tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/compiler.h>
  19 #include <linux/err.h>
  20 #include <linux/zalloc.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include <stdlib.h>
  26 #include <perf/evsel.h>
  27 #include "asm/bug.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "../perf-sys.h"
  49 #include "util/parse-branch-options.h"
  50 #include <internal/xyarray.h>
  51 #include <internal/lib.h>
  52
  53 #include <linux/ctype.h>
  54
  55 struct perf_missing_features perf_missing_features;
  56
  57 static clockid_t clockid;
  58
  59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  60 {
  61         return 0;
  62 }
  63
  64 void __weak test_attr__ready(void) { }
  65
  66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  67 {
  68 }
  69
  70 static struct {
  71         size_t  size;
  72         int     (*init)(struct evsel *evsel);
  73         void    (*fini)(struct evsel *evsel);
  74 } perf_evsel__object = {
  75         .size = sizeof(struct evsel),
  76         .init = evsel__no_extra_init,
  77         .fini = evsel__no_extra_fini,
  78 };
  79
  80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
  81                          void (*fini)(struct evsel *evsel))
  82 {
  83
  84         if (object_size == 0)
  85                 goto set_methods;
  86
  87         if (perf_evsel__object.size > object_size)
  88                 return -EINVAL;
  89
  90         perf_evsel__object.size = object_size;
  91
  92 set_methods:
  93         if (init != NULL)
  94                 perf_evsel__object.init = init;
  95
  96         if (fini != NULL)
  97                 perf_evsel__object.fini = fini;
  98
  99         return 0;
 100 }
 101
 102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 103
 104 int __evsel__sample_size(u64 sample_type)
 105 {
 106         u64 mask = sample_type & PERF_SAMPLE_MASK;
 107         int size = 0;
 108         int i;
 109
 110         for (i = 0; i < 64; i++) {
 111                 if (mask & (1ULL << i))
 112                         size++;
 113         }
 114
 115         size *= sizeof(u64);
 116
 117         return size;
 118 }
 119
 120 /**
 121  * __perf_evsel__calc_id_pos - calculate id_pos.
 122  * @sample_type: sample type
 123  *
 124  * This function returns the position of the event id (PERF_SAMPLE_ID or
 125  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 126  * perf_record_sample.
 127  */
 128 static int __perf_evsel__calc_id_pos(u64 sample_type)
 129 {
 130         int idx = 0;
 131
 132         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 133                 return 0;
 134
 135         if (!(sample_type & PERF_SAMPLE_ID))
 136                 return -1;
 137
 138         if (sample_type & PERF_SAMPLE_IP)
 139                 idx += 1;
 140
 141         if (sample_type & PERF_SAMPLE_TID)
 142                 idx += 1;
 143
 144         if (sample_type & PERF_SAMPLE_TIME)
 145                 idx += 1;
 146
 147         if (sample_type & PERF_SAMPLE_ADDR)
 148                 idx += 1;
 149
 150         return idx;
 151 }
 152
 153 /**
 154  * __perf_evsel__calc_is_pos - calculate is_pos.
 155  * @sample_type: sample type
 156  *
 157  * This function returns the position (counting backwards) of the event id
 158  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 159  * sample_id_all is used there is an id sample appended to non-sample events.
 160  */
 161 static int __perf_evsel__calc_is_pos(u64 sample_type)
 162 {
 163         int idx = 1;
 164
 165         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 166                 return 1;
 167
 168         if (!(sample_type & PERF_SAMPLE_ID))
 169                 return -1;
 170
 171         if (sample_type & PERF_SAMPLE_CPU)
 172                 idx += 1;
 173
 174         if (sample_type & PERF_SAMPLE_STREAM_ID)
 175                 idx += 1;
 176
 177         return idx;
 178 }
 179
 180 void evsel__calc_id_pos(struct evsel *evsel)
 181 {
 182         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 183         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 184 }
 185
 186 void __evsel__set_sample_bit(struct evsel *evsel,
 187                                   enum perf_event_sample_format bit)
 188 {
 189         if (!(evsel->core.attr.sample_type & bit)) {
 190                 evsel->core.attr.sample_type |= bit;
 191                 evsel->sample_size += sizeof(u64);
 192                 evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void __evsel__reset_sample_bit(struct evsel *evsel,
 197                                     enum perf_event_sample_format bit)
 198 {
 199         if (evsel->core.attr.sample_type & bit) {
 200                 evsel->core.attr.sample_type &= ~bit;
 201                 evsel->sample_size -= sizeof(u64);
 202                 evsel__calc_id_pos(evsel);
 203         }
 204 }
 205
 206 void evsel__set_sample_id(struct evsel *evsel,
 207                                bool can_sample_identifier)
 208 {
 209         if (can_sample_identifier) {
 210                 evsel__reset_sample_bit(evsel, ID);
 211                 evsel__set_sample_bit(evsel, IDENTIFIER);
 212         } else {
 213                 evsel__set_sample_bit(evsel, ID);
 214         }
 215         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 216 }
 217
 218 /**
 219  * evsel__is_function_event - Return whether given evsel is a function
 220  * trace event
 221  *
 222  * @evsel - evsel selector to be tested
 223  *
 224  * Return %true if event is function trace event
 225  */
 226 bool evsel__is_function_event(struct evsel *evsel)
 227 {
 228 #define FUNCTION_EVENT "ftrace:function"
 229
 230         return evsel->name &&
 231                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 232
 233 #undef FUNCTION_EVENT
 234 }
 235
 236 void evsel__init(struct evsel *evsel,
 237                  struct perf_event_attr *attr, int idx)
 238 {
 239         perf_evsel__init(&evsel->core, attr);
 240         evsel->idx         = idx;
 241         evsel->tracking    = !idx;
 242         evsel->leader      = evsel;
 243         evsel->unit        = "";
 244         evsel->scale       = 1.0;
 245         evsel->max_events  = ULONG_MAX;
 246         evsel->evlist      = NULL;
 247         evsel->bpf_obj     = NULL;
 248         evsel->bpf_fd      = -1;
 249         INIT_LIST_HEAD(&evsel->config_terms);
 250         perf_evsel__object.init(evsel);
 251         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 252         evsel__calc_id_pos(evsel);
 253         evsel->cmdline_group_boundary = false;
 254         evsel->metric_expr   = NULL;
 255         evsel->metric_name   = NULL;
 256         evsel->metric_events = NULL;
 257         evsel->per_pkg_mask  = NULL;
 258         evsel->collect_stat  = false;
 259         evsel->pmu_name      = NULL;
 260 }
 261
 262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 263 {
 264         struct evsel *evsel = zalloc(perf_evsel__object.size);
 265
 266         if (!evsel)
 267                 return NULL;
 268         evsel__init(evsel, attr, idx);
 269
 270         if (evsel__is_bpf_output(evsel)) {
 271                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 272                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 273                 evsel->core.attr.sample_period = 1;
 274         }
 275
 276         if (evsel__is_clock(evsel)) {
 277                 /*
 278                  * The evsel->unit points to static alias->unit
 279                  * so it's ok to use static string in here.
 280                  */
 281                 static const char *unit = "msec";
 282
 283                 evsel->unit = unit;
 284                 evsel->scale = 1e-6;
 285         }
 286
 287         return evsel;
 288 }
 289
 290 static bool perf_event_can_profile_kernel(void)
 291 {
 292         return perf_event_paranoid_check(1);
 293 }
 294
 295 struct evsel *evsel__new_cycles(bool precise)
 296 {
 297         struct perf_event_attr attr = {
 298                 .type   = PERF_TYPE_HARDWARE,
 299                 .config = PERF_COUNT_HW_CPU_CYCLES,
 300                 .exclude_kernel = !perf_event_can_profile_kernel(),
 301         };
 302         struct evsel *evsel;
 303
 304         event_attr_init(&attr);
 305
 306         if (!precise)
 307                 goto new_event;
 308
 309         /*
 310          * Now let the usual logic to set up the perf_event_attr defaults
 311          * to kick in when we return and before perf_evsel__open() is called.
 312          */
 313 new_event:
 314         evsel = evsel__new(&attr);
 315         if (evsel == NULL)
 316                 goto out;
 317
 318         evsel->precise_max = true;
 319
 320         /* use asprintf() because free(evsel) assumes name is allocated */
 321         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 322                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 323                      attr.exclude_kernel ? "u" : "",
 324                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 325                 goto error_free;
 326 out:
 327         return evsel;
 328 error_free:
 329         evsel__delete(evsel);
 330         evsel = NULL;
 331         goto out;
 332 }
 333
 334 /*
 335  * Returns pointer with encoded error via <linux/err.h> interface.
