perf top: Fall back to cpu-clock-tick hrtimer sampling if no cycle counter available
[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 "util/rbtree.h"
27 #include "util/parse-options.h"
28 #include "util/parse-events.h"
29
30 #include <assert.h>
31 #include <fcntl.h>
32
33 #include <stdio.h>
34
35 #include <errno.h>
36 #include <time.h>
37 #include <sched.h>
38 #include <pthread.h>
39
40 #include <sys/syscall.h>
41 #include <sys/ioctl.h>
42 #include <sys/poll.h>
43 #include <sys/prctl.h>
44 #include <sys/wait.h>
45 #include <sys/uio.h>
46 #include <sys/mman.h>
47
48 #include <linux/unistd.h>
49 #include <linux/types.h>
50
51 static int                      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
69 static char                     *sym_filter;
70 static unsigned long            filter_start;
71 static unsigned long            filter_end;
72
73 static int                      delay_secs                      =  2;
74 static int                      zero;
75 static int                      dump_symtab;
76
77 /*
78  * Symbols
79  */
80
81 static uint64_t                 min_ip;
82 static uint64_t                 max_ip = -1ll;
83
84 struct sym_entry {
85         struct rb_node          rb_node;
86         struct list_head        node;
87         unsigned long           count[MAX_COUNTERS];
88         unsigned long           snap_count;
89         double                  weight;
90         int                     skip;
91 };
92
93 struct sym_entry                *sym_filter_entry;
94
95 struct dso                      *kernel_dso;
96
97 /*
98  * Symbols will be added here in record_ip and will get out
99  * after decayed.
100  */
101 static LIST_HEAD(active_symbols);
102 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
103
104 /*
105  * Ordering weight: count-1 * count-2 * ... / count-n
106  */
107 static double sym_weight(const struct sym_entry *sym)
108 {
109         double weight = sym->snap_count;
110         int counter;
111
112         for (counter = 1; counter < nr_counters-1; counter++)
113                 weight *= sym->count[counter];
114
115         weight /= (sym->count[counter] + 1);
116
117         return weight;
118 }
119
120 static long                     samples;
121 static long                     userspace_samples;
122 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
123
124 static void __list_insert_active_sym(struct sym_entry *syme)
125 {
126         list_add(&syme->node, &active_symbols);
127 }
128
129 static void list_remove_active_sym(struct sym_entry *syme)
130 {
131         pthread_mutex_lock(&active_symbols_lock);
132         list_del_init(&syme->node);
133         pthread_mutex_unlock(&active_symbols_lock);
134 }
135
136 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
137 {
138         struct rb_node **p = &tree->rb_node;
139         struct rb_node *parent = NULL;
140         struct sym_entry *iter;
141
142         while (*p != NULL) {
143                 parent = *p;
144                 iter = rb_entry(parent, struct sym_entry, rb_node);
145
146                 if (se->weight > iter->weight)
147                         p = &(*p)->rb_left;
148                 else
149                         p = &(*p)->rb_right;
150         }
151
152         rb_link_node(&se->rb_node, parent, p);
153         rb_insert_color(&se->rb_node, tree);
154 }
155
156 static void print_sym_table(void)
157 {
158         int printed = 0, j;
159         int counter;
160         float samples_per_sec = samples/delay_secs;
161         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
162         float sum_ksamples = 0.0;
163         struct sym_entry *syme, *n;
164         struct rb_root tmp = RB_ROOT;
165         struct rb_node *nd;
166
167         samples = userspace_samples = 0;
168
169         /* Sort the active symbols */
170         pthread_mutex_lock(&active_symbols_lock);
171         syme = list_entry(active_symbols.next, struct sym_entry, node);
172         pthread_mutex_unlock(&active_symbols_lock);
173
174         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
175                 syme->snap_count = syme->count[0];
176                 if (syme->snap_count != 0) {
177                         syme->weight = sym_weight(syme);
178                         rb_insert_active_sym(&tmp, syme);
179                         sum_ksamples += syme->snap_count;
180
181                         for (j = 0; j < nr_counters; j++)
182                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
183                 } else
184                         list_remove_active_sym(syme);
185         }
186
187         puts(CONSOLE_CLEAR);
188
189         printf(
190 "------------------------------------------------------------------------------\n");
191         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
192                 samples_per_sec,
193                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
194
195         if (nr_counters == 1) {
196                 printf("%Ld", attrs[0].sample_period);
197                 if (freq)
198                         printf("Hz ");
199                 else
200                         printf(" ");
201         }
202
203         for (counter = 0; counter < nr_counters; counter++) {
204                 if (counter)
205                         printf("/");
206
207                 printf("%s", event_name(counter));
208         }
209
210         printf( "], ");
211
212         if (target_pid != -1)
213                 printf(" (target_pid: %d", target_pid);
214         else
215                 printf(" (all");
216
217         if (profile_cpu != -1)
218                 printf(", cpu: %d)\n", profile_cpu);
219         else {
220                 if (target_pid != -1)
221                         printf(")\n");
222                 else
223                         printf(", %d CPUs)\n", nr_cpus);
