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/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/event.h"
31 #include "util/machine.h"
32 #include "util/session.h"
33 #include "util/symbol.h"
34 #include "util/thread.h"
35 #include "util/thread_map.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/term.h"
44 #include "util/intlist.h"
45 #include "util/parse-branch-options.h"
46 #include "arch/common.h"
48 #include "util/debug.h"
64 #include <sys/syscall.h>
65 #include <sys/ioctl.h>
67 #include <sys/prctl.h>
70 #include <sys/utsname.h>
73 #include <linux/stringify.h>
74 #include <linux/time64.h>
75 #include <linux/types.h>
77 #include "sane_ctype.h"
79 static volatile int done
;
80 static volatile int resize
;
82 #define HEADER_LINE_NR 5
84 static void perf_top__update_print_entries(struct perf_top
*top
)
86 top
->print_entries
= top
->winsize
.ws_row
- HEADER_LINE_NR
;
89 static void winch_sig(int sig __maybe_unused
)
94 static void perf_top__resize(struct perf_top
*top
)
96 get_term_dimensions(&top
->winsize
);
97 perf_top__update_print_entries(top
);
100 static int perf_top__parse_source(struct perf_top
*top
, struct hist_entry
*he
)
102 struct perf_evsel
*evsel
= hists_to_evsel(he
->hists
);
104 struct annotation
*notes
;
108 if (!he
|| !he
->ms
.sym
)
115 * We can't annotate with just /proc/kallsyms
117 if (map
->dso
->symtab_type
== DSO_BINARY_TYPE__KALLSYMS
&&
118 !dso__is_kcore(map
->dso
)) {
119 pr_err("Can't annotate %s: No vmlinux file was found in the "
120 "path\n", sym
->name
);
125 notes
= symbol__annotation(sym
);
126 if (notes
->src
!= NULL
) {
127 pthread_mutex_lock(¬es
->lock
);
131 pthread_mutex_lock(¬es
->lock
);
133 if (symbol__alloc_hist(sym
) < 0) {
134 pthread_mutex_unlock(¬es
->lock
);
135 pr_err("Not enough memory for annotating '%s' symbol!\n",
141 err
= symbol__annotate(sym
, map
, evsel
, 0, NULL
);
144 top
->sym_filter_entry
= he
;
147 symbol__strerror_disassemble(sym
, map
, err
, msg
, sizeof(msg
));
148 pr_err("Couldn't annotate %s: %s\n", sym
->name
, msg
);
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 ui__warn_map_erange(struct map
*map
, struct symbol
*sym
, u64 ip
)
164 int err
= uname(&uts
);
166 ui__warning("Out of bounds address found:\n\n"
167 "Addr: %" PRIx64
"\n"
169 "Map: %" PRIx64
"-%" PRIx64
"\n"
170 "Symbol: %" PRIx64
"-%" PRIx64
" %c %s\n"
174 "Not all samples will be on the annotation output.\n\n"
175 "Please report to linux-kernel@vger.kernel.org\n",
176 ip
, map
->dso
->long_name
, dso__symtab_origin(map
->dso
),
177 map
->start
, map
->end
, sym
->start
, sym
->end
,
178 sym
->binding
== STB_GLOBAL
? 'g' :
179 sym
->binding
== STB_LOCAL
? 'l' : 'w', sym
->name
,
180 err
? "[unknown]" : uts
.machine
,
181 err
? "[unknown]" : uts
.release
, perf_version_string
);
182 if (use_browser
<= 0)
185 map
->erange_warned
= true;
188 static void perf_top__record_precise_ip(struct perf_top
*top
,
189 struct hist_entry
*he
,
190 struct perf_sample
*sample
,
193 struct annotation
*notes
;
194 struct symbol
*sym
= he
->ms
.sym
;
197 if (sym
== NULL
|| (use_browser
== 0 &&
198 (top
->sym_filter_entry
== NULL
||
199 top
->sym_filter_entry
->ms
.sym
!= sym
)))
202 notes
= symbol__annotation(sym
);
204 if (pthread_mutex_trylock(¬es
->lock
))
207 err
= hist_entry__inc_addr_samples(he
, sample
, counter
, ip
);
209 pthread_mutex_unlock(¬es
->lock
);
213 * This function is now called with he->hists->lock held.
214 * Release it before going to sleep.