 336  */
 337 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 338 {
 339         struct evsel *evsel = zalloc(perf_evsel__object.size);
 340         int err = -ENOMEM;
 341
 342         if (evsel == NULL) {
 343                 goto out_err;
 344         } else {
 345                 struct perf_event_attr attr = {
 346                         .type          = PERF_TYPE_TRACEPOINT,
 347                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 348                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 349                 };
 350
 351                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 352                         goto out_free;
 353
 354                 evsel->tp_format = trace_event__tp_format(sys, name);
 355                 if (IS_ERR(evsel->tp_format)) {
 356                         err = PTR_ERR(evsel->tp_format);
 357                         goto out_free;
 358                 }
 359
 360                 event_attr_init(&attr);
 361                 attr.config = evsel->tp_format->id;
 362                 attr.sample_period = 1;
 363                 evsel__init(evsel, &attr, idx);
 364         }
 365
 366         return evsel;
 367
 368 out_free:
 369         zfree(&evsel->name);
 370         free(evsel);
 371 out_err:
 372         return ERR_PTR(err);
 373 }
 374
 375 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
 376         "cycles",
 377         "instructions",
 378         "cache-references",
 379         "cache-misses",
 380         "branches",
 381         "branch-misses",
 382         "bus-cycles",
 383         "stalled-cycles-frontend",
 384         "stalled-cycles-backend",
 385         "ref-cycles",
 386 };
 387
 388 static const char *__evsel__hw_name(u64 config)
 389 {
 390         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 391                 return evsel__hw_names[config];
 392
 393         return "unknown-hardware";
 394 }
 395
 396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 397 {
 398         int colon = 0, r = 0;
 399         struct perf_event_attr *attr = &evsel->core.attr;
 400         bool exclude_guest_default = false;
 401
 402 #define MOD_PRINT(context, mod) do {                                    \
 403                 if (!attr->exclude_##context) {                         \
 404                         if (!colon) colon = ++r;                        \
 405                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 406                 } } while(0)
 407
 408         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 409                 MOD_PRINT(kernel, 'k');
 410                 MOD_PRINT(user, 'u');
 411                 MOD_PRINT(hv, 'h');
 412                 exclude_guest_default = true;
 413         }
 414
 415         if (attr->precise_ip) {
 416                 if (!colon)
 417                         colon = ++r;
 418                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 419                 exclude_guest_default = true;
 420         }
 421
 422         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 423                 MOD_PRINT(host, 'H');
 424                 MOD_PRINT(guest, 'G');
 425         }
 426 #undef MOD_PRINT
 427         if (colon)
 428                 bf[colon - 1] = ':';
 429         return r;
 430 }
 431
 432 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 433 {
 434         int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 435         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 436 }
 437
 438 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
 439         "cpu-clock",
 440         "task-clock",
 441         "page-faults",
 442         "context-switches",
 443         "cpu-migrations",
 444         "minor-faults",
 445         "major-faults",
 446         "alignment-faults",
 447         "emulation-faults",
 448         "dummy",
 449 };
 450
 451 static const char *__evsel__sw_name(u64 config)
 452 {
 453         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 454                 return evsel__sw_names[config];
 455         return "unknown-software";
 456 }
 457
 458 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 459 {
 460         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 461         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 462 }
 463
 464 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 465 {
 466         int r;
 467
 468         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 469
 470         if (type & HW_BREAKPOINT_R)
 471                 r += scnprintf(bf + r, size - r, "r");
 472
 473         if (type & HW_BREAKPOINT_W)
 474                 r += scnprintf(bf + r, size - r, "w");
 475
 476         if (type & HW_BREAKPOINT_X)
 477                 r += scnprintf(bf + r, size - r, "x");
 478
 479         return r;
 480 }
 481
 482 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 483 {
 484         struct perf_event_attr *attr = &evsel->core.attr;
 485         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 486         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 487 }
 488
 489 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 490  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 491  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 492  { "LLC",       "L2",                                                   },
 493  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 494  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 495  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 496  { "node",                                                              },
 497 };
 498
 499 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 500  { "load",      "loads",        "read",                                 },
 501  { "store",     "stores",       "write",                                },
 502  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 503 };
 504
 505 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 506  { "refs",      "Reference",    "ops",          "access",               },
 507  { "misses",    "miss",                                                 },
 508 };
 509
 510 #define C(x)            PERF_COUNT_HW_CACHE_##x
 511 #define CACHE_READ      (1 << C(OP_READ))
 512 #define CACHE_WRITE     (1 << C(OP_WRITE))
 513 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 514 #define COP(x)          (1 << x)
 515
 516 /*
 517  * cache operartion stat
 518  * L1I : Read and prefetch only
 519  * ITLB and BPU : Read-only
 520  */
 521 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
 522  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 523  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 524  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 526  [C(ITLB)]      = (CACHE_READ),
 527  [C(BPU)]       = (CACHE_READ),
 528  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 529 };
 530
 531 bool evsel__is_cache_op_valid(u8 type, u8 op)
 532 {
 533         if (evsel__hw_cache_stat[type] & COP(op))
 534                 return true;    /* valid */
 535         else
 536                 return false;   /* invalid */
 537 }
 538
 539 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 540 {
 541         if (result) {
 542                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 543                                  evsel__hw_cache_op[op][0],
 544                                  evsel__hw_cache_result[result][0]);
 545         }
 546
 547         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 548                          evsel__hw_cache_op[op][1]);
 549 }
 550
 551 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 552 {
 553         u8 op, result, type = (config >>  0) & 0xff;
 554         const char *err = "unknown-ext-hardware-cache-type";
 555
 556         if (type >= PERF_COUNT_HW_CACHE_MAX)
 557                 goto out_err;
 558
 559         op = (config >>  8) & 0xff;
 560         err = "unknown-ext-hardware-cache-op";
 561         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 562                 goto out_err;
 563
 564         result = (config >> 16) & 0xff;
 565         err = "unknown-ext-hardware-cache-result";
 566         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 567                 goto out_err;
 568
 569         err = "invalid-cache";
 570         if (!evsel__is_cache_op_valid(type, op))
 571                 goto out_err;
 572
 573         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 574 out_err:
 575         return scnprintf(bf, size, "%s", err);
 576 }
 577
 578 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 579 {
 580         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 581         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 582 }
 583
 584 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 585 {
 586         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 587         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 588 }
 589
 590 static int evsel__tool_name(char *bf, size_t size)
 591 {
 592         int ret = scnprintf(bf, size, "duration_time");
 593         return ret;
 594 }
 595
 596 const char *evsel__name(struct evsel *evsel)
 597 {
 598         char bf[128];
 599
 600         if (!evsel)
 601                 goto out_unknown;
 602
 603         if (evsel->name)
 604                 return evsel->name;
 605
 606         switch (evsel->core.attr.type) {
 607         case PERF_TYPE_RAW:
 608                 evsel__raw_name(evsel, bf, sizeof(bf));
 609                 break;
 610
 611         case PERF_TYPE_HARDWARE:
 612                 evsel__hw_name(evsel, bf, sizeof(bf));
 613                 break;
 614
 615         case PERF_TYPE_HW_CACHE:
 616                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 617                 break;
 618
 619         case PERF_TYPE_SOFTWARE:
 620                 if (evsel->tool_event)
 621                         evsel__tool_name(bf, sizeof(bf));
 622                 else
 623                         evsel__sw_name(evsel, bf, sizeof(bf));
 624                 break;
 625
 626         case PERF_TYPE_TRACEPOINT:
 627                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 628                 break;
 629
 630         case PERF_TYPE_BREAKPOINT:
 631                 evsel__bp_name(evsel, bf, sizeof(bf));
 632                 break;
 633
 634         default:
 635                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 636                           evsel->core.attr.type);
 637                 break;
 638         }
 639
 640         evsel->name = strdup(bf);
 641
 642         if (evsel->name)
 643                 return evsel->name;
 644 out_unknown:
 645         return "unknown";
 646 }
 647
 648 const char *evsel__group_name(struct evsel *evsel)
 649 {
 650         return evsel->group_name ?: "anon group";
 651 }
 652
 653 /*
 654  * Returns the group details for the specified leader,
 655  * with following rules.
