tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.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 "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_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->name)
 594                 return evsel->name;
 595
 596         switch (evsel->attr.type) {
 597         case PERF_TYPE_RAW:
 598                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 599                 break;
 600
 601         case PERF_TYPE_HARDWARE:
 602                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 603                 break;
 604
 605         case PERF_TYPE_HW_CACHE:
 606                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 607                 break;
 608
 609         case PERF_TYPE_SOFTWARE:
 610                 if (evsel->tool_event)
 611                         perf_evsel__tool_name(bf, sizeof(bf));
 612                 else
 613                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 614                 break;
 615
 616         case PERF_TYPE_TRACEPOINT:
 617                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 618                 break;
 619
 620         case PERF_TYPE_BREAKPOINT:
 621                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 622                 break;
 623
 624         default:
 625                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 626                           evsel->attr.type);
 627                 break;
 628         }
 629
 630         evsel->name = strdup(bf);
 631
 632         return evsel->name ?: "unknown";
 633 }
 634
 635 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 636 {
 637         return evsel->group_name ?: "anon group";
 638 }
 639
 640 /*
 641  * Returns the group details for the specified leader,
 642  * with following rules.
 643  *
 644  *  For record -e '{cycles,instructions}'
 645  *    'anon group { cycles:u, instructions:u }'
 646  *
 647  *  For record -e 'cycles,instructions' and report --group
 648  *    'cycles:u, instructions:u'
 649  */
 650 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 651 {
 652         int ret = 0;
 653         struct perf_evsel *pos;
 654         const char *group_name = perf_evsel__group_name(evsel);
 655
 656         if (!evsel->forced_leader)
 657                 ret = scnprintf(buf, size, "%s { ", group_name);
 658
 659         ret += scnprintf(buf + ret, size - ret, "%s",
 660                          perf_evsel__name(evsel));
 661
 662         for_each_group_member(pos, evsel)
 663                 ret += scnprintf(buf + ret, size - ret, ", %s",
 664                                  perf_evsel__name(pos));
 665
 666         if (!evsel->forced_leader)
 667                 ret += scnprintf(buf + ret, size - ret, " }");
 668
 669         return ret;
 670 }
 671
 672 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 673                                            struct record_opts *opts,
 674                                            struct callchain_param *param)
 675 {
 676         bool function = perf_evsel__is_function_event(evsel);
 677         struct perf_event_attr *attr = &evsel->attr;
 678
 679         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 680
 681         attr->sample_max_stack = param->max_stack;
 682
 683         if (param->record_mode == CALLCHAIN_LBR) {
 684                 if (!opts->branch_stack) {
 685                         if (attr->exclude_user) {
 686                                 pr_warning("LBR callstack option is only available "
 687                                            "to get user callchain information. "
 688                                            "Falling back to framepointers.\n");
 689                         } else {
 690                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 691                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 692                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 693                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 694                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 695                         }
 696                 } else
 697                          pr_warning("Cannot use LBR callstack with branch stack. "
 698                                     "Falling back to framepointers.\n");
 699         }
 700
 701         if (param->record_mode == CALLCHAIN_DWARF) {
 702                 if (!function) {
 703                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 704                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 705                         attr->sample_regs_user |= PERF_REGS_MASK;
 706                         attr->sample_stack_user = param->dump_size;
 707                         attr->exclude_callchain_user = 1;
 708                 } else {
 709                         pr_info("Cannot use DWARF unwind for function trace event,"
 710                                 " falling back to framepointers.\n");
 711                 }
 712         }
 713
 714         if (function) {
 715                 pr_info("Disabling user space callchains for function trace event.\n");
 716                 attr->exclude_callchain_user = 1;
 717         }
 718 }
 719
 720 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 721                                   struct record_opts *opts,
 722                                   struct callchain_param *param)
 723 {
 724         if (param->enabled)
 725                 return __perf_evsel__config_callchain(evsel, opts, param);
 726 }
 727
 728 static void
 729 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 730                             struct callchain_param *param)
 731 {
 732         struct perf_event_attr *attr = &evsel->attr;
 733
 734         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 735         if (param->record_mode == CALLCHAIN_LBR) {
 736                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 737                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 738                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 739         }
 740         if (param->record_mode == CALLCHAIN_DWARF) {
 741                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 742                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 743         }
 744 }
 745
 746 static void apply_config_terms(struct perf_evsel *evsel,
 747                                struct record_opts *opts, bool track)
 748 {
 749         struct perf_evsel_config_term *term;
 750         struct list_head *config_terms = &evsel->config_terms;
 751         struct perf_event_attr *attr = &evsel->attr;
