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 "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "cpumap.h"
  32 #include "thread_map.h"
  33 #include "target.h"
  34 #include "perf_regs.h"
  35 #include "debug.h"
  36 #include "trace-event.h"
  37 #include "stat.h"
  38 #include "string2.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include <linux/ctype.h>
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->max_events  = ULONG_MAX;
 236         evsel->evlist      = NULL;
 237         evsel->bpf_fd      = -1;
 238         INIT_LIST_HEAD(&evsel->node);
 239         INIT_LIST_HEAD(&evsel->config_terms);
 240         perf_evsel__object.init(evsel);
 241         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 242         perf_evsel__calc_id_pos(evsel);
 243         evsel->cmdline_group_boundary = false;
 244         evsel->metric_expr   = NULL;
 245         evsel->metric_name   = NULL;
 246         evsel->metric_events = NULL;
 247         evsel->collect_stat  = false;
 248         evsel->pmu_name      = NULL;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (!evsel)
 256                 return NULL;
 257         perf_evsel__init(evsel, attr, idx);
 258
 259         if (perf_evsel__is_bpf_output(evsel)) {
 260                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 261                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 262                 evsel->attr.sample_period = 1;
 263         }
 264
 265         if (perf_evsel__is_clock(evsel)) {
 266                 /*
 267                  * The evsel->unit points to static alias->unit
 268                  * so it's ok to use static string in here.
 269                  */
 270                 static const char *unit = "msec";
 271
 272                 evsel->unit = unit;
 273                 evsel->scale = 1e-6;
 274         }
 275
 276         return evsel;
 277 }
 278
 279 static bool perf_event_can_profile_kernel(void)
 280 {
 281         return geteuid() == 0 || perf_event_paranoid() == -1;
 282 }
 283
 284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 285 {
 286         struct perf_event_attr attr = {
 287                 .type   = PERF_TYPE_HARDWARE,
 288                 .config = PERF_COUNT_HW_CPU_CYCLES,
 289                 .exclude_kernel = !perf_event_can_profile_kernel(),
 290         };
 291         struct perf_evsel *evsel;
 292
 293         event_attr_init(&attr);
 294
 295         if (!precise)
 296                 goto new_event;
 297
 298         /*
 299          * Now let the usual logic to set up the perf_event_attr defaults
 300          * to kick in when we return and before perf_evsel__open() is called.
 301          */
 302 new_event:
 303         evsel = perf_evsel__new(&attr);
 304         if (evsel == NULL)
 305                 goto out;
 306
 307         evsel->precise_max = true;
 308
 309         /* use asprintf() because free(evsel) assumes name is allocated */
 310         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 311                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 312                      attr.exclude_kernel ? "u" : "",
 313                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 314                 goto error_free;
 315 out:
 316         return evsel;
 317 error_free:
 318         perf_evsel__delete(evsel);
 319         evsel = NULL;
 320         goto out;
 321 }
 322
 323 /*
 324  * Returns pointer with encoded error via <linux/err.h> interface.
 325  */
 326 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 327 {
 328         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 329         int err = -ENOMEM;
 330
 331         if (evsel == NULL) {
 332                 goto out_err;
 333         } else {
 334                 struct perf_event_attr attr = {
 335                         .type          = PERF_TYPE_TRACEPOINT,
 336                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 337                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 338                 };
 339
 340                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 341                         goto out_free;
 342
 343                 evsel->tp_format = trace_event__tp_format(sys, name);
 344                 if (IS_ERR(evsel->tp_format)) {
 345                         err = PTR_ERR(evsel->tp_format);
 346                         goto out_free;
 347                 }
 348
 349                 event_attr_init(&attr);
 350                 attr.config = evsel->tp_format->id;
 351                 attr.sample_period = 1;
 352                 perf_evsel__init(evsel, &attr, idx);
 353         }
 354
 355         return evsel;
 356
 357 out_free:
 358         zfree(&evsel->name);
 359         free(evsel);
 360 out_err:
 361         return ERR_PTR(err);
 362 }
 363
 364 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 365         "cycles",
 366         "instructions",
 367         "cache-references",
 368         "cache-misses",
 369         "branches",
 370         "branch-misses",
 371         "bus-cycles",
 372         "stalled-cycles-frontend",
 373         "stalled-cycles-backend",
 374         "ref-cycles",
 375 };
 376
 377 static const char *__perf_evsel__hw_name(u64 config)
 378 {
 379         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 380                 return perf_evsel__hw_names[config];
 381
 382         return "unknown-hardware";
 383 }
 384
 385 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 386 {
 387         int colon = 0, r = 0;
 388         struct perf_event_attr *attr = &evsel->attr;
 389         bool exclude_guest_default = false;
 390
 391 #define MOD_PRINT(context, mod) do {                                    \
 392                 if (!attr->exclude_##context) {                         \
 393                         if (!colon) colon = ++r;                        \
 394                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 395                 } } while(0)
 396
 397         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 398                 MOD_PRINT(kernel, 'k');
 399                 MOD_PRINT(user, 'u');
 400                 MOD_PRINT(hv, 'h');
 401                 exclude_guest_default = true;
 402         }
 403
 404         if (attr->precise_ip) {
 405                 if (!colon)
 406                         colon = ++r;
 407                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 412                 MOD_PRINT(host, 'H');
 413                 MOD_PRINT(guest, 'G');
 414         }
 415 #undef MOD_PRINT
 416         if (colon)
 417                 bf[colon - 1] = ':';
 418         return r;
 419 }
 420
 421 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 422 {
 423         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 424         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 425 }
 426
 427 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 428         "cpu-clock",
 429         "task-clock",
 430         "page-faults",
 431         "context-switches",
 432         "cpu-migrations",
 433         "minor-faults",
 434         "major-faults",
 435         "alignment-faults",
 436         "emulation-faults",
 437         "dummy",
 438 };
 439
 440 static const char *__perf_evsel__sw_name(u64 config)
 441 {
 442         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 443                 return perf_evsel__sw_names[config];
 444         return "unknown-software";
 445 }
 446
 447 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 448 {
 449         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 450         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 451 }
 452
 453 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 454 {
 455         int r;
 456
 457         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 458
 459         if (type & HW_BREAKPOINT_R)
 460                 r += scnprintf(bf + r, size - r, "r");
 461
 462         if (type & HW_BREAKPOINT_W)
 463                 r += scnprintf(bf + r, size - r, "w");
 464
 465         if (type & HW_BREAKPOINT_X)
 466                 r += scnprintf(bf + r, size - r, "x");
 467
 468         return r;
 469 }
 470
 471 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 472 {
 473         struct perf_event_attr *attr = &evsel->attr;
 474         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 475         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 476 }
 477
 478 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 479                                 [PERF_EVSEL__MAX_ALIASES] = {
 480  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 481  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 482  { "LLC",       "L2",                                                   },
 483  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 484  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 485  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 486  { "node",                                                              },
 487 };
 488
 489 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 490                                    [PERF_EVSEL__MAX_ALIASES] = {
 491  { "load",      "loads",        "read",                                 },
 492  { "store",     "stores",       "write",                                },
 493  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 494 };
 495
 496 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 497                                        [PERF_EVSEL__MAX_ALIASES] = {
 498  { "refs",      "Reference",    "ops",          "access",               },
 499  { "misses",    "miss",                                                 },
 500 };
 501
 502 #define C(x)            PERF_COUNT_HW_CACHE_##x
 503 #define CACHE_READ      (1 << C(OP_READ))
 504 #define CACHE_WRITE     (1 << C(OP_WRITE))
 505 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 506 #define COP(x)          (1 << x)
 507
 508 /*
 509  * cache operartion stat
 510  * L1I : Read and prefetch only
 511  * ITLB and BPU : Read-only
 512  */
 513 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 514  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 515  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 516  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 517  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 518  [C(ITLB)]      = (CACHE_READ),
 519  [C(BPU)]       = (CACHE_READ),
 520  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 521 };
 522
 523 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 524 {
 525         if (perf_evsel__hw_cache_stat[type] & COP(op))
 526                 return true;    /* valid */
 527         else
 528                 return false;   /* invalid */
 529 }
 530
 531 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 532                                             char *bf, size_t size)
 533 {
 534         if (result) {
 535                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 536                                  perf_evsel__hw_cache_op[op][0],
 537                                  perf_evsel__hw_cache_result[result][0]);
 538         }
 539
 540         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 541                          perf_evsel__hw_cache_op[op][1]);
 542 }
 543
 544 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 545 {
 546         u8 op, result, type = (config >>  0) & 0xff;
 547         const char *err = "unknown-ext-hardware-cache-type";
 548
 549         if (type >= PERF_COUNT_HW_CACHE_MAX)
 550                 goto out_err;
 551
 552         op = (config >>  8) & 0xff;
 553         err = "unknown-ext-hardware-cache-op";
 554         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 555                 goto out_err;
 556
 557         result = (config >> 16) & 0xff;
 558         err = "unknown-ext-hardware-cache-result";
 559         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 560                 goto out_err;
 561
 562         err = "invalid-cache";
 563         if (!perf_evsel__is_cache_op_valid(type, op))
 564                 goto out_err;
 565
 566         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 567 out_err:
 568         return scnprintf(bf, size, "%s", err);
 569 }
 570
 571 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 572 {
 573         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 574         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 575 }
 576
 577 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 578 {
 579         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 580         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 581 }
 582
 583 static int perf_evsel__tool_name(char *bf, size_t size)
 584 {
 585         int ret = scnprintf(bf, size, "duration_time");
 586         return ret;
 587 }
 588
 589 const char *perf_evsel__name(struct perf_evsel *evsel)
 590 {
 591         char bf[128];
 592
 593         if (!evsel)
 594                 goto out_unknown;
 595
 596         if (evsel->name)
 597                 return evsel->name;
 598
 599         switch (evsel->attr.type) {
 600         case PERF_TYPE_RAW:
 601                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 602                 break;
 603
 604         case PERF_TYPE_HARDWARE:
 605                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 606                 break;
 607
 608         case PERF_TYPE_HW_CACHE:
 609                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 610                 break;
 611
 612         case PERF_TYPE_SOFTWARE:
 613                 if (evsel->tool_event)
 614                         perf_evsel__tool_name(bf, sizeof(bf));
 615                 else
 616                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 617                 break;
 618
 619         case PERF_TYPE_TRACEPOINT:
 620                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 621                 break;
 622
 623         case PERF_TYPE_BREAKPOINT:
 624                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 625                 break;
 626
 627         default:
 628                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 629                           evsel->attr.type);
 630                 break;
 631         }
 632
 633         evsel->name = strdup(bf);
 634
 635         if (evsel->name)
 636                 return evsel->name;
 637 out_unknown:
 638         return "unknown";
 639 }
 640
 641 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 642 {
 643         return evsel->group_name ?: "anon group";
 644 }
 645
 646 /*
 647  * Returns the group details for the specified leader,
 648  * with following rules.
