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/session.h"
30 #include "util/symbol.h"
31 #include "util/thread.h"
32 #include "util/thread_map.h"
34 #include "util/util.h"
35 #include <linux/rbtree.h>
36 #include "util/parse-options.h"
37 #include "util/parse-events.h"
38 #include "util/cpumap.h"
39 #include "util/xyarray.h"
40 #include "util/sort.h"
42 #include "util/debug.h"
56 #include <sys/syscall.h>
57 #include <sys/ioctl.h>
59 #include <sys/prctl.h>
64 #include <linux/unistd.h>
65 #include <linux/types.h>
68 void get_term_dimensions(struct winsize
*ws
)
70 char *s
= getenv("LINES");
74 s
= getenv("COLUMNS");
77 if (ws
->ws_row
&& ws
->ws_col
)
82 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
83 ws
->ws_row
&& ws
->ws_col
)
90 static void perf_top__update_print_entries(struct perf_top
*top
)
92 top
->print_entries
= top
->winsize
.ws_row
;
94 if (top
->print_entries
> 9)
95 top
->print_entries
-= 9;
98 static void perf_top__sig_winch(int sig __used
, siginfo_t
*info __used
, void *arg
)
100 struct perf_top
*top
= arg
;
102 get_term_dimensions(&top
->winsize
);
103 perf_top__update_print_entries(top
);
106 static int perf_top__parse_source(struct perf_top
*top
, struct hist_entry
*he
)
109 struct annotation
*notes
;
113 if (!he
|| !he
->ms
.sym
)
120 * We can't annotate with just /proc/kallsyms
122 if (map
->dso
->symtab_type
== SYMTAB__KALLSYMS
) {
123 pr_err("Can't annotate %s: No vmlinux file was found in the "
124 "path\n", sym
->name
);
129 notes
= symbol__annotation(sym
);
130 if (notes
->src
!= NULL
) {
131 pthread_mutex_lock(¬es
->lock
);
135 pthread_mutex_lock(¬es
->lock
);
137 if (symbol__alloc_hist(sym
) < 0) {
138 pthread_mutex_unlock(¬es
->lock
);
139 pr_err("Not enough memory for annotating '%s' symbol!\n",
145 err
= symbol__annotate(sym
, map
, 0);
148 top
->sym_filter_entry
= he
;
151 pthread_mutex_unlock(¬es
->lock
);
155 static void __zero_source_counters(struct hist_entry
*he
)
157 struct symbol
*sym
= he
->ms
.sym
;
158 symbol__annotate_zero_histograms(sym
);
161 static void perf_top__record_precise_ip(struct perf_top
*top
,
162 struct hist_entry
*he
,
165 struct annotation
*notes
;
168 if (he
== NULL
|| he
->ms
.sym
== NULL
||
169 ((top
->sym_filter_entry
== NULL
||
170 top
->sym_filter_entry
->ms
.sym
!= he
->ms
.sym
) && use_browser
!= 1))
174 notes
= symbol__annotation(sym
);
176 if (pthread_mutex_trylock(¬es
->lock
))
179 if (notes
->src
== NULL
&& symbol__alloc_hist(sym
) < 0) {
180 pthread_mutex_unlock(¬es
->lock
);
181 pr_err("Not enough memory for annotating '%s' symbol!\n",
187 ip
= he
->ms
.map
->map_ip(he
->ms
.map
, ip
);
188 symbol__inc_addr_samples(sym
, he
->ms
.map
, counter
, ip
);
190 pthread_mutex_unlock(¬es
->lock
);
193 static void perf_top__show_details(struct perf_top
*top
)
195 struct hist_entry
*he
= top
->sym_filter_entry
;
196 struct annotation
*notes
;
197 struct symbol
*symbol
;
204 notes
= symbol__annotation(symbol
);
206 pthread_mutex_lock(¬es
->lock
);
208 if (notes
->src
== NULL
)
211 printf("Showing %s for %s\n", event_name(top
->sym_evsel
), symbol
->name
);
212 printf(" Events Pcnt (>=%d%%)\n", top
->sym_pcnt_filter
);
214 more
= symbol__annotate_printf(symbol
, he
->ms
.map
, top
->sym_evsel
->idx
,
215 0, top
->sym_pcnt_filter
, top
->print_entries
, 4);
217 symbol__annotate_zero_histogram(symbol
, top
->sym_evsel
->idx
);
219 symbol__annotate_decay_histogram(symbol
, top
->sym_evsel
->idx
);
221 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
223 pthread_mutex_unlock(¬es
->lock
);
226 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
228 static struct hist_entry
*perf_evsel__add_hist_entry(struct perf_evsel
*evsel
,
229 struct addr_location
