]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - Documentation/perf_counter/builtin-top.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf_counter tools: Remove the last nmi bits
[mirror_ubuntu-artful-kernel.git] / Documentation / perf_counter / builtin-top.c
1 /*
2 * kerneltop.c: show top kernel functions - performance counters showcase
3
4 Build with:
5
6 make -C Documentation/perf_counter/
7
8 Sample output:
9
10 ------------------------------------------------------------------------------
11 KernelTop: 2669 irqs/sec [cache-misses/cache-refs], (all, cpu: 2)
12 ------------------------------------------------------------------------------
13
14 weight RIP kernel function
15 ______ ________________ _______________
16
17 35.20 - ffffffff804ce74b : skb_copy_and_csum_dev
18 33.00 - ffffffff804cb740 : sock_alloc_send_skb
19 31.26 - ffffffff804ce808 : skb_push
20 22.43 - ffffffff80510004 : tcp_established_options
21 19.00 - ffffffff8027d250 : find_get_page
22 15.76 - ffffffff804e4fc9 : eth_type_trans
23 15.20 - ffffffff804d8baa : dst_release
24 14.86 - ffffffff804cf5d8 : skb_release_head_state
25 14.00 - ffffffff802217d5 : read_hpet
26 12.00 - ffffffff804ffb7f : __ip_local_out
27 11.97 - ffffffff804fc0c8 : ip_local_deliver_finish
28 8.54 - ffffffff805001a3 : ip_queue_xmit
29 */
30
31 /*
32 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
33 *
34 * Improvements and fixes by:
35 *
36 * Arjan van de Ven <arjan@linux.intel.com>
37 * Yanmin Zhang <yanmin.zhang@intel.com>
38 * Wu Fengguang <fengguang.wu@intel.com>
39 * Mike Galbraith <efault@gmx.de>
40 * Paul Mackerras <paulus@samba.org>
41 *
42 * Released under the GPL v2. (and only v2, not any later version)
43 */
44
45 #include "perf.h"
46 #include "builtin.h"
47 #include "util/symbol.h"
48 #include "util/util.h"
49 #include "util/rbtree.h"
50 #include "util/parse-options.h"
51 #include "util/parse-events.h"
52
53 #include <assert.h>
54 #include <fcntl.h>
55
56 #include <stdio.h>
57
58 #include <errno.h>
59 #include <time.h>
60 #include <sched.h>
61 #include <pthread.h>
62
63 #include <sys/syscall.h>
64 #include <sys/ioctl.h>
65 #include <sys/poll.h>
66 #include <sys/prctl.h>
67 #include <sys/wait.h>
68 #include <sys/uio.h>
69 #include <sys/mman.h>
70
71 #include <linux/unistd.h>
72 #include <linux/types.h>
73
74 static int system_wide = 0;
75
76 static __u64 default_event_id[MAX_COUNTERS] = {
77 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK),
78 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES),
79 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS),
80 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS),
81
82 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES),
83 EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS),
84 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES),
85 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES),
86 };
87 static int default_interval = 100000;
88 static int event_count[MAX_COUNTERS];
89 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
90
91 static __u64 count_filter = 100;
92
93 static int target_pid = -1;
94 static int profile_cpu = -1;
95 static int nr_cpus = 0;
96 static unsigned int realtime_prio = 0;
97 static int group = 0;
98 static unsigned int page_size;
99 static unsigned int mmap_pages = 16;
100 static int use_mmap = 0;
101 static int use_munmap = 0;
102 static int freq = 0;
103
104 static char *sym_filter;
105 static unsigned long filter_start;
106 static unsigned long filter_end;
107
108 static int delay_secs = 2;
109 static int zero;
110 static int dump_symtab;
111
112 static const unsigned int default_count[] = {
113 1000000,
114 1000000,
115 10000,
116 10000,
117 1000000,
118 10000,
119 };
120
121 /*
122 * Symbols
123 */
124
125 static uint64_t min_ip;
126 static uint64_t max_ip = -1ll;
127
128 struct sym_entry {
129 struct rb_node rb_node;
130 struct list_head node;
131 unsigned long count[MAX_COUNTERS];
132 unsigned long snap_count;
133 double weight;
134 int skip;
135 };
136
137 struct sym_entry *sym_filter_entry;
138
139 struct dso *kernel_dso;
140
141 /*
142 * Symbols will be added here in record_ip and will get out
143 * after decayed.
