1 #include <inttypes.h> 2 #include <math.h> 3 #include "stat.h" 4 #include "evlist.h" 5 #include "evsel.h" 6 #include "thread_map.h" 7 8 void update_stats(struct stats *stats, u64 val) 9 { 10 double delta; 11 12 stats->n++; 13 delta = val - stats->mean; 14 stats->mean += delta / stats->n; 15 stats->M2 += delta*(val - stats->mean); 16 17 if (val > stats->max) 18 stats->max = val; 19 20 if (val < stats->min) 21 stats->min = val; 22 } 23 24 double avg_stats(struct stats *stats) 25 { 26 return stats->mean; 27 } 28 29 /* 30 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 31 * 32 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 33 * s^2 = ------------------------------- 34 * n - 1 35 * 36 * http://en.wikipedia.org/wiki/Stddev 37 * 38 * The std dev of the mean is related to the std dev by: 39 * 40 * s 41 * s_mean = ------- 42 * sqrt(n) 43 * 44 */ 45 double stddev_stats(struct stats *stats) 46 { 47 double variance, variance_mean; 48 49 if (stats->n < 2) 50 return 0.0; 51 52 variance = stats->M2 / (stats->n - 1); 53 variance_mean = variance / stats->n; 54 55 return sqrt(variance_mean); 56 } 57 58 double rel_stddev_stats(double stddev, double avg) 59 { 60 double pct = 0.0; 61 62 if (avg) 63 pct = 100.0 * stddev/avg; 64 65 return pct; 66 } 67 68 bool __perf_evsel_stat__is(struct perf_evsel *evsel, 69 enum perf_stat_evsel_id id) 70 { 71 struct perf_stat_evsel *ps = evsel->priv; 72 73 return ps->id == id; 74 } 75 76 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name 77 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = { 78 ID(NONE, x), 79 ID(CYCLES_IN_TX, cpu/cycles-t/), 80 ID(TRANSACTION_START, cpu/tx-start/), 81 ID(ELISION_START, cpu/el-start/), 82 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/), 83 ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots), 84 ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued), 85 ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired), 86 ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles), 87 ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles), 88 }; 89 #undef ID 90 91 void perf_stat_evsel_id_init(struct perf_evsel *evsel) 92 { 93 struct perf_stat_evsel *ps = evsel->priv; 94 int i; 95 96 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */ 97 98 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) { 99 if (!strcmp(perf_evsel__name(evsel), id_str[i])) { 100 ps->id = i; 101 break; 102 } 103 } 104 } 105 106 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel) 107 { 108 int i; 109 struct perf_stat_evsel *ps = evsel->priv; 110 111 for (i = 0; i < 3; i++) 112 init_stats(&ps->res_stats[i]); 113 114 perf_stat_evsel_id_init(evsel); 115 } 116 117 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) 118 { 119 evsel->priv = zalloc(sizeof(struct perf_stat_evsel)); 120 if (evsel->priv == NULL) 121 return -ENOMEM; 122 perf_evsel__reset_stat_priv(evsel); 123 return 0; 124 } 125 126 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) 127 { 128 zfree(&evsel->priv); 129 } 130 131 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel, 132 int ncpus, int nthreads) 133 { 134 struct perf_counts *counts; 135 136 counts = perf_counts__new(ncpus, nthreads); 137 if (counts) 138 evsel->prev_raw_counts = counts; 139 140 return counts ? 0 : -ENOMEM; 141 } 142 143 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel) 144 { 145 perf_counts__delete(evsel->prev_raw_counts); 146 evsel->prev_raw_counts = NULL; 147 } 148 149 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw) 150 { 151 int ncpus = perf_evsel__nr_cpus(evsel); 152 int nthreads = thread_map__nr(evsel->threads); 153 154 if (perf_evsel__alloc_stat_priv(evsel) < 0 || 155 perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 || 156 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0)) 157 return -ENOMEM; 158 159 return 0; 160 } 161 162 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw) 163 { 164 struct perf_evsel *evsel; 165 166 evlist__for_each_entry(evlist, evsel) { 167 if (perf_evsel__alloc_stats(evsel, alloc_raw)) 168 goto out_free; 169 } 170 171 return 0; 172 173 out_free: 174 perf_evlist__free_stats(evlist); 175 return -1; 176 } 177 178 void perf_evlist__free_stats(struct perf_evlist *evlist) 179 { 180 struct perf_evsel *evsel; 181 182 evlist__for_each_entry(evlist, evsel) { 183 perf_evsel__free_stat_priv(evsel); 184 perf_evsel__free_counts(evsel); 185 perf_evsel__free_prev_raw_counts(evsel); 186 } 187 } 188 189 void perf_evlist__reset_stats(struct perf_evlist *evlist) 190 { 191 struct perf_evsel *evsel; 192 193 evlist__for_each_entry(evlist, evsel) { 194 perf_evsel__reset_stat_priv(evsel); 195 perf_evsel__reset_counts(evsel); 196 } 197 } 198 199 static void zero_per_pkg(struct perf_evsel *counter) 200 { 201 if (counter->per_pkg_mask) 202 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS); 203 } 204 205 static int check_per_pkg(struct perf_evsel *counter, 206 struct perf_counts_values *vals, int cpu, bool *skip) 207 { 208 unsigned long *mask = counter->per_pkg_mask; 209 struct cpu_map *cpus = perf_evsel__cpus(counter); 210 int s; 211 212 *skip = false; 213 214 if (!counter->per_pkg) 215 return 0; 216 217 if (cpu_map__empty(cpus)) 218 return 0; 219 220 if (!mask) { 221 mask = zalloc(MAX_NR_CPUS); 222 if (!mask) 223 return -ENOMEM; 224 225 counter->per_pkg_mask = mask; 226 } 227 228 /* 229 * we do not consider an event that has not run as a good 230 * instance to mark a package as used (skip=1). Otherwise 231 * we may run into a situation where the first CPU in a package 232 * is not running anything, yet the second is, and this function 233 * would mark the package as used after the first CPU and would 234 * not read the values from the second CPU. 235 */ 236 if (!