*al
,
230 struct perf_sample
*sample
)
232 struct hist_entry
*he
;
234 he
= __hists__add_entry(&evsel
->hists
, al
, NULL
, sample
->period
);
238 evsel
->hists
.stats
.total_period
+= sample
->period
;
239 hists__inc_nr_events(&evsel
->hists
, PERF_RECORD_SAMPLE
);
243 static void perf_top__print_sym_table(struct perf_top
*top
)
247 const int win_width
= top
->winsize
.ws_col
- 1;
251 perf_top__header_snprintf(top
, bf
, sizeof(bf
));
254 perf_top__reset_sample_counters(top
);
256 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
258 if (top
->sym_evsel
->hists
.stats
.nr_lost_warned
!=
259 top
->sym_evsel
->hists
.stats
.nr_events
[PERF_RECORD_LOST
]) {
260 top
->sym_evsel
->hists
.stats
.nr_lost_warned
=
261 top
->sym_evsel
->hists
.stats
.nr_events
[PERF_RECORD_LOST
];
262 color_fprintf(stdout
, PERF_COLOR_RED
,
263 "WARNING: LOST %d chunks, Check IO/CPU overload",
264 top
->sym_evsel
->hists
.stats
.nr_lost_warned
);
268 if (top
->sym_filter_entry
) {
269 perf_top__show_details(top
);
273 hists__collapse_resort_threaded(&top
->sym_evsel
->hists
);
274 hists__output_resort_threaded(&top
->sym_evsel
->hists
);
275 hists__decay_entries_threaded(&top
->sym_evsel
->hists
,
276 top
->hide_user_symbols
,
277 top
->hide_kernel_symbols
);
278 hists__output_recalc_col_len(&top
->sym_evsel
->hists
,
279 top
->winsize
.ws_row
- 3);
281 hists__fprintf(&top
->sym_evsel
->hists
, NULL
, false, false,
282 top
->winsize
.ws_row
- 4 - printed
, win_width
, stdout
);
285 static void prompt_integer(int *target
, const char *msg
)
287 char *buf
= malloc(0), *p
;
291 fprintf(stdout
, "\n%s: ", msg
);
292 if (getline(&buf
, &dummy
, stdin
) < 0)
295 p
= strchr(buf
, '\n');
305 tmp
= strtoul(buf
, NULL
, 10);
311 static void prompt_percent(int *target
, const char *msg
)
315 prompt_integer(&tmp
, msg
);
316 if (tmp
>= 0 && tmp
<= 100)
320 static void perf_top__prompt_symbol(struct perf_top
*top
, const char *msg
)
322 char *buf
= malloc(0), *p
;
323 struct hist_entry
*syme
= top
->sym_filter_entry
, *n
, *found
= NULL
;
324 struct rb_node
*next
;
327 /* zero counters of active symbol */
329 __zero_source_counters(syme
);
330 top
->sym_filter_entry
= NULL
;
333 fprintf(stdout
, "\n%s: ", msg
);
334 if (getline(&buf
, &dummy
, stdin
) < 0)
337 p
= strchr(buf
, '\n');
341 next
= rb_first(&top
->sym_evsel
->hists
.entries
);
343 n
= rb_entry(next
, struct hist_entry
, rb_node
);
344 if (n
->ms
.sym
&& !strcmp(buf
, n
->ms
.sym
->name
)) {
348 next
= rb_next(&n
->rb_node
);
352 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
356 perf_top__parse_source(top
, found
);
362 static void perf_top__print_mapped_keys(struct perf_top
*top
)
366 if (top
->sym_filter_entry
) {
367 struct symbol
*sym
= top
->sym_filter_entry
->ms
.sym
;
371 fprintf(stdout
, "\nMapped keys:\n");
372 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", top
->delay_secs
);
373 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", top
->print_entries
);
375 if (top
->evlist
->nr_entries
> 1)
376 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(top
->sym_evsel
));
378 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", top
->count_filter
);
380 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", top
->sym_pcnt_filter
);
381 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
382 fprintf(stdout
, "\t[S] stop annotation.\n");
385 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
386 top
->hide_kernel_symbols
? "yes" : "no");
388 "\t[U] hide user symbols. \t(%s)\n",
389 top
->hide_user_symbols
? "yes" : "no");
390 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", top
->zero
? 1 : 0);
391 fprintf(stdout
, "\t[qQ] quit.\n");
