tools/perf/util/stat-shadow.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <stdio.h>
   3 #include "evsel.h"
   4 #include "stat.h"
   5 #include "color.h"
   6 #include "pmu.h"
   7 #include "rblist.h"
   8 #include "evlist.h"
   9 #include "expr.h"
  10
  11 enum {
  12         CTX_BIT_USER    = 1 << 0,
  13         CTX_BIT_KERNEL  = 1 << 1,
  14         CTX_BIT_HV      = 1 << 2,
  15         CTX_BIT_HOST    = 1 << 3,
  16         CTX_BIT_IDLE    = 1 << 4,
  17         CTX_BIT_MAX     = 1 << 5,
  18 };
  19
  20 #define NUM_CTX CTX_BIT_MAX
  21
  22 /*
  23  * AGGR_GLOBAL: Use CPU 0
  24  * AGGR_SOCKET: Use first CPU of socket
  25  * AGGR_CORE: Use first CPU of core
  26  * AGGR_NONE: Use matching CPU
  27  * AGGR_THREAD: Not supported?
  28  */
  29 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
  30 static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
  31 static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
  32 static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
  33 static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
  34 static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
  35 static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
  36 static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
  37 static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
  38 static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
  39 static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
  40 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
  41 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
  42 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
  43 static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
  44 static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
  45 static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
  46 static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
  47 static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
  48 static struct stats runtime_smi_num_stats[NUM_CTX][MAX_NR_CPUS];
  49 static struct stats runtime_aperf_stats[NUM_CTX][MAX_NR_CPUS];
  50 static struct rblist runtime_saved_values;
  51 static bool have_frontend_stalled;
  52
  53 struct stats walltime_nsecs_stats;
  54
  55 struct saved_value {
  56         struct rb_node rb_node;
  57         struct perf_evsel *evsel;
  58         int cpu;
  59         int ctx;
  60         struct stats stats;
  61 };
  62
  63 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
  64 {
  65         struct saved_value *a = container_of(rb_node,
  66                                              struct saved_value,
  67                                              rb_node);
  68         const struct saved_value *b = entry;
  69
  70         if (a->ctx != b->ctx)
  71                 return a->ctx - b->ctx;
  72         if (a->cpu != b->cpu)
  73                 return a->cpu - b->cpu;
  74         if (a->evsel == b->evsel)
  75                 return 0;
  76         if ((char *)a->evsel < (char *)b->evsel)
  77                 return -1;
  78         return +1;
  79 }
  80
  81 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
  82                                      const void *entry)
  83 {
  84         struct saved_value *nd = malloc(sizeof(struct saved_value));
  85
  86         if (!nd)
  87                 return NULL;
  88         memcpy(nd, entry, sizeof(struct saved_value));
  89         return &nd->rb_node;
  90 }
  91
  92 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
  93                                               int cpu, int ctx,
  94                                               bool create)
  95 {
  96         struct rb_node *nd;
  97         struct saved_value dm = {
  98                 .cpu = cpu,
  99                 .ctx = ctx,
 100                 .evsel = evsel,
 101         };
 102         nd = rblist__find(&runtime_saved_values, &dm);
 103         if (nd)
 104                 return container_of(nd, struct saved_value, rb_node);
 105         if (create) {
 106                 rblist__add_node(&runtime_saved_values, &dm);
 107                 nd = rblist__find(&runtime_saved_values, &dm);
 108                 if (nd)
 109                         return container_of(nd, struct saved_value, rb_node);
 110         }
 111         return NULL;
 112 }
 113
 114 void perf_stat__init_shadow_stats(void)
 115 {
 116         have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
 117         rblist__init(&runtime_saved_values);
 118         runtime_saved_values.node_cmp = saved_value_cmp;
 119         runtime_saved_values.node_new = saved_value_new;
 120         /* No delete for now */
 121 }
 122
 123 static int evsel_context(struct perf_evsel *evsel)
 124 {
 125         int ctx = 0;
 126
 127         if (evsel->attr.exclude_kernel)
 128                 ctx |= CTX_BIT_KERNEL;
 129         if (evsel->attr.exclude_user)
 130                 ctx |= CTX_BIT_USER;
 131         if (evsel->attr.exclude_hv)
 132                 ctx |= CTX_BIT_HV;
 133         if (evsel->attr.exclude_host)
 134                 ctx |= CTX_BIT_HOST;
