perf_counter tools: Convert builtin-top to use libperf symbol routines
[pandora-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         int                     skip;
133 };
134
135 struct sym_entry                *sym_filter_entry;
136
137 struct dso *kernel_dso;
138
139 /*
140  * Symbols will be added here in record_ip and will get out
141  * after decayed.
142  */
143 static LIST_HEAD(active_symbols);
144
145 /*
146  * Ordering weight: count-1 * count-2 * ... / count-n
147  */
148 static double sym_weight(const struct sym_entry *sym)
149 {
150         double weight;
151         int counter;
152
153         weight = sym->count[0];
154
155         for (counter = 1; counter < nr_counters-1; counter++)
156                 weight *= sym->count[counter];
157
158         weight /= (sym->count[counter] + 1);
159
160         return weight;
161 }
162
163 static long                     events;
164 static long                     userspace_events;
165 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
166
167 static void list_insert_active_sym(struct sym_entry *syme)
168 {
169         list_add(&syme->node, &active_symbols);
170 }
171
172 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
173 {
174         struct rb_node **p = &tree->rb_node;
175         struct rb_node *parent = NULL;
176         struct sym_entry *iter;
177
178         while (*p != NULL) {
179                 parent = *p;
180                 iter = rb_entry(parent, struct sym_entry, rb_node);
181
182                 if (sym_weight(se) > sym_weight(iter))
183                         p = &(*p)->rb_left;
184                 else
185                         p = &(*p)->rb_right;
186         }
187
188         rb_link_node(&se->rb_node, parent, p);
189         rb_insert_color(&se->rb_node, tree);
190 }
191
192 static void print_sym_table(void)
193 {
194         int printed, j;
195         int counter;
196         float events_per_sec = events/delay_secs;
197         float kevents_per_sec = (events-userspace_events)/delay_secs;
198         float sum_kevents = 0.0;
199         struct sym_entry *syme, *n;
200         struct rb_root tmp = RB_ROOT;
201         struct rb_node *nd;
202
203         events = userspace_events = 0;
204
205         /* Sort the active symbols */
206         list_for_each_entry_safe(syme, n, &active_symbols, node) {
207                 if (syme->count[0] != 0) {
208                         rb_insert_active_sym(&tmp, syme);
209                         sum_kevents += syme->count[0];
210
211                         for (j = 0; j < nr_counters; j++)
212                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
213                 } else
214                         list_del_init(&syme->node);
215         }
216
217         write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR));
218
219         printf(
220 "------------------------------------------------------------------------------\n");
221         printf( " KernelTop:%8.0f irqs/sec  kernel:%4.1f%% [",
222                 events_per_sec,
223                 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)));
224
225         if (nr_counters == 1)
226                 printf("%d ", event_count[0]);
227
228         for (counter = 0; counter < nr_counters; counter++) {
229                 if (counter)
230                         printf("/");
231
232                 printf("%s", event_name(counter));
233         }
234
235         printf( "], ");
236
237         if (target_pid != -1)
238                 printf(" (target_pid: %d", target_pid);
239         else
240                 printf(" (all");
241
242         if (profile_cpu != -1)
243                 printf(", cpu: %d)\n", profile_cpu);
244         else {
245                 if (target_pid != -1)
246                         printf(")\n");
247                 else
248                         printf(", %d CPUs)\n", nr_cpus);
249         }
250
251         printf("------------------------------------------------------------------------------\n\n");
252
253         if (nr_counters == 1)
254                 printf("             events    pcnt");
255         else
256                 printf("  weight     events    pcnt");
257
258         printf("         RIP          kernel function\n"
259                        "  ______     ______   _____   ________________   _______________\n\n"
260         );
261
262         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
263                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
264                 struct symbol *sym = (struct symbol *)(syme + 1);
265                 float pcnt;
266
267                 if (++printed > 18 || syme->count[0] < count_filter)
268                         break;
269
270                 pcnt = 100.0 - (100.0 * ((sum_kevents - syme->count[0]) /
271                                          sum_kevents));
272
273                 if (nr_counters == 1)
274                         printf("%19.2f - %4.1f%% - %016llx : %s\n",
275                                 sym_weight(syme),
276                                 pcnt, sym->start, sym->name);
277                 else
278                         printf("%8.1f %10ld - %4.1f%% - %016llx : %s\n",
279                                 sym_weight(syme), syme->count[0],
