perf_counter tools: Enable kernel module symbol loading in tools
[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 static char                     *vmlinux                        =  NULL;
70
71 static char                     *sym_filter;
72 static unsigned long            filter_start;
73 static unsigned long            filter_end;
74
75 static int                      delay_secs                      =  2;
76 static int                      zero;
77 static int                      dump_symtab;
78
79 /*
80  * Symbols
81  */
82
83 static u64                      min_ip;
84 static u64                      max_ip = -1ll;
85
86 struct sym_entry {
87         struct rb_node          rb_node;
88         struct list_head        node;
89         unsigned long           count[MAX_COUNTERS];
90         unsigned long           snap_count;
91         double                  weight;
92         int                     skip;
93 };
94
95 struct sym_entry                *sym_filter_entry;
96
97 struct dso                      *kernel_dso;
98
99 /*
100  * Symbols will be added here in record_ip and will get out
101  * after decayed.
102  */
103 static LIST_HEAD(active_symbols);
104 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
105
106 /*
107  * Ordering weight: count-1 * count-2 * ... / count-n
108  */
109 static double sym_weight(const struct sym_entry *sym)
110 {
111         double weight = sym->snap_count;
112         int counter;
113
114         for (counter = 1; counter < nr_counters-1; counter++)
115                 weight *= sym->count[counter];
116
117         weight /= (sym->count[counter] + 1);
118
119         return weight;
120 }
121
122 static long                     samples;
123 static long                     userspace_samples;
124 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
125
126 static void __list_insert_active_sym(struct sym_entry *syme)
127 {
128         list_add(&syme->node, &active_symbols);
129 }
130
131 static void list_remove_active_sym(struct sym_entry *syme)
132 {
133         pthread_mutex_lock(&active_symbols_lock);
134         list_del_init(&syme->node);
135         pthread_mutex_unlock(&active_symbols_lock);
136 }
137
138 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
139 {
140         struct rb_node **p = &tree->rb_node;
141         struct rb_node *parent = NULL;
142         struct sym_entry *iter;
143
144         while (*p != NULL) {
145                 parent = *p;
146                 iter = rb_entry(parent, struct sym_entry, rb_node);
147
148                 if (se->weight > iter->weight)
149                         p = &(*p)->rb_left;
150                 else
151                         p = &(*p)->rb_right;
152         }
153
154         rb_link_node(&se->rb_node, parent, p);
155         rb_insert_color(&se->rb_node, tree);
156 }
157
158 static void print_sym_table(void)
159 {
160         int printed = 0, j;
161         int counter;
162         float samples_per_sec = samples/delay_secs;
163         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
164         float sum_ksamples = 0.0;
165         struct sym_entry *syme, *n;
166         struct rb_root tmp = RB_ROOT;
167         struct rb_node *nd;
168
169         samples = userspace_samples = 0;
170
171         /* Sort the active symbols */
172         pthread_mutex_lock(&active_symbols_lock);
173         syme = list_entry(active_symbols.next, struct sym_entry, node);
174         pthread_mutex_unlock(&active_symbols_lock);
175
176         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
177                 syme->snap_count = syme->count[0];
178                 if (syme->snap_count != 0) {
179                         syme->weight = sym_weight(syme);
180                         rb_insert_active_sym(&tmp, syme);
181                         sum_ksamples += syme->snap_count;
182
183                         for (j = 0; j < nr_counters; j++)
184                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
185                 } else
186                         list_remove_active_sym(syme);
187         }
188
189         puts(CONSOLE_CLEAR);
190
191         printf(
192 "------------------------------------------------------------------------------\n");
193         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
194                 samples_per_sec,
195                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
196
197         if (nr_counters == 1) {
198                 printf("%Ld", (u64)attrs[0].sample_period);
199                 if (freq)
200                         printf("Hz ");
201                 else
202                         printf(" ");
203         }
204
205         for (counter = 0; counter < nr_counters; counter++) {
206                 if (counter)
207                         printf("/");
208
209                 printf("%s", event_name(counter));
210         }
211
212         printf( "], ");
213
214         if (target_pid != -1)
215                 printf(" (target_pid: %d", target_pid);
216         else
217                 printf(" (all");
218
219         if (profile_cpu != -1)
220                 printf(", cpu: %d)\n", profile_cpu);
221         else {
222                 if (target_pid != -1)
223                         printf(")\n");
224                 else
225                         printf(", %d CPUs)\n", nr_cpus);
226         }
227
228         printf("------------------------------------------------------------------------------\n\n");
229
230         if (nr_counters == 1)
231                 printf("             samples    pcnt");
232         else
233                 printf("  weight     samples    pcnt");
234
235         printf("         RIP          kernel function\n"
236                        "  ______     _______   _____   ________________   _______________\n\n"
237         );
238
239         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
240                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
241                 struct symbol *sym = (struct symbol *)(syme + 1);
242                 char *color = PERF_COLOR_NORMAL;
243                 double pcnt;
244
245                 if (++printed > print_entries || syme->snap_count < count_filter)
246                         continue;
247
248                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
249                                          sum_ksamples));
250
251                 /*
252                  * We color high-overhead entries in red, mid-overhead
253                  * entries in green - and keep the low overhead places