 649  *
 650  *  For record -e '{cycles,instructions}'
 651  *    'anon group { cycles:u, instructions:u }'
 652  *
 653  *  For record -e 'cycles,instructions' and report --group
 654  *    'cycles:u, instructions:u'
 655  */
 656 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 657 {
 658         int ret = 0;
 659         struct perf_evsel *pos;
 660         const char *group_name = perf_evsel__group_name(evsel);
 661
 662         if (!evsel->forced_leader)
 663                 ret = scnprintf(buf, size, "%s { ", group_name);
 664
 665         ret += scnprintf(buf + ret, size - ret, "%s",
 666                          perf_evsel__name(evsel));
 667
 668         for_each_group_member(pos, evsel)
 669                 ret += scnprintf(buf + ret, size - ret, ", %s",
 670                                  perf_evsel__name(pos));
 671
 672         if (!evsel->forced_leader)
 673                 ret += scnprintf(buf + ret, size - ret, " }");
 674
 675         return ret;
 676 }
 677
 678 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 679                                            struct record_opts *opts,
 680                                            struct callchain_param *param)
 681 {
 682         bool function = perf_evsel__is_function_event(evsel);
 683         struct perf_event_attr *attr = &evsel->attr;
 684
 685         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 686
 687         attr->sample_max_stack = param->max_stack;
 688
 689         if (opts->kernel_callchains)
 690                 attr->exclude_callchain_user = 1;
 691         if (opts->user_callchains)
 692                 attr->exclude_callchain_kernel = 1;
 693         if (param->record_mode == CALLCHAIN_LBR) {
 694                 if (!opts->branch_stack) {
 695                         if (attr->exclude_user) {
 696                                 pr_warning("LBR callstack option is only available "
 697                                            "to get user callchain information. "
 698                                            "Falling back to framepointers.\n");
 699                         } else {
 700                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 701                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 702                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 703                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 704                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 705                         }
 706                 } else
 707                          pr_warning("Cannot use LBR callstack with branch stack. "
 708                                     "Falling back to framepointers.\n");
 709         }
 710
 711         if (param->record_mode == CALLCHAIN_DWARF) {
 712                 if (!function) {
 713                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 714                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 715                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 716                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 717                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 718                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 719                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 720                         } else {
 721                                 attr->sample_regs_user |= PERF_REGS_MASK;
 722                         }
 723                         attr->sample_stack_user = param->dump_size;
 724                         attr->exclude_callchain_user = 1;
 725                 } else {
 726                         pr_info("Cannot use DWARF unwind for function trace event,"
 727                                 " falling back to framepointers.\n");
 728                 }
 729         }
 730
 731         if (function) {
 732                 pr_info("Disabling user space callchains for function trace event.\n");
 733                 attr->exclude_callchain_user = 1;
 734         }
 735 }
 736
 737 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 738                                   struct record_opts *opts,
 739                                   struct callchain_param *param)
 740 {
 741         if (param->enabled)
 742                 return __perf_evsel__config_callchain(evsel, opts, param);
 743 }
 744
 745 static void
 746 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 747                             struct callchain_param *param)
 748 {
 749         struct perf_event_attr *attr = &evsel->attr;
 750
 751         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 752         if (param->record_mode == CALLCHAIN_LBR) {
 753                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 754                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 755                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 756         }
 757         if (param->record_mode == CALLCHAIN_DWARF) {
 758                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 759                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 760         }
 761 }
 762
 763 static void apply_config_terms(struct perf_evsel *evsel,
 764                                struct record_opts *opts, bool track)
 765 {
 766         struct perf_evsel_config_term *term;
 767         struct list_head *config_terms = &evsel->config_terms;
 768         struct perf_event_attr *attr = &evsel->attr;
 769         /* callgraph default */
 770         struct callchain_param param = {
 771                 .record_mode = callchain_param.record_mode,
 772         };
 773         u32 dump_size = 0;
 774         int max_stack = 0;
 775         const char *callgraph_buf = NULL;
 776
 777         list_for_each_entry(term, config_terms, list) {
 778                 switch (term->type) {
 779                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 780                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 781                                 attr->sample_period = term->val.period;
 782                                 attr->freq = 0;
 783                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 784                         }
 785                         break;
 786                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 787                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 788                                 attr->sample_freq = term->val.freq;
 789                                 attr->freq = 1;
 790                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 791                         }
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_TIME:
 794                         if (term->val.time)
 795                                 perf_evsel__set_sample_bit(evsel, TIME);
 796                         else
 797                                 perf_evsel__reset_sample_bit(evsel, TIME);
 798                         break;
 799                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 800                         callgraph_buf = term->val.callgraph;
 801                         break;
 802                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 803                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 804                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 805                                 parse_branch_str(term->val.branch,
 806                                                  &attr->branch_sample_type);
 807                         } else
 808                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 809                         break;
 810                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 811                         dump_size = term->val.stack_user;
 812                         break;
 813                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 814                         max_stack = term->val.max_stack;
 815                         break;
 816                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 817                         evsel->max_events = term->val.max_events;
 818                         break;
 819                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 820                         /*
 821                          * attr->inherit should has already been set by
 822                          * perf_evsel__config. If user explicitly set
 823                          * inherit using config terms, override global
 824                          * opt->no_inherit setting.
 825                          */
 826                         attr->inherit = term->val.inherit ? 1 : 0;
 827                         break;
 828                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 829                         attr->write_backward = term->val.overwrite ? 1 : 0;
 830                         break;
 831                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 832                         break;
 833                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 834                         break;
 835                 default:
 836                         break;
 837                 }
 838         }
 839
 840         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 841         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 842                 bool sample_address = false;
 843
 844                 if (max_stack) {
 845                         param.max_stack = max_stack;
 846                         if (callgraph_buf == NULL)
 847                                 callgraph_buf = "fp";
 848                 }
 849
 850                 /* parse callgraph parameters */
 851                 if (callgraph_buf != NULL) {
 852                         if (!strcmp(callgraph_buf, "no")) {
 853                                 param.enabled = false;
 854                                 param.record_mode = CALLCHAIN_NONE;
 855                         } else {
 856                                 param.enabled = true;
 857                                 if (parse_callchain_record(callgraph_buf, &param)) {
 858                                         pr_err("per-event callgraph setting for %s failed. "
 859                                                "Apply callgraph global setting for it\n",
 860                                                evsel->name);
 861                                         return;
 862                                 }
 863                                 if (param.record_mode == CALLCHAIN_DWARF)
 864                                         sample_address = true;
 865                         }
 866                 }
 867                 if (dump_size > 0) {
 868                         dump_size = round_up(dump_size, sizeof(u64));
 869                         param.dump_size = dump_size;
 870                 }
 871
 872                 /* If global callgraph set, clear it */
 873                 if (callchain_param.enabled)
 874                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 875
 876                 /* set perf-event callgraph */
 877                 if (param.enabled) {
 878                         if (sample_address) {
 879                                 perf_evsel__set_sample_bit(evsel, ADDR);
 880                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 881                                 evsel->attr.mmap_data = track;
 882                         }
 883                         perf_evsel__config_callchain(evsel, opts, &param);
 884                 }
 885         }
 886 }
 887
 888 static bool is_dummy_event(struct perf_evsel *evsel)
 889 {
 890         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 891                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 892 }
 893
 894 /*
 895  * The enable_on_exec/disabled value strategy:
 896  *
 897  *  1) For any type of traced program:
 898  *    - all independent events and group leaders are disabled
 899  *    - all group members are enabled
 900  *
 901  *     Group members are ruled by group leaders. They need to
 902  *     be enabled, because the group scheduling relies on that.
 903  *
 904  *  2) For traced programs executed by perf:
 905  *     - all independent events and group leaders have
 906  *       enable_on_exec set
 907  *     - we don't specifically enable or disable any event during
 908  *       the record command
 909  *
 910  *     Independent events and group leaders are initially disabled
 911  *     and get enabled by exec. Group members are ruled by group
 912  *     leaders as stated in 1).
 913  *
 914  *  3) For traced programs attached by perf (pid/tid):
 915  *     - we specifically enable or disable all events during
 916  *       the record command
 917  *
 918  *     When attaching events to already running traced we
 919  *     enable/disable events specifically, as there's no
 920  *     initial traced exec call.
 921  */
 922 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 923                         struct callchain_param *callchain)
 924 {
 925         struct perf_evsel *leader = evsel->leader;
 926         struct perf_event_attr *attr = &evsel->attr;
 927         int track = evsel->tracking;
 928         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 929
 930         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 931         attr->inherit       = !opts->no_inherit;
 932         attr->write_backward = opts->overwrite ? 1 : 0;
 933
 934         perf_evsel__set_sample_bit(evsel, IP);
 935         perf_evsel__set_sample_bit(evsel, TID);
 936
 937         if (evsel->sample_read) {
 938                 perf_evsel__set_sample_bit(evsel, READ);
 939
 940                 /*
 941                  * We need ID even in case of single event, because
 942                  * PERF_SAMPLE_READ process ID specific data.
 943                  */
 944                 perf_evsel__set_sample_id(evsel, false);
 945
 946                 /*
 947                  * Apply group format only if we belong to group
 948                  * with more than one members.
 949                  */
 950                 if (leader->nr_members > 1) {
 951                         attr->read_format |= PERF_FORMAT_GROUP;
 952                         attr->inherit = 0;
 953                 }
 954         }
 955
 956         /*
 957          * We default some events to have a default interval. But keep
 958          * it a weak assumption overridable by the user.
 959          */
 960         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 961                                      opts->user_interval != ULLONG_MAX)) {
 962                 if (opts->freq) {
 963                         perf_evsel__set_sample_bit(evsel, PERIOD);
 964                         attr->freq              = 1;
 965                         attr->sample_freq       = opts->freq;
 966                 } else {
 967                         attr->sample_period = opts->default_interval;
 968                 }
 969         }
 970
 971         /*
 972          * Disable sampling for all group members other
 973          * than leader in case leader 'leads' the sampling.
