perf_counter tools: Give perf top inherit option
[linux-2.6.git] / 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                      inherit                         =  0;
62 static int                      profile_cpu                     = -1;
63 static int                      nr_cpus                         =  0;
64 static unsigned int             realtime_prio                   =  0;
65 static int                      group                           =  0;
66 static unsigned int             page_size;
67 static unsigned int             mmap_pages                      = 16;
68 static int                      freq                            =  0;
69 static int                      verbose                         =  0;
70 static char                     *vmlinux                        =  NULL;
71
72 static char                     *sym_filter;
73 static unsigned long            filter_start;
74 static unsigned long            filter_end;
75
76 static int                      delay_secs                      =  2;
77 static int                      zero;
78 static int                      dump_symtab;
79
80 /*
81  * Symbols
82  */
83
84 static u64                      min_ip;
85 static u64                      max_ip = -1ll;
86
87 struct sym_entry {
88         struct rb_node          rb_node;
89         struct list_head        node;
90         unsigned long           count[MAX_COUNTERS];
91         unsigned long           snap_count;
92         double                  weight;
93         int                     skip;
94 };
95
96 struct sym_entry                *sym_filter_entry;
97
98 struct dso                      *kernel_dso;
99
100 /*
101  * Symbols will be added here in record_ip and will get out
102  * after decayed.
103  */
104 static LIST_HEAD(active_symbols);
105 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
106
107 /*
108  * Ordering weight: count-1 * count-2 * ... / count-n
109  */
110 static double sym_weight(const struct sym_entry *sym)
111 {
112         double weight = sym->snap_count;
113         int counter;
114
115         for (counter = 1; counter < nr_counters-1; counter++)
116                 weight *= sym->count[counter];
117
118         weight /= (sym->count[counter] + 1);
119
120         return weight;
121 }
122
123 static long                     samples;
124 static long                     userspace_samples;
125 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
126
127 static void __list_insert_active_sym(struct sym_entry *syme)
128 {
129         list_add(&syme->node, &active_symbols);
130 }
131
132 static void list_remove_active_sym(struct sym_entry *syme)
133 {
134         pthread_mutex_lock(&active_symbols_lock);
135         list_del_init(&syme->node);
136         pthread_mutex_unlock(&active_symbols_lock);
137 }
138
139 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
140 {
141         struct rb_node **p = &tree->rb_node;
142         struct rb_node *parent = NULL;
143         struct sym_entry *iter;
144
145         while (*p != NULL) {
146                 parent = *p;
147                 iter = rb_entry(parent, struct sym_entry, rb_node);
148
149                 if (se->weight > iter->weight)
150                         p = &(*p)->rb_left;
151                 else
152                         p = &(*p)->rb_right;
153         }
154
155         rb_link_node(&se->rb_node, parent, p);
156         rb_insert_color(&se->rb_node, tree);
157 }
158
159 static void print_sym_table(void)
160 {
161         int printed = 0, j;
162         int counter;
163         float samples_per_sec = samples/delay_secs;
164         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
165         float sum_ksamples = 0.0;
166         struct sym_entry *syme, *n;
167         struct rb_root tmp = RB_ROOT;
168         struct rb_node *nd;
169
170         samples = userspace_samples = 0;
171
172         /* Sort the active symbols */
173         pthread_mutex_lock(&active_symbols_lock);
174         syme = list_entry(active_symbols.next, struct sym_entry, node);
175         pthread_mutex_unlock(&active_symbols_lock);
176
177         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
178                 syme->snap_count = syme->count[0];
179                 if (syme->snap_count != 0) {
180                         syme->weight = sym_weight(syme);
181                         rb_insert_active_sym(&tmp, syme);
182                         sum_ksamples += syme->snap_count;
183
184                         for (j = 0; j < nr_counters; j++)
185                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
186                 } else
187                         list_remove_active_sym(syme);
188         }
189
190         puts(CONSOLE_CLEAR);
191
192         printf(
193 "------------------------------------------------------------------------------\n");
194         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
195                 samples_per_sec,
196                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
197
198         if (nr_counters == 1) {
199                 printf("%Ld", (u64)attrs[0].sample_period);
200                 if (freq)
201                         printf("Hz ");
202                 else