 656  *
 657  *  For record -e '{cycles,instructions}'
 658  *    'anon group { cycles:u, instructions:u }'
 659  *
 660  *  For record -e 'cycles,instructions' and report --group
 661  *    'cycles:u, instructions:u'
 662  */
 663 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 664 {
 665         int ret = 0;
 666         struct evsel *pos;
 667         const char *group_name = evsel__group_name(evsel);
 668
 669         if (!evsel->forced_leader)
 670                 ret = scnprintf(buf, size, "%s { ", group_name);
 671
 672         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 673
 674         for_each_group_member(pos, evsel)
 675                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 676
 677         if (!evsel->forced_leader)
 678                 ret += scnprintf(buf + ret, size - ret, " }");
 679
 680         return ret;
 681 }
 682
 683 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 684                                       struct callchain_param *param)
 685 {
 686         bool function = evsel__is_function_event(evsel);
 687         struct perf_event_attr *attr = &evsel->core.attr;
 688
 689         evsel__set_sample_bit(evsel, CALLCHAIN);
 690
 691         attr->sample_max_stack = param->max_stack;
 692
 693         if (opts->kernel_callchains)
 694                 attr->exclude_callchain_user = 1;
 695         if (opts->user_callchains)
 696                 attr->exclude_callchain_kernel = 1;
 697         if (param->record_mode == CALLCHAIN_LBR) {
 698                 if (!opts->branch_stack) {
 699                         if (attr->exclude_user) {
 700                                 pr_warning("LBR callstack option is only available "
 701                                            "to get user callchain information. "
 702                                            "Falling back to framepointers.\n");
 703                         } else {
 704                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 705                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 706                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 707                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 708                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 709                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 710                         }
 711                 } else
 712                          pr_warning("Cannot use LBR callstack with branch stack. "
 713                                     "Falling back to framepointers.\n");
 714         }
 715
 716         if (param->record_mode == CALLCHAIN_DWARF) {
 717                 if (!function) {
 718                         evsel__set_sample_bit(evsel, REGS_USER);
 719                         evsel__set_sample_bit(evsel, STACK_USER);
 720                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 721                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 722                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 723                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 724                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 725                         } else {
 726                                 attr->sample_regs_user |= PERF_REGS_MASK;
 727                         }
 728                         attr->sample_stack_user = param->dump_size;
 729                         attr->exclude_callchain_user = 1;
 730                 } else {
 731                         pr_info("Cannot use DWARF unwind for function trace event,"
 732                                 " falling back to framepointers.\n");
 733                 }
 734         }
 735
 736         if (function) {
 737                 pr_info("Disabling user space callchains for function trace event.\n");
 738                 attr->exclude_callchain_user = 1;
 739         }
 740 }
 741
 742 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 743                              struct callchain_param *param)
 744 {
 745         if (param->enabled)
 746                 return __evsel__config_callchain(evsel, opts, param);
 747 }
 748
 749 static void
 750 perf_evsel__reset_callgraph(struct evsel *evsel,
 751                             struct callchain_param *param)
 752 {
 753         struct perf_event_attr *attr = &evsel->core.attr;
 754
 755         evsel__reset_sample_bit(evsel, CALLCHAIN);
 756         if (param->record_mode == CALLCHAIN_LBR) {
 757                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 758                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 759                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 760                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 761         }
 762         if (param->record_mode == CALLCHAIN_DWARF) {
 763                 evsel__reset_sample_bit(evsel, REGS_USER);
 764                 evsel__reset_sample_bit(evsel, STACK_USER);
 765         }
 766 }
 767
 768 static void evsel__apply_config_terms(struct evsel *evsel,
 769                                       struct record_opts *opts, bool track)
 770 {
 771         struct evsel_config_term *term;
 772         struct list_head *config_terms = &evsel->config_terms;
 773         struct perf_event_attr *attr = &evsel->core.attr;
 774         /* callgraph default */
 775         struct callchain_param param = {
 776                 .record_mode = callchain_param.record_mode,
 777         };
 778         u32 dump_size = 0;
 779         int max_stack = 0;
 780         const char *callgraph_buf = NULL;
 781
 782         list_for_each_entry(term, config_terms, list) {
 783                 switch (term->type) {
 784                 case EVSEL__CONFIG_TERM_PERIOD:
 785                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 786                                 attr->sample_period = term->val.period;
 787                                 attr->freq = 0;
 788                                 evsel__reset_sample_bit(evsel, PERIOD);
 789                         }
 790                         break;
 791                 case EVSEL__CONFIG_TERM_FREQ:
 792                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 793                                 attr->sample_freq = term->val.freq;
 794                                 attr->freq = 1;
 795                                 evsel__set_sample_bit(evsel, PERIOD);
 796                         }
 797                         break;
 798                 case EVSEL__CONFIG_TERM_TIME:
 799                         if (term->val.time)
 800                                 evsel__set_sample_bit(evsel, TIME);
 801                         else
 802                                 evsel__reset_sample_bit(evsel, TIME);
 803                         break;
 804                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 805                         callgraph_buf = term->val.str;
 806                         break;
 807                 case EVSEL__CONFIG_TERM_BRANCH:
 808                         if (term->val.str && strcmp(term->val.str, "no")) {
 809                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 810                                 parse_branch_str(term->val.str,
 811                                                  &attr->branch_sample_type);
 812                         } else
 813                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 814                         break;
 815                 case EVSEL__CONFIG_TERM_STACK_USER:
 816                         dump_size = term->val.stack_user;
 817                         break;
 818                 case EVSEL__CONFIG_TERM_MAX_STACK:
 819                         max_stack = term->val.max_stack;
 820                         break;
 821                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 822                         evsel->max_events = term->val.max_events;
 823                         break;
 824                 case EVSEL__CONFIG_TERM_INHERIT:
 825                         /*
 826                          * attr->inherit should has already been set by
 827                          * evsel__config. If user explicitly set
 828                          * inherit using config terms, override global
 829                          * opt->no_inherit setting.
 830                          */
 831                         attr->inherit = term->val.inherit ? 1 : 0;
 832                         break;
 833                 case EVSEL__CONFIG_TERM_OVERWRITE:
 834                         attr->write_backward = term->val.overwrite ? 1 : 0;
 835                         break;
 836                 case EVSEL__CONFIG_TERM_DRV_CFG:
 837                         break;
 838                 case EVSEL__CONFIG_TERM_PERCORE:
 839                         break;
 840                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
 841                         attr->aux_output = term->val.aux_output ? 1 : 0;
 842                         break;
 843                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 844                         /* Already applied by auxtrace */
 845                         break;
 846                 case EVSEL__CONFIG_TERM_CFG_CHG:
 847                         break;
 848                 default:
 849                         break;
 850                 }
 851         }
 852
 853         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 854         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 855                 bool sample_address = false;
 856
 857                 if (max_stack) {
 858                         param.max_stack = max_stack;
 859                         if (callgraph_buf == NULL)
 860                                 callgraph_buf = "fp";
 861                 }
 862
 863                 /* parse callgraph parameters */
 864                 if (callgraph_buf != NULL) {
 865                         if (!strcmp(callgraph_buf, "no")) {
 866                                 param.enabled = false;
 867                                 param.record_mode = CALLCHAIN_NONE;
 868                         } else {
 869                                 param.enabled = true;
 870                                 if (parse_callchain_record(callgraph_buf, &param)) {
 871                                         pr_err("per-event callgraph setting for %s failed. "
 872                                                "Apply callgraph global setting for it\n",
 873                                                evsel->name);
 874                                         return;
 875                                 }
 876                                 if (param.record_mode == CALLCHAIN_DWARF)
 877                                         sample_address = true;
 878                         }
 879                 }
 880                 if (dump_size > 0) {
 881                         dump_size = round_up(dump_size, sizeof(u64));
 882                         param.dump_size = dump_size;
 883                 }
 884
 885                 /* If global callgraph set, clear it */
 886                 if (callchain_param.enabled)
 887                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 888
 889                 /* set perf-event callgraph */
 890                 if (param.enabled) {
 891                         if (sample_address) {
 892                                 evsel__set_sample_bit(evsel, ADDR);
 893                                 evsel__set_sample_bit(evsel, DATA_SRC);
 894                                 evsel->core.attr.mmap_data = track;
 895                         }
 896                         evsel__config_callchain(evsel, opts, &param);
 897                 }
 898         }
 899 }
 900
 901 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
 902 {
 903         struct evsel_config_term *term, *found_term = NULL;
 904
 905         list_for_each_entry(term, &evsel->config_terms, list) {
 906                 if (term->type == type)
 907                         found_term = term;
 908         }
 909
 910         return found_term;
 911 }
 912
 913 /*
 914  * The enable_on_exec/disabled value strategy:
 915  *
 916  *  1) For any type of traced program:
 917  *    - all independent events and group leaders are disabled
 918  *    - all group members are enabled
 919  *
 920  *     Group members are ruled by group leaders. They need to
 921  *     be enabled, because the group scheduling relies on that.
 922  *
 923  *  2) For traced programs executed by perf:
 924  *     - all independent events and group leaders have
 925  *       enable_on_exec set
 926  *     - we don't specifically enable or disable any event during
 927  *       the record command
 928  *
 929  *     Independent events and group leaders are initially disabled
 930  *     and get enabled by exec. Group members are ruled by group
 931  *     leaders as stated in 1).
 932  *
 933  *  3) For traced programs attached by perf (pid/tid):
 934  *     - we specifically enable or disable all events during
 935  *       the record command
 936  *
 937  *     When attaching events to already running traced we
 938  *     enable/disable events specifically, as there's no
 939  *     initial traced exec call.
 940  */
 941 void evsel__config(struct evsel *evsel, struct record_opts *opts,
 942                    struct callchain_param *callchain)
 943 {
 944         struct evsel *leader = evsel->leader;
 945         struct perf_event_attr *attr = &evsel->core.attr;
 946         int track = evsel->tracking;
 947         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 948
 949         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 950         attr->inherit       = !opts->no_inherit;
 951         attr->write_backward = opts->overwrite ? 1 : 0;
 952
 953         evsel__set_sample_bit(evsel, IP);
 954         evsel__set_sample_bit(evsel, TID);
 955
 956         if (evsel->sample_read) {
 957                 evsel__set_sample_bit(evsel, READ);
 958
 959                 /*
 960                  * We need ID even in case of single event, because
 961                  * PERF_SAMPLE_READ process ID specific data.
 962                  */
 963                 evsel__set_sample_id(evsel, false);
 964
 965                 /*
 966                  * Apply group format only if we belong to group
 967                  * with more than one members.
 968                  */
 969                 if (leader->core.nr_members > 1) {
 970                         attr->read_format |= PERF_FORMAT_GROUP;
 971                         attr->inherit = 0;
 972                 }
 973         }
 974
 975         /*
 976          * We default some events to have a default interval. But keep
 977          * it a weak assumption overridable by the user.
 978          */
 979         if (!attr->sample_period) {
 980                 if (opts->freq) {
 981                         attr->freq              = 1;
 982                         attr->sample_freq       = opts->freq;
 983                 } else {
 984                         attr->sample_period = opts->default_interval;
 985                 }
 986         }
 987         /*
 988          * If attr->freq was set (here or earlier), ask for period
 989          * to be sampled.
 990          */
 991         if (attr->freq)
 992                 evsel__set_sample_bit(evsel, PERIOD);
 993
 994         if (opts->no_samples)
 995                 attr->sample_freq = 0;
 996
 997         if (opts->inherit_stat) {
 998                 evsel->core.attr.read_format |=
 999                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1000                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1001                         PERF_FORMAT_ID;
1002                 attr->inherit_stat = 1;
1003         }
1004
1005         if (opts->sample_address) {
1006                 evsel__set_sample_bit(evsel, ADDR);
1007                 attr->mmap_data = track;
1008         }
1009
1010         /*
1011          * We don't allow user space callchains for  function trace
1012          * event, due to issues with page faults while tracing page
1013          * fault handler and its overall trickiness nature.