 974          */
 975         if ((leader != evsel) && leader->sample_read) {
 976                 attr->freq           = 0;
 977                 attr->sample_freq    = 0;
 978                 attr->sample_period  = 0;
 979                 attr->write_backward = 0;
 980
 981                 /*
 982                  * We don't get sample for slave events, we make them
 983                  * when delivering group leader sample. Set the slave
 984                  * event to follow the master sample_type to ease up
 985                  * report.
 986                  */
 987                 attr->sample_type = leader->attr.sample_type;
 988         }
 989
 990         if (opts->no_samples)
 991                 attr->sample_freq = 0;
 992
 993         if (opts->inherit_stat) {
 994                 evsel->attr.read_format |=
 995                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 996                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 997                         PERF_FORMAT_ID;
 998                 attr->inherit_stat = 1;
 999         }
1000
1001         if (opts->sample_address) {
1002                 perf_evsel__set_sample_bit(evsel, ADDR);
1003                 attr->mmap_data = track;
1004         }
1005
1006         /*
1007          * We don't allow user space callchains for  function trace
1008          * event, due to issues with page faults while tracing page
1009          * fault handler and its overall trickiness nature.
1010          */
1011         if (perf_evsel__is_function_event(evsel))
1012                 evsel->attr.exclude_callchain_user = 1;
1013
1014         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1015                 perf_evsel__config_callchain(evsel, opts, callchain);
1016
1017         if (opts->sample_intr_regs) {
1018                 attr->sample_regs_intr = opts->sample_intr_regs;
1019                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1020         }
1021
1022         if (opts->sample_user_regs) {
1023                 attr->sample_regs_user |= opts->sample_user_regs;
1024                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1025         }
1026
1027         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1028                 perf_evsel__set_sample_bit(evsel, CPU);
1029
1030         /*
1031          * When the user explicitly disabled time don't force it here.
1032          */
1033         if (opts->sample_time &&
1034             (!perf_missing_features.sample_id_all &&
1035             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1036              opts->sample_time_set)))
1037                 perf_evsel__set_sample_bit(evsel, TIME);
1038
1039         if (opts->raw_samples && !evsel->no_aux_samples) {
1040                 perf_evsel__set_sample_bit(evsel, TIME);
1041                 perf_evsel__set_sample_bit(evsel, RAW);
1042                 perf_evsel__set_sample_bit(evsel, CPU);
1043         }
1044
1045         if (opts->sample_address)
1046                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1047
1048         if (opts->sample_phys_addr)
1049                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1050
1051         if (opts->no_buffering) {
1052                 attr->watermark = 0;
1053                 attr->wakeup_events = 1;
1054         }
1055         if (opts->branch_stack && !evsel->no_aux_samples) {
1056                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1057                 attr->branch_sample_type = opts->branch_stack;
1058         }
1059
1060         if (opts->sample_weight)
1061                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1062
1063         attr->task  = track;
1064         attr->mmap  = track;
1065         attr->mmap2 = track && !perf_missing_features.mmap2;
1066         attr->comm  = track;
1067         attr->ksymbol = track && !perf_missing_features.ksymbol;
1068         attr->bpf_event = track && !opts->no_bpf_event &&
1069                 !perf_missing_features.bpf_event;
1070
1071         if (opts->record_namespaces)
1072                 attr->namespaces  = track;
1073
1074         if (opts->record_switch_events)
1075                 attr->context_switch = track;
1076
1077         if (opts->sample_transaction)
1078                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1079
1080         if (opts->running_time) {
1081                 evsel->attr.read_format |=
1082                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1083                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1084         }
1085
1086         /*
1087          * XXX see the function comment above
1088          *
1089          * Disabling only independent events or group leaders,
1090          * keeping group members enabled.
1091          */
1092         if (perf_evsel__is_group_leader(evsel))
1093                 attr->disabled = 1;
1094
1095         /*
1096          * Setting enable_on_exec for independent events and
1097          * group leaders for traced executed by perf.
1098          */
1099         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1100                 !opts->initial_delay)
1101                 attr->enable_on_exec = 1;
1102
1103         if (evsel->immediate) {
1104                 attr->disabled = 0;
1105                 attr->enable_on_exec = 0;
1106         }
1107
1108         clockid = opts->clockid;
1109         if (opts->use_clockid) {
1110                 attr->use_clockid = 1;
1111                 attr->clockid = opts->clockid;
1112         }
1113
1114         if (evsel->precise_max)
1115                 attr->precise_ip = 3;
1116
1117         if (opts->all_user) {
1118                 attr->exclude_kernel = 1;
1119                 attr->exclude_user   = 0;
1120         }
1121
1122         if (opts->all_kernel) {
1123                 attr->exclude_kernel = 0;
1124                 attr->exclude_user   = 1;
1125         }
1126
1127         if (evsel->own_cpus || evsel->unit)
1128                 evsel->attr.read_format |= PERF_FORMAT_ID;
1129
1130         /*
1131          * Apply event specific term settings,
1132          * it overloads any global configuration.
1133          */
1134         apply_config_terms(evsel, opts, track);
1135
1136         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1137
1138         /* The --period option takes the precedence. */
1139         if (opts->period_set) {
1140                 if (opts->period)
1141                         perf_evsel__set_sample_bit(evsel, PERIOD);
1142                 else
1143                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1144         }
1145
1146         /*
1147          * For initial_delay, a dummy event is added implicitly.
1148          * The software event will trigger -EOPNOTSUPP error out,
1149          * if BRANCH_STACK bit is set.
1150          */
1151         if (opts->initial_delay && is_dummy_event(evsel))
1152                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1153 }
1154
1155 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1156 {
1157         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1158
1159         if (evsel->fd) {
1160                 int cpu, thread;
1161                 for (cpu = 0; cpu < ncpus; cpu++) {
1162                         for (thread = 0; thread < nthreads; thread++) {
1163                                 FD(evsel, cpu, thread) = -1;
1164                         }
1165                 }
1166         }
1167
1168         return evsel->fd != NULL ? 0 : -ENOMEM;
1169 }
1170
1171 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1172                           int ioc,  void *arg)
1173 {
1174         int cpu, thread;
1175
1176         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1177                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1178                         int fd = FD(evsel, cpu, thread),
1179                             err = ioctl(fd, ioc, arg);
1180
1181                         if (err)
1182                                 return err;
1183                 }
1184         }
1185
1186         return 0;
1187 }
1188
1189 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1190 {
1191         return perf_evsel__run_ioctl(evsel,
1192                                      PERF_EVENT_IOC_SET_FILTER,
1193                                      (void *)filter);
1194 }
1195
1196 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1197 {
1198         char *new_filter = strdup(filter);
1199
1200         if (new_filter != NULL) {
1201                 free(evsel->filter);
1202                 evsel->filter = new_filter;
1203                 return 0;
1204         }
1205
1206         return -1;
1207 }
1208
1209 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1210                                      const char *fmt, const char *filter)
1211 {
1212         char *new_filter;
1213
1214         if (evsel->filter == NULL)
1215                 return perf_evsel__set_filter(evsel, filter);
1216
1217         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1218                 free(evsel->filter);
1219                 evsel->filter = new_filter;
1220                 return 0;
1221         }
1222
1223         return -1;
1224 }
1225
1226 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1227 {
1228         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1229 }
1230
1231 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1232 {
1233         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1234 }
1235
1236 int perf_evsel__enable(struct perf_evsel *evsel)
1237 {
1238         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1239
1240         if (!err)
1241                 evsel->disabled = false;
1242
1243         return err;
1244 }
1245
1246 int perf_evsel__disable(struct perf_evsel *evsel)
1247 {
1248         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1249         /*
1250          * We mark it disabled here so that tools that disable a event can
1251          * ignore events after they disable it. I.e. the ring buffer may have
1252          * already a few more events queued up before the kernel got the stop
1253          * request.
