Linux Audio

Check our new training course

Loading...
v4.6
   1/*
   2 * builtin-top.c
   3 *
   4 * Builtin top command: Display a continuously updated profile of
   5 * any workload, CPU or specific PID.
   6 *
   7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8 *		 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9 *
  10 * Improvements and fixes by:
  11 *
  12 *   Arjan van de Ven <arjan@linux.intel.com>
  13 *   Yanmin Zhang <yanmin.zhang@intel.com>
  14 *   Wu Fengguang <fengguang.wu@intel.com>
  15 *   Mike Galbraith <efault@gmx.de>
  16 *   Paul Mackerras <paulus@samba.org>
  17 *
  18 * Released under the GPL v2. (and only v2, not any later version)
  19 */
  20#include "builtin.h"
  21
  22#include "perf.h"
  23
  24#include "util/annotate.h"
  25#include "util/cache.h"
  26#include "util/color.h"
 
  27#include "util/evlist.h"
  28#include "util/evsel.h"
  29#include "util/machine.h"
  30#include "util/session.h"
  31#include "util/symbol.h"
  32#include "util/thread.h"
  33#include "util/thread_map.h"
  34#include "util/top.h"
  35#include "util/util.h"
  36#include <linux/rbtree.h>
  37#include <subcmd/parse-options.h>
  38#include "util/parse-events.h"
  39#include "util/cpumap.h"
  40#include "util/xyarray.h"
  41#include "util/sort.h"
  42#include "util/intlist.h"
  43#include "util/parse-branch-options.h"
  44#include "arch/common.h"
  45
  46#include "util/debug.h"
  47
  48#include <assert.h>
  49#include <elf.h>
  50#include <fcntl.h>
  51
  52#include <stdio.h>
  53#include <termios.h>
  54#include <unistd.h>
  55#include <inttypes.h>
  56
  57#include <errno.h>
  58#include <time.h>
  59#include <sched.h>
  60
  61#include <sys/syscall.h>
  62#include <sys/ioctl.h>
  63#include <poll.h>
  64#include <sys/prctl.h>
  65#include <sys/wait.h>
  66#include <sys/uio.h>
  67#include <sys/utsname.h>
  68#include <sys/mman.h>
  69
  70#include <linux/stringify.h>
 
  71#include <linux/types.h>
  72
  73static volatile int done;
  74
  75#define HEADER_LINE_NR  5
  76
  77static void perf_top__update_print_entries(struct perf_top *top)
  78{
  79	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  80}
  81
  82static void perf_top__sig_winch(int sig __maybe_unused,
  83				siginfo_t *info __maybe_unused, void *arg)
  84{
  85	struct perf_top *top = arg;
  86
  87	get_term_dimensions(&top->winsize);
  88	perf_top__update_print_entries(top);
  89}
  90
  91static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  92{
  93	struct symbol *sym;
  94	struct annotation *notes;
  95	struct map *map;
  96	int err = -1;
  97
  98	if (!he || !he->ms.sym)
  99		return -1;
 100
 101	sym = he->ms.sym;
 102	map = he->ms.map;
 103
 104	/*
 105	 * We can't annotate with just /proc/kallsyms
 106	 */
 107	if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 108	    !dso__is_kcore(map->dso)) {
 109		pr_err("Can't annotate %s: No vmlinux file was found in the "
 110		       "path\n", sym->name);
 111		sleep(1);
 112		return -1;
 113	}
 114
 115	notes = symbol__annotation(sym);
 116	if (notes->src != NULL) {
 117		pthread_mutex_lock(&notes->lock);
 118		goto out_assign;
 119	}
 120
 121	pthread_mutex_lock(&notes->lock);
 122
 123	if (symbol__alloc_hist(sym) < 0) {
 124		pthread_mutex_unlock(&notes->lock);
 125		pr_err("Not enough memory for annotating '%s' symbol!\n",
 126		       sym->name);
 127		sleep(1);
 128		return err;
 129	}
 130
 131	err = symbol__annotate(sym, map, 0);
 132	if (err == 0) {
 133out_assign:
 134		top->sym_filter_entry = he;
 
 
 
