?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf_counter tools: Handle kernels with !CONFIG_PERF_COUNTER
[linux-2.6/verdex.git] / tools / perf / builtin-top.c
blobbe1698f1189b7546659adb38bf90a6dc8928f562
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
51 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
52
53 static int system_wide = 0;
54
55 static int default_interval = 100000;
56
57 static __u64 count_filter = 5;
58 static int print_entries = 15;
59
60 static int target_pid = -1;
61 static int profile_cpu = -1;
62 static int nr_cpus = 0;
63 static unsigned int realtime_prio = 0;
64 static int group = 0;
65 static unsigned int page_size;
66 static unsigned int mmap_pages = 16;
67 static int freq = 0;
68 static int verbose = 0;
69
70 static char *sym_filter;
71 static unsigned long filter_start;
72 static unsigned long filter_end;
73
74 static int delay_secs = 2;
75 static int zero;
76 static int dump_symtab;
77
78 /*
79 * Symbols
80 */
81
82 static uint64_t min_ip;
83 static uint64_t max_ip = -1ll;
84
85 struct sym_entry {
86 struct rb_node rb_node;
87 struct list_head node;
88 unsigned long count[MAX_COUNTERS];
89 unsigned long snap_count;
90 double weight;
91 int skip;
92 };
93
94 struct sym_entry *sym_filter_entry;
95
96 struct dso *kernel_dso;
97
98 /*
99 * Symbols will be added here in record_ip and will get out
100 * after decayed.
101 */
102 static LIST_HEAD(active_symbols);
103 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
104
105 /*
106 * Ordering weight: count-1 * count-2 * ... / count-n
107 */
108 static double sym_weight(const struct sym_entry *sym)
109 {
110 double weight = sym->snap_count;
111 int counter;
112
113 for (counter = 1; counter < nr_counters-1; counter++)
114 weight *= sym->count[counter];
115
116 weight /= (sym->count[counter] + 1);
117
118 return weight;
119 }
120
121 static long samples;
122 static long userspace_samples;
123 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
124
125 static void __list_insert_active_sym(struct sym_entry *syme)
126 {
127 list_add(&syme->node, &active_symbols);
128 }
129
130 static void list_remove_active_sym(struct sym_entry *syme)
131 {
132 pthread_mutex_lock(&active_symbols_lock);
133 list_del_init(&syme->node);
134 pthread_mutex_unlock(&active_symbols_lock);
135 }
136
137 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
138 {
139 struct rb_node **p = &tree->rb_node;
140 struct rb_node *parent = NULL;
141 struct sym_entry *iter;
142
143 while (*p != NULL) {
144 parent = *p;
145 iter = rb_entry(parent, struct sym_entry, rb_node);
146
147 if (se->weight > iter->weight)
148 p = &(*p)->rb_left;
149 else
150 p = &(*p)->rb_right;
151 }
152
153 rb_link_node(&se->rb_node, parent, p);
154 rb_insert_color(&se->rb_node, tree);
155 }
156
157 static void print_sym_table(void)
158 {
159 int printed = 0, j;
160 int counter;
161 float samples_per_sec = samples/delay_secs;
162 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
163 float sum_ksamples = 0.0;
164 struct sym_entry *syme, *n;
165 struct rb_root tmp = RB_ROOT;
166 struct rb_node *nd;
167
168 samples = userspace_samples = 0;
169
170 /* Sort the active symbols */
171 pthread_mutex_lock(&active_symbols_lock);
172 syme = list_entry(active_symbols.next, struct sym_entry, node);
173 pthread_mutex_unlock(&active_symbols_lock);
174
175 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
176 syme->snap_count = syme->count[0];
177 if (syme->snap_count != 0) {
178 syme->weight = sym_weight(syme);
179 rb_insert_active_sym(&tmp, syme);
180 sum_ksamples += syme->snap_count;
181
182 for (j = 0; j < nr_counters; j++)
183 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
184 } else
185 list_remove_active_sym(syme);
186 }
187
188 puts(CONSOLE_CLEAR);
189
190 printf(
191 "------------------------------------------------------------------------------\n");
192 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
193 samples_per_sec,
194 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
195
196 if (nr_counters == 1) {
197 printf("%Ld", attrs[0].sample_period);
198 if (freq)
199 printf("Hz ");
200 else
201 printf(" ");
202 }
203
204 for (counter = 0; counter < nr_counters; counter++) {
205 if (counter)
206 printf("/");
207
208 printf("%s", event_name(counter));
209 }
210
211 printf( "], ");
212
213 if (target_pid != -1)
214 printf(" (target_pid: %d", target_pid);
215 else
216 printf(" (all");
217
218 if (profile_cpu != -1)
219 printf(", cpu: %d)\n", profile_cpu);
220 else {
221 if (target_pid != -1)
222 printf(")\n");
223 else
224 printf(", %d CPUs)\n", nr_cpus);
225 }
226
227 printf("------------------------------------------------------------------------------\n\n");
228
229 if (nr_counters == 1)
230 printf(" samples pcnt");
231 else
232 printf(" weight samples pcnt");
233
234 printf(" RIP kernel function\n"
235 " ______ _______ _____ ________________ _______________\n\n"