1254          */
1255         if (!err)
1256                 evsel->disabled = true;
1257
1258         return err;
1259 }
1260
1261 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1262 {
1263         if (ncpus == 0 || nthreads == 0)
1264                 return 0;
1265
1266         if (evsel->system_wide)
1267                 nthreads = 1;
1268
1269         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1270         if (evsel->sample_id == NULL)
1271                 return -ENOMEM;
1272
1273         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1274         if (evsel->id == NULL) {
1275                 xyarray__delete(evsel->sample_id);
1276                 evsel->sample_id = NULL;
1277                 return -ENOMEM;
1278         }
1279
1280         return 0;
1281 }
1282
1283 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1284 {
1285         xyarray__delete(evsel->fd);
1286         evsel->fd = NULL;
1287 }
1288
1289 static void perf_evsel__free_id(struct perf_evsel *evsel)
1290 {
1291         xyarray__delete(evsel->sample_id);
1292         evsel->sample_id = NULL;
1293         zfree(&evsel->id);
1294         evsel->ids = 0;
1295 }
1296
1297 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1298 {
1299         struct perf_evsel_config_term *term, *h;
1300
1301         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1302                 list_del_init(&term->list);
1303                 free(term);
1304         }
1305 }
1306
1307 void perf_evsel__close_fd(struct perf_evsel *evsel)
1308 {
1309         int cpu, thread;
1310
1311         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1312                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1313                         close(FD(evsel, cpu, thread));
1314                         FD(evsel, cpu, thread) = -1;
1315                 }
1316 }
1317
1318 void perf_evsel__exit(struct perf_evsel *evsel)
1319 {
1320         assert(list_empty(&evsel->node));
1321         assert(evsel->evlist == NULL);
1322         perf_evsel__free_counts(evsel);
1323         perf_evsel__free_fd(evsel);
1324         perf_evsel__free_id(evsel);
1325         perf_evsel__free_config_terms(evsel);
1326         cgroup__put(evsel->cgrp);
1327         cpu_map__put(evsel->cpus);
1328         cpu_map__put(evsel->own_cpus);
1329         thread_map__put(evsel->threads);
1330         zfree(&evsel->group_name);
1331         zfree(&evsel->name);
1332         perf_evsel__object.fini(evsel);
1333 }
1334
1335 void perf_evsel__delete(struct perf_evsel *evsel)
1336 {
1337         perf_evsel__exit(evsel);
1338         free(evsel);
1339 }
1340
1341 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1342                                 struct perf_counts_values *count)
1343 {
1344         struct perf_counts_values tmp;
1345
1346         if (!evsel->prev_raw_counts)
1347                 return;
1348
1349         if (cpu == -1) {
1350                 tmp = evsel->prev_raw_counts->aggr;
1351                 evsel->prev_raw_counts->aggr = *count;
1352         } else {
1353                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1354                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1355         }
1356
1357         count->val = count->val - tmp.val;
1358         count->ena = count->ena - tmp.ena;
1359         count->run = count->run - tmp.run;
1360 }
1361
1362 void perf_counts_values__scale(struct perf_counts_values *count,
1363                                bool scale, s8 *pscaled)
1364 {
1365         s8 scaled = 0;
1366
1367         if (scale) {
1368                 if (count->run == 0) {
1369                         scaled = -1;
1370                         count->val = 0;
1371                 } else if (count->run < count->ena) {
1372                         scaled = 1;
1373                         count->val = (u64)((double) count->val * count->ena / count->run);
1374                 }
1375         }
1376
1377         if (pscaled)
1378                 *pscaled = scaled;
1379 }
1380
1381 static int perf_evsel__read_size(struct perf_evsel *evsel)
1382 {
1383         u64 read_format = evsel->attr.read_format;
1384         int entry = sizeof(u64); /* value */
1385         int size = 0;
1386         int nr = 1;
1387
1388         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1389                 size += sizeof(u64);
1390
1391         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1392                 size += sizeof(u64);
1393
1394         if (read_format & PERF_FORMAT_ID)
1395                 entry += sizeof(u64);
1396
1397         if (read_format & PERF_FORMAT_GROUP) {
1398                 nr = evsel->nr_members;
1399                 size += sizeof(u64);
1400         }
1401
1402         size += entry * nr;
1403         return size;
1404 }
1405
1406 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1407                      struct perf_counts_values *count)
1408 {
1409         size_t size = perf_evsel__read_size(evsel);
1410
1411         memset(count, 0, sizeof(*count));
1412
1413         if (FD(evsel, cpu, thread) < 0)
1414                 return -EINVAL;
1415
1416         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1417                 return -errno;
1418
1419         return 0;
1420 }
1421
1422 static int
1423 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1424 {
1425         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1426
1427         return perf_evsel__read(evsel, cpu, thread, count);
1428 }
1429
1430 static void
1431 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1432                       u64 val, u64 ena, u64 run)
1433 {
1434         struct perf_counts_values *count;
1435
1436         count = perf_counts(counter->counts, cpu, thread);
1437
1438         count->val    = val;
1439         count->ena    = ena;
1440         count->run    = run;
1441         count->loaded = true;
1442 }
1443
1444 static int
1445 perf_evsel__process_group_data(struct perf_evsel *leader,
1446                                int cpu, int thread, u64 *data)
1447 {
1448         u64 read_format = leader->attr.read_format;
1449         struct sample_read_value *v;
1450         u64 nr, ena = 0, run = 0, i;
1451
1452         nr = *data++;
1453
1454         if (nr != (u64) leader->nr_members)
1455                 return -EINVAL;
1456
1457         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1458                 ena = *data++;
1459
1460         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1461                 run = *data++;
1462
1463         v = (struct sample_read_value *) data;
1464
1465         perf_evsel__set_count(leader, cpu, thread,
1466                               v[0].value, ena, run);
1467
1468         for (i = 1; i < nr; i++) {
1469                 struct perf_evsel *counter;
1470
1471                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1472                 if (!counter)
1473                         return -EINVAL;
1474
1475                 perf_evsel__set_count(counter, cpu, thread,
1476                                       v[i].value, ena, run);
1477         }
1478
1479         return 0;
1480 }
1481
1482 static int
1483 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1484 {
1485         struct perf_stat_evsel *ps = leader->stats;
1486         u64 read_format = leader->attr.read_format;
1487         int size = perf_evsel__read_size(leader);
1488         u64 *data = ps->group_data;
1489
1490         if (!(read_format & PERF_FORMAT_ID))
1491                 return -EINVAL;
1492
1493         if (!perf_evsel__is_group_leader(leader))
1494                 return -EINVAL;
1495
1496         if (!data) {
1497                 data = zalloc(size);
1498                 if (!data)
1499                         return -ENOMEM;
1500
1501                 ps->group_data = data;
1502         }
1503
1504         if (FD(leader, cpu, thread) < 0)
1505                 return -EINVAL;
1506
1507         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1508                 return -errno;
1509
1510         return perf_evsel__process_group_data(leader, cpu, thread, data);
1511 }
1512
1513 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1514 {
1515         u64 read_format = evsel->attr.read_format;
1516
1517         if (read_format & PERF_FORMAT_GROUP)
1518                 return perf_evsel__read_group(evsel, cpu, thread);
1519         else
1520                 return perf_evsel__read_one(evsel, cpu, thread);
1521 }
1522
1523 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1524                               int cpu, int thread, bool scale)
1525 {
1526         struct perf_counts_values count;
1527         size_t nv = scale ? 3 : 1;
1528
1529         if (FD(evsel, cpu, thread) < 0)
1530                 return -EINVAL;
1531
1532         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1533                 return -ENOMEM;
1534
1535         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1536                 return -errno;
1537
1538         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1539         perf_counts_values__scale(&count, scale, NULL);
1540         *perf_counts(evsel->counts, cpu, thread) = count;
1541         return 0;
1542 }
1543
1544 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1545 {
1546         struct perf_evsel *leader = evsel->leader;
1547         int fd;
1548
1549         if (perf_evsel__is_group_leader(evsel))
1550                 return -1;
1551
1552         /*
1553          * Leader must be already processed/open,
1554          * if not it's a bug.
1555          */
1556         BUG_ON(!leader->fd);
1557
1558         fd = FD(leader, cpu, thread);
1559         BUG_ON(fd == -1);
1560
1561         return fd;
1562 }
1563
1564 struct bit_names {
1565         int bit;
1566         const char *name;
1567 };
1568
1569 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1570 {
1571         bool first_bit = true;
1572         int i = 0;
1573
1574         do {
1575                 if (value & bits[i].bit) {
1576                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1577                         first_bit = false;
1578                 }
1579         } while (bits[++i].name != NULL);
1580 }
1581
1582 static void __p_sample_type(char *buf, size_t size, u64 value)
1583 {
1584 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1585         struct bit_names bits[] = {
1586                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1587                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1588                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1589                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1590                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1591                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1592                 { .name = NULL, }
1593         };
1594 #undef bit_name
1595         __p_bits(buf, size, value, bits);
1596 }
1597
1598 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1599 {
1600 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1601         struct bit_names bits[] = {
1602                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1603                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1604                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1605                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1606                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1607                 { .name = NULL, }
1608         };
1609 #undef bit_name
1610         __p_bits(buf, size, value, bits);
1611 }
1612
1613 static void __p_read_format(char *buf, size_t size, u64 value)
1614 {
1615 #define bit_name(n) { PERF_FORMAT_##n, #n }
1616         struct bit_names bits[] = {
1617                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1618                 bit_name(ID), bit_name(GROUP),
1619                 { .name = NULL, }
1620         };
1621 #undef bit_name
1622         __p_bits(buf, size, value, bits);
1623 }
1624
1625 #define BUF_SIZE                1024
1626
1627 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1628 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1629 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1630 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1631 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1632 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1633
1634 #define PRINT_ATTRn(_n, _f, _p)                         \
1635 do {                                                    \
1636         if (attr->_f) {                                 \
1637                 _p(attr->_f);                           \
1638                 ret += attr__fprintf(fp, _n, buf, priv);\
1639         }                                               \
1640 } while (0)
1641
1642 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1643
1644 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1645                              attr__fprintf_f attr__fprintf, void *priv)
1646 {
1647         char buf[BUF_SIZE];
1648         int ret = 0;
1649
1650         PRINT_ATTRf(type, p_unsigned);
1651         PRINT_ATTRf(size, p_unsigned);
1652         PRINT_ATTRf(config, p_hex);
1653         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1654         PRINT_ATTRf(sample_type, p_sample_type);
1655         PRINT_ATTRf(read_format, p_read_format);
1656
1657         PRINT_ATTRf(disabled, p_unsigned);
1658         PRINT_ATTRf(inherit, p_unsigned);
1659         PRINT_ATTRf(pinned, p_unsigned);
1660         PRINT_ATTRf(exclusive, p_unsigned);
1661         PRINT_ATTRf(exclude_user, p_unsigned);
1662         PRINT_ATTRf(exclude_kernel, p_unsigned);
1663         PRINT_ATTRf(exclude_hv, p_unsigned);
1664         PRINT_ATTRf(exclude_idle, p_unsigned);
1665         PRINT_ATTRf(mmap, p_unsigned);
1666         PRINT_ATTRf(comm, p_unsigned);
1667         PRINT_ATTRf(freq, p_unsigned);
1668         PRINT_ATTRf(inherit_stat, p_unsigned);
1669         PRINT_ATTRf(enable_on_exec, p_unsigned);
1670         PRINT_ATTRf(task, p_unsigned);
1671         PRINT_ATTRf(watermark, p_unsigned);
1672         PRINT_ATTRf(precise_ip, p_unsigned);
1673         PRINT_ATTRf(mmap_data, p_unsigned);
1674         PRINT_ATTRf(sample_id_all, p_unsigned);
1675         PRINT_ATTRf(exclude_host, p_unsigned);
1676         PRINT_ATTRf(exclude_guest, p_unsigned);
1677         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1678         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1679         PRINT_ATTRf(mmap2, p_unsigned);
1680         PRINT_ATTRf(comm_exec, p_unsigned);
1681         PRINT_ATTRf(use_clockid, p_unsigned);
1682         PRINT_ATTRf(context_switch, p_unsigned);
1683         PRINT_ATTRf(write_backward, p_unsigned);
1684         PRINT_ATTRf(namespaces, p_unsigned);
1685         PRINT_ATTRf(ksymbol, p_unsigned);
1686         PRINT_ATTRf(bpf_event, p_unsigned);
1687
1688         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1689         PRINT_ATTRf(bp_type, p_unsigned);
1690         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1691         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1692         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1693         PRINT_ATTRf(sample_regs_user, p_hex);
1694         PRINT_ATTRf(sample_stack_user, p_unsigned);
1695         PRINT_ATTRf(clockid, p_signed);
1696         PRINT_ATTRf(sample_regs_intr, p_hex);
1697         PRINT_ATTRf(aux_watermark, p_unsigned);
1698         PRINT_ATTRf(sample_max_stack, p_unsigned);
1699
1700         return ret;
1701 }
1702
1703 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1704                                 void *priv __maybe_unused)
1705 {
1706         return fprintf(fp, "  %-32s %s\n", name, val);
1707 }
1708
1709 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1710                                   int nr_cpus, int nr_threads,
1711                                   int thread_idx)
1712 {
1713         for (int cpu = 0; cpu < nr_cpus; cpu++)
1714                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1715                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1716 }
1717
1718 static int update_fds(struct perf_evsel *evsel,
1719                       int nr_cpus, int cpu_idx,
1720                       int nr_threads, int thread_idx)
1721 {
1722         struct perf_evsel *pos;
1723
1724         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1725                 return -EINVAL;
1726
1727         evlist__for_each_entry(evsel->evlist, pos) {
1728                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1729
1730                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1731
1732                 /*
1733                  * Since fds for next evsel has not been created,
1734                  * there is no need to iterate whole event list.