 
 135	}
 136
 137	pthread_mutex_unlock(&notes->lock);
 138	return err;
 139}
 140
 141static void __zero_source_counters(struct hist_entry *he)
 142{
 143	struct symbol *sym = he->ms.sym;
 144	symbol__annotate_zero_histograms(sym);
 145}
 146
 147static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 148{
 149	struct utsname uts;
 150	int err = uname(&uts);
 151
 152	ui__warning("Out of bounds address found:\n\n"
 153		    "Addr:   %" PRIx64 "\n"
 154		    "DSO:    %s %c\n"
 155		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
 156		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 157		    "Arch:   %s\n"
 158		    "Kernel: %s\n"
 159		    "Tools:  %s\n\n"
 160		    "Not all samples will be on the annotation output.\n\n"
 161		    "Please report to linux-kernel@vger.kernel.org\n",
 162		    ip, map->dso->long_name, dso__symtab_origin(map->dso),
 163		    map->start, map->end, sym->start, sym->end,
 164		    sym->binding == STB_GLOBAL ? 'g' :
 165		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 166		    err ? "[unknown]" : uts.machine,
 167		    err ? "[unknown]" : uts.release, perf_version_string);
 168	if (use_browser <= 0)
 169		sleep(5);
 170
 171	map->erange_warned = true;
 172}
 173
 174static void perf_top__record_precise_ip(struct perf_top *top,
 175					struct hist_entry *he,
 176					int counter, u64 ip)
 177{
 178	struct annotation *notes;
 179	struct symbol *sym = he->ms.sym;
 180	int err = 0;
 181
 182	if (sym == NULL || (use_browser == 0 &&
 183			    (top->sym_filter_entry == NULL ||
 184			     top->sym_filter_entry->ms.sym != sym)))
 185		return;
 186
 187	notes = symbol__annotation(sym);
 188
 189	if (pthread_mutex_trylock(&notes->lock))
 190		return;
 191
 192	err = hist_entry__inc_addr_samples(he, counter, ip);
 193
 194	pthread_mutex_unlock(&notes->lock);
 195
 196	if (unlikely(err)) {
 197		/*
 198		 * This function is now called with he->hists->lock held.
 199		 * Release it before going to sleep.
 200		 */
 201		pthread_mutex_unlock(&he->hists->lock);
 202
 203		if (err == -ERANGE && !he->ms.map->erange_warned)
 204			ui__warn_map_erange(he->ms.map, sym, ip);
 205		else if (err == -ENOMEM) {
 206			pr_err("Not enough memory for annotating '%s' symbol!\n",
 207			       sym->name);
 208			sleep(1);
 209		}
 210
 211		pthread_mutex_lock(&he->hists->lock);
 212	}
 213}
 214
 215static void perf_top__show_details(struct perf_top *top)
 216{
 217	struct hist_entry *he = top->sym_filter_entry;
 218	struct annotation *notes;
 219	struct symbol *symbol;
 220	int more;
 221
 222	if (!he)
 223		return;
 224
 225	symbol = he->ms.sym;
 226	notes = symbol__annotation(symbol);
 227
 228	pthread_mutex_lock(&notes->lock);
 229
 230	if (notes->src == NULL)
 231		goto out_unlock;
 232
 233	printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 234	printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 235
 236	more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 237				       0, top->sym_pcnt_filter, top->print_entries, 4);
 238
 239	if (top->evlist->enabled) {
 240		if (top->zero)
 241			symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 242		else
 243			symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 244	}
 245	if (more != 0)
 246		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 247out_unlock:
 248	pthread_mutex_unlock(&notes->lock);
 249}
 250
 251static void perf_top__print_sym_table(struct perf_top *top)
 252{
 253	char bf[160];
 254	int printed = 0;
 255	const int win_width = top->winsize.ws_col - 1;
 256	struct perf_evsel *evsel = top->sym_evsel;
 257	struct hists *hists = evsel__hists(evsel);
 258
 259	puts(CONSOLE_CLEAR);
 260
 261	perf_top__header_snprintf(top, bf, sizeof(bf));
 262	printf("%s\n", bf);
 263
 264	perf_top__reset_sample_counters(top);
 265
 266	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 267
 268	if (hists->stats.nr_lost_warned !=
 269	    hists->stats.nr_events[PERF_RECORD_LOST]) {
 270		hists->stats.nr_lost_warned =
 271			      hists->stats.nr_events[PERF_RECORD_LOST];
 272		color_fprintf(stdout, PERF_COLOR_RED,
 273			      "WARNING: LOST %d chunks, Check IO/CPU overload",
 274			      hists->stats.nr_lost_warned);
 275		++printed;
 276	}
 277
 278	if (top->sym_filter_entry) {
 279		perf_top__show_details(top);
 280		return;
 281	}
 282
 283	if (top->evlist->enabled) {
 284		if (top->zero) {
 285			hists__delete_entries(hists);
 286		} else {
 287			hists__decay_entries(hists, top->hide_user_symbols,
 288					     top->hide_kernel_symbols);
 289		}
 290	}
 291
 292	hists__collapse_resort(hists, NULL);
 293	perf_evsel__output_resort(evsel, NULL);
 294
 295	hists__output_recalc_col_len(hists, top->print_entries - printed);
 296	putchar('\n');
 297	hists__fprintf(hists, false, top->print_entries - printed, win_width,
 298		       top->min_percent, stdout);
 299}
 300
 301static void prompt_integer(int *target, const char *msg)
 302{
 303	char *buf = malloc(0), *p;
 304	size_t dummy = 0;
 305	int tmp;
 306
 307	fprintf(stdout, "\n%s: ", msg);
 308	if (getline(&buf, &dummy, stdin) < 0)
 309		return;
 310
 311	p = strchr(buf, '\n');
 312	if (p)
 313		*p = 0;
 314
 315	p = buf;
 316	while(*p) {
 317		if (!isdigit(*p))
 318			goto out_free;
 319		p++;
 320	}
 321	tmp = strtoul(buf, NULL, 10);
 322	*target = tmp;
 323out_free:
 324	free(buf);
 325}
 326
 327static void prompt_percent(int *target, const char *msg)
 328{
 329	int tmp = 0;
 330
 331	prompt_integer(&tmp, msg);
 332	if (tmp >= 0 && tmp <= 100)
 333		*target = tmp;
 334}
 335
 336static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 337{
 338	char *buf = malloc(0), *p;
 339	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 340	struct hists *hists = evsel__hists(top->sym_evsel);
 341	struct rb_node *next;
 342	size_t dummy = 0;
 343
 344	/* zero counters of active symbol */
 345	if (syme) {
 346		__zero_source_counters(syme);
 347		top->sym_filter_entry = NULL;
 348	}
 349
 350	fprintf(stdout, "\n%s: ", msg);
 351	if (getline(&buf, &dummy, stdin) < 0)
 352		goto out_free;
 353
 354	p = strchr(buf, '\n');
 355	if (p)
 356		*p = 0;
 357
 358	next = rb_first(&hists->entries);
 359	while (next) {
 360		n = rb_entry(next, struct hist_entry, rb_node);
 361		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 362			found = n;
 363			break;
 364		}
 365		next = rb_next(&n->rb_node);
 366	}
 367
 368	if (!found) {
 369		fprintf(stderr, "Sorry, %s is not active.\n", buf);
 370		sleep(1);
 371	} else
 372		perf_top__parse_source(top, found);
 373
 374out_free:
 375	free(buf);
 376}
 377
 378static void perf_top__print_mapped_keys(struct perf_top *top)
 379{
 380	char *name = NULL;
 381
 382	if (top->sym_filter_entry) {
 383		struct symbol *sym = top->sym_filter_entry->ms.sym;
 384		name = sym->name;
 385	}
 386
 387	fprintf(stdout, "\nMapped keys:\n");
 388	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 389	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 390
 391	if (top->evlist->nr_entries > 1)
 392		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 393
 394	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 395
 396	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 397	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 398	fprintf(stdout, "\t[S]     stop annotation.\n");
 399
 400	fprintf(stdout,
 401		"\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 402		top->hide_kernel_symbols ? "yes" : "no");
 403	fprintf(stdout,
 404		"\t[U]     hide user symbols.               \t(%s)\n",
 405		top->hide_user_symbols ? "yes" : "no");
 406	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 407	fprintf(stdout, "\t[qQ]    quit.\n");
 408}
 409
 410static int perf_top__key_mapped(struct perf_top *top, int c)
 411{
 412	switch (c) {
 413		case 'd':
 414		case 'e':
 415		case 'f':
 416		case 'z':
 417		case 'q':
 418		case 'Q':
 419		case 'K':
 420		case 'U':
 421		case 'F':
 422		case 's':
 423		case 'S':
 424			return 1;
 425		case 'E':
 426			return top->evlist->nr_entries > 1 ? 1 : 0;
 427		default:
 428			break;
 429	}
 430
 431	return 0;
 432}
 433
 434static bool perf_top__handle_keypress(struct perf_top *top, int c)
 435{
 436	bool ret = true;
 437
 438	if (!perf_top__key_mapped(top, c)) {
 439		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 440		struct termios save;
 441
 442		perf_top__print_mapped_keys(top);
 443		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 444		fflush(stdout);
 445
 446		set_term_quiet_input(&save);
 447
 448		poll(&stdin_poll, 1, -1);
 449		c = getc(stdin);
 450
 451		tcsetattr(0, TCSAFLUSH, &save);
 452		if (!perf_top__key_mapped(top, c))
 453			return ret;
 454	}
 455
 456	switch (c) {
 457		case 'd':
 458			prompt_integer(&top->delay_secs, "Enter display delay");
 459			if (top->delay_secs < 1)
 460				top->delay_secs = 1;
 461			break;
 462		case 'e':
 463			prompt_integer(&top->print_entries, "Enter display entries (lines)");
 464			if (top->print_entries == 0) {
 465				struct sigaction act = {
 466					.sa_sigaction = perf_top__sig_winch,
 467					.sa_flags     = SA_SIGINFO,
 468				};
 469				perf_top__sig_winch(SIGWINCH, NULL, top);
 470				sigaction(SIGWINCH, &act, NULL);
 471			} else {
 472				signal(SIGWINCH, SIG_DFL);
 473			}
 474			break;
 475		case 'E':
 476			if (top->evlist->nr_entries > 1) {
 477				/* Select 0 as the default event: */
 478				int counter = 0;
 479
 480				fprintf(stderr, "\nAvailable events:");
 481
 482				evlist__for_each(top->evlist, top->sym_evsel)
 483					fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 484
 485				prompt_integer(&counter, "Enter details event counter");
 486
 487				if (counter >= top->evlist->nr_entries) {
 488					top->sym_evsel = perf_evlist__first(top->evlist);
 489					fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 490					sleep(1);
 491					break;
 492				}
 493				evlist__for_each(top->evlist, top->sym_evsel)
 494					if (top->sym_evsel->idx == counter)
 495						break;
 496			} else
 497				top->sym_evsel = perf_evlist__first(top->evlist);
 498			break;
 499		case 'f':
 500			prompt_integer(&top->count_filter, "Enter display event count filter");
 501			break;
 502		case 'F':
 503			prompt_percent(&top->sym_pcnt_filter,
 504				       "Enter details display event filter (percent)");
 505			break;
 506		case 'K':
 507			top->hide_kernel_symbols = !top->hide_kernel_symbols;
 508			break;
 509		case 'q':
 510		case 'Q':
 511			printf("exiting.\n");
 512			if (top->dump_symtab)
 513				perf_session__fprintf_dsos(top->session, stderr);
 514			ret = false;
 515			break;
 516		case 's':
 517			perf_top__prompt_symbol(top, "Enter details symbol");
 518			break;
 519		case 'S':
 520			if (!top->sym_filter_entry)
 521				break;
 522			else {
 523				struct hist_entry *syme = top->sym_filter_entry;
 524
 525				top->sym_filter_entry = NULL;
 526				__zero_source_counters(syme);
 527			}
 528			break;
 529		case 'U':
 530			top->hide_user_symbols = !top->hide_user_symbols;
 531			break;
 532		case 'z':
 533			top->zero = !top->zero;
 534			break;
 535		default:
 536			break;
 537	}
 538
 539	return ret;
 540}
 541
 542static void perf_top__sort_new_samples(void *arg)
 543{
 544	struct perf_top *t = arg;
