '%s' symbol!\n",
 146		       sym->name);
 147		sleep(1);
 148		return err;
 149	}
 150
 151	err = symbol__annotate(&he->ms, evsel, NULL);
 152	if (err == 0) {
 153		top->sym_filter_entry = he;
 154	} else {
 155		char msg[BUFSIZ];
 156		symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
 157		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 158	}
 159
 160	annotation__unlock(notes);
 161	return err;
 162}
 163
 164static void __zero_source_counters(struct hist_entry *he)
 165{
 166	struct symbol *sym = he->ms.sym;
 167	symbol__annotate_zero_histograms(sym);
 168}
 169
 170static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 171{
 172	struct utsname uts;
 173	int err = uname(&uts);
 174	struct dso *dso = map__dso(map);
 175
 176	ui__warning("Out of bounds address found:\n\n"
 177		    "Addr:   %" PRIx64 "\n"
 178		    "DSO:    %s %c\n"
 179		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
 180		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 181		    "Arch:   %s\n"
 182		    "Kernel: %s\n"
 183		    "Tools:  %s\n\n"
 184		    "Not all samples will be on the annotation output.\n\n"
 185		    "Please report to linux-kernel@vger.kernel.org\n",
 186		    ip, dso__long_name(dso), dso__symtab_origin(dso),
 187		    map__start(map), map__end(map), sym->start, sym->end,
 188		    sym->binding == STB_GLOBAL ? 'g' :
 189		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 190		    err ? "[unknown]" : uts.machine,
 191		    err ? "[unknown]" : uts.release, perf_version_string);
 192	if (use_browser <= 0)
 193		sleep(5);
 194
 195	map__set_erange_warned(map);
 196}
 197
 198static void perf_top__record_precise_ip(struct perf_top *top,
 199					struct hist_entry *he,
 200					struct perf_sample *sample,
 201					struct evsel *evsel, u64 ip)
 202	EXCLUSIVE_LOCKS_REQUIRED(he->hists->lock)
 203{
 204	struct annotation *notes;
 205	struct symbol *sym = he->ms.sym;
 206	int err = 0;
 207
 208	if (sym == NULL || (use_browser == 0 &&
 209			    (top->sym_filter_entry == NULL ||
 210			     top->sym_filter_entry->ms.sym != sym)))
 211		return;
 212
 213	notes = symbol__annotation(sym);
 214
 215	if (!annotation__trylock(notes))
 216		return;
 217
 218	err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 219
 220	annotation__unlock(notes);
 221
 222	if (unlikely(err)) {
 223		/*
 224		 * This function is now called with he->hists->lock held.
 225		 * Release it before going to sleep.
 226		 */
 227		mutex_unlock(&he->hists->lock);
 228
 229		if (err == -ERANGE && !map__erange_warned(he->ms.map))
 230			ui__warn_map_erange(he->ms.map, sym, ip);
 231		else if (err == -ENOMEM) {
 232			pr_err("Not enough memory for annotating '%s' symbol!\n",
 233			       sym->name);
 234			sleep(1);
 235		}
 236
 237		mutex_lock(&he->hists->lock);
 238	}
 239}
 240
 241static void perf_top__show_details(struct perf_top *top)
 242{
 243	struct hist_entry *he = top->sym_filter_entry;
 244	struct evsel *evsel;
 245	struct annotation *notes;
 246	struct symbol *symbol;
 247	int more;
 248
 249	if (!he)
 250		return;
 251
 252	evsel = hists_to_evsel(he->hists);
 253
 254	symbol = he->ms.sym;
 255	notes = symbol__annotation(symbol);
 256
 257	annotation__lock(notes);
 258
 259	symbol__calc_percent(symbol, evsel);
 260
 261	if (notes->src == NULL)
 262		goto out_unlock;
 263
 264	printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
 265	printf("  Events  Pcnt (>=%d%%)\n", annotate_opts.min_pcnt);
 266
 267	more = hist_entry__annotate_printf(he, top->sym_evsel);
 268
 269	if (top->evlist->enabled) {
 270		if (top->zero)
 271			symbol__annotate_zero_histogram(symbol, top->sym_evsel);
 272		else
 273			symbol__annotate_decay_histogram(symbol, top->sym_evsel);
 274	}
 275	if (more != 0)
 276		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 277out_unlock:
 278	annotation__unlock(notes);
 279}
 280
 281static void perf_top__resort_hists(struct perf_top *t)
 282{
 283	struct evlist *evlist = t->evlist;
 284	struct evsel *pos;
 285
 286	evlist__for_each_entry(evlist, pos) {
 287		struct hists *hists = evsel__hists(pos);
 288
 289		/*
 290		 * unlink existing entries so that they can be linked
 291		 * in a correct order in hists__match() below.
 292		 */
 293		hists__unlink(hists);
 294
 295		if (evlist->enabled) {
 296			if (t->zero) {
 297				hists__delete_entries(hists);
 298			} else {
 299				hists__decay_entries(hists, t->hide_user_symbols,
 300						     t->hide_kernel_symbols);
 301			}
 302		}
 303
 304		hists__collapse_resort(hists, NULL);
 305
 306		/* Non-group events are considered as leader */
 307		if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
 308			struct hists *leader_hists = evsel__hists(evsel__leader(pos));
 309
 310			hists__match(leader_hists, hists);
 311			hists__link(leader_hists, hists);
 312		}
 313	}
 314
 315	evlist__for_each_entry(evlist, pos) {
 316		evsel__output_resort(pos, NULL);
 317	}
 318}
 319
 320static void perf_top__print_sym_table(struct perf_top *top)
 321{
 322	char bf[160];
 323	int printed = 0;
 324	const int win_width = top->winsize.ws_col - 1;
 325	struct evsel *evsel = top->sym_evsel;
 326	struct hists *hists = evsel__hists(evsel);
 327
 328	puts(CONSOLE_CLEAR);
 329
 330	perf_top__header_snprintf(top, bf, sizeof(bf));
 331	printf("%s\n", bf);
 332
 333	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 334
 335	if (!top->record_opts.overwrite &&
 336	    (top->evlist->stats.nr_lost_warned !=
 337	     top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
 338		top->evlist->stats.nr_lost_warned =
 339			      top->evlist->stats.nr_events[PERF_RECORD_LOST];
 340		color_fprintf(stdout, PERF_COLOR_RED,
 341			      "WARNING: LOST %d chunks, Check IO/CPU overload",
 342			      top->evlist->stats.nr_lost_warned);
 343		++printed;
 344	}
 345
 346	if (top->sym_filter_entry) {
 347		perf_top__show_details(top);
 348		return;
 349	}
 350
 351	perf_top__resort_hists(top);
 352
 353	hists__output_recalc_col_len(hists, top->print_entries - printed);
 354	putchar('\n');
 355	hists__fprintf(hists, false, top->print_entries - printed, win_width,
 356		       top->min_percent, stdout, !symbol_conf.use_callchain);
 357}
 358
 359static void prompt_integer(int *target, const char *msg)
 360{
 361	char *buf = NULL, *p;
 362	size_t dummy = 0;
 363	int tmp;
 364
 365	fprintf(stdout, "\n%s: ", msg);
 366	if (getline(&buf, &dummy, stdin) < 0)
 367		return;
 368
 369	p = strchr(buf, '\n');
 370	if (p)
 371		*p = 0;
 372
 373	p = buf;
 374	while(*p) {
 375		if (!isdigit(*p))
 376			goto out_free;
 377		p++;
 378	}
 379	tmp = strtoul(buf, NULL, 10);
 380	*target = tmp;
 381out_free:
 382	free(buf);
 383}
