perf_counter tools: Make symbol loading consistently return number of loaded symbols
[safe/jmp/linux-2.6] / tools / perf / 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 <linux/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 u64                      min_ip;
83 static u64                      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", (u64)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, mid-overhead
252                  * entries in green - and keep the low overhead places
253                  * normal:
254                  */
255                 if (pcnt >= 5.0) {
256                         color = PERF_COLOR_RED;
257                 } else {
258                         if (pcnt >= 0.5)
259                                 color = PERF_COLOR_GREEN;
260                 }
261
262                 if (nr_counters == 1)
263                         printf("%20.2f - ", syme->weight);
264                 else
265                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
266
267                 color_fprintf(stdout, color, "%4.1f%%", pcnt);
268                 printf(" - %016llx : %s\n", sym->start, sym->name);
269         }
270 }
271
272 static void *display_thread(void *arg __used)
273 {
274         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
275         int delay_msecs = delay_secs * 1000;
276
277         printf("PerfTop refresh period: %d seconds\n", delay_secs);
278
279         do {
280                 print_sym_table();
281         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
282
283         printf("key pressed - exiting.\n");
284         exit(0);
285
286         return NULL;
287 }
288
289 /* Tag samples to be skipped. */
290 static const char *skip_symbols[] = {
291         "default_idle",
292         "cpu_idle",
293         "enter_idle",
294         "exit_idle",
295         "mwait_idle",
296         "ppc64_runlatch_off",
297         "pseries_dedicated_idle_sleep",
298         NULL
299 };
300
301 static int symbol_filter(struct dso *self, struct symbol *sym)
302 {
303         static int filter_match;
304         struct sym_entry *syme;
305         const char *name = sym->name;
306         int i;
307
308         /*
309          * ppc64 uses function descriptors and appends a '.' to the
310          * start of every instruction address. Remove it.
311          */
312         if (name[0] == '.')
313                 name++;
314
315         if (!strcmp(name, "_text") ||
316             !strcmp(name, "_etext") ||
317             !strcmp(name, "_sinittext") ||
318             !strncmp("init_module", name, 11) ||
319             !strncmp("cleanup_module", name, 14) ||
320             strstr(name, "_text_start") ||
321             strstr(name, "_text_end"))
322                 return 1;
323
324         syme = dso__sym_priv(self, sym);
325         for (i = 0; skip_symbols[i]; i++) {
326                 if (!strcmp(skip_symbols[i], name)) {
327                         syme->skip = 1;
328                         break;
329                 }
330         }
331
332         if (filter_match == 1) {
333                 filter_end = sym->start;
334                 filter_match = -1;
335                 if (filter_end - filter_start > 10000) {
336                         fprintf(stderr,
337                                 "hm, too large filter symbol <%s> - skipping.\n",
338                                 sym_filter);
339                         fprintf(stderr, "symbol filter start: %016lx\n",
340                                 filter_start);
341                         fprintf(stderr, "                end: %016lx\n",
342                                 filter_end);
343                         filter_end = filter_start = 0;
344                         sym_filter = NULL;
345                         sleep(1);
346                 }
347         }
348
349         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
350                 filter_match = 1;
351                 filter_start = sym->start;
352         }
353
354
355         return 0;
356 }
357
358 static int parse_symbols(void)
359 {
360         struct rb_node *node;
361         struct symbol  *sym;
362
363         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
364         if (kernel_dso == NULL)
365                 return -1;
366
367         if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) <= 0)
368                 goto out_delete_dso;
369
370         node = rb_first(&kernel_dso->syms);
371         sym = rb_entry(node, struct symbol, rb_node);
372         min_ip = sym->start;
373
374         node = rb_last(&kernel_dso->syms);
375         sym = rb_entry(node, struct symbol, rb_node);
376         max_ip = sym->end;
377
378         if (dump_symtab)
379                 dso__fprintf(kernel_dso, stderr);
380
381         return 0;
382
383 out_delete_dso:
384         dso__delete(kernel_dso);
385         kernel_dso = NULL;
386         return -1;
387 }
388
389 #define TRACE_COUNT     3
390
391 /*
392  * Binary search in the histogram table and record the hit:
393  */
394 static void record_ip(u64 ip, int counter)
395 {
396         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
397
398         if (sym != NULL) {
399                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
400
401                 if (!syme->skip) {
402                         syme->count[counter]++;
403                         pthread_mutex_lock(&active_symbols_lock);
404                         if (list_empty(&syme->node) || !syme->node.next)
405                                 __list_insert_active_sym(syme);
406                         pthread_mutex_unlock(&active_symbols_lock);
407                         return;
408                 }
409         }
410
411         samples--;
412 }
413
414 static void process_event(u64 ip, int counter, int user)
415 {
416         samples++;
417
418         if (user) {
419                 userspace_samples++;
420                 return;
421         }
422
423         record_ip(ip, counter);
424 }
425
426 struct mmap_data {
427         int                     counter;
428         void                    *base;
429         int                     mask;
430         unsigned int            prev;
431 };
432
433 static unsigned int mmap_read_head(struct mmap_data *md)
434 {
435         struct perf_counter_mmap_page *pc = md->base;
436         int head;
437
438         head = pc->data_head;
439         rmb();
440
441         return head;
442 }
443
444 struct timeval last_read, this_read;
445
446 static void mmap_read_counter(struct mmap_data *md)
447 {
448         unsigned int head = mmap_read_head(md);
449         unsigned int old = md->prev;
450         unsigned char *data = md->base + page_size;
451         int diff;
452
453         gettimeofday(&this_read, NULL);
454
455         /*
456          * If we're further behind than half the buffer, there's a chance
457          * the writer will bite our tail and mess up the samples under us.