 545	struct perf_evsel *evsel = t->sym_evsel;
 546	struct hists *hists;
 547
 548	perf_top__reset_sample_counters(t);
 549
 550	if (t->evlist->selected != NULL)
 551		t->sym_evsel = t->evlist->selected;
 552
 553	hists = evsel__hists(evsel);
 554
 555	if (t->evlist->enabled) {
 556		if (t->zero) {
 557			hists__delete_entries(hists);
 558		} else {
 559			hists__decay_entries(hists, t->hide_user_symbols,
 560					     t->hide_kernel_symbols);
 561		}
 562	}
 563
 564	hists__collapse_resort(hists, NULL);
 565	perf_evsel__output_resort(evsel, NULL);
 566}
 567
 568static void *display_thread_tui(void *arg)
 569{
 570	struct perf_evsel *pos;
 571	struct perf_top *top = arg;
 572	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 573	struct hist_browser_timer hbt = {
 574		.timer		= perf_top__sort_new_samples,
 575		.arg		= top,
 576		.refresh	= top->delay_secs,
 577	};
 578
 579	perf_top__sort_new_samples(top);
 580
 581	/*
 582	 * Initialize the uid_filter_str, in the future the TUI will allow
 583	 * Zooming in/out UIDs. For now juse use whatever the user passed
 584	 * via --uid.
 585	 */
 586	evlist__for_each(top->evlist, pos) {
 587		struct hists *hists = evsel__hists(pos);
 588		hists->uid_filter_str = top->record_opts.target.uid_str;
 589	}
 590
 591	perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 592				      top->min_percent,
 593				      &top->session->header.env);
 594
 595	done = 1;
 596	return NULL;
 597}
 598
 599static void display_sig(int sig __maybe_unused)
 600{
 601	done = 1;
 602}
 603
 604static void display_setup_sig(void)
 605{
 606	signal(SIGSEGV, sighandler_dump_stack);
 607	signal(SIGFPE, sighandler_dump_stack);
 608	signal(SIGINT,  display_sig);
 609	signal(SIGQUIT, display_sig);
 610	signal(SIGTERM, display_sig);
 611}
 612
 613static void *display_thread(void *arg)
 614{
 615	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 616	struct termios save;
 617	struct perf_top *top = arg;
 618	int delay_msecs, c;
 619
 620	display_setup_sig();
 621	pthread__unblock_sigwinch();
 622repeat:
 623	delay_msecs = top->delay_secs * 1000;
 624	set_term_quiet_input(&save);
 625	/* trash return*/
 626	getc(stdin);
 627
 628	while (!done) {
 629		perf_top__print_sym_table(top);
 630		/*
 631		 * Either timeout expired or we got an EINTR due to SIGWINCH,
 632		 * refresh screen in both cases.
 633		 */
 634		switch (poll(&stdin_poll, 1, delay_msecs)) {
 635		case 0:
 636			continue;
 637		case -1:
 638			if (errno == EINTR)
 639				continue;
 640			/* Fall trhu */
 641		default:
 642			c = getc(stdin);
 643			tcsetattr(0, TCSAFLUSH, &save);
 644
 645			if (perf_top__handle_keypress(top, c))
 646				goto repeat;
 647			done = 1;
 648		}
 649	}
 650
 651	tcsetattr(0, TCSAFLUSH, &save);
 652	return NULL;
 653}
 654
 655static int symbol_filter(struct map *map, struct symbol *sym)
 656{
 657	const char *name = sym->name;
 658
 659	if (!__map__is_kernel(map))
 660		return 0;
 661	/*
 662	 * ppc64 uses function descriptors and appends a '.' to the
 663	 * start of every instruction address. Remove it.
 664	 */
 665	if (name[0] == '.')
 666		name++;
 667
 668	if (!strcmp(name, "_text") ||
 669	    !strcmp(name, "_etext") ||
 670	    !strcmp(name, "_sinittext") ||
 671	    !strncmp("init_module", name, 11) ||
 672	    !strncmp("cleanup_module", name, 14) ||
 673	    strstr(name, "_text_start") ||
 674	    strstr(name, "_text_end"))
 675		return 1;
 676
 677	if (symbol__is_idle(sym))
 678		sym->ignore = true;
 679
 680	return 0;
 681}
 682
 683static int hist_iter__top_callback(struct hist_entry_iter *iter,
 684				   struct addr_location *al, bool single,
 685				   void *arg)
 686{
 687	struct perf_top *top = arg;
 688	struct hist_entry *he = iter->he;
 689	struct perf_evsel *evsel = iter->evsel;
 690
 691	if (sort__has_sym && single)
 692		perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
 693
 694	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 695		     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 696	return 0;
 697}
 698
 699static void perf_event__process_sample(struct perf_tool *tool,
 700				       const union perf_event *event,
 701				       struct perf_evsel *evsel,
 702				       struct perf_sample *sample,
 703				       struct machine *machine)
 704{
 705	struct perf_top *top = container_of(tool, struct perf_top, tool);
 706	struct addr_location al;
 707	int err;
 708
 709	if (!machine && perf_guest) {
 710		static struct intlist *seen;
 711
 712		if (!seen)
 713			seen = intlist__new(NULL);
 714
 715		if (!intlist__has_entry(seen, sample->pid)) {
 716			pr_err("Can't find guest [%d]'s kernel information\n",
 717				sample->pid);
 718			intlist__add(seen, sample->pid);
 719		}
 720		return;
 721	}
 722
 723	if (!machine) {
 724		pr_err("%u unprocessable samples recorded.\r",
 725		       top->session->evlist->stats.nr_unprocessable_samples++);
 726		return;
 727	}
 728
 729	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 730		top->exact_samples++;
 731
 732	if (machine__resolve(machine, &al, sample) < 0)
 733		return;
 734
 735	if (!top->kptr_restrict_warned &&
 736	    symbol_conf.kptr_restrict &&
 737	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
 738		ui__warning(
 739"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 740"Check /proc/sys/kernel/kptr_restrict.\n\n"
 741"Kernel%s samples will not be resolved.\n",
 742			  al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 743			  " modules" : "");
 744		if (use_browser <= 0)
 745			sleep(5);
 746		top->kptr_restrict_warned = true;
 747	}
 748
 749	if (al.sym == NULL) {
 750		const char *msg = "Kernel samples will not be resolved.\n";
 751		/*
 752		 * As we do lazy loading of symtabs we only will know if the
 753		 * specified vmlinux file is invalid when we actually have a
 754		 * hit in kernel space and then try to load it. So if we get
 755		 * here and there are _no_ symbols in the DSO backing the
 756		 * kernel map, bail out.
 757		 *
 758		 * We may never get here, for instance, if we use -K/
 759		 * --hide-kernel-symbols, even if the user specifies an
 760		 * invalid --vmlinux ;-)
 761		 */
 762		if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 763		    al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 764		    RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 765			if (symbol_conf.vmlinux_name) {
 766				char serr[256];
 767				dso__strerror_load(al.map->dso, serr, sizeof(serr));
 768				ui__warning("The %s file can't be used: %s\n%s",
 769					    symbol_conf.vmlinux_name, serr, msg);
 770			} else {
 771				ui__warning("A vmlinux file was not found.\n%s",
 772					    msg);
 773			}
 774
 775			if (use_browser <= 0)
 776				sleep(5);
 777			top->vmlinux_warned = true;
 778		}
 779	}
 780
 781	if (al.sym == NULL || !al.sym->ignore) {
 782		struct hists *hists = evsel__hists(evsel);
 783		struct hist_entry_iter iter = {
 784			.evsel		= evsel,
 785			.sample 	= sample,
 786			.add_entry_cb 	= hist_iter__top_callback,
 787		};
 788
 789		if (symbol_conf.cumulate_callchain)
 790			iter.ops = &hist_iter_cumulative;
 791		else
 792			iter.ops = &hist_iter_normal;
 793
 794		pthread_mutex_lock(&hists->lock);
 795
 796		err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 797		if (err < 0)
 798			pr_err("Problem incrementing symbol period, skipping event\n");
 799
 800		pthread_mutex_unlock(&hists->lock);
 801	}
 802
 803	addr_location__put(&al);
 804}
 805
 806static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 807{
 808	struct perf_sample sample;
 809	struct perf_evsel *evsel;
 810	struct perf_session *session = top->session;
 811	union perf_event *event;
 812	struct machine *machine;
 813	int ret;
 814
 815	while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 816		ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 817		if (ret) {
 818			pr_err("Can't parse sample, err = %d\n", ret);
 819			goto next_event;
 820		}
 821
 822		evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 823		assert(evsel != NULL);
 824
 825		if (event->header.type == PERF_RECORD_SAMPLE)
 826			++top->samples;
 827
 828		switch (sample.cpumode) {
 829		case PERF_RECORD_MISC_USER:
 830			++top->us_samples;
 831			if (top->hide_user_symbols)
 832				goto next_event;
 833			machine = &session->machines.host;
 834			break;
 835		case PERF_RECORD_MISC_KERNEL:
 836			++top->kernel_samples;
 837			if (top->hide_kernel_symbols)
 838				goto next_event;
 839			machine = &session->machines.host;
 840			break;
 841		case PERF_RECORD_MISC_GUEST_KERNEL:
 842			++top->guest_kernel_samples;
 843			machine = perf_session__find_machine(session,
 844							     sample.pid);
 845			break;
 846		case PERF_RECORD_MISC_GUEST_USER:
 847			++top->guest_us_samples;
 848			/*
 849			 * TODO: we don't process guest user from host side
 850			 * except simple counting.
 851			 */
 852			goto next_event;
 853		default:
 854			if (event->header.type == PERF_RECORD_SAMPLE)
 855				goto next_event;
 856			machine = &session->machines.host;
 857			break;
 858		}
 859
 860
 861		if (event->header.type == PERF_RECORD_SAMPLE) {
 862			perf_event__process_sample(&top->tool, event, evsel,
 863						   &sample, machine);
 864		} else if (event->header.type < PERF_RECORD_MAX) {
 865			hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 866			machine__process_event(machine, event, &sample);
 867		} else
 868			++session->evlist->stats.nr_unknown_events;
 869next_event:
 870		perf_evlist__mmap_consume(top->evlist, idx);
 871	}
 872}
 873
 874static void perf_top__mmap_read(struct perf_top *top)
 875{
 876	int i;
 877
 878	for (i = 0; i < top->evlist->nr_mmaps; i++)
 879		perf_top__mmap_read_idx(top, i);
 880}
 881
 882static int perf_top__start_counters(struct perf_top *top)
 883{
 884	char msg[512];
 885	struct perf_evsel *counter;
 886	struct perf_evlist *evlist = top->evlist;
 887	struct record_opts *opts = &top->record_opts;
 888
 889	perf_evlist__config(evlist, opts);
 890
 891	evlist__for_each(evlist, counter) {
 892try_again:
 893		if (perf_evsel__open(counter, top->evlist->cpus,
 894				     top->evlist->threads) < 0) {
 895			if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 896				if (verbose)
 897					ui__warning("%s\n", msg);
 898				goto try_again;
 899			}
 900
 901			perf_evsel__open_strerror(counter, &opts->target,
 902						  errno, msg, sizeof(msg));
 903			ui__error("%s\n", msg);
 904			goto out_err;
 905		}
 906	}
 907
 908	if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 909		ui__error("Failed to mmap with %d (%s)\n",
 910			    errno, strerror_r(errno, msg, sizeof(msg)));
 911		goto out_err;
 912	}
 913
 914	return 0;
 915
 916out_err:
 917	return -1;
 918}
 919
 920static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 921{
 922	if (!sort__has_sym) {
 923		if (symbol_conf.use_callchain) {
 924			ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 925			return -EINVAL;
 926		}
 927	} else if (callchain_param.mode != CHAIN_NONE) {
 928		if (callchain_register_param(&callchain_param) < 0) {
 929			ui__error("Can't register callchain params.\n");
 930			return -EINVAL;
 931		}
 932	}
 933
 934	return 0;
 935}
 936
 937static int __cmd_top(struct perf_top *top)
 938{
 