 752         /* callgraph default */
 753         struct callchain_param param = {
 754                 .record_mode = callchain_param.record_mode,
 755         };
 756         u32 dump_size = 0;
 757         int max_stack = 0;
 758         const char *callgraph_buf = NULL;
 759
 760         list_for_each_entry(term, config_terms, list) {
 761                 switch (term->type) {
 762                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 763                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 764                                 attr->sample_period = term->val.period;
 765                                 attr->freq = 0;
 766                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 767                         }
 768                         break;
 769                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 770                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 771                                 attr->sample_freq = term->val.freq;
 772                                 attr->freq = 1;
 773                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 774                         }
 775                         break;
 776                 case PERF_EVSEL__CONFIG_TERM_TIME:
 777                         if (term->val.time)
 778                                 perf_evsel__set_sample_bit(evsel, TIME);
 779                         else
 780                                 perf_evsel__reset_sample_bit(evsel, TIME);
 781                         break;
 782                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 783                         callgraph_buf = term->val.callgraph;
 784                         break;
 785                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 786                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 787                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 788                                 parse_branch_str(term->val.branch,
 789                                                  &attr->branch_sample_type);
 790                         } else
 791                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 794                         dump_size = term->val.stack_user;
 795                         break;
 796                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 797                         max_stack = term->val.max_stack;
 798                         break;
 799                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 800                         evsel->max_events = term->val.max_events;
 801                         break;
 802                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 803                         /*
 804                          * attr->inherit should has already been set by
 805                          * perf_evsel__config. If user explicitly set
 806                          * inherit using config terms, override global
 807                          * opt->no_inherit setting.
 808                          */
 809                         attr->inherit = term->val.inherit ? 1 : 0;
 810                         break;
 811                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 812                         attr->write_backward = term->val.overwrite ? 1 : 0;
 813                         break;
 814                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 815                         break;
 816                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 817                         break;
 818                 default:
 819                         break;
 820                 }
 821         }
 822
 823         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 824         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 825                 bool sample_address = false;
 826
 827                 if (max_stack) {
 828                         param.max_stack = max_stack;
 829                         if (callgraph_buf == NULL)
 830                                 callgraph_buf = "fp";
 831                 }
 832
 833                 /* parse callgraph parameters */
 834                 if (callgraph_buf != NULL) {
 835                         if (!strcmp(callgraph_buf, "no")) {
 836                                 param.enabled = false;
 837                                 param.record_mode = CALLCHAIN_NONE;
 838                         } else {
 839                                 param.enabled = true;
 840                                 if (parse_callchain_record(callgraph_buf, &param)) {
 841                                         pr_err("per-event callgraph setting for %s failed. "
 842                                                "Apply callgraph global setting for it\n",
 843                                                evsel->name);
 844                                         return;
 845                                 }
 846                                 if (param.record_mode == CALLCHAIN_DWARF)
 847                                         sample_address = true;
 848                         }
 849                 }
 850                 if (dump_size > 0) {
 851                         dump_size = round_up(dump_size, sizeof(u64));
 852                         param.dump_size = dump_size;
 853                 }
 854
 855                 /* If global callgraph set, clear it */
 856                 if (callchain_param.enabled)
 857                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 858
 859                 /* set perf-event callgraph */
 860                 if (param.enabled) {
 861                         if (sample_address) {
 862                                 perf_evsel__set_sample_bit(evsel, ADDR);
 863                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 864                                 evsel->attr.mmap_data = track;
 865                         }
 866                         perf_evsel__config_callchain(evsel, opts, &param);
 867                 }
 868         }
 869 }
 870
 871 static bool is_dummy_event(struct perf_evsel *evsel)
 872 {
 873         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 874                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 875 }
 876
 877 /*
 878  * The enable_on_exec/disabled value strategy:
 879  *
 880  *  1) For any type of traced program:
 881  *    - all independent events and group leaders are disabled
 882  *    - all group members are enabled
 883  *
 884  *     Group members are ruled by group leaders. They need to
 885  *     be enabled, because the group scheduling relies on that.