216 pthread_mutex_unlock(&he
->hists
->lock
);
218 if (err
== -ERANGE
&& !he
->ms
.map
->erange_warned
)
219 ui__warn_map_erange(he
->ms
.map
, sym
, ip
);
220 else if (err
== -ENOMEM
) {
221 pr_err("Not enough memory for annotating '%s' symbol!\n",
226 pthread_mutex_lock(&he
->hists
->lock
);
230 static void perf_top__show_details(struct perf_top
*top
)
232 struct hist_entry
*he
= top
->sym_filter_entry
;
233 struct perf_evsel
*evsel
= hists_to_evsel(he
->hists
);
234 struct annotation
*notes
;
235 struct symbol
*symbol
;
242 notes
= symbol__annotation(symbol
);
244 pthread_mutex_lock(¬es
->lock
);
246 symbol__calc_percent(symbol
, evsel
);
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
,
255 0, top
->sym_pcnt_filter
, top
->print_entries
, 4);
257 if (top
->evlist
->enabled
) {
259 symbol__annotate_zero_histogram(symbol
, top
->sym_evsel
->idx
);
261 symbol__annotate_decay_histogram(symbol
, top
->sym_evsel
->idx
);
264 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
266 pthread_mutex_unlock(¬es
->lock
);
269 static void perf_top__print_sym_table(struct perf_top
*top
)
273 const int win_width
= top
->winsize
.ws_col
- 1;
274 struct perf_evsel
*evsel
= top
->sym_evsel
;
275 struct hists
*hists
= evsel__hists(evsel
);
279 perf_top__header_snprintf(top
, bf
, sizeof(bf
));
282 perf_top__reset_sample_counters(top
);
284 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
286 if (!top
->record_opts
.overwrite
&&
287 (hists
->stats
.nr_lost_warned
!=
288 hists
->stats
.nr_events
[PERF_RECORD_LOST
])) {
289 hists
->stats
.nr_lost_warned
=
290 hists
->stats
.nr_events
[PERF_RECORD_LOST
];
291 color_fprintf(stdout
, PERF_COLOR_RED
,
292 "WARNING: LOST %d chunks, Check IO/CPU overload",
293 hists
->stats
.nr_lost_warned
);
297 if (top
->sym_filter_entry
) {
298 perf_top__show_details(top
);
302 if (top
->evlist
->enabled
) {
304 hists__delete_entries(hists
);
306 hists__decay_entries(hists
, top
->hide_user_symbols
,
307 top
->hide_kernel_symbols
);
311 hists__collapse_resort(hists
, NULL
);
312 perf_evsel__output_resort(evsel
, NULL
);
314 hists__output_recalc_col_len(hists
, top
->print_entries
- printed
);
316 hists__fprintf(hists
, false, top
->print_entries
- printed
, win_width
,
317 top
->min_percent
, stdout
, symbol_conf
.use_callchain
);
320 static void prompt_integer(int *target
, const char *msg
)
322 char *buf
= malloc(0), *p
;
326 fprintf(stdout
, "\n%s: ", msg
);
327 if (getline(&buf
, &dummy
, stdin
) < 0)
330 p
= strchr(buf
, '\n');
340 tmp
= strtoul(buf
, NULL
, 10);
346 static void prompt_percent(int *target
, const char *msg
)
350 prompt_integer(&tmp
, msg
);
351 if (tmp
>= 0 && tmp
<= 100)
355 static void perf_top__prompt_symbol(struct perf_top
*top
, const char *msg
)
357 char *buf
= malloc(0), *p
;
358 struct hist_entry
*syme
= top
->sym_filter_entry
, *n
, *found
= NULL
;
359 struct hists
*hists
= evsel__hists(top
->sym_evsel
);
360 struct rb_node
*next
;
363 /* zero counters of active symbol */
365 __zero_source_counters(syme
);
366 top
->sym_filter_entry
= NULL
;
369 fprintf(stdout
, "\n%s: ", msg
);
370 if (getline(&buf
, &dummy
, stdin
) < 0)
373 p
= strchr(buf
, '\n');
377 next
= rb_first(&hists
->entries
);
379 n
= rb_entry(next
, struct hist_entry
, rb_node
);
380 if (n
->ms
.sym
&& !strcmp(buf
, n
->ms
.sym
->name
)) {
384 next
= rb_next(&n
->rb_node
);
388 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
391 perf_top__parse_source(top
, found
);
397 static void perf_top__print_mapped_keys(struct perf_top
*top
)
401 if (top
->sym_filter_entry
) {
402 struct symbol
*sym
= top
->sym_filter_entry
->ms
.sym
;
406 fprintf(stdout
, "\nMapped keys:\n");
407 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", top
->delay_secs
);
408 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", top
->print_entries
);
410 if (top
->evlist
->nr_entries
> 1)
411 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top
->sym_evsel
));
413 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", top
->count_filter
);
415 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", top
->sym_pcnt_filter
);
416 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
417 fprintf(stdout
, "\t[S] stop annotation.\n");
420 "\t[K] hide kernel symbols. \t(%s)\n",
421 top
->hide_kernel_symbols
? "yes" : "no");
423 "\t[U] hide user symbols. \t(%s)\n",
424 top
->hide_user_symbols
? "yes" : "no");
425 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", top
->zero
? 1 : 0);
426 fprintf(stdout
, "\t[qQ] quit.\n");
429 static int perf_top__key_mapped(struct perf_top
*top
, int c
)
445 return top
->evlist