1735                  */
1736                 if (pos == evsel)
1737                         break;
1738         }
1739         return 0;
1740 }
1741
1742 static bool ignore_missing_thread(struct perf_evsel *evsel,
1743                                   int nr_cpus, int cpu,
1744                                   struct thread_map *threads,
1745                                   int thread, int err)
1746 {
1747         pid_t ignore_pid = thread_map__pid(threads, thread);
1748
1749         if (!evsel->ignore_missing_thread)
1750                 return false;
1751
1752         /* The system wide setup does not work with threads. */
1753         if (evsel->system_wide)
1754                 return false;
1755
1756         /* The -ESRCH is perf event syscall errno for pid's not found. */
1757         if (err != -ESRCH)
1758                 return false;
1759
1760         /* If there's only one thread, let it fail. */
1761         if (threads->nr == 1)
1762                 return false;
1763
1764         /*
1765          * We should remove fd for missing_thread first
1766          * because thread_map__remove() will decrease threads->nr.
1767          */
1768         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1769                 return false;
1770
1771         if (thread_map__remove(threads, thread))
1772                 return false;
1773
1774         pr_warning("WARNING: Ignored open failure for pid %d\n",
1775                    ignore_pid);
1776         return true;
1777 }
1778
1779 static void display_attr(struct perf_event_attr *attr)
1780 {
1781         if (verbose >= 2) {
1782                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1783                 fprintf(stderr, "perf_event_attr:\n");
1784                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1785                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1786         }
1787 }
1788
1789 static int perf_event_open(struct perf_evsel *evsel,
1790                            pid_t pid, int cpu, int group_fd,
1791                            unsigned long flags)
1792 {
1793         int precise_ip = evsel->attr.precise_ip;
1794         int fd;
1795
1796         while (1) {
1797                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1798                           pid, cpu, group_fd, flags);
1799
1800                 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1801                 if (fd >= 0)
1802                         break;
1803
1804                 /* Do not try less precise if not requested. */
1805                 if (!evsel->precise_max)
1806                         break;
1807
1808                 /*
1809                  * We tried all the precise_ip values, and it's
1810                  * still failing, so leave it to standard fallback.
1811                  */
1812                 if (!evsel->attr.precise_ip) {
1813                         evsel->attr.precise_ip = precise_ip;
1814                         break;
1815                 }
1816
1817                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1818                 evsel->attr.precise_ip--;
1819                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1820                 display_attr(&evsel->attr);
1821         }
1822
1823         return fd;
1824 }
1825
1826 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1827                      struct thread_map *threads)
1828 {
1829         int cpu, thread, nthreads;
1830         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1831         int pid = -1, err;
1832         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1833
1834         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1835                 return -EINVAL;
1836
1837         if (cpus == NULL) {
1838                 static struct cpu_map *empty_cpu_map;
1839
1840                 if (empty_cpu_map == NULL) {
1841                         empty_cpu_map = cpu_map__dummy_new();
1842                         if (empty_cpu_map == NULL)
1843                                 return -ENOMEM;
1844                 }
1845
1846                 cpus = empty_cpu_map;
1847         }
1848
1849         if (threads == NULL) {
1850                 static struct thread_map *empty_thread_map;
1851
1852                 if (empty_thread_map == NULL) {
1853                         empty_thread_map = thread_map__new_by_tid(-1);
1854                         if (empty_thread_map == NULL)
1855                                 return -ENOMEM;
1856                 }
1857
1858                 threads = empty_thread_map;
1859         }
1860
1861         if (evsel->system_wide)
1862                 nthreads = 1;
1863         else
1864                 nthreads = threads->nr;
1865
1866         if (evsel->fd == NULL &&
1867             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1868                 return -ENOMEM;
1869
1870         if (evsel->cgrp) {
1871                 flags |= PERF_FLAG_PID_CGROUP;
1872                 pid = evsel->cgrp->fd;
1873         }
1874
1875 fallback_missing_features:
1876         if (perf_missing_features.clockid_wrong)
1877                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1878         if (perf_missing_features.clockid) {
1879                 evsel->attr.use_clockid = 0;
1880                 evsel->attr.clockid = 0;
1881         }
1882         if (perf_missing_features.cloexec)
1883                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1884         if (perf_missing_features.mmap2)
1885                 evsel->attr.mmap2 = 0;
1886         if (perf_missing_features.exclude_guest)
1887                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1888         if (perf_missing_features.lbr_flags)
1889                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1890                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1891         if (perf_missing_features.group_read && evsel->attr.inherit)
1892                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1893         if (perf_missing_features.ksymbol)
1894                 evsel->attr.ksymbol = 0;
1895         if (perf_missing_features.bpf_event)
1896                 evsel->attr.bpf_event = 0;
1897 retry_sample_id:
1898         if (perf_missing_features.sample_id_all)
1899                 evsel->attr.sample_id_all = 0;
1900
1901         display_attr(&evsel->attr);
1902
1903         for (cpu = 0; cpu < cpus->nr; cpu++) {
1904
1905                 for (thread = 0; thread < nthreads; thread++) {
1906                         int fd, group_fd;
1907
1908                         if (!evsel->cgrp && !evsel->system_wide)
1909                                 pid = thread_map__pid(threads, thread);
1910
1911                         group_fd = get_group_fd(evsel, cpu, thread);
1912 retry_open:
1913                         test_attr__ready();
1914
1915                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1916                                              group_fd, flags);
1917
1918                         FD(evsel, cpu, thread) = fd;
1919
1920                         if (fd < 0) {
1921                                 err = -errno;
1922
1923                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1924                                         /*
1925                                          * We just removed 1 thread, so take a step
1926                                          * back on thread index and lower the upper
1927                                          * nthreads limit.
1928                                          */
1929                                         nthreads--;
1930                                         thread--;
1931
1932                                         /* ... and pretend like nothing have happened. */
1933                                         err = 0;
1934                                         continue;
1935                                 }
1936
1937                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1938                                           err);
1939                                 goto try_fallback;
1940                         }
1941
1942                         pr_debug2(" = %d\n", fd);
1943
1944                         if (evsel->bpf_fd >= 0) {
1945                                 int evt_fd = fd;
1946                                 int bpf_fd = evsel->bpf_fd;
1947
1948                                 err = ioctl(evt_fd,
1949                                             PERF_EVENT_IOC_SET_BPF,
1950                                             bpf_fd);
1951                                 if (err && errno != EEXIST) {
1952                                         pr_err("failed to attach bpf fd %d: %s\n",
1953                                                bpf_fd, strerror(errno));
1954                                         err = -EINVAL;
1955                                         goto out_close;
1956                                 }
1957                         }
1958
1959                         set_rlimit = NO_CHANGE;
1960
1961                         /*
1962                          * If we succeeded but had to kill clockid, fail and
1963                          * have perf_evsel__open_strerror() print us a nice
1964                          * error.
1965                          */
1966                         if (perf_missing_features.clockid ||
1967                             perf_missing_features.clockid_wrong) {
1968                                 err = -EINVAL;
1969                                 goto out_close;
1970                         }
1971                 }
1972         }
1973
1974         return 0;
1975
1976 try_fallback:
1977         /*
1978          * perf stat needs between 5 and 22 fds per CPU. When we run out
1979          * of them try to increase the limits.
1980          */
1981         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1982                 struct rlimit l;
1983                 int old_errno = errno;
1984
1985                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1986                         if (set_rlimit == NO_CHANGE)
1987                                 l.rlim_cur = l.rlim_max;
1988                         else {
1989                                 l.rlim_cur = l.rlim_max + 1000;
1990                                 l.rlim_max = l.rlim_cur;
1991                         }
1992                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1993                                 set_rlimit++;
1994                                 errno = old_errno;
1995                                 goto retry_open;
1996                         }
1997                 }
1998                 errno = old_errno;
1999         }
2000
2001         if (err != -EINVAL || cpu > 0 || thread > 0)
2002                 goto out_close;
2003
2004         /*
2005          * Must probe features in the order they were added to the
2006          * perf_event_attr interface.