 886  *
 887  *  2) For traced programs executed by perf:
 888  *     - all independent events and group leaders have
 889  *       enable_on_exec set
 890  *     - we don't specifically enable or disable any event during
 891  *       the record command
 892  *
 893  *     Independent events and group leaders are initially disabled
 894  *     and get enabled by exec. Group members are ruled by group
 895  *     leaders as stated in 1).
 896  *
 897  *  3) For traced programs attached by perf (pid/tid):
 898  *     - we specifically enable or disable all events during
 899  *       the record command
 900  *
 901  *     When attaching events to already running traced we
 902  *     enable/disable events specifically, as there's no
 903  *     initial traced exec call.
 904  */
 905 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 906                         struct callchain_param *callchain)
 907 {
 908         struct perf_evsel *leader = evsel->leader;
 909         struct perf_event_attr *attr = &evsel->attr;
 910         int track = evsel->tracking;
 911         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 912
 913         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 914         attr->inherit       = !opts->no_inherit;
 915         attr->write_backward = opts->overwrite ? 1 : 0;
 916
 917         perf_evsel__set_sample_bit(evsel, IP);
 918         perf_evsel__set_sample_bit(evsel, TID);
 919
 920         if (evsel->sample_read) {
 921                 perf_evsel__set_sample_bit(evsel, READ);
 922
 923                 /*
 924                  * We need ID even in case of single event, because
 925                  * PERF_SAMPLE_READ process ID specific data.
 926                  */
 927                 perf_evsel__set_sample_id(evsel, false);
 928
 929                 /*
 930                  * Apply group format only if we belong to group
 931                  * with more than one members.
 932                  */
 933                 if (leader->nr_members > 1) {
 934                         attr->read_format |= PERF_FORMAT_GROUP;
 935                         attr->inherit = 0;
 936                 }
 937         }
 938
 939         /*
 940          * We default some events to have a default interval. But keep
 941          * it a weak assumption overridable by the user.
 942          */
 943         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 944                                      opts->user_interval != ULLONG_MAX)) {
 945                 if (opts->freq) {
 946                         perf_evsel__set_sample_bit(evsel, PERIOD);
 947                         attr->freq              = 1;
 948                         attr->sample_freq       = opts->freq;
 949                 } else {
 950                         attr->sample_period = opts->default_interval;
 951                 }
 952         }
 953
 954         /*
 955          * Disable sampling for all group members other
 956          * than leader in case leader 'leads' the sampling.
 957          */
 958         if ((leader != evsel) && leader->sample_read) {
 959                 attr->freq           = 0;
 960                 attr->sample_freq    = 0;
 961                 attr->sample_period  = 0;
 962                 attr->write_backward = 0;
 963
 964                 /*
 965                  * We don't get sample for slave events, we make them
 966                  * when delivering group leader sample. Set the slave
 967                  * event to follow the master sample_type to ease up
 968                  * report.
 969                  */
 970                 attr->sample_type = leader->attr.sample_type;
 971         }
 972
 973         if (opts->no_samples)
 974                 attr->sample_freq = 0;
 975
 976         if (opts->inherit_stat) {
 977                 evsel->attr.read_format |=
 978                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 979                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 980                         PERF_FORMAT_ID;
 981                 attr->inherit_stat = 1;
 982         }
 983
 984         if (opts->sample_address) {
 985                 perf_evsel__set_sample_bit(evsel, ADDR);
 986                 attr->mmap_data = track;
 987         }
 988
 989         /*
 990          * We don't allow user space callchains for  function trace
 991          * event, due to issues with page faults while tracing page
 992          * fault handler and its overall trickiness nature.
 993          */
 994         if (perf_evsel__is_function_event(evsel))
 995                 evsel->attr.exclude_callchain_user = 1;
 996
 997         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 998                 perf_evsel__config_callchain(evsel, opts, callchain);
 999
1000         if (opts->sample_intr_regs) {
1001                 attr->sample_regs_intr = opts->sample_intr_regs;
1002                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1003         }
1004
1005         if (opts->sample_user_regs) {
1006                 attr->sample_regs_user |= opts->sample_user_regs;
1007                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1008         }
1009
1010         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1011                 perf_evsel__set_sample_bit(evsel, CPU);
1012
1013         /*
1014          * When the user explicitly disabled time don't force it here.