144 */
145 static LIST_HEAD(active_symbols);
146 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
147
148 /*
149 * Ordering weight: count-1 * count-2 * ... / count-n
150 */
151 static double sym_weight(const struct sym_entry *sym)
152 {
153 double weight = sym->snap_count;
154 int counter;
155
156 for (counter = 1; counter < nr_counters-1; counter++)
157 weight *= sym->count[counter];
158
159 weight /= (sym->count[counter] + 1);
160
161 return weight;
162 }
163
164 static long events;
165 static long userspace_events;
166 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
167
168 static void __list_insert_active_sym(struct sym_entry *syme)
169 {
170 list_add(&syme->node, &active_symbols);
171 }
172
173 static void list_remove_active_sym(struct sym_entry *syme)
174 {
175 pthread_mutex_lock(&active_symbols_lock);
176 list_del_init(&syme->node);
177 pthread_mutex_unlock(&active_symbols_lock);
178 }
179
180 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
181 {
182 struct rb_node **p = &tree->rb_node;
183 struct rb_node *parent = NULL;
184 struct sym_entry *iter;
185
186 while (*p != NULL) {
187 parent = *p;
188 iter = rb_entry(parent, struct sym_entry, rb_node);
189
190 if (se->weight > iter->weight)
191 p = &(*p)->rb_left;
192 else
193 p = &(*p)->rb_right;
194 }
195
196 rb_link_node(&se->rb_node, parent, p);
197 rb_insert_color(&se->rb_node, tree);
198 }
199
200 static void print_sym_table(void)
201 {
202 int printed, j;
203 int counter;
204 float events_per_sec = events/delay_secs;
205 float kevents_per_sec = (events-userspace_events)/delay_secs;
206 float sum_kevents = 0.0;
207 struct sym_entry *syme, *n;
208 struct rb_root tmp = RB_ROOT;
209 struct rb_node *nd;
210
211 events = userspace_events = 0;
212
213 /* Sort the active symbols */
214 pthread_mutex_lock(&active_symbols_lock);
215 syme = list_entry(active_symbols.next, struct sym_entry, node);
216 pthread_mutex_unlock(&active_symbols_lock);
217
218 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
219 syme->snap_count = syme->count[0];
220 if (syme->snap_count != 0) {
221 syme->weight = sym_weight(syme);
222 rb_insert_active_sym(&tmp, syme);
223 sum_kevents += syme->snap_count;
224
225 for (j = 0; j < nr_counters; j++)
226 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
227 } else
228 list_remove_active_sym(syme);
229 }
230
231 write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR));
232
233 printf(
234 "------------------------------------------------------------------------------\n");
235 printf( " KernelTop:%8.0f irqs/sec kernel:%4.1f%% [",
236 events_per_sec,
237 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)));
238
239 if (nr_counters == 1)
240 printf("%d ", event_count[0]);
241
242 for (counter = 0; counter < nr_counters; counter++) {
243 if (counter)
244 printf("/");
245
246 printf("%s", event_name(counter));
247 }
248
249 printf( "], ");
250
251 if (target_pid != -1)
252 printf(" (target_pid: %d", target_pid);
253 else
254 printf(" (all");
255
256 if (profile_cpu != -1)
257 printf(", cpu: %d)\n", profile_cpu);
258 else {
259 if (target_pid != -1)
260 printf(")\n");
261 else
262 printf(", %d CPUs)\n", nr_cpus);
263 }
264
265 printf("------------------------------------------------------------------------------\n\n");
266
267 if (nr_counters == 1)
268 printf(" events pcnt");
269 else
270 printf(" weight events pcnt");
271
272 printf(" RIP kernel function\n"
273 " ______ ______ _____ ________________ _______________\n\n"
274 );
275
276 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
277 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
278 struct symbol *sym = (struct symbol *)(syme + 1);
279 float pcnt;
280
281 if (++printed > 18 || syme->snap_count < count_filter)
282 continue;
283
284 pcnt = 100.0 - (100.0 * ((sum_kevents - syme->snap_count) /
285 sum_kevents));
286
287 if (nr_counters == 1)
288 printf("%19.2f - %4.1f%% - %016llx : %s\n",
289 syme->weight, pcnt, sym->start, sym->name);
290 else
291 printf("%8.1f %10ld - %4.1f%% - %016llx : %s\n",
