aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf/builtin-top.c
diff options
context:
space:
mode:
Diffstat (limited to 'tools/perf/builtin-top.c')
-rw-r--r--tools/perf/builtin-top.c692
1 files changed, 692 insertions, 0 deletions
diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
new file mode 100644
index 000000000000..f2e7312f85c9
--- /dev/null
+++ b/tools/perf/builtin-top.c
@@ -0,0 +1,692 @@
1/*
2 * builtin-top.c
3 *
4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
6 *
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 *
9 * Improvements and fixes by:
10 *
11 * Arjan van de Ven <arjan@linux.intel.com>
12 * Yanmin Zhang <yanmin.zhang@intel.com>
13 * Wu Fengguang <fengguang.wu@intel.com>
14 * Mike Galbraith <efault@gmx.de>
15 * Paul Mackerras <paulus@samba.org>
16 *
17 * Released under the GPL v2. (and only v2, not any later version)
18 */
19#include "builtin.h"
20
21#include "perf.h"
22
23#include "util/symbol.h"
24#include "util/color.h"
25#include "util/util.h"
26#include "util/rbtree.h"
27#include "util/parse-options.h"
28#include "util/parse-events.h"
29
30#include <assert.h>
31#include <fcntl.h>
32
33#include <stdio.h>
34
35#include <errno.h>
36#include <time.h>
37#include <sched.h>
38#include <pthread.h>
39
40#include <sys/syscall.h>
41#include <sys/ioctl.h>
42#include <sys/poll.h>
43#include <sys/prctl.h>
44#include <sys/wait.h>
45#include <sys/uio.h>
46#include <sys/mman.h>
47
48#include <linux/unistd.h>
49#include <linux/types.h>
50
51static int fd[MAX_NR_CPUS][MAX_COUNTERS];
52
53static int system_wide = 0;
54
55static int default_interval = 100000;
56
57static __u64 count_filter = 5;
58static int print_entries = 15;
59
60static int target_pid = -1;
61static int profile_cpu = -1;
62static int nr_cpus = 0;
63static unsigned int realtime_prio = 0;
64static int group = 0;
65static unsigned int page_size;
66static unsigned int mmap_pages = 16;
67static int freq = 0;
68
69static char *sym_filter;
70static unsigned long filter_start;
71static unsigned long filter_end;
72
73static int delay_secs = 2;
74static int zero;
75static int dump_symtab;
76
77/*
78 * Symbols
79 */
80
81static uint64_t min_ip;
82static uint64_t max_ip = -1ll;
83
84struct sym_entry {
85 struct rb_node rb_node;
86 struct list_head node;
87 unsigned long count[MAX_COUNTERS];
88 unsigned long snap_count;
89 double weight;
90 int skip;
91};
92
93struct sym_entry *sym_filter_entry;
94
95struct dso *kernel_dso;
96
97/*
98 * Symbols will be added here in record_ip and will get out
99 * after decayed.
100 */
101static LIST_HEAD(active_symbols);
102static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
103
104/*
105 * Ordering weight: count-1 * count-2 * ... / count-n
106 */
107static double sym_weight(const struct sym_entry *sym)
108{
109 double weight = sym->snap_count;
110 int counter;
111
112 for (counter = 1; counter < nr_counters-1; counter++)
113 weight *= sym->count[counter];
114
115 weight /= (sym->count[counter] + 1);
116
117 return weight;
118}
119
120static long samples;
121static long userspace_samples;
122static const char CONSOLE_CLEAR[] = "";
123
124static void __list_insert_active_sym(struct sym_entry *syme)
125{
126 list_add(&syme->node, &active_symbols);
127}
128
129static void list_remove_active_sym(struct sym_entry *syme)
130{
131 pthread_mutex_lock(&active_symbols_lock);
132 list_del_init(&syme->node);
133 pthread_mutex_unlock(&active_symbols_lock);
134}
135
136static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
137{
138 struct rb_node **p = &tree->rb_node;
139 struct rb_node *parent = NULL;
140 struct sym_entry *iter;
141
142 while (*p != NULL) {
143 parent = *p;
144 iter = rb_entry(parent, struct sym_entry, rb_node);
145
146 if (se->weight > iter->weight)
147 p = &(*p)->rb_left;
148 else
149 p = &(*p)->rb_right;
150 }
151
152 rb_link_node(&se->rb_node, parent, p);
153 rb_insert_color(&se->rb_node, tree);
154}
155
156static void print_sym_table(void)
157{
158 int printed = 0, j;
159 int counter;
160 float samples_per_sec = samples/delay_secs;