203                         printf(" ");
204         }
205
206         for (counter = 0; counter < nr_counters; counter++) {
207                 if (counter)
208                         printf("/");
209
210                 printf("%s", event_name(counter));
211         }
212
213         printf( "], ");
214
215         if (target_pid != -1)
216                 printf(" (target_pid: %d", target_pid);
217         else
218                 printf(" (all");
219
220         if (profile_cpu != -1)
221                 printf(", cpu: %d)\n", profile_cpu);
222         else {
223                 if (target_pid != -1)
224                         printf(")\n");
225                 else
226                         printf(", %d CPUs)\n", nr_cpus);
227         }
228
229         printf("------------------------------------------------------------------------------\n\n");
230
231         if (nr_counters == 1)
232                 printf("             samples    pcnt");
233         else
234                 printf("  weight     samples    pcnt");
235
236         printf("         RIP          kernel function\n"
237                        "  ______     _______   _____   ________________   _______________\n\n"
238         );
239
240         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
241                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
242                 struct symbol *sym = (struct symbol *)(syme + 1);
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                 if (nr_counters == 1)
252                         printf("%20.2f - ", syme->weight);
253                 else
254                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
255
256                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
257                 printf(" - %016llx : %s", sym->start, sym->name);
258                 if (sym->module)
259                         printf("\t[%s]", sym->module->name);
260                 printf("\n");
261         }
262 }
263
264 static void *display_thread(void *arg __used)
265 {
266         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
267         int delay_msecs = delay_secs * 1000;
268
269         printf("PerfTop refresh period: %d seconds\n", delay_secs);
270
271         do {
272                 print_sym_table();
273         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
274
275         printf("key pressed - exiting.\n");
276         exit(0);
277
278         return NULL;
279 }
280
281 /* Tag samples to be skipped. */
282 static const char *skip_symbols[] = {
283         "default_idle",
284         "cpu_idle",
285         "enter_idle",
286         "exit_idle",
287         "mwait_idle",
288         "ppc64_runlatch_off",
289         "pseries_dedicated_idle_sleep",
290         NULL
291 };
292
293 static int symbol_filter(struct dso *self, struct symbol *sym)
294 {
295         static int filter_match;
296         struct sym_entry *syme;
297         const char *name = sym->name;
298         int i;
299
300         /*
301          * ppc64 uses function descriptors and appends a '.' to the
302          * start of every instruction address. Remove it.
303          */
304         if (name[0] == '.')
305                 name++;
306
307         if (!strcmp(name, "_text") ||
308             !strcmp(name, "_etext") ||
309             !strcmp(name, "_sinittext") ||
310             !strncmp("init_module", name, 11) ||
311             !strncmp("cleanup_module", name, 14) ||
312             strstr(name, "_text_start") ||
313             strstr(name, "_text_end"))
314                 return 1;
315
316         syme = dso__sym_priv(self, sym);
317         for (i = 0; skip_symbols[i]; i++) {
318                 if (!strcmp(skip_symbols[i], name)) {
319                         syme->skip = 1;
320                         break;
321                 }
322         }
323
324         if (filter_match == 1) {
325                 filter_end = sym->start;
326                 filter_match = -1;
327                 if (filter_end - filter_start > 10000) {
328                         fprintf(stderr,
329                                 "hm, too large filter symbol <%s> - skipping.\n",
330                                 sym_filter);
331                         fprintf(stderr, "symbol filter start: %016lx\n",
332                                 filter_start);
333                         fprintf(stderr, "                end: %016lx\n",
334                                 filter_end);
335                         filter_end = filter_start = 0;
336                         sym_filter = NULL;
337                         sleep(1);
338                 }
339         }
340
341         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
342                 filter_match = 1;
343                 filter_start = sym->start;
344         }
345
346
347         return 0;
348 }
349
350 static int parse_symbols(void)
351 {
352         struct rb_node *node;
353         struct symbol  *sym;
354         int modules = vmlinux ? 1 : 0;
355
356         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
357         if (kernel_dso == NULL)
358                 return -1;
359
360         if (dso__load_kernel(kernel_dso, vmlinux, symbol_filter, verbose, modules) <= 0)