292 syme->weight, syme->snap_count,
293 pcnt, sym->start, sym->name);
294 }
295
296 {
297 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
298
299 if (poll(&stdin_poll, 1, 0) == 1) {
300 printf("key pressed - exiting.\n");
301 exit(0);
302 }
303 }
304 }
305
306 static void *display_thread(void *arg)
307 {
308 printf("KernelTop refresh period: %d seconds\n", delay_secs);
309
310 while (!sleep(delay_secs))
311 print_sym_table();
312
313 return NULL;
314 }
315
316 static int symbol_filter(struct dso *self, struct symbol *sym)
317 {
318 static int filter_match;
319 struct sym_entry *syme;
320 const char *name = sym->name;
321
322 if (!strcmp(name, "_text") ||
323 !strcmp(name, "_etext") ||
324 !strcmp(name, "_sinittext") ||
325 !strncmp("init_module", name, 11) ||
326 !strncmp("cleanup_module", name, 14) ||
327 strstr(name, "_text_start") ||
328 strstr(name, "_text_end"))
329 return 1;
330
331 syme = dso__sym_priv(self, sym);
332 /* Tag events to be skipped. */
333 if (!strcmp("default_idle", name) ||
334 !strcmp("cpu_idle", name) ||
335 !strcmp("enter_idle", name) ||
336 !strcmp("exit_idle", name) ||
337 !strcmp("mwait_idle", name))
338 syme->skip = 1;
339
340 if (filter_match == 1) {
341 filter_end = sym->start;
342 filter_match = -1;
343 if (filter_end - filter_start > 10000) {
344 fprintf(stderr,
345 "hm, too large filter symbol <%s> - skipping.\n",
346 sym_filter);
347 fprintf(stderr, "symbol filter start: %016lx\n",
348 filter_start);
349 fprintf(stderr, " end: %016lx\n",
350 filter_end);
351 filter_end = filter_start = 0;
352 sym_filter = NULL;
353 sleep(1);
354 }
355 }
356
357 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
358 filter_match = 1;
359 filter_start = sym->start;
360 }
361
362
363 return 0;
364 }
365
366 static int parse_symbols(void)
367 {
368 struct rb_node *node;
369 struct symbol *sym;
370
371 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
372 if (kernel_dso == NULL)
373 return -1;
374
375 if (dso__load_kernel(kernel_dso, NULL, symbol_filter) != 0)
376 goto out_delete_dso;
377
378 node = rb_first(&kernel_dso->syms);
379 sym = rb_entry(node, struct symbol, rb_node);
380 min_ip = sym->start;
381
382 node = rb_last(&kernel_dso->syms);
383 sym = rb_entry(node, struct symbol, rb_node);
384 max_ip = sym->end;
385
386 if (dump_symtab)
387 dso__fprintf(kernel_dso, stderr);
388
389 return 0;
390
391 out_delete_dso:
392 dso__delete(kernel_dso);
393 kernel_dso = NULL;
394 return -1;
395 }
396
397 #define TRACE_COUNT 3
398
399 /*
400 * Binary search in the histogram table and record the hit:
401 */
402 static void record_ip(uint64_t ip, int counter)
403 {
404 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
405
406 if (sym != NULL) {
407 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
408
409 if (!syme->skip) {
410 syme->count[counter]++;
411 pthread_mutex_lock(&active_symbols_lock);
412 if (list_empty(&syme->node) || !syme->node.next)
413 __list_insert_active_sym(syme);
414 pthread_mutex_unlock(&active_symbols_lock);
415 return;
416 }
417 }
418
419 events--;
420 }
421
422 static void process_event(uint64_t ip, int counter)
423 {
424 events++;
425
426 if (ip < min_ip || ip > max_ip) {
427 userspace_events++;
428 return;
429 }
430
431 record_ip(ip, counter);
432 }
433
434 struct mmap_data {
435 int counter;
436 void *base;
437 unsigned int mask;
438 unsigned int prev;
439 };
440
441 static unsigned int mmap_read_head(struct mmap_data *md)
442 {
443 struct perf_counter_mmap_page *pc = md->base;
444 int head;
445
446 head = pc->data_head;
447 rmb();
448
449 return head;
450 }
451
452 struct timeval last_read, this_read;
453
454 static void mmap_read(struct mmap_data *md)
455 {
456 unsigned int head = mmap_read_head(md);
457 unsigned int old = md->prev;
458 unsigned char *data = md->base + page_size;
459 int diff;
460
461 gettimeofday(&this_read, NULL);
462
463 /*
464 * If we're further behind than half the buffer, there's a chance
465 * the writer will bite our tail and screw up the events under us.
