4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10 * Improvements and fixes by:
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>
18 * Released under the GPL v2. (and only v2, not any later version)
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"
35 #include "util/util.h"
36 #include <linux/rbtree.h>
37 #include "util/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"
44 #include "util/debug.h"
59 #include <sys/syscall.h>
60 #include <sys/ioctl.h>
62 #include <sys/prctl.h>
65 #include <sys/utsname.h>
68 #include <linux/unistd.h>
69 #include <linux/types.h>
71 void get_term_dimensions(struct winsize
*ws
)
73 char *s
= getenv("LINES");
77 s
= getenv("COLUMNS");
80 if (ws
->ws_row
&& ws
->ws_col
)
85 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
86 ws
->ws_row
&& ws
->ws_col
)
93 static void perf_top__update_print_entries(struct perf_top
*top
)
95 if (top
->print_entries
> 9)
96 top
->print_entries
-= 9;
99 static void perf_top__sig_winch(int sig __maybe_unused
,
100 siginfo_t
*info __maybe_unused
, void *arg
)
102 struct perf_top
*top
= arg
;
104 get_term_dimensions(&top
->winsize
);
105 if (!top
->print_entries
106 || (top
->print_entries
+4) > top
->winsize
.ws_row
) {
107 top
->print_entries
= top
->winsize
.ws_row
;
109 top
->print_entries
+= 4;
110 top
->winsize
.ws_row
= top
->print_entries
;
112 perf_top__update_print_entries(top
);
115 static int perf_top__parse_source(struct perf_top
*top
, struct hist_entry
*he
)
118 struct annotation
*notes
;
122 if (!he
|| !he
->ms
.sym
)
129 * We can't annotate with just /proc/kallsyms
131 if (map
->dso
->symtab_type
== DSO_BINARY_TYPE__KALLSYMS
) {
132 pr_err("Can't annotate %s: No vmlinux file was found in the "
133 "path\n", sym
->name
);
138 notes
= symbol__annotation(sym
);
139 if (notes
->src
!= NULL
) {
140 pthread_mutex_lock(¬es
->lock
);
144 pthread_mutex_lock(¬es
->lock
);
146 if (symbol__alloc_hist(sym
) < 0) {
147 pthread_mutex_unlock(¬es
->lock
);
148 pr_err("Not enough memory for annotating '%s' symbol!\n",
154 err
= symbol__annotate(sym
, map
, 0);
157 top
->sym_filter_entry
= he
;
160 pthread_mutex_unlock(¬es
->lock
);
164 static void __zero_source_counters(struct hist_entry
*he
)
166 struct symbol
*sym
= he
->ms
.sym
;
167 symbol__annotate_zero_histograms(sym
);
170 static void ui__warn_map_erange(struct map
*map
, struct symbol
*sym
, u64 ip
)
173 int err
= uname(&uts
);
175 ui__warning("Out of bounds address found:\n\n"
176 "Addr: %" PRIx64
"\n"
178 "Map: %" PRIx64
"-%" PRIx64
"\n"
179 "Symbol: %" PRIx64
"-%" PRIx64
" %c %s\n"
183 "Not all samples will be on the annotation output.\n\n"
184 "Please report to linux-kernel@vger.kernel.org\n",
185 ip
, map
->dso
->long_name
, dso__symtab_origin(map
->dso
),
186 map
->start
, map
->end
, sym
->start
, sym
->end
,
187 sym
->binding
== STB_GLOBAL
? 'g' :
188 sym
->binding
== STB_LOCAL
? 'l' : 'w', sym
->name
,
189 err
? "[unknown]" : uts
.machine
,
190 err
? "[unknown]" : uts
.release
, perf_version_string
);
191 if (use_browser
<= 0)
194 map
->erange_warned
= true;
197 static void perf_top__record_precise_ip(struct perf_top
*top
,
198 struct hist_entry
*he
,
201 struct annotation
*notes
;
205 if (he
== NULL
|| he
->ms
.sym
== NULL
||
206 ((top
->sym_filter_entry
== NULL
||
207 top
->sym_filter_entry
->ms
.sym
!= he
->ms
.sym
) && use_browser
!= 1))
211 notes
= symbol__annotation(sym
);
213 if (pthread_mutex_trylock(¬es
->lock
))
216 if (notes
->src
== NULL
&& symbol__alloc_hist(sym
) < 0) {
217 pthread_mutex_unlock(¬es
->lock
);
218 pr_err("Not enough memory for annotating '%s' symbol!\n",
224 ip
= he
->ms
.map
->map_ip(he
->ms
.map
, ip
);
225 err
= symbol__inc_addr_samples(sym
, he
->ms
.map
, counter
, ip
);
227 pthread_mutex_unlock(¬es
->lock
);
229 if (err
== -ERANGE