361                 goto out_delete_dso;
362
363         node = rb_first(&kernel_dso->syms);
364         sym = rb_entry(node, struct symbol, rb_node);
365         min_ip = sym->start;
366
367         node = rb_last(&kernel_dso->syms);
368         sym = rb_entry(node, struct symbol, rb_node);
369         max_ip = sym->end;
370
371         if (dump_symtab)
372                 dso__fprintf(kernel_dso, stderr);
373
374         return 0;
375
376 out_delete_dso:
377         dso__delete(kernel_dso);
378         kernel_dso = NULL;
379         return -1;
380 }
381
382 #define TRACE_COUNT     3
383
384 /*
385  * Binary search in the histogram table and record the hit:
386  */
387 static void record_ip(u64 ip, int counter)
388 {
389         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
390
391         if (sym != NULL) {
392                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
393
394                 if (!syme->skip) {
395                         syme->count[counter]++;
396                         pthread_mutex_lock(&active_symbols_lock);
397                         if (list_empty(&syme->node) || !syme->node.next)
398                                 __list_insert_active_sym(syme);
399                         pthread_mutex_unlock(&active_symbols_lock);
400                         return;
401                 }
402         }
403
404         samples--;
405 }
406
407 static void process_event(u64 ip, int counter, int user)
408 {
409         samples++;
410
411         if (user) {
412                 userspace_samples++;
413                 return;
414         }
415
416         record_ip(ip, counter);
417 }
418
419 struct mmap_data {
420         int                     counter;
421         void                    *base;
422         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_counter(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 mess up the samples 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.type == PERF_EVENT_SAMPLE) {
525                         int user =
526         (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
527                         process_event(event->ip.ip, md->counter, user);
528                 }
529         }
530
531         md->prev = old;
532 }
533
534 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
535 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
536
537 static void mmap_read(void)
538 {
539         int i, counter;
540
541         for (i = 0; i < nr_cpus; i++) {
542                 for (counter = 0; counter < nr_counters; counter++)
543                         mmap_read_counter(&mmap_array[i][counter]);
544         }
545 }
546
547 int nr_poll;
548 int group_fd;
549
550 static void start_counter(int i, int counter)
551 {
552         struct perf_counter_attr *attr;
553         int cpu;
554
555         cpu = profile_cpu;
556         if (target_pid == -1 && profile_cpu == -1)
557                 cpu = i;
558
559         attr = attrs + counter;
560
561         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
562         attr->freq              = freq;
563         attr->inherit           = (cpu < 0) && inherit;
564
565 try_again:
566         fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
567
568         if (fd[i][counter] < 0) {
569                 int err = errno;
570
571                 if (err == EPERM)
572                         die("No permission - are you root?\n");
573                 /*
574                  * If it's cycles then fall back to hrtimer
575                  * based cpu-clock-tick sw counter, which
576                  * is always available even if no PMU support:
577                  */
578                 if (attr->type == PERF_TYPE_HARDWARE
579                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
580
581                         if (verbose)
582                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
583
584                         attr->type = PERF_TYPE_SOFTWARE;
585                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
586                         goto try_again;
587                 }
588                 printf("\n");
589                 error("perfcounter syscall returned with %d (%s)\n",
590                         fd[i][counter], strerror(err));
591                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
592                 exit(-1);
593         }
594         assert(fd[i][counter] >= 0);
595         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
596
597         /*
598          * First counter acts as the group leader:
599          */
600         if (group && group_fd == -1)
601                 group_fd = fd[i][counter];
602
603         event_array[nr_poll].fd = fd[i][counter];
604         event_array[nr_poll].events = POLLIN;
605         nr_poll++;
606
607         mmap_array[i][counter].counter = counter;
608         mmap_array[i][counter].prev = 0;
609         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
610         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
611                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
612         if (mmap_array[i][counter].base == MAP_FAILED)