1015          */
1016         if (opts->sample_time &&
1017             (!perf_missing_features.sample_id_all &&
1018             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1019              opts->sample_time_set)))
1020                 perf_evsel__set_sample_bit(evsel, TIME);
1021
1022         if (opts->raw_samples && !evsel->no_aux_samples) {
1023                 perf_evsel__set_sample_bit(evsel, TIME);
1024                 perf_evsel__set_sample_bit(evsel, RAW);
1025                 perf_evsel__set_sample_bit(evsel, CPU);
1026         }
1027
1028         if (opts->sample_address)
1029                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1030
1031         if (opts->sample_phys_addr)
1032                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1033
1034         if (opts->no_buffering) {
1035                 attr->watermark = 0;
1036                 attr->wakeup_events = 1;
1037         }
1038         if (opts->branch_stack && !evsel->no_aux_samples) {
1039                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1040                 attr->branch_sample_type = opts->branch_stack;
1041         }
1042
1043         if (opts->sample_weight)
1044                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1045
1046         attr->task  = track;
1047         attr->mmap  = track;
1048         attr->mmap2 = track && !perf_missing_features.mmap2;
1049         attr->comm  = track;
1050         attr->ksymbol = track && !perf_missing_features.ksymbol;
1051         attr->bpf_event = track && !opts->no_bpf_event &&
1052                 !perf_missing_features.bpf_event;
1053
1054         if (opts->record_namespaces)
1055                 attr->namespaces  = track;
1056
1057         if (opts->record_switch_events)
1058                 attr->context_switch = track;
1059
1060         if (opts->sample_transaction)
1061                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1062
1063         if (opts->running_time) {
1064                 evsel->attr.read_format |=
1065                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1066                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1067         }
1068
1069         /*
1070          * XXX see the function comment above
1071          *
1072          * Disabling only independent events or group leaders,
1073          * keeping group members enabled.
1074          */
1075         if (perf_evsel__is_group_leader(evsel))
1076                 attr->disabled = 1;
1077
1078         /*
1079          * Setting enable_on_exec for independent events and
1080          * group leaders for traced executed by perf.
1081          */
1082         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1083                 !opts->initial_delay)
1084                 attr->enable_on_exec = 1;
1085
1086         if (evsel->immediate) {
1087                 attr->disabled = 0;
1088                 attr->enable_on_exec = 0;
1089         }
1090
1091         clockid = opts->clockid;
1092         if (opts->use_clockid) {
1093                 attr->use_clockid = 1;
1094                 attr->clockid = opts->clockid;
1095         }
1096
1097         if (evsel->precise_max)
1098                 attr->precise_ip = 3;
1099
1100         if (opts->all_user) {
1101                 attr->exclude_kernel = 1;
1102                 attr->exclude_user   = 0;
1103         }
1104
1105         if (opts->all_kernel) {
1106                 attr->exclude_kernel = 0;
1107                 attr->exclude_user   = 1;
1108         }
1109
1110         if (evsel->own_cpus || evsel->unit)
1111                 evsel->attr.read_format |= PERF_FORMAT_ID;
1112
1113         /*
1114          * Apply event specific term settings,
1115          * it overloads any global configuration.
1116          */
1117         apply_config_terms(evsel, opts, track);
1118
1119         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1120
1121         /* The --period option takes the precedence. */
1122         if (opts->period_set) {
1123                 if (opts->period)
1124                         perf_evsel__set_sample_bit(evsel, PERIOD);
1125                 else
1126                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1127         }
1128
1129         /*
1130          * For initial_delay, a dummy event is added implicitly.
1131          * The software event will trigger -EOPNOTSUPP error out,
1132          * if BRANCH_STACK bit is set.