 135         if (evsel->attr.exclude_idle)
 136                 ctx |= CTX_BIT_IDLE;
 137
 138         return ctx;
 139 }
 140
 141 void perf_stat__reset_shadow_stats(void)
 142 {
 143         struct rb_node *pos, *next;
 144
 145         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 146         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 147         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 148         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 149         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 150         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 151         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 152         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 153         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 154         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 155         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 156         memset(runtime_cycles_in_tx_stats, 0,
 157                         sizeof(runtime_cycles_in_tx_stats));
 158         memset(runtime_transaction_stats, 0,
 159                 sizeof(runtime_transaction_stats));
 160         memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
 161         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
 162         memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
 163         memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
 164         memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
 165         memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
 166         memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
 167         memset(runtime_smi_num_stats, 0, sizeof(runtime_smi_num_stats));
 168         memset(runtime_aperf_stats, 0, sizeof(runtime_aperf_stats));
 169
 170         next = rb_first(&runtime_saved_values.entries);
 171         while (next) {
 172                 pos = next;
 173                 next = rb_next(pos);
 174                 memset(&container_of(pos, struct saved_value, rb_node)->stats,
 175                        0,
 176                        sizeof(struct stats));
 177         }
 178 }
 179
 180 /*
 181  * Update various tracking values we maintain to print
 182  * more semantic information such as miss/hit ratios,
 183  * instruction rates, etc:
 184  */
 185 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
 186                                     int cpu)
 187 {
 188         int ctx = evsel_context(counter);
 189
 190         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
 191             perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
 192                 update_stats(&runtime_nsecs_stats[cpu], count[0]);
 193         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 194                 update_stats(&runtime_cycles_stats[ctx][cpu], count[0]);
 195         else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
 196                 update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count[0]);
 197         else if (perf_stat_evsel__is(counter, TRANSACTION_START))
 198                 update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
 199         else if (perf_stat_evsel__is(counter, ELISION_START))
 200                 update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
 201         else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
 202                 update_stats(&runtime_topdown_total_slots[ctx][cpu], count[0]);
 203         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
 204                 update_stats(&runtime_topdown_slots_issued[ctx][cpu], count[0]);
 205         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
 206                 update_stats(&runtime_topdown_slots_retired[ctx][cpu], count[0]);
 207         else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
 208                 update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu],count[0]);
 209         else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
 210                 update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count[0]);
 211         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 212                 update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
 213         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 214                 update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count[0]);
 215         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 216                 update_stats(&runtime_branches_stats[ctx][cpu], count[0]);
 217         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 218                 update_stats(&runtime_cacherefs_stats[ctx][cpu], count[0]);
 219         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 220                 update_stats(&runtime_l1_dcache_stats[ctx][cpu], count[0]);
 221         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 222                 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
 223         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 224                 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