161 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
162 float sum_ksamples = 0.0;
163 struct sym_entry *syme, *n;
164 struct rb_root tmp = RB_ROOT;
165 struct rb_node *nd;
166
167 samples = userspace_samples = 0;
168
169 /* Sort the active symbols */
170 pthread_mutex_lock(&active_symbols_lock);
171 syme = list_entry(active_symbols.next, struct sym_entry, node);
172 pthread_mutex_unlock(&active_symbols_lock);
173
174 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
175 syme->snap_count = syme->count[0];
176 if (syme->snap_count != 0) {
177 syme->weight = sym_weight(syme);
178 rb_insert_active_sym(&tmp, syme);
179 sum_ksamples += syme->snap_count;
180
181 for (j = 0; j < nr_counters; j++)
182 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
183 } else
184 list_remove_active_sym(syme);
185 }
186
187 puts(CONSOLE_CLEAR);
188
189 printf(
190"------------------------------------------------------------------------------\n");
191 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
192 samples_per_sec,
193 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
194
195 if (nr_counters == 1) {
196 printf("%Ld", attrs[0].sample_period);
197 if (freq)
198 printf("Hz ");
199 else
200 printf(" ");
201 }
202
203 for (counter = 0; counter < nr_counters; counter++) {
204 if (counter)
205 printf("/");
206
207 printf("%s", event_name(counter));
208 }
209
210 printf( "], ");
211
212 if (target_pid != -1)
213 printf(" (target_pid: %d", target_pid);
214 else
215 printf(" (all");
216
217 if (profile_cpu != -1)
218 printf(", cpu: %d)\n", profile_cpu);
219 else {
220 if (target_pid != -1)
221 printf(")\n");
222 else
223 printf(", %d CPUs)\n", nr_cpus);
224 }
225
226 printf("------------------------------------------------------------------------------\n\n");
227
228 if (nr_counters == 1)
229 printf(" samples pcnt");
230 else
231 printf(" weight samples pcnt");
232
233 printf(" RIP kernel function\n"
234 " ______ _______ _____ ________________ _______________\n\n"
235 );
236
237 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
238 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
239 struct symbol *sym = (struct symbol *)(syme + 1);
240 char *color = PERF_COLOR_NORMAL;
241 double pcnt;
242
243 if (++printed > print_entries || syme->snap_count < count_filter)
244 continue;
245
246 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
247 sum_ksamples));
248
249 /*
250 * We color high-overhead entries in red, low-overhead
251 * entries in green - and keep the middle ground normal:
252 */
253 if (pcnt >= 5.0)
254 color = PERF_COLOR_RED;
255 if (pcnt < 0.5)
256 color = PERF_COLOR_GREEN;
257
258 if (nr_counters == 1)
259 printf("%20.2f - ", syme->weight);
260 else
261 printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
262
263 color_fprintf(stdout, color, "%4.1f%%", pcnt);
264 printf(" - %016llx : %s\n", sym->start, sym->name);
265 }
266}
267
268static void *display_thread(void *arg)
269{
270 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
271 int delay_msecs = delay_secs * 1000;
272
273 printf("PerfTop refresh period: %d seconds\n", delay_secs);
274
275 do {
276 print_sym_table();
277 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
278
279 printf("key pressed - exiting.\n");
280 exit(0);
281
282 return NULL;
283}
284
285static int symbol_filter(struct dso *self, struct symbol *sym)
286{
287 static int filter_match;
288 struct sym_entry *syme;
289 const char *name = sym->name;
290
291 if (!strcmp(name, "_text") ||
292 !strcmp(name, "_etext") ||
293 !strcmp(name, "_sinittext") ||
294 !strncmp("init_module", name, 11) ||
295 !strncmp("cleanup_module", name, 14) ||
296 strstr(name, "_text_start") ||
297 strstr(name, "_text_end"))
298 return 1;
299
300 syme = dso__sym_priv(self, sym);
301 /* Tag samples to be skipped. */
302 if (!strcmp("default_idle", name) ||
303 !strcmp("cpu_idle", name) ||
304 !strcmp("enter_idle", name) ||
305 !strcmp("exit_idle", name) ||
306 !strcmp("mwait_idle", name))
307 syme->skip = 1;
308
309 if (filter_match == 1) {
310 filter_end = sym->start;