1014          */
1015         if (evsel__is_function_event(evsel))
1016                 evsel->core.attr.exclude_callchain_user = 1;
1017
1018         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1019                 evsel__config_callchain(evsel, opts, callchain);
1020
1021         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1022             !evsel__is_dummy_event(evsel)) {
1023                 attr->sample_regs_intr = opts->sample_intr_regs;
1024                 evsel__set_sample_bit(evsel, REGS_INTR);
1025         }
1026
1027         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1028             !evsel__is_dummy_event(evsel)) {
1029                 attr->sample_regs_user |= opts->sample_user_regs;
1030                 evsel__set_sample_bit(evsel, REGS_USER);
1031         }
1032
1033         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1034                 evsel__set_sample_bit(evsel, CPU);
1035
1036         /*
1037          * When the user explicitly disabled time don't force it here.
1038          */
1039         if (opts->sample_time &&
1040             (!perf_missing_features.sample_id_all &&
1041             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1042              opts->sample_time_set)))
1043                 evsel__set_sample_bit(evsel, TIME);
1044
1045         if (opts->raw_samples && !evsel->no_aux_samples) {
1046                 evsel__set_sample_bit(evsel, TIME);
1047                 evsel__set_sample_bit(evsel, RAW);
1048                 evsel__set_sample_bit(evsel, CPU);
1049         }
1050
1051         if (opts->sample_address)
1052                 evsel__set_sample_bit(evsel, DATA_SRC);
1053
1054         if (opts->sample_phys_addr)
1055                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1056
1057         if (opts->no_buffering) {
1058                 attr->watermark = 0;
1059                 attr->wakeup_events = 1;
1060         }
1061         if (opts->branch_stack && !evsel->no_aux_samples) {
1062                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1063                 attr->branch_sample_type = opts->branch_stack;
1064         }
1065
1066         if (opts->sample_weight)
1067                 evsel__set_sample_bit(evsel, WEIGHT);
1068
1069         attr->task  = track;
1070         attr->mmap  = track;
1071         attr->mmap2 = track && !perf_missing_features.mmap2;
1072         attr->comm  = track;
1073         /*
1074          * ksymbol is tracked separately with text poke because it needs to be
1075          * system wide and enabled immediately.
1076          */
1077         if (!opts->text_poke)
1078                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1079         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1080
1081         if (opts->record_namespaces)
1082                 attr->namespaces  = track;
1083
1084         if (opts->record_cgroup) {
1085                 attr->cgroup = track && !perf_missing_features.cgroup;
1086                 evsel__set_sample_bit(evsel, CGROUP);
1087         }
1088
1089         if (opts->record_switch_events)
1090                 attr->context_switch = track;
1091
1092         if (opts->sample_transaction)
1093                 evsel__set_sample_bit(evsel, TRANSACTION);
1094
1095         if (opts->running_time) {
1096                 evsel->core.attr.read_format |=
1097                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1098                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1099         }
1100
1101         /*
1102          * XXX see the function comment above
1103          *
1104          * Disabling only independent events or group leaders,
1105          * keeping group members enabled.
1106          */
1107         if (evsel__is_group_leader(evsel))
1108                 attr->disabled = 1;
1109
1110         /*
1111          * Setting enable_on_exec for independent events and
1112          * group leaders for traced executed by perf.
1113          */
1114         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1115             !opts->initial_delay)
1116                 attr->enable_on_exec = 1;
1117
1118         if (evsel->immediate) {
1119                 attr->disabled = 0;
1120                 attr->enable_on_exec = 0;
1121         }
1122
1123         clockid = opts->clockid;
1124         if (opts->use_clockid) {
1125                 attr->use_clockid = 1;
1126                 attr->clockid = opts->clockid;
1127         }
1128
1129         if (evsel->precise_max)
1130                 attr->precise_ip = 3;
1131
1132         if (opts->all_user) {
1133                 attr->exclude_kernel = 1;
1134                 attr->exclude_user   = 0;
1135         }
1136
1137         if (opts->all_kernel) {
1138                 attr->exclude_kernel = 0;
1139                 attr->exclude_user   = 1;
1140         }
1141
1142         if (evsel->core.own_cpus || evsel->unit)
1143                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1144
1145         /*
1146          * Apply event specific term settings,
1147          * it overloads any global configuration.
1148          */
1149         evsel__apply_config_terms(evsel, opts, track);
1150
1151         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1152
1153         /* The --period option takes the precedence. */
1154         if (opts->period_set) {
1155                 if (opts->period)
1156                         evsel__set_sample_bit(evsel, PERIOD);
1157                 else
1158                         evsel__reset_sample_bit(evsel, PERIOD);
1159         }
1160
1161         /*
1162          * A dummy event never triggers any actual counter and therefore
1163          * cannot be used with branch_stack.
1164          *
1165          * For initial_delay, a dummy event is added implicitly.
1166          * The software event will trigger -EOPNOTSUPP error out,
1167          * if BRANCH_STACK bit is set.
1168          */
1169         if (evsel__is_dummy_event(evsel))
1170                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1171 }
1172
1173 int evsel__set_filter(struct evsel *evsel, const char *filter)
1174 {
1175         char *new_filter = strdup(filter);
1176
1177         if (new_filter != NULL) {
1178                 free(evsel->filter);
1179                 evsel->filter = new_filter;
1180                 return 0;
1181         }
1182
1183         return -1;
1184 }
1185
1186 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1187 {
1188         char *new_filter;
1189
1190         if (evsel->filter == NULL)
1191                 return evsel__set_filter(evsel, filter);
1192
1193         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1194                 free(evsel->filter);
1195                 evsel->filter = new_filter;
1196                 return 0;
1197         }
1198
1199         return -1;
1200 }
1201
1202 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1203 {
1204         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1205 }
1206
1207 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1208 {
1209         return evsel__append_filter(evsel, "%s,%s", filter);
1210 }
1211
1212 /* Caller has to clear disabled after going through all CPUs. */
1213 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1214 {
1215         return perf_evsel__enable_cpu(&evsel->core, cpu);
1216 }
1217
1218 int evsel__enable(struct evsel *evsel)
1219 {
1220         int err = perf_evsel__enable(&evsel->core);
1221
1222         if (!err)
1223                 evsel->disabled = false;
1224         return err;
1225 }
1226
1227 /* Caller has to set disabled after going through all CPUs. */
1228 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1229 {
1230         return perf_evsel__disable_cpu(&evsel->core, cpu);
1231 }
1232
1233 int evsel__disable(struct evsel *evsel)
1234 {
1235         int err = perf_evsel__disable(&evsel->core);
1236         /*
1237          * We mark it disabled here so that tools that disable a event can
1238          * ignore events after they disable it. I.e. the ring buffer may have
1239          * already a few more events queued up before the kernel got the stop
1240          * request.
1241          */
1242         if (!err)
1243                 evsel->disabled = true;
1244
1245         return err;
1246 }
1247
1248 static void evsel__free_config_terms(struct evsel *evsel)
1249 {
1250         struct evsel_config_term *term, *h;
1251
1252         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1253                 list_del_init(&term->list);
1254                 if (term->free_str)
1255                         zfree(&term->val.str);
1256                 free(term);
1257         }
1258 }
1259
1260 void evsel__exit(struct evsel *evsel)
1261 {
1262         assert(list_empty(&evsel->core.node));
1263         assert(evsel->evlist == NULL);
1264         evsel__free_counts(evsel);
1265         perf_evsel__free_fd(&evsel->core);
1266         perf_evsel__free_id(&evsel->core);
1267         evsel__free_config_terms(evsel);
1268         cgroup__put(evsel->cgrp);
1269         perf_cpu_map__put(evsel->core.cpus);
1270         perf_cpu_map__put(evsel->core.own_cpus);
1271         perf_thread_map__put(evsel->core.threads);
1272         zfree(&evsel->group_name);
1273         zfree(&evsel->name);
1274         zfree(&evsel->pmu_name);
1275         zfree(&evsel->per_pkg_mask);
1276         zfree(&evsel->metric_events);
1277         perf_evsel__object.fini(evsel);
1278 }
1279
1280 void evsel__delete(struct evsel *evsel)
1281 {
1282         evsel__exit(evsel);
1283         free(evsel);
1284 }
1285
1286 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1287                            struct perf_counts_values *count)
1288 {
1289         struct perf_counts_values tmp;
1290
1291         if (!evsel->prev_raw_counts)
1292                 return;
1293
1294         if (cpu == -1) {
1295                 tmp = evsel->prev_raw_counts->aggr;
1296                 evsel->prev_raw_counts->aggr = *count;
1297         } else {
1298                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1299                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1300         }
1301
1302         count->val = count->val - tmp.val;
1303         count->ena = count->ena - tmp.ena;
1304         count->run = count->run - tmp.run;
1305 }
1306
1307 void perf_counts_values__scale(struct perf_counts_values *count,
1308                                bool scale, s8 *pscaled)
1309 {
1310         s8 scaled = 0;
1311
1312         if (scale) {
1313                 if (count->run == 0) {
1314                         scaled = -1;
1315                         count->val = 0;
1316                 } else if (count->run < count->ena) {
1317                         scaled = 1;
1318                         count->val = (u64)((double) count->val * count->ena / count->run);
1319                 }
1320         }
1321
1322         if (pscaled)
1323                 *pscaled = scaled;
1324 }
1325
1326 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1327 {
1328         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1329
1330         return perf_evsel__read(&evsel->core, cpu, thread, count);
1331 }
1332
1333 static void
1334 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1335                       u64 val, u64 ena, u64 run)
1336 {
1337         struct perf_counts_values *count;
1338
1339         count = perf_counts(counter->counts, cpu, thread);
1340
1341         count->val    = val;
1342         count->ena    = ena;
1343         count->run    = run;
1344
1345         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1346 }
1347
1348 static int
1349 perf_evsel__process_group_data(struct evsel *leader,
1350                                int cpu, int thread, u64 *data)
1351 {
1352         u64 read_format = leader->core.attr.read_format;
1353         struct sample_read_value *v;
1354         u64 nr, ena = 0, run = 0, i;
1355
1356         nr = *data++;
1357
1358         if (nr != (u64) leader->core.nr_members)
1359                 return -EINVAL;
1360
1361         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1362                 ena = *data++;
1363
1364         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1365                 run = *data++;
1366
1367         v = (struct sample_read_value *) data;
1368
1369         perf_evsel__set_count(leader, cpu, thread,