(vals->run && vals->ena)) 237 return 0; 238 239 s = cpu_map__get_socket(cpus, cpu, NULL); 240 if (s < 0) 241 return -1; 242 243 *skip = test_and_set_bit(s, mask) == 1; 244 return 0; 245 } 246 247 static int 248 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel, 249 int cpu, int thread, 250 struct perf_counts_values *count) 251 { 252 struct perf_counts_values *aggr = &evsel->counts->aggr; 253 static struct perf_counts_values zero; 254 bool skip = false; 255 256 if (check_per_pkg(evsel, count, cpu, &skip)) { 257 pr_err("failed to read per-pkg counter\n"); 258 return -1; 259 } 260 261 if (skip) 262 count = &zero; 263 264 switch (config->aggr_mode) { 265 case AGGR_THREAD: 266 case AGGR_CORE: 267 case AGGR_SOCKET: 268 case AGGR_NONE: 269 if (!evsel->snapshot) 270 perf_evsel__compute_deltas(evsel, cpu, thread, count); 271 perf_counts_values__scale(count, config->scale, NULL); 272 if (config->aggr_mode == AGGR_NONE) 273 perf_stat__update_shadow_stats(evsel, count->values, cpu); 274 break; 275 case AGGR_GLOBAL: 276 aggr->val += count->val; 277 if (config->scale) { 278 aggr->ena += count->ena; 279 aggr->run += count->run; 280 } 281 case AGGR_UNSET: 282 default: 283 break; 284 } 285 286 return 0; 287 } 288 289 static int process_counter_maps(struct perf_stat_config *config, 290 struct perf_evsel *counter) 291 { 292 int nthreads = thread_map__nr(counter->threads); 293 int ncpus = perf_evsel__nr_cpus(counter); 294 int cpu, thread; 295 296 if (counter->system_wide) 297 nthreads = 1; 298 299 for (thread = 0; thread < nthreads; thread++) { 300 for (cpu = 0; cpu < ncpus; cpu++) { 301 if (process_counter_values(config, counter, cpu, thread, 302 perf_counts(counter->counts, cpu, thread))) 303 return -1; 304 } 305 } 306 307 return 0; 308 } 309 310 int perf_stat_process_counter(struct perf_stat_config *config, 311 struct perf_evsel *counter) 312 { 313 struct perf_counts_values *aggr = &counter->counts->aggr; 314 struct perf_stat_evsel *ps = counter->priv; 315 u64 *count = counter->counts->aggr.values; 316 u64 val; 317 int i, ret; 318 319 aggr->val = aggr->ena = aggr->run = 0; 320 321 /* 322 * We calculate counter's data every interval, 323 * and the display code shows ps->res_stats 324 * avg value. We need to zero the stats for 325 * interval mode, otherwise overall avg running 326 * averages will be shown for each interval. 327 */ 328 if (config->interval) 329 init_stats(ps->res_stats); 330 331 if (counter->per_pkg) 332 zero_per_pkg(counter); 333 334 ret = process_counter_maps(config, counter); 335 if (ret) 336 return ret; 337 338 if (config->aggr_mode != AGGR_GLOBAL) 339 return 0; 340 341 if (!counter->snapshot) 342 perf_evsel__compute_deltas(counter, -1, -1, aggr); 343 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled); 344 345 for (i = 0; i < 3; i++) 346 update_stats(&ps->res_stats[i], count[i]); 347 348 if (verbose > 0) { 349 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 350 perf_evsel__name(counter), count[0], count[1], count[2]); 351 } 352 353 /* 354 * Save the full runtime - to allow normalization during printout: 355 */ 356 val = counter->scale * *count; 357 perf_stat__update_shadow_stats(counter, &val, 0); 358 359 return 0; 360 } 361 362 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused, 363 union perf_event *event, 364 struct perf_session *session) 365 { 366 struct perf_counts_values count; 367 struct stat_event *st = &event->stat; 368 struct perf_evsel *counter; 369 370 count.val = st->val; 371 count.ena = st->ena; 372 count.run = st->run; 373 374 counter = perf_evlist__id2evsel(session->evlist, st->id); 375 if (!counter) { 376 pr_err("Failed to resolve counter for stat event.\n"); 377 return -EINVAL; 378 } 379 380 *perf_counts(counter->counts, st->cpu, st->thread) = count; 381 counter->supported = true; 382 return 0; 383 } 384 385 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp) 386 { 387 struct stat_event *st = (struct stat_event *) event; 388 size_t ret; 389 390 ret = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n", 391 st->id, st->cpu, st->thread); 392 ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n", 393 st->val, st->ena, st->run); 394 395 return ret; 396 } 397 398 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp) 399 { 400 struct stat_round_event *rd = (struct stat_round_event *)event; 401 size_t ret; 402 403 ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time, 404 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL"); 405 406 return ret; 407 } 408 409 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp) 410 { 411 struct perf_stat_config sc; 412 size_t ret; 413 414 perf_event__read_stat_config(&sc, &event->stat_config); 415 416 ret = fprintf(fp, "\n"); 417 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode); 418 ret += fprintf(fp, "... scale %d\n", sc.scale); 419 ret += fprintf(fp, "... interval %u\n", sc.interval); 420 421 return ret; 422 } 423