1133          */
1134         if (opts->initial_delay && is_dummy_event(evsel))
1135                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1136 }
1137
1138 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1139 {
1140         if (evsel->system_wide)
1141                 nthreads = 1;
1142
1143         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1144
1145         if (evsel->fd) {
1146                 int cpu, thread;
1147                 for (cpu = 0; cpu < ncpus; cpu++) {
1148                         for (thread = 0; thread < nthreads; thread++) {
1149                                 FD(evsel, cpu, thread) = -1;
1150                         }
1151                 }
1152         }
1153
1154         return evsel->fd != NULL ? 0 : -ENOMEM;
1155 }
1156
1157 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1158                           int ioc,  void *arg)
1159 {
1160         int cpu, thread;
1161
1162         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1163                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1164                         int fd = FD(evsel, cpu, thread),
1165                             err = ioctl(fd, ioc, arg);
1166
1167                         if (err)
1168                                 return err;
1169                 }
1170         }
1171
1172         return 0;
1173 }
1174
1175 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1176 {
1177         return perf_evsel__run_ioctl(evsel,
1178                                      PERF_EVENT_IOC_SET_FILTER,
1179                                      (void *)filter);
1180 }
1181
1182 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1183 {
1184         char *new_filter = strdup(filter);
1185
1186         if (new_filter != NULL) {
1187                 free(evsel->filter);
1188                 evsel->filter = new_filter;
1189                 return 0;
1190         }
1191
1192         return -1;
1193 }
1194
1195 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1196                                      const char *fmt, const char *filter)
1197 {
1198         char *new_filter;
1199
1200         if (evsel->filter == NULL)
1201                 return perf_evsel__set_filter(evsel, filter);
1202
1203         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1204                 free(evsel->filter);
1205                 evsel->filter = new_filter;
1206                 return 0;
1207         }
1208
1209         return -1;
1210 }
1211
1212 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1213 {
1214         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1215 }
1216
1217 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1218 {
1219         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1220 }
1221
1222 int perf_evsel__enable(struct perf_evsel *evsel)
1223 {
1224         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1225
1226         if (!err)
1227                 evsel->disabled = false;
1228
1229         return err;
1230 }
1231
1232 int perf_evsel__disable(struct perf_evsel *evsel)
1233 {
1234         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1235         /*
1236          * We mark it disabled here so that tools that disable a event can
1237          * ignore events after they disable it. I.e. the ring buffer may have
1238          * already a few more events queued up before the kernel got the stop
1239          * request.
1240          */
1241         if (!err)
1242                 evsel->disabled = true;
1243
1244         return err;
1245 }
1246
1247 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1248 {
1249         if (ncpus == 0 || nthreads == 0)
1250                 return 0;
1251
1252         if (evsel->system_wide)
1253                 nthreads = 1;
1254
1255         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1256         if (evsel->sample_id == NULL)
1257                 return -ENOMEM;
1258
1259         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1260         if (evsel->id == NULL) {
1261                 xyarray__delete(evsel->sample_id);
1262                 evsel->sample_id = NULL;
1263                 return -ENOMEM;
1264         }
1265
1266         return 0;
1267 }
1268
1269 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1270 {
1271         xyarray__delete(evsel->fd);
1272         evsel->fd = NULL;
1273 }
1274
1275 static void perf_evsel__free_id(struct perf_evsel *evsel)
1276 {
1277         xyarray__delete(evsel->sample_id);
1278         evsel->sample_id = NULL;
1279         zfree(&evsel->id);
1280 }
1281
1282 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1283 {
1284         struct perf_evsel_config_term *term, *h;
1285
1286         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1287                 list_del(&term->list);
1288                 free(term);
1289         }
1290 }
1291
1292 void perf_evsel__close_fd(struct perf_evsel *evsel)
1293 {
1294         int cpu, thread;
1295
1296         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1297                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1298                         close(FD(evsel, cpu, thread));
1299                         FD(evsel, cpu, thread) = -1;
1300                 }
1301 }
1302
1303 void perf_evsel__exit(struct perf_evsel *evsel)
1304 {
1305         assert(list_empty(&evsel->node));