1370                               v[0].value, ena, run);
1371
1372         for (i = 1; i < nr; i++) {
1373                 struct evsel *counter;
1374
1375                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1376                 if (!counter)
1377                         return -EINVAL;
1378
1379                 perf_evsel__set_count(counter, cpu, thread,
1380                                       v[i].value, ena, run);
1381         }
1382
1383         return 0;
1384 }
1385
1386 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1387 {
1388         struct perf_stat_evsel *ps = leader->stats;
1389         u64 read_format = leader->core.attr.read_format;
1390         int size = perf_evsel__read_size(&leader->core);
1391         u64 *data = ps->group_data;
1392
1393         if (!(read_format & PERF_FORMAT_ID))
1394                 return -EINVAL;
1395
1396         if (!evsel__is_group_leader(leader))
1397                 return -EINVAL;
1398
1399         if (!data) {
1400                 data = zalloc(size);
1401                 if (!data)
1402                         return -ENOMEM;
1403
1404                 ps->group_data = data;
1405         }
1406
1407         if (FD(leader, cpu, thread) < 0)
1408                 return -EINVAL;
1409
1410         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1411                 return -errno;
1412
1413         return perf_evsel__process_group_data(leader, cpu, thread, data);
1414 }
1415
1416 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1417 {
1418         u64 read_format = evsel->core.attr.read_format;
1419
1420         if (read_format & PERF_FORMAT_GROUP)
1421                 return evsel__read_group(evsel, cpu, thread);
1422
1423         return evsel__read_one(evsel, cpu, thread);
1424 }
1425
1426 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1427 {
1428         struct perf_counts_values count;
1429         size_t nv = scale ? 3 : 1;
1430
1431         if (FD(evsel, cpu, thread) < 0)
1432                 return -EINVAL;
1433
1434         if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1435                 return -ENOMEM;
1436
1437         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1438                 return -errno;
1439
1440         evsel__compute_deltas(evsel, cpu, thread, &count);
1441         perf_counts_values__scale(&count, scale, NULL);
1442         *perf_counts(evsel->counts, cpu, thread) = count;
1443         return 0;
1444 }
1445
1446 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1447 {
1448         struct evsel *leader = evsel->leader;
1449         int fd;
1450
1451         if (evsel__is_group_leader(evsel))
1452                 return -1;
1453
1454         /*
1455          * Leader must be already processed/open,
1456          * if not it's a bug.
1457          */
1458         BUG_ON(!leader->core.fd);
1459
1460         fd = FD(leader, cpu, thread);
1461         BUG_ON(fd == -1);
1462
1463         return fd;
1464 }
1465
1466 static void perf_evsel__remove_fd(struct evsel *pos,
1467                                   int nr_cpus, int nr_threads,
1468                                   int thread_idx)
1469 {
1470         for (int cpu = 0; cpu < nr_cpus; cpu++)
1471                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1472                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1473 }
1474
1475 static int update_fds(struct evsel *evsel,
1476                       int nr_cpus, int cpu_idx,
1477                       int nr_threads, int thread_idx)
1478 {
1479         struct evsel *pos;
1480
1481         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1482                 return -EINVAL;
1483
1484         evlist__for_each_entry(evsel->evlist, pos) {
1485                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1486
1487                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1488
1489                 /*
1490                  * Since fds for next evsel has not been created,
1491                  * there is no need to iterate whole event list.
1492                  */
1493                 if (pos == evsel)
1494                         break;
1495         }
1496         return 0;
1497 }
1498
1499 static bool ignore_missing_thread(struct evsel *evsel,
1500                                   int nr_cpus, int cpu,
1501                                   struct perf_thread_map *threads,
1502                                   int thread, int err)
1503 {
1504         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1505
1506         if (!evsel->ignore_missing_thread)
1507                 return false;
1508
1509         /* The system wide setup does not work with threads. */
1510         if (evsel->core.system_wide)
1511                 return false;
1512
1513         /* The -ESRCH is perf event syscall errno for pid's not found. */
1514         if (err != -ESRCH)
1515                 return false;
1516
1517         /* If there's only one thread, let it fail. */
1518         if (threads->nr == 1)
1519                 return false;
1520
1521         /*
1522          * We should remove fd for missing_thread first
1523          * because thread_map__remove() will decrease threads->nr.
1524          */
1525         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1526                 return false;
1527
1528         if (thread_map__remove(threads, thread))
1529                 return false;
1530
1531         pr_warning("WARNING: Ignored open failure for pid %d\n",
1532                    ignore_pid);
1533         return true;
1534 }
1535
1536 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1537                                 void *priv __maybe_unused)
1538 {
1539         return fprintf(fp, "  %-32s %s\n", name, val);
1540 }
1541
1542 static void display_attr(struct perf_event_attr *attr)
1543 {
1544         if (verbose >= 2 || debug_peo_args) {
1545                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1546                 fprintf(stderr, "perf_event_attr:\n");
1547                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1548                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1549         }
1550 }
1551
1552 static int perf_event_open(struct evsel *evsel,
1553                            pid_t pid, int cpu, int group_fd,
1554                            unsigned long flags)
1555 {
1556         int precise_ip = evsel->core.attr.precise_ip;
1557         int fd;
1558
1559         while (1) {
1560                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1561                           pid, cpu, group_fd, flags);
1562
1563                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1564                 if (fd >= 0)
1565                         break;
1566
1567                 /* Do not try less precise if not requested. */
1568                 if (!evsel->precise_max)
1569                         break;
1570
1571                 /*
1572                  * We tried all the precise_ip values, and it's
1573                  * still failing, so leave it to standard fallback.
1574                  */
1575                 if (!evsel->core.attr.precise_ip) {
1576                         evsel->core.attr.precise_ip = precise_ip;
1577                         break;
1578                 }
1579
1580                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1581                 evsel->core.attr.precise_ip--;
1582                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1583                 display_attr(&evsel->core.attr);
1584         }
1585
1586         return fd;
1587 }
1588
1589 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1590                 struct perf_thread_map *threads,
1591                 int start_cpu, int end_cpu)
1592 {
1593         int cpu, thread, nthreads;
1594         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1595         int pid = -1, err, old_errno;
1596         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1597
1598         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1599             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1600                 return -EINVAL;
1601
1602         if (cpus == NULL) {
1603                 static struct perf_cpu_map *empty_cpu_map;
1604
1605                 if (empty_cpu_map == NULL) {
1606                         empty_cpu_map = perf_cpu_map__dummy_new();
1607                         if (empty_cpu_map == NULL)
1608                                 return -ENOMEM;
1609                 }
1610
1611                 cpus = empty_cpu_map;
1612         }
1613
1614         if (threads == NULL) {
1615                 static struct perf_thread_map *empty_thread_map;
1616
1617                 if (empty_thread_map == NULL) {
1618                         empty_thread_map = thread_map__new_by_tid(-1);
1619                         if (empty_thread_map == NULL)
1620                                 return -ENOMEM;
1621                 }
1622
1623                 threads = empty_thread_map;
1624         }
1625
1626         if (evsel->core.system_wide)
1627                 nthreads = 1;
1628         else
1629                 nthreads = threads->nr;
1630
1631         if (evsel->core.fd == NULL &&
1632             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1633                 return -ENOMEM;
1634
1635         if (evsel->cgrp) {
1636                 flags |= PERF_FLAG_PID_CGROUP;
1637                 pid = evsel->cgrp->fd;
1638         }
1639
1640 fallback_missing_features:
1641         if (perf_missing_features.clockid_wrong)
1642                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1643         if (perf_missing_features.clockid) {
1644                 evsel->core.attr.use_clockid = 0;
1645                 evsel->core.attr.clockid = 0;
1646         }
1647         if (perf_missing_features.cloexec)
1648                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1649         if (perf_missing_features.mmap2)
1650                 evsel->core.attr.mmap2 = 0;
1651         if (perf_missing_features.exclude_guest)
1652                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1653         if (perf_missing_features.lbr_flags)
1654                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1655                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1656         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1657                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1658         if (perf_missing_features.ksymbol)
1659                 evsel->core.attr.ksymbol = 0;
1660         if (perf_missing_features.bpf)
1661                 evsel->core.attr.bpf_event = 0;
1662         if (perf_missing_features.branch_hw_idx)
1663                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1664 retry_sample_id:
1665         if (perf_missing_features.sample_id_all)
1666                 evsel->core.attr.sample_id_all = 0;
1667
1668         display_attr(&evsel->core.attr);
1669
1670         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1671
1672                 for (thread = 0; thread < nthreads; thread++) {
1673                         int fd, group_fd;
1674
1675                         if (!evsel->cgrp && !evsel->core.system_wide)
1676                                 pid = perf_thread_map__pid(threads, thread);
1677
1678                         group_fd = get_group_fd(evsel, cpu, thread);
1679 retry_open:
1680                         test_attr__ready();
1681
1682                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1683                                              group_fd, flags);
1684
1685                         FD(evsel, cpu, thread) = fd;
1686
1687                         if (unlikely(test_attr__enabled)) {
1688                                 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1689                                                 fd, group_fd, flags);
1690                         }
1691
1692                         if (fd < 0) {
1693                                 err = -errno;
1694
1695                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1696                                         /*
1697                                          * We just removed 1 thread, so take a step
1698                                          * back on thread index and lower the upper
1699                                          * nthreads limit.