311 filter_match = -1;
312 if (filter_end - filter_start > 10000) {
313 fprintf(stderr,
314 "hm, too large filter symbol <%s> - skipping.\n",
315 sym_filter);
316 fprintf(stderr, "symbol filter start: %016lx\n",
317 filter_start);
318 fprintf(stderr, " end: %016lx\n",
319 filter_end);
320 filter_end = filter_start = 0;
321 sym_filter = NULL;
322 sleep(1);
323 }
324 }
325
326 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
327 filter_match = 1;
328 filter_start = sym->start;
329 }
330
331
332 return 0;
333}
334
335static int parse_symbols(void)
336{
337 struct rb_node *node;
338 struct symbol *sym;
339
340 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
341 if (kernel_dso == NULL)
342 return -1;
343
344 if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0)
345 goto out_delete_dso;
346
347 node = rb_first(&kernel_dso->syms);
348 sym = rb_entry(node, struct symbol, rb_node);
349 min_ip = sym->start;
350
351 node = rb_last(&kernel_dso->syms);
352 sym = rb_entry(node, struct symbol, rb_node);
353 max_ip = sym->end;
354
355 if (dump_symtab)
356 dso__fprintf(kernel_dso, stderr);
357
358 return 0;
359
360out_delete_dso:
361 dso__delete(kernel_dso);
362 kernel_dso = NULL;
363 return -1;
364}
365
366#define TRACE_COUNT 3
367
368/*
369 * Binary search in the histogram table and record the hit:
370 */
371static void record_ip(uint64_t ip, int counter)
372{
373 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
374
375 if (sym != NULL) {
376 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
377
378 if (!syme->skip) {
379 syme->count[counter]++;
380 pthread_mutex_lock(&active_symbols_lock);
381 if (list_empty(&syme->node) || !syme->node.next)
382 __list_insert_active_sym(syme);
383 pthread_mutex_unlock(&active_symbols_lock);
384 return;
385 }
386 }
387
388 samples--;
389}
390
391static void process_event(uint64_t ip, int counter)
392{
393 samples++;
394
395 if (ip < min_ip || ip > max_ip) {
396 userspace_samples++;
397 return;
398 }
399
400 record_ip(ip, counter);
401}
402
403struct mmap_data {
404 int counter;
405 void *base;
406 unsigned int mask;
407 unsigned int prev;
408};
409
410static unsigned int mmap_read_head(struct mmap_data *md)
411{
412 struct perf_counter_mmap_page *pc = md->base;
413 int head;
414
415 head = pc->data_head;
416 rmb();
417
418 return head;
419}
420
421struct timeval last_read, this_read;
422
423static void mmap_read(struct mmap_data *md)
424{
425 unsigned int head = mmap_read_head(md);
426 unsigned int old = md->prev;
427 unsigned char *data = md->base + page_size;
428 int diff;
429
430 gettimeofday(&this_read, NULL);
431
432 /*
433 * If we're further behind than half the buffer, there's a chance
434 * the writer will bite our tail and mess up the samples under us.
435 *
436 * If we somehow ended up ahead of the head, we got messed up.
437 *
438 * In either case, truncate and restart at head.
439 */
440 diff = head - old;
441 if (diff > md->mask / 2 || diff < 0) {
442 struct timeval iv;
443 unsigned long msecs;
444
445 timersub(&this_read, &last_read, &iv);
446 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
447
448 fprintf(stderr, "WARNING: failed to keep up with mmap data."
449 " Last read %lu msecs ago.\n", msecs);
450
451 /*
452 * head points to a known good entry, start there.
453 */
454 old = head;
455 }
456
457 last_read = this_read;
458
459 for (; old != head;) {
460 struct ip_event {
461 struct perf_event_header header;
462 __u64 ip;
463 __u32 pid, target_pid;
464 };
465 struct mmap_event {
466 struct perf_event_header header;
467 __u32 pid, target_pid;
468 __u64 start;
469 __u64 len;
470 __u64 pgoff;
471 char filename[PATH_MAX];
472 };
473
474 typedef union event_union {
475 struct perf_event_header header;
476 struct ip_event ip;
477 struct mmap_event mmap;
478 } event_t;
479
480 event_t *event = (event_t *)&data[old & md->mask];
481
482 event_t event_copy;
483
484 size_t size = event->header.size;
485
486 /*
487 * Event straddles the mmap boundary -- header should always
488 * be inside due to u64 alignment of output.