236 );
237
238 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
239 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
240 struct symbol *sym = (struct symbol *)(syme + 1);
241 char *color = PERF_COLOR_NORMAL;
242 double pcnt;
243
244 if (++printed > print_entries || syme->snap_count < count_filter)
245 continue;
246
247 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
248 sum_ksamples));
249
250 /*
251 * We color high-overhead entries in red, low-overhead
252 * entries in green - and keep the middle ground normal:
253 */
254 if (pcnt >= 5.0)
255 color = PERF_COLOR_RED;
256 if (pcnt < 0.5)
257 color = PERF_COLOR_GREEN;
258
259 if (nr_counters == 1)
260 printf("%20.2f - ", syme->weight);
261 else
262 printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
263
264 color_fprintf(stdout, color, "%4.1f%%", pcnt);
265 printf(" - %016llx : %s\n", sym->start, sym->name);
266 }
267 }
268
269 static void *display_thread(void *arg)
270 {
271 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
272 int delay_msecs = delay_secs * 1000;
273
274 printf("PerfTop refresh period: %d seconds\n", delay_secs);
275
276 do {
277 print_sym_table();
278 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
279
280 printf("key pressed - exiting.\n");
281 exit(0);
282
283 return NULL;
284 }
285
286 static int symbol_filter(struct dso *self, struct symbol *sym)
287 {
288 static int filter_match;
289 struct sym_entry *syme;
290 const char *name = sym->name;
291
292 if (!strcmp(name, "_text") ||
293 !strcmp(name, "_etext") ||
294 !strcmp(name, "_sinittext") ||
295 !strncmp("init_module", name, 11) ||
296 !strncmp("cleanup_module", name, 14) ||
297 strstr(name, "_text_start") ||
298 strstr(name, "_text_end"))
299 return 1;
300
301 syme = dso__sym_priv(self, sym);
302 /* Tag samples to be skipped. */
303 if (!strcmp("default_idle", name) ||
304 !strcmp("cpu_idle", name) ||
305 !strcmp("enter_idle", name) ||
306 !strcmp("exit_idle", name) ||
307 !strcmp("mwait_idle", name))
308 syme->skip = 1;
309
310 if (filter_match == 1) {
311 filter_end = sym->start;
312 filter_match = -1;
313 if (filter_end - filter_start > 10000) {
314 fprintf(stderr,
315 "hm, too large filter symbol <%s> - skipping.\n",
316 sym_filter);
317 fprintf(stderr, "symbol filter start: %016lx\n",
318 filter_start);
319 fprintf(stderr, " end: %016lx\n",
320 filter_end);
321 filter_end = filter_start = 0;
322 sym_filter = NULL;
323 sleep(1);
324 }
325 }
326
327 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
328 filter_match = 1;
329 filter_start = sym->start;
330 }
331
332
333 return 0;
334 }
335
336 static int parse_symbols(void)
337 {
338 struct rb_node *node;
339 struct symbol *sym;
340
341 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
342 if (kernel_dso == NULL)
343 return -1;
344
345 if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0)
346 goto out_delete_dso;
347
348 node = rb_first(&kernel_dso->syms);
349 sym = rb_entry(node, struct symbol, rb_node);
350 min_ip = sym->start;
351
352 node = rb_last(&kernel_dso->syms);
353 sym = rb_entry(node, struct symbol, rb_node);
354 max_ip = sym->end;
355
356 if (dump_symtab)
357 dso__fprintf(kernel_dso, stderr);
358
359 return 0;
360
361 out_delete_dso:
362 dso__delete(kernel_dso);
363 kernel_dso = NULL;
364 return -1;
365 }
366
367 #define TRACE_COUNT 3
368
369 /*
370 * Binary search in the histogram table and record the hit:
371 */
372 static void record_ip(uint64_t ip, int counter)
373 {
374 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
375
376 if (sym != NULL) {
377 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
378
379 if (!syme->skip) {
380 syme->count[counter]++;
381 pthread_mutex_lock(&active_symbols_lock);
382 if (list_empty(&syme->node) || !syme->node.next)
383 __list_insert_active_sym(syme);
384 pthread_mutex_unlock(&active_symbols_lock);
385 return;
386 }
387 }
388
389 samples--;
390 }
391
392 static void process_event(uint64_t ip, int counter)
393 {
394 samples++;
395
396 if (ip < min_ip || ip > max_ip) {
397 userspace_samples++;
398 return;
399 }
400
401 record_ip(ip, counter);
402 }
403
404 struct mmap_data {
405 int counter;
406 void *base;
407 unsigned int mask;
408 unsigned int prev;
409 };
410
411 static unsigned int mmap_read_head(struct mmap_data *md)
412 {
413 struct perf_counter_mmap_page *pc = md->base;
414 int head;
415
416 head = pc->data_head;
417 rmb();
418
419 return head;
420 }
421
422 struct timeval last_read, this_read;
423
424 static void mmap_read_counter(struct mmap_data *md)
425 {
426 unsigned int head = mmap_read_head(md);
427 unsigned int old = md->prev;
428 unsigned char *data = md->base + page_size;
429 int diff;
430
431 gettimeofday(&this_read, NULL);
432
433 /*
434 * If we're further behind than half the buffer, there's a chance
435 * the writer will bite our tail and mess up the samples under us.