&& !he
->ms
.map
->erange_warned
)
230 ui__warn_map_erange(he
->ms
.map
, sym
, ip
);
233 static void perf_top__show_details(struct perf_top
*top
)
235 struct hist_entry
*he
= top
->sym_filter_entry
;
236 struct annotation
*notes
;
237 struct symbol
*symbol
;
244 notes
= symbol__annotation(symbol
);
246 pthread_mutex_lock(¬es
->lock
);
248 if (notes
->src
== NULL
)
251 printf("Showing %s for %s\n", perf_evsel__name(top
->sym_evsel
), symbol
->name
);
252 printf(" Events Pcnt (>=%d%%)\n", top
->sym_pcnt_filter
);
254 more
= symbol__annotate_printf(symbol
, he
->ms
.map
, top
->sym_evsel
->idx
,
255 0, top
->sym_pcnt_filter
, top
->print_entries
, 4);
257 symbol__annotate_zero_histogram(symbol
, top
->sym_evsel
->idx
);
259 symbol__annotate_decay_histogram(symbol
, top
->sym_evsel
->idx
);
261 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
263 pthread_mutex_unlock(¬es
->lock
);
266 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
268 static struct hist_entry
*perf_evsel__add_hist_entry(struct perf_evsel
*evsel
,
269 struct addr_location
*al
,
270 struct perf_sample
*sample
)
272 struct hist_entry
*he
;
274 he
= __hists__add_entry(&evsel
->hists
, al
, NULL
, sample
->period
);
278 hists__inc_nr_events(&evsel
->hists
, PERF_RECORD_SAMPLE
);
282 static void perf_top__print_sym_table(struct perf_top
*top
)
286 const int win_width
= top
->winsize
.ws_col
- 1;
290 perf_top__header_snprintf(top
, bf
, sizeof(bf
));
293 perf_top__reset_sample_counters(top
);
295 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
297 if (top
->sym_evsel
->hists
.stats
.nr_lost_warned
!=
298 top
->sym_evsel
->hists
.stats
.nr_events
[PERF_RECORD_LOST
]) {
299 top
->sym_evsel
->hists
.stats
.nr_lost_warned
=
300 top
->sym_evsel
->hists
.stats
.nr_events
[PERF_RECORD_LOST
];
301 color_fprintf(stdout
, PERF_COLOR_RED
,
302 "WARNING: LOST %d chunks, Check IO/CPU overload",
303 top
->sym_evsel
->hists
.stats
.nr_lost_warned
);
307 if (top
->sym_filter_entry
) {
308 perf_top__show_details(top
);
312 hists__collapse_resort_threaded(&top
->sym_evsel
->hists
);
313 hists__output_resort_threaded(&top
->sym_evsel
->hists
);
314 hists__decay_entries_threaded(&top
->sym_evsel
->hists
,
315 top
->hide_user_symbols
,
316 top
->hide_kernel_symbols
);
317 hists__output_recalc_col_len(&top
->sym_evsel
->hists
,
318 top
->winsize
.ws_row
- 3);
320 hists__fprintf(&top
->sym_evsel
->hists
, false,
321 top
->winsize
.ws_row
- 4 - printed
, win_width
, stdout
);
324 static void prompt_integer(int *target
, const char *msg
)
326 char *buf
= malloc(0), *p
;
330 fprintf(stdout
, "\n%s: ", msg
);
331 if (getline(&buf
, &dummy
, stdin
) < 0)
334 p
= strchr(buf
, '\n');
344 tmp
= strtoul(buf
, NULL
, 10);
350 static void prompt_percent(int *target
, const char *msg
)
354 prompt_integer(&tmp
, msg
);
355 if (tmp
>= 0 && tmp
<= 100)
359 static void perf_top__prompt_symbol(struct perf_top
*top
, const char *msg
)
361 char *buf
= malloc(0), *p
;
362 struct hist_entry
*syme
= top
->sym_filter_entry
, *n
, *found
= NULL
;
363 struct rb_node
*next
;
366 /* zero counters of active symbol */
368 __zero_source_counters(syme
);
369 top
->sym_filter_entry
= NULL
;
372 fprintf(stdout
, "\n%s: ", msg
);
373 if (getline(&buf
, &dummy
, stdin
) < 0)
376 p
= strchr(buf
, '\n');
380 next
= rb_first(&top
->sym_evsel
->hists
.entries
);
382 n
= rb_entry(next
, struct hist_entry
, rb_node
);
383 if (n
->ms
.sym
&& !strcmp(buf
, n
->ms
.sym
->name
)) {
387 next
= rb_next(&n
->rb_node
);
391 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
394 perf_top__parse_source(top
, found
);
400 static void perf_top__print_mapped_keys(struct perf_top
*top
)
404 if (top
->sym_filter_entry
) {
405 struct symbol
*sym
= top
->sym_filter_entry
->ms
.sym
;
409 fprintf(stdout