 384
 385static void prompt_percent(int *target, const char *msg)
 386{
 387	int tmp = 0;
 388
 389	prompt_integer(&tmp, msg);
 390	if (tmp >= 0 && tmp <= 100)
 391		*target = tmp;
 392}
 393
 394static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 395{
 396	char *buf = NULL, *p;
 397	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 398	struct hists *hists = evsel__hists(top->sym_evsel);
 399	struct rb_node *next;
 400	size_t dummy = 0;
 401
 402	/* zero counters of active symbol */
 403	if (syme) {
 404		__zero_source_counters(syme);
 405		top->sym_filter_entry = NULL;
 406	}
 407
 408	fprintf(stdout, "\n%s: ", msg);
 409	if (getline(&buf, &dummy, stdin) < 0)
 410		goto out_free;
 411
 412	p = strchr(buf, '\n');
 413	if (p)
 414		*p = 0;
 415
 416	next = rb_first_cached(&hists->entries);
 417	while (next) {
 418		n = rb_entry(next, struct hist_entry, rb_node);
 419		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 420			found = n;
 421			break;
 422		}
 423		next = rb_next(&n->rb_node);
 424	}
 425
 426	if (!found) {
 427		fprintf(stderr, "Sorry, %s is not active.\n", buf);
 428		sleep(1);
 429	} else
 430		perf_top__parse_source(top, found);
 431
 432out_free:
 433	free(buf);
 434}
 435
 436static void perf_top__print_mapped_keys(struct perf_top *top)
 437{
 438	char *name = NULL;
 439
 440	if (top->sym_filter_entry) {
 441		struct symbol *sym = top->sym_filter_entry->ms.sym;
 442		name = sym->name;
 443	}
 444
 445	fprintf(stdout, "\nMapped keys:\n");
 446	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 447	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 448
 449	if (top->evlist->core.nr_entries > 1)
 450		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
 451
 452	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 453
 454	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", annotate_opts.min_pcnt);
 455	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 456	fprintf(stdout, "\t[S]     stop annotation.\n");
 457
 458	fprintf(stdout,
 459		"\t[K]     hide kernel symbols.             \t(%s)\n",
 460		top->hide_kernel_symbols ? "yes" : "no");
 461	fprintf(stdout,
 462		"\t[U]     hide user symbols.               \t(%s)\n",
 463		top->hide_user_symbols ? "yes" : "no");
 464	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 465	fprintf(stdout, "\t[qQ]    quit.\n");
 466}
 467
 468static int perf_top__key_mapped(struct perf_top *top, int c)
 469{
 470	switch (c) {
 471		case 'd':
 472		case 'e':
 473		case 'f':
 474		case 'z':
 475		case 'q':
 476		case 'Q':
 477		case 'K':
 478		case 'U':
 479		case 'F':
 480		case 's':
 481		case 'S':
 482			return 1;
 483		case 'E':
 484			return top->evlist->core.nr_entries > 1 ? 1 : 0;
 485		default:
 486			break;
 487	}
 488
 489	return 0;
 490}
 491
 492static bool perf_top__handle_keypress(struct perf_top *top, int c)
 493{
 494	bool ret = true;
 495
 496	if (!perf_top__key_mapped(top, c)) {
 497		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 498		struct termios save;
 499
 500		perf_top__print_mapped_keys(top);
 501		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 502		fflush(stdout);
 503
 504		set_term_quiet_input(&save);
 505
 506		poll(&stdin_poll, 1, -1);
 507		c = getc(stdin);
 508
 509		tcsetattr(0, TCSAFLUSH, &save);
 510		if (!perf_top__key_mapped(top, c))
 511			return ret;
 512	}
 513
 514	switch (c) {
 515		case 'd':
 516			prompt_integer(&top->delay_secs, "Enter display delay");
 517			if (top->delay_secs < 1)
 518				top->delay_secs = 1;
 519			break;
 520		case 'e':
 521			prompt_integer(&top->print_entries, "Enter display entries (lines)");
 522			if (top->print_entries == 0) {
 523				perf_top__resize(top);
 524				signal(SIGWINCH, winch_sig);
 525			} else {
 526				signal(SIGWINCH, SIG_DFL);
 527			}
 528			break;
 529		case 'E':
 530			if (top->evlist->core.nr_entries > 1) {
 531				/* Select 0 as the default event: */
 532				int counter = 0;
 533
 534				fprintf(stderr, "\nAvailable events:");
 535
 536				evlist__for_each_entry(top->evlist, top->sym_evsel)
 537					fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));
 538
 539				prompt_integer(&counter, "Enter details event counter");
 540
 541				if (counter >= top->evlist->core.nr_entries) {
 542					top->sym_evsel = evlist__first(top->evlist);
 543					fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
 544					sleep(1);
 545					break;
 546				}
 547				evlist__for_each_entry(top->evlist, top->sym_evsel)
 548					if (top->sym_evsel->core.idx == counter)
 549						break;
 550			} else
 551				top->sym_evsel = evlist__first(top->evlist);
 552			break;
 553		case 'f':
 554			prompt_integer(&top->count_filter, "Enter display event count filter");
 555			break;
 556		case 'F':
 557			prompt_percent(&annotate_opts.min_pcnt,
 558				       "Enter details display event filter (percent)");
 559			break;
 560		case 'K':
 561			top->hide_kernel_symbols = !top->hide_kernel_symbols;
 562			break;
 563		case 'q':
 564		case 'Q':
 565			printf("exiting.\n");
 566			if (top->dump_symtab)
 567				perf_session__fprintf_dsos(top->session, stderr);
 568			ret = false;
 569			break;
 570		case 's':
 571			perf_top__prompt_symbol(top, "Enter details symbol");
 572			break;
 573		case 'S':
 574			if (!top->sym_filter_entry)
 575				break;
 576			else {
 577				struct hist_entry *syme = top->sym_filter_entry;
 578
 579				top->sym_filter_entry = NULL;
 580				__zero_source_counters(syme);
 581			}
 582			break;
 583		case 'U':
 584			top->hide_user_symbols = !top->hide_user_symbols;
 585			break;
 586		case 'z':
 587			top->zero = !top->zero;
 588			break;
 589		default:
 590			break;
 591	}
 592
 593	return ret;
 594}
 595
 596static void perf_top__sort_new_samples(void *arg)
 597{
 598	struct perf_top *t = arg;
 599
 600	if (t->evlist->selected != NULL)
 601		t->sym_evsel = t->evlist->selected;
 602
 603	perf_top__resort_hists(t);
 604
 605	if (t->lost || t->drop)
 606		pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 607}
 608
 609static void stop_top(void)
 610{
 611	session_done = 1;
 612	done = 1;
 613}
 614
 615static void *display_thread_tui(void *arg)
 616{
 617	struct evsel *pos;
 618	struct perf_top *top = arg;
 619	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 620	struct hist_browser_timer hbt = {
 621		.timer		= perf_top__sort_new_samples,
 622		.arg		= top,
 623		.refresh	= top->delay_secs,
 624	};
 625	int ret;
 626
 627	/* In order to read symbols from other namespaces perf to  needs to call
 628	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 629	 * unshare(2) the fs so that we may continue to setns into namespaces
 630	 * that we're observing.
