tools/perf/util/stat-shadow.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <math.h>
   3 #include <stdio.h>
   4 #include "evsel.h"
   5 #include "stat.h"
   6 #include "color.h"
   7 #include "debug.h"
   8 #include "pmu.h"
   9 #include "rblist.h"
  10 #include "evlist.h"
  11 #include "expr.h"
  12 #include "metricgroup.h"
  13 #include "cgroup.h"
  14 #include "units.h"
  15 #include <linux/zalloc.h>
  16 #include "iostat.h"
  17 #include "util/hashmap.h"
  18 #include "tool_pmu.h"
  19
  20 struct stats walltime_nsecs_stats;
  21 struct rusage_stats ru_stats;
  22
  23 enum {
  24         CTX_BIT_USER    = 1 << 0,
  25         CTX_BIT_KERNEL  = 1 << 1,
  26         CTX_BIT_HV      = 1 << 2,
  27         CTX_BIT_HOST    = 1 << 3,
  28         CTX_BIT_IDLE    = 1 << 4,
  29         CTX_BIT_MAX     = 1 << 5,
  30 };
  31
  32 enum stat_type {
  33         STAT_NONE = 0,
  34         STAT_NSECS,
  35         STAT_CYCLES,
  36         STAT_INSTRUCTIONS,
  37         STAT_STALLED_CYCLES_FRONT,
  38         STAT_STALLED_CYCLES_BACK,
  39         STAT_BRANCHES,
  40         STAT_BRANCH_MISS,
  41         STAT_CACHE_REFS,
  42         STAT_CACHE_MISSES,
  43         STAT_L1_DCACHE,
  44         STAT_L1_ICACHE,
  45         STAT_LL_CACHE,
  46         STAT_ITLB_CACHE,
  47         STAT_DTLB_CACHE,
  48         STAT_L1D_MISS,
  49         STAT_L1I_MISS,
  50         STAT_LL_MISS,
  51         STAT_DTLB_MISS,
  52         STAT_ITLB_MISS,
  53         STAT_MAX
  54 };
  55
  56 static int evsel_context(const struct evsel *evsel)
  57 {
  58         int ctx = 0;
  59
  60         if (evsel->core.attr.exclude_kernel)
  61                 ctx |= CTX_BIT_KERNEL;
  62         if (evsel->core.attr.exclude_user)
  63                 ctx |= CTX_BIT_USER;
  64         if (evsel->core.attr.exclude_hv)
  65                 ctx |= CTX_BIT_HV;
  66         if (evsel->core.attr.exclude_host)
  67                 ctx |= CTX_BIT_HOST;
  68         if (evsel->core.attr.exclude_idle)
  69                 ctx |= CTX_BIT_IDLE;
  70
  71         return ctx;
  72 }
  73
  74 void perf_stat__reset_shadow_stats(void)
  75 {
  76         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
  77         memset(&ru_stats, 0, sizeof(ru_stats));
  78 }
  79
  80 static enum stat_type evsel__stat_type(struct evsel *evsel)
  81 {
  82         /* Fake perf_hw_cache_op_id values for use with evsel__match. */
  83         u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
  84                 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
  85                 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
  86         u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
  87                 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
  88                 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
  89         u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
  90                 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
  91                 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
  92         u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
  93                 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
  94                 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
  95         u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
  96                 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
  97                 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
  98
  99         if (evsel__is_clock(evsel))
 100                 return STAT_NSECS;
 101         else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
 102                 return STAT_CYCLES;
 103         else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
 104                 return STAT_INSTRUCTIONS;
 105         else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 106                 return STAT_STALLED_CYCLES_FRONT;
 107         else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 108                 return STAT_STALLED_CYCLES_BACK;
 109         else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 110                 return STAT_BRANCHES;
 111         else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
 112                 return STAT_BRANCH_MISS;
 113         else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
 114                 return STAT_CACHE_REFS;
 115         else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
 116                 return STAT_CACHE_MISSES;
 117         else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
 118                 return STAT_L1_DCACHE;
 119         else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
 120                 return STAT_L1_ICACHE;
 121         else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
 122                 return STAT_LL_CACHE;
 123         else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
 124                 return STAT_DTLB_CACHE;
 125         else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
 126                 return STAT_ITLB_CACHE;
 127         else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
 128                 return STAT_L1D_MISS;
 129         else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
 130                 return STAT_L1I_MISS;
 131         else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
 132                 return STAT_LL_MISS;
 133         else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
 134                 return STAT_DTLB_MISS;
 135         else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
 136                 return STAT_ITLB_MISS;
 137         return STAT_NONE;
 138 }
 139
 140 static enum metric_threshold_classify get_ratio_thresh(const double ratios[3], double val)
 141 {
 142         assert(ratios[0] > ratios[1]);
 143         assert(ratios[1] > ratios[2]);
 144
 145         return val > ratios[1]
 146                 ? (val > ratios[0] ? METRIC_THRESHOLD_BAD : METRIC_THRESHOLD_NEARLY_BAD)