->nr_entries
> 1 ? 1 : 0;
453 static bool perf_top__handle_keypress(struct perf_top
*top
, int c
)
457 if (!perf_top__key_mapped(top
, c
)) {
458 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
461 perf_top__print_mapped_keys(top
);
462 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
465 set_term_quiet_input(&save
);
467 poll(&stdin_poll
, 1, -1);
470 tcsetattr(0, TCSAFLUSH
, &save
);
471 if (!perf_top__key_mapped(top
, c
))
477 prompt_integer(&top
->delay_secs
, "Enter display delay");
478 if (top
->delay_secs
< 1)
482 prompt_integer(&top
->print_entries
, "Enter display entries (lines)");
483 if (top
->print_entries
== 0) {
484 perf_top__resize(top
);
485 signal(SIGWINCH
, winch_sig
);
487 signal(SIGWINCH
, SIG_DFL
);
491 if (top
->evlist
->nr_entries
> 1) {
492 /* Select 0 as the default event: */
495 fprintf(stderr
, "\nAvailable events:");
497 evlist__for_each_entry(top
->evlist
, top
->sym_evsel
)
498 fprintf(stderr
, "\n\t%d %s", top
->sym_evsel
->idx
, perf_evsel__name(top
->sym_evsel
));
500 prompt_integer(&counter
, "Enter details event counter");
502 if (counter
>= top
->evlist
->nr_entries
) {
503 top
->sym_evsel
= perf_evlist__first(top
->evlist
);
504 fprintf(stderr
, "Sorry, no such event, using %s.\n", perf_evsel__name(top
->sym_evsel
));
508 evlist__for_each_entry(top
->evlist
, top
->sym_evsel
)
509 if (top
->sym_evsel
->idx
== counter
)
512 top
->sym_evsel
= perf_evlist__first(top
->evlist
);
515 prompt_integer(&top
->count_filter
, "Enter display event count filter");
518 prompt_percent(&top
->sym_pcnt_filter
,
519 "Enter details display event filter (percent)");
522 top
->hide_kernel_symbols
= !top
->hide_kernel_symbols
;
526 printf("exiting.\n");
527 if (top
->dump_symtab
)
528 perf_session__fprintf_dsos(top
->session
, stderr
);
532 perf_top__prompt_symbol(top
, "Enter details symbol");
535 if (!top
->sym_filter_entry
)
538 struct hist_entry
*syme
= top
->sym_filter_entry
;
540 top
->sym_filter_entry
= NULL
;
541 __zero_source_counters(syme
);
545 top
->hide_user_symbols
= !top
->hide_user_symbols
;
548 top
->zero
= !top
->zero
;
557 static void perf_top__sort_new_samples(void *arg
)
559 struct perf_top
*t
= arg
;
560 struct perf_evsel
*evsel
= t
->sym_evsel
;
563 perf_top__reset_sample_counters(t
);
565 if (t
->evlist
->selected
!= NULL
)
566 t
->sym_evsel
= t
->evlist
->selected
;
568 hists
= evsel__hists(evsel
);
570 if (t
->evlist
->enabled
) {
572 hists__delete_entries(hists
);
574 hists__decay_entries(hists
, t
->hide_user_symbols
,
575 t
->hide_kernel_symbols
);
579 hists__collapse_resort(hists
, NULL
);
580 perf_evsel__output_resort(evsel
, NULL
);
583 static void *display_thread_tui(void *arg
)
585 struct perf_evsel
*pos
;
586 struct perf_top
*top
= arg
;
587 const char *help
= "For a higher level overview, try: perf top --sort comm,dso";
588 struct hist_browser_timer hbt
= {
589 .timer
= perf_top__sort_new_samples
,
591 .refresh
= top
->delay_secs
,
594 /* In order to read symbols from other namespaces perf to needs to call
595 * setns(2). This isn't permitted if the struct_fs has multiple users.
596 * unshare(2) the fs so that we may continue to setns into namespaces
597 * that we're observing.
601 perf_top__sort_new_samples(top
);
604 * Initialize the uid_filter_str, in the future the TUI will allow
605 * Zooming in/out UIDs. For now juse use whatever the user passed
608 evlist__for_each_entry(top
->evlist
, pos
) {
609 struct hists
*hists
= evsel__hists(pos
);
610 hists
->uid_filter_str
= top
->record_opts
.target
.uid_str
;
613 perf_evlist__tui_browse_hists(top
->evlist
, help
, &hbt
,
615 &top
->session
->header
.env
,
616 !top
->record_opts
.overwrite
);
622 static void display_sig(int sig __maybe_unused
)
627 static void display_setup_sig(void)
629 signal(SIGSEGV
, sighandler_dump_stack
);
630 signal(SIGFPE
, sighandler_dump_stack
);
631 signal(SIGINT
, display_sig
);
632 signal(SIGQUIT
, display_sig
);
633 signal(SIGTERM
, display_sig
);
636 static void *display_thread(void *arg
)
638 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
640 struct perf_top
*top
= arg
;
643 /* In order to read symbols from other namespaces perf to needs to call
644 * setns(2). This isn't permitted if the struct_fs has multiple users.
645 * unshare(2) the fs so that we may continue to setns into namespaces
646 * that we're observing.
651 pthread__unblock_sigwinch();
653 delay_msecs
= top
->delay_secs
* MSEC_PER_SEC
;
654 set_term_quiet_input(&save
);
659 perf_top__print_sym_table(top
);
661 * Either timeout expired or we got an EINTR due to SIGWINCH,
662 * refresh screen in both cases.