224         }
225
226         printf("------------------------------------------------------------------------------\n\n");
227
228         if (nr_counters == 1)
229                 printf("             samples    pcnt");
230         else
231                 printf("  weight     samples    pcnt");
232
233         printf("         RIP          kernel function\n"
234                        "  ______     _______   _____   ________________   _______________\n\n"
235         );
236
237         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
238                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
239                 struct symbol *sym = (struct symbol *)(syme + 1);
240                 char *color = PERF_COLOR_NORMAL;
241                 double pcnt;
242
243                 if (++printed > print_entries || syme->snap_count < count_filter)
244                         continue;
245
246                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
247                                          sum_ksamples));
248
249                 /*
250                  * We color high-overhead entries in red, low-overhead
251                  * entries in green - and keep the middle ground normal:
252                  */
253                 if (pcnt >= 5.0)
254                         color = PERF_COLOR_RED;
255                 if (pcnt < 0.5)
256                         color = PERF_COLOR_GREEN;
257
258                 if (nr_counters == 1)
259                         printf("%20.2f - ", syme->weight);
260                 else
261                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
262
263                 color_fprintf(stdout, color, "%4.1f%%", pcnt);
264                 printf(" - %016llx : %s\n", sym->start, sym->name);
265         }
266 }
267
268 static void *display_thread(void *arg)
269 {
270         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
271         int delay_msecs = delay_secs * 1000;
272
273         printf("PerfTop refresh period: %d seconds\n", delay_secs);
274
275         do {
276                 print_sym_table();
277         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
278
279         printf("key pressed - exiting.\n");
280         exit(0);
281
282         return NULL;
283 }
284
285 static int symbol_filter(struct dso *self, struct symbol *sym)
286 {
287         static int filter_match;
288         struct sym_entry *syme;
289         const char *name = sym->name;
290
291         if (!strcmp(name, "_text") ||
292             !strcmp(name, "_etext") ||
293             !strcmp(name, "_sinittext") ||
294             !strncmp("init_module", name, 11) ||
295             !strncmp("cleanup_module", name, 14) ||
296             strstr(name, "_text_start") ||
297             strstr(name, "_text_end"))
298                 return 1;
299
300         syme = dso__sym_priv(self, sym);
301         /* Tag samples to be skipped. */
302         if (!strcmp("default_idle", name) ||
303             !strcmp("cpu_idle", name) ||
304             !strcmp("enter_idle", name) ||
305             !strcmp("exit_idle", name) ||
306             !strcmp("mwait_idle", name))
307                 syme->skip = 1;
308
309         if (filter_match == 1) {
310                 filter_end = sym->start;
311                 filter_match = -1;
312                 if (filter_end - filter_start > 10000) {
313                         fprintf(stderr,
314                                 "hm, too large filter symbol <%s> - skipping.\n",
315                                 sym_filter);
316                         fprintf(stderr, "symbol filter start: %016lx\n",
317                                 filter_start);
318                         fprintf(stderr, "                end: %016lx\n",
319                                 filter_end);
320                         filter_end = filter_start = 0;
321                         sym_filter = NULL;
322                         sleep(1);
323                 }
324         }
325
326         if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
327                 filter_match = 1;
328                 filter_start = sym->start;
329         }
330
331
332         return 0;
333 }
334
335 static int parse_symbols(void)
336 {
337         struct rb_node *node;
338         struct symbol  *sym;
339
340         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
341         if (kernel_dso == NULL)
342                 return -1;
343
344         if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0)
345                 goto out_delete_dso;
346
347         node = rb_first(&kernel_dso->syms);
348         sym = rb_entry(node, struct symbol, rb_node);
349         min_ip = sym->start;
350
351         node = rb_last(&kernel_dso->syms);
352         sym = rb_entry(node, struct symbol, rb_node);
353         max_ip = sym->end;
354
355         if (dump_symtab)
356                 dso__fprintf(kernel_dso, stderr);
357
358         return 0;
359
360 out_delete_dso:
361         dso__delete(kernel_dso);
362         kernel_dso = NULL;
363         return -1;
364 }
365
366 #define TRACE_COUNT     3
367
368 /*
369  * Binary search in the histogram table and record the hit:
370  */
371 static void record_ip(uint64_t ip, int counter)
372 {
373         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
374
375         if (sym != NULL) {
376                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
377
378                 if (!syme->skip) {
379                         syme->count[counter]++;
380                         pthread_mutex_lock(&active_symbols_lock);
381                         if (list_empty(&syme->node) || !syme->node.next)
382                                 __list_insert_active_sym(syme);
383                         pthread_mutex_unlock(&active_symbols_lock);
