tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ~/hackbench 10
  10    Time: 0.104
  11
  12     Performance counter stats for '/home/mingo/hackbench':
  13
  14        1255.538611  task clock ticks     #      10.143 CPU utilization factor
  15              54011  context switches     #       0.043 M/sec
  16                385  CPU migrations       #       0.000 M/sec
  17              17755  pagefaults           #       0.014 M/sec
  18         3808323185  CPU cycles           #    3033.219 M/sec
  19         1575111190  instructions         #    1254.530 M/sec
  20           17367895  cache references     #      13.833 M/sec
  21            7674421  cache misses         #       6.112 M/sec
  22
  23     Wall-clock time elapsed:   123.786620 msecs
  24
  25  *
  26  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  27  *
  28  * Improvements and fixes by:
  29  *
  30  *   Arjan van de Ven <arjan@linux.intel.com>
  31  *   Yanmin Zhang <yanmin.zhang@intel.com>
  32  *   Wu Fengguang <fengguang.wu@intel.com>
  33  *   Mike Galbraith <efault@gmx.de>
  34  *   Paul Mackerras <paulus@samba.org>
  35  *
  36  * Released under the GPL v2. (and only v2, not any later version)
  37  */
  38
  39 #include "perf.h"
  40 #include "builtin.h"
  41 #include "util/util.h"
  42 #include "util/parse-options.h"
  43 #include "util/parse-events.h"
  44
  45 #include <sys/prctl.h>
  46
  47 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
  48
  49   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  50   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  53
  54   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  55   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  58
  59 };
  60
  61 static int                      system_wide                     =  0;
  62 static int                      inherit                         =  1;
  63 static int                      verbose                         =  0;
  64
  65 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  66
  67 static int                      target_pid                      = -1;
  68 static int                      nr_cpus                         =  0;
  69 static unsigned int             page_size;
  70
  71 static int                      scale                           =  1;
  72
  73 static const unsigned int default_count[] = {
  74         1000000,
  75         1000000,
  76           10000,
  77           10000,
  78         1000000,
  79           10000,
  80 };
  81
  82 static __u64                    event_res[MAX_COUNTERS][3];
  83 static __u64                    event_scaled[MAX_COUNTERS];
  84
  85 static __u64                    runtime_nsecs;
  86 static __u64                    walltime_nsecs;
  87 static __u64                    runtime_cycles;
  88
  89 static void create_perf_stat_counter(int counter)
  90 {
  91         struct perf_counter_attr *attr = attrs + counter;
  92
  93         if (scale)
  94                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
  95                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
  96
  97         if (system_wide) {
  98                 int cpu;
  99                 for (cpu = 0; cpu < nr_cpus; cpu ++) {
 100                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 101                         if (fd[cpu][counter] < 0 && verbose) {
 102                                 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno));
 103                         }
 104                 }
 105         } else {
 106                 attr->inherit   = inherit;
 107                 attr->disabled  = 1;
 108
 109                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 110                 if (fd[0][counter] < 0 && verbose) {
 111                         printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno));
 112                 }
 113         }
 114 }
 115
 116 /*
 117  * Does the counter have nsecs as a unit?
 118  */
 119 static inline int nsec_counter(int counter)
 120 {
 121         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 122                 return 0;
 123
 124         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 125                 return 1;
 126
 127         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 128                 return 1;
 129
 130         return 0;
 131 }
 132
 133 /*
 134  * Read out the results of a single counter:
 135  */
 136 static void read_counter(int counter)
 137 {
 138         __u64 *count, single_count[3];
 139         ssize_t res;
 140         int cpu, nv;
 141         int scaled;
 142
 143         count = event_res[counter];
 144
 145         count[0] = count[1] = count[2] = 0;
 146
 147         nv = scale ? 3 : 1;
 148         for (cpu = 0; cpu < nr_cpus; cpu ++) {
 149                 if (fd[cpu][counter] < 0)
 150                         continue;
 151
 152                 res = read(fd[cpu][counter], single_count, nv * sizeof(__u64));
 153                 assert(res == nv * sizeof(__u64));
 154
 155                 count[0] += single_count[0];
 156                 if (scale) {
 157                         count[1] += single_count[1];
 158                         count[2] += single_count[2];
 159                 }
 160         }
 161
 162         scaled = 0;
 163         if (scale) {
 164                 if (count[2] == 0) {
 165                         event_scaled[counter] = -1;
 166                         count[0] = 0;
 167                         return;
 168                 }
 169
 170                 if (count[2] < count[1]) {
 171                         event_scaled[counter] = 1;
 172                         count[0] = (unsigned long long)
 173                                 ((double)count[0] * count[1] / count[2] + 0.5);
 174                 }
 175         }
 176         /*
 177          * Save the full runtime - to allow normalization during printout:
 178          */
 179         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 180                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 181                 runtime_nsecs = count[0];
 182         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 183                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 184                 runtime_cycles = count[0];
 185 }