, "\nMapped keys:\n");
410 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", top
->delay_secs
);
411 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", top
->print_entries
);
413 if (top
->evlist
->nr_entries
> 1)
414 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top
->sym_evsel
));
416 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", top
->count_filter
);
418 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", top
->sym_pcnt_filter
);
419 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
420 fprintf(stdout
, "\t[S] stop annotation.\n");
423 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
424 top
->hide_kernel_symbols
? "yes" : "no");
426 "\t[U] hide user symbols. \t(%s)\n",
427 top
->hide_user_symbols
? "yes" : "no");
428 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", top
->zero
? 1 : 0);
429 fprintf(stdout
, "\t[qQ] quit.\n");
432 static int perf_top__key_mapped(struct perf_top
*top
, int c
)
448 return top
->evlist
->nr_entries
> 1 ? 1 : 0;
456 static void perf_top__handle_keypress(struct perf_top
*top
, int c
)
458 if (!perf_top__key_mapped(top
, c
)) {
459 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
460 struct termios tc
, save
;
462 perf_top__print_mapped_keys(top
);
463 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
468 tc
.c_lflag
&= ~(ICANON
| ECHO
);
471 tcsetattr(0, TCSANOW
, &tc
);
473 poll(&stdin_poll
, 1, -1);
476 tcsetattr(0, TCSAFLUSH
, &save
);
477 if (!perf_top__key_mapped(top
, c
))
483 prompt_integer(&top
->delay_secs
, "Enter display delay");
484 if (top
->delay_secs
< 1)
488 prompt_integer(&top
->print_entries
, "Enter display entries (lines)");
489 if (top
->print_entries
== 0) {
490 struct sigaction act
= {
491 .sa_sigaction
= perf_top__sig_winch
,
492 .sa_flags
= SA_SIGINFO
,
494 perf_top__sig_winch(SIGWINCH
, NULL
, top
);
495 sigaction(SIGWINCH
, &act
, NULL
);
497 perf_top__sig_winch(SIGWINCH
, NULL
, top
);
498 signal(SIGWINCH
, SIG_DFL
);
502 if (top
->evlist
->nr_entries
> 1) {
503 /* Select 0 as the default event: */
506 fprintf(stderr
, "\nAvailable events:");
508 list_for_each_entry(top
->sym_evsel
, &top
->evlist
->entries
, node
)
509 fprintf(stderr
, "\n\t%d %s", top
->sym_evsel
->idx
, perf_evsel__name(top
->sym_evsel
));
511 prompt_integer(&counter
, "Enter details event counter");
513 if (counter
>= top
->evlist
->nr_entries
) {
514 top
->sym_evsel
= perf_evlist__first(top
->evlist
);
515 fprintf(stderr
, "Sorry, no such event, using %s.\n", perf_evsel__name(top
->sym_evsel
));
519 list_for_each_entry(top
->sym_evsel
, &top
->evlist
->entries
, node
)
520 if (top
->sym_evsel
->idx
== counter
)
523 top
->sym_evsel
= perf_evlist__first(top
->evlist
);
526 prompt_integer(&top
->count_filter
, "Enter display event count filter");
529 prompt_percent(&top
->sym_pcnt_filter
,
530 "Enter details display event filter (percent)");
533 top
->hide_kernel_symbols
= !top
->hide_kernel_symbols
;
537 printf("exiting.\n");
538 if (top
->dump_symtab
)
539 perf_session__fprintf_dsos(top
->session
, stderr
);
542 perf_top__prompt_symbol(top
, "Enter details symbol");
545 if (!top
->sym_filter_entry
)
548 struct hist_entry
*syme
= top
->sym_filter_entry
;
550 top
->sym_filter_entry
= NULL
;
551 __zero_source_counters(syme
);
555 top
->hide_user_symbols
= !top
->hide_user_symbols
;
558 top
->zero
= !top
->zero
;
565 static void perf_top__sort_new_samples(void *arg
)
567 struct perf_top
*t
= arg
;
568 perf_top__reset_sample_counters(t
);
570 if (t
->evlist
->selected
!= NULL
)
571 t
->sym_evsel
= t
->evlist
->selected
;
573 hists__collapse_resort_threaded(&t
->sym_evsel
->hists
);
574 hists__output_resort_threaded(&t
->sym_evsel
->hists
);
575 hists__decay_entries_threaded(&t
->sym_evsel
->hists
,
576 t
->hide_user_symbols
,