394 static int perf_top__key_mapped(struct perf_top
*top
, int c
)
410 return top
->evlist
->nr_entries
> 1 ? 1 : 0;
418 static void perf_top__handle_keypress(struct perf_top
*top
, int c
)
420 if (!perf_top__key_mapped(top
, c
)) {
421 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
422 struct termios tc
, save
;
424 perf_top__print_mapped_keys(top
);
425 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
430 tc
.c_lflag
&= ~(ICANON
| ECHO
);
433 tcsetattr(0, TCSANOW
, &tc
);
435 poll(&stdin_poll
, 1, -1);
438 tcsetattr(0, TCSAFLUSH
, &save
);
439 if (!perf_top__key_mapped(top
, c
))
445 prompt_integer(&top
->delay_secs
, "Enter display delay");
446 if (top
->delay_secs
< 1)
450 prompt_integer(&top
->print_entries
, "Enter display entries (lines)");
451 if (top
->print_entries
== 0) {
452 struct sigaction act
= {
453 .sa_sigaction
= perf_top__sig_winch
,
454 .sa_flags
= SA_SIGINFO
,
456 perf_top__sig_winch(SIGWINCH
, NULL
, top
);
457 sigaction(SIGWINCH
, &act
, NULL
);
459 signal(SIGWINCH
, SIG_DFL
);
462 if (top
->evlist
->nr_entries
> 1) {
463 /* Select 0 as the default event: */
466 fprintf(stderr
, "\nAvailable events:");
468 list_for_each_entry(top
->sym_evsel
, &top
->evlist
->entries
, node
)
469 fprintf(stderr
, "\n\t%d %s", top
->sym_evsel
->idx
, event_name(top
->sym_evsel
));
471 prompt_integer(&counter
, "Enter details event counter");
473 if (counter
>= top
->evlist
->nr_entries
) {
474 top
->sym_evsel
= list_entry(top
->evlist
->entries
.next
, struct perf_evsel
, node
);
475 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(top
->sym_evsel
));
479 list_for_each_entry(top
->sym_evsel
, &top
->evlist
->entries
, node
)
480 if (top
->sym_evsel
->idx
== counter
)
483 top
->sym_evsel
= list_entry(top
->evlist
->entries
.next
, struct perf_evsel
, node
);
486 prompt_integer(&top
->count_filter
, "Enter display event count filter");
489 prompt_percent(&top
->sym_pcnt_filter
,
490 "Enter details display event filter (percent)");
493 top
->hide_kernel_symbols
= !top
->hide_kernel_symbols
;
497 printf("exiting.\n");
498 if (top
->dump_symtab
)
499 perf_session__fprintf_dsos(top
->session
, stderr
);
502 perf_top__prompt_symbol(top
, "Enter details symbol");
505 if (!top
->sym_filter_entry
)
508 struct hist_entry
*syme
= top
->sym_filter_entry
;
510 top
->sym_filter_entry
= NULL
;
511 __zero_source_counters(syme
);
515 top
->hide_user_symbols
= !top
->hide_user_symbols
;
518 top
->zero
= !top
->zero
;
525 static void perf_top__sort_new_samples(void *arg
)
527 struct perf_top
*t
= arg
;
528 perf_top__reset_sample_counters(t
);
530 if (t
->evlist
->selected
!= NULL
)
531 t
->sym_evsel
= t
->evlist
->selected
;
533 hists__collapse_resort_threaded(&t
->sym_evsel
->hists
);
534 hists__output_resort_threaded(&t
->sym_evsel
->hists
);
535 hists__decay_entries_threaded(&t
->sym_evsel
->hists
,
536 t
->hide_user_symbols
,
537 t
->hide_kernel_symbols
);
540 static void *display_thread_tui(void *arg
)
542 struct perf_top
*top
= arg
;
543 const char *help
= "For a higher level overview, try: perf top --sort comm,dso";
545 perf_top__sort_new_samples(top
);
546 perf_evlist__tui_browse_hists(top
->evlist
, help
,
547 perf_top__sort_new_samples
,
548 top
, top
->delay_secs
);
555 static void *display_thread(void *arg
)
557 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
558 struct termios tc
, save
;
559 struct perf_top
*top
= arg
;
564 tc
.c_lflag
&= ~(ICANON
| ECHO
);
568 pthread__unblock_sigwinch();
570 delay_msecs
= top
->delay_secs
* 1000;
571 tcsetattr(0, TCSANOW
, &tc
);
576 perf_top__print_sym_table(top
);
578 * Either timeout expired or we got an EINTR due to SIGWINCH,
579 * refresh screen in both cases.