466 *
467 * If we somehow ended up ahead of the head, we got messed up.
468 *
469 * In either case, truncate and restart at head.
470 */
471 diff = head - old;
472 if (diff > md->mask / 2 || diff < 0) {
473 struct timeval iv;
474 unsigned long msecs;
475
476 timersub(&this_read, &last_read, &iv);
477 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
478
479 fprintf(stderr, "WARNING: failed to keep up with mmap data."
480 " Last read %lu msecs ago.\n", msecs);
481
482 /*
483 * head points to a known good entry, start there.
484 */
485 old = head;
486 }
487
488 last_read = this_read;
489
490 for (; old != head;) {
491 struct ip_event {
492 struct perf_event_header header;
493 __u64 ip;
494 __u32 pid, target_pid;
495 };
496 struct mmap_event {
497 struct perf_event_header header;
498 __u32 pid, target_pid;
499 __u64 start;
500 __u64 len;
501 __u64 pgoff;
502 char filename[PATH_MAX];
503 };
504
505 typedef union event_union {
506 struct perf_event_header header;
507 struct ip_event ip;
508 struct mmap_event mmap;
509 } event_t;
510
511 event_t *event = (event_t *)&data[old & md->mask];
512
513 event_t event_copy;
514
515 size_t size = event->header.size;
516
517 /*
518 * Event straddles the mmap boundary -- header should always
519 * be inside due to u64 alignment of output.
520 */
521 if ((old & md->mask) + size != ((old + size) & md->mask)) {
522 unsigned int offset = old;
523 unsigned int len = min(sizeof(*event), size), cpy;
524 void *dst = &event_copy;
525
526 do {
527 cpy = min(md->mask + 1 - (offset & md->mask), len);
528 memcpy(dst, &data[offset & md->mask], cpy);
529 offset += cpy;
530 dst += cpy;
531 len -= cpy;
532 } while (len);
533
534 event = &event_copy;
535 }
536
537 old += size;
538
539 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
540 if (event->header.type & PERF_RECORD_IP)
541 process_event(event->ip.ip, md->counter);
542 } else {
543 switch (event->header.type) {
544 case PERF_EVENT_MMAP:
545 case PERF_EVENT_MUNMAP:
546 printf("%s: %Lu %Lu %Lu %s\n",
547 event->header.type == PERF_EVENT_MMAP
548 ? "mmap" : "munmap",
549 event->mmap.start,
550 event->mmap.len,
551 event->mmap.pgoff,
552 event->mmap.filename);
553 break;
554 }
555 }
556 }
557
558 md->prev = old;
559 }
560
561 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
562 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
563
564 static int __cmd_top(void)
565 {
566 struct perf_counter_hw_event hw_event;
567 pthread_t thread;
568 int i, counter, group_fd, nr_poll = 0;
569 unsigned int cpu;
570 int ret;
571
572 for (i = 0; i < nr_cpus; i++) {
573 group_fd = -1;
574 for (counter = 0; counter < nr_counters; counter++) {
575
576 cpu = profile_cpu;
577 if (target_pid == -1 && profile_cpu == -1)
578 cpu = i;
579
580 memset(&hw_event, 0, sizeof(hw_event));
581 hw_event.config = event_id[counter];
582 hw_event.irq_period = event_count[counter];
583 hw_event.record_type = PERF_RECORD_IP | PERF_RECORD_TID;
584 hw_event.mmap = use_mmap;
585 hw_event.munmap = use_munmap;
586 hw_event.freq = freq;
587
588 fd[i][counter] = sys_perf_counter_open(&hw_event, target_pid, cpu, group_fd, 0);
589 if (fd[i][counter] < 0) {
590 int err = errno;
591 printf("kerneltop error: syscall returned with %d (%s)\n",
592 fd[i][counter], strerror(err));
593 if (err == EPERM)
594 printf("Are you root?\n");
595 exit(-1);
596 }
597 assert(fd[i][counter] >= 0);
598 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
599
600 /*
601 * First counter acts as the group leader:
602 */
603 if (group && group_fd == -1)
604 group_fd = fd[i][counter];
605
606 event_array[nr_poll].fd = fd[i][counter];
607 event_array[nr_poll].events = POLLIN;
608 nr_poll++;
609
610 mmap_array[i][counter].counter = counter;
611 mmap_array[i][counter].prev = 0;