489 */
490 if ((old & md->mask) + size != ((old + size) & md->mask)) {
491 unsigned int offset = old;
492 unsigned int len = min(sizeof(*event), size), cpy;
493 void *dst = &event_copy;
494
495 do {
496 cpy = min(md->mask + 1 - (offset & md->mask), len);
497 memcpy(dst, &data[offset & md->mask], cpy);
498 offset += cpy;
499 dst += cpy;
500 len -= cpy;
501 } while (len);
502
503 event = &event_copy;
504 }
505
506 old += size;
507
508 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
509 if (event->header.type & PERF_SAMPLE_IP)
510 process_event(event->ip.ip, md->counter);
511 }
512 }
513
514 md->prev = old;
515}
516
517static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
518static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
519
520static int __cmd_top(void)
521{
522 struct perf_counter_attr *attr;
523 pthread_t thread;
524 int i, counter, group_fd, nr_poll = 0;
525 unsigned int cpu;
526 int ret;
527
528 for (i = 0; i < nr_cpus; i++) {
529 group_fd = -1;
530 for (counter = 0; counter < nr_counters; counter++) {
531
532 cpu = profile_cpu;
533 if (target_pid == -1 && profile_cpu == -1)
534 cpu = i;
535
536 attr = attrs + counter;
537
538 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
539 attr->freq = freq;
540
541 fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
542 if (fd[i][counter] < 0) {
543 int err = errno;
544
545 error("syscall returned with %d (%s)\n",
546 fd[i][counter], strerror(err));
547 if (err == EPERM)
548 printf("Are you root?\n");
549 exit(-1);
550 }
551 assert(fd[i][counter] >= 0);
552 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
553
554 /*
555 * First counter acts as the group leader:
556 */
557 if (group && group_fd == -1)
558 group_fd = fd[i][counter];
559
560 event_array[nr_poll].fd = fd[i][counter];
561 event_array[nr_poll].events = POLLIN;
562 nr_poll++;
563
564 mmap_array[i][counter].counter = counter;
565 mmap_array[i][counter].prev = 0;
566 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
567 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
568 PROT_READ, MAP_SHARED, fd[i][counter], 0);
569 if (mmap_array[i][counter].base == MAP_FAILED)
570 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
571 }
572 }
573
574 if (pthread_create(&thread, NULL, display_thread, NULL)) {
575 printf("Could not create display thread.\n");
576 exit(-1);
577 }
578
579 if (realtime_prio) {
580 struct sched_param param;
581
582 param.sched_priority = realtime_prio;
583 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
584 printf("Could not set realtime priority.\n");
585 exit(-1);
586 }
587 }
588
589 while (1) {
590 int hits = samples;
591
592 for (i = 0; i < nr_cpus; i++) {
593 for (counter = 0; counter < nr_counters; counter++)
594 mmap_read(&mmap_array[i][counter]);
595 }
596
597 if (hits == samples)
598 ret = poll(event_array, nr_poll, 100);
599 }
600
601 return 0;
602}
603
604static const char * const top_usage[] = {
605 "perf top [<options>]",
606 NULL
607};
608
609static const struct option options[] = {
610 OPT_CALLBACK('e', "event", NULL, "event",
611 "event selector. use 'perf list' to list available events",
612 parse_events),
613 OPT_INTEGER('c', "count", &default_interval,
614 "event period to sample"),
615 OPT_INTEGER('p', "pid", &target_pid,
616 "profile events on existing pid"),
617 OPT_BOOLEAN('a', "all-cpus", &system_wide,
618 "system-wide collection from all CPUs"),
619 OPT_INTEGER('C', "CPU", &profile_cpu,
620 "CPU to profile on"),
621 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
622 "number of mmap data pages"),
623 OPT_INTEGER('r', "realtime", &realtime_prio,
624 "collect data with this RT SCHED_FIFO priority"),
625 OPT_INTEGER('d', "delay", &delay_secs,
626 "number of seconds to delay between refreshes"),
627 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
628 "dump the symbol table used for profiling"),
629 OPT_INTEGER('f', "count-filter", &count_filter,
630 "only display functions with more events than this"),
631 OPT_BOOLEAN('g', "group", &group,
632 "put the counters into a counter group"),
633 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
634 "only display symbols matchig this pattern"),
635 OPT_BOOLEAN('z', "zero", &group,
636 "zero history across updates"),
637 OPT_INTEGER('F', "freq", &freq,
638 "profile at this frequency"),
639 OPT_INTEGER('E', "entries", &print_entries,
640 "display this many functions"),
641 OPT_END()
642};
643
644int cmd_top(int argc, const char **argv, const char *prefix)
645{
646 int counter;
647
648 page_size = sysconf(_SC_PAGE_SIZE);
649
650 argc = parse_options(argc, argv, options, top_usage, 0);
651 if (argc)
652 usage_with_options(top_usage, options);
653
654 if (freq) {
655 default_interval = freq;
656 freq = 1;
657 }
658
659 /* CPU and PID are mutually exclusive */
660 if (target_pid != -1 && profile_cpu != -1) {
661 printf("WARNING: PID switch overriding CPU\n");
662 sleep(1);
663 profile_cpu = -1;
664 }
665
666 if (!nr_counters)
667 nr_counters = 1;
668
669 if (delay_secs < 1)
670 delay_secs = 1;
671
672 parse_symbols();
673
674 /*
675 * Fill in the ones not specifically initialized via -c:
676 */
677 for (counter = 0; counter < nr_counters; counter++) {
678 if (attrs[counter].sample_period)
679 continue;
680
681 attrs[counter].sample_period = default_interval;
682 }
683
684 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
685 assert(nr_cpus <= MAX_NR_CPUS);
686 assert(nr_cpus >= 0);
687
688 if (target_pid != -1 || profile_cpu != -1)
689 nr_cpus = 1;
690
691 return __cmd_top();
692}
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-21 21:04:14 -0500