254                  * normal:
255                  */
256                 if (pcnt >= 5.0) {
257                         color = PERF_COLOR_RED;
258                 } else {
259                         if (pcnt >= 0.5)
260                                 color = PERF_COLOR_GREEN;
261                 }
262
263                 if (nr_counters == 1)
264                         printf("%20.2f - ", syme->weight);
265                 else
266                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
267
268                 color_fprintf(stdout, color, "%4.1f%%", pcnt);
269                 printf(" - %016llx : %s", sym->start, sym->name);
270                 if (sym->module)
271                         printf("\t[%s]", sym->module->name);
272                 printf("\n");
273         }
274 }
275
276 static void *display_thread(void *arg __used)
277 {
278         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
279         int delay_msecs = delay_secs * 1000;
280
281         printf("PerfTop refresh period: %d seconds\n", delay_secs);
282
283         do {
284                 print_sym_table();
285         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
286
287         printf("key pressed - exiting.\n");
288         exit(0);
289
290         return NULL;
291 }
292
293 /* Tag samples to be skipped. */
294 static const char *skip_symbols[] = {
295         "default_idle",
296         "cpu_idle",
297         "enter_idle",
298         "exit_idle",
299         "mwait_idle",
300         "ppc64_runlatch_off",
301         "pseries_dedicated_idle_sleep",
302         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         int i;
311
312         /*
313          * ppc64 uses function descriptors and appends a '.' to the
314          * start of every instruction address. Remove it.
315          */
316         if (name[0] == '.')
317                 name++;
318
319         if (!strcmp(name, "_text") ||
320             !strcmp(name, "_etext") ||
321             !strcmp(name, "_sinittext") ||
322             !strncmp("init_module", name, 11) ||
323             !strncmp("cleanup_module", name, 14) ||
324             strstr(name, "_text_start") ||
325             strstr(name, "_text_end"))
326                 return 1;
327
328         syme = dso__sym_priv(self, sym);
329         for (i = 0; skip_symbols[i]; i++) {
330                 if (!strcmp(skip_symbols[i], name)) {
331                         syme->skip = 1;
332                         break;
333                 }
334         }
335
336         if (filter_match == 1) {
337                 filter_end = sym->start;
338                 filter_match = -1;
339                 if (filter_end - filter_start > 10000) {
340                         fprintf(stderr,
341                                 "hm, too large filter symbol <%s> - skipping.\n",
342                                 sym_filter);
343                         fprintf(stderr, "symbol filter start: %016lx\n",
344                                 filter_start);
345                         fprintf(stderr, "                end: %016lx\n",
346                                 filter_end);
347                         filter_end = filter_start = 0;
348                         sym_filter = NULL;
349                         sleep(1);
350                 }
351         }
352
353         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
354                 filter_match = 1;
355                 filter_start = sym->start;
356         }
357
358
359         return 0;
360 }
361
362 static int parse_symbols(void)
363 {
364         struct rb_node *node;
365         struct symbol  *sym;
366         int modules = vmlinux ? 1 : 0;
367
368         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
369         if (kernel_dso == NULL)
370                 return -1;
371
372         if (dso__load_kernel(kernel_dso, vmlinux, symbol_filter, verbose, modules) <= 0)
373                 goto out_delete_dso;
374
375         node = rb_first(&kernel_dso->syms);
376         sym = rb_entry(node, struct symbol, rb_node);
377         min_ip = sym->start;
378
379         node = rb_last(&kernel_dso->syms);
380         sym = rb_entry(node, struct symbol, rb_node);
381         max_ip = sym->end;
382
383         if (dump_symtab)
384                 dso__fprintf(kernel_dso, stderr);
385
386         return 0;
387
388 out_delete_dso:
389         dso__delete(kernel_dso);
390         kernel_dso = NULL;
391         return -1;
392 }
393
394 #define TRACE_COUNT     3
395
396 /*
397  * Binary search in the histogram table and record the hit:
398  */
399 static void record_ip(u64 ip, int counter)
400 {
401         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
402
403         if (sym != NULL) {
404                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
405
406                 if (!syme->skip) {
407                         syme->count[counter]++;
408                         pthread_mutex_lock(&active_symbols_lock);
409                         if (list_empty(&syme->node) || !syme->node.next)
410                                 __list_insert_active_sym(syme);
411                         pthread_mutex_unlock(&active_symbols_lock);
412                         return;
413                 }
414         }
415
416         samples--;
417 }
418
419 static void process_event(u64 ip, int counter, int user)
420 {
421         samples++;
422
423         if (user) {
424                 userspace_samples++;
425                 return;
426         }
427
428         record_ip(ip, counter);
429 }
430
431 struct mmap_data {
432         int                     counter;
433         void                    *base;
434         int                     mask;
435         unsigned int            prev;
436 };
437
438 static unsigned int mmap_read_head(struct mmap_data *md)
439 {
440         struct perf_counter_mmap_page *pc = md->base;
441         int head;
442
443         head = pc->data_head;
444         rmb();
445
446         return head;
447 }
448
449 struct timeval last_read, this_read;
450
451 static void mmap_read_counter(struct mmap_data *md)
452 {
453         unsigned int head = mmap_read_head(md);
454         unsigned int old = md->prev;
455         unsigned char *data = md->base + page_size;
456         int diff;
457
458         gettimeofday(&this_read, NULL);
459
460         /*
461          * If we're further behind than half the buffer, there's a chance
462          * the writer will bite our tail and mess up the samples under us.