458          *
459          * If we somehow ended up ahead of the head, we got messed up.
460          *
461          * In either case, truncate and restart at head.
462          */
463         diff = head - old;
464         if (diff > md->mask / 2 || diff < 0) {
465                 struct timeval iv;
466                 unsigned long msecs;
467
468                 timersub(&this_read, &last_read, &iv);
469                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
470
471                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
472                                 "  Last read %lu msecs ago.\n", msecs);
473
474                 /*
475                  * head points to a known good entry, start there.
476                  */
477                 old = head;
478         }
479
480         last_read = this_read;
481
482         for (; old != head;) {
483                 struct ip_event {
484                         struct perf_event_header header;
485                         u64 ip;
486                         u32 pid, target_pid;
487                 };
488                 struct mmap_event {
489                         struct perf_event_header header;
490                         u32 pid, target_pid;
491                         u64 start;
492                         u64 len;
493                         u64 pgoff;
494                         char filename[PATH_MAX];
495                 };
496
497                 typedef union event_union {
498                         struct perf_event_header header;
499                         struct ip_event ip;
500                         struct mmap_event mmap;
501                 } event_t;
502
503                 event_t *event = (event_t *)&data[old & md->mask];
504
505                 event_t event_copy;
506
507                 size_t size = event->header.size;
508
509                 /*
510                  * Event straddles the mmap boundary -- header should always
511                  * be inside due to u64 alignment of output.
512                  */
513                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
514                         unsigned int offset = old;
515                         unsigned int len = min(sizeof(*event), size), cpy;
516                         void *dst = &event_copy;
517
518                         do {
519                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
520                                 memcpy(dst, &data[offset & md->mask], cpy);
521                                 offset += cpy;
522                                 dst += cpy;
523                                 len -= cpy;
524                         } while (len);
525
526                         event = &event_copy;
527                 }
528
529                 old += size;
530
531                 if (event->header.type == PERF_EVENT_SAMPLE) {
532                         int user =
533         (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
534                         process_event(event->ip.ip, md->counter, user);
535                 }
536         }
537
538         md->prev = old;
539 }
540
541 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
542 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
543
544 static void mmap_read(void)
545 {
546         int i, counter;
547
548         for (i = 0; i < nr_cpus; i++) {
549                 for (counter = 0; counter < nr_counters; counter++)
550                         mmap_read_counter(&mmap_array[i][counter]);
551         }
552 }
553
554 int nr_poll;
555 int group_fd;
556
557 static void start_counter(int i, int counter)
558 {
559         struct perf_counter_attr *attr;
560         unsigned int cpu;
561
562         cpu = profile_cpu;
563         if (target_pid == -1 && profile_cpu == -1)
564                 cpu = i;
565
566         attr = attrs + counter;
567
568         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
569         attr->freq              = freq;
570
571 try_again:
572         fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
573
574         if (fd[i][counter] < 0) {
575                 int err = errno;
576
577                 if (err == EPERM)
578                         die("No permission - are you root?\n");
579                 /*
580                  * If it's cycles then fall back to hrtimer
581                  * based cpu-clock-tick sw counter, which
582                  * is always available even if no PMU support:
583                  */
584                 if (attr->type == PERF_TYPE_HARDWARE
585                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
586
587                         if (verbose)
588                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
589
590                         attr->type = PERF_TYPE_SOFTWARE;
591                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
592                         goto try_again;
593                 }
594                 printf("\n");
595                 error("perfcounter syscall returned with %d (%s)\n",
596                         fd[i][counter], strerror(err));
597                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
598                 exit(-1);
599         }
600         assert(fd[i][counter] >= 0);
601         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
602
603         /*
604          * First counter acts as the group leader:
605          */
606         if (group && group_fd == -1)
607                 group_fd = fd[i][counter];
608