577 t
->hide_kernel_symbols
);
580 static void *display_thread_tui(void *arg
)
582 struct perf_evsel
*pos
;
583 struct perf_top
*top
= arg
;
584 const char *help
= "For a higher level overview, try: perf top --sort comm,dso";
585 struct hist_browser_timer hbt
= {
586 .timer
= perf_top__sort_new_samples
,
588 .refresh
= top
->delay_secs
,
591 perf_top__sort_new_samples(top
);
594 * Initialize the uid_filter_str, in the future the TUI will allow
595 * Zooming in/out UIDs. For now juse use whatever the user passed
598 list_for_each_entry(pos
, &top
->evlist
->entries
, node
)
599 pos
->hists
.uid_filter_str
= top
->target
.uid_str
;
601 perf_evlist__tui_browse_hists(top
->evlist
, help
, &hbt
,
602 &top
->session
->header
.env
);
609 static void *display_thread(void *arg
)
611 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
612 struct termios tc
, save
;
613 struct perf_top
*top
= arg
;
618 tc
.c_lflag
&= ~(ICANON
| ECHO
);
622 pthread__unblock_sigwinch();
624 delay_msecs
= top
->delay_secs
* 1000;
625 tcsetattr(0, TCSANOW
, &tc
);
630 perf_top__print_sym_table(top
);
632 * Either timeout expired or we got an EINTR due to SIGWINCH,
633 * refresh screen in both cases.
635 switch (poll(&stdin_poll
, 1, delay_msecs
)) {
648 tcsetattr(0, TCSAFLUSH
, &save
);
650 perf_top__handle_keypress(top
, c
);
656 /* Tag samples to be skipped. */
657 static const char *skip_symbols
[] = {
665 "mwait_idle_with_hints",
667 "ppc64_runlatch_off",
668 "pseries_dedicated_idle_sleep",
672 static int symbol_filter(struct map
*map __maybe_unused
, struct symbol
*sym
)
674 const char *name
= sym
->name
;
678 * ppc64 uses function descriptors and appends a '.' to the
679 * start of every instruction address. Remove it.
684 if (!strcmp(name
, "_text") ||
685 !strcmp(name
, "_etext") ||
686 !strcmp(name
, "_sinittext") ||
687 !strncmp("init_module", name
, 11) ||
688 !strncmp("cleanup_module", name
, 14) ||
689 strstr(name
, "_text_start") ||
690 strstr(name
, "_text_end"))
693 for (i
= 0; skip_symbols
[i
]; i
++) {
694 if (!strcmp(skip_symbols
[i
], name
)) {
703 static void perf_event__process_sample(struct perf_tool
*tool
,
704 const union perf_event
*event
,
705 struct perf_evsel
*evsel
,
706 struct perf_sample
*sample
,
707 struct machine
*machine
)
709 struct perf_top
*top
= container_of(tool
, struct perf_top
, tool
);
710 struct symbol
*parent
= NULL
;
711 u64 ip
= event
->ip
.ip
;
712 struct addr_location al
;
715 if (!machine
&& perf_guest
) {
716 static struct intlist
*seen
;
719 seen
= intlist__new();
721 if (!intlist__has_entry(seen
, event
->ip
.pid
)) {
722 pr_err("Can't find guest [%d]'s kernel information\n",
724 intlist__add(seen
, event
->ip
.pid
);
730 pr_err("%u unprocessable samples recorded.\n",
731 top
->session
->hists
.stats
.nr_unprocessable_samples
++);
735 if (event
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
736 top
->exact_samples
++;
738 if (perf_event__preprocess_sample(event
, machine
, &al
, sample
,
739 symbol_filter
) < 0 ||
743 if (!top
->kptr_restrict_warned
&&
744 symbol_conf
.kptr_restrict
&&
745 al
.cpumode
== PERF_RECORD_MISC_KERNEL
) {
747 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
748 "Check /proc/sys/kernel/kptr_restrict.\n\n"
749 "Kernel%s samples will not be resolved.\n",
750 !RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
]) ?
752 if (use_browser
<= 0)
754 top
->kptr_restrict_warned
= true;
757 if (al
.sym
== NULL
) {
758 const char *msg
= "Kernel samples will not be resolved.\n";
760 * As we do lazy loading of symtabs we only will know if the
761 * specified vmlinux file is invalid when we actually have a
762 * hit in kernel space and then try to load it. So if we get
763 * here and there are _no_ symbols in the DSO backing the
764 * kernel map, bail out.