613                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
614 }
615
616 static int __cmd_top(void)
617 {
618         pthread_t thread;
619         int i, counter;
620         int ret;
621
622         for (i = 0; i < nr_cpus; i++) {
623                 group_fd = -1;
624                 for (counter = 0; counter < nr_counters; counter++)
625                         start_counter(i, counter);
626         }
627
628         /* Wait for a minimal set of events before starting the snapshot */
629         poll(event_array, nr_poll, 100);
630
631         mmap_read();
632
633         if (pthread_create(&thread, NULL, display_thread, NULL)) {
634                 printf("Could not create display thread.\n");
635                 exit(-1);
636         }
637
638         if (realtime_prio) {
639                 struct sched_param param;
640
641                 param.sched_priority = realtime_prio;
642                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
643                         printf("Could not set realtime priority.\n");
644                         exit(-1);
645                 }
646         }
647
648         while (1) {
649                 int hits = samples;
650
651                 mmap_read();
652
653                 if (hits == samples)
654                         ret = poll(event_array, nr_poll, 100);
655         }
656
657         return 0;
658 }
659
660 static const char * const top_usage[] = {
661         "perf top [<options>]",
662         NULL
663 };
664
665 static const struct option options[] = {
666         OPT_CALLBACK('e', "event", NULL, "event",
667                      "event selector. use 'perf list' to list available events",
668                      parse_events),
669         OPT_INTEGER('c', "count", &default_interval,
670                     "event period to sample"),
671         OPT_INTEGER('p', "pid", &target_pid,
672                     "profile events on existing pid"),
673         OPT_BOOLEAN('a', "all-cpus", &system_wide,
674                             "system-wide collection from all CPUs"),
675         OPT_INTEGER('C', "CPU", &profile_cpu,
676                     "CPU to profile on"),
677         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
678         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
679                     "number of mmap data pages"),
680         OPT_INTEGER('r', "realtime", &realtime_prio,
681                     "collect data with this RT SCHED_FIFO priority"),
682         OPT_INTEGER('d', "delay", &delay_secs,
683                     "number of seconds to delay between refreshes"),
684         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
685                             "dump the symbol table used for profiling"),
686         OPT_INTEGER('f', "count-filter", &count_filter,
687                     "only display functions with more events than this"),
688         OPT_BOOLEAN('g', "group", &group,
689                             "put the counters into a counter group"),
690         OPT_BOOLEAN('i', "inherit", &inherit,
691                     "child tasks inherit counters"),
692         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
693                     "only display symbols matchig this pattern"),
694         OPT_BOOLEAN('z', "zero", &zero,
695                     "zero history across updates"),
696         OPT_INTEGER('F', "freq", &freq,
697                     "profile at this frequency"),
698         OPT_INTEGER('E', "entries", &print_entries,
699                     "display this many functions"),
700         OPT_BOOLEAN('v', "verbose", &verbose,
701                     "be more verbose (show counter open errors, etc)"),
702         OPT_END()
703 };
704
705 int cmd_top(int argc, const char **argv, const char *prefix __used)
706 {
707         int counter;
708
709         symbol__init();
710
711         page_size = sysconf(_SC_PAGE_SIZE);
712
713         argc = parse_options(argc, argv, options, top_usage, 0);
714         if (argc)
715                 usage_with_options(top_usage, options);
716
717         if (freq) {
718                 default_interval = freq;
719                 freq = 1;
720         }
721
722         /* CPU and PID are mutually exclusive */
723         if (target_pid != -1 && profile_cpu != -1) {
724                 printf("WARNING: PID switch overriding CPU\n");
725                 sleep(1);
726                 profile_cpu = -1;
727         }
728
729         if (!nr_counters)
730                 nr_counters = 1;
731
732         if (delay_secs < 1)
733                 delay_secs = 1;
734
735         parse_symbols();
736
737         /*
738          * Fill in the ones not specifically initialized via -c:
739          */
740         for (counter = 0; counter < nr_counters; counter++) {
741                 if (attrs[counter].sample_period)
742                         continue;
743
744                 attrs[counter].sample_period = default_interval;
745         }
746
747         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
748         assert(nr_cpus <= MAX_NR_CPUS);
749         assert(nr_cpus >= 0);
750
751         if (target_pid != -1 || profile_cpu != -1)
752                 nr_cpus = 1;
753
754         return __cmd_top();
755 }