 147                 : (val > ratios[2] ? METRIC_THRESHOLD_LESS_GOOD : METRIC_THRESHOLD_GOOD);
 148 }
 149
 150 static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
 151 {
 152         struct evsel *cur;
 153         int evsel_ctx = evsel_context(evsel);
 154
 155         evlist__for_each_entry(evsel->evlist, cur) {
 156                 struct perf_stat_aggr *aggr;
 157
 158                 /* Ignore the evsel that is being searched from. */
 159                 if (evsel == cur)
 160                         continue;
 161
 162                 /* Ignore evsels that are part of different groups. */
 163                 if (evsel->core.leader->nr_members > 1 &&
 164                     evsel->core.leader != cur->core.leader)
 165                         continue;
 166                 /* Ignore evsels with mismatched modifiers. */
 167                 if (evsel_ctx != evsel_context(cur))
 168                         continue;
 169                 /* Ignore if not the cgroup we're looking for. */
 170                 if (evsel->cgrp != cur->cgrp)
 171                         continue;
 172                 /* Ignore if not the stat we're looking for. */
 173                 if (type != evsel__stat_type(cur))
 174                         continue;
 175
 176                 /*
 177                  * Except the SW CLOCK events,
 178                  * ignore if not the PMU we're looking for.
 179                  */
 180                 if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
 181                         continue;
 182
 183                 aggr = &cur->stats->aggr[aggr_idx];
 184                 if (type == STAT_NSECS)
 185                         return aggr->counts.val;
 186                 return aggr->counts.val * cur->scale;
 187         }
 188         return 0.0;
 189 }
 190
 191 static void print_ratio(struct perf_stat_config *config,
 192                         const struct evsel *evsel, int aggr_idx,
 193                         double numerator, struct perf_stat_output_ctx *out,
 194                         enum stat_type denominator_type,
 195                         const double thresh_ratios[3], const char *_unit)
 196 {
 197         double denominator = find_stat(evsel, aggr_idx, denominator_type);
 198         double ratio = 0;
 199         enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
 200         const char *fmt = NULL;
 201         const char *unit = NULL;
 202
 203         if (numerator && denominator) {
 204                 ratio = numerator / denominator * 100.0;
 205                 thresh = get_ratio_thresh(thresh_ratios, ratio);
 206                 fmt = "%7.2f%%";
 207                 unit = _unit;
 208         }
 209         out->print_metric(config, out->ctx, thresh, fmt, unit, ratio);
 210 }
 211
 212 static void print_stalled_cycles_front(struct perf_stat_config *config,
 213                                 const struct evsel *evsel,
 214                                 int aggr_idx, double stalled,
 215                                 struct perf_stat_output_ctx *out)
 216 {
 217         const double thresh_ratios[3] = {50.0, 30.0, 10.0};
 218
 219         print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
 220                     "frontend cycles idle");
 221 }
 222
 223 static void print_stalled_cycles_back(struct perf_stat_config *config,
 224                                 const struct evsel *evsel,
 225                                 int aggr_idx, double stalled,
 226                                 struct perf_stat_output_ctx *out)
 227 {
 228         const double thresh_ratios[3] = {75.0, 50.0, 20.0};
 229
 230         print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
 231                     "backend cycles idle");
 232 }
 233
 234 static void print_branch_miss(struct perf_stat_config *config,
 235                         const struct evsel *evsel,
 236                         int aggr_idx, double misses,
 237                         struct perf_stat_output_ctx *out)
 238 {
 239         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 240
 241         print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, thresh_ratios,
 242                     "of all branches");
 243 }
 244
 245 static void print_l1d_miss(struct perf_stat_config *config,
 246                         const struct evsel *evsel,
 247                         int aggr_idx, double misses,
 248                         struct perf_stat_output_ctx *out)
 249 {
 250         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 251
 252         print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, thresh_ratios,
 253                     "of all L1-dcache accesses");
 254 }
 255
 256 static void print_l1i_miss(struct perf_stat_config *config,
 257                         const struct evsel *evsel,
 258                         int aggr_idx, double misses,
 259                         struct perf_stat_output_ctx *out)
 260 {
 261         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 262
 263         print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, thresh_ratios,
 264                     "of all L1-icache accesses");
 265 }
 266
 267 static void print_ll_miss(struct perf_stat_config *config,
 268                         const struct evsel *evsel,
 269                         int aggr_idx, double misses,
 270                         struct perf_stat_output_ctx *out)
 271 {
 272         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 273
 274         print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, thresh_ratios,
 275                     "of all LL-cache accesses");
 276 }
 277
 278 static void print_dtlb_miss(struct perf_stat_config *config,
 279                         const struct evsel *evsel,
 280                         int aggr_idx, double misses,
 281                         struct perf_stat_output_ctx *out)
 282 {
 283         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 284