280                                 pcnt, sym->start, sym->name);
281         }
282
283         {
284                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
285
286                 if (poll(&stdin_poll, 1, 0) == 1) {
287                         printf("key pressed - exiting.\n");
288                         exit(0);
289                 }
290         }
291 }
292
293 static void *display_thread(void *arg)
294 {
295         printf("KernelTop refresh period: %d seconds\n", delay_secs);
296
297         while (!sleep(delay_secs))
298                 print_sym_table();
299
300         return NULL;
301 }
302
303 static int symbol_filter(struct dso *self, struct symbol *sym)
304 {
305         static int filter_match;
306         struct sym_entry *syme;
307         const char *name = sym->name;
308
309         if (!strcmp(name, "_text") ||
310             !strcmp(name, "_etext") ||
311             !strcmp(name, "_sinittext") ||
312             !strncmp("init_module", name, 11) ||
313             !strncmp("cleanup_module", name, 14) ||
314             strstr(name, "_text_start") ||
315             strstr(name, "_text_end"))
316                 return 1;
317
318         syme = dso__sym_priv(self, sym);
319         /* Tag events to be skipped. */
320         if (!strcmp("default_idle", name) ||
321             !strcmp("cpu_idle", name) ||
322             !strcmp("enter_idle", name) ||
323             !strcmp("exit_idle", name) ||
324             !strcmp("mwait_idle", name))
325                 syme->skip = 1;
326
327         if (filter_match == 1) {
328                 filter_end = sym->start;
329                 filter_match = -1;
330                 if (filter_end - filter_start > 10000) {
331                         fprintf(stderr,
332                                 "hm, too large filter symbol <%s> - skipping.\n",
333                                 sym_filter);
334                         fprintf(stderr, "symbol filter start: %016lx\n",
335                                 filter_start);
336                         fprintf(stderr, "                end: %016lx\n",
337                                 filter_end);
338                         filter_end = filter_start = 0;
339                         sym_filter = NULL;
340                         sleep(1);
341                 }
342         }
343
344         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
345                 filter_match = 1;
346                 filter_start = sym->start;
347         }
348
349
350         return 0;
351 }
352
353 static int parse_symbols(void)
354 {
355         struct rb_node *node;
356         struct symbol  *sym;
357
358         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
359         if (kernel_dso == NULL)
360                 return -1;
361
362         if (dso__load_kernel(kernel_dso, NULL, symbol_filter) != 0)
363                 goto out_delete_dso;
364
365         node = rb_first(&kernel_dso->syms);
366         sym = rb_entry(node, struct symbol, rb_node);
367         min_ip = sym->start;
368
369         node = rb_last(&kernel_dso->syms);
370         sym = rb_entry(node, struct symbol, rb_node);
371         max_ip = sym->start;
372
373         if (dump_symtab)
374                 dso__fprintf(kernel_dso, stdout);
375
376         return 0;
377
378 out_delete_dso:
379         dso__delete(kernel_dso);
380         kernel_dso = NULL;
381         return -1;
382 }
383
384 #define TRACE_COUNT     3
385
386 /*
387  * Binary search in the histogram table and record the hit:
388  */
389 static void record_ip(uint64_t ip, int counter)
390 {
391         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
392
393         if (sym != NULL) {
394                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
395
396                 if (!syme->skip) {
397                         syme->count[counter]++;
398                         if (list_empty(&syme->node) || !syme->node.next)
399                                 list_insert_active_sym(syme);
400                         return;
401                 }
402         }
403
404         events--;
405 }
406
407 static void process_event(uint64_t ip, int counter)
408 {
409         events++;
410
411         if (ip < min_ip || ip > max_ip) {
412                 userspace_events++;
413                 return;
414         }
415
416         record_ip(ip, counter);
417 }
418
419 struct mmap_data {
420         int counter;
421         void *base;
422         unsigned int mask;
423         unsigned int prev;
424 };
425
426 static unsigned int mmap_read_head(struct mmap_data *md)
427 {
428         struct perf_counter_mmap_page *pc = md->base;
429         int head;
430
431         head = pc->data_head;
432         rmb();
433
434         return head;
435 }
436
437 struct timeval last_read, this_read;
438
439 static void mmap_read(struct mmap_data *md)
440 {
441         unsigned int head = mmap_read_head(md);
442         unsigned int old = md->prev;
443         unsigned char *data = md->base + page_size;
444         int diff;
445
446         gettimeofday(&this_read, NULL);
447
448         /*
449          * If we're further behind than half the buffer, there's a chance
450          * the writer will bite our tail and screw up the events under us.