581 switch (poll(&stdin_poll
, 1, delay_msecs
)) {
594 tcsetattr(0, TCSAFLUSH
, &save
);
596 perf_top__handle_keypress(top
, c
);
602 /* Tag samples to be skipped. */
603 static const char *skip_symbols
[] = {
610 "mwait_idle_with_hints",
612 "ppc64_runlatch_off",
613 "pseries_dedicated_idle_sleep",
617 static int symbol_filter(struct map
*map __used
, struct symbol
*sym
)
619 const char *name
= sym
->name
;
623 * ppc64 uses function descriptors and appends a '.' to the
624 * start of every instruction address. Remove it.
629 if (!strcmp(name
, "_text") ||
630 !strcmp(name
, "_etext") ||
631 !strcmp(name
, "_sinittext") ||
632 !strncmp("init_module", name
, 11) ||
633 !strncmp("cleanup_module", name
, 14) ||
634 strstr(name
, "_text_start") ||
635 strstr(name
, "_text_end"))
638 for (i
= 0; skip_symbols
[i
]; i
++) {
639 if (!strcmp(skip_symbols
[i
], name
)) {
648 static void perf_event__process_sample(struct perf_tool
*tool
,
649 const union perf_event
*event
,
650 struct perf_evsel
*evsel
,
651 struct perf_sample
*sample
,
652 struct machine
*machine
)
654 struct perf_top
*top
= container_of(tool
, struct perf_top
, tool
);
655 struct symbol
*parent
= NULL
;
656 u64 ip
= event
->ip
.ip
;
657 struct addr_location al
;
660 if (!machine
&& perf_guest
) {
661 pr_err("Can't find guest [%d]'s kernel information\n",
666 if (event
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
667 top
->exact_samples
++;
669 if (perf_event__preprocess_sample(event
, machine
, &al
, sample
,
670 symbol_filter
) < 0 ||
674 if (!top
->kptr_restrict_warned
&&
675 symbol_conf
.kptr_restrict
&&
676 al
.cpumode
== PERF_RECORD_MISC_KERNEL
) {
678 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
679 "Check /proc/sys/kernel/kptr_restrict.\n\n"
680 "Kernel%s samples will not be resolved.\n",
681 !RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
]) ?
683 if (use_browser
<= 0)
685 top
->kptr_restrict_warned
= true;
688 if (al
.sym
== NULL
) {
689 const char *msg
= "Kernel samples will not be resolved.\n";
691 * As we do lazy loading of symtabs we only will know if the
692 * specified vmlinux file is invalid when we actually have a
693 * hit in kernel space and then try to load it. So if we get
694 * here and there are _no_ symbols in the DSO backing the
695 * kernel map, bail out.