1306         assert(evsel->evlist == NULL);
1307         perf_evsel__free_counts(evsel);
1308         perf_evsel__free_fd(evsel);
1309         perf_evsel__free_id(evsel);
1310         perf_evsel__free_config_terms(evsel);
1311         cgroup__put(evsel->cgrp);
1312         cpu_map__put(evsel->cpus);
1313         cpu_map__put(evsel->own_cpus);
1314         thread_map__put(evsel->threads);
1315         zfree(&evsel->group_name);
1316         zfree(&evsel->name);
1317         perf_evsel__object.fini(evsel);
1318 }
1319
1320 void perf_evsel__delete(struct perf_evsel *evsel)
1321 {
1322         perf_evsel__exit(evsel);
1323         free(evsel);
1324 }
1325
1326 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1327                                 struct perf_counts_values *count)
1328 {
1329         struct perf_counts_values tmp;
1330
1331         if (!evsel->prev_raw_counts)
1332                 return;
1333
1334         if (cpu == -1) {
1335                 tmp = evsel->prev_raw_counts->aggr;
1336                 evsel->prev_raw_counts->aggr = *count;
1337         } else {
1338                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1339                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1340         }
1341
1342         count->val = count->val - tmp.val;
1343         count->ena = count->ena - tmp.ena;
1344         count->run = count->run - tmp.run;
1345 }
1346
1347 void perf_counts_values__scale(struct perf_counts_values *count,
1348                                bool scale, s8 *pscaled)
1349 {
1350         s8 scaled = 0;
1351
1352         if (scale) {
1353                 if (count->run == 0) {
1354                         scaled = -1;
1355                         count->val = 0;
1356                 } else if (count->run < count->ena) {
1357                         scaled = 1;
1358                         count->val = (u64)((double) count->val * count->ena / count->run);
1359                 }
1360         }
1361
1362         if (pscaled)
1363                 *pscaled = scaled;
1364 }
1365
1366 static int perf_evsel__read_size(struct perf_evsel *evsel)
1367 {
1368         u64 read_format = evsel->attr.read_format;
1369         int entry = sizeof(u64); /* value */
1370         int size = 0;
1371         int nr = 1;
1372
1373         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1374                 size += sizeof(u64);
1375
1376         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1377                 size += sizeof(u64);
1378
1379         if (read_format & PERF_FORMAT_ID)
1380                 entry += sizeof(u64);
1381
1382         if (read_format & PERF_FORMAT_GROUP) {
1383                 nr = evsel->nr_members;
1384                 size += sizeof(u64);
1385         }
1386
1387         size += entry * nr;
1388         return size;
1389 }
1390
1391 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1392                      struct perf_counts_values *count)
1393 {
1394         size_t size = perf_evsel__read_size(evsel);
1395
1396         memset(count, 0, sizeof(*count));
1397
1398         if (FD(evsel, cpu, thread) < 0)
1399                 return -EINVAL;
1400
1401         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1402                 return -errno;
1403
1404         return 0;
1405 }
1406
1407 static int
1408 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1409 {
1410         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1411
1412         return perf_evsel__read(evsel, cpu, thread, count);
1413 }
1414
1415 static void
1416 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1417                       u64 val, u64 ena, u64 run)
1418 {
1419         struct perf_counts_values *count;
1420
1421         count = perf_counts(counter->counts, cpu, thread);
1422
1423         count->val    = val;
1424         count->ena    = ena;
1425         count->run    = run;
1426         count->loaded = true;
1427 }
1428
1429 static int
1430 perf_evsel__process_group_data(struct perf_evsel *leader,
1431                                int cpu, int thread, u64 *data)
1432 {
1433         u64 read_format = leader->attr.read_format;
1434         struct sample_read_value *v;
1435         u64 nr, ena = 0, run = 0, i;
1436
1437         nr = *data++;
1438
1439         if (nr != (u64) leader->nr_members)
1440                 return -EINVAL;
1441
1442         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1443                 ena = *data++;
1444
1445         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1446                 run = *data++;
1447
1448         v = (struct sample_read_value *) data;
1449
1450         perf_evsel__set_count(leader, cpu, thread,
1451                               v[0].value, ena, run);
1452
1453         for (i = 1; i < nr; i++) {
1454                 struct perf_evsel *counter;
1455
1456                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1457                 if (!counter)
1458                         return -EINVAL;
1459
1460                 perf_evsel__set_count(counter, cpu, thread,
1461                                       v[i].value, ena, run);
1462         }
1463
1464         return 0;
1465 }
1466
1467 static int
1468 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1469 {
1470         struct perf_stat_evsel *ps = leader->stats;
1471         u64 read_format = leader->attr.read_format;