384                         return;
385                 }
386         }
387
388         samples--;
389 }
390
391 static void process_event(uint64_t ip, int counter)
392 {
393         samples++;
394
395         if (ip < min_ip || ip > max_ip) {
396                 userspace_samples++;
397                 return;
398         }
399
400         record_ip(ip, counter);
401 }
402
403 struct mmap_data {
404         int                     counter;
405         void                    *base;
406         unsigned int            mask;
407         unsigned int            prev;
408 };
409
410 static unsigned int mmap_read_head(struct mmap_data *md)
411 {
412         struct perf_counter_mmap_page *pc = md->base;
413         int head;
414
415         head = pc->data_head;
416         rmb();
417
418         return head;
419 }
420
421 struct timeval last_read, this_read;
422
423 static void mmap_read_counter(struct mmap_data *md)
424 {
425         unsigned int head = mmap_read_head(md);
426         unsigned int old = md->prev;
427         unsigned char *data = md->base + page_size;
428         int diff;
429
430         gettimeofday(&this_read, NULL);
431
432         /*
433          * If we're further behind than half the buffer, there's a chance
434          * the writer will bite our tail and mess up the samples under us.
435          *
436          * If we somehow ended up ahead of the head, we got messed up.
437          *
438          * In either case, truncate and restart at head.
439          */
440         diff = head - old;
441         if (diff > md->mask / 2 || diff < 0) {
442                 struct timeval iv;
443                 unsigned long msecs;
444
445                 timersub(&this_read, &last_read, &iv);
446                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
447
448                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
449                                 "  Last read %lu msecs ago.\n", msecs);
450
451                 /*
452                  * head points to a known good entry, start there.
453                  */
454                 old = head;
455         }
456
457         last_read = this_read;
458
459         for (; old != head;) {
460                 struct ip_event {
461                         struct perf_event_header header;
462                         __u64 ip;
463                         __u32 pid, target_pid;
464                 };
465                 struct mmap_event {
466                         struct perf_event_header header;
467                         __u32 pid, target_pid;
468                         __u64 start;
469                         __u64 len;
470                         __u64 pgoff;
471                         char filename[PATH_MAX];
472                 };
473
474                 typedef union event_union {
475                         struct perf_event_header header;
476                         struct ip_event ip;
477                         struct mmap_event mmap;
478                 } event_t;
479
480                 event_t *event = (event_t *)&data[old & md->mask];
481
482                 event_t event_copy;
483
484                 size_t size = event->header.size;
485
486                 /*
487                  * Event straddles the mmap boundary -- header should always
488                  * be inside due to u64 alignment of output.
489                  */
490                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
491                         unsigned int offset = old;
492                         unsigned int len = min(sizeof(*event), size), cpy;
493                         void *dst = &event_copy;
494
495                         do {
496                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
497                                 memcpy(dst, &data[offset & md->mask], cpy);
498                                 offset += cpy;
499                                 dst += cpy;
500                                 len -= cpy;
501                         } while (len);
502
503                         event = &event_copy;
504                 }
505
506                 old += size;
507
508                 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
509                         if (event->header.type & PERF_SAMPLE_IP)
510                                 process_event(event->ip.ip, md->counter);
511                 }
512         }
513
514         md->prev = old;
515 }
516
517 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
518 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
519
520 static void mmap_read(void)
521 {
522         int i, counter;
523
524         for (i = 0; i < nr_cpus; i++) {
525                 for (counter = 0; counter < nr_counters; counter++)
526                         mmap_read_counter(&mmap_array[i][counter]);
527         }
528 }
529
530 int nr_poll;
531 int group_fd;
532
533 static void start_counter(int i, int counter)
534 {
535         struct perf_counter_attr *attr;
536         unsigned int cpu;
537
538         cpu = profile_cpu;
539         if (target_pid == -1 && profile_cpu == -1)
540                 cpu = i;
541
542         attr = attrs + counter;
543
544         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
545         attr->freq              = freq;
546
547 try_again:
548         fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
549
550         if (fd[i][counter] < 0) {
551                 int err = errno;
552
553                 error("sys_perf_counter_open() syscall returned with %d (%s)\n",
554                         fd[i][counter], strerror(err));
555
556                 if (err == EPERM)
557                         die(" No permission - are you root?\n");
558                 /*
559                  * If it's cycles then fall back to hrtimer
560                  * based cpu-clock-tick sw counter, which
