perf_counter tools: Clarify events/samples naming
[safe/jmp/linux-2.6] / Documentation / perf_counter / builtin-top.c
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                      system_wide                     =  0;
52
53 static __u64                    default_event_id[MAX_COUNTERS]          = {
54         EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK),
55         EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES),
56         EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS),
57         EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS),
58
59         EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES),
60         EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS),
61         EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES),
62         EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES),
63 };
64 static int                      default_interval = 100000;
65 static int                      event_count[MAX_COUNTERS];
66 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
67
68 static __u64                    count_filter                    =  5;
69 static int                      print_entries                   = 15;
70
71 static int                      target_pid                      = -1;
72 static int                      profile_cpu                     = -1;
73 static int                      nr_cpus                         =  0;
74 static unsigned int             realtime_prio                   =  0;
75 static int                      group                           =  0;
76 static unsigned int             page_size;
77 static unsigned int             mmap_pages                      = 16;
78 static int                      freq                            =  0;
79
80 static char                     *sym_filter;
81 static unsigned long            filter_start;
82 static unsigned long            filter_end;
83
84 static int                      delay_secs                      =  2;
85 static int                      zero;
86 static int                      dump_symtab;
87
88 static const unsigned int default_count[] = {
89         1000000,
90         1000000,
91           10000,
92           10000,
93         1000000,
94           10000,
95 };
96
97 /*
98  * Symbols
99  */
100
101 static uint64_t                 min_ip;
102 static uint64_t                 max_ip = -1ll;
103
104 struct sym_entry {
105         struct rb_node          rb_node;
106         struct list_head        node;
107         unsigned long           count[MAX_COUNTERS];
108         unsigned long           snap_count;
109         double                  weight;
110         int                     skip;
111 };
112
113 struct sym_entry                *sym_filter_entry;
114
115 struct dso *kernel_dso;
116
117 /*
118  * Symbols will be added here in record_ip and will get out
119  * after decayed.
120  */
121 static LIST_HEAD(active_symbols);
122 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
123
124 /*
125  * Ordering weight: count-1 * count-2 * ... / count-n
126  */
127 static double sym_weight(const struct sym_entry *sym)
128 {
129         double weight = sym->snap_count;
130         int counter;
131
132         for (counter = 1; counter < nr_counters-1; counter++)
133                 weight *= sym->count[counter];
134
135         weight /= (sym->count[counter] + 1);
136
137         return weight;
138 }
139
140 static long                     samples;
141 static long                     userspace_samples;
142 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
143
144 static void __list_insert_active_sym(struct sym_entry *syme)
145 {
146         list_add(&syme->node, &active_symbols);
147 }
148
149 static void list_remove_active_sym(struct sym_entry *syme)
150 {
151         pthread_mutex_lock(&active_symbols_lock);
152         list_del_init(&syme->node);
153         pthread_mutex_unlock(&active_symbols_lock);
154 }
155
156 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
157 {
158         struct rb_node **p = &tree->rb_node;
159         struct rb_node *parent = NULL;
160         struct sym_entry *iter;
161
162         while (*p != NULL) {
163                 parent = *p;
164                 iter = rb_entry(parent, struct sym_entry, rb_node);
165
166                 if (se->weight > iter->weight)
167                         p = &(*p)->rb_left;
168                 else
169                         p = &(*p)->rb_right;
170         }
171
172         rb_link_node(&se->rb_node, parent, p);
173         rb_insert_color(&se->rb_node, tree);
174 }
175
176 static void print_sym_table(void)
177 {
178         int printed = 0, j;
179         int counter;
180         float samples_per_sec = samples/delay_secs;
181         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
182         float sum_ksamples = 0.0;
183         struct sym_entry *syme, *n;
184         struct rb_root tmp = RB_ROOT;
185         struct rb_node *nd;
186
187         samples = userspace_samples = 0;
188
189         /* Sort the active symbols */
190         pthread_mutex_lock(&active_symbols_lock);
191         syme = list_entry(active_symbols.next, struct sym_entry, node);
192         pthread_mutex_unlock(&active_symbols_lock);
193
194         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
195                 syme->snap_count = syme->count[0];
196                 if (syme->snap_count != 0) {
197                         syme->weight = sym_weight(syme);
198                         rb_insert_active_sym(&tmp, syme);
199                         sum_ksamples += syme->snap_count;
200
201                         for (j = 0; j < nr_counters; j++)
202                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
203                 } else
204                         list_remove_active_sym(syme);
205         }
206
207         puts(CONSOLE_CLEAR);
208
209         printf(
210 "------------------------------------------------------------------------------\n");
211         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
212                 samples_per_sec,
213                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