664 switch (poll(&stdin_poll
, 1, delay_msecs
)) {
673 tcsetattr(0, TCSAFLUSH
, &save
);
675 if (perf_top__handle_keypress(top
, c
))
681 tcsetattr(0, TCSAFLUSH
, &save
);
685 static int hist_iter__top_callback(struct hist_entry_iter
*iter
,
686 struct addr_location
*al
, bool single
,
689 struct perf_top
*top
= arg
;
690 struct hist_entry
*he
= iter
->he
;
691 struct perf_evsel
*evsel
= iter
->evsel
;
693 if (perf_hpp_list
.sym
&& single
)
694 perf_top__record_precise_ip(top
, he
, iter
->sample
, evsel
->idx
, al
->addr
);
696 hist__account_cycles(iter
->sample
->branch_stack
, al
, iter
->sample
,
697 !(top
->record_opts
.branch_stack
& PERF_SAMPLE_BRANCH_ANY
));
701 static void perf_event__process_sample(struct perf_tool
*tool
,
702 const union perf_event
*event
,
703 struct perf_evsel
*evsel
,
704 struct perf_sample
*sample
,
705 struct machine
*machine
)
707 struct perf_top
*top
= container_of(tool
, struct perf_top
, tool
);
708 struct addr_location al
;
711 if (!machine
&& perf_guest
) {
712 static struct intlist
*seen
;
715 seen
= intlist__new(NULL
);
717 if (!intlist__has_entry(seen
, sample
->pid
)) {
718 pr_err("Can't find guest [%d]'s kernel information\n",
720 intlist__add(seen
, sample
->pid
);
726 pr_err("%u unprocessable samples recorded.\r",
727 top
->session
->evlist
->stats
.nr_unprocessable_samples
++);
731 if (event
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
732 top
->exact_samples
++;
734 if (machine__resolve(machine
, &al
, sample
) < 0)
737 if (!machine
->kptr_restrict_warned
&&
738 symbol_conf
.kptr_restrict
&&
739 al
.cpumode
== PERF_RECORD_MISC_KERNEL
) {
740 if (!perf_evlist__exclude_kernel(top
->session
->evlist
)) {
742 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
743 "Check /proc/sys/kernel/kptr_restrict.\n\n"
744 "Kernel%s samples will not be resolved.\n",
745 al
.map
&& !RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
]) ?
747 if (use_browser
<= 0)
750 machine
->kptr_restrict_warned
= true;
753 if (al
.sym
== NULL
) {
754 const char *msg
= "Kernel samples will not be resolved.\n";
756 * As we do lazy loading of symtabs we only will know if the
757 * specified vmlinux file is invalid when we actually have a
758 * hit in kernel space and then try to load it. So if we get
759 * here and there are _no_ symbols in the DSO backing the
760 * kernel map, bail out.
762 * We may never get here, for instance, if we use -K/
763 * --hide-kernel-symbols, even if the user specifies an
764 * invalid --vmlinux ;-)
766 if (!machine
->kptr_restrict_warned
&& !top
->vmlinux_warned
&&
767 al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
768 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
769 if (symbol_conf
.vmlinux_name
) {
771 dso__strerror_load(al
.map
->dso
, serr
, sizeof(serr
));
772 ui__warning("The %s file can't be used: %s\n%s",
773 symbol_conf
.vmlinux_name
, serr
, msg
);
775 ui__warning("A vmlinux file was not found.\n%s",
779 if (use_browser
<= 0)
781 top
->vmlinux_warned
= true;
785 if (al
.sym
== NULL
|| !al
.sym
->idle
) {
786 struct hists
*hists
= evsel__hists(evsel
);
787 struct hist_entry_iter iter
= {
790 .add_entry_cb
= hist_iter__top_callback
,
793 if (symbol_conf
.cumulate_callchain
)
794 iter
.ops
= &hist_iter_cumulative
;
796 iter
.ops
= &hist_iter_normal
;
798 pthread_mutex_lock(&hists
->lock
);
800 err
= hist_entry_iter__add(&iter
, &al
, top
->max_stack
, top
);
802 pr_err("Problem incrementing symbol period, skipping event\n");
804 pthread_mutex_unlock(&hists
->lock
);
807 addr_location__put(&al
);
810 static void perf_top__mmap_read_idx(struct perf_top
*top
, int idx
)
812 struct record_opts
*opts
= &top
->record_opts
;
813 struct perf_evlist
*evlist
= top
->evlist
;
814 struct perf_sample sample
;
815 struct perf_evsel
*evsel
;
816 struct perf_mmap
*md
;
817 struct perf_session
*session
= top
->session
;
818 union perf_event
*event
;
819 struct machine
*machine
;
822 md
= opts
->overwrite
? &evlist
->overwrite_mmap
[idx
] : &evlist
->mmap
[idx
];
823 if (perf_mmap__read_init(md
) < 0)
826 while ((event
= perf_mmap__read_event(md
)) != NULL
) {
827 ret
= perf_evlist__parse_sample(evlist
, event
, &sample
);
829 pr_err("Can't parse sample, err = %d\n", ret
);
833 evsel
= perf_evlist__id2evsel(session
->evlist
, sample
.id
);
834 assert(evsel
!= NULL
);
836 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
839 switch (sample
.cpumode
) {
840 case PERF_RECORD_MISC_USER
:
842 if (top
->hide_user_symbols
)
844 machine
= &session
->machines
.host
;
846 case PERF_RECORD_MISC_KERNEL
:
847 ++top
->kernel_samples
;
848 if (top
->hide_kernel_symbols
)
850 machine
= &session
->machines
.host
;
852 case PERF_RECORD_MISC_GUEST_KERNEL
:
853 ++top
->guest_kernel_samples
;
854 machine
= perf_session__find_machine(session
,
857 case PERF_RECORD_MISC_GUEST_USER
:
858 ++top
->guest_us_samples
;
860 * TODO: we don't process guest user from host side
861 * except simple counting.