766 * We may never get here, for instance, if we use -K/
767 * --hide-kernel-symbols, even if the user specifies an
768 * invalid --vmlinux ;-)
770 if (!top
->kptr_restrict_warned
&& !top
->vmlinux_warned
&&
771 al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
772 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
773 if (symbol_conf
.vmlinux_name
) {
774 ui__warning("The %s file can't be used.\n%s",
775 symbol_conf
.vmlinux_name
, msg
);
777 ui__warning("A vmlinux file was not found.\n%s",
781 if (use_browser
<= 0)
783 top
->vmlinux_warned
= true;
787 if (al
.sym
== NULL
|| !al
.sym
->ignore
) {
788 struct hist_entry
*he
;
790 if ((sort__has_parent
|| symbol_conf
.use_callchain
) &&
792 err
= machine__resolve_callchain(machine
, evsel
,
800 he
= perf_evsel__add_hist_entry(evsel
, &al
, sample
);
802 pr_err("Problem incrementing symbol period, skipping event\n");
806 if (symbol_conf
.use_callchain
) {
807 err
= callchain_append(he
->callchain
, &callchain_cursor
,
813 if (top
->sort_has_symbols
)
814 perf_top__record_precise_ip(top
, he
, evsel
->idx
, ip
);
820 static void perf_top__mmap_read_idx(struct perf_top
*top
, int idx
)
822 struct perf_sample sample
;
823 struct perf_evsel
*evsel
;
824 struct perf_session
*session
= top
->session
;
825 union perf_event
*event
;
826 struct machine
*machine
;
830 while ((event
= perf_evlist__mmap_read(top
->evlist
, idx
)) != NULL
) {
831 ret
= perf_evlist__parse_sample(top
->evlist
, event
, &sample
);
833 pr_err("Can't parse sample, err = %d\n", ret
);
837 evsel
= perf_evlist__id2evsel(session
->evlist
, sample
.id
);
838 assert(evsel
!= NULL
);
840 origin
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
842 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
846 case PERF_RECORD_MISC_USER
:
848 if (top
->hide_user_symbols
)
850 machine
= perf_session__find_host_machine(session
);
852 case PERF_RECORD_MISC_KERNEL
:
853 ++top
->kernel_samples
;
854 if (top
->hide_kernel_symbols
)
856 machine
= perf_session__find_host_machine(session
);
858 case PERF_RECORD_MISC_GUEST_KERNEL
:
859 ++top
->guest_kernel_samples
;
860 machine
= perf_session__find_machine(session
, event
->ip
.pid
);
862 case PERF_RECORD_MISC_GUEST_USER
:
863 ++top
->guest_us_samples
;
865 * TODO: we don't process guest user from host side
866 * except simple counting.
874 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
875 perf_event__process_sample(&top
->tool
, event
, evsel
,
877 } else if (event
->header
.type
< PERF_RECORD_MAX
) {
878 hists__inc_nr_events(&evsel
->hists
, event
->header
.type
);
879 machine__process_event(machine
, event
);
881 ++session
->hists
.stats
.nr_unknown_events
;
885 static void perf_top__mmap_read(struct perf_top
*top
)
889 for (i
= 0; i
< top
->evlist
->nr_mmaps
; i
++)
890 perf_top__mmap_read_idx(top
, i
);
893 static void perf_top__start_counters(struct perf_top
*top
)
895 struct perf_evsel
*counter
;
896 struct perf_evlist
*evlist
= top
->evlist
;
899 perf_evlist__set_leader(evlist
);
901 list_for_each_entry(counter
, &evlist
->entries
, node
) {
902 struct perf_event_attr
*attr
= &counter
->attr
;
904 perf_evsel__set_sample_bit(counter
, IP
);
905 perf_evsel__set_sample_bit(counter
, TID
);
908 perf_evsel__set_sample_bit(counter
, PERIOD
);
910 attr
->sample_freq
= top
->freq
;
913 if (evlist
->nr_entries
> 1)
914 perf_evsel__set_sample_id(counter
);
916 if (perf_target__has_cpu(&top
->target
))
917 perf_evsel__set_sample_bit(counter
, CPU
);
919 if (symbol_conf
.use_callchain
)
920 perf_evsel__set_sample_bit(counter
, CALLCHAIN
);
924 attr
->inherit
= top
->inherit
;
925 fallback_missing_features
:
926 if (top
->exclude_guest_missing
)
927 attr
->exclude_guest
= attr
->exclude_host
= 0;
929 attr
->sample_id_all
= top
->sample_id_all_missing
? 0 : 1;
931 if (perf_evsel__open(counter
, top
->evlist
->cpus
,
932 top
->evlist
->threads
) < 0) {
935 if (err
== EPERM