 
 
 
 939	struct record_opts *opts = &top->record_opts;
 940	pthread_t thread;
 941	int ret;
 942
 943	top->session = perf_session__new(NULL, false, NULL);
 944	if (top->session == NULL)
 945		return -1;
 946
 947	machines__set_symbol_filter(&top->session->machines, symbol_filter);
 948
 949	if (!objdump_path) {
 950		ret = perf_env__lookup_objdump(&top->session->header.env);
 951		if (ret)
 952			goto out_delete;
 953	}
 954
 955	ret = perf_top__setup_sample_type(top);
 956	if (ret)
 957		goto out_delete;
 958
 959	if (perf_session__register_idle_thread(top->session) < 0)
 960		goto out_delete;
 961
 962	machine__synthesize_threads(&top->session->machines.host, &opts->target,
 963				    top->evlist->threads, false, opts->proc_map_timeout);
 964
 965	if (sort__has_socket) {
 966		ret = perf_env__read_cpu_topology_map(&perf_env);
 967		if (ret < 0)
 968			goto out_err_cpu_topo;
 969	}
 970
 971	ret = perf_top__start_counters(top);
 972	if (ret)
 973		goto out_delete;
 974
 
 
 
 
 
 
 
 
 975	top->session->evlist = top->evlist;
 976	perf_session__set_id_hdr_size(top->session);
 977
 978	/*
 979	 * When perf is starting the traced process, all the events (apart from
 980	 * group members) have enable_on_exec=1 set, so don't spoil it by
 981	 * prematurely enabling them.
 982	 *
 983	 * XXX 'top' still doesn't start workloads like record, trace, but should,
 984	 * so leave the check here.
 985	 */
 986        if (!target__none(&opts->target))
 987                perf_evlist__enable(top->evlist);
 988
 989	/* Wait for a minimal set of events before starting the snapshot */
 990	perf_evlist__poll(top->evlist, 100);
 991
 992	perf_top__mmap_read(top);
 993
 994	ret = -1;
 995	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 996							    display_thread), top)) {
 997		ui__error("Could not create display thread.\n");
 998		goto out_delete;
 999	}
1000
1001	if (top->realtime_prio) {
1002		struct sched_param param;
1003
1004		param.sched_priority = top->realtime_prio;
1005		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1006			ui__error("Could not set realtime priority.\n");
1007			goto out_join;
1008		}
1009	}
1010
1011	while (!done) {
1012		u64 hits = top->samples;
1013
1014		perf_top__mmap_read(top);
1015
1016		if (hits == top->samples)
1017			ret = perf_evlist__poll(top->evlist, 100);
1018	}
1019
1020	ret = 0;
1021out_join:
1022	pthread_join(thread, NULL);
1023out_delete:
1024	perf_session__delete(top->session);
1025	top->session = NULL;
1026
1027	return ret;
1028
1029out_err_cpu_topo: {
1030	char errbuf[BUFSIZ];
1031	const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1032
1033	ui__error("Could not read the CPU topology map: %s\n", err);
1034	goto out_delete;
1035}
1036}
1037
1038static int
1039callchain_opt(const struct option *opt, const char *arg, int unset)
1040{
1041	symbol_conf.use_callchain = true;
1042	return record_callchain_opt(opt, arg, unset);
1043}
1044
1045static int
1046parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1047{
1048	struct record_opts *record = (struct record_opts *)opt->value;
1049
1050	record->callgraph_set = true;
1051	callchain_param.enabled = !unset;
1052	callchain_param.record_mode = CALLCHAIN_FP;
1053
1054	/*
1055	 * --no-call-graph
1056	 */
1057	if (unset) {
1058		symbol_conf.use_callchain = false;
1059		callchain_param.record_mode = CALLCHAIN_NONE;
1060		return 0;
1061	}
1062
1063	return parse_callchain_top_opt(arg);
1064}
1065
1066static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1067{
1068	if (!strcmp(var, "top.call-graph"))
1069		var = "call-graph.record-mode"; /* fall-through */
1070	if (!strcmp(var, "top.children")) {
1071		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1072		return 0;
1073	}
1074
1075	return 0;
1076}
1077
1078static int
1079parse_percent_limit(const struct option *opt, const char *arg,
1080		    int unset __maybe_unused)
1081{
1082	struct perf_top *top = opt->value;
1083
1084	top->min_percent = strtof(arg, NULL);
1085	return 0;
1086}
1087
1088const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1089	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1090
1091int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1092{
1093	char errbuf[BUFSIZ];
1094	struct perf_top top = {
1095		.count_filter	     = 5,
1096		.delay_secs	     = 2,
1097		.record_opts = {
1098			.mmap_pages	= UINT_MAX,
1099			.user_freq	= UINT_MAX,
1100			.user_interval	= ULLONG_MAX,
1101			.freq		= 4000, /* 4 KHz */
1102			.target		= {
1103				.uses_mmap   = true,
1104			},
1105			.proc_map_timeout    = 500,
1106		},
1107		.max_stack	     = PERF_MAX_STACK_DEPTH,
1108		.sym_pcnt_filter     = 5,
1109	};
1110	struct record_opts *opts = &top.record_opts;
1111	struct target *target = &opts->target;
1112	const struct option options[] = {
1113	OPT_CALLBACK('e', "event", &top.evlist, "event",
1114		     "event selector. use 'perf list' to list available events",
1115		     parse_events_option),
1116	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1117	OPT_STRING('p', "pid", &target->pid, "pid",
1118		    "profile events on existing process id"),
1119	OPT_STRING('t', "tid", &target->tid, "tid",
1120		    "profile events on existing thread id"),
1121	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1122			    "system-wide collection from all CPUs"),
1123	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1124		    "list of cpus to monitor"),
1125	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1126		   "file", "vmlinux pathname"),
1127	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1128		    "don't load vmlinux even if found"),
1129	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1130		    "hide kernel symbols"),
1131	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1132		     "number of mmap data pages",
1133		     perf_evlist__parse_mmap_pages),
1134	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1135		    "collect data with this RT SCHED_FIFO priority"),
1136	OPT_INTEGER('d', "delay", &top.delay_secs,
1137		    "number of seconds to delay between refreshes"),
1138	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1139			    "dump the symbol table used for profiling"),
1140	OPT_INTEGER('f', "count-filter", &top.count_filter,
1141		    "only display functions with more events than this"),
1142	OPT_BOOLEAN(0, "group", &opts->group,
1143			    "put the counters into a counter group"),
1144	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1145		    "child tasks do not inherit counters"),
1146	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1147		    "symbol to annotate"),
1148	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1149	OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1150	OPT_INTEGER('E', "entries", &top.print_entries,
1151		    "display this many functions"),
1152	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1153		    "hide user symbols"),
1154	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1155	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1156	OPT_INCR('v', "verbose", &verbose,
1157		    "be more verbose (show counter open errors, etc)"),
1158	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1159		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1160		   " Please refer the man page for the complete list."),
1161	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1162		   "output field(s): overhead, period, sample plus all of sort keys"),
1163	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1164		    "Show a column with the number of samples"),
1165	OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1166			   NULL, "enables call-graph recording and display",
1167			   &callchain_opt),
1168	OPT_CALLBACK(0, "call-graph", &top.record_opts,
1169		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1170		     top_callchain_help, &parse_callchain_opt),
1171	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1172		    "Accumulate callchains of children and show total overhead as well"),
1173	OPT_INTEGER(0, "max-stack", &top.max_stack,
1174		    "Set the maximum stack depth when parsing the callchain. "
1175		    "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1176	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1177		   "ignore callees of these functions in call graphs",
1178		   report_parse_ignore_callees_opt),
1179	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1180		    "Show a column with the sum of periods"),
1181	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1182		   "only consider symbols in these dsos"),
1183	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1184		   "only consider symbols in these comms"),
1185	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1186		   "only consider these symbols"),
1187	OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1188		    "Interleave source code with assembly code (default)"),
1189	OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1190		    "Display raw encoding of assembly instructions (default)"),
1191	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1192		    "Enable kernel symbol demangling"),
1193	OPT_STRING(0, "objdump", &objdump_path, "path",
1194		    "objdump binary to use for disassembly and annotations"),
1195	OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1196		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1197	OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1198	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1199		     "Don't show entries under that percent", parse_percent_limit),
1200	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1201		     "How to display percentage of filtered entries", parse_filter_percentage),
1202	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1203		   "width[,width...]",
1204		   "don't try to adjust column width, use these fixed values"),
1205	OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1206			"per thread proc mmap processing timeout in ms"),
1207	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1208		     "branch any", "sample any taken branches",
1209		     parse_branch_stack),
1210	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1211		     "branch filter mask", "branch stack filter modes",
1212		     parse_branch_stack),
1213	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1214		    "Show raw trace event output (do not use print fmt or plugins)"),
1215	OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1216		    "Show entries in a hierarchy"),
1217	OPT_END()
1218	};
1219	const char * const top_usage[] = {
1220		"perf top [<options>]",
1221		NULL
1222	};
1223	int status = hists__init();
1224
1225	if (status < 0)
1226		return status;
1227
1228	top.evlist = perf_evlist__new();
1229	if (top.evlist == NULL)
1230		return -ENOMEM;
1231
1232	perf_config(perf_top_config, &top);
1233
1234	argc = parse_options(argc, argv, options, top_usage, 0);
1235	if (argc)
1236		usage_with_options(top_usage, options);
1237
1238	if (!top.evlist->nr_entries &&
1239	    perf_evlist__add_default(top.evlist) < 0) {
1240		pr_err("Not enough memory for event selector list\n");
1241		goto out_delete_evlist;
1242	}
1243
1244	if (symbol_conf.report_hierarchy) {
1245		/* disable incompatible options */
1246		symbol_conf.event_group = false;
1247		symbol_conf.cumulate_callchain = false;
1248
1249		if (field_order) {
1250			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1251			parse_options_usage(top_usage, options, "fields", 0);
1252			parse_options_usage(NULL, options, "hierarchy", 0);
1253			goto out_delete_evlist;
1254		}
1255	}
1256
1257	sort__mode = SORT_MODE__TOP;
1258	/* display thread wants entries to be collapsed in a different tree */
1259	sort__need_collapse = 1;
1260
1261	if (top.use_stdio)
1262		use_browser = 0;
1263	else if (top.use_tui)
1264		use_browser = 1;
1265
1266	setup_browser(false);
1267
1268	if (setup_sorting(top.evlist) < 0) {
1269		if (sort_order)
1270			parse_options_usage(top_usage, options, "s", 1);
1271		if (field_order)
1272			parse_options_usage(sort_order ? NULL : top_usage,
1273					    options, "fields", 0);
1274		goto out_delete_evlist;
1275	}
1276
1277	status = target__validate(target);
1278	if (status) {
1279		target__strerror(target, status, errbuf, BUFSIZ);
1280		ui__warning("%s\n", errbuf);
1281	}
1282
1283	status = target__parse_uid(target);
1284	if (status) {
1285		int saved_errno = errno;
1286
1287		target__strerror(target, status, errbuf, BUFSIZ);
1288		ui__error("%s\n", errbuf);
1289
1290		status = -saved_errno;
1291		goto out_delete_evlist;
1292	}
1293
1294	if (target__none(target))
1295		target->system_wide = true;
1296
1297	if (perf_evlist__create_maps(top.evlist, target) < 0) {
1298		ui__error("Couldn't create thread/CPU maps: %s\n",
1299			  errno == ENOENT ? "No such process" : strerror_r(errno, errbuf, sizeof(errbuf)));
1300		goto out_delete_evlist;
1301	}
1302
1303	symbol_conf.nr_events = top.evlist->nr_entries;
1304
1305	if (top.delay_secs < 1)
1306		top.delay_secs = 1;
1307
1308	if (record_opts__config(opts)) {
1309		status = -EINVAL;
1310		goto out_delete_evlist;
1311	}
1312
1313	top.sym_evsel = perf_evlist__first(top.evlist);
1314
1315	if (!symbol_conf.use_callchain) {
1316		symbol_conf.cumulate_callchain = false;
1317		perf_hpp__cancel_cumulate();
1318	}
1319
1320	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1321		callchain_param.order = ORDER_CALLER;
1322
1323	symbol_conf.priv_size = sizeof(struct annotation);
 