865 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
867 machine
= &session
->machines
.host
;
872 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
873 perf_event__process_sample(&top
->tool
, event
, evsel
,
875 } else if (event
->header
.type
< PERF_RECORD_MAX
) {
876 hists__inc_nr_events(evsel__hists(evsel
), event
->header
.type
);
877 machine__process_event(machine
, event
, &sample
);
879 ++session
->evlist
->stats
.nr_unknown_events
;
881 perf_mmap__consume(md
);
884 perf_mmap__read_done(md
);
887 static void perf_top__mmap_read(struct perf_top
*top
)
889 bool overwrite
= top
->record_opts
.overwrite
;
890 struct perf_evlist
*evlist
= top
->evlist
;
891 unsigned long long start
, end
;
896 perf_evlist__toggle_bkw_mmap(evlist
, BKW_MMAP_DATA_PENDING
);
898 for (i
= 0; i
< top
->evlist
->nr_mmaps
; i
++)
899 perf_top__mmap_read_idx(top
, i
);
902 perf_evlist__toggle_bkw_mmap(evlist
, BKW_MMAP_EMPTY
);
903 perf_evlist__toggle_bkw_mmap(evlist
, BKW_MMAP_RUNNING
);
907 if ((end
- start
) > (unsigned long long)top
->delay_secs
* NSEC_PER_SEC
)
908 ui__warning("Too slow to read ring buffer.\n"
909 "Please try increasing the period (-c) or\n"
910 "decreasing the freq (-F) or\n"
911 "limiting the number of CPUs (-C)\n");
915 * Check per-event overwrite term.
916 * perf top should support consistent term for all events.
917 * - All events don't have per-event term
918 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
919 * Nothing change, return 0.
920 * - All events have same per-event term
921 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
922 * Using the per-event setting to replace the opts->overwrite if
923 * they are different, then return 0.
924 * - Events have different per-event term
925 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
927 * - Some of the event set per-event term, but some not.
928 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
931 static int perf_top__overwrite_check(struct perf_top
*top
)
933 struct record_opts
*opts
= &top
->record_opts
;
934 struct perf_evlist
*evlist
= top
->evlist
;
935 struct perf_evsel_config_term
*term
;
936 struct list_head
*config_terms
;
937 struct perf_evsel
*evsel
;
938 int set
, overwrite
= -1;
940 evlist__for_each_entry(evlist
, evsel
) {
942 config_terms
= &evsel
->config_terms
;
943 list_for_each_entry(term
, config_terms
, list
) {
944 if (term
->type
== PERF_EVSEL__CONFIG_TERM_OVERWRITE
)
945 set
= term
->val
.overwrite
? 1 : 0;
948 /* no term for current and previous event (likely) */
949 if ((overwrite
< 0) && (set
< 0))
952 /* has term for both current and previous event, compare */
953 if ((overwrite
>= 0) && (set
>= 0) && (overwrite
!= set
))
956 /* no term for current event but has term for previous one */
957 if ((overwrite
>= 0) && (set
< 0))
960 /* has term for current event */
961 if ((overwrite
< 0) && (set
>= 0)) {
962 /* if it's first event, set overwrite */
963 if (evsel
== perf_evlist__first(evlist
))
970 if ((overwrite
>= 0) && (opts
->overwrite
!= overwrite
))
971 opts
->overwrite
= overwrite
;
976 static int perf_top_overwrite_fallback(struct perf_top
*top
,
977 struct perf_evsel
*evsel
)
979 struct record_opts
*opts
= &top
->record_opts
;
980 struct perf_evlist
*evlist
= top
->evlist
;
981 struct perf_evsel
*counter
;
983 if (!opts
->overwrite
)
986 /* only fall back when first event fails */
987 if (evsel
!= perf_evlist__first(evlist
))
990 evlist__for_each_entry(evlist
, counter
)
991 counter
->attr
.write_backward
= false;
992 opts
->overwrite
= false;
993 pr_debug2("fall back to non-overwrite mode\n");
997 static int perf_top__start_counters(struct perf_top
*top
)
1000 struct perf_evsel
*counter
;
1001 struct perf_evlist
*evlist
= top
->evlist
;
1002 struct record_opts
*opts
= &top
->record_opts
;
1004 if (perf_top__overwrite_check(top
)) {
1005 ui__error("perf top only support consistent per-event "
1006 "overwrite setting for all events\n");
1010 perf_evlist__config(evlist
, opts
, &callchain_param
);
1012 evlist__for_each_entry(evlist
, counter
) {
1014 if (perf_evsel__open(counter
, top
->evlist
->cpus
,
1015 top
->evlist
->threads
) < 0) {
1018 * Specially handle overwrite fall back.
1019 * Because perf top is the only tool which has
1020 * overwrite mode by default, support
1021 * both overwrite and non-overwrite mode, and
1022 * require consistent mode for all events.
1024 * May move it to generic code with more tools
1025 * have similar attribute.