 631	 */
 632	unshare(CLONE_FS);
 633
 634	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 635
 636repeat:
 637	perf_top__sort_new_samples(top);
 638
 639	/*
 640	 * Initialize the uid_filter_str, in the future the TUI will allow
 641	 * Zooming in/out UIDs. For now just use whatever the user passed
 642	 * via --uid.
 643	 */
 644	evlist__for_each_entry(top->evlist, pos) {
 645		struct hists *hists = evsel__hists(pos);
 646		hists->uid_filter_str = top->uid_str;
 647	}
 648
 649	ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 650				       perf_session__env(top->session),
 651				       !top->record_opts.overwrite);
 652	if (ret == K_RELOAD) {
 653		top->zero = true;
 654		goto repeat;
 655	} else
 656		stop_top();
 657
 658	return NULL;
 659}
 660
 661static void display_sig(int sig __maybe_unused)
 662{
 663	stop_top();
 664}
 665
 666static void display_setup_sig(void)
 667{
 668	signal(SIGSEGV, sighandler_dump_stack);
 669	signal(SIGFPE, sighandler_dump_stack);
 670	signal(SIGINT,  display_sig);
 671	signal(SIGQUIT, display_sig);
 672	signal(SIGTERM, display_sig);
 673}
 674
 675static void *display_thread(void *arg)
 676{
 677	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 678	struct termios save;
 679	struct perf_top *top = arg;
 680	int delay_msecs, c;
 681
 682	/* In order to read symbols from other namespaces perf to  needs to call
 683	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 684	 * unshare(2) the fs so that we may continue to setns into namespaces
 685	 * that we're observing.
 686	 */
 687	unshare(CLONE_FS);
 688
 689	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 690
 691	display_setup_sig();
 692	pthread__unblock_sigwinch();
 693repeat:
 694	delay_msecs = top->delay_secs * MSEC_PER_SEC;
 695	set_term_quiet_input(&save);
 696	/* trash return*/
 697	clearerr(stdin);
 698	if (poll(&stdin_poll, 1, 0) > 0)
 699		getc(stdin);
 700
 701	while (!done) {
 702		perf_top__print_sym_table(top);
 703		/*
 704		 * Either timeout expired or we got an EINTR due to SIGWINCH,
 705		 * refresh screen in both cases.
 706		 */
 707		switch (poll(&stdin_poll, 1, delay_msecs)) {
 708		case 0:
 709			continue;
 710		case -1:
 711			if (errno == EINTR)
 712				continue;
 713			fallthrough;
 714		default:
 715			c = getc(stdin);
 716			tcsetattr(0, TCSAFLUSH, &save);
 717
 718			if (perf_top__handle_keypress(top, c))
 719				goto repeat;
 720			stop_top();
 721		}
 722	}
 723
 724	tcsetattr(0, TCSAFLUSH, &save);
 725	return NULL;
 726}
 727
 728static int hist_iter__top_callback(struct hist_entry_iter *iter,
 729				   struct addr_location *al, bool single,
 730				   void *arg)
 731	EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
 732{
 733	struct perf_top *top = arg;
 734	struct evsel *evsel = iter->evsel;
 735
 736	if (perf_hpp_list.sym && single)
 737		perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);
 738
 739	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 740			     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 741			     NULL, evsel);
 742	return 0;
 743}
 744
 745static void perf_event__process_sample(const struct perf_tool *tool,
 746				       const union perf_event *event,
 747				       struct evsel *evsel,
 748				       struct perf_sample *sample,
 749				       struct machine *machine)
 750{
 751	struct perf_top *top = container_of(tool, struct perf_top, tool);
 752	struct addr_location al;
 753
 754	if (!machine && perf_guest) {
 755		static struct intlist *seen;
 756
 757		if (!seen)
 758			seen = intlist__new(NULL);
 759
 760		if (!intlist__has_entry(seen, sample->pid)) {
 761			pr_err("Can't find guest [%d]'s kernel information\n",
 762				sample->pid);
 763			intlist__add(seen, sample->pid);
 764		}
 765		return;
 766	}
 767
 768	if (!machine) {
 769		pr_err("%u unprocessable samples recorded.\r",
 770		       top->session->evlist->stats.nr_unprocessable_samples++);
 771		return;
 772	}
 773
 774	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 775		top->exact_samples++;
 776
 777	addr_location__init(&al);
 778	if (machine__resolve(machine, &al, sample) < 0)
 779		goto out;
 780
 781	if (top->stitch_lbr)
 782		thread__set_lbr_stitch_enable(al.thread, true);
 783
 784	if (!machine->kptr_restrict_warned &&
 785	    symbol_conf.kptr_restrict &&
 786	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
 787		if (!evlist__exclude_kernel(top->session->evlist)) {
 788			ui__warning(
 789"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 790"Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 791"Kernel%s samples will not be resolved.\n",
 792			  al.map && map__has_symbols(al.map) ?
 793			  " modules" : "");
 794			if (use_browser <= 0)
 795				sleep(5);
 796		}
 797		machine->kptr_restrict_warned = true;
 798	}
 799
 800	if (al.sym == NULL && al.map != NULL) {
 801		const char *msg = "Kernel samples will not be resolved.\n";
 802		/*
 803		 * As we do lazy loading of symtabs we only will know if the
 804		 * specified vmlinux file is invalid when we actually have a
 805		 * hit in kernel space and then try to load it. So if we get
 806		 * here and there are _no_ symbols in the DSO backing the
 807		 * kernel map, bail out.