1700                                          */
1701                                         nthreads--;
1702                                         thread--;
1703
1704                                         /* ... and pretend like nothing have happened. */
1705                                         err = 0;
1706                                         continue;
1707                                 }
1708
1709                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1710                                           err);
1711                                 goto try_fallback;
1712                         }
1713
1714                         pr_debug2_peo(" = %d\n", fd);
1715
1716                         if (evsel->bpf_fd >= 0) {
1717                                 int evt_fd = fd;
1718                                 int bpf_fd = evsel->bpf_fd;
1719
1720                                 err = ioctl(evt_fd,
1721                                             PERF_EVENT_IOC_SET_BPF,
1722                                             bpf_fd);
1723                                 if (err && errno != EEXIST) {
1724                                         pr_err("failed to attach bpf fd %d: %s\n",
1725                                                bpf_fd, strerror(errno));
1726                                         err = -EINVAL;
1727                                         goto out_close;
1728                                 }
1729                         }
1730
1731                         set_rlimit = NO_CHANGE;
1732
1733                         /*
1734                          * If we succeeded but had to kill clockid, fail and
1735                          * have evsel__open_strerror() print us a nice error.
1736                          */
1737                         if (perf_missing_features.clockid ||
1738                             perf_missing_features.clockid_wrong) {
1739                                 err = -EINVAL;
1740                                 goto out_close;
1741                         }
1742                 }
1743         }
1744
1745         return 0;
1746
1747 try_fallback:
1748         /*
1749          * perf stat needs between 5 and 22 fds per CPU. When we run out
1750          * of them try to increase the limits.
1751          */
1752         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1753                 struct rlimit l;
1754
1755                 old_errno = errno;
1756                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1757                         if (set_rlimit == NO_CHANGE)
1758                                 l.rlim_cur = l.rlim_max;
1759                         else {
1760                                 l.rlim_cur = l.rlim_max + 1000;
1761                                 l.rlim_max = l.rlim_cur;
1762                         }
1763                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1764                                 set_rlimit++;
1765                                 errno = old_errno;
1766                                 goto retry_open;
1767                         }
1768                 }
1769                 errno = old_errno;
1770         }
1771
1772         if (err != -EINVAL || cpu > 0 || thread > 0)
1773                 goto out_close;
1774
1775         /*
1776          * Must probe features in the order they were added to the
1777          * perf_event_attr interface.
1778          */
1779         if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1780                 perf_missing_features.cgroup = true;
1781                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1782                 goto out_close;
1783         } else if (!perf_missing_features.branch_hw_idx &&
1784             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1785                 perf_missing_features.branch_hw_idx = true;
1786                 pr_debug2("switching off branch HW index support\n");
1787                 goto fallback_missing_features;
1788         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1789                 perf_missing_features.aux_output = true;
1790                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1791                 goto out_close;
1792         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1793                 perf_missing_features.bpf = true;
1794                 pr_debug2_peo("switching off bpf_event\n");
1795                 goto fallback_missing_features;
1796         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1797                 perf_missing_features.ksymbol = true;
1798                 pr_debug2_peo("switching off ksymbol\n");
1799                 goto fallback_missing_features;
1800         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1801                 perf_missing_features.write_backward = true;
1802                 pr_debug2_peo("switching off write_backward\n");
1803                 goto out_close;
1804         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1805                 perf_missing_features.clockid_wrong = true;
1806                 pr_debug2_peo("switching off clockid\n");
1807                 goto fallback_missing_features;
1808         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1809                 perf_missing_features.clockid = true;
1810                 pr_debug2_peo("switching off use_clockid\n");
1811                 goto fallback_missing_features;
1812         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1813                 perf_missing_features.cloexec = true;
1814                 pr_debug2_peo("switching off cloexec flag\n");
1815                 goto fallback_missing_features;
1816         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1817                 perf_missing_features.mmap2 = true;
1818                 pr_debug2_peo("switching off mmap2\n");
1819                 goto fallback_missing_features;
1820         } else if (!perf_missing_features.exclude_guest &&
1821                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1822                 perf_missing_features.exclude_guest = true;
1823                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1824                 goto fallback_missing_features;
1825         } else if (!perf_missing_features.sample_id_all) {
1826                 perf_missing_features.sample_id_all = true;
1827                 pr_debug2_peo("switching off sample_id_all\n");
1828                 goto retry_sample_id;
1829         } else if (!perf_missing_features.lbr_flags &&
1830                         (evsel->core.attr.branch_sample_type &
1831                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1832                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1833                 perf_missing_features.lbr_flags = true;
1834                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1835                 goto fallback_missing_features;
1836         } else if (!perf_missing_features.group_read &&
1837                     evsel->core.attr.inherit &&
1838                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1839                    evsel__is_group_leader(evsel)) {
1840                 perf_missing_features.group_read = true;
1841                 pr_debug2_peo("switching off group read\n");
1842                 goto fallback_missing_features;
1843         }
1844 out_close:
1845         if (err)
1846                 threads->err_thread = thread;
1847
1848         old_errno = errno;
1849         do {
1850                 while (--thread >= 0) {
1851                         if (FD(evsel, cpu, thread) >= 0)
1852                                 close(FD(evsel, cpu, thread));
1853                         FD(evsel, cpu, thread) = -1;
1854                 }
1855                 thread = nthreads;
1856         } while (--cpu >= 0);
1857         errno = old_errno;
1858         return err;
1859 }
1860
1861 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1862                 struct perf_thread_map *threads)
1863 {
1864         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1865 }
1866
1867 void evsel__close(struct evsel *evsel)
1868 {
1869         perf_evsel__close(&evsel->core);
1870         perf_evsel__free_id(&evsel->core);
1871 }
1872
1873 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1874 {
1875         if (cpu == -1)
1876                 return evsel__open_cpu(evsel, cpus, NULL, 0,
1877                                         cpus ? cpus->nr : 1);
1878
1879         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1880 }
1881
1882 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1883 {
1884         return evsel__open(evsel, NULL, threads);
1885 }
1886
1887 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1888                                        const union perf_event *event,
1889                                        struct perf_sample *sample)
1890 {
1891         u64 type = evsel->core.attr.sample_type;
1892         const __u64 *array = event->sample.array;
1893         bool swapped = evsel->needs_swap;
1894         union u64_swap u;
1895
1896         array += ((event->header.size -
1897                    sizeof(event->header)) / sizeof(u64)) - 1;
1898
1899         if (type & PERF_SAMPLE_IDENTIFIER) {
1900                 sample->id = *array;
1901                 array--;
1902         }
1903
1904         if (type & PERF_SAMPLE_CPU) {
1905                 u.val64 = *array;
1906                 if (swapped) {
1907                         /* undo swap of u64, then swap on individual u32s */
1908                         u.val64 = bswap_64(u.val64);
1909                         u.val32[0] = bswap_32(u.val32[0]);
1910                 }
1911
1912                 sample->cpu = u.val32[0];
1913                 array--;
1914         }
1915
1916         if (type & PERF_SAMPLE_STREAM_ID) {
1917                 sample->stream_id = *array;
1918                 array--;
1919         }
1920
1921         if (type & PERF_SAMPLE_ID) {
1922                 sample->id = *array;
1923                 array--;
1924         }
1925
1926         if (type & PERF_SAMPLE_TIME) {
1927                 sample->time = *array;
1928                 array--;
1929         }
1930
1931         if (type & PERF_SAMPLE_TID) {
1932                 u.val64 = *array;
1933                 if (swapped) {
1934                         /* undo swap of u64, then swap on individual u32s */
1935                         u.val64 = bswap_64(u.val64);
1936                         u.val32[0] = bswap_32(u.val32[0]);
1937                         u.val32[1] = bswap_32(u.val32[1]);
1938                 }
1939
1940                 sample->pid = u.val32[0];
1941                 sample->tid = u.val32[1];
1942                 array--;
1943         }
1944
1945         return 0;
1946 }
1947
1948 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1949                             u64 size)
1950 {
1951         return size > max_size || offset + size > endp;
1952 }
1953
1954 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1955         do {                                                            \
1956                 if (overflow(endp, (max_size), (offset), (size)))       \
1957                         return -EFAULT;                                 \
1958         } while (0)
1959
1960 #define OVERFLOW_CHECK_u64(offset) \
1961         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1962
1963 static int
1964 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1965 {
1966         /*
1967          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1968          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1969          * check the format does not go past the end of the event.
1970          */
1971         if (sample_size + sizeof(event->header) > event->header.size)
1972                 return -EFAULT;
1973
1974         return 0;
1975 }
1976
1977 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1978                         struct perf_sample *data)
1979 {
1980         u64 type = evsel->core.attr.sample_type;
1981         bool swapped = evsel->needs_swap;
1982         const __u64 *array;
1983         u16 max_size = event->header.size;
1984         const void *endp = (void *)event + max_size;
1985         u64 sz;
1986
1987         /*
1988          * used for cross-endian analysis. See git commit 65014ab3
1989          * for why this goofiness is needed.