 285         print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, thresh_ratios,
 286                     "of all dTLB cache accesses");
 287 }
 288
 289 static void print_itlb_miss(struct perf_stat_config *config,
 290                         const struct evsel *evsel,
 291                         int aggr_idx, double misses,
 292                         struct perf_stat_output_ctx *out)
 293 {
 294         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 295
 296         print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, thresh_ratios,
 297                     "of all iTLB cache accesses");
 298 }
 299
 300 static void print_cache_miss(struct perf_stat_config *config,
 301                         const struct evsel *evsel,
 302                         int aggr_idx, double misses,
 303                         struct perf_stat_output_ctx *out)
 304 {
 305         const double thresh_ratios[3] = {20.0, 10.0, 5.0};
 306
 307         print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, thresh_ratios,
 308                     "of all cache refs");
 309 }
 310
 311 static void print_instructions(struct perf_stat_config *config,
 312                         const struct evsel *evsel,
 313                         int aggr_idx, double instructions,
 314                         struct perf_stat_output_ctx *out)
 315 {
 316         print_metric_t print_metric = out->print_metric;
 317         void *ctxp = out->ctx;
 318         double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
 319         double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
 320                                 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
 321
 322         if (cycles) {
 323                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
 324                              "insn per cycle", instructions / cycles);
 325         } else {
 326                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
 327                              "insn per cycle", 0);
 328         }
 329         if (max_stalled && instructions) {
 330                 if (out->new_line)
 331                         out->new_line(config, ctxp);
 332                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
 333                              "stalled cycles per insn", max_stalled / instructions);
 334         }
 335 }
 336
 337 static void print_cycles(struct perf_stat_config *config,
 338                         const struct evsel *evsel,
 339                         int aggr_idx, double cycles,
 340                         struct perf_stat_output_ctx *out)
 341 {
 342         double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
 343
 344         if (cycles && nsecs) {
 345                 double ratio = cycles / nsecs;
 346
 347                 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
 348                                   "GHz", ratio);
 349         } else {
 350                 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
 351                                   "GHz", 0);
 352         }
 353 }
 354
 355 static void print_nsecs(struct perf_stat_config *config,
 356                         const struct evsel *evsel,
 357                         int aggr_idx __maybe_unused, double nsecs,
 358                         struct perf_stat_output_ctx *out)
 359 {
 360         print_metric_t print_metric = out->print_metric;
 361         void *ctxp = out->ctx;
 362         double wall_time = avg_stats(&walltime_nsecs_stats);
 363
 364         if (wall_time) {
 365                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f", "CPUs utilized",
 366                         nsecs / (wall_time * evsel->scale));
 367         } else {
 368                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
 369                              "CPUs utilized", 0);
 370         }
 371 }
 372
 373 static int prepare_metric(const struct metric_expr *mexp,
 374                           const struct evsel *evsel,
 375                           struct expr_parse_ctx *pctx,
 376                           int aggr_idx)
 377 {
 378         struct evsel * const *metric_events = mexp->metric_events;
 379         struct metric_ref *metric_refs = mexp->metric_refs;
 380         int i;
 381
 382         for (i = 0; metric_events[i]; i++) {
 383                 char *n;
 384                 double val;
 385                 int source_count = 0;
 386
 387                 if (evsel__is_tool(metric_events[i])) {
 388                         struct stats *stats;
 389                         double scale;
 390
 391                         switch (evsel__tool_event(metric_events[i])) {
 392                         case TOOL_PMU__EVENT_DURATION_TIME:
 393                                 stats = &walltime_nsecs_stats;
 394                                 scale = 1e-9;
 395                                 break;
 396                         case TOOL_PMU__EVENT_USER_TIME:
 397                                 stats = &ru_stats.ru_utime_usec_stat;
 398                                 scale = 1e-6;
 399                                 break;
 400                         case TOOL_PMU__EVENT_SYSTEM_TIME:
 401                                 stats = &ru_stats.ru_stime_usec_stat;
 402                                 scale = 1e-6;
 403                                 break;
 404                         case TOOL_PMU__EVENT_NONE:
 405                                 pr_err("Invalid tool event 'none'");
 406                                 abort();
 407                         case TOOL_PMU__EVENT_MAX:
 408                                 pr_err("Invalid tool event 'max'");
 409                                 abort();
 410                         case TOOL_PMU__EVENT_HAS_PMEM:
 411                         case TOOL_PMU__EVENT_NUM_CORES:
 412                         case TOOL_PMU__EVENT_NUM_CPUS:
 413                         case TOOL_PMU__EVENT_NUM_CPUS_ONLINE:
 414                         case TOOL_PMU__EVENT_NUM_DIES:
 415                         case TOOL_PMU__EVENT_NUM_PACKAGES:
 416                         case TOOL_PMU__EVENT_SLOTS:
 417                         case TOOL_PMU__EVENT_SMT_ON:
 418                         case TOOL_PMU__EVENT_SYSTEM_TSC_FREQ:
 419                         default:
 420                                 pr_err("Unexpected tool event '%s'", evsel__name(metric_events[i]));
 421                                 abort();
 422                         }
 423                         val = avg_stats(stats) * scale;
 424                         source_count = 1;
 425                 } else {
 426                         struct perf_stat_evsel *ps = metric_events[i]->stats;
 427                         struct perf_stat_aggr *aggr;
 428
 429                         /*
 430                          * If there are multiple uncore PMUs and we're not
 431                          * reading the leader's stats, determine the stats for
 432                          * the appropriate uncore PMU.
 433                          */
 434                         if (evsel && evsel->metric_leader &&
 435                             evsel->pmu != evsel->metric_leader->pmu &&
 436                             mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
 437                                 struct evsel *pos;
 438
 439                                 evlist__for_each_entry(evsel->evlist, pos) {
 440                                         if (pos->pmu != evsel->pmu)
 441                                                 continue;
 442                                         if (pos->metric_leader != mexp->metric_events[i])
 443                                                 continue;
 444                                         ps = pos->stats;
 445                                         source_count = 1;
 446                                         break;
 447                                 }
 448                         }
 449                         aggr = &ps->aggr[aggr_idx];
 450                         if (!aggr)
 451                                 break;
 452
 453                         if (!metric_events[i]->supported) {
 454                                 /*
 455                                  * Not supported events will have a count of 0,
 456                                  * which can be confusing in a
 457                                  * metric. Explicitly set the value to NAN. Not
 458                                  * counted events (enable time of 0) are read as
 459                                  * 0.
 460                                  */
 461                                 val = NAN;
 462                                 source_count = 0;
 463                         } else {
 464                                 val = aggr->counts.val;
 465                                 if (!source_count)
 466                                         source_count = evsel__source_count(metric_events[i]);
 467                         }
 468                 }
 469                 n = strdup(evsel__metric_id(metric_events[i]));
 470                 if (!n)
 471                         return -ENOMEM;
 472
 473                 expr__add_id_val_source_count(pctx, n, val, source_count);
 474         }
 475
 476         for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
 477                 int ret = expr__add_ref(pctx, &metric_refs[j]);
 478
 479                 if (ret)
 480                         return ret;
 481         }
 482
 483         return i;
 484 }
 485
 486 static void generic_metric(struct perf_stat_config *config,
 487                            struct metric_expr *mexp,
 488                            struct evsel *evsel,
 489                            int aggr_idx,
 490                            struct perf_stat_output_ctx *out)
 491 {
 492         print_metric_t print_metric = out->print_metric;
 493         const char *metric_name = mexp->metric_name;
 494         const char *metric_expr = mexp->metric_expr;
 495         const char *metric_threshold = mexp->metric_threshold;
 496         const char *metric_unit = mexp->metric_unit;
 497         struct evsel * const *metric_events = mexp->metric_events;
 498         int runtime = mexp->runtime;
 499         struct expr_parse_ctx *pctx;
 500         double ratio, scale, threshold;
 501         int i;
 502         void *ctxp = out->ctx;
 503         enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
 504
 505         pctx = expr__ctx_new();
 506         if (!pctx)
 507                 return;
 508
 509         if (config->user_requested_cpu_list)
 510                 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
 511         pctx->sctx.runtime = runtime;
 512         pctx->sctx.system_wide = config->system_wide;
 513         i = prepare_metric(mexp, evsel, pctx, aggr_idx);
 514         if (i < 0) {
 515                 expr__ctx_free(pctx);
 516                 return;
 517         }
 518         if (!metric_events[i]) {
 519                 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
 520                         char *unit;
 521                         char metric_bf[128];
 522
 523                         if (metric_threshold &&
 524                             expr__parse(&threshold, pctx, metric_threshold) == 0 &&
 525                             !isnan(threshold)) {
 526                                 thresh = fpclassify(threshold) == FP_ZERO
 527                                         ? METRIC_THRESHOLD_GOOD : METRIC_THRESHOLD_BAD;
 528                         }
 529
 530                         if (metric_unit && metric_name) {
 531                                 if (perf_pmu__convert_scale(metric_unit,
 532                                         &unit, &scale) >= 0) {
 533                                         ratio *= scale;
 534                                 }
 535                                 if (strstr(metric_expr, "?"))