463          *
464          * If we somehow ended up ahead of the head, we got messed up.
465          *
466          * In either case, truncate and restart at head.
467          */
468         diff = head - old;
469         if (diff > md->mask / 2 || diff < 0) {
470                 struct timeval iv;
471                 unsigned long msecs;
472
473                 timersub(&this_read, &last_read, &iv);
474                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
475
476                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
477                                 "  Last read %lu msecs ago.\n", msecs);
478
479                 /*
480                  * head points to a known good entry, start there.
481                  */
482                 old = head;
483         }
484
485         last_read = this_read;
486
487         for (; old != head;) {
488                 struct ip_event {
489                         struct perf_event_header header;
490                         u64 ip;
491                         u32 pid, target_pid;
492                 };
493                 struct mmap_event {
494                         struct perf_event_header header;
495                         u32 pid, target_pid;
496                         u64 start;
497                         u64 len;
498                         u64 pgoff;
499                         char filename[PATH_MAX];
500                 };
501
502                 typedef union event_union {
503                         struct perf_event_header header;
504                         struct ip_event ip;
505                         struct mmap_event mmap;
506                 } event_t;
507
508                 event_t *event = (event_t *)&data[old & md->mask];
509
510                 event_t event_copy;
511
512                 size_t size = event->header.size;
513
514                 /*
515                  * Event straddles the mmap boundary -- header should always
516                  * be inside due to u64 alignment of output.
517                  */
518                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
519                         unsigned int offset = old;
520                         unsigned int len = min(sizeof(*event), size), cpy;
521                         void *dst = &event_copy;
522
523                         do {
524                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
525                                 memcpy(dst, &data[offset & md->mask], cpy);
526                                 offset += cpy;
527                                 dst += cpy;
528                                 len -= cpy;
529                         } while (len);
530
531                         event = &event_copy;
532                 }
533
534                 old += size;
535
536                 if (event->header.type == PERF_EVENT_SAMPLE) {
537                         int user =
538         (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
539                         process_event(event->ip.ip, md->counter, user);
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 void mmap_read(void)
550 {
551         int i, counter;
552
553         for (i = 0; i < nr_cpus; i++) {
554                 for (counter = 0; counter < nr_counters; counter++)
555                         mmap_read_counter(&mmap_array[i][counter]);
556         }
557 }
558
559 int nr_poll;
560 int group_fd;
561
562 static void start_counter(int i, int counter)
563 {
564         struct perf_counter_attr *attr;
565         unsigned int cpu;
566
567         cpu = profile_cpu;
568         if (target_pid == -1 && profile_cpu == -1)
569                 cpu = i;
570
571         attr = attrs + counter;
572
573         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
574         attr->freq              = freq;
575
576 try_again:
577         fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
578
579         if (fd[i][counter] < 0) {
580                 int err = errno;
581
582                 if (err == EPERM)
583                         die("No permission - are you root?\n");
584                 /*
585                  * If it's cycles then fall back to hrtimer
586                  * based cpu-clock-tick sw counter, which
587                  * is always available even if no PMU support:
588                  */
589                 if (attr->type == PERF_TYPE_HARDWARE
590                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
591
592                         if (verbose)
593                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
594
595                         attr->type = PERF_TYPE_SOFTWARE;
596                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
597                         goto try_again;
598                 }
599                 printf("\n");
600                 error("perfcounter syscall returned with %d (%s)\n",
601                         fd[i][counter], strerror(err));
602                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
603                 exit(-1);
604         }
605         assert(fd[i][counter] >= 0);
606         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
607
608         /*
609          * First counter acts as the group leader:
610          */
611         if (group && group_fd == -1)
612                 group_fd = fd[i][counter];
613
614         event_array[nr_poll].fd = fd[i][counter];
615         event_array[nr_poll].events = POLLIN;
616         nr_poll++;
617