 
1324
1325	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1326	if (symbol__init(NULL) < 0)
1327		return -1;
1328
1329	sort__setup_elide(stdout);
1330
1331	get_term_dimensions(&top.winsize);
1332	if (top.print_entries == 0) {
1333		struct sigaction act = {
1334			.sa_sigaction = perf_top__sig_winch,
1335			.sa_flags     = SA_SIGINFO,
1336		};
1337		perf_top__update_print_entries(&top);
1338		sigaction(SIGWINCH, &act, NULL);
1339	}
1340
1341	status = __cmd_top(&top);
1342
1343out_delete_evlist:
1344	perf_evlist__delete(top.evlist);
1345
1346	return status;
1347}
v4.10.11
   1/*
   2 * builtin-top.c
   3 *
   4 * Builtin top command: Display a continuously updated profile of
   5 * any workload, CPU or specific PID.
   6 *
   7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8 *		 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9 *
  10 * Improvements and fixes by:
  11 *
  12 *   Arjan van de Ven <arjan@linux.intel.com>
  13 *   Yanmin Zhang <yanmin.zhang@intel.com>
  14 *   Wu Fengguang <fengguang.wu@intel.com>
  15 *   Mike Galbraith <efault@gmx.de>
  16 *   Paul Mackerras <paulus@samba.org>
  17 *
  18 * Released under the GPL v2. (and only v2, not any later version)
  19 */
  20#include "builtin.h"
  21
  22#include "perf.h"
  23
  24#include "util/annotate.h"
  25#include "util/config.h"
  26#include "util/color.h"
  27#include "util/drv_configs.h"
  28#include "util/evlist.h"
  29#include "util/evsel.h"
  30#include "util/machine.h"
  31#include "util/session.h"
  32#include "util/symbol.h"
  33#include "util/thread.h"
  34#include "util/thread_map.h"
  35#include "util/top.h"
  36#include "util/util.h"
  37#include <linux/rbtree.h>
  38#include <subcmd/parse-options.h>
  39#include "util/parse-events.h"
  40#include "util/cpumap.h"
  41#include "util/xyarray.h"
  42#include "util/sort.h"
  43#include "util/intlist.h"
  44#include "util/parse-branch-options.h"
  45#include "arch/common.h"
  46
  47#include "util/debug.h"
  48
  49#include <assert.h>
  50#include <elf.h>
  51#include <fcntl.h>
  52
  53#include <stdio.h>
  54#include <termios.h>
  55#include <unistd.h>
  56#include <inttypes.h>
  57
  58#include <errno.h>
  59#include <time.h>
  60#include <sched.h>
  61
  62#include <sys/syscall.h>
  63#include <sys/ioctl.h>
  64#include <poll.h>
  65#include <sys/prctl.h>
  66#include <sys/wait.h>
  67#include <sys/uio.h>
  68#include <sys/utsname.h>
  69#include <sys/mman.h>
  70
  71#include <linux/stringify.h>
  72#include <linux/time64.h>
  73#include <linux/types.h>
  74
  75static volatile int done;
  76
  77#define HEADER_LINE_NR  5
  78
  79static void perf_top__update_print_entries(struct perf_top *top)
  80{
  81	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  82}
  83
  84static void perf_top__sig_winch(int sig __maybe_unused,
  85				siginfo_t *info __maybe_unused, void *arg)
  86{
  87	struct perf_top *top = arg;
  88
  89	get_term_dimensions(&top->winsize);
  90	perf_top__update_print_entries(top);
  91}
  92
  93static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  94{
  95	struct symbol *sym;
  96	struct annotation *notes;
  97	struct map *map;
  98	int err = -1;
  99
 100	if (!he || !he->ms.sym)
 101		return -1;
 102
 103	sym = he->ms.sym;
 104	map = he->ms.map;
 105
 106	/*
 107	 * We can't annotate with just /proc/kallsyms
 108	 */
 109	if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 110	    !dso__is_kcore(map->dso)) {
 111		pr_err("Can't annotate %s: No vmlinux file was found in the "
 112		       "path\n", sym->name);
 113		sleep(1);
 114		return -1;
 115	}
 116
 117	notes = symbol__annotation(sym);
 118	if (notes->src != NULL) {
 119		pthread_mutex_lock(&notes->lock);
 120		goto out_assign;
 121	}
 122
 123	pthread_mutex_lock(&notes->lock);
 124
 125	if (symbol__alloc_hist(sym) < 0) {
 126		pthread_mutex_unlock(&notes->lock);
 127		pr_err("Not enough memory for annotating '%s' symbol!\n",
 128		       sym->name);
 129		sleep(1);
 130		return err;
 131	}
 132
 133	err = symbol__disassemble(sym, map, NULL, 0);
 134	if (err == 0) {
 135out_assign:
 136		top->sym_filter_entry = he;
 137	} else {
 138		char msg[BUFSIZ];
 139		symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 140		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 141	}
 142
 143	pthread_mutex_unlock(&notes->lock);
 144	return err;
 145}
 146
 147static void __zero_source_counters(struct hist_entry *he)
 148{
 149	struct symbol *sym = he->ms.sym;
 150	symbol__annotate_zero_histograms(sym);
 151}
 152
 153static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 154{
 155	struct utsname uts;
 156	int err = uname(&uts);
 157
 158	ui__warning("Out of bounds address found:\n\n"
 159		    "Addr:   %" PRIx64 "\n"
 160		    "DSO:    %s %c\n"
 161		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
 162		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 163		    "Arch:   %s\n"
 164		    "Kernel: %s\n"
 165		    "Tools:  %s\n\n"
 166		    "Not all samples will be on the annotation output.\n\n"
 167		    "Please report to linux-kernel@vger.kernel.org\n",
 168		    ip, map->dso->long_name, dso__symtab_origin(map->dso),
 169		    map->start, map->end, sym->start, sym->end,
 170		    sym->binding == STB_GLOBAL ? 'g' :
 171		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 172		    err ? "[unknown]" : uts.machine,
 173		    err ? "[unknown]" : uts.release, perf_version_string);
 174	if (use_browser <= 0)
 175		sleep(5);
 176
 177	map->erange_warned = true;
 178}
 179
 180static void perf_top__record_precise_ip(struct perf_top *top,
 181					struct hist_entry *he,
 182					int counter, u64 ip)
 183{
 184	struct annotation *notes;
 185	struct symbol *sym = he->ms.sym;
 186	int err = 0;
 187
 188	if (sym == NULL || (use_browser == 0 &&
 189			    (top->sym_filter_entry == NULL ||
 190			     top->sym_filter_entry->ms.sym != sym)))
 191		return;
 192
 193	notes = symbol__annotation(sym);
 194
 195	if (pthread_mutex_trylock(&notes->lock))
 196		return;
 197
 198	err = hist_entry__inc_addr_samples(he, counter, ip);
 199
 200	pthread_mutex_unlock(&notes->lock);
 201
 202	if (unlikely(err)) {
 203		/*
 204		 * This function is now called with he->hists->lock held.
 205		 * Release it before going to sleep.
 206		 */
 207		pthread_mutex_unlock(&he->hists->lock);
 208
 209		if (err == -ERANGE && !he->ms.map->erange_warned)
 210			ui__warn_map_erange(he->ms.map, sym, ip);
 211		else if (err == -ENOMEM) {
 212			pr_err("Not enough memory for annotating '%s' symbol!\n",
 213			       sym->name);
 214			sleep(1);
 215		}
 216
 217		pthread_mutex_lock(&he->hists->lock);
 218	}
 219}
 220
 221static void perf_top__show_details(struct perf_top *top)
 222{
 223	struct hist_entry *he = top->sym_filter_entry;
 224	struct annotation *notes;
 225	struct symbol *symbol;
 226	int more;
 227
 228	if (!he)
 229		return;
 230
 231	symbol = he->ms.sym;
 232	notes = symbol__annotation(symbol);
 233
 234	pthread_mutex_lock(&notes->lock);
 235
 236	if (notes->src == NULL)
 237		goto out_unlock;
 238
 239	printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 240	printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 241
 242	more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 243				       0, top->sym_pcnt_filter, top->print_entries, 4);
 244
 245	if (top->evlist->enabled) {
 246		if (top->zero)
 247			symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 248		else
 249			symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 250	}
 251	if (more != 0)
 252		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 253out_unlock:
 254	pthread_mutex_unlock(&notes->lock);
 255}
 256
 257static void perf_top__print_sym_table(struct perf_top *top)
 258{
 259	char bf[160];
 260	int printed = 0;
 261	const int win_width = top->winsize.ws_col - 1;
 262	struct perf_evsel *evsel = top->sym_evsel;
 263	struct hists *hists = evsel__hists(evsel);
 264
 265	puts(CONSOLE_CLEAR);
 266
 267	perf_top__header_snprintf(top, bf, sizeof(bf));
 268	printf("%s\n", bf);
 269
 270	perf_top__reset_sample_counters(top);
 271
 272	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 273
 274	if (hists->stats.nr_lost_warned !=
 275	    hists->stats.nr_events[PERF_RECORD_LOST]) {
 276		hists->stats.nr_lost_warned =
 277			      hists->stats.nr_events[PERF_RECORD_LOST];
 278		color_fprintf(stdout, PERF_COLOR_RED,
 279			      "WARNING: LOST %d chunks, Check IO/CPU overload",
 280			      hists->stats.nr_lost_warned);
 281		++printed;
 282	}
 283
 284	if (top->sym_filter_entry) {
 285		perf_top__show_details(top);
 286		return;
 287	}
 288
 289	if (top->evlist->enabled) {
 290		if (top->zero) {
 291			hists__delete_entries(hists);
 292		} else {
 293			hists__decay_entries(hists, top->hide_user_symbols,
 294					     top->hide_kernel_symbols);
 295		}
 296	}
 297
 298	hists__collapse_resort(hists, NULL);
 299	perf_evsel__output_resort(evsel, NULL);
 300
 301	hists__output_recalc_col_len(hists, top->print_entries - printed);
 302	putchar('\n');
 303	hists__fprintf(hists, false, top->print_entries - printed, win_width,
 304		       top->min_percent, stdout, symbol_conf.use_callchain);
 305}
 306
 307static void prompt_integer(int *target, const char *msg)
 308{
 309	char *buf = malloc(0), *p;
 310	size_t dummy = 0;
 311	int tmp;
 312
 313	fprintf(stdout, "\n%s: ", msg);
 314	if (getline(&buf, &dummy, stdin) < 0)
 315		return;
 316
 317	p = strchr(buf, '\n');
 318	if (p)
 319		*p = 0;
 320
 321	p = buf;
 322	while(*p) {
 323		if (!isdigit(*p))
 324			goto out_free;
 325		p++;
 326	}
 327	tmp = strtoul(buf, NULL, 10);
 328	*target = tmp;
 329out_free:
 330	free(buf);
 331}
 332
 333static void prompt_percent(int *target, const char *msg)
 334{
 335	int tmp = 0;
 336
 337	prompt_integer(&tmp, msg);
 338	if (tmp >= 0 && tmp <= 100)
 339		*target = tmp;
 340}
 341
 342static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 343{
 344	char *buf = malloc(0), *p;
 345	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 346	struct hists *hists = evsel__hists(top->sym_evsel);
 347	struct rb_node *next;
 348	size_t dummy = 0;
 349
 350	/* zero counters of active symbol */
 351	if (syme) {
 352		__zero_source_counters(syme);
 353		top->sym_filter_entry = NULL;
 354	}
 355
 356	fprintf(stdout, "\n%s: ", msg);
 357	if (getline(&buf, &dummy, stdin) < 0)
 358		goto out_free;
 359
 360	p = strchr(buf, '\n');
 361	if (p)
 362		*p = 0;
 363
 364	next = rb_first(&hists->entries);
 365	while (next) {
 366		n = rb_entry(next, struct hist_entry, rb_node);
 367		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 368			found = n;
 369			break;
 370		}
 371		next = rb_next(&n->rb_node);
 372	}
 373
 374	if (!found) {
 375		fprintf(stderr, "Sorry, %s is not active.\n", buf);
 376		sleep(1);
 377	} else
 378		perf_top__parse_source(top, found);
 379
 380out_free:
 381	free(buf);
 382}
 383
 384static void perf_top__print_mapped_keys(struct perf_top *top)
 385{
 386	char *name = NULL;
 387
 388	if (top->sym_filter_entry) {
 389		struct symbol *sym = top->sym_filter_entry->ms.sym;
 390		name = sym->name;
 391	}
 392
 393	fprintf(stdout, "\nMapped keys:\n");
 394	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 395	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 396
 397	if (top->evlist->nr_entries > 1)
 398		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 399
 400	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 401
 402	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 403	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 404	fprintf(stdout, "\t[S]     stop annotation.\n");
 405
 406	fprintf(stdout,
 407		"\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 408		top->hide_kernel_symbols ? "yes" : "no");
 409	fprintf(stdout,
 410		"\t[U]     hide user symbols.               \t(%s)\n",
 411		top->hide_user_symbols ? "yes" : "no");
 412	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 413	fprintf(stdout, "\t[qQ]    quit.\n");
 414}
 415
 416static int perf_top__key_mapped(struct perf_top *top, int c)
 417{
 418	switch (c) {
 419		case 'd':
 420		case 'e':
 421		case 'f':
 422		case 'z':
 423		case 'q':
 424		case 'Q':
 425		case 'K':
 426		case 'U':
 427		case 'F':
 428		case 's':
 429		case 'S':
 430			return 1;
 431		case 'E':
 432			return top->evlist->nr_entries > 1 ? 1 : 0;
 433		default:
 434			break;
 435	}
 436
 437	return 0;
 438}
 439
 440static bool perf_top__handle_keypress(struct perf_top *top, int c)
 441{
 442	bool ret = true;
 443
 444	if (!perf_top__key_mapped(top, c)) {
 445		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 446		struct termios save;
 447
 448		perf_top__print_mapped_keys(top);
 449		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 450		fflush(stdout);
 451
 452		set_term_quiet_input(&save);
 453
 454		poll(&stdin_poll, 1, -1);
 455		c = getc(stdin);
 456
 457		tcsetattr(0, TCSAFLUSH, &save);
 458		if (!perf_top__key_mapped(top, c))
 459			return ret;
 460	}
 461
 462	switch (c) {
 463		case 'd':
 464			prompt_integer(&top->delay_secs, "Enter display delay");
 465			if (top->delay_secs < 1)
 466				top->delay_secs = 1;
 467			break;
 468		case 'e':
 469			prompt_integer(&top->print_entries, "Enter display entries (lines)");