 808		 *
 809		 * We may never get here, for instance, if we use -K/
 810		 * --hide-kernel-symbols, even if the user specifies an
 811		 * invalid --vmlinux ;-)
 812		 */
 813		if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 814		    __map__is_kernel(al.map) && !map__has_symbols(al.map)) {
 815			if (symbol_conf.vmlinux_name) {
 816				char serr[256];
 817
 818				dso__strerror_load(map__dso(al.map), serr, sizeof(serr));
 819				ui__warning("The %s file can't be used: %s\n%s",
 820					    symbol_conf.vmlinux_name, serr, msg);
 821			} else {
 822				ui__warning("A vmlinux file was not found.\n%s",
 823					    msg);
 824			}
 825
 826			if (use_browser <= 0)
 827				sleep(5);
 828			top->vmlinux_warned = true;
 829		}
 830	}
 831
 832	if (al.sym == NULL || !al.sym->idle) {
 833		struct hists *hists = evsel__hists(evsel);
 834		struct hist_entry_iter iter = {
 835			.evsel		= evsel,
 836			.sample 	= sample,
 837			.add_entry_cb 	= hist_iter__top_callback,
 838		};
 839
 840		if (symbol_conf.cumulate_callchain)
 841			iter.ops = &hist_iter_cumulative;
 842		else
 843			iter.ops = &hist_iter_normal;
 844
 845		mutex_lock(&hists->lock);
 846
 847		if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
 848			pr_err("Problem incrementing symbol period, skipping event\n");
 849
 850		mutex_unlock(&hists->lock);
 851	}
 852
 853out:
 854	addr_location__exit(&al);
 855}
 856
 857static void
 858perf_top__process_lost(struct perf_top *top, union perf_event *event,
 859		       struct evsel *evsel)
 860{
 861	top->lost += event->lost.lost;
 862	top->lost_total += event->lost.lost;
 863	evsel->evlist->stats.total_lost += event->lost.lost;
 864}
 865
 866static void
 867perf_top__process_lost_samples(struct perf_top *top,
 868			       union perf_event *event,
 869			       struct evsel *evsel)
 870{
 871	top->lost += event->lost_samples.lost;
 872	top->lost_total += event->lost_samples.lost;
 873	evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
 874}
 875
 876static u64 last_timestamp;
 877
 878static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 879{
 880	struct record_opts *opts = &top->record_opts;
 881	struct evlist *evlist = top->evlist;
 882	struct mmap *md;
 883	union perf_event *event;
 884
 885	md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 886	if (perf_mmap__read_init(&md->core) < 0)
 887		return;
 888
 889	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 890		int ret;
 891
 892		ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 893		if (ret && ret != -1)
 894			break;
 895
 896		ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
 897		if (ret)
 898			break;
 899
 900		perf_mmap__consume(&md->core);
 901
 902		if (top->qe.rotate) {
 903			mutex_lock(&top->qe.mutex);
 904			top->qe.rotate = false;
 905			cond_signal(&top->qe.cond);
 906			mutex_unlock(&top->qe.mutex);
 907		}
 908	}
 909
 910	perf_mmap__read_done(&md->core);
 911}
 912
 913static void perf_top__mmap_read(struct perf_top *top)
 914{
 915	bool overwrite = top->record_opts.overwrite;
 916	struct evlist *evlist = top->evlist;
 917	int i;
 918
 919	if (overwrite)
 920		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 921
 922	for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 923		perf_top__mmap_read_idx(top, i);
 924
 925	if (overwrite) {
 926		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 927		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 928	}
 929}
 930
 931/*
 932 * Check per-event overwrite term.
 933 * perf top should support consistent term for all events.
 934 * - All events don't have per-event term
 935 *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 936 *   Nothing change, return 0.
 937 * - All events have same per-event term
 938 *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 939 *   Using the per-event setting to replace the opts->overwrite if
 940 *   they are different, then return 0.
 941 * - Events have different per-event term
 942 *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 943 *   Return -1
 944 * - Some of the event set per-event term, but some not.
 945 *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 946 *   Return -1
 947 */
 948static int perf_top__overwrite_check(struct perf_top *top)
 949{
 950	struct record_opts *opts = &top->record_opts;
 951	struct evlist *evlist = top->evlist;
 952	struct evsel_config_term *term;
 953	struct list_head *config_terms;
 954	struct evsel *evsel;
 955	int set, overwrite = -1;
 956
 957	evlist__for_each_entry(evlist, evsel) {
 958		set = -1;
 959		config_terms = &evsel->config_terms;
 960		list_for_each_entry(term, config_terms, list) {
 961			if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
 962				set = term->val.overwrite ? 1 : 0;
 963		}
 964
 965		/* no term for current and previous event (likely) */
 966		if ((overwrite < 0) && (set < 0))
 967			continue;
 968
 969		/* has term for both current and previous event, compare */
 970		if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 971			return -1;
 972
 973		/* no term for current event but has term for previous one */
 974		if ((overwrite >= 0) && (set < 0))
 975			return -1;
 976
 977		/* has term for current event */
 978		if ((overwrite < 0) && (set >= 0)) {
 979			/* if it's first event, set overwrite */
 980			if (evsel == evlist__first(evlist))
 981				overwrite = set;
 982			else
 983				return -1;
 984		}
 985	}
 986
 987	if ((overwrite >= 0) && (opts->overwrite != overwrite))
 988		opts->overwrite = overwrite;
 989
 990	return 0;
 991}
 992
 993static int perf_top_overwrite_fallback(struct perf_top *top,
 994				       struct evsel *evsel)
 995{
 996	struct record_opts *opts = &top->record_opts;
 997	struct evlist *evlist = top->evlist;
 998	struct evsel *counter;
 999
1000	if (!opts->overwrite)
1001		return 0;
1002
1003	/* only fall back when first event fails */
1004	if (evsel != evlist__first(evlist))
1005		return 0;
1006
1007	evlist__for_each_entry(evlist, counter)
1008		counter->core.attr.write_backward = false;
1009	opts->overwrite = false;
1010	pr_debug2("fall back to non-overwrite mode\n");
1011	return 1;
1012}
1013
1014static int perf_top__start_counters(struct perf_top *top)
1015{
1016	char msg[BUFSIZ];
1017	struct evsel *counter;
1018	struct evlist *evlist = top->evlist;
1019	struct record_opts *opts = &top->record_opts;
1020
1021	if (perf_top__overwrite_check(top)) {
1022		ui__error("perf top only support consistent per-event "
1023			  "overwrite setting for all events\n");
1024		goto out_err;
1025	}
1026
1027	evlist__config(evlist, opts, &callchain_param);
1028
1029	evlist__for_each_entry(evlist, counter) {
1030try_again:
1031		if (evsel__open(counter, counter->core.cpus,
1032				counter->core.threads) < 0) {
1033
1034			/*
1035			 * Specially handle overwrite fall back.
1036			 * Because perf top is the only tool which has
1037			 * overwrite mode by default, support
1038			 * both overwrite and non-overwrite mode, and
1039			 * require consistent mode for all events.
1040			 *
1041			 * May move it to generic code with more tools
1042			 * have similar attribute.