697 * We may never get here, for instance, if we use -K/
698 * --hide-kernel-symbols, even if the user specifies an
699 * invalid --vmlinux ;-)
701 if (!top
->kptr_restrict_warned
&& !top
->vmlinux_warned
&&
702 al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
703 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
704 if (symbol_conf
.vmlinux_name
) {
705 ui__warning("The %s file can't be used.\n%s",
706 symbol_conf
.vmlinux_name
, msg
);
708 ui__warning("A vmlinux file was not found.\n%s",
712 if (use_browser
<= 0)
714 top
->vmlinux_warned
= true;
718 if (al
.sym
== NULL
|| !al
.sym
->ignore
) {
719 struct hist_entry
*he
;
721 if ((sort__has_parent
|| symbol_conf
.use_callchain
) &&
723 err
= machine__resolve_callchain(machine
, evsel
, al
.thread
,
724 sample
->callchain
, &parent
);
729 he
= perf_evsel__add_hist_entry(evsel
, &al
, sample
);
731 pr_err("Problem incrementing symbol period, skipping event\n");
735 if (symbol_conf
.use_callchain
) {
736 err
= callchain_append(he
->callchain
, &evsel
->hists
.callchain_cursor
,
742 if (top
->sort_has_symbols
)
743 perf_top__record_precise_ip(top
, he
, evsel
->idx
, ip
);
749 static void perf_top__mmap_read_idx(struct perf_top
*top
, int idx
)
751 struct perf_sample sample
;
752 struct perf_evsel
*evsel
;
753 struct perf_session
*session
= top
->session
;
754 union perf_event
*event
;
755 struct machine
*machine
;
759 while ((event
= perf_evlist__mmap_read(top
->evlist
, idx
)) != NULL
) {
760 ret
= perf_session__parse_sample(session
, event
, &sample
);
762 pr_err("Can't parse sample, err = %d\n", ret
);
766 evsel
= perf_evlist__id2evsel(session
->evlist
, sample
.id
);
767 assert(evsel
!= NULL
);
769 origin
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
771 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
775 case PERF_RECORD_MISC_USER
:
777 if (top
->hide_user_symbols
)
779 machine
= perf_session__find_host_machine(session
);
781 case PERF_RECORD_MISC_KERNEL
:
782 ++top
->kernel_samples
;
783 if (top
->hide_kernel_symbols
)
785 machine
= perf_session__find_host_machine(session
);
787 case PERF_RECORD_MISC_GUEST_KERNEL
:
788 ++top
->guest_kernel_samples
;
789 machine
= perf_session__find_machine(session
, event
->ip
.pid
);
791 case PERF_RECORD_MISC_GUEST_USER
:
792 ++top
->guest_us_samples
;
794 * TODO: we don't process guest user from host side
795 * except simple counting.
803 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
804 perf_event__process_sample(&top
->tool
, event
, evsel
,
806 } else if (event
->header
.type
< PERF_RECORD_MAX
) {
807 hists__inc_nr_events(&evsel
->hists
, event
->header
.type
);
808 perf_event__process(&top
->tool
, event
, &sample
, machine
);
810 ++session
->hists
.stats
.nr_unknown_events
;
814 static void perf_top__mmap_read(struct perf_top
*top
)
818 for (i
= 0; i
< top
->evlist
->nr_mmaps
; i
++)
819 perf_top__mmap_read_idx(top
, i
);
822 static void perf_top__start_counters(struct perf_top
*top
)
824 struct perf_evsel
*counter
, *first
;
825 struct perf_evlist
*evlist
= top
->evlist
;
827 first
= list_entry(evlist
->entries
.next
, struct perf_evsel
, node
);
829 list_for_each_entry(counter
, &evlist
->entries
, node
) {
830 struct perf_event_attr
*attr
= &counter
->attr
;
831 struct xyarray
*group_fd
= NULL
;
833 if (top
->group
&& counter
!= first
)
834 group_fd
= first
->fd
;
836 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
839 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
841 attr
->sample_freq
= top
->freq
;
844 if (evlist
->nr_entries
> 1) {
845 attr
->sample_type
|= PERF_SAMPLE_ID
;
846 attr
->read_format
|= PERF_FORMAT_ID
;
849 if (symbol_conf
.use_callchain
)
850 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
