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