1043			 */
1044			if (perf_missing_features.write_backward &&
1045			    perf_top_overwrite_fallback(top, counter))
1046				goto try_again;
1047
1048			if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
1049				if (verbose > 0)
1050					ui__warning("%s\n", msg);
1051				goto try_again;
1052			}
1053
1054			evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1055			ui__error("%s\n", msg);
1056			goto out_err;
1057		}
1058	}
1059
1060	if (evlist__apply_filters(evlist, &counter, &opts->target)) {
1061		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
1062			counter->filter ?: "BPF", evsel__name(counter), errno,
1063			str_error_r(errno, msg, sizeof(msg)));
1064		goto out_err;
1065	}
1066
1067	if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1068		ui__error("Failed to mmap with %d (%s)\n",
1069			    errno, str_error_r(errno, msg, sizeof(msg)));
1070		goto out_err;
1071	}
1072
1073	return 0;
1074
1075out_err:
1076	return -1;
1077}
1078
1079static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1080{
1081	if (callchain->mode != CHAIN_NONE) {
1082		if (callchain_register_param(callchain) < 0) {
1083			ui__error("Can't register callchain params.\n");
1084			return -EINVAL;
1085		}
1086	}
1087
1088	return 0;
1089}
1090
1091static struct ordered_events *rotate_queues(struct perf_top *top)
1092{
1093	struct ordered_events *in = top->qe.in;
1094
1095	if (top->qe.in == &top->qe.data[1])
1096		top->qe.in = &top->qe.data[0];
1097	else
1098		top->qe.in = &top->qe.data[1];
1099
1100	return in;
1101}
1102
1103static void *process_thread(void *arg)
1104{
1105	struct perf_top *top = arg;
1106
1107	while (!done) {
1108		struct ordered_events *out, *in = top->qe.in;
1109
1110		if (!in->nr_events) {
1111			usleep(100);
1112			continue;
1113		}
1114
1115		out = rotate_queues(top);
1116
1117		mutex_lock(&top->qe.mutex);
1118		top->qe.rotate = true;
1119		cond_wait(&top->qe.cond, &top->qe.mutex);
1120		mutex_unlock(&top->qe.mutex);
1121
1122		if (ordered_events__flush(out, OE_FLUSH__TOP))
1123			pr_err("failed to process events\n");
1124	}
1125
1126	return NULL;
1127}
1128
1129/*
1130 * Allow only 'top->delay_secs' seconds behind samples.
1131 */
1132static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1133{
1134	union perf_event *event = qevent->event;
1135	u64 delay_timestamp;
1136
1137	if (event->header.type != PERF_RECORD_SAMPLE)
1138		return false;
1139
1140	delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1141	return delay_timestamp < last_timestamp;
1142}
1143
1144static int deliver_event(struct ordered_events *qe,
1145			 struct ordered_event *qevent)
1146{
1147	struct perf_top *top = qe->data;
1148	struct evlist *evlist = top->evlist;
1149	struct perf_session *session = top->session;
1150	union perf_event *event = qevent->event;
1151	struct perf_sample sample;
1152	struct evsel *evsel;
1153	struct machine *machine;
1154	int ret = -1;
1155
1156	if (should_drop(qevent, top)) {
1157		top->drop++;
1158		top->drop_total++;
1159		return 0;
1160	}
1161
1162	perf_sample__init(&sample, /*all=*/false);
1163	ret = evlist__parse_sample(evlist, event, &sample);
1164	if (ret) {
1165		pr_err("Can't parse sample, err = %d\n", ret);
1166		goto next_event;
1167	}
1168
1169	evsel = evlist__id2evsel(session->evlist, sample.id);
1170	assert(evsel != NULL);
1171
1172	if (event->header.type == PERF_RECORD_SAMPLE) {
1173		if (evswitch__discard(&top->evswitch, evsel)) {
1174			ret = 0;
1175			goto next_event;
1176		}
1177		++top->samples;
1178	}
1179
1180	switch (sample.cpumode) {
1181	case PERF_RECORD_MISC_USER:
1182		++top->us_samples;
1183		if (top->hide_user_symbols)
1184			goto next_event;
1185		machine = &session->machines.host;
1186		break;
1187	case PERF_RECORD_MISC_KERNEL:
1188		++top->kernel_samples;
1189		if (top->hide_kernel_symbols)
1190			goto next_event;
1191		machine = &session->machines.host;
1192		break;
1193	case PERF_RECORD_MISC_GUEST_KERNEL:
1194		++top->guest_kernel_samples;
1195		machine = perf_session__find_machine(session,
1196						     sample.pid);
1197		break;
1198	case PERF_RECORD_MISC_GUEST_USER:
1199		++top->guest_us_samples;
1200		/*
1201		 * TODO: we don't process guest user from host side
1202		 * except simple counting.
1203		 */
1204		goto next_event;
1205	default:
1206		if (event->header.type == PERF_RECORD_SAMPLE)
1207			goto next_event;
1208		machine = &session->machines.host;
1209		break;
1210	}
1211
1212	if (event->header.type == PERF_RECORD_SAMPLE) {
1213		perf_event__process_sample(&top->tool, event, evsel,
1214					   &sample, machine);
1215	} else if (event->header.type == PERF_RECORD_LOST) {
1216		perf_top__process_lost(top, event, evsel);
1217	} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1218		perf_top__process_lost_samples(top, event, evsel);
1219	} else if (event->header.type < PERF_RECORD_MAX) {
1220		events_stats__inc(&session->evlist->stats, event->header.type);
1221		machine__process_event(machine, event, &sample);
1222	} else
1223		++session->evlist->stats.nr_unknown_events;
1224
1225	ret = 0;
1226next_event:
1227	perf_sample__exit(&sample);
1228	return ret;
1229}
1230
1231static void init_process_thread(struct perf_top *top)
1232{
1233	ordered_events__init(&top->qe.data[0], deliver_event, top);
1234	ordered_events__init(&top->qe.data[1], deliver_event, top);
1235	ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1236	ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1237	top->qe.in = &top->qe.data[0];
1238	mutex_init(&top->qe.mutex);
1239	cond_init(&top->qe.cond);
1240}
1241
1242static void exit_process_thread(struct perf_top *top)
1243{
1244	ordered_events__free(&top->qe.data[0]);
1245	ordered_events__free(&top->qe.data[1]);
1246	mutex_destroy(&top->qe.mutex);
1247	cond_destroy(&top->qe.cond);
1248}
1249
1250static int __cmd_top(struct perf_top *top)
1251{
1252	struct record_opts *opts = &top->record_opts;
1253	pthread_t thread, thread_process;
1254	int ret;
1255
1256	if (!annotate_opts.objdump_path) {
1257		ret = perf_env__lookup_objdump(perf_session__env(top->session),
1258					       &annotate_opts.objdump_path);
1259		if (ret)
1260			return ret;
1261	}
1262
1263	ret = callchain_param__setup_sample_type(&callchain_param);
1264	if (ret)
1265		return ret;
1266
1267	if (perf_session__register_idle_thread(top->session) < 0)
1268		return ret;
1269
1270	if (top->nr_threads_synthesize > 1)
1271		perf_set_multithreaded();
1272
1273	init_process_thread(top);
1274
1275	if (opts->record_namespaces)
1276		top->tool.namespace_events = true;
1277	if (opts->record_cgroup) {
1278#ifdef HAVE_FILE_HANDLE
1279		top->tool.cgroup_events = true;
1280#else
1281		pr_err("cgroup tracking is not supported.\n");
1282		return -1;
1283#endif
1284	}
1285
1286	ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1287						&top->session->machines.host,
1288						&top->record_opts);
1289	if (ret < 0)
1290		pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1291
1292	ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1293					     &top->session->machines.host);
1294	if (ret < 0)
1295		pr_debug("Couldn't synthesize cgroup events.\n");
1296
1297	machine__synthesize_threads(&top->session->machines.host, &opts->target,
1298				    top->evlist->core.threads, true, false,
1299				    top->nr_threads_synthesize);
1300
1301	perf_set_multithreaded();
1302
1303	if (perf_hpp_list.socket) {
1304		ret = perf_env__read_cpu_topology_map(perf_session__env(top->session));
1305		if (ret < 0) {
1306			char errbuf[BUFSIZ];
1307			const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1308
1309			ui__error("Could not read the CPU topology map: %s\n", err);
1310			return ret;
1311		}
1312	}
1313
1314	/*
1315	 * Use global stat_config that is zero meaning aggr_mode is AGGR_NONE
1316	 * and hybrid_merge is false.