2007          */
2008         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
2009                 perf_missing_features.bpf_event = true;
2010                 pr_debug2("switching off bpf_event\n");
2011                 goto fallback_missing_features;
2012         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
2013                 perf_missing_features.ksymbol = true;
2014                 pr_debug2("switching off ksymbol\n");
2015                 goto fallback_missing_features;
2016         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
2017                 perf_missing_features.write_backward = true;
2018                 pr_debug2("switching off write_backward\n");
2019                 goto out_close;
2020         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2021                 perf_missing_features.clockid_wrong = true;
2022                 pr_debug2("switching off clockid\n");
2023                 goto fallback_missing_features;
2024         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2025                 perf_missing_features.clockid = true;
2026                 pr_debug2("switching off use_clockid\n");
2027                 goto fallback_missing_features;
2028         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2029                 perf_missing_features.cloexec = true;
2030                 pr_debug2("switching off cloexec flag\n");
2031                 goto fallback_missing_features;
2032         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2033                 perf_missing_features.mmap2 = true;
2034                 pr_debug2("switching off mmap2\n");
2035                 goto fallback_missing_features;
2036         } else if (!perf_missing_features.exclude_guest &&
2037                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2038                 perf_missing_features.exclude_guest = true;
2039                 pr_debug2("switching off exclude_guest, exclude_host\n");
2040                 goto fallback_missing_features;
2041         } else if (!perf_missing_features.sample_id_all) {
2042                 perf_missing_features.sample_id_all = true;
2043                 pr_debug2("switching off sample_id_all\n");
2044                 goto retry_sample_id;
2045         } else if (!perf_missing_features.lbr_flags &&
2046                         (evsel->attr.branch_sample_type &
2047                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2048                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2049                 perf_missing_features.lbr_flags = true;
2050                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2051                 goto fallback_missing_features;
2052         } else if (!perf_missing_features.group_read &&
2053                     evsel->attr.inherit &&
2054                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2055                    perf_evsel__is_group_leader(evsel)) {
2056                 perf_missing_features.group_read = true;
2057                 pr_debug2("switching off group read\n");
2058                 goto fallback_missing_features;
2059         }
2060 out_close:
2061         if (err)
2062                 threads->err_thread = thread;
2063
2064         do {
2065                 while (--thread >= 0) {
2066                         close(FD(evsel, cpu, thread));
2067                         FD(evsel, cpu, thread) = -1;
2068                 }
2069                 thread = nthreads;
2070         } while (--cpu >= 0);
2071         return err;
2072 }
2073
2074 void perf_evsel__close(struct perf_evsel *evsel)
2075 {
2076         if (evsel->fd == NULL)
2077                 return;
2078
2079         perf_evsel__close_fd(evsel);
2080         perf_evsel__free_fd(evsel);
2081         perf_evsel__free_id(evsel);
2082 }
2083
2084 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2085                              struct cpu_map *cpus)
2086 {
2087         return perf_evsel__open(evsel, cpus, NULL);
2088 }
2089
2090 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2091                                 struct thread_map *threads)
2092 {
2093         return perf_evsel__open(evsel, NULL, threads);
2094 }
2095
2096 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2097                                        const union perf_event *event,
2098                                        struct perf_sample *sample)
2099 {
2100         u64 type = evsel->attr.sample_type;
2101         const u64 *array = event->sample.array;
2102         bool swapped = evsel->needs_swap;
2103         union u64_swap u;
2104
2105         array += ((event->header.size -
2106                    sizeof(event->header)) / sizeof(u64)) - 1;
2107
2108         if (type & PERF_SAMPLE_IDENTIFIER) {
2109                 sample->id = *array;
2110                 array--;
2111         }
2112
2113         if (type & PERF_SAMPLE_CPU) {
2114                 u.val64 = *array;
2115                 if (swapped) {
2116                         /* undo swap of u64, then swap on individual u32s */
2117                         u.val64 = bswap_64(u.val64);
2118                         u.val32[0] = bswap_32(u.val32[0]);
2119                 }
2120
2121                 sample->cpu = u.val32[0];
2122                 array--;
2123         }
2124
2125         if (type & PERF_SAMPLE_STREAM_ID) {
2126                 sample->stream_id = *array;
2127                 array--;
2128         }
2129
2130         if (type & PERF_SAMPLE_ID) {
2131                 sample->id = *array;
2132                 array--;
2133         }
2134
2135         if (type & PERF_SAMPLE_TIME) {
2136                 sample->time = *array;
2137                 array--;
2138         }
2139
2140         if (type & PERF_SAMPLE_TID) {
2141                 u.val64 = *array;
2142                 if (swapped) {
2143                         /* undo swap of u64, then swap on individual u32s */
2144                         u.val64 = bswap_64(u.val64);
2145                         u.val32[0] = bswap_32(u.val32[0]);
2146                         u.val32[1] = bswap_32(u.val32[1]);
2147                 }
2148
2149                 sample->pid = u.val32[0];
2150                 sample->tid = u.val32[1];
2151                 array--;
2152         }
2153
2154         return 0;
2155 }
2156
2157 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2158                             u64 size)
2159 {
2160         return size > max_size || offset + size > endp;
2161 }
2162
2163 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2164         do {                                                            \
2165                 if (overflow(endp, (max_size), (offset), (size)))       \
2166                         return -EFAULT;                                 \
2167         } while (0)
2168
2169 #define OVERFLOW_CHECK_u64(offset) \
2170         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2171
2172 static int
2173 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2174 {
2175         /*
2176          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2177          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2178          * check the format does not go past the end of the event.
2179          */
2180         if (sample_size + sizeof(event->header) > event->header.size)
2181                 return -EFAULT;
2182
2183         return 0;
2184 }
2185
2186 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2187                              struct perf_sample *data)
2188 {
2189         u64 type = evsel->attr.sample_type;
2190         bool swapped = evsel->needs_swap;
2191         const u64 *array;
2192         u16 max_size = event->header.size;
2193         const void *endp = (void *)event + max_size;
2194         u64 sz;
2195
2196         /*
2197          * used for cross-endian analysis. See git commit 65014ab3
2198          * for why this goofiness is needed.
2199          */
2200         union u64_swap u;
2201
2202         memset(data, 0, sizeof(*data));
2203         data->cpu = data->pid = data->tid = -1;
2204         data->stream_id = data->id = data->time = -1ULL;
2205         data->period = evsel->attr.sample_period;
2206         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2207         data->misc    = event->header.misc;
2208         data->id = -1ULL;
2209         data->data_src = PERF_MEM_DATA_SRC_NONE;
2210
2211         if (event->header.type != PERF_RECORD_SAMPLE) {
2212                 if (!evsel->attr.sample_id_all)
2213                         return 0;
2214                 return perf_evsel__parse_id_sample(evsel, event, data);
2215         }
2216
2217         array = event->sample.array;
2218
2219         if (perf_event__check_size(event, evsel->sample_size))
2220                 return -EFAULT;
2221
2222         if (type & PERF_SAMPLE_IDENTIFIER) {
2223                 data->id = *array;
2224                 array++;
2225         }
2226
2227         if (type & PERF_SAMPLE_IP) {
2228                 data->ip = *array;
2229                 array++;
2230         }
2231
2232         if (type & PERF_SAMPLE_TID) {
2233                 u.val64 = *array;
2234                 if (swapped) {
2235                         /* undo swap of u64, then swap on individual u32s */
2236                         u.val64 = bswap_64(u.val64);
2237                         u.val32[0] = bswap_32(u.val32[0]);
2238                         u.val32[1] = bswap_32(u.val32[1]);
2239                 }
2240
2241                 data->pid = u.val32[0];
2242                 data->tid = u.val32[1];
2243                 array++;
2244         }
2245
2246         if (type & PERF_SAMPLE_TIME) {
2247                 data->time = *array;
2248                 array++;
2249         }
2250
2251         if (type & PERF_SAMPLE_ADDR) {
2252                 data->addr = *array;
2253                 array++;
2254         }
2255
2256         if (type & PERF_SAMPLE_ID) {
2257                 data->id = *array;
2258                 array++;
2259         }
2260
2261         if (type & PERF_SAMPLE_STREAM_ID) {
2262                 data->stream_id = *array;
2263                 array++;
2264         }
2265
2266         if (type & PERF_SAMPLE_CPU) {
2267
2268                 u.val64 = *array;
2269                 if (swapped) {
2270                         /* undo swap of u64, then swap on individual u32s */
2271                         u.val64 = bswap_64(u.val64);
2272                         u.val32[0] = bswap_32(u.val32[0]);
2273                 }
2274
2275                 data->cpu = u.val32[0];
2276                 array++;
2277         }
2278
2279         if (type & PERF_SAMPLE_PERIOD) {
2280                 data->period = *array;
2281                 array++;
2282         }
2283
2284         if (type & PERF_SAMPLE_READ) {
2285                 u64 read_format = evsel->attr.read_format;
2286
2287                 OVERFLOW_CHECK_u64(array);
2288                 if (read_format & PERF_FORMAT_GROUP)
2289                         data->read.group.nr = *array;
2290                 else
2291                         data->read.one.value = *array;
2292
2293                 array++;
2294
2295                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2296                         OVERFLOW_CHECK_u64(array);
2297                         data->read.time_enabled = *array;
2298                         array++;
2299                 }
2300
2301                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2302                         OVERFLOW_CHECK_u64(array);
2303                         data->read.time_running = *array;
2304                         array++;
2305                 }
2306
2307                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2308                 if (read_format & PERF_FORMAT_GROUP) {
2309                         const u64 max_group_nr = UINT64_MAX /
2310                                         sizeof(struct sample_read_value);
2311
2312                         if (data->read.group.nr > max_group_nr)
2313                                 return -EFAULT;
2314                         sz = data->read.group.nr *
2315                              sizeof(struct sample_read_value);
2316                         OVERFLOW_CHECK(array, sz, max_size);
2317                         data->read.group.values =
2318                                         (struct sample_read_value *)array;
2319                         array = (void *)array + sz;
2320                 } else {
2321                         OVERFLOW_CHECK_u64(array);
2322                         data->read.one.id = *array;
2323                         array++;
2324                 }
2325         }
2326
2327         if (evsel__has_callchain(evsel)) {
2328                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2329
2330                 OVERFLOW_CHECK_u64(array);
2331                 data->callchain = (struct ip_callchain *)array++;
2332                 if (data->callchain->nr > max_callchain_nr)
2333                         return -EFAULT;
2334                 sz = data->callchain->nr * sizeof(u64);
2335                 OVERFLOW_CHECK(array, sz, max_size);
2336                 array = (void *)array + sz;
2337         }
2338
2339         if (type & PERF_SAMPLE_RAW) {
2340                 OVERFLOW_CHECK_u64(array);
2341                 u.val64 = *array;
2342
2343                 /*
2344                  * Undo swap of u64, then swap on individual u32s,
2345                  * get the size of the raw area and undo all of the
2346                  * swap. The pevent interface handles endianity by
2347                  * itself.
2348                  */
2349                 if (swapped) {
2350                         u.val64 = bswap_64(u.val64);
2351                         u.val32[0] = bswap_32(u.val32[0]);
2352                         u.val32[1] = bswap_32(u.val32[1]);
2353                 }
2354                 data->raw_size = u.val32[0];
2355
2356                 /*
2357                  * The raw data is aligned on 64bits including the
2358                  * u32 size, so it's safe to use mem_bswap_64.