451          *
452          * If we somehow ended up ahead of the head, we got messed up.
453          *
454          * In either case, truncate and restart at head.
455          */
456         diff = head - old;
457         if (diff > md->mask / 2 || diff < 0) {
458                 struct timeval iv;
459                 unsigned long msecs;
460
461                 timersub(&this_read, &last_read, &iv);
462                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
463
464                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
465                                 "  Last read %lu msecs ago.\n", msecs);
466
467                 /*
468                  * head points to a known good entry, start there.
469                  */
470                 old = head;
471         }
472
473         last_read = this_read;
474
475         for (; old != head;) {
476                 struct ip_event {
477                         struct perf_event_header header;
478                         __u64 ip;
479                         __u32 pid, target_pid;
480                 };
481                 struct mmap_event {
482                         struct perf_event_header header;
483                         __u32 pid, target_pid;
484                         __u64 start;
485                         __u64 len;
486                         __u64 pgoff;
487                         char filename[PATH_MAX];
488                 };
489
490                 typedef union event_union {
491                         struct perf_event_header header;
492                         struct ip_event ip;
493                         struct mmap_event mmap;
494                 } event_t;
495
496                 event_t *event = (event_t *)&data[old & md->mask];
497
498                 event_t event_copy;
499
500                 size_t size = event->header.size;
501
502                 /*
503                  * Event straddles the mmap boundary -- header should always
504                  * be inside due to u64 alignment of output.
505                  */
506                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
507                         unsigned int offset = old;
508                         unsigned int len = min(sizeof(*event), size), cpy;
509                         void *dst = &event_copy;
510
511                         do {
512                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
513                                 memcpy(dst, &data[offset & md->mask], cpy);
514                                 offset += cpy;
515                                 dst += cpy;
516                                 len -= cpy;
517                         } while (len);
518
519                         event = &event_copy;
520                 }
521
522                 old += size;
523
524                 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
525                         if (event->header.type & PERF_RECORD_IP)
526                                 process_event(event->ip.ip, md->counter);
527                 } else {
528                         switch (event->header.type) {
529                                 case PERF_EVENT_MMAP:
530                                 case PERF_EVENT_MUNMAP:
531                                         printf("%s: %Lu %Lu %Lu %s\n",
532                                                         event->header.type == PERF_EVENT_MMAP
533                                                         ? "mmap" : "munmap",
534                                                         event->mmap.start,
535                                                         event->mmap.len,
536                                                         event->mmap.pgoff,
537                                                         event->mmap.filename);
538                                         break;
539                         }
540                 }
541         }
542
543         md->prev = old;
544 }
545
546 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
547 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
548
549 static int __cmd_top(void)
550 {
551         struct perf_counter_hw_event hw_event;
552         pthread_t thread;
553         int i, counter, group_fd, nr_poll = 0;
554         unsigned int cpu;
555         int ret;
556
557         for (i = 0; i < nr_cpus; i++) {
558                 group_fd = -1;
559                 for (counter = 0; counter < nr_counters; counter++) {
560
561                         cpu     = profile_cpu;
562                         if (target_pid == -1 && profile_cpu == -1)
563                                 cpu = i;
564
565                         memset(&hw_event, 0, sizeof(hw_event));
566                         hw_event.config         = event_id[counter];
567                         hw_event.irq_period     = event_count[counter];
568                         hw_event.record_type    = PERF_RECORD_IP | PERF_RECORD_TID;
569                         hw_event.nmi            = 1;
570                         hw_event.mmap           = use_mmap;
571                         hw_event.munmap         = use_munmap;
572                         hw_event.freq           = freq;
573
574                         fd[i][counter] = sys_perf_counter_open(&hw_event, target_pid, cpu, group_fd, 0);
575                         if (fd[i][counter] < 0) {
576                                 int err = errno;
577                                 printf("kerneltop error: syscall returned with %d (%s)\n",
578                                         fd[i][counter], strerror(err));