1317	 */
1318	evlist__uniquify_evsel_names(top->evlist, &stat_config);
1319	ret = perf_top__start_counters(top);
1320	if (ret)
1321		return ret;
1322
1323	top->session->evlist = top->evlist;
1324	perf_session__set_id_hdr_size(top->session);
1325
1326	/*
1327	 * When perf is starting the traced process, all the events (apart from
1328	 * group members) have enable_on_exec=1 set, so don't spoil it by
1329	 * prematurely enabling them.
1330	 *
1331	 * XXX 'top' still doesn't start workloads like record, trace, but should,
1332	 * so leave the check here.
1333	 */
1334        if (!target__none(&opts->target))
1335		evlist__enable(top->evlist);
1336
1337	ret = -1;
1338	if (pthread_create(&thread_process, NULL, process_thread, top)) {
1339		ui__error("Could not create process thread.\n");
1340		return ret;
1341	}
1342
1343	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1344							    display_thread), top)) {
1345		ui__error("Could not create display thread.\n");
1346		goto out_join_thread;
1347	}
1348
1349	if (top->realtime_prio) {
1350		struct sched_param param;
1351
1352		param.sched_priority = top->realtime_prio;
1353		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1354			ui__error("Could not set realtime priority.\n");
1355			goto out_join;
1356		}
1357	}
1358
1359	/* Wait for a minimal set of events before starting the snapshot */
1360	evlist__poll(top->evlist, 100);
1361
1362	perf_top__mmap_read(top);
1363
1364	while (!done) {
1365		u64 hits = top->samples;
1366
1367		perf_top__mmap_read(top);
1368
1369		if (opts->overwrite || (hits == top->samples))
1370			ret = evlist__poll(top->evlist, 100);
1371
1372		if (resize) {
1373			perf_top__resize(top);
1374			resize = 0;
1375		}
1376	}
1377
1378	ret = 0;
1379out_join:
1380	pthread_join(thread, NULL);
1381out_join_thread:
1382	cond_signal(&top->qe.cond);
1383	pthread_join(thread_process, NULL);
1384	perf_set_singlethreaded();
1385	exit_process_thread(top);
1386	return ret;
1387}
1388
1389static int
1390callchain_opt(const struct option *opt, const char *arg, int unset)
1391{
1392	symbol_conf.use_callchain = true;
1393	return record_callchain_opt(opt, arg, unset);
1394}
1395
1396static int
1397parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1398{
1399	struct callchain_param *callchain = opt->value;
1400
1401	callchain->enabled = !unset;
1402	callchain->record_mode = CALLCHAIN_FP;
1403
1404	/*
1405	 * --no-call-graph
1406	 */
1407	if (unset) {
1408		symbol_conf.use_callchain = false;
1409		callchain->record_mode = CALLCHAIN_NONE;
1410		return 0;
1411	}
1412
1413	return parse_callchain_top_opt(arg);
1414}
1415
1416static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1417{
1418	if (!strcmp(var, "top.call-graph")) {
1419		var = "call-graph.record-mode";
1420		return perf_default_config(var, value, cb);
1421	}
1422	if (!strcmp(var, "top.children")) {
1423		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1424		return 0;
1425	}
1426
1427	return 0;
1428}
1429
1430static int
1431parse_percent_limit(const struct option *opt, const char *arg,
1432		    int unset __maybe_unused)
1433{
1434	struct perf_top *top = opt->value;
1435
1436	top->min_percent = strtof(arg, NULL);
1437	return 0;
1438}
1439
1440const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1441	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1442
1443int cmd_top(int argc, const char **argv)
1444{
1445	char errbuf[BUFSIZ];
1446	struct perf_top top = {
1447		.count_filter	     = 5,
1448		.delay_secs	     = 2,
1449		.record_opts = {
1450			.mmap_pages	= UINT_MAX,
1451			.user_freq	= UINT_MAX,
1452			.user_interval	= ULLONG_MAX,
1453			.freq		= 4000, /* 4 KHz */
1454			.target		= {
1455				.uses_mmap   = true,
1456			},
1457			/*
1458			 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1459			 * when we pause, fix that and reenable. Probably using a
1460			 * separate evlist with a dummy event, i.e. a non-overwrite
1461			 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1462			 * stays in overwrite mode. -acme
1463			 * */
1464			.overwrite	= 0,
1465			.sample_time	= true,
1466			.sample_time_set = true,
1467		},
1468		.max_stack	     = sysctl__max_stack(),
1469		.nr_threads_synthesize = UINT_MAX,
1470	};
1471	struct parse_events_option_args parse_events_option_args = {
1472		.evlistp = &top.evlist,
1473	};
1474	bool branch_call_mode = false;
1475	struct record_opts *opts = &top.record_opts;
1476	struct target *target = &opts->target;
1477	const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
1478	const struct option options[] = {
1479	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1480		     "event selector. use 'perf list' to list available events",
1481		     parse_events_option),
1482	OPT_CALLBACK(0, "filter", &top.evlist, "filter",
1483		     "event filter", parse_filter),
1484	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1485	OPT_STRING('p', "pid", &target->pid, "pid",
1486		    "profile events on existing process id"),
1487	OPT_STRING('t', "tid", &target->tid, "tid",
1488		    "profile events on existing thread id"),
1489	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1490			    "system-wide collection from all CPUs"),
1491	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1492		    "list of cpus to monitor"),
1493	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1494		   "file", "vmlinux pathname"),
1495	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1496		    "don't load vmlinux even if found"),
1497	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1498		   "file", "kallsyms pathname"),
1499	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1500		    "hide kernel symbols"),
1501	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1502		     "number of mmap data pages", evlist__parse_mmap_pages),
1503	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1504		    "collect data with this RT SCHED_FIFO priority"),
1505	OPT_INTEGER('d', "delay", &top.delay_secs,
1506		    "number of seconds to delay between refreshes"),
1507	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1508			    "dump the symbol table used for profiling"),
1509	OPT_INTEGER('f', "count-filter", &top.count_filter,
1510		    "only display functions with more events than this"),
1511	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1512		    "child tasks do not inherit counters"),
1513	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1514		    "symbol to annotate"),
1515	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1516	OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1517		     "profile at this frequency",