1472         int size = perf_evsel__read_size(leader);
1473         u64 *data = ps->group_data;
1474
1475         if (!(read_format & PERF_FORMAT_ID))
1476                 return -EINVAL;
1477
1478         if (!perf_evsel__is_group_leader(leader))
1479                 return -EINVAL;
1480
1481         if (!data) {
1482                 data = zalloc(size);
1483                 if (!data)
1484                         return -ENOMEM;
1485
1486                 ps->group_data = data;
1487         }
1488
1489         if (FD(leader, cpu, thread) < 0)
1490                 return -EINVAL;
1491
1492         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1493                 return -errno;
1494
1495         return perf_evsel__process_group_data(leader, cpu, thread, data);
1496 }
1497
1498 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1499 {
1500         u64 read_format = evsel->attr.read_format;
1501
1502         if (read_format & PERF_FORMAT_GROUP)
1503                 return perf_evsel__read_group(evsel, cpu, thread);
1504         else
1505                 return perf_evsel__read_one(evsel, cpu, thread);
1506 }
1507
1508 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1509                               int cpu, int thread, bool scale)
1510 {
1511         struct perf_counts_values count;
1512         size_t nv = scale ? 3 : 1;
1513
1514         if (FD(evsel, cpu, thread) < 0)
1515                 return -EINVAL;
1516
1517         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1518                 return -ENOMEM;
1519
1520         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1521                 return -errno;
1522
1523         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1524         perf_counts_values__scale(&count, scale, NULL);
1525         *perf_counts(evsel->counts, cpu, thread) = count;
1526         return 0;
1527 }
1528
1529 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1530 {
1531         struct perf_evsel *leader = evsel->leader;
1532         int fd;
1533
1534         if (perf_evsel__is_group_leader(evsel))
1535                 return -1;
1536
1537         /*
1538          * Leader must be already processed/open,
1539          * if not it's a bug.
1540          */
1541         BUG_ON(!leader->fd);
1542
1543         fd = FD(leader, cpu, thread);
1544         BUG_ON(fd == -1);
1545
1546         return fd;
1547 }
1548
1549 struct bit_names {
1550         int bit;
1551         const char *name;
1552 };
1553
1554 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1555 {
1556         bool first_bit = true;
1557         int i = 0;
1558
1559         do {
1560                 if (value & bits[i].bit) {
1561                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1562                         first_bit = false;
1563                 }
1564         } while (bits[++i].name != NULL);
1565 }
1566
1567 static void __p_sample_type(char *buf, size_t size, u64 value)
1568 {
1569 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1570         struct bit_names bits[] = {
1571                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1572                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1573                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1574                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1575                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1576                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1577                 { .name = NULL, }
1578         };
1579 #undef bit_name
1580         __p_bits(buf, size, value, bits);
1581 }
1582
1583 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1584 {
1585 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1586         struct bit_names bits[] = {
1587                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1588                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1589                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1590                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1591                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1592                 { .name = NULL, }
1593         };
1594 #undef bit_name
1595         __p_bits(buf, size, value, bits);
1596 }
1597
1598 static void __p_read_format(char *buf, size_t size, u64 value)
1599 {
1600 #define bit_name(n) { PERF_FORMAT_##n, #n }
1601         struct bit_names bits[] = {
1602                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1603                 bit_name(ID), bit_name(GROUP),
1604                 { .name = NULL, }
1605         };
1606 #undef bit_name
1607         __p_bits(buf, size, value, bits);
1608 }
1609
1610 #define BUF_SIZE                1024
1611
1612 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1613 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1614 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1615 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1616 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1617 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1618
1619 #define PRINT_ATTRn(_n, _f, _p)                         \
1620 do {                                                    \
1621         if (attr->_f) {                                 \
1622                 _p(attr->_f);                           \