|| err
== EACCES
) {
936 ui__error_paranoid();
938 } else if (err
== EINVAL
) {
939 if (!top
->exclude_guest_missing
&&
940 (attr
->exclude_guest
|| attr
->exclude_host
)) {
941 pr_debug("Old kernel, cannot exclude "
942 "guest or host samples.\n");
943 top
->exclude_guest_missing
= true;
944 goto fallback_missing_features
;
945 } else if (!top
->sample_id_all_missing
) {
947 * Old kernel, no attr->sample_id_type_all field
949 top
->sample_id_all_missing
= true;
950 goto retry_sample_id
;
954 * If it's cycles then fall back to hrtimer
955 * based cpu-clock-tick sw counter, which
956 * is always available even if no PMU support:
958 if ((err
== ENOENT
|| err
== ENXIO
) &&
959 (attr
->type
== PERF_TYPE_HARDWARE
) &&
960 (attr
->config
== PERF_COUNT_HW_CPU_CYCLES
)) {
963 ui__warning("Cycles event not supported,\n"
964 "trying to fall back to cpu-clock-ticks\n");
966 attr
->type
= PERF_TYPE_SOFTWARE
;
967 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
970 counter
->name
= NULL
;
976 ui__error("The %s event is not supported.\n",
977 perf_evsel__name(counter
));
979 } else if (err
== EMFILE
) {
980 ui__error("Too many events are opened.\n"
981 "Try again after reducing the number of events\n");
983 } else if ((err
== EOPNOTSUPP
) && (attr
->precise_ip
)) {
984 ui__error("\'precise\' request may not be supported. "
985 "Try removing 'p' modifier\n");
989 ui__error("The sys_perf_event_open() syscall "
990 "returned with %d (%s). /bin/dmesg "
991 "may provide additional information.\n"
992 "No CONFIG_PERF_EVENTS=y kernel support "
993 "configured?\n", err
, strerror(err
));
998 if (perf_evlist__mmap(evlist
, top
->mmap_pages
, false) < 0) {
999 ui__error("Failed to mmap with %d (%s)\n",
1000 errno
, strerror(errno
));
1011 static int perf_top__setup_sample_type(struct perf_top
*top
)
1013 if (!top
->sort_has_symbols
) {
1014 if (symbol_conf
.use_callchain
) {
1015 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1018 } else if (!top
->dont_use_callchains
&& callchain_param
.mode
!= CHAIN_NONE
) {
1019 if (callchain_register_param(&callchain_param
) < 0) {
1020 ui__error("Can't register callchain params.\n");
1028 static int __cmd_top(struct perf_top
*top
)
1033 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1034 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1036 top
->session
= perf_session__new(NULL
, O_WRONLY
, false, false, NULL
);
1037 if (top
->session
== NULL
)
1040 ret
= perf_top__setup_sample_type(top
);
1044 if (perf_target__has_task(&top
->target
))
1045 perf_event__synthesize_thread_map(&top
->tool
, top
->evlist
->threads
,
1046 perf_event__process
,
1047 &top
->session
->host_machine
);
1049 perf_event__synthesize_threads(&top
->tool
, perf_event__process
,
1050 &top
->session
->host_machine
);
1051 perf_top__start_counters(top
);
1052 top
->session
->evlist
= top
->evlist
;
1053 perf_session__set_id_hdr_size(top
->session
);
1055 /* Wait for a minimal set of events before starting the snapshot */
1056 poll(top
->evlist
->pollfd
, top
->evlist
->nr_fds
, 100);
1058 perf_top__mmap_read(top
);
1060 if (pthread_create(&thread
, NULL
, (use_browser
> 0 ? display_thread_tui
:
1061 display_thread
), top
)) {
1062 ui__error("Could not create display thread.\n");
1066 if (top
->realtime_prio
) {
1067 struct sched_param param
;
1069 param
.sched_priority
= top
->realtime_prio
;
1070 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1071 ui__error("Could not set realtime priority.\n");
1077 u64 hits
= top
->samples
;
1079 perf_top__mmap_read(top
);
1081 if (hits
== top
->samples
)
1082 ret
= poll(top
->evlist
->pollfd
, top
->evlist
->nr_fds
, 100);
1086 perf_session__delete(top
->session
);
1087 top
->session
= NULL
;
1093 parse_callchain_opt(const struct option
*opt
, const char *arg
, int unset
)
1095 struct perf_top
*top
= (struct perf_top
*)opt
->value
;
1103 top
->dont_use_callchains
= true;
1107 symbol_conf