 225         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 226                 update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count[0]);
 227         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 228                 update_stats(&runtime_itlb_cache_stats[ctx][cpu], count[0]);
 229         else if (perf_stat_evsel__is(counter, SMI_NUM))
 230                 update_stats(&runtime_smi_num_stats[ctx][cpu], count[0]);
 231         else if (perf_stat_evsel__is(counter, APERF))
 232                 update_stats(&runtime_aperf_stats[ctx][cpu], count[0]);
 233
 234         if (counter->collect_stat) {
 235                 struct saved_value *v = saved_value_lookup(counter, cpu, ctx,
 236                                                            true);
 237                 update_stats(&v->stats, count[0]);
 238         }
 239 }
 240
 241 /* used for get_ratio_color() */
 242 enum grc_type {
 243         GRC_STALLED_CYCLES_FE,
 244         GRC_STALLED_CYCLES_BE,
 245         GRC_CACHE_MISSES,
 246         GRC_MAX_NR
 247 };
 248
 249 static const char *get_ratio_color(enum grc_type type, double ratio)
 250 {
 251         static const double grc_table[GRC_MAX_NR][3] = {
 252                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 253                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 254                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 255         };
 256         const char *color = PERF_COLOR_NORMAL;
 257
 258         if (ratio > grc_table[type][0])
 259                 color = PERF_COLOR_RED;
 260         else if (ratio > grc_table[type][1])
 261                 color = PERF_COLOR_MAGENTA;
 262         else if (ratio > grc_table[type][2])
 263                 color = PERF_COLOR_YELLOW;
 264
 265         return color;
 266 }
 267
 268 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
 269                                                 const char *name)
 270 {
 271         struct perf_evsel *c2;
 272
 273         evlist__for_each_entry (evsel_list, c2) {
 274                 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
 275                         return c2;
 276         }
 277         return NULL;
 278 }
 279
 280 /* Mark MetricExpr target events and link events using them to them. */
 281 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
 282 {
 283         struct perf_evsel *counter, *leader, **metric_events, *oc;
 284         bool found;
 285         const char **metric_names;
 286         int i;
 287         int num_metric_names;
 288
 289         evlist__for_each_entry(evsel_list, counter) {
 290                 bool invalid = false;
 291
 292                 leader = counter->leader;
 293                 if (!counter->metric_expr)
 294                         continue;
 295                 metric_events = counter->metric_events;
 296                 if (!metric_events) {
 297                         if (expr__find_other(counter->metric_expr, counter->name,
 298                                                 &metric_names, &num_metric_names) < 0)
 299                                 continue;
 300
 301                         metric_events = calloc(sizeof(struct perf_evsel *),
 302                                                num_metric_names + 1);
 303                         if (!metric_events)
 304                                 return;
 305                         counter->metric_events = metric_events;
 306                 }
 307
 308                 for (i = 0; i < num_metric_names; i++) {
 309                         found = false;
 310                         if (leader) {
 311                                 /* Search in group */
 312                                 for_each_group_member (oc, leader) {
 313                                         if (!strcasecmp(oc->name, metric_names[i]) &&
 314                                                 !oc->collect_stat) {
 315                                                 found = true;
 316                                                 break;
 317                                         }
 318                                 }
 319                         }
 320                         if (!found) {
 321                                 /* Search ignoring groups */
 322                                 oc = perf_stat__find_event(evsel_list, metric_names[i]);
 323                         }
 324                         if (!oc) {
 325                                 /* Deduping one is good enough to handle duplicated PMUs. */
 326                                 static char *printed;
 327
 328                                 /*
 329                                  * Adding events automatically would be difficult, because
 330                                  * it would risk creating groups that are not schedulable.
 331                                  * perf stat doesn't understand all the scheduling constraints
 332                                  * of events. So we ask the user instead to add the missing
 333                                  * events.