854 attr
->inherit
= top
->inherit
;
856 attr
->sample_id_all
= top
->sample_id_all_avail
? 1 : 0;
858 if (perf_evsel__open(counter
, top
->evlist
->cpus
,
859 top
->evlist
->threads
, top
->group
,
863 if (err
== EPERM
|| err
== EACCES
) {
864 ui__error_paranoid();
866 } else if (err
== EINVAL
&& top
->sample_id_all_avail
) {
868 * Old kernel, no attr->sample_id_type_all field
870 top
->sample_id_all_avail
= false;
871 goto retry_sample_id
;
874 * If it's cycles then fall back to hrtimer
875 * based cpu-clock-tick sw counter, which
876 * is always available even if no PMU support:
878 if (attr
->type
== PERF_TYPE_HARDWARE
&&
879 attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
881 ui__warning("Cycles event not supported,\n"
882 "trying to fall back to cpu-clock-ticks\n");
884 attr
->type
= PERF_TYPE_SOFTWARE
;
885 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
890 ui__warning("The %s event is not supported.\n",
891 event_name(counter
));
895 ui__warning("The sys_perf_event_open() syscall "
896 "returned with %d (%s). /bin/dmesg "
897 "may provide additional information.\n"
898 "No CONFIG_PERF_EVENTS=y kernel support "
899 "configured?\n", err
, strerror(err
));
904 if (perf_evlist__mmap(evlist
, top
->mmap_pages
, false) < 0) {
905 ui__warning("Failed to mmap with %d (%s)\n",
906 errno
, strerror(errno
));
917 static int perf_top__setup_sample_type(struct perf_top
*top
)
919 if (!top
->sort_has_symbols
) {
920 if (symbol_conf
.use_callchain
) {
921 ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
924 } else if (!top
->dont_use_callchains
&& callchain_param
.mode
!= CHAIN_NONE
) {
925 if (callchain_register_param(&callchain_param
) < 0) {
926 ui__warning("Can't register callchain params.\n");
934 static int __cmd_top(struct perf_top
*top
)
939 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
940 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
942 top
->session
= perf_session__new(NULL
, O_WRONLY
, false, false, NULL
);
943 if (top
->session
== NULL
)
946 ret
= perf_top__setup_sample_type(top
);
950 if (top
->target_tid
!= -1)
951 perf_event__synthesize_thread_map(&top
->tool
, top
->evlist
->threads
,
953 &top
->session
->host_machine
);
955 perf_event__synthesize_threads(&top
->tool
, perf_event__process
,
956 &top
->session
->host_machine
);
957 perf_top__start_counters(top
);
958 top
->session
->evlist
= top
->evlist
;
959 perf_session__update_sample_type(top
->session
);
961 /* Wait for a minimal set of events before starting the snapshot */
962 poll(top
->evlist
->pollfd
, top
->evlist
->nr_fds
, 100);
964 perf_top__mmap_read(top
);
966 if (pthread_create(&thread
, NULL
, (use_browser
> 0 ? display_thread_tui
:
967 display_thread
), top
)) {
968 printf("Could not create display thread.\n");
972 if (top
->realtime_prio
) {
973 struct sched_param param
;
975 param
.sched_priority
= top
->realtime_prio
;
976 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
977 printf("Could not set realtime priority.\n");
983 u64 hits
= top
->samples
;
985 perf_top__mmap_read(top
);
987 if (hits
== top
->samples
)
988 ret
= poll(top
->evlist
->pollfd
, top
->evlist
->nr_fds
, 100);
992 perf_session__delete(top
->session
);
999 parse_callchain_opt(const struct option
*opt
, const char *arg
, int unset
)
1001 struct perf_top
*top
= (struct perf_top
*)opt
->value
;
1009 top
->dont_use_callchains
= true;
1013 symbol_conf
.use_callchain
= true;
1018 tok
= strtok((char *)arg
, ",");
1022 /* get the output mode */
1023 if (!strncmp(tok
, "graph", strlen(arg
)))
1024 callchain_param
.mode
= CHAIN_GRAPH_ABS
;
1026 else if (!strncmp(tok
, "flat", strlen(arg
)))
1027 callchain_param
.mode