612 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
613 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
614 PROT_READ, MAP_SHARED, fd[i][counter], 0);
615 if (mmap_array[i][counter].base == MAP_FAILED) {
616 printf("kerneltop error: failed to mmap with %d (%s)\n",
617 errno, strerror(errno));
618 exit(-1);
619 }
620 }
621 }
622
623 if (pthread_create(&thread, NULL, display_thread, NULL)) {
624 printf("Could not create display thread.\n");
625 exit(-1);
626 }
627
628 if (realtime_prio) {
629 struct sched_param param;
630
631 param.sched_priority = realtime_prio;
632 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
633 printf("Could not set realtime priority.\n");
634 exit(-1);
635 }
636 }
637
638 while (1) {
639 int hits = events;
640
641 for (i = 0; i < nr_cpus; i++) {
642 for (counter = 0; counter < nr_counters; counter++)
643 mmap_read(&mmap_array[i][counter]);
644 }
645
646 if (hits == events)
647 ret = poll(event_array, nr_poll, 100);
648 }
649
650 return 0;
651 }
652
653 static const char * const top_usage[] = {
654 "perf top [<options>]",
655 NULL
656 };
657
658 static char events_help_msg[EVENTS_HELP_MAX];
659
660 static const struct option options[] = {
661 OPT_CALLBACK('e', "event", NULL, "event",
662 events_help_msg, parse_events),
663 OPT_INTEGER('c', "count", &default_interval,
664 "event period to sample"),
665 OPT_INTEGER('p', "pid", &target_pid,
666 "profile events on existing pid"),
667 OPT_BOOLEAN('a', "all-cpus", &system_wide,
668 "system-wide collection from all CPUs"),
669 OPT_INTEGER('C', "CPU", &profile_cpu,
670 "CPU to profile on"),
671 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
672 "number of mmap data pages"),
673 OPT_INTEGER('r', "realtime", &realtime_prio,
674 "collect data with this RT SCHED_FIFO priority"),
675 OPT_INTEGER('d', "delay", &delay_secs,
676 "number of seconds to delay between refreshes"),
677 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
678 "dump the symbol table used for profiling"),
679 OPT_INTEGER('f', "--count-filter", &count_filter,
680 "only display functions with more events than this"),
681 OPT_BOOLEAN('g', "group", &group,
682 "put the counters into a counter group"),
683 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
684 "only display symbols matchig this pattern"),
685 OPT_BOOLEAN('z', "zero", &group,
686 "zero history across updates"),
687 OPT_BOOLEAN('M', "use-mmap", &use_mmap,
688 "track mmap events"),
689 OPT_BOOLEAN('U', "use-munmap", &use_munmap,
690 "track munmap events"),
691 OPT_INTEGER('F', "--freq", &freq,
692 "profile at this frequency"),
693 OPT_END()
694 };
695
696 int cmd_top(int argc, const char **argv, const char *prefix)
697 {
698 int counter;
699
700 page_size = sysconf(_SC_PAGE_SIZE);
701
702 create_events_help(events_help_msg);
703 memcpy(event_id, default_event_id, sizeof(default_event_id));
704
705 argc = parse_options(argc, argv, options, top_usage, 0);
706 if (argc)
707 usage_with_options(top_usage, options);
708
709 if (freq) {
710 default_interval = freq;
711 freq = 1;
712 }
713
714 /* CPU and PID are mutually exclusive */
715 if (target_pid != -1 && profile_cpu != -1) {
716 printf("WARNING: PID switch overriding CPU\n");
717 sleep(1);
718 profile_cpu = -1;
719 }
720
721 if (!nr_counters) {
722 nr_counters = 1;
723 event_id[0] = 0;
724 }
725
726 for (counter = 0; counter < nr_counters; counter++) {
727 if (event_count[counter])
728 continue;
729
730 event_count[counter] = default_interval;
731 }
732
733 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
734 assert(nr_cpus <= MAX_NR_CPUS);
735 assert(nr_cpus >= 0);
736
737 if (target_pid != -1 || profile_cpu != -1)
738 nr_cpus = 1;
739
740 parse_symbols();
741
742 return __cmd_top();
743 }
Ubuntu Artful kernel mirror