 334                                  */
 335                                 if (!printed || strcasecmp(printed, metric_names[i])) {
 336                                         fprintf(stderr,
 337                                                 "Add %s event to groups to get metric expression for %s\n",
 338                                                 metric_names[i],
 339                                                 counter->name);
 340                                         printed = strdup(metric_names[i]);
 341                                 }
 342                                 invalid = true;
 343                                 continue;
 344                         }
 345                         metric_events[i] = oc;
 346                         oc->collect_stat = true;
 347                 }
 348                 metric_events[i] = NULL;
 349                 free(metric_names);
 350                 if (invalid) {
 351                         free(metric_events);
 352                         counter->metric_events = NULL;
 353                         counter->metric_expr = NULL;
 354                 }
 355         }
 356 }
 357
 358 static void print_stalled_cycles_frontend(int cpu,
 359                                           struct perf_evsel *evsel, double avg,
 360                                           struct perf_stat_output_ctx *out)
 361 {
 362         double total, ratio = 0.0;
 363         const char *color;
 364         int ctx = evsel_context(evsel);
 365
 366         total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 367
 368         if (total)
 369                 ratio = avg / total * 100.0;
 370
 371         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 372
 373         if (ratio)
 374                 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
 375                                   ratio);
 376         else
 377                 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
 378 }
 379
 380 static void print_stalled_cycles_backend(int cpu,
 381                                          struct perf_evsel *evsel, double avg,
 382                                          struct perf_stat_output_ctx *out)
 383 {
 384         double total, ratio = 0.0;
 385         const char *color;
 386         int ctx = evsel_context(evsel);
 387
 388         total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 389
 390         if (total)
 391                 ratio = avg / total * 100.0;
 392
 393         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 394
 395         out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
 396 }
 397
 398 static void print_branch_misses(int cpu,
 399                                 struct perf_evsel *evsel,
 400                                 double avg,
 401                                 struct perf_stat_output_ctx *out)
 402 {
 403         double total, ratio = 0.0;
 404         const char *color;
 405         int ctx = evsel_context(evsel);
 406
 407         total = avg_stats(&runtime_branches_stats[ctx][cpu]);
 408
 409         if (total)
 410                 ratio = avg / total * 100.0;
 411
 412         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 413
 414         out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
 415 }
 416
 417 static void print_l1_dcache_misses(int cpu,
 418                                    struct perf_evsel *evsel,
 419                                    double avg,
 420                                    struct perf_stat_output_ctx *out)
 421 {
 422         double total, ratio = 0.0;
 423         const char *color;
 424         int ctx = evsel_context(evsel);
 425
 426         total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
 427
 428         if (total)
 429                 ratio = avg / total * 100.0;
 430
 431         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 432
 433         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
 434 }
 435
 436 static void print_l1_icache_misses(int cpu,
 437                                    struct perf_evsel *evsel,
 438                                    double avg,
 439                                    struct perf_stat_output_ctx *out)
 440 {
 441         double total, ratio = 0.0;
 442         const char *color;
 443         int ctx = evsel_context(evsel);
 444
 445         total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
 446
 447         if (total)
 448                 ratio = avg / total * 100.0;
 449
 450         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 451         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
 452 }
 453
 454 static void print_dtlb_cache_misses(int cpu,
 455                                     struct perf_evsel *evsel,
 456                                     double avg,
 457                                     struct perf_stat_output_ctx *out)
 458 {
 459         double total, ratio = 0.0;
 460         const char *color;
 461         int ctx = evsel_context(evsel);
 462
 463         total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
 464
 465         if (total)
 466                 ratio = avg / total * 100.0;
 467
 468         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 469         out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
 470 }
 471
 472 static void print_itlb_cache_misses(int cpu,
 473                                     struct perf_evsel *evsel,
 474                                     double avg,
 475                                     struct perf_stat_output_ctx *out)
 476 {
 477         double total, ratio = 0.0;
 478         const char *color;
 479         int ctx = evsel_context(evsel);
 480
 481         total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
 482
 483         if (total)
 484                 ratio = avg / total * 100.0;
 485
 486         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 487         out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
 488 }
 489
 490 static void print_ll_cache_misses(int cpu,
 491                                   struct perf_evsel *evsel,
 492                                   double avg,
 493                                   struct perf_stat_output_ctx *out)
 494 {
 495         double total, ratio = 0.0;
 496         const char *color;
 497         int ctx = evsel_context(evsel);
 498
 499         total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
 500
 501         if (total)
 502                 ratio = avg / total * 100.0;
 503
 504         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 505         out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
 506 }
 507
 508 /*
 509  * High level "TopDown" CPU core pipe line bottleneck break down.