= CHAIN_FLAT
;
1029 else if (!strncmp(tok
, "fractal", strlen(arg
)))
1030 callchain_param
.mode
= CHAIN_GRAPH_REL
;
1032 else if (!strncmp(tok
, "none", strlen(arg
))) {
1033 callchain_param
.mode
= CHAIN_NONE
;
1034 symbol_conf
.use_callchain
= false;
1042 /* get the min percentage */
1043 tok
= strtok(NULL
, ",");
1047 callchain_param
.min_percent
= strtod(tok
, &endptr
);
1051 /* get the print limit */
1052 tok2
= strtok(NULL
, ",");
1056 if (tok2
[0] != 'c') {
1057 callchain_param
.print_limit
= strtod(tok2
, &endptr
);
1058 tok2
= strtok(NULL
, ",");
1063 /* get the call chain order */
1064 if (!strcmp(tok2
, "caller"))
1065 callchain_param
.order
= ORDER_CALLER
;
1066 else if (!strcmp(tok2
, "callee"))
1067 callchain_param
.order
= ORDER_CALLEE
;
1071 if (callchain_register_param(&callchain_param
) < 0) {
1072 fprintf(stderr
, "Can't register callchain params\n");
1078 static const char * const top_usage
[] = {
1079 "perf top [<options>]",
1083 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1085 struct perf_evsel
*pos
;
1086 int status
= -ENOMEM
;
1087 struct perf_top top
= {
1092 .freq
= 1000, /* 1 KHz */
1093 .sample_id_all_avail
= true,
1095 .sym_pcnt_filter
= 5,
1097 char callchain_default_opt
[] = "fractal,0.5,callee";
1098 const struct option options
[] = {
1099 OPT_CALLBACK('e', "event", &top
.evlist
, "event",
1100 "event selector. use 'perf list' to list available events",
1101 parse_events_option
),
1102 OPT_INTEGER('c', "count", &top
.default_interval
,
1103 "event period to sample"),
1104 OPT_INTEGER('p', "pid", &top
.target_pid
,
1105 "profile events on existing process id"),
1106 OPT_INTEGER('t', "tid", &top
.target_tid
,
1107 "profile events on existing thread id"),
1108 OPT_BOOLEAN('a', "all-cpus", &top
.system_wide
,
1109 "system-wide collection from all CPUs"),
1110 OPT_STRING('C', "cpu", &top
.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('K', "hide_kernel_symbols", &top
.hide_kernel_symbols
,
1115 "hide kernel symbols"),
1116 OPT_UINTEGER('m', "mmap-pages", &top
.mmap_pages
, "number of mmap data pages"),
1117 OPT_INTEGER('r', "realtime", &top
.realtime_prio
,
1118 "collect data with this RT SCHED_FIFO priority"),
1119 OPT_INTEGER('d', "delay", &top
.delay_secs
,
1120 "number of seconds to delay between refreshes"),
1121 OPT_BOOLEAN('D', "dump-symtab", &top
.dump_symtab
,
1122 "dump the symbol table used for profiling"),
1123 OPT_INTEGER('f', "count-filter", &top
.count_filter
,
1124 "only display functions with more events than this"),
1125 OPT_BOOLEAN('g', "group", &top
.group
,
1126 "put the counters into a counter group"),
1127 OPT_BOOLEAN('i', "inherit", &top
.inherit
,
1128 "child tasks inherit counters"),
1129 OPT_STRING(0, "sym-annotate", &top
.sym_filter
, "symbol name",
1130 "symbol to annotate"),
1131 OPT_BOOLEAN('z', "zero", &top
.zero
,
1132 "zero history across updates"),
1133 OPT_INTEGER('F', "freq", &top
.freq
,
1134 "profile at this frequency"),
1135 OPT_INTEGER('E', "entries", &top
.print_entries
,
1136 "display this many functions"),
1137 OPT_BOOLEAN('U', "hide_user_symbols", &top
.hide_user_symbols
,
1138 "hide user symbols"),
1139 OPT_BOOLEAN(0, "tui", &top
.use_tui
, "Use the TUI interface"),
1140 OPT_BOOLEAN(0, "stdio", &top
.use_stdio
, "Use the stdio interface"),
1141 OPT_INCR('v', "verbose", &verbose
,
1142 "be more verbose (show counter open errors, etc)"),
1143 OPT_STRING('s', "sort", &sort_order
, "key[,key2...]",
1144 "sort by key(s): pid, comm, dso, symbol, parent"),
1145 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf
.show_nr_samples
,
1146 "Show a column with the number of samples"),
1147 OPT_CALLBACK_DEFAULT('G', "call-graph", &top
, "output_type,min_percent, call_order",