561                  * is always available even if no PMU support:
562                  */
563                 if (attr->type == PERF_TYPE_HARDWARE
564                         && attr->config == PERF_COUNT_CPU_CYCLES) {
565
566                         warning(" ... trying to fall back to cpu-clock-ticks\n");
567                         attr->type = PERF_TYPE_SOFTWARE;
568                         attr->config = PERF_COUNT_CPU_CLOCK;
569                         goto try_again;
570                 }
571                 exit(-1);
572         }
573         assert(fd[i][counter] >= 0);
574         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
575
576         /*
577          * First counter acts as the group leader:
578          */
579         if (group && group_fd == -1)
580                 group_fd = fd[i][counter];
581
582         event_array[nr_poll].fd = fd[i][counter];
583         event_array[nr_poll].events = POLLIN;
584         nr_poll++;
585
586         mmap_array[i][counter].counter = counter;
587         mmap_array[i][counter].prev = 0;
588         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
589         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
590                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
591         if (mmap_array[i][counter].base == MAP_FAILED)
592                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
593 }
594
595 static int __cmd_top(void)
596 {
597         pthread_t thread;
598         int i, counter;
599         int ret;
600
601         for (i = 0; i < nr_cpus; i++) {
602                 group_fd = -1;
603                 for (counter = 0; counter < nr_counters; counter++)
604                         start_counter(i, counter);
605         }
606
607         /* Wait for a minimal set of events before starting the snapshot */
608         poll(event_array, nr_poll, 100);
609
610         mmap_read();
611
612         if (pthread_create(&thread, NULL, display_thread, NULL)) {
613                 printf("Could not create display thread.\n");
614                 exit(-1);
615         }
616
617         if (realtime_prio) {
618                 struct sched_param param;
619
620                 param.sched_priority = realtime_prio;
621                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
622                         printf("Could not set realtime priority.\n");
623                         exit(-1);
624                 }
625         }
626
627         while (1) {
628                 int hits = samples;
629
630                 mmap_read();
631
632                 if (hits == samples)
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 const struct option options[] = {
645         OPT_CALLBACK('e', "event", NULL, "event",
646                      "event selector. use 'perf list' to list available events",
647                      parse_events),
648         OPT_INTEGER('c', "count", &default_interval,
649                     "event period to sample"),
650         OPT_INTEGER('p', "pid", &target_pid,
651                     "profile events on existing pid"),
652         OPT_BOOLEAN('a', "all-cpus", &system_wide,
653                             "system-wide collection from all CPUs"),
654         OPT_INTEGER('C', "CPU", &profile_cpu,
655                     "CPU to profile on"),
656         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
657                     "number of mmap data pages"),
658         OPT_INTEGER('r', "realtime", &realtime_prio,
659                     "collect data with this RT SCHED_FIFO priority"),
660         OPT_INTEGER('d', "delay", &delay_secs,
661                     "number of seconds to delay between refreshes"),
662         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
663                             "dump the symbol table used for profiling"),
664         OPT_INTEGER('f', "count-filter", &count_filter,
665                     "only display functions with more events than this"),
666         OPT_BOOLEAN('g', "group", &group,
667                             "put the counters into a counter group"),
668         OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
669                     "only display symbols matchig this pattern"),
670         OPT_BOOLEAN('z', "zero", &group,
671                     "zero history across updates"),
672         OPT_INTEGER('F', "freq", &freq,
673                     "profile at this frequency"),
674         OPT_INTEGER('E', "entries", &print_entries,
675                     "display this many functions"),
676         OPT_END()
677 };
678
679 int cmd_top(int argc, const char **argv, const char *prefix)
680 {
681         int counter;
682
683         page_size = sysconf(_SC_PAGE_SIZE);
684
685         argc = parse_options(argc, argv, options, top_usage, 0);
686         if (argc)
687                 usage_with_options(top_usage, options);
688
689         if (freq) {
690                 default_interval = freq;
691                 freq = 1;
692         }
693
694         /* CPU and PID are mutually exclusive */
695         if (target_pid != -1 && profile_cpu != -1) {
696                 printf("WARNING: PID switch overriding CPU\n");
697                 sleep(1);
698                 profile_cpu = -1;
699         }
700
701         if (!nr_counters)
702                 nr_counters = 1;
703
704         if (delay_secs < 1)
705                 delay_secs = 1;
706
707         parse_symbols();
708
709         /*
710          * Fill in the ones not specifically initialized via -c:
711          */
712         for (counter = 0; counter < nr_counters; counter++) {
713                 if (attrs[counter].sample_period)
714                         continue;
715
716                 attrs[counter].sample_period = 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         return __cmd_top();
727 }