 510  *
 511  * Basic concept following
 512  * Yasin, A Top Down Method for Performance analysis and Counter architecture
 513  * ISPASS14
 514  *
 515  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
 516  *
 517  * Frontend -> Backend -> Retiring
 518  * BadSpeculation in addition means out of order execution that is thrown away
 519  * (for example branch mispredictions)
 520  * Frontend is instruction decoding.
 521  * Backend is execution, like computation and accessing data in memory
 522  * Retiring is good execution that is not directly bottlenecked
 523  *
 524  * The formulas are computed in slots.
 525  * A slot is an entry in the pipeline each for the pipeline width
 526  * (for example a 4-wide pipeline has 4 slots for each cycle)
 527  *
 528  * Formulas:
 529  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
 530  *                      TotalSlots
 531  * Retiring = SlotsRetired / TotalSlots
 532  * FrontendBound = FetchBubbles / TotalSlots
 533  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
 534  *
 535  * The kernel provides the mapping to the low level CPU events and any scaling
 536  * needed for the CPU pipeline width, for example:
 537  *
 538  * TotalSlots = Cycles * 4
 539  *
 540  * The scaling factor is communicated in the sysfs unit.
 541  *
 542  * In some cases the CPU may not be able to measure all the formulas due to
 543  * missing events. In this case multiple formulas are combined, as possible.
 544  *
 545  * Full TopDown supports more levels to sub-divide each area: for example
 546  * BackendBound into computing bound and memory bound. For now we only
 547  * support Level 1 TopDown.
 548  */
 549
 550 static double sanitize_val(double x)
 551 {
 552         if (x < 0 && x >= -0.02)
 553                 return 0.0;
 554         return x;
 555 }
 556
 557 static double td_total_slots(int ctx, int cpu)
 558 {
 559         return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
 560 }
 561
 562 static double td_bad_spec(int ctx, int cpu)
 563 {
 564         double bad_spec = 0;
 565         double total_slots;
 566         double total;
 567
 568         total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
 569                 avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
 570                 avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
 571         total_slots = td_total_slots(ctx, cpu);
 572         if (total_slots)
 573                 bad_spec = total / total_slots;
 574         return sanitize_val(bad_spec);
 575 }
 576
 577 static double td_retiring(int ctx, int cpu)
 578 {
 579         double retiring = 0;
 580         double total_slots = td_total_slots(ctx, cpu);
 581         double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
 582
 583         if (total_slots)
 584                 retiring = ret_slots / total_slots;
 585         return retiring;
 586 }
 587
 588 static double td_fe_bound(int ctx, int cpu)
 589 {
 590         double fe_bound = 0;
 591         double total_slots = td_total_slots(ctx, cpu);
 592         double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
 593
 594         if (total_slots)
 595                 fe_bound = fetch_bub / total_slots;
 596         return fe_bound;
 597 }
 598
 599 static double td_be_bound(int ctx, int cpu)
 600 {
 601         double sum = (td_fe_bound(ctx, cpu) +
 602                       td_bad_spec(ctx, cpu) +
 603                       td_retiring(ctx, cpu));
 604         if (sum == 0)
 605                 return 0;
 606         return sanitize_val(1.0 - sum);
 607 }
 608
 609 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
 610                            struct perf_stat_output_ctx *out)
 611 {
 612         double smi_num, aperf, cycles, cost = 0.0;
 613         int ctx = evsel_context(evsel);
 614         const char *color = NULL;
 615
 616         smi_num = avg_stats(&runtime_smi_num_stats[ctx][cpu]);
 617         aperf = avg_stats(&runtime_aperf_stats[ctx][cpu]);
 618         cycles = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 619
 620         if ((cycles == 0) || (aperf == 0))
 621                 return;
 622
 623         if (smi_num)
 624                 cost = (aperf - cycles) / aperf * 100.00;
 625
 626         if (cost > 10)
 627                 color = PERF_COLOR_RED;
 628         out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
 629         out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
 630 }
 631
 632 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 633                                    double avg, int cpu,
 634                                    struct perf_stat_output_ctx *out)
 635 {
 636         void *ctxp = out->ctx;
 637         print_metric_t print_metric = out->print_metric;
 638         double total, ratio = 0.0, total2;
 639         const char *color = NULL;
 640         int ctx = evsel_context(evsel);
 641
 642         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 643                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 644                 if (total) {
 645                         ratio = avg / total;
 646                         print_metric(ctxp, NULL, "%7.2f ",
 647                                         "insn per cycle", ratio);
 648                 } else {
 649                         print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
 650                 }
 651                 total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
 652                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
 653
 654                 if (total && avg) {
 655                         out->new_line(ctxp);
 656                         ratio = total / avg;
 657                         print_metric(ctxp, NULL, "%7.2f ",
 658                                         "stalled cycles per insn",
