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