1990          */
1991         union u64_swap u;
1992
1993         memset(data, 0, sizeof(*data));
1994         data->cpu = data->pid = data->tid = -1;
1995         data->stream_id = data->id = data->time = -1ULL;
1996         data->period = evsel->core.attr.sample_period;
1997         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1998         data->misc    = event->header.misc;
1999         data->id = -1ULL;
2000         data->data_src = PERF_MEM_DATA_SRC_NONE;
2001
2002         if (event->header.type != PERF_RECORD_SAMPLE) {
2003                 if (!evsel->core.attr.sample_id_all)
2004                         return 0;
2005                 return perf_evsel__parse_id_sample(evsel, event, data);
2006         }
2007
2008         array = event->sample.array;
2009
2010         if (perf_event__check_size(event, evsel->sample_size))
2011                 return -EFAULT;
2012
2013         if (type & PERF_SAMPLE_IDENTIFIER) {
2014                 data->id = *array;
2015                 array++;
2016         }
2017
2018         if (type & PERF_SAMPLE_IP) {
2019                 data->ip = *array;
2020                 array++;
2021         }
2022
2023         if (type & PERF_SAMPLE_TID) {
2024                 u.val64 = *array;
2025                 if (swapped) {
2026                         /* undo swap of u64, then swap on individual u32s */
2027                         u.val64 = bswap_64(u.val64);
2028                         u.val32[0] = bswap_32(u.val32[0]);
2029                         u.val32[1] = bswap_32(u.val32[1]);
2030                 }
2031
2032                 data->pid = u.val32[0];
2033                 data->tid = u.val32[1];
2034                 array++;
2035         }
2036
2037         if (type & PERF_SAMPLE_TIME) {
2038                 data->time = *array;
2039                 array++;
2040         }
2041
2042         if (type & PERF_SAMPLE_ADDR) {
2043                 data->addr = *array;
2044                 array++;
2045         }
2046
2047         if (type & PERF_SAMPLE_ID) {
2048                 data->id = *array;
2049                 array++;
2050         }
2051
2052         if (type & PERF_SAMPLE_STREAM_ID) {
2053                 data->stream_id = *array;
2054                 array++;
2055         }
2056
2057         if (type & PERF_SAMPLE_CPU) {
2058
2059                 u.val64 = *array;
2060                 if (swapped) {
2061                         /* undo swap of u64, then swap on individual u32s */
2062                         u.val64 = bswap_64(u.val64);
2063                         u.val32[0] = bswap_32(u.val32[0]);
2064                 }
2065
2066                 data->cpu = u.val32[0];
2067                 array++;
2068         }
2069
2070         if (type & PERF_SAMPLE_PERIOD) {
2071                 data->period = *array;
2072                 array++;
2073         }
2074
2075         if (type & PERF_SAMPLE_READ) {
2076                 u64 read_format = evsel->core.attr.read_format;
2077
2078                 OVERFLOW_CHECK_u64(array);
2079                 if (read_format & PERF_FORMAT_GROUP)
2080                         data->read.group.nr = *array;
2081                 else
2082                         data->read.one.value = *array;
2083
2084                 array++;
2085
2086                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2087                         OVERFLOW_CHECK_u64(array);
2088                         data->read.time_enabled = *array;
2089                         array++;
2090                 }
2091
2092                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2093                         OVERFLOW_CHECK_u64(array);
2094                         data->read.time_running = *array;
2095                         array++;
2096                 }
2097
2098                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2099                 if (read_format & PERF_FORMAT_GROUP) {
2100                         const u64 max_group_nr = UINT64_MAX /
2101                                         sizeof(struct sample_read_value);
2102
2103                         if (data->read.group.nr > max_group_nr)
2104                                 return -EFAULT;
2105                         sz = data->read.group.nr *
2106                              sizeof(struct sample_read_value);
2107                         OVERFLOW_CHECK(array, sz, max_size);
2108                         data->read.group.values =
2109                                         (struct sample_read_value *)array;
2110                         array = (void *)array + sz;
2111                 } else {
2112                         OVERFLOW_CHECK_u64(array);
2113                         data->read.one.id = *array;
2114                         array++;
2115                 }
2116         }
2117
2118         if (type & PERF_SAMPLE_CALLCHAIN) {
2119                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2120
2121                 OVERFLOW_CHECK_u64(array);
2122                 data->callchain = (struct ip_callchain *)array++;
2123                 if (data->callchain->nr > max_callchain_nr)
2124                         return -EFAULT;
2125                 sz = data->callchain->nr * sizeof(u64);
2126                 OVERFLOW_CHECK(array, sz, max_size);
2127                 array = (void *)array + sz;
2128         }
2129
2130         if (type & PERF_SAMPLE_RAW) {
2131                 OVERFLOW_CHECK_u64(array);
2132                 u.val64 = *array;
2133
2134                 /*
2135                  * Undo swap of u64, then swap on individual u32s,
2136                  * get the size of the raw area and undo all of the
2137                  * swap. The pevent interface handles endianity by
2138                  * itself.
2139                  */
2140                 if (swapped) {
2141                         u.val64 = bswap_64(u.val64);
2142                         u.val32[0] = bswap_32(u.val32[0]);
2143                         u.val32[1] = bswap_32(u.val32[1]);
2144                 }
2145                 data->raw_size = u.val32[0];
2146
2147                 /*
2148                  * The raw data is aligned on 64bits including the
2149                  * u32 size, so it's safe to use mem_bswap_64.
2150                  */
2151                 if (swapped)
2152                         mem_bswap_64((void *) array, data->raw_size);
2153
2154                 array = (void *)array + sizeof(u32);
2155
2156                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2157                 data->raw_data = (void *)array;
2158                 array = (void *)array + data->raw_size;
2159         }
2160
2161         if (type & PERF_SAMPLE_BRANCH_STACK) {
2162                 const u64 max_branch_nr = UINT64_MAX /
2163                                           sizeof(struct branch_entry);
2164
2165                 OVERFLOW_CHECK_u64(array);
2166                 data->branch_stack = (struct branch_stack *)array++;
2167
2168                 if (data->branch_stack->nr > max_branch_nr)
2169                         return -EFAULT;
2170
2171                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2172                 if (evsel__has_branch_hw_idx(evsel))
2173                         sz += sizeof(u64);
2174                 else
2175                         data->no_hw_idx = true;
2176                 OVERFLOW_CHECK(array, sz, max_size);
2177                 array = (void *)array + sz;
2178         }
2179
2180         if (type & PERF_SAMPLE_REGS_USER) {
2181                 OVERFLOW_CHECK_u64(array);
2182                 data->user_regs.abi = *array;
2183                 array++;
2184
2185                 if (data->user_regs.abi) {
2186                         u64 mask = evsel->core.attr.sample_regs_user;
2187
2188                         sz = hweight64(mask) * sizeof(u64);
2189                         OVERFLOW_CHECK(array, sz, max_size);
2190                         data->user_regs.mask = mask;
2191                         data->user_regs.regs = (u64 *)array;
2192                         array = (void *)array + sz;
2193                 }
2194         }
2195
2196         if (type & PERF_SAMPLE_STACK_USER) {
2197                 OVERFLOW_CHECK_u64(array);
2198                 sz = *array++;
2199
2200                 data->user_stack.offset = ((char *)(array - 1)
2201                                           - (char *) event);
2202
2203                 if (!sz) {
2204                         data->user_stack.size = 0;
2205                 } else {
2206                         OVERFLOW_CHECK(array, sz, max_size);
2207                         data->user_stack.data = (char *)array;
2208                         array = (void *)array + sz;
2209                         OVERFLOW_CHECK_u64(array);
2210                         data->user_stack.size = *array++;
2211                         if (WARN_ONCE(data->user_stack.size > sz,
2212                                       "user stack dump failure\n"))
2213                                 return -EFAULT;
2214                 }
2215         }
2216
2217         if (type & PERF_SAMPLE_WEIGHT) {
2218                 OVERFLOW_CHECK_u64(array);
2219                 data->weight = *array;
2220                 array++;
2221         }
2222
2223         if (type & PERF_SAMPLE_DATA_SRC) {
2224                 OVERFLOW_CHECK_u64(array);
2225                 data->data_src = *array;
2226                 array++;
2227         }
2228
2229         if (type & PERF_SAMPLE_TRANSACTION) {
2230                 OVERFLOW_CHECK_u64(array);
2231                 data->transaction = *array;
2232                 array++;
2233         }
2234
2235         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2236         if (type & PERF_SAMPLE_REGS_INTR) {
2237                 OVERFLOW_CHECK_u64(array);
2238                 data->intr_regs.abi = *array;
2239                 array++;
2240
2241                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2242                         u64 mask = evsel->core.attr.sample_regs_intr;
2243
2244                         sz = hweight64(mask) * sizeof(u64);
2245                         OVERFLOW_CHECK(array, sz, max_size);
2246                         data->intr_regs.mask = mask;
2247                         data->intr_regs.regs = (u64 *)array;
2248                         array = (void *)array + sz;
2249                 }
2250         }
2251
2252         data->phys_addr = 0;
2253         if (type & PERF_SAMPLE_PHYS_ADDR) {
2254                 data->phys_addr = *array;
2255                 array++;
2256         }
2257
2258         data->cgroup = 0;
2259         if (type & PERF_SAMPLE_CGROUP) {
2260                 data->cgroup = *array;
2261                 array++;
2262         }
2263
2264         if (type & PERF_SAMPLE_AUX) {
2265                 OVERFLOW_CHECK_u64(array);
2266                 sz = *array++;
2267
2268                 OVERFLOW_CHECK(array, sz, max_size);
2269                 /* Undo swap of data */
2270                 if (swapped)
2271                         mem_bswap_64((char *)array, sz);
2272                 data->aux_sample.size = sz;
2273                 data->aux_sample.data = (char *)array;
2274                 array = (void *)array + sz;
2275         }
2276
2277         return 0;
2278 }
2279
2280 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2281                                   u64 *timestamp)
2282 {
2283         u64 type = evsel->core.attr.sample_type;
2284         const __u64 *array;
2285
2286         if (!(type & PERF_SAMPLE_TIME))
2287                 return -1;
2288
2289         if (event->header.type != PERF_RECORD_SAMPLE) {
2290                 struct perf_sample data = {
2291                         .time = -1ULL,
2292                 };
2293
2294                 if (!evsel->core.attr.sample_id_all)
2295                         return -1;
2296                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2297                         return -1;
2298
2299                 *timestamp = data.time;
2300                 return 0;
2301         }
2302
2303         array = event->sample.array;
2304
2305         if (perf_event__check_size(event, evsel->sample_size))
2306                 return -EFAULT;
2307
2308         if (type & PERF_SAMPLE_IDENTIFIER)
2309                 array++;
2310
2311         if (type & PERF_SAMPLE_IP)
2312                 array++;
2313
2314         if (type & PERF_SAMPLE_TID)
2315                 array++;
2316
2317         if (type & PERF_SAMPLE_TIME)
2318                 *timestamp = *array;
2319
2320         return 0;