2359                  */
2360                 if (swapped)
2361                         mem_bswap_64((void *) array, data->raw_size);
2362
2363                 array = (void *)array + sizeof(u32);
2364
2365                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2366                 data->raw_data = (void *)array;
2367                 array = (void *)array + data->raw_size;
2368         }
2369
2370         if (type & PERF_SAMPLE_BRANCH_STACK) {
2371                 const u64 max_branch_nr = UINT64_MAX /
2372                                           sizeof(struct branch_entry);
2373
2374                 OVERFLOW_CHECK_u64(array);
2375                 data->branch_stack = (struct branch_stack *)array++;
2376
2377                 if (data->branch_stack->nr > max_branch_nr)
2378                         return -EFAULT;
2379                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2380                 OVERFLOW_CHECK(array, sz, max_size);
2381                 array = (void *)array + sz;
2382         }
2383
2384         if (type & PERF_SAMPLE_REGS_USER) {
2385                 OVERFLOW_CHECK_u64(array);
2386                 data->user_regs.abi = *array;
2387                 array++;
2388
2389                 if (data->user_regs.abi) {
2390                         u64 mask = evsel->attr.sample_regs_user;
2391
2392                         sz = hweight64(mask) * sizeof(u64);
2393                         OVERFLOW_CHECK(array, sz, max_size);
2394                         data->user_regs.mask = mask;
2395                         data->user_regs.regs = (u64 *)array;
2396                         array = (void *)array + sz;
2397                 }
2398         }
2399
2400         if (type & PERF_SAMPLE_STACK_USER) {
2401                 OVERFLOW_CHECK_u64(array);
2402                 sz = *array++;
2403
2404                 data->user_stack.offset = ((char *)(array - 1)
2405                                           - (char *) event);
2406
2407                 if (!sz) {
2408                         data->user_stack.size = 0;
2409                 } else {
2410                         OVERFLOW_CHECK(array, sz, max_size);
2411                         data->user_stack.data = (char *)array;
2412                         array = (void *)array + sz;
2413                         OVERFLOW_CHECK_u64(array);
2414                         data->user_stack.size = *array++;
2415                         if (WARN_ONCE(data->user_stack.size > sz,
2416                                       "user stack dump failure\n"))
2417                                 return -EFAULT;
2418                 }
2419         }
2420
2421         if (type & PERF_SAMPLE_WEIGHT) {
2422                 OVERFLOW_CHECK_u64(array);
2423                 data->weight = *array;
2424                 array++;
2425         }
2426
2427         if (type & PERF_SAMPLE_DATA_SRC) {
2428                 OVERFLOW_CHECK_u64(array);
2429                 data->data_src = *array;
2430                 array++;
2431         }
2432
2433         if (type & PERF_SAMPLE_TRANSACTION) {
2434                 OVERFLOW_CHECK_u64(array);
2435                 data->transaction = *array;
2436                 array++;
2437         }
2438
2439         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2440         if (type & PERF_SAMPLE_REGS_INTR) {
2441                 OVERFLOW_CHECK_u64(array);
2442                 data->intr_regs.abi = *array;
2443                 array++;
2444
2445                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2446                         u64 mask = evsel->attr.sample_regs_intr;
2447
2448                         sz = hweight64(mask) * sizeof(u64);
2449                         OVERFLOW_CHECK(array, sz, max_size);
2450                         data->intr_regs.mask = mask;
2451                         data->intr_regs.regs = (u64 *)array;
2452                         array = (void *)array + sz;
2453                 }
2454         }
2455
2456         data->phys_addr = 0;
2457         if (type & PERF_SAMPLE_PHYS_ADDR) {
2458                 data->phys_addr = *array;
2459                 array++;
2460         }
2461
2462         return 0;
2463 }
2464
2465 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2466                                        union perf_event *event,
2467                                        u64 *timestamp)
2468 {
2469         u64 type = evsel->attr.sample_type;
2470         const u64 *array;
2471
2472         if (!(type & PERF_SAMPLE_TIME))
2473                 return -1;
2474
2475         if (event->header.type != PERF_RECORD_SAMPLE) {
2476                 struct perf_sample data = {
2477                         .time = -1ULL,
2478                 };
2479
2480                 if (!evsel->attr.sample_id_all)
2481                         return -1;
2482                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2483                         return -1;
2484
2485                 *timestamp = data.time;
2486                 return 0;
2487         }
2488
2489         array = event->sample.array;
2490
2491         if (perf_event__check_size(event, evsel->sample_size))
2492                 return -EFAULT;
2493
2494         if (type & PERF_SAMPLE_IDENTIFIER)
2495                 array++;
2496
2497         if (type & PERF_SAMPLE_IP)
2498                 array++;
2499
2500         if (type & PERF_SAMPLE_TID)
2501                 array++;
2502
2503         if (type & PERF_SAMPLE_TIME)
2504                 *timestamp = *array;
2505
2506         return 0;
2507 }
2508
2509 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2510                                      u64 read_format)
2511 {
2512         size_t sz, result = sizeof(struct sample_event);
2513
2514         if (type & PERF_SAMPLE_IDENTIFIER)
2515                 result += sizeof(u64);
2516
2517         if (type & PERF_SAMPLE_IP)
2518                 result += sizeof(u64);
2519
2520         if (type & PERF_SAMPLE_TID)
2521                 result += sizeof(u64);
2522
2523         if (type & PERF_SAMPLE_TIME)
2524                 result += sizeof(u64);
2525
2526         if (type & PERF_SAMPLE_ADDR)
2527                 result += sizeof(u64);
2528
2529         if (type & PERF_SAMPLE_ID)
2530                 result += sizeof(u64);
2531
2532         if (type & PERF_SAMPLE_STREAM_ID)
2533                 result += sizeof(u64);
2534
2535         if (type & PERF_SAMPLE_CPU)
2536                 result += sizeof(u64);
2537
2538         if (type & PERF_SAMPLE_PERIOD)
2539                 result += sizeof(u64);
2540
2541         if (type & PERF_SAMPLE_READ) {
2542                 result += sizeof(u64);
2543                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2544                         result += sizeof(u64);
2545                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2546                         result += sizeof(u64);
2547                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2548                 if (read_format & PERF_FORMAT_GROUP) {
2549                         sz = sample->read.group.nr *
2550                              sizeof(struct sample_read_value);
2551                         result += sz;
2552                 } else {
2553                         result += sizeof(u64);
2554                 }
2555         }
2556
2557         if (type & PERF_SAMPLE_CALLCHAIN) {
2558                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2559                 result += sz;
2560         }
2561
2562         if (type & PERF_SAMPLE_RAW) {
2563                 result += sizeof(u32);
2564                 result += sample->raw_size;
2565         }
2566
2567         if (type & PERF_SAMPLE_BRANCH_STACK) {
2568                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2569                 sz += sizeof(u64);
2570                 result += sz;
2571         }
2572
2573         if (type & PERF_SAMPLE_REGS_USER) {
2574                 if (sample->user_regs.abi) {
2575                         result += sizeof(u64);
2576                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2577                         result += sz;
2578                 } else {
2579                         result += sizeof(u64);
2580                 }
2581         }
2582
2583         if (type & PERF_SAMPLE_STACK_USER) {
2584                 sz = sample->user_stack.size;
2585                 result += sizeof(u64);
2586                 if (sz) {
2587                         result += sz;
2588                         result += sizeof(u64);
2589                 }
2590         }
2591
2592         if (type & PERF_SAMPLE_WEIGHT)
2593                 result += sizeof(u64);
2594
2595         if (type & PERF_SAMPLE_DATA_SRC)
2596                 result += sizeof(u64);
2597
2598         if (type & PERF_SAMPLE_TRANSACTION)
2599                 result += sizeof(u64);
2600
2601         if (type & PERF_SAMPLE_REGS_INTR) {
2602                 if (sample->intr_regs.abi) {
2603                         result += sizeof(u64);
2604                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2605                         result += sz;
2606                 } else {
2607                         result += sizeof(u64);
2608                 }
2609         }
2610
2611         if (type & PERF_SAMPLE_PHYS_ADDR)
2612                 result += sizeof(u64);
2613
2614         return result;
2615 }
2616
2617 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2618                                   u64 read_format,
2619                                   const struct perf_sample *sample)
2620 {
2621         u64 *array;
2622         size_t sz;
2623         /*
2624          * used for cross-endian analysis. See git commit 65014ab3
2625          * for why this goofiness is needed.