 470			if (top->print_entries == 0) {
 471				struct sigaction act = {
 472					.sa_sigaction = perf_top__sig_winch,
 473					.sa_flags     = SA_SIGINFO,
 474				};
 475				perf_top__sig_winch(SIGWINCH, NULL, top);
 476				sigaction(SIGWINCH, &act, NULL);
 477			} else {
 478				signal(SIGWINCH, SIG_DFL);
 479			}
 480			break;
 481		case 'E':
 482			if (top->evlist->nr_entries > 1) {
 483				/* Select 0 as the default event: */
 484				int counter = 0;
 485
 486				fprintf(stderr, "\nAvailable events:");
 487
 488				evlist__for_each_entry(top->evlist, top->sym_evsel)
 489					fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 490
 491				prompt_integer(&counter, "Enter details event counter");
 492
 493				if (counter >= top->evlist->nr_entries) {
 494					top->sym_evsel = perf_evlist__first(top->evlist);
 495					fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 496					sleep(1);
 497					break;
 498				}
 499				evlist__for_each_entry(top->evlist, top->sym_evsel)
 500					if (top->sym_evsel->idx == counter)
 501						break;
 502			} else
 503				top->sym_evsel = perf_evlist__first(top->evlist);
 504			break;
 505		case 'f':
 506			prompt_integer(&top->count_filter, "Enter display event count filter");
 507			break;
 508		case 'F':
 509			prompt_percent(&top->sym_pcnt_filter,
 510				       "Enter details display event filter (percent)");
 511			break;
 512		case 'K':
 513			top->hide_kernel_symbols = !top->hide_kernel_symbols;
 514			break;
 515		case 'q':
 516		case 'Q':
 517			printf("exiting.\n");
 518			if (top->dump_symtab)
 519				perf_session__fprintf_dsos(top->session, stderr);
 520			ret = false;
 521			break;
 522		case 's':
 523			perf_top__prompt_symbol(top, "Enter details symbol");
 524			break;
 525		case 'S':
 526			if (!top->sym_filter_entry)
 527				break;
 528			else {
 529				struct hist_entry *syme = top->sym_filter_entry;
 530
 531				top->sym_filter_entry = NULL;
 532				__zero_source_counters(syme);
 533			}
 534			break;
 535		case 'U':
 536			top->hide_user_symbols = !top->hide_user_symbols;
 537			break;
 538		case 'z':
 539			top->zero = !top->zero;
 540			break;
 541		default:
 542			break;
 543	}
 544
 545	return ret;
 546}
 547
 548static void perf_top__sort_new_samples(void *arg)
 549{
 550	struct perf_top *t = arg;
 551	struct perf_evsel *evsel = t->sym_evsel;
 552	struct hists *hists;
 553
 554	perf_top__reset_sample_counters(t);
 555
 556	if (t->evlist->selected != NULL)
 557		t->sym_evsel = t->evlist->selected;
 558
 559	hists = evsel__hists(evsel);
 560
 561	if (t->evlist->enabled) {
 562		if (t->zero) {
 563			hists__delete_entries(hists);
 564		} else {
 565			hists__decay_entries(hists, t->hide_user_symbols,
 566					     t->hide_kernel_symbols);
 567		}
 568	}
 569
 570	hists__collapse_resort(hists, NULL);
 571	perf_evsel__output_resort(evsel, NULL);
 572}
 573
 574static void *display_thread_tui(void *arg)
 575{
 576	struct perf_evsel *pos;
 577	struct perf_top *top = arg;
 578	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 579	struct hist_browser_timer hbt = {
 580		.timer		= perf_top__sort_new_samples,
 581		.arg		= top,
 582		.refresh	= top->delay_secs,
 583	};
 584
 585	perf_top__sort_new_samples(top);
 586
 587	/*
 588	 * Initialize the uid_filter_str, in the future the TUI will allow
 589	 * Zooming in/out UIDs. For now juse use whatever the user passed
 590	 * via --uid.
 591	 */
 592	evlist__for_each_entry(top->evlist, pos) {
 593		struct hists *hists = evsel__hists(pos);
 594		hists->uid_filter_str = top->record_opts.target.uid_str;
 595	}
 596
 597	perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 598				      top->min_percent,
 599				      &top->session->header.env);
 600
 601	done = 1;
 602	return NULL;
 603}
 604
 605static void display_sig(int sig __maybe_unused)
 606{
 607	done = 1;
 608}
 609
 610static void display_setup_sig(void)
 611{
 612	signal(SIGSEGV, sighandler_dump_stack);
 613	signal(SIGFPE, sighandler_dump_stack);
 614	signal(SIGINT,  display_sig);
 615	signal(SIGQUIT, display_sig);
 616	signal(SIGTERM, display_sig);
 617}
 618
 619static void *display_thread(void *arg)
 620{
 621	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 622	struct termios save;
 623	struct perf_top *top = arg;
 624	int delay_msecs, c;
 625
 626	display_setup_sig();
 627	pthread__unblock_sigwinch();
 628repeat:
 629	delay_msecs = top->delay_secs * MSEC_PER_SEC;
 630	set_term_quiet_input(&save);
 631	/* trash return*/
 632	getc(stdin);
 633
 634	while (!done) {
 635		perf_top__print_sym_table(top);
 636		/*
 637		 * Either timeout expired or we got an EINTR due to SIGWINCH,
 638		 * refresh screen in both cases.
 639		 */
 640		switch (poll(&stdin_poll, 1, delay_msecs)) {
 641		case 0:
 642			continue;
 643		case -1:
 644			if (errno == EINTR)
 645				continue;
 646			/* Fall trhu */
 647		default:
 648			c = getc(stdin);
 649			tcsetattr(0, TCSAFLUSH, &save);
 650
 651			if (perf_top__handle_keypress(top, c))
 652				goto repeat;
 653			done = 1;
 654		}
 655	}
 656
 657	tcsetattr(0, TCSAFLUSH, &save);
 658	return NULL;
 659}
 660
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 661static int hist_iter__top_callback(struct hist_entry_iter *iter,
 662				   struct addr_location *al, bool single,
 663				   void *arg)
 664{
 665	struct perf_top *top = arg;
 666	struct hist_entry *he = iter->he;
 667	struct perf_evsel *evsel = iter->evsel;
 668
 669	if (perf_hpp_list.sym && single)
 670		perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
 671
 672	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 673		     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 674	return 0;
 675}
 676
 677static void perf_event__process_sample(struct perf_tool *tool,
 678				       const union perf_event *event,
 679				       struct perf_evsel *evsel,
 680				       struct perf_sample *sample,
 681				       struct machine *machine)
 682{
 683	struct perf_top *top = container_of(tool, struct perf_top, tool);
 684	struct addr_location al;
 685	int err;
 686
 687	if (!machine && perf_guest) {
 688		static struct intlist *seen;
 689
 690		if (!seen)
 691			seen = intlist__new(NULL);
 692
 693		if (!intlist__has_entry(seen, sample->pid)) {
 694			pr_err("Can't find guest [%d]'s kernel information\n",
 695				sample->pid);
 696			intlist__add(seen, sample->pid);
 697		}
 698		return;
 699	}
 700
 701	if (!machine) {
 702		pr_err("%u unprocessable samples recorded.\r",
 703		       top->session->evlist->stats.nr_unprocessable_samples++);
 704		return;
 705	}
 706
 707	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 708		top->exact_samples++;
 709
 710	if (machine__resolve(machine, &al, sample) < 0)
 711		return;
 712
 713	if (!machine->kptr_restrict_warned &&
 714	    symbol_conf.kptr_restrict &&
 715	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
 716		ui__warning(
 717"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 718"Check /proc/sys/kernel/kptr_restrict.\n\n"
 719"Kernel%s samples will not be resolved.\n",
 720			  al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 721			  " modules" : "");
 722		if (use_browser <= 0)
 723			sleep(5);
 724		machine->kptr_restrict_warned = true;
 725	}
 726
 727	if (al.sym == NULL) {
 728		const char *msg = "Kernel samples will not be resolved.\n";
 729		/*
 730		 * As we do lazy loading of symtabs we only will know if the
 731		 * specified vmlinux file is invalid when we actually have a
 732		 * hit in kernel space and then try to load it. So if we get
 733		 * here and there are _no_ symbols in the DSO backing the
 734		 * kernel map, bail out.
 735		 *
 736		 * We may never get here, for instance, if we use -K/
 737		 * --hide-kernel-symbols, even if the user specifies an
 738		 * invalid --vmlinux ;-)
 739		 */
 740		if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 741		    al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 742		    RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 743			if (symbol_conf.vmlinux_name) {
 744				char serr[256];
 745				dso__strerror_load(al.map->dso, serr, sizeof(serr));
 746				ui__warning("The %s file can't be used: %s\n%s",
 747					    symbol_conf.vmlinux_name, serr, msg);
 748			} else {
 749				ui__warning("A vmlinux file was not found.\n%s",
 750					    msg);
 751			}
 752
 753			if (use_browser <= 0)
 754				sleep(5);
 755			top->vmlinux_warned = true;
 756		}
 757	}
 758
 759	if (al.sym == NULL || !al.sym->idle) {
 760		struct hists *hists = evsel__hists(evsel);
 761		struct hist_entry_iter iter = {
 762			.evsel		= evsel,
 763			.sample 	= sample,
 764			.add_entry_cb 	= hist_iter__top_callback,
 765		};
 766
 767		if (symbol_conf.cumulate_callchain)
 768			iter.ops = &hist_iter_cumulative;
 769		else
 770			iter.ops = &hist_iter_normal;
 771
 772		pthread_mutex_lock(&hists->lock);
 773
 774		err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 775		if (err < 0)
 776			pr_err("Problem incrementing symbol period, skipping event\n");
 777
 778		pthread_mutex_unlock(&hists->lock);
 779	}
 780
 781	addr_location__put(&al);
 782}
 783
 784static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 785{
 786	struct perf_sample sample;
 787	struct perf_evsel *evsel;
 788	struct perf_session *session = top->session;
 789	union perf_event *event;
 790	struct machine *machine;
 791	int ret;
 792
 793	while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 794		ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 795		if (ret) {
 796			pr_err("Can't parse sample, err = %d\n", ret);
 797			goto next_event;
 798		}
 799
 800		evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 801		assert(evsel != NULL);
 802
 803		if (event->header.type == PERF_RECORD_SAMPLE)
 804			++top->samples;
 805
 806		switch (sample.cpumode) {
 807		case PERF_RECORD_MISC_USER:
 808			++top->us_samples;
 809			if (top->hide_user_symbols)
 810				goto next_event;
 811			machine = &session->machines.host;
 812			break;
 813		case PERF_RECORD_MISC_KERNEL:
 814			++top->kernel_samples;
 815			if (top->hide_kernel_symbols)
 816				goto next_event;
 817			machine = &session->machines.host;
 818			break;
 819		case PERF_RECORD_MISC_GUEST_KERNEL:
 820			++top->guest_kernel_samples;
 821			machine = perf_session__find_machine(session,
 822							     sample.pid);
 823			break;
 824		case PERF_RECORD_MISC_GUEST_USER:
 825			++top->guest_us_samples;
 826			/*
 827			 * TODO: we don't process guest user from host side
 828			 * except simple counting.
 829			 */
 830			goto next_event;
 831		default:
 832			if (event->header.type == PERF_RECORD_SAMPLE)
 833				goto next_event;
 834			machine = &session->machines.host;
 835			break;
 836		}
 837
 838
 839		if (event->header.type == PERF_RECORD_SAMPLE) {
 840			perf_event__process_sample(&top->tool, event, evsel,
 841						   &sample, machine);
 842		} else if (event->header.type < PERF_RECORD_MAX) {
 843			hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 844			machine__process_event(machine, event, &sample);
 845		} else
 846			++session->evlist->stats.nr_unknown_events;
 847next_event:
 848		perf_evlist__mmap_consume(top->evlist, idx);
 849	}
 850}
 851
 852static void perf_top__mmap_read(struct perf_top *top)
 853{
 854	int i;
 855
 856	for (i = 0; i < top->evlist->nr_mmaps; i++)
 857		perf_top__mmap_read_idx(top, i);
 858}
 859
 860static int perf_top__start_counters(struct perf_top *top)
 861{
 862	char msg[512];
 863	struct perf_evsel *counter;
 864	struct perf_evlist *evlist = top->evlist;
 865	struct record_opts *opts = &top->record_opts;
 866
 867	perf_evlist__config(evlist, opts, &callchain_param);
 868
 869	evlist__for_each_entry(evlist, counter) {
 870try_again:
 871		if (perf_evsel__open(counter, top->evlist->cpus,
 872				     top->evlist->threads) < 0) {
 873			if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 874				if (verbose)
 875					ui__warning("%s\n", msg);
 876				goto try_again;
 877			}
 878
 879			perf_evsel__open_strerror(counter, &opts->target,
 880						  errno, msg, sizeof(msg));
 881			ui__error("%s\n", msg);
 882			goto out_err;
 883		}
 884	}
 885
 886	if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 887		ui__error("Failed to mmap with %d (%s)\n",
 888			    errno, str_error_r(errno, msg, sizeof(msg)));
 889		goto out_err;
 890	}
 891
 892	return 0;
 893
 894out_err:
 895	return -1;
 896}
 897
 898static int callchain_param__setup_sample_type(struct callchain_param *callchain)
 899{
 900	if (!perf_hpp_list.sym) {
 901		if (callchain->enabled) {
 902			ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 903			return -EINVAL;
 904		}
 905	} else if (callchain->mode != CHAIN_NONE) {
 906		if (callchain_register_param(callchain) < 0) {
 907			ui__error("Can't register callchain params.\n");
 908			return -EINVAL;
 909		}
 910	}
 911
 912	return 0;
 913}
 914
 915static int __cmd_top(struct perf_top *top)
 916{
 917	char msg[512];
 918	struct perf_evsel *pos;
 919	struct perf_evsel_config_term *err_term;
 920	struct perf_evlist *evlist = top->evlist;
 921	struct record_opts *opts = &top->record_opts;
 922	pthread_t thread;
 923	int ret;
 924
 925	top->session = perf_session__new(NULL, false, NULL);
 926	if (top->session == NULL)
 927		return -1;
 928
 