.use_callchain
= true;
1112 tok
= strtok((char *)arg
, ",");
1116 /* get the output mode */
1117 if (!strncmp(tok
, "graph", strlen(arg
)))
1118 callchain_param
.mode
= CHAIN_GRAPH_ABS
;
1120 else if (!strncmp(tok
, "flat", strlen(arg
)))
1121 callchain_param
.mode
= CHAIN_FLAT
;
1123 else if (!strncmp(tok
, "fractal", strlen(arg
)))
1124 callchain_param
.mode
= CHAIN_GRAPH_REL
;
1126 else if (!strncmp(tok
, "none", strlen(arg
))) {
1127 callchain_param
.mode
= CHAIN_NONE
;
1128 symbol_conf
.use_callchain
= false;
1134 /* get the min percentage */
1135 tok
= strtok(NULL
, ",");
1139 callchain_param
.min_percent
= strtod(tok
, &endptr
);
1143 /* get the print limit */
1144 tok2
= strtok(NULL
, ",");
1148 if (tok2
[0] != 'c') {
1149 callchain_param
.print_limit
= strtod(tok2
, &endptr
);
1150 tok2
= strtok(NULL
, ",");
1155 /* get the call chain order */
1156 if (!strcmp(tok2
, "caller"))
1157 callchain_param
.order
= ORDER_CALLER
;
1158 else if (!strcmp(tok2
, "callee"))
1159 callchain_param
.order
= ORDER_CALLEE
;
1163 if (callchain_register_param(&callchain_param
) < 0) {
1164 fprintf(stderr
, "Can't register callchain params\n");
1170 int cmd_top(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1172 struct perf_evsel
*pos
;
1174 char errbuf
[BUFSIZ
];
1175 struct perf_top top
= {
1178 .freq
= 4000, /* 4 KHz */
1180 .sym_pcnt_filter
= 5,
1185 char callchain_default_opt
[] = "fractal,0.5,callee";
1186 const struct option options
[] = {
1187 OPT_CALLBACK('e', "event", &top
.evlist
, "event",
1188 "event selector. use 'perf list' to list available events",
1189 parse_events_option
),
1190 OPT_INTEGER('c', "count", &top
.default_interval
,
1191 "event period to sample"),
1192 OPT_STRING('p', "pid", &top
.target
.pid
, "pid",
1193 "profile events on existing process id"),
1194 OPT_STRING('t', "tid", &top
.target
.tid
, "tid",
1195 "profile events on existing thread id"),
1196 OPT_BOOLEAN('a', "all-cpus", &top
.target
.system_wide
,
1197 "system-wide collection from all CPUs"),
1198 OPT_STRING('C', "cpu", &top
.target
.cpu_list
, "cpu",
1199 "list of cpus to monitor"),
1200 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1201 "file", "vmlinux pathname"),
1202 OPT_BOOLEAN('K', "hide_kernel_symbols", &top
.hide_kernel_symbols
,
1203 "hide kernel symbols"),
1204 OPT_UINTEGER('m', "mmap-pages", &top
.mmap_pages
, "number of mmap data pages"),
1205 OPT_INTEGER('r', "realtime", &top
.realtime_prio
,
1206 "collect data with this RT SCHED_FIFO priority"),
1207 OPT_INTEGER('d', "delay", &top
.delay_secs
,
1208 "number of seconds to delay between refreshes"),
1209 OPT_BOOLEAN('D', "dump-symtab", &top
.dump_symtab
,
1210 "dump the symbol table used for profiling"),
1211 OPT_INTEGER('f', "count-filter", &top
.count_filter
,
1212 "only display functions with more events than this"),
1213 OPT_BOOLEAN('g', "group", &top
.group
,
1214 "put the counters into a counter group"),
1215 OPT_BOOLEAN('i', "inherit", &top
.inherit
,
1216 "child tasks inherit counters"),
1217 OPT_STRING(0, "sym-annotate", &top
.sym_filter
, "symbol name",
1218 "symbol to annotate"),
1219 OPT_BOOLEAN('z', "zero", &top
.zero
,
1220 "zero history across updates"),
1221 OPT_INTEGER('F', "freq", &top
.freq
,
1222 "profile at this frequency"),
1223 OPT_INTEGER('E', "entries", &top
.print_entries
,
1224 "display this many functions"),
1225 OPT_BOOLEAN('U', "hide_user_symbols", &top
.hide_user_symbols
,
1226 "hide user symbols"),
1227 OPT_BOOLEAN(0, "tui", &top
.use_tui
, "Use the TUI interface"),
1228 OPT_BOOLEAN(0, "stdio", &top
.use_stdio
, "Use the stdio interface"),
1229 OPT_INCR('v', "verbose", &verbose
,
1230 "be more verbose (show counter open errors, etc)"),
1231 OPT_STRING('s', "sort", &sort_order
, "key[,key2...]",
1232 "sort by key(s): pid, comm, dso, symbol, parent"),
1233 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf
.show_nr_samples
,