 659                                         ratio);
 660                 } else if (have_frontend_stalled) {
 661                         print_metric(ctxp, NULL, NULL,
 662                                      "stalled cycles per insn", 0);
 663                 }
 664         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
 665                 if (runtime_branches_stats[ctx][cpu].n != 0)
 666                         print_branch_misses(cpu, evsel, avg, out);
 667                 else
 668                         print_metric(ctxp, NULL, NULL, "of all branches", 0);
 669         } else if (
 670                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 671                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 672                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 673                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
 674                 if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
 675                         print_l1_dcache_misses(cpu, evsel, avg, out);
 676                 else
 677                         print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
 678         } else if (
 679                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 680                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 681                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 682                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
 683                 if (runtime_l1_icache_stats[ctx][cpu].n != 0)
 684                         print_l1_icache_misses(cpu, evsel, avg, out);
 685                 else
 686                         print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
 687         } else if (
 688                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 689                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 690                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 691                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
 692                 if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
 693                         print_dtlb_cache_misses(cpu, evsel, avg, out);
 694                 else
 695                         print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
 696         } else if (
 697                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 698                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 699                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 700                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
 701                 if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
 702                         print_itlb_cache_misses(cpu, evsel, avg, out);
 703                 else
 704                         print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
 705         } else if (
 706                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 707                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 708                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 709                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
 710                 if (runtime_ll_cache_stats[ctx][cpu].n != 0)
 711                         print_ll_cache_misses(cpu, evsel, avg, out);
 712                 else
 713                         print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
 714         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
 715                 total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
 716
 717                 if (total)
 718                         ratio = avg * 100 / total;
 719
 720                 if (runtime_cacherefs_stats[ctx][cpu].n != 0)
 721                         print_metric(ctxp, NULL, "%8.3f %%",
 722                                      "of all cache refs", ratio);
 723                 else
 724                         print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
 725         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 726                 print_stalled_cycles_frontend(cpu, evsel, avg, out);
 727         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 728                 print_stalled_cycles_backend(cpu, evsel, avg, out);
 729         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 730                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 731
 732                 if (total) {
 733                         ratio = avg / total;
 734                         print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
 735                 } else {
 736                         print_metric(ctxp, NULL, NULL, "Ghz", 0);
 737                 }
 738         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
 739                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 740                 if (total)
 741                         print_metric(ctxp, NULL,
 742                                         "%7.2f%%", "transactional cycles",
 743                                         100.0 * (avg / total));
 744                 else
 745                         print_metric(ctxp, NULL, NULL, "transactional cycles",
 746                                      0);
 747         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
 748                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 749                 total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 750                 if (total2 < avg)
 751                         total2 = avg;
 752                 if (total)
 753                         print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
 754                                 100.0 * ((total2-avg) / total));
 755                 else