 
 929	if (!objdump_path) {
 930		ret = perf_env__lookup_objdump(&top->session->header.env);
 931		if (ret)
 932			goto out_delete;
 933	}
 934
 935	ret = callchain_param__setup_sample_type(&callchain_param);
 936	if (ret)
 937		goto out_delete;
 938
 939	if (perf_session__register_idle_thread(top->session) < 0)
 940		goto out_delete;
 941
 942	machine__synthesize_threads(&top->session->machines.host, &opts->target,
 943				    top->evlist->threads, false, opts->proc_map_timeout);
 944
 945	if (perf_hpp_list.socket) {
 946		ret = perf_env__read_cpu_topology_map(&perf_env);
 947		if (ret < 0)
 948			goto out_err_cpu_topo;
 949	}
 950
 951	ret = perf_top__start_counters(top);
 952	if (ret)
 953		goto out_delete;
 954
 955	ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
 956	if (ret) {
 957		error("failed to set config \"%s\" on event %s with %d (%s)\n",
 958			err_term->val.drv_cfg, perf_evsel__name(pos), errno,
 959			str_error_r(errno, msg, sizeof(msg)));
 960		goto out_delete;
 961	}
 962
 963	top->session->evlist = top->evlist;
 964	perf_session__set_id_hdr_size(top->session);
 965
 966	/*
 967	 * When perf is starting the traced process, all the events (apart from
 968	 * group members) have enable_on_exec=1 set, so don't spoil it by
 969	 * prematurely enabling them.
 970	 *
 971	 * XXX 'top' still doesn't start workloads like record, trace, but should,
 972	 * so leave the check here.
 973	 */
 974        if (!target__none(&opts->target))
 975                perf_evlist__enable(top->evlist);
 976
 977	/* Wait for a minimal set of events before starting the snapshot */
 978	perf_evlist__poll(top->evlist, 100);
 979
 980	perf_top__mmap_read(top);
 981
 982	ret = -1;
 983	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 984							    display_thread), top)) {
 985		ui__error("Could not create display thread.\n");
 986		goto out_delete;
 987	}
 988
 989	if (top->realtime_prio) {
 990		struct sched_param param;
 991
 992		param.sched_priority = top->realtime_prio;
 993		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 994			ui__error("Could not set realtime priority.\n");
 995			goto out_join;
 996		}
 997	}
 998
 999	while (!done) {
1000		u64 hits = top->samples;
1001
1002		perf_top__mmap_read(top);
1003
1004		if (hits == top->samples)
1005			ret = perf_evlist__poll(top->evlist, 100);
1006	}
1007
1008	ret = 0;
1009out_join:
1010	pthread_join(thread, NULL);
1011out_delete:
1012	perf_session__delete(top->session);
1013	top->session = NULL;
1014
1015	return ret;
1016
1017out_err_cpu_topo: {
1018	char errbuf[BUFSIZ];
1019	const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1020
1021	ui__error("Could not read the CPU topology map: %s\n", err);
1022	goto out_delete;
1023}
1024}
1025
1026static int
1027callchain_opt(const struct option *opt, const char *arg, int unset)
1028{
1029	symbol_conf.use_callchain = true;
1030	return record_callchain_opt(opt, arg, unset);
1031}
1032
1033static int
1034parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1035{
1036	struct callchain_param *callchain = opt->value;
1037
1038	callchain->enabled = !unset;
1039	callchain->record_mode = CALLCHAIN_FP;
 