 536                                         scnprintf(metric_bf, sizeof(metric_bf),
 537                                           "%s  %s_%d", unit, metric_name, runtime);
 538                                 else
 539                                         scnprintf(metric_bf, sizeof(metric_bf),
 540                                           "%s  %s", unit, metric_name);
 541
 542                                 print_metric(config, ctxp, thresh, "%8.1f",
 543                                              metric_bf, ratio);
 544                         } else {
 545                                 print_metric(config, ctxp, thresh, "%8.2f",
 546                                         metric_name ?
 547                                         metric_name :
 548                                         out->force_header ?  evsel->name : "",
 549                                         ratio);
 550                         }
 551                 } else {
 552                         print_metric(config, ctxp, thresh, /*fmt=*/NULL,
 553                                      out->force_header ?
 554                                      (metric_name ?: evsel->name) : "", 0);
 555                 }
 556         } else {
 557                 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
 558                              out->force_header ?
 559                              (metric_name ?: evsel->name) : "", 0);
 560         }
 561
 562         expr__ctx_free(pctx);
 563 }
 564
 565 double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
 566 {
 567         struct expr_parse_ctx *pctx;
 568         double ratio = 0.0;
 569
 570         pctx = expr__ctx_new();
 571         if (!pctx)
 572                 return NAN;
 573
 574         if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
 575                 goto out;
 576
 577         if (expr__parse(&ratio, pctx, mexp->metric_expr))
 578                 ratio = 0.0;
 579
 580 out:
 581         expr__ctx_free(pctx);
 582         return ratio;
 583 }
 584
 585 static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
 586                                                 struct evsel *evsel,
 587                                                 void *ctxp,
 588                                                 const char *name,
 589                                                 struct perf_stat_output_ctx *out)
 590 {
 591         bool need_full_name = perf_pmus__num_core_pmus() > 1;
 592         static const char *last_name;
 593         static const struct perf_pmu *last_pmu;
 594         char full_name[64];
 595
 596         /*
 597          * A metricgroup may have several metric events,
 598          * e.g.,TopdownL1 on e-core of ADL.
 599          * The name has been output by the first metric
 600          * event. Only align with other metics from
 601          * different metric events.
 602          */
 603         if (last_name && !strcmp(last_name, name)) {
 604                 if (!need_full_name || last_pmu != evsel->pmu) {
 605                         out->print_metricgroup_header(config, ctxp, NULL);
 606                         return;
 607                 }
 608         }
 609
 610         if (need_full_name && evsel->pmu)
 611                 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu->name);
 612         else
 613                 scnprintf(full_name, sizeof(full_name), "%s", name);
 614
 615         out->print_metricgroup_header(config, ctxp, full_name);
 616
 617         last_name = name;
 618         last_pmu = evsel->pmu;
 619 }
 620
 621 /**
 622  * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
 623  *                                             For the non-default, all metrics associated
 624  *                                             with the evsel are printed.
 625  *                                             For the default mode, only the metrics from
 626  *                                             the same metricgroup and the name of the
 627  *                                             metricgroup are printed. To print the metrics
 628  *                                             from the next metricgroup (if available),
 629  *                                             invoke the function with correspoinding
 630  *                                             metric_expr.