214
215         if (nr_counters == 1) {
216                 printf("%d", event_count[0]);
217                 if (freq)
218                         printf("Hz ");
219                 else
220                         printf(" ");
221         }
222
223         for (counter = 0; counter < nr_counters; counter++) {
224                 if (counter)
225                         printf("/");
226
227                 printf("%s", event_name(counter));
228         }
229
230         printf( "], ");
231
232         if (target_pid != -1)
233                 printf(" (target_pid: %d", target_pid);
234         else
235                 printf(" (all");
236
237         if (profile_cpu != -1)
238                 printf(", cpu: %d)\n", profile_cpu);
239         else {
240                 if (target_pid != -1)
241                         printf(")\n");
242                 else
243                         printf(", %d CPUs)\n", nr_cpus);
244         }
245
246         printf("------------------------------------------------------------------------------\n\n");
247
248         if (nr_counters == 1)
249                 printf("             samples    pcnt");
250         else
251                 printf("  weight     samples    pcnt");
252
253         printf("         RIP          kernel function\n"
254                        "  ______     _______   _____   ________________   _______________\n\n"
255         );
256
257         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
258                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
259                 struct symbol *sym = (struct symbol *)(syme + 1);
260                 char *color = PERF_COLOR_NORMAL;
261                 double pcnt;
262
263                 if (++printed > print_entries || syme->snap_count < count_filter)
264                         continue;
265
266                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
267                                          sum_ksamples));
268
269                 /*
270                  * We color high-overhead entries in red, low-overhead
271                  * entries in green - and keep the middle ground normal:
272                  */
273                 if (pcnt >= 5.0)
274                         color = PERF_COLOR_RED;
275                 if (pcnt < 0.5)
276                         color = PERF_COLOR_GREEN;
277
278                 if (nr_counters == 1)
279                         printf("%20.2f - ", syme->weight);
280                 else
281                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
282
283                 color_fprintf(stdout, color, "%4.1f%%", pcnt);
284                 printf(" - %016llx : %s\n", sym->start, sym->name);
285         }
286 }
287
288 static void *display_thread(void *arg)
289 {
290         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
291         int delay_msecs = delay_secs * 1000;
292
293         printf("PerfTop refresh period: %d seconds\n", delay_secs);
294
295         do {
296                 print_sym_table();
297         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
298
299         printf("key pressed - exiting.\n");
300         exit(0);
301
302         return NULL;
303 }
304
305 static int symbol_filter(struct dso *self, struct symbol *sym)
306 {
307         static int filter_match;
308         struct sym_entry *syme;
309         const char *name = sym->name;
310
311         if (!strcmp(name, "_text") ||
312             !strcmp(name, "_etext") ||
313             !strcmp(name, "_sinittext") ||
314             !strncmp("init_module", name, 11) ||
315             !strncmp("cleanup_module", name, 14) ||
316             strstr(name, "_text_start") ||
317             strstr(name, "_text_end"))
318                 return 1;
319
320         syme = dso__sym_priv(self, sym);
321         /* Tag samples to be skipped. */
322         if (!strcmp("default_idle", name) ||
323             !strcmp("cpu_idle", name) ||
324             !strcmp("enter_idle", name) ||
325             !strcmp("exit_idle", name) ||
326             !strcmp("mwait_idle", name))
327                 syme->skip = 1;
328
329         if (filter_match == 1) {
330                 filter_end = sym->start;
331                 filter_match = -1;
332                 if (filter_end - filter_start > 10000) {
333                         fprintf(stderr,
334                                 "hm, too large filter symbol <%s> - skipping.\n",
335                                 sym_filter);
336                         fprintf(stderr, "symbol filter start: %016lx\n",
337                                 filter_start);
338                         fprintf(stderr, "                end: %016lx\n",
339                                 filter_end);
340                         filter_end = filter_start = 0;
341                         sym_filter = NULL;
342                         sleep(1);
343                 }
344         }
345
346         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
347                 filter_match = 1;
348                 filter_start = sym->start;
349         }
350
351
352         return 0;
353 }
354
355 static int parse_symbols(void)
356 {
357         struct rb_node *node;
358         struct symbol  *sym;
359
360         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
361         if (kernel_dso == NULL)
362                 return -1;
363
364         if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0)
365                 goto out_delete_dso;
366
367         node = rb_first(&kernel_dso->syms);
368         sym = rb_entry(node, struct symbol, rb_node);
369         min_ip = sym->start;
370
371         node = rb_last(&kernel_dso->syms);
372         sym = rb_entry(node, struct symbol, rb_node);
373         max_ip = sym->end;
374
375         if (dump_symtab)
376                 dso__fprintf(kernel_dso, stderr);
377
378         return 0;
379
380 out_delete_dso:
381         dso__delete(kernel_dso);
382         kernel_dso = NULL;
383         return -1;
384 }
385
386 #define TRACE_COUNT     3
387
388 /*
389  * Binary search in the histogram table and record the hit:
390  */
391 static void record_ip(uint64_t ip, int counter)
392 {
393         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
394
395         if (sym != NULL) {
396                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
397
398                 if (!syme->skip) {
399                         syme->count[counter]++;
400                         pthread_mutex_lock(&active_symbols_lock);
401                         if (list_empty(&syme->node) || !syme->node.next)
402                                 __list_insert_active_sym(syme);
403                         pthread_mutex_unlock(&active_symbols_lock);
404                         return;
405                 }
406         }
407
408         samples--;
409 }
410
411 static void process_event(uint64_t ip, int counter)
412 {
413         samples++;
414
415         if (ip < min_ip || ip > max_ip) {
416                 userspace_samples++;
417                 return;
418         }
419
420         record_ip(ip, counter);
421 }
422
423 struct mmap_data {
424         int counter;
425         void *base;
426         unsigned int mask;
427         unsigned int prev;
428 };
429
430 static unsigned int mmap_read_head(struct mmap_data *md)
431 {
432         struct perf_counter_mmap_page *pc = md->base;
433         int head;
434
435         head = pc->data_head;
436         rmb();
437
438         return head;
439 }
440
441 struct timeval last_read, this_read;
442
443 static void mmap_read(struct mmap_data *md)
444 {
445         unsigned int head = mmap_read_head(md);
446         unsigned int old = md->prev;
447         unsigned char *data = md->base + page_size;
448         int diff;
449
450         gettimeofday(&this_read, NULL);
451
452         /*
453          * If we're further behind than half the buffer, there's a chance
454          * the writer will bite our tail and mess up the samples under us.
455          *
456          * If we somehow ended up ahead of the head, we got messed up.
457          *
458          * In either case, truncate and restart at head.
459          */
460         diff = head - old;
461         if (diff > md->mask / 2 || diff < 0) {
462                 struct timeval iv;
463                 unsigned long msecs;
464
465                 timersub(&this_read, &last_read, &iv);
466                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
467
468                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
469                                 "  Last read %lu msecs ago.\n", msecs);
470
471                 /*
472                  * head points to a known good entry, start there.
473                  */
474                 old = head;
475         }
476
477         last_read = this_read;
478
479         for (; old != head;) {
480                 struct ip_event {
481                         struct perf_event_header header;
482                         __u64 ip;
483                         __u32 pid, target_pid;
484                 };
485                 struct mmap_event {
486                         struct perf_event_header header;
487                         __u32 pid, target_pid;
488                         __u64 start;
489                         __u64 len;
490                         __u64 pgoff;
491                         char filename[PATH_MAX];
492                 };
493
494                 typedef union event_union {
495                         struct perf_event_header header;
496                         struct ip_event ip;
497                         struct mmap_event mmap;
498                 } event_t;
499
500                 event_t *event = (event_t *)&data[old & md->mask];
501
502                 event_t event_copy;
503
504                 size_t size = event->header.size;
505
506                 /*
507                  * Event straddles the mmap boundary -- header should always
508                  * be inside due to u64 alignment of output.
509                  */
510                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
511                         unsigned int offset = old;
512                         unsigned int len = min(sizeof(*event), size), cpy;
513                         void *dst = &event_copy;
514
515                         do {
516                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
517                                 memcpy(dst, &data[offset & md->mask], cpy);
518                                 offset += cpy;
519                                 dst += cpy;
520                                 len -= cpy;
521                         } while (len);
522
523                         event = &event_copy;
524                 }
525
526                 old += size;
527
528                 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
529                         if (event->header.type & PERF_SAMPLE_IP)
530                                 process_event(event->ip.ip, md->counter);
531                 }
532         }
533
534         md->prev = old;
535 }
536
537 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
538 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
539
540 static int __cmd_top(void)
541 {
542         struct perf_counter_attr attr;
543         pthread_t thread;
544         int i, counter, group_fd, nr_poll = 0;
545         unsigned int cpu;
546         int ret;
547
548         for (i = 0; i < nr_cpus; i++) {
549                 group_fd = -1;
550                 for (counter = 0; counter < nr_counters; counter++) {
551
552                         cpu     = profile_cpu;
553                         if (target_pid == -1 && profile_cpu == -1)
554                                 cpu = i;
555
556                         memset(&attr, 0, sizeof(attr));
557                         attr.config             = event_id[counter];
558                         attr.sample_period      = event_count[counter];
559                         attr.sample_type        = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
560                         attr.freq               = freq;
561
562                         fd[i][counter] = sys_perf_counter_open(&attr, target_pid, cpu, group_fd, 0);