 186
 187 /*
 188  * Print out the results of a single counter:
 189  */
 190 static void print_counter(int counter)
 191 {
 192         __u64 *count;
 193         int scaled;
 194
 195         count = event_res[counter];
 196         scaled = event_scaled[counter];
 197
 198         if (scaled == -1) {
 199                 fprintf(stderr, " %14s  %-20s\n",
 200                         "<not counted>", event_name(counter));
 201                 return;
 202         }
 203
 204         if (nsec_counter(counter)) {
 205                 double msecs = (double)count[0] / 1000000;
 206
 207                 fprintf(stderr, " %14.6f  %-20s",
 208                         msecs, event_name(counter));
 209                 if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 210                         attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 211
 212                         if (walltime_nsecs)
 213                                 fprintf(stderr, " # %11.3f CPU utilization factor",
 214                                         (double)count[0] / (double)walltime_nsecs);
 215                 }
 216         } else {
 217                 fprintf(stderr, " %14Ld  %-20s",
 218                         count[0], event_name(counter));
 219                 if (runtime_nsecs)
 220                         fprintf(stderr, " # %11.3f M/sec",
 221                                 (double)count[0]/runtime_nsecs*1000.0);
 222                 if (runtime_cycles &&
 223                         attrs[counter].type == PERF_TYPE_HARDWARE &&
 224                                 attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 225
 226                         fprintf(stderr, " # %1.3f per cycle",
 227                                 (double)count[0] / (double)runtime_cycles);
 228                 }
 229         }
 230         if (scaled)
 231                 fprintf(stderr, "  (scaled from %.2f%%)",
 232                         (double) count[2] / count[1] * 100);
 233         fprintf(stderr, "\n");
 234 }
 235
 236 static int do_perf_stat(int argc, const char **argv)
 237 {
 238         unsigned long long t0, t1;
 239         int counter;
 240         int status;
 241         int pid;
 242         int i;
 243
 244         if (!system_wide)
 245                 nr_cpus = 1;
 246
 247         for (counter = 0; counter < nr_counters; counter++)
 248                 create_perf_stat_counter(counter);
 249
 250         /*
 251          * Enable counters and exec the command:
 252          */
 253         t0 = rdclock();
 254         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 255
 256         if ((pid = fork()) < 0)
 257                 perror("failed to fork");
 258
 259         if (!pid) {
 260                 if (execvp(argv[0], (char **)argv)) {
 261                         perror(argv[0]);
 262                         exit(-1);
 263                 }
 264         }
 265
 266         while (wait(&status) >= 0)
 267                 ;
 268
 269         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 270         t1 = rdclock();
 271
 272         walltime_nsecs = t1 - t0;
 273
 274         fflush(stdout);
 275
 276         fprintf(stderr, "\n");
 277         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 278
 279         for (i = 1; i < argc; i++)
 280                 fprintf(stderr, " %s", argv[i]);
 281
 282         fprintf(stderr, "\':\n");
 283         fprintf(stderr, "\n");
 284
 285         for (counter = 0; counter < nr_counters; counter++)
 286                 read_counter(counter);
 287
 288         for (counter = 0; counter < nr_counters; counter++)
 289                 print_counter(counter);
 290
 291
 292         fprintf(stderr, "\n");
 293         fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
 294                         (double)(t1-t0)/1e6);
 295         fprintf(stderr, "\n");
 296
 297         return 0;
 298 }
 299
 300 static volatile int signr = -1;
 301
 302 static void skip_signal(int signo)
 303 {
 304         signr = signo;
 305 }
 306
 307 static void sig_atexit(void)
 308 {
 309         if (signr == -1)
 310                 return;
 311
 312         signal(signr, SIG_DFL);
 313         kill(getpid(), signr);
 314 }
 315
 316 static const char * const stat_usage[] = {
 317         "perf stat [<options>] <command>",
 318         NULL
 319 };
 320
 321 static const struct option options[] = {
 322         OPT_CALLBACK('e', "event", NULL, "event",
 323                      "event selector. use 'perf list' to list available events",
 324                      parse_events),
 325         OPT_BOOLEAN('i', "inherit", &inherit,
 326                     "child tasks inherit counters"),
 327         OPT_INTEGER('p', "pid", &target_pid,
 328                     "stat events on existing pid"),
 329         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 330                             "system-wide collection from all CPUs"),
 331         OPT_BOOLEAN('S', "scale", &scale,
 332                             "scale/normalize counters"),
 333         OPT_BOOLEAN('v', "verbose", &verbose,
 334                     "be more verbose (show counter open errors, etc)"),
 335         OPT_END()
 336 };
 337
 338 int cmd_stat(int argc, const char **argv, const char *prefix)
 339 {
 340         page_size = sysconf(_SC_PAGE_SIZE);
 341
 342         memcpy(attrs, default_attrs, sizeof(attrs));
 343
 344         argc = parse_options(argc, argv, options, stat_usage, 0);
 345         if (!argc)
 346                 usage_with_options(stat_usage, options);
 347
 348         if (!nr_counters)
 349                 nr_counters = 8;
 350
 351         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 352         assert(nr_cpus <= MAX_NR_CPUS);
 353         assert(nr_cpus >= 0);
 354
 355         /*
 356          * We dont want to block the signals - that would cause
 357          * child tasks to inherit that and Ctrl-C would not work.
 358          * What we want is for Ctrl-C to work in the exec()-ed
 359          * task, but being ignored by perf stat itself:
 360          */
 361         atexit(sig_atexit);
 362         signal(SIGINT,  skip_signal);
 363         signal(SIGALRM, skip_signal);
 364         signal(SIGABRT, skip_signal);
 365
 366         return do_perf_stat(argc, argv);
 367 }