 631  */
 632 void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
 633                                                 struct evsel *evsel,
 634                                                 int aggr_idx,
 635                                                 int *num,
 636                                                 void *from,
 637                                                 struct perf_stat_output_ctx *out,
 638                                                 struct rblist *metric_events)
 639 {
 640         struct metric_event *me;
 641         struct metric_expr *mexp = from;
 642         void *ctxp = out->ctx;
 643         bool header_printed = false;
 644         const char *name = NULL;
 645
 646         me = metricgroup__lookup(metric_events, evsel, false);
 647         if (me == NULL)
 648                 return NULL;
 649
 650         if (!mexp)
 651                 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
 652
 653         list_for_each_entry_from(mexp, &me->head, nd) {
 654                 /* Print the display name of the Default metricgroup */
 655                 if (!config->metric_only && me->is_default) {
 656                         if (!name)
 657                                 name = mexp->default_metricgroup_name;
 658                         /*
 659                          * Two or more metricgroup may share the same metric
 660                          * event, e.g., TopdownL1 and TopdownL2 on SPR.
 661                          * Return and print the prefix, e.g., noise, running
 662                          * for the next metricgroup.
 663                          */
 664                         if (strcmp(name, mexp->default_metricgroup_name))
 665                                 return (void *)mexp;
 666                         /* Only print the name of the metricgroup once */
 667                         if (!header_printed) {
 668                                 header_printed = true;
 669                                 perf_stat__print_metricgroup_header(config, evsel, ctxp,
 670                                                                     name, out);
 671                         }
 672                 }
 673
 674                 if ((*num)++ > 0 && out->new_line)
 675                         out->new_line(config, ctxp);
 676                 generic_metric(config, mexp, evsel, aggr_idx, out);
 677         }
 678
 679         return NULL;
 680 }
 681
 682 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
 683                                    struct evsel *evsel,
 684                                    double avg, int aggr_idx,
 685                                    struct perf_stat_output_ctx *out,
 686                                    struct rblist *metric_events)
 687 {
 688         typedef void (*stat_print_function_t)(struct perf_stat_config *config,
 689                                         const struct evsel *evsel,
 690                                         int aggr_idx, double misses,
 691                                         struct perf_stat_output_ctx *out);
 692         static const stat_print_function_t stat_print_function[STAT_MAX] = {
 693                 [STAT_INSTRUCTIONS] = print_instructions,
 694                 [STAT_BRANCH_MISS] = print_branch_miss,
 695                 [STAT_L1D_MISS] = print_l1d_miss,
 696                 [STAT_L1I_MISS] = print_l1i_miss,
 697                 [STAT_DTLB_MISS] = print_dtlb_miss,
 698                 [STAT_ITLB_MISS] = print_itlb_miss,
 699                 [STAT_LL_MISS] = print_ll_miss,
 700                 [STAT_CACHE_MISSES] = print_cache_miss,
 701                 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
 702                 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
 703                 [STAT_CYCLES] = print_cycles,
 704                 [STAT_NSECS] = print_nsecs,
 705         };
 706         print_metric_t print_metric = out->print_metric;
 707         void *ctxp = out->ctx;
 708         int num = 1;
 709
 710         if (config->iostat_run) {
 711                 iostat_print_metric(config, evsel, out);
 712         } else {
 713                 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
 714
 715                 if (fn)
 716                         fn(config, evsel, aggr_idx, avg, out);
 717                 else {
 718                         double nsecs =  find_stat(evsel, aggr_idx, STAT_NSECS);
 719
 720                         if (nsecs) {
 721                                 char unit = ' ';
 722                                 char unit_buf[10] = "/sec";
 723                                 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
 724                                                                    &unit);
 725
 726                                 if (unit != ' ')
 727                                         snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
 728                                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
 729                                              unit_buf, ratio);
 730                         } else {
 731                                 num = 0;
 732                         }
 733                 }
 734         }
 735
 736         perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
 737                                                   &num, NULL, out, metric_events);
 738
 739         if (num == 0) {
 740                 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN,
 741                              /*fmt=*/NULL, /*unit=*/NULL, 0);
 742         }
 743 }
 744
 745 /**
 746  * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
 747  *                                if it's not running or not the metric event.
 748  */
 749 bool perf_stat__skip_metric_event(struct evsel *evsel,
 750                                   struct rblist *metric_events,
 751                                   u64 ena, u64 run)
 752 {
 753         if (!evsel->default_metricgroup)
 754                 return false;
 755
 756         if (!ena || !run)
 757                 return true;
 758
 759         return !metricgroup__lookup(metric_events, evsel, false);
 760 }