2321 }
2322
2323 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2324 {
2325         return tep_find_field(evsel->tp_format, name);
2326 }
2327
2328 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2329 {
2330         struct tep_format_field *field = evsel__field(evsel, name);
2331         int offset;
2332
2333         if (!field)
2334                 return NULL;
2335
2336         offset = field->offset;
2337
2338         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2339                 offset = *(int *)(sample->raw_data + field->offset);
2340                 offset &= 0xffff;
2341         }
2342
2343         return sample->raw_data + offset;
2344 }
2345
2346 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2347                          bool needs_swap)
2348 {
2349         u64 value;
2350         void *ptr = sample->raw_data + field->offset;
2351
2352         switch (field->size) {
2353         case 1:
2354                 return *(u8 *)ptr;
2355         case 2:
2356                 value = *(u16 *)ptr;
2357                 break;
2358         case 4:
2359                 value = *(u32 *)ptr;
2360                 break;
2361         case 8:
2362                 memcpy(&value, ptr, sizeof(u64));
2363                 break;
2364         default:
2365                 return 0;
2366         }
2367
2368         if (!needs_swap)
2369                 return value;
2370
2371         switch (field->size) {
2372         case 2:
2373                 return bswap_16(value);
2374         case 4:
2375                 return bswap_32(value);
2376         case 8:
2377                 return bswap_64(value);
2378         default:
2379                 return 0;
2380         }
2381
2382         return 0;
2383 }
2384
2385 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2386 {
2387         struct tep_format_field *field = evsel__field(evsel, name);
2388
2389         if (!field)
2390                 return 0;
2391
2392         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2393 }
2394
2395 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2396 {
2397         int paranoid;
2398
2399         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2400             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2401             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2402                 /*
2403                  * If it's cycles then fall back to hrtimer based
2404                  * cpu-clock-tick sw counter, which is always available even if
2405                  * no PMU support.
2406                  *
2407                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2408                  * b0a873e).
2409                  */
2410                 scnprintf(msg, msgsize, "%s",
2411 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2412
2413                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2414                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2415
2416                 zfree(&evsel->name);
2417                 return true;
2418         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2419                    (paranoid = perf_event_paranoid()) > 1) {
2420                 const char *name = evsel__name(evsel);
2421                 char *new_name;
2422                 const char *sep = ":";
2423
2424                 /* If event has exclude user then don't exclude kernel. */
2425                 if (evsel->core.attr.exclude_user)
2426                         return false;
2427
2428                 /* Is there already the separator in the name. */
2429                 if (strchr(name, '/') ||
2430                     (strchr(name, ':') && !evsel->is_libpfm_event))
2431                         sep = "";
2432
2433                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2434                         return false;
2435
2436                 if (evsel->name)
2437                         free(evsel->name);
2438                 evsel->name = new_name;
2439                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2440                           "to fall back to excluding kernel and hypervisor "
2441                           " samples", paranoid);
2442                 evsel->core.attr.exclude_kernel = 1;
2443                 evsel->core.attr.exclude_hv     = 1;
2444
2445                 return true;
2446         }
2447
2448         return false;
2449 }
2450
2451 static bool find_process(const char *name)
2452 {
2453         size_t len = strlen(name);
2454         DIR *dir;
2455         struct dirent *d;
2456         int ret = -1;
2457
2458         dir = opendir(procfs__mountpoint());
2459         if (!dir)
2460                 return false;
2461
2462         /* Walk through the directory. */
2463         while (ret && (d = readdir(dir)) != NULL) {
2464                 char path[PATH_MAX];
2465                 char *data;
2466                 size_t size;
2467
2468                 if ((d->d_type != DT_DIR) ||
2469                      !strcmp(".", d->d_name) ||
2470                      !strcmp("..", d->d_name))
2471                         continue;
2472
2473                 scnprintf(path, sizeof(path), "%s/%s/comm",
2474                           procfs__mountpoint(), d->d_name);
2475
2476                 if (filename__read_str(path, &data, &size))
2477                         continue;
2478
2479                 ret = strncmp(name, data, len);
2480                 free(data);
2481         }
2482
2483         closedir(dir);
2484         return ret ? false : true;
2485 }
2486
2487 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2488                          int err, char *msg, size_t size)
2489 {
2490         char sbuf[STRERR_BUFSIZE];
2491         int printed = 0, enforced = 0;
2492
2493         switch (err) {
2494         case EPERM:
2495         case EACCES:
2496                 printed += scnprintf(msg + printed, size - printed,
2497                         "Access to performance monitoring and observability operations is limited.\n");
2498
2499                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2500                         if (enforced) {
2501                                 printed += scnprintf(msg + printed, size - printed,
2502                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2503                                         "monitoring and observability operations. Inspect system audit records for\n"
2504                                         "more perf_event access control information and adjusting the policy.\n");
2505                         }
2506                 }
2507
2508                 if (err == EPERM)
2509                         printed += scnprintf(msg, size,
2510                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2511
2512                 return scnprintf(msg + printed, size - printed,
2513                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2514                  "access to performance monitoring and observability operations for processes\n"
2515                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2516                  "More information can be found at 'Perf events and tool security' document:\n"
2517                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2518                  "perf_event_paranoid setting is %d:\n"
2519                  "  -1: Allow use of (almost) all events by all users\n"
2520                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2521                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2522                  ">= 1: Disallow CPU event access\n"
2523                  ">= 2: Disallow kernel profiling\n"
2524                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2525                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2526                  perf_event_paranoid());
2527         case ENOENT:
2528                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2529         case EMFILE:
2530                 return scnprintf(msg, size, "%s",
2531                          "Too many events are opened.\n"
2532                          "Probably the maximum number of open file descriptors has been reached.\n"
2533                          "Hint: Try again after reducing the number of events.\n"
2534                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2535         case ENOMEM:
2536                 if (evsel__has_callchain(evsel) &&
2537                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2538                         return scnprintf(msg, size,
2539                                          "Not enough memory to setup event with callchain.\n"
2540                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2541                                          "Hint: Current value: %d", sysctl__max_stack());
2542                 break;
2543         case ENODEV:
2544                 if (target->cpu_list)
2545                         return scnprintf(msg, size, "%s",
2546          "No such device - did you specify an out-of-range profile CPU?");
2547                 break;
2548         case EOPNOTSUPP:
2549                 if (evsel->core.attr.aux_output)
2550                         return scnprintf(msg, size,
2551         "%s: PMU Hardware doesn't support 'aux_output' feature",
2552                                          evsel__name(evsel));
2553                 if (evsel->core.attr.sample_period != 0)
2554                         return scnprintf(msg, size,
2555         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2556                                          evsel__name(evsel));
2557                 if (evsel->core.attr.precise_ip)
2558                         return scnprintf(msg, size, "%s",
2559         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2560 #if defined(__i386__) || defined(__x86_64__)
2561                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2562                         return scnprintf(msg, size, "%s",
2563         "No hardware sampling interrupt available.\n");
2564 #endif
2565                 break;
2566         case EBUSY:
2567                 if (find_process("oprofiled"))
2568                         return scnprintf(msg, size,
2569         "The PMU counters are busy/taken by another profiler.\n"
2570         "We found oprofile daemon running, please stop it and try again.");
2571                 break;
2572         case EINVAL:
2573                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2574                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2575                 if (perf_missing_features.clockid)
2576                         return scnprintf(msg, size, "clockid feature not supported.");
2577                 if (perf_missing_features.clockid_wrong)
2578                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2579                 if (perf_missing_features.aux_output)
2580                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2581                 break;
2582         default:
2583                 break;
2584         }
2585
2586         return scnprintf(msg, size,
2587         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2588         "/bin/dmesg | grep -i perf may provide additional information.\n",
2589                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2590 }
2591
2592 struct perf_env *evsel__env(struct evsel *evsel)
2593 {
2594         if (evsel && evsel->evlist)
2595                 return evsel->evlist->env;
2596         return &perf_env;
2597 }
2598
2599 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2600 {
2601         int cpu, thread;
2602
2603         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2604                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2605                      thread++) {
2606                         int fd = FD(evsel, cpu, thread);
2607
2608                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2609                                                    cpu, thread, fd) < 0)
2610                                 return -1;
2611                 }
2612         }
2613
2614         return 0;
2615 }
2616
2617 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2618 {
2619         struct perf_cpu_map *cpus = evsel->core.cpus;
2620         struct perf_thread_map *threads = evsel->core.threads;
2621
2622         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2623                 return -ENOMEM;
2624
2625         return store_evsel_ids(evsel, evlist);
2626 }