SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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_TASK_CLOCK         },
  50   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_CONTEXT_SWITCHES   },
  51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_CPU_MIGRATIONS     },
  52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_PAGE_FAULTS        },
  53
  54   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CPU_CYCLES         },
  55   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_INSTRUCTIONS       },
  56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CACHE_REFERENCES   },
  57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CACHE_MISSES       },
  58 };
  59
  60 static int                      system_wide                     =  0;
  61 static int                      inherit                         =  1;
  62 static int                      verbose                         =  0;
  63
  64 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  65
  66 static int                      target_pid                      = -1;
  67 static int                      nr_cpus                         =  0;
  68 static unsigned int             page_size;
  69
  70 static int                      scale                           =  1;
  71
  72 static const unsigned int default_count[] = {
  73         1000000,
  74         1000000,
  75           10000,
  76           10000,
  77         1000000,
  78           10000,
  79 };
  80
  81 static __u64                    event_res[MAX_COUNTERS][3];
  82 static __u64                    event_scaled[MAX_COUNTERS];
  83
  84 static __u64                    runtime_nsecs;
  85 static __u64                    walltime_nsecs;
  86 static __u64                    runtime_cycles;
  87
  88 static void create_perf_stat_counter(int counter)
  89 {
  90         struct perf_counter_attr *attr = attrs + counter;
  91
  92         if (scale)
  93                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
  94                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
  95
  96         if (system_wide) {
  97                 int cpu;
  98                 for (cpu = 0; cpu < nr_cpus; cpu ++) {
  99                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 100                         if (fd[cpu][counter] < 0 && verbose) {
 101                                 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno));
 102                         }
 103                 }
 104         } else {
 105                 attr->inherit   = inherit;
 106                 attr->disabled  = 1;
 107
 108                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 109                 if (fd[0][counter] < 0 && verbose) {
 110                         printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno));
 111                 }
 112         }
 113 }
 114
 115 /*
 116  * Does the counter have nsecs as a unit?
 117  */
 118 static inline int nsec_counter(int counter)
 119 {
 120         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 121                 return 0;
 122
 123         if (attrs[counter].config == PERF_COUNT_CPU_CLOCK)
 124                 return 1;
 125
 126         if (attrs[counter].config == PERF_COUNT_TASK_CLOCK)
 127                 return 1;
 128
 129         return 0;
 130 }
 131
 132 /*
 133  * Read out the results of a single counter:
 134  */
 135 static void read_counter(int counter)
 136 {
 137         __u64 *count, single_count[3];
 138         ssize_t res;
 139         int cpu, nv;
 140         int scaled;
 141
 142         count = event_res[counter];
 143
 144         count[0] = count[1] = count[2] = 0;
 145
 146         nv = scale ? 3 : 1;
 147         for (cpu = 0; cpu < nr_cpus; cpu ++) {
 148                 if (fd[cpu][counter] < 0)
 149                         continue;
 150
 151                 res = read(fd[cpu][counter], single_count, nv * sizeof(__u64));
 152                 assert(res == nv * sizeof(__u64));
 153
 154                 count[0] += single_count[0];
 155                 if (scale) {
 156                         count[1] += single_count[1];
 157                         count[2] += single_count[2];
 158                 }
 159         }
 160
 161         scaled = 0;
 162         if (scale) {
 163                 if (count[2] == 0) {
 164                         event_scaled[counter] = -1;
 165                         count[0] = 0;
 166                         return;
 167                 }
 168
 169                 if (count[2] < count[1]) {
 170                         event_scaled[counter] = 1;
 171                         count[0] = (unsigned long long)
 172                                 ((double)count[0] * count[1] / count[2] + 0.5);
 173                 }
 174         }
 175         /*
 176          * Save the full runtime - to allow normalization during printout:
 177          */
 178         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 179                 attrs[counter].config == PERF_COUNT_TASK_CLOCK)
 180                 runtime_nsecs = count[0];
 181         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 182                 attrs[counter].config == PERF_COUNT_CPU_CYCLES)
 183                 runtime_cycles = count[0];
 184 }
 185
 186 /*
 187  * Print out the results of a single counter:
 188  */
 189 static void print_counter(int counter)
 190 {
 191         __u64 *count;
 192         int scaled;
 193
 194         count = event_res[counter];
 195         scaled = event_scaled[counter];
 196
 197         if (scaled == -1) {
 198                 fprintf(stderr, " %14s  %-20s\n",
 199                         "<not counted>", event_name(counter));
 200                 return;
 201         }
 202
 203         if (nsec_counter(counter)) {