1027 if (perf_missing_features
.write_backward
&&
1028 perf_top_overwrite_fallback(top
, counter
))
1031 if (perf_evsel__fallback(counter
, errno
, msg
, sizeof(msg
))) {
1033 ui__warning("%s\n", msg
);
1037 perf_evsel__open_strerror(counter
, &opts
->target
,
1038 errno
, msg
, sizeof(msg
));
1039 ui__error("%s\n", msg
);
1044 if (perf_evlist__mmap(evlist
, opts
->mmap_pages
) < 0) {
1045 ui__error("Failed to mmap with %d (%s)\n",
1046 errno
, str_error_r(errno
, msg
, sizeof(msg
)));
1056 static int callchain_param__setup_sample_type(struct callchain_param
*callchain
)
1058 if (!perf_hpp_list
.sym
) {
1059 if (callchain
->enabled
) {
1060 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1063 } else if (callchain
->mode
!= CHAIN_NONE
) {
1064 if (callchain_register_param(callchain
) < 0) {
1065 ui__error("Can't register callchain params.\n");
1073 static int __cmd_top(struct perf_top
*top
)
1076 struct perf_evsel
*pos
;
1077 struct perf_evsel_config_term
*err_term
;
1078 struct perf_evlist
*evlist
= top
->evlist
;
1079 struct record_opts
*opts
= &top
->record_opts
;
1083 top
->session
= perf_session__new(NULL
, false, NULL
);
1084 if (top
->session
== NULL
)
1087 if (!objdump_path
) {
1088 ret
= perf_env__lookup_objdump(&top
->session
->header
.env
);
1093 ret
= callchain_param__setup_sample_type(&callchain_param
);
1097 if (perf_session__register_idle_thread(top
->session
) < 0)
1100 if (top
->nr_threads_synthesize
> 1)
1101 perf_set_multithreaded();
1103 machine__synthesize_threads(&top
->session
->machines
.host
, &opts
->target
,
1104 top
->evlist
->threads
, false,
1105 opts
->proc_map_timeout
,
1106 top
->nr_threads_synthesize
);
1108 if (top
->nr_threads_synthesize
> 1)
1109 perf_set_singlethreaded();
1111 if (perf_hpp_list
.socket
) {
1112 ret
= perf_env__read_cpu_topology_map(&perf_env
);
1114 goto out_err_cpu_topo
;
1117 ret
= perf_top__start_counters(top
);
1121 ret
= perf_evlist__apply_drv_configs(evlist
, &pos
, &err_term
);
1123 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1124 err_term
->val
.drv_cfg
, perf_evsel__name(pos
), errno
,
1125 str_error_r(errno
, msg
, sizeof(msg
)));
1129 top
->session
->evlist
= top
->evlist
;
1130 perf_session__set_id_hdr_size(top
->session
);
1133 * When perf is starting the traced process, all the events (apart from
1134 * group members) have enable_on_exec=1 set, so don't spoil it by
1135 * prematurely enabling them.
1137 * XXX 'top' still doesn't start workloads like record, trace, but should,
1138 * so leave the check here.
1140 if (!target__none(&opts
->target
))
1141 perf_evlist__enable(top
->evlist
);
1143 /* Wait for a minimal set of events before starting the snapshot */
1144 perf_evlist__poll(top
->evlist
, 100);
1146 perf_top__mmap_read(top
);
1149 if (pthread_create(&thread
, NULL
, (use_browser
> 0 ? display_thread_tui
:
1150 display_thread
), top
)) {
1151 ui__error("Could not create display thread.\n");
1155 if (top
->realtime_prio
) {
1156 struct sched_param param
;
1158 param
.sched_priority
= top
->realtime_prio
;
1159 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1160 ui__error("Could not set realtime priority.\n");
1166 u64 hits
= top
->samples
;
1168 perf_top__mmap_read(top
);
1170 if (opts
->overwrite
|| (hits
== top
->samples
))
1171 ret
= perf_evlist__poll(top
->evlist
, 100);
1174 perf_top__resize(top
);
1181 pthread_join(thread
, NULL
);
1183 perf_session__delete(top
->session
);
1184 top
->session
= NULL
;
1189 char errbuf
[BUFSIZ
];
1190 const char *err
= str_error_r(-ret
, errbuf
, sizeof(errbuf
));
1192 ui__error("Could not read the CPU topology map: %s\n", err
);
1198 callchain_opt(const struct option
*opt
, const char *arg
, int unset
)
1200 symbol_conf
.use_callchain
= true;
1201 return record_callchain_opt(opt
, arg
, unset
);
1205 parse_callchain_opt(const struct option
*opt
, const char *arg
, int unset
)
1207 struct callchain_param
*callchain
= opt
->value
;
1209 callchain
->enabled
= !unset
;
1210 callchain
->record_mode
= CALLCHAIN_FP
;
1216 symbol_conf
.use_callchain
= false;
1217 callchain
->record_mode
= CALLCHAIN_NONE
;
1221 return parse_callchain_top_opt(arg
);
1224 static int perf_top_config(const char *var
, const char *value
, void *cb __maybe_unused
)
1226 if (!strcmp(var
, "top.call-graph")) {
1227 var
= "call-graph.record-mode";
1228 return perf_default_config(var
, value
, cb
);
1230 if (!strcmp(var
, "top.children")) {
1231 symbol_conf
.cumulate_callchain
= perf_config_bool(var
, value
);
1239 parse_percent_limit(const struct option
*opt
, const char *arg
,
1240 int unset __maybe_unused
)
1242 struct perf_top
*top
= opt
->value
;
1244 top
->min_percent
= strtof(arg
, NULL
);
1248 const char top_callchain_help