436 *
437 * If we somehow ended up ahead of the head, we got messed up.
438 *
439 * In either case, truncate and restart at head.
440 */
441 diff = head - old;
442 if (diff > md->mask / 2 || diff < 0) {
443 struct timeval iv;
444 unsigned long msecs;
445
446 timersub(&this_read, &last_read, &iv);
447 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
448
449 fprintf(stderr, "WARNING: failed to keep up with mmap data."
450 " Last read %lu msecs ago.\n", msecs);
451
452 /*
453 * head points to a known good entry, start there.
454 */
455 old = head;
456 }
457
458 last_read = this_read;
459
460 for (; old != head;) {
461 struct ip_event {
462 struct perf_event_header header;
463 __u64 ip;
464 __u32 pid, target_pid;
465 };
466 struct mmap_event {
467 struct perf_event_header header;
468 __u32 pid, target_pid;
469 __u64 start;
470 __u64 len;
471 __u64 pgoff;
472 char filename[PATH_MAX];
473 };
474
475 typedef union event_union {
476 struct perf_event_header header;
477 struct ip_event ip;
478 struct mmap_event mmap;
479 } event_t;
480
481 event_t *event = (event_t *)&data[old & md->mask];
482
483 event_t event_copy;
484
485 size_t size = event->header.size;
486
487 /*
488 * Event straddles the mmap boundary -- header should always
489 * be inside due to u64 alignment of output.
490 */
491 if ((old & md->mask) + size != ((old + size) & md->mask)) {
492 unsigned int offset = old;
493 unsigned int len = min(sizeof(*event), size), cpy;
494 void *dst = &event_copy;
495
496 do {
497 cpy = min(md->mask + 1 - (offset & md->mask), len);
498 memcpy(dst, &data[offset & md->mask], cpy);
499 offset += cpy;
500 dst += cpy;
501 len -= cpy;
502 } while (len);
503
504 event = &event_copy;
505 }
506
507 old += size;
508
509 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
510 if (event->header.type & PERF_SAMPLE_IP)
511 process_event(event->ip.ip, md->counter);
512 }
513 }
514
515 md->prev = old;
516 }
517
518 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
519 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
520
521 static void mmap_read(void)
522 {
523 int i, counter;
524
525 for (i = 0; i < nr_cpus; i++) {
526 for (counter = 0; counter < nr_counters; counter++)
527 mmap_read_counter(&mmap_array[i][counter]);
528 }
529 }
530
531 int nr_poll;
532 int group_fd;
533
534 static void start_counter(int i, int counter)
535 {
536 struct perf_counter_attr *attr;
537 unsigned int cpu;
538
539 cpu = profile_cpu;
540 if (target_pid == -1 && profile_cpu == -1)
541 cpu = i;
542
543 attr = attrs + counter;
544
545 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
546 attr->freq = freq;
547
548 try_again:
549 fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
550
551 if (fd[i][counter] < 0) {
552 int err = errno;
553
554 if (err == EPERM)
555 die("No permission - are you root?\n");
556 /*
557 * If it's cycles then fall back to hrtimer
558 * based cpu-clock-tick sw counter, which
559 * is always available even if no PMU support:
560 */
561 if (attr->type == PERF_TYPE_HARDWARE
562 && attr->config == PERF_COUNT_CPU_CYCLES) {
563
564 if (verbose)
565 warning(" ... trying to fall back to cpu-clock-ticks\n");
566
567 attr->type = PERF_TYPE_SOFTWARE;
568 attr->config = PERF_COUNT_CPU_CLOCK;
569 goto try_again;
570 }
571 printf("\n");
572 error("perfcounter syscall returned with %d (%s)\n",
573 fd[i][counter], strerror(err));
574 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
575 exit(-1);
576 }
577 assert(fd[i][counter] >= 0);
578 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
579
580 /*
581 * First counter acts as the group leader:
582 */
583 if (group && group_fd == -1)
584 group_fd = fd[i][counter];
585
586 event_array[nr_poll].fd = fd[i][counter];
587 event_array[nr_poll].events = POLLIN;
588 nr_poll++;
589
590 mmap_array[i][counter].counter = counter;
591 mmap_array[i][counter].prev = 0;
592 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