563                         if (fd[i][counter] < 0) {
564                                 int err = errno;
565
566                                 error("syscall returned with %d (%s)\n",
567                                         fd[i][counter], strerror(err));
568                                 if (err == EPERM)
569                                         printf("Are you root?\n");
570                                 exit(-1);
571                         }
572                         assert(fd[i][counter] >= 0);
573                         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
574
575                         /*
576                          * First counter acts as the group leader:
577                          */
578                         if (group && group_fd == -1)
579                                 group_fd = fd[i][counter];
580
581                         event_array[nr_poll].fd = fd[i][counter];
582                         event_array[nr_poll].events = POLLIN;
583                         nr_poll++;
584
585                         mmap_array[i][counter].counter = counter;
586                         mmap_array[i][counter].prev = 0;
587                         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
588                         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
589                                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
590                         if (mmap_array[i][counter].base == MAP_FAILED)
591                                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
592                 }
593         }
594
595         if (pthread_create(&thread, NULL, display_thread, NULL)) {
596                 printf("Could not create display thread.\n");
597                 exit(-1);
598         }
599
600         if (realtime_prio) {
601                 struct sched_param param;
602
603                 param.sched_priority = realtime_prio;
604                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
605                         printf("Could not set realtime priority.\n");
606                         exit(-1);
607                 }
608         }
609
610         while (1) {
611                 int hits = samples;
612
613                 for (i = 0; i < nr_cpus; i++) {
614                         for (counter = 0; counter < nr_counters; counter++)
615                                 mmap_read(&mmap_array[i][counter]);
616                 }
617
618                 if (hits == samples)
619                         ret = poll(event_array, nr_poll, 100);
620         }
621
622         return 0;
623 }
624
625 static const char * const top_usage[] = {
626         "perf top [<options>]",
627         NULL
628 };
629
630 static char events_help_msg[EVENTS_HELP_MAX];
631
632 static const struct option options[] = {
633         OPT_CALLBACK('e', "event", NULL, "event",
634                      events_help_msg, parse_events),
635         OPT_INTEGER('c', "count", &default_interval,
636                     "event period to sample"),
637         OPT_INTEGER('p', "pid", &target_pid,
638                     "profile events on existing pid"),
639         OPT_BOOLEAN('a', "all-cpus", &system_wide,
640                             "system-wide collection from all CPUs"),
641         OPT_INTEGER('C', "CPU", &profile_cpu,
642                     "CPU to profile on"),
643         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
644                     "number of mmap data pages"),
645         OPT_INTEGER('r', "realtime", &realtime_prio,
646                     "collect data with this RT SCHED_FIFO priority"),
647         OPT_INTEGER('d', "delay", &delay_secs,
648                     "number of seconds to delay between refreshes"),
649         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
650                             "dump the symbol table used for profiling"),
651         OPT_INTEGER('f', "count-filter", &count_filter,
652                     "only display functions with more events than this"),
653         OPT_BOOLEAN('g', "group", &group,
654                             "put the counters into a counter group"),
655         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
656                     "only display symbols matchig this pattern"),
657         OPT_BOOLEAN('z', "zero", &group,
658                     "zero history across updates"),
659         OPT_INTEGER('F', "freq", &freq,
660                     "profile at this frequency"),
661         OPT_INTEGER('E', "entries", &print_entries,
662                     "display this many functions"),
663         OPT_END()
664 };
665
666 int cmd_top(int argc, const char **argv, const char *prefix)
667 {
668         int counter;
669
670         page_size = sysconf(_SC_PAGE_SIZE);
671
672         create_events_help(events_help_msg);
673         memcpy(event_id, default_event_id, sizeof(default_event_id));
674
675         argc = parse_options(argc, argv, options, top_usage, 0);
676         if (argc)
677                 usage_with_options(top_usage, options);
678
679         if (freq) {
680                 default_interval = freq;
681                 freq = 1;
682         }
683
684         /* CPU and PID are mutually exclusive */
685         if (target_pid != -1 && profile_cpu != -1) {
686                 printf("WARNING: PID switch overriding CPU\n");
687                 sleep(1);
688                 profile_cpu = -1;
689         }
690
691         if (!nr_counters) {
692                 nr_counters = 1;
693                 event_id[0] = 0;
694         }
695
696         for (counter = 0; counter < nr_counters; counter++) {
697                 if (event_count[counter])
698                         continue;
699
700                 event_count[counter] = default_interval;
701         }
702
703         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
704         assert(nr_cpus <= MAX_NR_CPUS);
705         assert(nr_cpus >= 0);
706
707         if (target_pid != -1 || profile_cpu != -1)
708                 nr_cpus = 1;
709
710         parse_symbols();
711
712         return __cmd_top();
713 }