 756                         print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
 757         } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
 758                 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 759
 760                 if (avg)
 761                         ratio = total / avg;
 762
 763                 if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
 764                         print_metric(ctxp, NULL, "%8.0f",
 765                                      "cycles / transaction", ratio);
 766                 else
 767                         print_metric(ctxp, NULL, NULL, "cycles / transaction",
 768                                      0);
 769         } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
 770                 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 771
 772                 if (avg)
 773                         ratio = total / avg;
 774
 775                 print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
 776         } else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
 777                    perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
 778                 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
 779                         print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
 780                                      avg / ratio);
 781                 else
 782                         print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
 783         } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
 784                 double fe_bound = td_fe_bound(ctx, cpu);
 785
 786                 if (fe_bound > 0.2)
 787                         color = PERF_COLOR_RED;
 788                 print_metric(ctxp, color, "%8.1f%%", "frontend bound",
 789                                 fe_bound * 100.);
 790         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
 791                 double retiring = td_retiring(ctx, cpu);
 792
 793                 if (retiring > 0.7)
 794                         color = PERF_COLOR_GREEN;
 795                 print_metric(ctxp, color, "%8.1f%%", "retiring",
 796                                 retiring * 100.);
 797         } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
 798                 double bad_spec = td_bad_spec(ctx, cpu);
 799
 800                 if (bad_spec > 0.1)
 801                         color = PERF_COLOR_RED;
 802                 print_metric(ctxp, color, "%8.1f%%", "bad speculation",
 803                                 bad_spec * 100.);
 804         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
 805                 double be_bound = td_be_bound(ctx, cpu);
 806                 const char *name = "backend bound";
 807                 static int have_recovery_bubbles = -1;
 808
 809                 /* In case the CPU does not support topdown-recovery-bubbles */
 810                 if (have_recovery_bubbles < 0)
 811                         have_recovery_bubbles = pmu_have_event("cpu",
 812                                         "topdown-recovery-bubbles");
 813                 if (!have_recovery_bubbles)
 814                         name = "backend bound/bad spec";
 815
 816                 if (be_bound > 0.2)
 817                         color = PERF_COLOR_RED;
 818                 if (td_total_slots(ctx, cpu) > 0)
 819                         print_metric(ctxp, color, "%8.1f%%", name,
 820                                         be_bound * 100.);
 821                 else
 822                         print_metric(ctxp, NULL, NULL, name, 0);
 823         } else if (evsel->metric_expr) {
 824                 struct parse_ctx pctx;
 825                 int i;
 826
 827                 expr__ctx_init(&pctx);
 828                 expr__add_id(&pctx, evsel->name, avg);
 829                 for (i = 0; evsel->metric_events[i]; i++) {
 830                         struct saved_value *v;
 831
 832                         v = saved_value_lookup(evsel->metric_events[i], cpu, ctx, false);
 833                         if (!v)
 834                                 break;
 835                         expr__add_id(&pctx, evsel->metric_events[i]->name,
 836                                              avg_stats(&v->stats));
 837                 }
 838                 if (!evsel->metric_events[i]) {
 839                         const char *p = evsel->metric_expr;
 840
 841                         if (expr__parse(&ratio, &pctx, &p) == 0)
 842                                 print_metric(ctxp, NULL, "%8.1f",
 843                                         evsel->metric_name ?
 844                                         evsel->metric_name :
 845                                         out->force_header ?  evsel->name : "",
 846                                         ratio);
 847                         else
 848                                 print_metric(ctxp, NULL, NULL, "", 0);
 849                 } else
 850                         print_metric(ctxp, NULL, NULL, "", 0);
 851         } else if (runtime_nsecs_stats[cpu].n != 0) {
 852                 char unit = 'M';
 853                 char unit_buf[10];
 854
 855                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 856
 857                 if (total)
 858                         ratio = 1000.0 * avg / total;
 859                 if (ratio < 0.001) {
 860                         ratio *= 1000;
 861                         unit = 'K';
 862                 }
 863                 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
 864                 print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
 865         } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
 866                 print_smi_cost(cpu, evsel, out);
 867         } else {
 868                 print_metric(ctxp, NULL, NULL, NULL, 0);
 869         }
 870 }