1148 "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1149 "Default: fractal,0.5,callee", &parse_callchain_opt
,
1150 callchain_default_opt
),
1151 OPT_BOOLEAN(0, "show-total-period", &symbol_conf
.show_total_period
,
1152 "Show a column with the sum of periods"),
1153 OPT_STRING(0, "dsos", &symbol_conf
.dso_list_str
, "dso[,dso...]",
1154 "only consider symbols in these dsos"),
1155 OPT_STRING(0, "comms", &symbol_conf
.comm_list_str
, "comm[,comm...]",
1156 "only consider symbols in these comms"),
1157 OPT_STRING(0, "symbols", &symbol_conf
.sym_list_str
, "symbol[,symbol...]",
1158 "only consider these symbols"),
1159 OPT_BOOLEAN(0, "source", &symbol_conf
.annotate_src
,
1160 "Interleave source code with assembly code (default)"),
1161 OPT_BOOLEAN(0, "asm-raw", &symbol_conf
.annotate_asm_raw
,
1162 "Display raw encoding of assembly instructions (default)"),
1163 OPT_STRING('M', "disassembler-style", &disassembler_style
, "disassembler style",
1164 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1168 top
.evlist
= perf_evlist__new(NULL
, NULL
);
1169 if (top
.evlist
== NULL
)
1172 symbol_conf
.exclude_other
= false;
1174 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1176 usage_with_options(top_usage
, options
);
1178 if (sort_order
== default_sort_order
)
1179 sort_order
= "dso,symbol";
1181 setup_sorting(top_usage
, options
);
1185 else if (top
.use_tui
)
1188 setup_browser(false);
1190 /* CPU and PID are mutually exclusive */
1191 if (top
.target_tid
> 0 && top
.cpu_list
) {
1192 printf("WARNING: PID switch overriding CPU\n");
1194 top
.cpu_list
= NULL
;
1197 if (top
.target_pid
!= -1)
1198 top
.target_tid
= top
.target_pid
;
1200 if (perf_evlist__create_maps(top
.evlist
, top
.target_pid
,
1201 top
.target_tid
, top
.cpu_list
) < 0)
1202 usage_with_options(top_usage
, options
);
1204 if (!top
.evlist
->nr_entries
&&
1205 perf_evlist__add_default(top
.evlist
) < 0) {
1206 pr_err("Not enough memory for event selector list\n");
1210 symbol_conf
.nr_events
= top
.evlist
->nr_entries
;
1212 if (top
.delay_secs
< 1)
1216 * User specified count overrides default frequency.
1218 if (top
.default_interval
)
1220 else if (top
.freq
) {
1221 top
.default_interval
= top
.freq
;
1223 fprintf(stderr
, "frequency and count are zero, aborting\n");
1227 list_for_each_entry(pos
, &top
.evlist
->entries
, node
) {
1228 if (perf_evsel__alloc_fd(pos
, top
.evlist
->cpus
->nr
,
1229 top
.evlist
->threads
->nr
) < 0)
1232 * Fill in the ones not specifically initialized via -c:
1234 if (pos
->attr
.sample_period
)
1237 pos
->attr
.sample_period
= top
.default_interval
;
1240 if (perf_evlist__alloc_pollfd(top
.evlist
) < 0 ||
1241 perf_evlist__alloc_mmap(top
.evlist
) < 0)
1244 top
.sym_evsel
= list_entry(top
.evlist
->entries
.next
, struct perf_evsel
, node
);
1246 symbol_conf
.priv_size
= sizeof(struct annotation
);
1248 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1249 if (symbol__init() < 0)
1252 sort_entry__setup_elide(&sort_dso
, symbol_conf
.dso_list
, "dso", stdout
);
1253 sort_entry__setup_elide(&sort_comm
, symbol_conf
.comm_list
, "comm", stdout
);
1254 sort_entry__setup_elide(&sort_sym
, symbol_conf
.sym_list
, "symbol", stdout
);
1257 * Avoid annotation data structures overhead when symbols aren't on the
1260 top
.sort_has_symbols
= sort_sym
.list
.next
!= NULL
;
1262 get_term_dimensions(&top
.winsize
);
1263 if (top
.print_entries
== 0) {
1264 struct sigaction act
= {
1265 .sa_sigaction
= perf_top__sig_winch
,
1266 .sa_flags
= SA_SIGINFO
,
1268 perf_top__update_print_entries(&top
);
1269 sigaction(SIGWINCH
, &act
, NULL
);
1272 status
= __cmd_top(&top
);
1274 perf_evlist__delete(top
.evlist
);