579                                 if (err == EPERM)
580                                         printf("Are you root?\n");
581                                 exit(-1);
582                         }
583                         assert(fd[i][counter] >= 0);
584                         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
585
586                         /*
587                          * First counter acts as the group leader:
588                          */
589                         if (group && group_fd == -1)
590                                 group_fd = fd[i][counter];
591
592                         event_array[nr_poll].fd = fd[i][counter];
593                         event_array[nr_poll].events = POLLIN;
594                         nr_poll++;
595
596                         mmap_array[i][counter].counter = counter;
597                         mmap_array[i][counter].prev = 0;
598                         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
599                         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
600                                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
601                         if (mmap_array[i][counter].base == MAP_FAILED) {
602                                 printf("kerneltop error: failed to mmap with %d (%s)\n",
603                                                 errno, strerror(errno));
604                                 exit(-1);
605                         }
606                 }
607         }
608
609         if (pthread_create(&thread, NULL, display_thread, NULL)) {
610                 printf("Could not create display thread.\n");
611                 exit(-1);
612         }
613
614         if (realtime_prio) {
615                 struct sched_param param;
616
617                 param.sched_priority = realtime_prio;
618                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
619                         printf("Could not set realtime priority.\n");
620                         exit(-1);
621                 }
622         }
623
624         while (1) {
625                 int hits = events;
626
627                 for (i = 0; i < nr_cpus; i++) {
628                         for (counter = 0; counter < nr_counters; counter++)
629                                 mmap_read(&mmap_array[i][counter]);
630                 }
631
632                 if (hits == events)
633                         ret = poll(event_array, nr_poll, 100);
634         }
635
636         return 0;
637 }
638
639 static const char * const top_usage[] = {
640         "perf top [<options>]",
641         NULL
642 };
643
644 static char events_help_msg[EVENTS_HELP_MAX];
645
646 static const struct option options[] = {
647         OPT_CALLBACK('e', "event", NULL, "event",
648                      events_help_msg, parse_events),
649         OPT_INTEGER('c', "count", &default_interval,
650                     "event period to sample"),
651         OPT_INTEGER('p', "pid", &target_pid,
652                     "profile events on existing pid"),
653         OPT_BOOLEAN('a', "all-cpus", &system_wide,
654                             "system-wide collection from all CPUs"),
655         OPT_INTEGER('C', "CPU", &profile_cpu,
656                     "CPU to profile on"),
657         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
658                     "number of mmap data pages"),
659         OPT_INTEGER('r', "realtime", &realtime_prio,
660                     "collect data with this RT SCHED_FIFO priority"),
661         OPT_INTEGER('d', "delay", &delay_secs,
662                     "number of seconds to delay between refreshes"),
663         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
664                             "dump the symbol table used for profiling"),
665         OPT_INTEGER('f', "--count-filter", &count_filter,
666                     "only display functions with more events than this"),
667         OPT_BOOLEAN('g', "group", &group,
668                             "put the counters into a counter group"),
669         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
670                     "only display symbols matchig this pattern"),
671         OPT_BOOLEAN('z', "zero", &group,
672                     "zero history across updates"),
673         OPT_BOOLEAN('M', "use-mmap", &use_mmap,
674                     "track mmap events"),
675         OPT_BOOLEAN('U', "use-munmap", &use_munmap,
676                     "track munmap events"),
677         OPT_INTEGER('F', "--freq", &freq,
678                     "profile at this frequency"),
679         OPT_END()
680 };
681
682 int cmd_top(int argc, const char **argv, const char *prefix)
683 {
684         int counter;
685
686         page_size = sysconf(_SC_PAGE_SIZE);
687
688         create_events_help(events_help_msg);
689         memcpy(event_id, default_event_id, sizeof(default_event_id));
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                 event_id[0] = 0;
710         }
711
712         for (counter = 0; counter < nr_counters; counter++) {
713                 if (event_count[counter])
714                         continue;
715
716                 event_count[counter] = default_interval;
717         }
718
719         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
720         assert(nr_cpus <= MAX_NR_CPUS);
721         assert(nr_cpus >= 0);
722
723         if (target_pid != -1 || profile_cpu != -1)
724                 nr_cpus = 1;
725
726         parse_symbols();
727
728         return __cmd_top();
729 }