1040
1041	/*
1042	 * --no-call-graph
1043	 */
1044	if (unset) {
1045		symbol_conf.use_callchain = false;
1046		callchain->record_mode = CALLCHAIN_NONE;
1047		return 0;
1048	}
1049
1050	return parse_callchain_top_opt(arg);
1051}
1052
1053static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1054{
1055	if (!strcmp(var, "top.call-graph"))
1056		var = "call-graph.record-mode"; /* fall-through */
1057	if (!strcmp(var, "top.children")) {
1058		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1059		return 0;
1060	}
1061
1062	return 0;
1063}
1064
1065static int
1066parse_percent_limit(const struct option *opt, const char *arg,
1067		    int unset __maybe_unused)
1068{
1069	struct perf_top *top = opt->value;
1070
1071	top->min_percent = strtof(arg, NULL);
1072	return 0;
1073}
1074
1075const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1076	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1077
1078int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1079{
1080	char errbuf[BUFSIZ];
1081	struct perf_top top = {
1082		.count_filter	     = 5,
1083		.delay_secs	     = 2,
1084		.record_opts = {
1085			.mmap_pages	= UINT_MAX,
1086			.user_freq	= UINT_MAX,
1087			.user_interval	= ULLONG_MAX,
1088			.freq		= 4000, /* 4 KHz */
1089			.target		= {
1090				.uses_mmap   = true,
1091			},
1092			.proc_map_timeout    = 500,
1093		},
1094		.max_stack	     = sysctl_perf_event_max_stack,
1095		.sym_pcnt_filter     = 5,
1096	};
1097	struct record_opts *opts = &top.record_opts;
1098	struct target *target = &opts->target;
1099	const struct option options[] = {
1100	OPT_CALLBACK('e', "event", &top.evlist, "event",
1101		     "event selector. use 'perf list' to list available events",
1102		     parse_events_option),
1103	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1104	OPT_STRING('p', "pid", &target->pid, "pid",
1105		    "profile events on existing process id"),
1106	OPT_STRING('t', "tid", &target->tid, "tid",
1107		    "profile events on existing thread id"),
1108	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1109			    "system-wide collection from all CPUs"),
1110	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1111		    "list of cpus to monitor"),
1112	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1113		   "file", "vmlinux pathname"),
1114	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1115		    "don't load vmlinux even if found"),
1116	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1117		    "hide kernel symbols"),
1118	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1119		     "number of mmap data pages",
1120		     perf_evlist__parse_mmap_pages),
1121	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1122		    "collect data with this RT SCHED_FIFO priority"),
1123	OPT_INTEGER('d', "delay", &top.delay_secs,
1124		    "number of seconds to delay between refreshes"),
1125	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1126			    "dump the symbol table used for profiling"),
1127	OPT_INTEGER('f', "count-filter", &top.count_filter,
1128		    "only display functions with more events than this"),
1129	OPT_BOOLEAN(0, "group", &opts->group,
1130			    "put the counters into a counter group"),
1131	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1132		    "child tasks do not inherit counters"),
1133	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1134		    "symbol to annotate"),
1135	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1136	OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1137	OPT_INTEGER('E', "entries", &top.print_entries,
1138		    "display this many functions"),
1139	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1140		    "hide user symbols"),
1141	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1142	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1143	OPT_INCR('v', "verbose", &verbose,
1144		    "be more verbose (show counter open errors, etc)"),
1145	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1146		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1147		   " Please refer the man page for the complete list."),
1148	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1149		   "output field(s): overhead, period, sample plus all of sort keys"),
1150	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1151		    "Show a column with the number of samples"),
1152	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1153			   NULL, "enables call-graph recording and display",
1154			   &callchain_opt),
1155	OPT_CALLBACK(0, "call-graph", &callchain_param,
1156		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1157		     top_callchain_help, &parse_callchain_opt),
1158	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1159		    "Accumulate callchains of children and show total overhead as well"),
1160	OPT_INTEGER(0, "max-stack", &top.max_stack,
1161		    "Set the maximum stack depth when parsing the callchain. "
1162		    "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1163	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1164		   "ignore callees of these functions in call graphs",
1165		   report_parse_ignore_callees_opt),
1166	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1167		    "Show a column with the sum of periods"),
1168	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1169		   "only consider symbols in these dsos"),
1170	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1171		   "only consider symbols in these comms"),
1172	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1173		   "only consider these symbols"),
1174	OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1175		    "Interleave source code with assembly code (default)"),
1176	OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1177		    "Display raw encoding of assembly instructions (default)"),
1178	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1179		    "Enable kernel symbol demangling"),
1180	OPT_STRING(0, "objdump", &objdump_path, "path",
1181		    "objdump binary to use for disassembly and annotations"),
1182	OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1183		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1184	OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1185	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1186		     "Don't show entries under that percent", parse_percent_limit),
1187	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1188		     "How to display percentage of filtered entries", parse_filter_percentage),
1189	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1190		   "width[,width...]",
1191		   "don't try to adjust column width, use these fixed values"),
1192	OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1193			"per thread proc mmap processing timeout in ms"),
1194	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1195		     "branch any", "sample any taken branches",
1196		     parse_branch_stack),
1197	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1198		     "branch filter mask", "branch stack filter modes",
1199		     parse_branch_stack),
1200	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1201		    "Show raw trace event output (do not use print fmt or plugins)"),
1202	OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1203		    "Show entries in a hierarchy"),
1204	OPT_END()
1205	};
1206	const char * const top_usage[] = {
1207		"perf top [<options>]",
1208		NULL
1209	};
1210	int status = hists__init();
1211
1212	if (status < 0)
1213		return status;
1214
1215	top.evlist = perf_evlist__new();
1216	if (top.evlist == NULL)
1217		return -ENOMEM;
1218
1219	perf_config(perf_top_config, &top);
1220
1221	argc = parse_options(argc, argv, options, top_usage, 0);
1222	if (argc)
1223		usage_with_options(top_usage, options);
1224
1225	if (!top.evlist->nr_entries &&
1226	    perf_evlist__add_default(top.evlist) < 0) {
1227		pr_err("Not enough memory for event selector list\n");
1228		goto out_delete_evlist;
1229	}
1230
1231	if (symbol_conf.report_hierarchy) {
1232		/* disable incompatible options */
1233		symbol_conf.event_group = false;
1234		symbol_conf.cumulate_callchain = false;
1235
1236		if (field_order) {
1237			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1238			parse_options_usage(top_usage, options, "fields", 0);
1239			parse_options_usage(NULL, options, "hierarchy", 0);
1240			goto out_delete_evlist;
1241		}
1242	}
1243
1244	sort__mode = SORT_MODE__TOP;
1245	/* display thread wants entries to be collapsed in a different tree */
1246	perf_hpp_list.need_collapse = 1;
1247
1248	if (top.use_stdio)
1249		use_browser = 0;
1250	else if (top.use_tui)
1251		use_browser = 1;
1252
1253	setup_browser(false);
1254
1255	if (setup_sorting(top.evlist) < 0) {
1256		if (sort_order)
1257			parse_options_usage(top_usage, options, "s", 1);
1258		if (field_order)
1259			parse_options_usage(sort_order ? NULL : top_usage,
1260					    options, "fields", 0);
1261		goto out_delete_evlist;
1262	}
1263
1264	status = target__validate(target);
1265	if (status) {
1266		target__strerror(target, status, errbuf, BUFSIZ);
1267		ui__warning("%s\n", errbuf);
1268	}
1269
1270	status = target__parse_uid(target);
1271	if (status) {
1272		int saved_errno = errno;
1273
1274		target__strerror(target, status, errbuf, BUFSIZ);
1275		ui__error("%s\n", errbuf);
1276
1277		status = -saved_errno;
1278		goto out_delete_evlist;
1279	}
1280
1281	if (target__none(target))
1282		target->system_wide = true;
1283
1284	if (perf_evlist__create_maps(top.evlist, target) < 0) {
1285		ui__error("Couldn't create thread/CPU maps: %s\n",
1286			  errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1287		goto out_delete_evlist;
1288	}
1289
1290	symbol_conf.nr_events = top.evlist->nr_entries;
1291
1292	if (top.delay_secs < 1)
1293		top.delay_secs = 1;
1294
1295	if (record_opts__config(opts)) {
1296		status = -EINVAL;
1297		goto out_delete_evlist;
1298	}
1299
1300	top.sym_evsel = perf_evlist__first(top.evlist);
1301
1302	if (!callchain_param.enabled) {
1303		symbol_conf.cumulate_callchain = false;
1304		perf_hpp__cancel_cumulate();
1305	}
1306
1307	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1308		callchain_param.order = ORDER_CALLER;
1309
1310	status = symbol__annotation_init();
1311	if (status < 0)
1312		goto out_delete_evlist;
1313
1314	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1315	if (symbol__init(NULL) < 0)
1316		return -1;
1317
1318	sort__setup_elide(stdout);
1319
1320	get_term_dimensions(&top.winsize);
1321	if (top.print_entries == 0) {
1322		struct sigaction act = {
1323			.sa_sigaction = perf_top__sig_winch,
1324			.sa_flags     = SA_SIGINFO,
1325		};
1326		perf_top__update_print_entries(&top);
1327		sigaction(SIGWINCH, &act, NULL);
1328	}
1329
1330	status = __cmd_top(&top);
1331
1332out_delete_evlist:
1333	perf_evlist__delete(top.evlist);
1334
1335	return status;
1336}