1234 "Show a column with the number of samples"),
1235 OPT_CALLBACK_DEFAULT('G', "call-graph", &top
, "output_type,min_percent, call_order",
1236 "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1237 "Default: fractal,0.5,callee", &parse_callchain_opt
,
1238 callchain_default_opt
),
1239 OPT_BOOLEAN(0, "show-total-period", &symbol_conf
.show_total_period
,
1240 "Show a column with the sum of periods"),
1241 OPT_STRING(0, "dsos", &symbol_conf
.dso_list_str
, "dso[,dso...]",
1242 "only consider symbols in these dsos"),
1243 OPT_STRING(0, "comms", &symbol_conf
.comm_list_str
, "comm[,comm...]",
1244 "only consider symbols in these comms"),
1245 OPT_STRING(0, "symbols", &symbol_conf
.sym_list_str
, "symbol[,symbol...]",
1246 "only consider these symbols"),
1247 OPT_BOOLEAN(0, "source", &symbol_conf
.annotate_src
,
1248 "Interleave source code with assembly code (default)"),
1249 OPT_BOOLEAN(0, "asm-raw", &symbol_conf
.annotate_asm_raw
,
1250 "Display raw encoding of assembly instructions (default)"),
1251 OPT_STRING('M', "disassembler-style", &disassembler_style
, "disassembler style",
1252 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1253 OPT_STRING('u', "uid", &top
.target
.uid_str
, "user", "user to profile"),
1256 const char * const top_usage
[] = {
1257 "perf top [<options>]",
1261 top
.evlist
= perf_evlist__new(NULL
, NULL
);
1262 if (top
.evlist
== NULL
)
1265 symbol_conf
.exclude_other
= false;
1267 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1269 usage_with_options(top_usage
, options
);
1271 if (sort_order
== default_sort_order
)
1272 sort_order
= "dso,symbol";
1274 setup_sorting(top_usage
, options
);
1278 else if (top
.use_tui
)
1281 setup_browser(false);
1283 status
= perf_target__validate(&top
.target
);
1285 perf_target__strerror(&top
.target
, status
, errbuf
, BUFSIZ
);
1286 ui__warning("%s", errbuf
);
1289 status
= perf_target__parse_uid(&top
.target
);
1291 int saved_errno
= errno
;
1293 perf_target__strerror(&top
.target
, status
, errbuf
, BUFSIZ
);
1294 ui__error("%s", errbuf
);
1296 status
= -saved_errno
;
1297 goto out_delete_evlist
;
1300 if (perf_target__none(&top
.target
))
1301 top
.target
.system_wide
= true;
1303 if (perf_evlist__create_maps(top
.evlist
, &top
.target
) < 0)
1304 usage_with_options(top_usage
, options
);
1306 if (!top
.evlist
->nr_entries
&&
1307 perf_evlist__add_default(top
.evlist
) < 0) {
1308 ui__error("Not enough memory for event selector list\n");
1312 symbol_conf
.nr_events
= top
.evlist
->nr_entries
;
1314 if (top
.delay_secs
< 1)
1318 * User specified count overrides default frequency.
1320 if (top
.default_interval
)
1322 else if (top
.freq
) {
1323 top
.default_interval
= top
.freq
;
1325 ui__error("frequency and count are zero, aborting\n");
1329 list_for_each_entry(pos
, &top
.evlist
->entries
, node
) {
1331 * Fill in the ones not specifically initialized via -c:
1333 if (!pos
->attr
.sample_period
)
1334 pos
->attr
.sample_period
= top
.default_interval
;
1337 top
.sym_evsel
= perf_evlist__first(top
.evlist
);
1339 symbol_conf
.priv_size
= sizeof(struct annotation
);
1341 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1342 if (symbol__init() < 0)
1345 sort_entry__setup_elide(&sort_dso
, symbol_conf
.dso_list
, "dso", stdout
);
1346 sort_entry__setup_elide(&sort_comm
, symbol_conf
.comm_list
, "comm", stdout
);
1347 sort_entry__setup_elide(&sort_sym
, symbol_conf
.sym_list
, "symbol", stdout
);
1350 * Avoid annotation data structures overhead when symbols aren't on the
1353 top
.sort_has_symbols
= sort_sym
.list
.next
!= NULL
;
1355 get_term_dimensions(&top
.winsize
);
1356 if (top
.print_entries
== 0) {
1357 struct sigaction act
= {
1358 .sa_sigaction
= perf_top__sig_winch
,
1359 .sa_flags
= SA_SIGINFO
,
1361 perf_top__update_print_entries(&top
);
1362 sigaction(SIGWINCH
, &act
, NULL
);
1365 status
= __cmd_top(&top
);
1368 perf_evlist__delete(top
.evlist
);