 204                 double msecs = (double)count[0] / 1000000;
 205
 206                 fprintf(stderr, " %14.6f  %-20s",
 207                         msecs, event_name(counter));
 208                 if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 209                         attrs[counter].config == PERF_COUNT_TASK_CLOCK) {
 210
 211                         if (walltime_nsecs)
 212                                 fprintf(stderr, " # %11.3f CPU utilization factor",
 213                                         (double)count[0] / (double)walltime_nsecs);
 214                 }
 215         } else {
 216                 fprintf(stderr, " %14Ld  %-20s",
 217                         count[0], event_name(counter));
 218                 if (runtime_nsecs)
 219                         fprintf(stderr, " # %11.3f M/sec",
 220                                 (double)count[0]/runtime_nsecs*1000.0);
 221                 if (runtime_cycles &&
 222                         attrs[counter].type == PERF_TYPE_HARDWARE &&
 223                                 attrs[counter].config == PERF_COUNT_INSTRUCTIONS) {
 224
 225                         fprintf(stderr, " # %1.3f per cycle",
 226                                 (double)count[0] / (double)runtime_cycles);
 227                 }
 228         }
 229         if (scaled)
 230                 fprintf(stderr, "  (scaled from %.2f%%)",
 231                         (double) count[2] / count[1] * 100);
 232         fprintf(stderr, "\n");
 233 }
 234
 235 static int do_perf_stat(int argc, const char **argv)
 236 {
 237         unsigned long long t0, t1;
 238         int counter;
 239         int status;
 240         int pid;
 241         int i;
 242
 243         if (!system_wide)
 244                 nr_cpus = 1;
 245
 246         for (counter = 0; counter < nr_counters; counter++)
 247                 create_perf_stat_counter(counter);
 248
 249         /*
 250          * Enable counters and exec the command:
 251          */
 252         t0 = rdclock();
 253         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 254
 255         if ((pid = fork()) < 0)
 256                 perror("failed to fork");
 257
 258         if (!pid) {
 259                 if (execvp(argv[0], (char **)argv)) {
 260                         perror(argv[0]);
 261                         exit(-1);
 262                 }
 263         }
 264
 265         while (wait(&status) >= 0)
 266                 ;
 267
 268         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 269         t1 = rdclock();
 270
 271         walltime_nsecs = t1 - t0;
 272
 273         fflush(stdout);
 274
 275         fprintf(stderr, "\n");
 276         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 277
 278         for (i = 1; i < argc; i++)
 279                 fprintf(stderr, " %s", argv[i]);
 280
 281         fprintf(stderr, "\':\n");
 282         fprintf(stderr, "\n");
 283
 284         for (counter = 0; counter < nr_counters; counter++)
 285                 read_counter(counter);
 286
 287         for (counter = 0; counter < nr_counters; counter++)
 288                 print_counter(counter);
 289
 290
 291         fprintf(stderr, "\n");
 292         fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
 293                         (double)(t1-t0)/1e6);
 294         fprintf(stderr, "\n");
 295
 296         return 0;
 297 }
 298
 299 static volatile int signr = -1;
 300
 301 static void skip_signal(int signo)
 302 {
 303         signr = signo;
 304 }
 305
 306 static void sig_atexit(void)
 307 {
 308         if (signr == -1)
 309                 return;
 310
 311         signal(signr, SIG_DFL);
 312         kill(getpid(), signr);
 313 }
 314
 315 static const char * const stat_usage[] = {
 316         "perf stat [<options>] <command>",
 317         NULL
 318 };
 319
 320 static const struct option options[] = {
 321         OPT_CALLBACK('e', "event", NULL, "event",
 322                      "event selector. use 'perf list' to list available events",
 323                      parse_events),
 324         OPT_BOOLEAN('i', "inherit", &inherit,
 325                     "child tasks inherit counters"),
 326         OPT_INTEGER('p', "pid", &target_pid,
 327                     "stat events on existing pid"),
 328         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 329                             "system-wide collection from all CPUs"),
 330         OPT_BOOLEAN('S', "scale", &scale,
 331                             "scale/normalize counters"),
 332         OPT_BOOLEAN('v', "verbose", &verbose,
 333                     "be more verbose (show counter open errors, etc)"),
 334         OPT_END()
 335 };
 336
 337 int cmd_stat(int argc, const char **argv, const char *prefix)
 338 {
 339         page_size = sysconf(_SC_PAGE_SIZE);
 340
 341         memcpy(attrs, default_attrs, sizeof(attrs));
 342
 343         argc = parse_options(argc, argv, options, stat_usage, 0);
 344         if (!argc)
 345                 usage_with_options(stat_usage, options);
 346
 347         if (!nr_counters)
 348                 nr_counters = 8;
 349
 350         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 351         assert(nr_cpus <= MAX_NR_CPUS);
 352         assert(nr_cpus >= 0);
 353
 354         /*
 355          * We dont want to block the signals - that would cause
 356          * child tasks to inherit that and Ctrl-C would not work.
 357          * What we want is for Ctrl-C to work in the exec()-ed
 358          * task, but being ignored by perf stat itself:
 359          */
 360         atexit(sig_atexit);
 361         signal(SIGINT,  skip_signal);
 362         signal(SIGALRM, skip_signal);
 363         signal(SIGABRT, skip_signal);
 364
 365         return do_perf_stat(argc, argv);
 366 }