[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1249 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1251 int cmd_top(int argc
, const char **argv
)
1253 char errbuf
[BUFSIZ
];
1254 struct perf_top top
= {
1258 .mmap_pages
= UINT_MAX
,
1259 .user_freq
= UINT_MAX
,
1260 .user_interval
= ULLONG_MAX
,
1261 .freq
= 4000, /* 4 KHz */
1265 .proc_map_timeout
= 500,
1268 .max_stack
= sysctl_perf_event_max_stack
,
1269 .sym_pcnt_filter
= 5,
1270 .nr_threads_synthesize
= UINT_MAX
,
1272 struct record_opts
*opts
= &top
.record_opts
;
1273 struct target
*target
= &opts
->target
;
1274 const struct option options
[] = {
1275 OPT_CALLBACK('e', "event", &top
.evlist
, "event",
1276 "event selector. use 'perf list' to list available events",
1277 parse_events_option
),
1278 OPT_U64('c', "count", &opts
->user_interval
, "event period to sample"),
1279 OPT_STRING('p', "pid", &target
->pid
, "pid",
1280 "profile events on existing process id"),
1281 OPT_STRING('t', "tid", &target
->tid
, "tid",
1282 "profile events on existing thread id"),
1283 OPT_BOOLEAN('a', "all-cpus", &target
->system_wide
,
1284 "system-wide collection from all CPUs"),
1285 OPT_STRING('C', "cpu", &target
->cpu_list
, "cpu",
1286 "list of cpus to monitor"),
1287 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1288 "file", "vmlinux pathname"),
1289 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf
.ignore_vmlinux
,
1290 "don't load vmlinux even if found"),
1291 OPT_BOOLEAN('K', "hide_kernel_symbols", &top
.hide_kernel_symbols
,
1292 "hide kernel symbols"),
1293 OPT_CALLBACK('m', "mmap-pages", &opts
->mmap_pages
, "pages",
1294 "number of mmap data pages",
1295 perf_evlist__parse_mmap_pages
),
1296 OPT_INTEGER('r', "realtime", &top
.realtime_prio
,
1297 "collect data with this RT SCHED_FIFO priority"),
1298 OPT_INTEGER('d', "delay", &top
.delay_secs
,
1299 "number of seconds to delay between refreshes"),
1300 OPT_BOOLEAN('D', "dump-symtab", &top
.dump_symtab
,
1301 "dump the symbol table used for profiling"),
1302 OPT_INTEGER('f', "count-filter", &top
.count_filter
,
1303 "only display functions with more events than this"),
1304 OPT_BOOLEAN(0, "group", &opts
->group
,
1305 "put the counters into a counter group"),
1306 OPT_BOOLEAN('i', "no-inherit", &opts
->no_inherit
,
1307 "child tasks do not inherit counters"),
1308 OPT_STRING(0, "sym-annotate", &top
.sym_filter
, "symbol name",
1309 "symbol to annotate"),
1310 OPT_BOOLEAN('z', "zero", &top
.zero
, "zero history across updates"),
1311 OPT_CALLBACK('F', "freq", &top
.record_opts
, "freq or 'max'",
1312 "profile at this frequency",
1313 record__parse_freq
),
1314 OPT_INTEGER('E', "entries", &top
.print_entries
,
1315 "display this many functions"),
1316 OPT_BOOLEAN('U', "hide_user_symbols", &top
.hide_user_symbols
,
1317 "hide user symbols"),
1318 OPT_BOOLEAN(0, "tui", &top
.use_tui
, "Use the TUI interface"),
1319 OPT_BOOLEAN(0, "stdio", &top
.use_stdio
, "Use the stdio interface"),
1320 OPT_INCR('v', "verbose", &verbose
,
1321 "be more verbose (show counter open errors, etc)"),
1322 OPT_STRING('s', "sort", &sort_order
, "key[,key2...]",
1323 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1324 " Please refer the man page for the complete list."),
1325 OPT_STRING(0, "fields", &field_order
, "key[,keys...]",
1326 "output field(s): overhead, period, sample plus all of sort keys"),
1327 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf
.show_nr_samples
,
1328 "Show a column with the number of samples"),
1329 OPT_CALLBACK_NOOPT('g', NULL
, &callchain_param
,
1330 NULL
, "enables call-graph recording and display",
1332 OPT_CALLBACK(0, "call-graph", &callchain_param
,
1333 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1334 top_callchain_help
, &parse_callchain_opt
),
1335 OPT_BOOLEAN(0, "children", &symbol_conf
.cumulate_callchain
,
1336 "Accumulate callchains of children and show total overhead as well"),
1337 OPT_INTEGER(0, "max-stack", &top
.max_stack
,
1338 "Set the maximum stack depth when parsing the callchain. "
1339 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH
)),
1340 OPT_CALLBACK(0, "ignore-callees", NULL
, "regex",
1341 "ignore callees of these functions in call graphs",
1342 report_parse_ignore_callees_opt
),
1343 OPT_BOOLEAN(0, "show-total-period", &symbol_conf
.show_total_period
,
1344 "Show a column with the sum of periods"),
1345 OPT_STRING(0, "dsos", &symbol_conf
.dso_list_str
, "dso[,dso...]",
1346 "only consider symbols in these dsos"),
1347 OPT_STRING(0, "comms", &symbol_conf
.comm_list_str
, "comm[,comm...]",
1348 "only consider symbols in these comms"),
1349 OPT_STRING(0, "symbols", &symbol_conf
.sym_list_str
, "symbol[,symbol...]",
1350 "only consider these symbols"),
1351 OPT_BOOLEAN(0, "source", &symbol_conf
.annotate_src
,