593 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
594 PROT_READ, MAP_SHARED, fd[i][counter], 0);
595 if (mmap_array[i][counter].base == MAP_FAILED)
596 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
597 }
598
599 static int __cmd_top(void)
600 {
601 pthread_t thread;
602 int i, counter;
603 int ret;
604
605 for (i = 0; i < nr_cpus; i++) {
606 group_fd = -1;
607 for (counter = 0; counter < nr_counters; counter++)
608 start_counter(i, counter);
609 }
610
611 /* Wait for a minimal set of events before starting the snapshot */
612 poll(event_array, nr_poll, 100);
613
614 mmap_read();
615
616 if (pthread_create(&thread, NULL, display_thread, NULL)) {
617 printf("Could not create display thread.\n");
618 exit(-1);
619 }
620
621 if (realtime_prio) {
622 struct sched_param param;
623
624 param.sched_priority = realtime_prio;
625 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
626 printf("Could not set realtime priority.\n");
627 exit(-1);
628 }
629 }
630
631 while (1) {
632 int hits = samples;
633
634 mmap_read();
635
636 if (hits == samples)
637 ret = poll(event_array, nr_poll, 100);
638 }
639
640 return 0;
641 }
642
643 static const char * const top_usage[] = {
644 "perf top [<options>]",
645 NULL
646 };
647
648 static const struct option options[] = {
649 OPT_CALLBACK('e', "event", NULL, "event",
650 "event selector. use 'perf list' to list available events",
651 parse_events),
652 OPT_INTEGER('c', "count", &default_interval,
653 "event period to sample"),
654 OPT_INTEGER('p', "pid", &target_pid,
655 "profile events on existing pid"),
656 OPT_BOOLEAN('a', "all-cpus", &system_wide,
657 "system-wide collection from all CPUs"),
658 OPT_INTEGER('C', "CPU", &profile_cpu,
659 "CPU to profile on"),
660 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
661 "number of mmap data pages"),
662 OPT_INTEGER('r', "realtime", &realtime_prio,
663 "collect data with this RT SCHED_FIFO priority"),
664 OPT_INTEGER('d', "delay", &delay_secs,
665 "number of seconds to delay between refreshes"),
666 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
667 "dump the symbol table used for profiling"),
668 OPT_INTEGER('f', "count-filter", &count_filter,
669 "only display functions with more events than this"),
670 OPT_BOOLEAN('g', "group", &group,
671 "put the counters into a counter group"),
672 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
673 "only display symbols matchig this pattern"),
674 OPT_BOOLEAN('z', "zero", &group,
675 "zero history across updates"),
676 OPT_INTEGER('F', "freq", &freq,
677 "profile at this frequency"),
678 OPT_INTEGER('E', "entries", &print_entries,
679 "display this many functions"),
680 OPT_BOOLEAN('v', "verbose", &verbose,
681 "be more verbose (show counter open errors, etc)"),
682 OPT_END()
683 };
684
685 int cmd_top(int argc, const char **argv, const char *prefix)
686 {
687 int counter;
688
689 page_size = sysconf(_SC_PAGE_SIZE);
690
691 argc = parse_options(argc, argv, options, top_usage, 0);
692 if (argc)
693 usage_with_options(top_usage, options);
694
695 if (freq) {
696 default_interval = freq;
697 freq = 1;
698 }
699
700 /* CPU and PID are mutually exclusive */
701 if (target_pid != -1 && profile_cpu != -1) {
702 printf("WARNING: PID switch overriding CPU\n");
703 sleep(1);
704 profile_cpu = -1;
705 }
706
707 if (!nr_counters)
708 nr_counters = 1;
709
710 if (delay_secs < 1)
711 delay_secs = 1;
712
713 parse_symbols();
714
715 /*
716 * Fill in the ones not specifically initialized via -c:
717 */
718 for (counter = 0; counter < nr_counters; counter++) {
719 if (attrs[counter].sample_period)
720 continue;
721
722 attrs[counter].sample_period = default_interval;
723 }
724
725 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
726 assert(nr_cpus <= MAX_NR_CPUS);
727 assert(nr_cpus >= 0);
728
729 if (target_pid != -1 || profile_cpu != -1)
730 nr_cpus = 1;
731
732 return __cmd_top();
733 }
Verdex Pro support