1518		      record__parse_freq),
1519	OPT_INTEGER('E', "entries", &top.print_entries,
1520		    "display this many functions"),
1521	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1522		    "hide user symbols"),
1523#ifdef HAVE_SLANG_SUPPORT
1524	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1525#endif
1526	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1527	OPT_INCR('v', "verbose", &verbose,
1528		    "be more verbose (show counter open errors, etc)"),
1529	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1530		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1531		   " Please refer the man page for the complete list."),
1532	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1533		   "output field(s): overhead, period, sample plus all of sort keys"),
1534	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1535		    "Show a column with the number of samples"),
1536	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1537			   NULL, "enables call-graph recording and display",
1538			   &callchain_opt),
1539	OPT_CALLBACK(0, "call-graph", &callchain_param,
1540		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1541		     top_callchain_help, &parse_callchain_opt),
1542	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1543		    "Accumulate callchains of children and show total overhead as well"),
1544	OPT_INTEGER(0, "max-stack", &top.max_stack,
1545		    "Set the maximum stack depth when parsing the callchain. "
1546		    "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1547	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1548		   "ignore callees of these functions in call graphs",
1549		   report_parse_ignore_callees_opt),
1550	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1551		    "Show a column with the sum of periods"),
1552	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1553		   "only consider symbols in these dsos"),
1554	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1555		   "only consider symbols in these comms"),
1556	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1557		   "only consider these symbols"),
1558	OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
1559		    "Interleave source code with assembly code (default)"),
1560	OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
1561		    "Display raw encoding of assembly instructions (default)"),
1562	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1563		    "Enable kernel symbol demangling"),
1564	OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1565	OPT_STRING(0, "objdump", &objdump_path, "path",
1566		    "objdump binary to use for disassembly and annotations"),
1567	OPT_STRING(0, "addr2line", &addr2line_path, "path",
1568		   "addr2line binary to use for line numbers"),
1569	OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1570		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1571	OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
1572		    "Add prefix to source file path names in programs (with --prefix-strip)"),
1573	OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
1574		    "Strip first N entries of source file path name in programs (with --prefix)"),
1575	OPT_STRING('u', "uid", &top.uid_str, "user", "user to profile"),
1576	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1577		     "Don't show entries under that percent", parse_percent_limit),
1578	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1579		     "How to display percentage of filtered entries", parse_filter_percentage),
1580	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1581		   "width[,width...]",
1582		   "don't try to adjust column width, use these fixed values"),
1583	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1584			"per thread proc mmap processing timeout in ms"),
1585	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1586		     "branch any", "sample any taken branches",
1587		     parse_branch_stack),
1588	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1589		     "branch filter mask", "branch stack filter modes",
1590		     parse_branch_stack),
1591	OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1592		    "add last branch records to call history"),
1593	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1594		    "Show raw trace event output (do not use print fmt or plugins)"),
1595	OPT_BOOLEAN('H', "hierarchy", &symbol_conf.report_hierarchy,
1596		    "Show entries in a hierarchy"),
1597	OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1598		    "Use a backward ring buffer, default: no"),
1599	OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1600	OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1601			"number of thread to run event synthesize"),
1602	OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1603		     "monitor event in cgroup name only", parse_cgroups),
1604	OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1605		    "Record namespaces events"),
1606	OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1607		    "Record cgroup events"),
1608	OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1609		    "Sort the output by the event at the index n in group. "
1610		    "If n is invalid, sort by the first event. "
1611		    "WARNING: should be used on grouped events."),
1612	OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1613		    "Enable LBR callgraph stitching approach"),
1614#ifdef HAVE_LIBPFM
1615	OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1616		"libpfm4 event selector. use 'perf list' to list available events",
1617		parse_libpfm_events_option),
1618#endif
1619	OPTS_EVSWITCH(&top.evswitch),
1620	OPT_END()
1621	};
1622	const char * const top_usage[] = {
1623		"perf top [<options>]",
1624		NULL
1625	};
1626	int status = hists__init();
1627	struct perf_env host_env;
1628
1629	if (status < 0)
1630		return status;
1631
1632	annotation_options__init();
1633
1634	annotate_opts.min_pcnt = 5;
1635	annotate_opts.context  = 4;
1636
1637	top.evlist = evlist__new();
1638	if (top.evlist == NULL)
1639		return -ENOMEM;
1640
1641	perf_env__init(&host_env);
1642	status = perf_config(perf_top_config, &top);
1643	if (status)
1644		goto out_delete_evlist;
1645	/*
1646	 * Since the per arch annotation init routine may need the cpuid, read
1647	 * it here, since we are not getting this from the perf.data header.
1648	 */
1649	status = perf_env__set_cmdline(&host_env, argc, argv);
1650	if (status)
1651		goto out_delete_evlist;
1652
1653	status = perf_env__read_cpuid(&host_env);
1654	if (status) {
1655		/*
1656		 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1657		 * warn the user explicitly.