1352 "Interleave source code with assembly code (default)"),
1353 OPT_BOOLEAN(0, "asm-raw", &symbol_conf
.annotate_asm_raw
,
1354 "Display raw encoding of assembly instructions (default)"),
1355 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf
.demangle_kernel
,
1356 "Enable kernel symbol demangling"),
1357 OPT_STRING(0, "objdump", &objdump_path
, "path",
1358 "objdump binary to use for disassembly and annotations"),
1359 OPT_STRING('M', "disassembler-style", &disassembler_style
, "disassembler style",
1360 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1361 OPT_STRING('u', "uid", &target
->uid_str
, "user", "user to profile"),
1362 OPT_CALLBACK(0, "percent-limit", &top
, "percent",
1363 "Don't show entries under that percent", parse_percent_limit
),
1364 OPT_CALLBACK(0, "percentage", NULL
, "relative|absolute",
1365 "How to display percentage of filtered entries", parse_filter_percentage
),
1366 OPT_STRING('w', "column-widths", &symbol_conf
.col_width_list_str
,
1368 "don't try to adjust column width, use these fixed values"),
1369 OPT_UINTEGER(0, "proc-map-timeout", &opts
->proc_map_timeout
,
1370 "per thread proc mmap processing timeout in ms"),
1371 OPT_CALLBACK_NOOPT('b', "branch-any", &opts
->branch_stack
,
1372 "branch any", "sample any taken branches",
1373 parse_branch_stack
),
1374 OPT_CALLBACK('j', "branch-filter", &opts
->branch_stack
,
1375 "branch filter mask", "branch stack filter modes",
1376 parse_branch_stack
),
1377 OPT_BOOLEAN(0, "raw-trace", &symbol_conf
.raw_trace
,
1378 "Show raw trace event output (do not use print fmt or plugins)"),
1379 OPT_BOOLEAN(0, "hierarchy", &symbol_conf
.report_hierarchy
,
1380 "Show entries in a hierarchy"),
1381 OPT_BOOLEAN(0, "force", &symbol_conf
.force
, "don't complain, do it"),
1382 OPT_UINTEGER(0, "num-thread-synthesize", &top
.nr_threads_synthesize
,
1383 "number of thread to run event synthesize"),
1386 const char * const top_usage
[] = {
1387 "perf top [<options>]",
1390 int status
= hists__init();
1395 top
.evlist
= perf_evlist__new();
1396 if (top
.evlist
== NULL
)
1399 status
= perf_config(perf_top_config
, &top
);
1403 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1405 usage_with_options(top_usage
, options
);
1407 if (!top
.evlist
->nr_entries
&&
1408 perf_evlist__add_default(top
.evlist
) < 0) {
1409 pr_err("Not enough memory for event selector list\n");
1410 goto out_delete_evlist
;
1413 if (symbol_conf
.report_hierarchy
) {
1414 /* disable incompatible options */
1415 symbol_conf
.event_group
= false;
1416 symbol_conf
.cumulate_callchain
= false;
1419 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1420 parse_options_usage(top_usage
, options
, "fields", 0);
1421 parse_options_usage(NULL
, options
, "hierarchy", 0);
1422 goto out_delete_evlist
;
1426 sort__mode
= SORT_MODE__TOP
;
1427 /* display thread wants entries to be collapsed in a different tree */
1428 perf_hpp_list
.need_collapse
= 1;
1432 else if (top
.use_tui
)
1435 setup_browser(false);
1437 if (setup_sorting(top
.evlist
) < 0) {
1439 parse_options_usage(top_usage
, options
, "s", 1);
1441 parse_options_usage(sort_order
? NULL
: top_usage
,
1442 options
, "fields", 0);
1443 goto out_delete_evlist
;
1446 status
= target__validate(target
);
1448 target__strerror(target
, status
, errbuf
, BUFSIZ
);
1449 ui__warning("%s\n", errbuf
);
1452 status
= target__parse_uid(target
);
1454 int saved_errno
= errno
;
1456 target__strerror(target
, status
, errbuf
, BUFSIZ
);
1457 ui__error("%s\n", errbuf
);
1459 status
= -saved_errno
;
1460 goto out_delete_evlist
;
1463 if (target__none(target
))
1464 target
->system_wide
= true;
1466 if (perf_evlist__create_maps(top
.evlist
, target
) < 0) {
1467 ui__error("Couldn't create thread/CPU maps: %s\n",
1468 errno
== ENOENT
? "No such process" : str_error_r(errno
, errbuf
, sizeof(errbuf
)));
1469 goto out_delete_evlist
;
1472 symbol_conf
.nr_events
= top
.evlist
->nr_entries
;
1474 if (top
.delay_secs
< 1)
1477 if (record_opts__config(opts
)) {
1479 goto out_delete_evlist
;
1482 top
.sym_evsel
= perf_evlist__first(top
.evlist
);
1484 if (!callchain_param
.enabled
) {
1485 symbol_conf
.cumulate_callchain
= false;
1486 perf_hpp__cancel_cumulate();
1489 if (symbol_conf
.cumulate_callchain
&& !callchain_param
.order_set
)
1490 callchain_param
.order
= ORDER_CALLER
;
1492 status
= symbol__annotation_init();
1494 goto out_delete_evlist
;
1496 annotation_config__init();
1498 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1499 if (symbol__init(NULL
) < 0)
1502 sort__setup_elide(stdout
);
1504 get_term_dimensions(&top
.winsize
);
1505 if (top
.print_entries
== 0) {
1506 perf_top__update_print_entries(&top
);
1507 signal(SIGWINCH
, winch_sig
);
1510 status
= __cmd_top(&top
);
1513 perf_evlist__delete(top
.evlist
);