2626          */
2627         union u64_swap u;
2628
2629         array = event->sample.array;
2630
2631         if (type & PERF_SAMPLE_IDENTIFIER) {
2632                 *array = sample->id;
2633                 array++;
2634         }
2635
2636         if (type & PERF_SAMPLE_IP) {
2637                 *array = sample->ip;
2638                 array++;
2639         }
2640
2641         if (type & PERF_SAMPLE_TID) {
2642                 u.val32[0] = sample->pid;
2643                 u.val32[1] = sample->tid;
2644                 *array = u.val64;
2645                 array++;
2646         }
2647
2648         if (type & PERF_SAMPLE_TIME) {
2649                 *array = sample->time;
2650                 array++;
2651         }
2652
2653         if (type & PERF_SAMPLE_ADDR) {
2654                 *array = sample->addr;
2655                 array++;
2656         }
2657
2658         if (type & PERF_SAMPLE_ID) {
2659                 *array = sample->id;
2660                 array++;
2661         }
2662
2663         if (type & PERF_SAMPLE_STREAM_ID) {
2664                 *array = sample->stream_id;
2665                 array++;
2666         }
2667
2668         if (type & PERF_SAMPLE_CPU) {
2669                 u.val32[0] = sample->cpu;
2670                 u.val32[1] = 0;
2671                 *array = u.val64;
2672                 array++;
2673         }
2674
2675         if (type & PERF_SAMPLE_PERIOD) {
2676                 *array = sample->period;
2677                 array++;
2678         }
2679
2680         if (type & PERF_SAMPLE_READ) {
2681                 if (read_format & PERF_FORMAT_GROUP)
2682                         *array = sample->read.group.nr;
2683                 else
2684                         *array = sample->read.one.value;
2685                 array++;
2686
2687                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2688                         *array = sample->read.time_enabled;
2689                         array++;
2690                 }
2691
2692                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2693                         *array = sample->read.time_running;
2694                         array++;
2695                 }
2696
2697                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2698                 if (read_format & PERF_FORMAT_GROUP) {
2699                         sz = sample->read.group.nr *
2700                              sizeof(struct sample_read_value);
2701                         memcpy(array, sample->read.group.values, sz);
2702                         array = (void *)array + sz;
2703                 } else {
2704                         *array = sample->read.one.id;
2705                         array++;
2706                 }
2707         }
2708
2709         if (type & PERF_SAMPLE_CALLCHAIN) {
2710                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2711                 memcpy(array, sample->callchain, sz);
2712                 array = (void *)array + sz;
2713         }
2714
2715         if (type & PERF_SAMPLE_RAW) {
2716                 u.val32[0] = sample->raw_size;
2717                 *array = u.val64;
2718                 array = (void *)array + sizeof(u32);
2719
2720                 memcpy(array, sample->raw_data, sample->raw_size);
2721                 array = (void *)array + sample->raw_size;
2722         }
2723
2724         if (type & PERF_SAMPLE_BRANCH_STACK) {
2725                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2726                 sz += sizeof(u64);
2727                 memcpy(array, sample->branch_stack, sz);
2728                 array = (void *)array + sz;
2729         }
2730
2731         if (type & PERF_SAMPLE_REGS_USER) {
2732                 if (sample->user_regs.abi) {
2733                         *array++ = sample->user_regs.abi;
2734                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2735                         memcpy(array, sample->user_regs.regs, sz);
2736                         array = (void *)array + sz;
2737                 } else {
2738                         *array++ = 0;
2739                 }
2740         }
2741
2742         if (type & PERF_SAMPLE_STACK_USER) {
2743                 sz = sample->user_stack.size;
2744                 *array++ = sz;
2745                 if (sz) {
2746                         memcpy(array, sample->user_stack.data, sz);
2747                         array = (void *)array + sz;
2748                         *array++ = sz;
2749                 }
2750         }
2751
2752         if (type & PERF_SAMPLE_WEIGHT) {
2753                 *array = sample->weight;
2754                 array++;
2755         }
2756
2757         if (type & PERF_SAMPLE_DATA_SRC) {
2758                 *array = sample->data_src;
2759                 array++;
2760         }
2761
2762         if (type & PERF_SAMPLE_TRANSACTION) {
2763                 *array = sample->transaction;
2764                 array++;
2765         }
2766
2767         if (type & PERF_SAMPLE_REGS_INTR) {
2768                 if (sample->intr_regs.abi) {
2769                         *array++ = sample->intr_regs.abi;
2770                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2771                         memcpy(array, sample->intr_regs.regs, sz);
2772                         array = (void *)array + sz;
2773                 } else {
2774                         *array++ = 0;
2775                 }
2776         }
2777
2778         if (type & PERF_SAMPLE_PHYS_ADDR) {
2779                 *array = sample->phys_addr;
2780                 array++;
2781         }
2782
2783         return 0;
2784 }
2785
2786 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2787 {
2788         return tep_find_field(evsel->tp_format, name);
2789 }
2790
2791 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2792                          const char *name)
2793 {
2794         struct tep_format_field *field = perf_evsel__field(evsel, name);
2795         int offset;
2796
2797         if (!field)
2798                 return NULL;
2799
2800         offset = field->offset;
2801
2802         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2803                 offset = *(int *)(sample->raw_data + field->offset);
2804                 offset &= 0xffff;
2805         }
2806
2807         return sample->raw_data + offset;
2808 }
2809
2810 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2811                          bool needs_swap)
2812 {
2813         u64 value;
2814         void *ptr = sample->raw_data + field->offset;
2815
2816         switch (field->size) {
2817         case 1:
2818                 return *(u8 *)ptr;
2819         case 2:
2820                 value = *(u16 *)ptr;
2821                 break;
2822         case 4:
2823                 value = *(u32 *)ptr;
2824                 break;
2825         case 8:
2826                 memcpy(&value, ptr, sizeof(u64));
2827                 break;
2828         default:
2829                 return 0;
2830         }
2831
2832         if (!needs_swap)
2833                 return value;
2834
2835         switch (field->size) {
2836         case 2:
2837                 return bswap_16(value);
2838         case 4:
2839                 return bswap_32(value);
2840         case 8:
2841                 return bswap_64(value);
2842         default:
2843                 return 0;
2844         }
2845
2846         return 0;
2847 }
2848
2849 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2850                        const char *name)
2851 {
2852         struct tep_format_field *field = perf_evsel__field(evsel, name);
2853
2854         if (!field)
2855                 return 0;
2856
2857         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2858 }
2859
2860 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2861                           char *msg, size_t msgsize)
2862 {
2863         int paranoid;
2864
2865         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2866             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2867             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2868                 /*
2869                  * If it's cycles then fall back to hrtimer based
2870                  * cpu-clock-tick sw counter, which is always available even if
2871                  * no PMU support.
2872                  *
2873                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2874                  * b0a873e).
2875                  */
2876                 scnprintf(msg, msgsize, "%s",
2877 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2878
2879                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2880                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2881
2882                 zfree(&evsel->name);
2883                 return true;
2884         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2885                    (paranoid = perf_event_paranoid()) > 1) {
2886                 const char *name = perf_evsel__name(evsel);
2887                 char *new_name;
2888                 const char *sep = ":";
2889
2890                 /* Is there already the separator in the name. */
2891                 if (strchr(name, '/') ||
2892                     strchr(name, ':'))
2893                         sep = "";
2894
2895                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2896                         return false;
2897
2898                 if (evsel->name)
2899                         free(evsel->name);
2900                 evsel->name = new_name;
2901                 scnprintf(msg, msgsize,
2902 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2903                 evsel->attr.exclude_kernel = 1;
2904
2905                 return true;
2906         }
2907
2908         return false;
2909 }
2910
2911 static bool find_process(const char *name)
2912 {
2913         size_t len = strlen(name);
2914         DIR *dir;
2915         struct dirent *d;
2916         int ret = -1;
2917
2918         dir = opendir(procfs__mountpoint());
2919         if (!dir)
2920                 return false;
2921
2922         /* Walk through the directory. */
2923         while (ret && (d = readdir(dir)) != NULL) {
2924                 char path[PATH_MAX];
2925                 char *data;
2926                 size_t size;
2927
2928                 if ((d->d_type != DT_DIR) ||
2929                      !strcmp(".", d->d_name) ||
2930                      !strcmp("..", d->d_name))
2931                         continue;
2932
2933                 scnprintf(path, sizeof(path), "%s/%s/comm",
2934                           procfs__mountpoint(), d->d_name);
2935
2936                 if (filename__read_str(path, &data, &size))
2937                         continue;
2938
2939                 ret = strncmp(name, data, len);
2940                 free(data);
2941         }
2942
2943         closedir(dir);
2944         return ret ? false : true;
2945 }
2946
2947 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2948                               int err, char *msg, size_t size)
2949 {
2950         char sbuf[STRERR_BUFSIZE];
2951         int printed = 0;
2952
2953         switch (err) {
2954         case EPERM:
2955         case EACCES:
2956                 if (err == EPERM)
2957                         printed = scnprintf(msg, size,
2958                                 "No permission to enable %s event.\n\n",
2959                                 perf_evsel__name(evsel));
2960
2961                 return scnprintf(msg + printed, size - printed,
2962                  "You may not have permission to collect %sstats.\n\n"
2963                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2964                  "which controls use of the performance events system by\n"
2965                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2966                  "The current value is %d:\n\n"
2967                  "  -1: Allow use of (almost) all events by all users\n"
2968                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2969                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2970                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2971                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2972                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2973                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2974                  "      kernel.perf_event_paranoid = -1\n" ,
2975                                  target->system_wide ? "system-wide " : "",
2976                                  perf_event_paranoid());
2977         case ENOENT:
2978                 return scnprintf(msg, size, "The %s event is not supported.",
2979                                  perf_evsel__name(evsel));
2980         case EMFILE:
2981                 return scnprintf(msg, size, "%s",
2982                          "Too many events are opened.\n"
2983                          "Probably the maximum number of open file descriptors has been reached.\n"
2984                          "Hint: Try again after reducing the number of events.\n"
2985                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2986         case ENOMEM:
2987                 if (evsel__has_callchain(evsel) &&
2988                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2989                         return scnprintf(msg, size,
2990                                          "Not enough memory to setup event with callchain.\n"
2991                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2992                                          "Hint: Current value: %d", sysctl__max_stack());
2993                 break;
2994         case ENODEV:
2995                 if (target->cpu_list)
2996                         return scnprintf(msg, size, "%s",
2997          "No such device - did you specify an out-of-range profile CPU?");
2998                 break;
2999         case EOPNOTSUPP:
3000                 if (evsel->attr.sample_period != 0)
3001                         return scnprintf(msg, size,
3002         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3003                                          perf_evsel__name(evsel));
3004                 if (evsel->attr.precise_ip)
3005                         return scnprintf(msg, size, "%s",
3006         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3007 #if defined(__i386__) || defined(__x86_64__)
3008                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
3009                         return scnprintf(msg, size, "%s",
3010         "No hardware sampling interrupt available.\n");
3011 #endif
3012                 break;
3013         case EBUSY:
3014                 if (find_process("oprofiled"))
3015                         return scnprintf(msg, size,
3016         "The PMU counters are busy/taken by another profiler.\n"
3017         "We found oprofile daemon running, please stop it and try again.");
3018                 break;
3019         case EINVAL:
3020                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3021                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3022                 if (perf_missing_features.clockid)
3023                         return scnprintf(msg, size, "clockid feature not supported.");
3024                 if (perf_missing_features.clockid_wrong)
3025                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3026                 break;
3027         default:
3028                 break;
3029         }
3030
3031         return scnprintf(msg, size,
3032         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3033         "/bin/dmesg | grep -i perf may provide additional information.\n",
3034                          err, str_error_r(err, sbuf, sizeof(sbuf)),
3035                          perf_evsel__name(evsel));
3036 }
3037
3038 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3039 {
3040         if (evsel && evsel->evlist)
3041                 return evsel->evlist->env;
3042         return NULL;
3043 }
3044
3045 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3046 {
3047         int cpu, thread;
3048
3049         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3050                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3051                      thread++) {
3052                         int fd = FD(evsel, cpu, thread);
3053
3054                         if (perf_evlist__id_add_fd(evlist, evsel,
3055                                                    cpu, thread, fd) < 0)
3056                                 return -1;
3057                 }
3058         }
3059
3060         return 0;
3061 }
3062
3063 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3064 {
3065         struct cpu_map *cpus = evsel->cpus;
3066         struct thread_map *threads = evsel->threads;
3067
3068         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3069                 return -ENOMEM;
3070
3071         return store_evsel_ids(evsel, evlist);
3072 }