1658		 */
1659		eprintf(status == ENOSYS ? 1 : 0, verbose,
1660			"Couldn't read the cpuid for this machine: %s\n",
1661			str_error_r(errno, errbuf, sizeof(errbuf)));
1662	}
1663
1664	argc = parse_options(argc, argv, options, top_usage, 0);
1665	if (argc)
1666		usage_with_options(top_usage, options);
1667
1668	if (disassembler_style) {
1669		annotate_opts.disassembler_style = strdup(disassembler_style);
1670		if (!annotate_opts.disassembler_style) {
1671			status = -ENOMEM;
1672			goto out_delete_evlist;
1673		}
1674	}
1675	if (objdump_path) {
1676		annotate_opts.objdump_path = strdup(objdump_path);
1677		if (!annotate_opts.objdump_path) {
1678			status = -ENOMEM;
1679			goto out_delete_evlist;
1680		}
1681	}
1682	if (addr2line_path) {
1683		symbol_conf.addr2line_path = strdup(addr2line_path);
1684		if (!symbol_conf.addr2line_path) {
1685			status = -ENOMEM;
1686			goto out_delete_evlist;
1687		}
1688	}
1689
1690	status = symbol__validate_sym_arguments();
1691	if (status)
1692		goto out_delete_evlist;
1693
1694	if (annotate_check_args() < 0)
1695		goto out_delete_evlist;
1696
1697	if (!top.evlist->core.nr_entries) {
1698		struct evlist *def_evlist = evlist__new_default();
1699
1700		if (!def_evlist)
1701			goto out_delete_evlist;
1702
1703		evlist__splice_list_tail(top.evlist, &def_evlist->core.entries);
1704		evlist__delete(def_evlist);
1705	}
1706
1707	status = evswitch__init(&top.evswitch, top.evlist, stderr);
1708	if (status)
1709		goto out_delete_evlist;
1710
1711	if (symbol_conf.report_hierarchy) {
1712		/* disable incompatible options */
1713		symbol_conf.event_group = false;
1714		symbol_conf.cumulate_callchain = false;
1715
1716		if (field_order) {
1717			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1718			parse_options_usage(top_usage, options, "fields", 0);
1719			parse_options_usage(NULL, options, "hierarchy", 0);
1720			goto out_delete_evlist;
1721		}
1722	}
1723
1724	if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1725		pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1726		goto out_delete_evlist;
1727	}
1728
1729	if (nr_cgroups > 0 && opts->record_cgroup) {
1730		pr_err("--cgroup and --all-cgroups cannot be used together\n");
1731		goto out_delete_evlist;
1732	}
1733
1734	if (branch_call_mode) {
1735		if (!opts->branch_stack)
1736			opts->branch_stack = PERF_SAMPLE_BRANCH_ANY;
1737		symbol_conf.use_callchain = true;
1738		callchain_param.key = CCKEY_ADDRESS;
1739		callchain_param.branch_callstack = true;
1740		callchain_param.enabled = true;
1741		if (callchain_param.record_mode == CALLCHAIN_NONE)
1742			callchain_param.record_mode = CALLCHAIN_FP;
1743		callchain_register_param(&callchain_param);
1744		if (!sort_order)
1745			sort_order = "srcline,symbol,dso";
1746	}
1747
1748	if (opts->branch_stack && callchain_param.enabled)
1749		symbol_conf.show_branchflag_count = true;
1750
1751	if (opts->branch_stack) {
1752		status = perf_env__read_core_pmu_caps(&host_env);
1753		if (status) {
1754			pr_err("PMU capability data is not available\n");
1755			goto out_delete_evlist;
1756		}
1757	}
1758
1759	sort__mode = SORT_MODE__TOP;
1760	/* display thread wants entries to be collapsed in a different tree */
1761	perf_hpp_list.need_collapse = 1;
1762
1763	if (top.use_stdio)
1764		use_browser = 0;
1765#ifdef HAVE_SLANG_SUPPORT
1766	else if (top.use_tui)
1767		use_browser = 1;
1768#endif
1769
1770	setup_browser(false);
1771
1772	top.session = __perf_session__new(/*data=*/NULL, /*tool=*/NULL,
1773					  /*trace_event_repipe=*/false,
1774					  &host_env);
1775	if (IS_ERR(top.session)) {
1776		status = PTR_ERR(top.session);
1777		top.session = NULL;
1778		goto out_delete_evlist;
1779	}
1780	top.evlist->session = top.session;
1781
1782	if (setup_sorting(top.evlist, perf_session__env(top.session)) < 0) {
1783		if (sort_order)
1784			parse_options_usage(top_usage, options, "s", 1);
1785		if (field_order)
1786			parse_options_usage(sort_order ? NULL : top_usage,
1787					    options, "fields", 0);
1788		goto out_delete_evlist;
1789	}
1790
1791	status = target__validate(target);
1792	if (status) {
1793		target__strerror(target, status, errbuf, BUFSIZ);
1794		ui__warning("%s\n", errbuf);
1795	}
1796
1797	if (top.uid_str) {
1798		uid_t uid = parse_uid(top.uid_str);
1799
1800		if (uid == UINT_MAX) {
1801			ui__error("Invalid User: %s", top.uid_str);
1802			status = -EINVAL;
1803			goto out_delete_evlist;
1804		}
1805		status = parse_uid_filter(top.evlist, uid);
1806		if (status)
1807			goto out_delete_evlist;
1808	}
1809
1810	if (target__none(target))
1811		target->system_wide = true;
1812
1813	if (evlist__create_maps(top.evlist, target) < 0) {
1814		ui__error("Couldn't create thread/CPU maps: %s\n",
1815			  errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1816		status = -errno;
1817		goto out_delete_evlist;
1818	}
1819
1820	if (top.delay_secs < 1)
1821		top.delay_secs = 1;
1822
1823	if (record_opts__config(opts)) {
1824		status = -EINVAL;
1825		goto out_delete_evlist;
1826	}
1827
1828	top.sym_evsel = evlist__first(top.evlist);
1829
1830	if (!callchain_param.enabled) {
1831		symbol_conf.cumulate_callchain = false;
1832		perf_hpp__cancel_cumulate(top.evlist);
1833	}
1834
1835	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1836		callchain_param.order = ORDER_CALLER;
1837
1838	status = symbol__annotation_init();
1839	if (status < 0)
1840		goto out_delete_evlist;
1841
1842	annotation_config__init();
1843
1844	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1845	status = symbol__init(NULL);
1846	if (status < 0)
1847		goto out_delete_evlist;
1848
1849	sort__setup_elide(stdout);
1850
1851	get_term_dimensions(&top.winsize);
1852	if (top.print_entries == 0) {
1853		perf_top__update_print_entries(&top);
1854		signal(SIGWINCH, winch_sig);
1855	}
1856
1857	if (!evlist__needs_bpf_sb_event(top.evlist))
1858		top.record_opts.no_bpf_event = true;
1859
1860#ifdef HAVE_LIBBPF_SUPPORT
1861	if (!top.record_opts.no_bpf_event) {
1862		top.sb_evlist = evlist__new();
1863
1864		if (top.sb_evlist == NULL) {
1865			pr_err("Couldn't create side band evlist.\n.");
1866			status = -EINVAL;
1867			goto out_delete_evlist;
1868		}
1869
1870		if (evlist__add_bpf_sb_event(top.sb_evlist, &host_env)) {
1871			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1872			status = -EINVAL;
1873			goto out_delete_evlist;
1874		}
1875	}
1876#endif
1877
1878	if (evlist__start_sb_thread(top.sb_evlist, target)) {
1879		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1880		opts->no_bpf_event = true;
1881	}
1882
1883	status = __cmd_top(&top);
1884
1885	if (!opts->no_bpf_event)
1886		evlist__stop_sb_thread(top.sb_evlist);
1887
1888out_delete_evlist:
1889	evlist__delete(top.evlist);
1890	perf_session__delete(top.session);
1891	annotation_options__exit();
1892	perf_env__exit(&host_env);
1893
1894	return status;
1895}