618         mmap_array[i][counter].counter = counter;
619         mmap_array[i][counter].prev = 0;
620         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
621         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
622                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
623         if (mmap_array[i][counter].base == MAP_FAILED)
624                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
625 }
626
627 static int __cmd_top(void)
628 {
629         pthread_t thread;
630         int i, counter;
631         int ret;
632
633         for (i = 0; i < nr_cpus; i++) {
634                 group_fd = -1;
635                 for (counter = 0; counter < nr_counters; counter++)
636                         start_counter(i, counter);
637         }
638
639         /* Wait for a minimal set of events before starting the snapshot */
640         poll(event_array, nr_poll, 100);
641
642         mmap_read();
643
644         if (pthread_create(&thread, NULL, display_thread, NULL)) {
645                 printf("Could not create display thread.\n");
646                 exit(-1);
647         }
648
649         if (realtime_prio) {
650                 struct sched_param param;
651
652                 param.sched_priority = realtime_prio;
653                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
654                         printf("Could not set realtime priority.\n");
655                         exit(-1);
656                 }
657         }
658
659         while (1) {
660                 int hits = samples;
661
662                 mmap_read();
663
664                 if (hits == samples)
665                         ret = poll(event_array, nr_poll, 100);
666         }
667
668         return 0;
669 }
670
671 static const char * const top_usage[] = {
672         "perf top [<options>]",
673         NULL
674 };
675
676 static const struct option options[] = {
677         OPT_CALLBACK('e', "event", NULL, "event",
678                      "event selector. use 'perf list' to list available events",
679                      parse_events),
680         OPT_INTEGER('c', "count", &default_interval,
681                     "event period to sample"),
682         OPT_INTEGER('p', "pid", &target_pid,
683                     "profile events on existing pid"),
684         OPT_BOOLEAN('a', "all-cpus", &system_wide,
685                             "system-wide collection from all CPUs"),
686         OPT_INTEGER('C', "CPU", &profile_cpu,
687                     "CPU to profile on"),
688         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
689         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
690                     "number of mmap data pages"),
691         OPT_INTEGER('r', "realtime", &realtime_prio,
692                     "collect data with this RT SCHED_FIFO priority"),
693         OPT_INTEGER('d', "delay", &delay_secs,
694                     "number of seconds to delay between refreshes"),
695         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
696                             "dump the symbol table used for profiling"),
697         OPT_INTEGER('f', "count-filter", &count_filter,
698                     "only display functions with more events than this"),
699         OPT_BOOLEAN('g', "group", &group,
700                             "put the counters into a counter group"),
701         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
702                     "only display symbols matchig this pattern"),
703         OPT_BOOLEAN('z', "zero", &zero,
704                     "zero history across updates"),
705         OPT_INTEGER('F', "freq", &freq,
706                     "profile at this frequency"),
707         OPT_INTEGER('E', "entries", &print_entries,
708                     "display this many functions"),
709         OPT_BOOLEAN('v', "verbose", &verbose,
710                     "be more verbose (show counter open errors, etc)"),
711         OPT_END()
712 };
713
714 int cmd_top(int argc, const char **argv, const char *prefix __used)
715 {
716         int counter;
717
718         symbol__init();
719
720         page_size = sysconf(_SC_PAGE_SIZE);
721
722         argc = parse_options(argc, argv, options, top_usage, 0);
723         if (argc)
724                 usage_with_options(top_usage, options);
725
726         if (freq) {
727                 default_interval = freq;
728                 freq = 1;
729         }
730
731         /* CPU and PID are mutually exclusive */
732         if (target_pid != -1 && profile_cpu != -1) {
733                 printf("WARNING: PID switch overriding CPU\n");
734                 sleep(1);
735                 profile_cpu = -1;
736         }
737
738         if (!nr_counters)
739                 nr_counters = 1;
740
741         if (delay_secs < 1)
742                 delay_secs = 1;
743
744         parse_symbols();
745
746         /*
747          * Fill in the ones not specifically initialized via -c:
748          */
749         for (counter = 0; counter < nr_counters; counter++) {
750                 if (attrs[counter].sample_period)
751                         continue;
752
753                 attrs[counter].sample_period = default_interval;
754         }
755
756         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
757         assert(nr_cpus <= MAX_NR_CPUS);
758         assert(nr_cpus >= 0);
759
760         if (target_pid != -1 || profile_cpu != -1)
761                 nr_cpus = 1;
762
763         return __cmd_top();
764 }