609         event_array[nr_poll].fd = fd[i][counter];
610         event_array[nr_poll].events = POLLIN;
611         nr_poll++;
612
613         mmap_array[i][counter].counter = counter;
614         mmap_array[i][counter].prev = 0;
615         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
616         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
617                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
618         if (mmap_array[i][counter].base == MAP_FAILED)
619                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
620 }
621
622 static int __cmd_top(void)
623 {
624         pthread_t thread;
625         int i, counter;
626         int ret;
627
628         for (i = 0; i < nr_cpus; i++) {
629                 group_fd = -1;
630                 for (counter = 0; counter < nr_counters; counter++)
631                         start_counter(i, counter);
632         }
633
634         /* Wait for a minimal set of events before starting the snapshot */
635         poll(event_array, nr_poll, 100);
636
637         mmap_read();
638
639         if (pthread_create(&thread, NULL, display_thread, NULL)) {
640                 printf("Could not create display thread.\n");
641                 exit(-1);
642         }
643
644         if (realtime_prio) {
645                 struct sched_param param;
646
647                 param.sched_priority = realtime_prio;
648                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
649                         printf("Could not set realtime priority.\n");
650                         exit(-1);
651                 }
652         }
653
654         while (1) {
655                 int hits = samples;
656
657                 mmap_read();
658
659                 if (hits == samples)
660                         ret = poll(event_array, nr_poll, 100);
661         }
662
663         return 0;
664 }
665
666 static const char * const top_usage[] = {
667         "perf top [<options>]",
668         NULL
669 };
670
671 static const struct option options[] = {
672         OPT_CALLBACK('e', "event", NULL, "event",
673                      "event selector. use 'perf list' to list available events",
674                      parse_events),
675         OPT_INTEGER('c', "count", &default_interval,
676                     "event period to sample"),
677         OPT_INTEGER('p', "pid", &target_pid,
678                     "profile events on existing pid"),
679         OPT_BOOLEAN('a', "all-cpus", &system_wide,
680                             "system-wide collection from all CPUs"),
681         OPT_INTEGER('C', "CPU", &profile_cpu,
682                     "CPU to profile on"),
683         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
684                     "number of mmap data pages"),
685         OPT_INTEGER('r', "realtime", &realtime_prio,
686                     "collect data with this RT SCHED_FIFO priority"),
687         OPT_INTEGER('d', "delay", &delay_secs,
688                     "number of seconds to delay between refreshes"),
689         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
690                             "dump the symbol table used for profiling"),
691         OPT_INTEGER('f', "count-filter", &count_filter,
692                     "only display functions with more events than this"),
693         OPT_BOOLEAN('g', "group", &group,
694                             "put the counters into a counter group"),
695         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
696                     "only display symbols matchig this pattern"),
697         OPT_BOOLEAN('z', "zero", &zero,
698                     "zero history across updates"),
699         OPT_INTEGER('F', "freq", &freq,
700                     "profile at this frequency"),
701         OPT_INTEGER('E', "entries", &print_entries,
702                     "display this many functions"),
703         OPT_BOOLEAN('v', "verbose", &verbose,
704                     "be more verbose (show counter open errors, etc)"),
705         OPT_END()
706 };
707
708 int cmd_top(int argc, const char **argv, const char *prefix __used)
709 {
710         int counter;
711
712         page_size = sysconf(_SC_PAGE_SIZE);
713
714         argc = parse_options(argc, argv, options, top_usage, 0);
715         if (argc)
716                 usage_with_options(top_usage, options);
717
718         if (freq) {
719                 default_interval = freq;
720                 freq = 1;
721         }
722
723         /* CPU and PID are mutually exclusive */
724         if (target_pid != -1 && profile_cpu != -1) {
725                 printf("WARNING: PID switch overriding CPU\n");
726                 sleep(1);
727                 profile_cpu = -1;
728         }
729
730         if (!nr_counters)
731                 nr_counters = 1;
732
733         if (delay_secs < 1)
734                 delay_secs = 1;
735
736         parse_symbols();
737
738         /*
739          * Fill in the ones not specifically initialized via -c:
740          */
741         for (counter = 0; counter < nr_counters; counter++) {
742                 if (attrs[counter].sample_period)
743                         continue;
744
745                 attrs[counter].sample_period = default_interval;
746         }
747
748         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
749         assert(nr_cpus <= MAX_NR_CPUS);
750         assert(nr_cpus >= 0);
751
752         if (target_pid != -1 || profile_cpu != -1)
753                 nr_cpus = 1;
754
755         return __cmd_top();
756 }