1623                 ret += attr__fprintf(fp, _n, buf, priv);\
1624         }                                               \
1625 } while (0)
1626
1627 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1628
1629 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1630                              attr__fprintf_f attr__fprintf, void *priv)
1631 {
1632         char buf[BUF_SIZE];
1633         int ret = 0;
1634
1635         PRINT_ATTRf(type, p_unsigned);
1636         PRINT_ATTRf(size, p_unsigned);
1637         PRINT_ATTRf(config, p_hex);
1638         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1639         PRINT_ATTRf(sample_type, p_sample_type);
1640         PRINT_ATTRf(read_format, p_read_format);
1641
1642         PRINT_ATTRf(disabled, p_unsigned);
1643         PRINT_ATTRf(inherit, p_unsigned);
1644         PRINT_ATTRf(pinned, p_unsigned);
1645         PRINT_ATTRf(exclusive, p_unsigned);
1646         PRINT_ATTRf(exclude_user, p_unsigned);
1647         PRINT_ATTRf(exclude_kernel, p_unsigned);
1648         PRINT_ATTRf(exclude_hv, p_unsigned);
1649         PRINT_ATTRf(exclude_idle, p_unsigned);
1650         PRINT_ATTRf(mmap, p_unsigned);
1651         PRINT_ATTRf(comm, p_unsigned);
1652         PRINT_ATTRf(freq, p_unsigned);
1653         PRINT_ATTRf(inherit_stat, p_unsigned);
1654         PRINT_ATTRf(enable_on_exec, p_unsigned);
1655         PRINT_ATTRf(task, p_unsigned);
1656         PRINT_ATTRf(watermark, p_unsigned);
1657         PRINT_ATTRf(precise_ip, p_unsigned);
1658         PRINT_ATTRf(mmap_data, p_unsigned);
1659         PRINT_ATTRf(sample_id_all, p_unsigned);
1660         PRINT_ATTRf(exclude_host, p_unsigned);
1661         PRINT_ATTRf(exclude_guest, p_unsigned);
1662         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1663         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1664         PRINT_ATTRf(mmap2, p_unsigned);
1665         PRINT_ATTRf(comm_exec, p_unsigned);
1666         PRINT_ATTRf(use_clockid, p_unsigned);
1667         PRINT_ATTRf(context_switch, p_unsigned);
1668         PRINT_ATTRf(write_backward, p_unsigned);
1669         PRINT_ATTRf(namespaces, p_unsigned);
1670         PRINT_ATTRf(ksymbol, p_unsigned);
1671         PRINT_ATTRf(bpf_event, p_unsigned);
1672
1673         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1674         PRINT_ATTRf(bp_type, p_unsigned);
1675         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1676         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1677         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1678         PRINT_ATTRf(sample_regs_user, p_hex);
1679         PRINT_ATTRf(sample_stack_user, p_unsigned);
1680         PRINT_ATTRf(clockid, p_signed);
1681         PRINT_ATTRf(sample_regs_intr, p_hex);
1682         PRINT_ATTRf(aux_watermark, p_unsigned);
1683         PRINT_ATTRf(sample_max_stack, p_unsigned);
1684
1685         return ret;
1686 }
1687
1688 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1689                                 void *priv __maybe_unused)
1690 {
1691         return fprintf(fp, "  %-32s %s\n", name, val);
1692 }
1693
1694 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1695                                   int nr_cpus, int nr_threads,
1696                                   int thread_idx)
1697 {
1698         for (int cpu = 0; cpu < nr_cpus; cpu++)
1699                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1700                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1701 }
1702
1703 static int update_fds(struct perf_evsel *evsel,
1704                       int nr_cpus, int cpu_idx,
1705                       int nr_threads, int thread_idx)
1706 {
1707         struct perf_evsel *pos;
1708
1709         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1710                 return -EINVAL;
1711
1712         evlist__for_each_entry(evsel->evlist, pos) {
1713                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1714
1715                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1716
1717                 /*
1718                  * Since fds for next evsel has not been created,
1719                  * there is no need to iterate whole event list.
1720                  */
1721                 if (pos == evsel)
1722                         break;
1723         }
1724         return 0;
1725 }
1726
1727 static bool ignore_missing_thread(struct perf_evsel *evsel,
1728                                   int nr_cpus, int cpu,
1729                                   struct thread_map *threads,
1730                                   int thread, int err)
1731 {
1732         pid_t ignore_pid = thread_map__pid(threads, thread);
1733
1734         if (!evsel->ignore_missing_thread)
1735                 return false;
1736
1737         /* The system wide setup does not work with threads. */
1738         if (evsel->system_wide)
1739                 return false;
1740
1741         /* The -ESRCH is perf event syscall errno for pid's not found. */
1742         if (err != -ESRCH)
1743                 return false;
1744
1745         /* If there's only one thread, let it fail. */
1746         if (threads->nr == 1)
1747                 return false;
1748
1749         /*
1750          * We should remove fd for missing_thread first
1751          * because thread_map__remove() will decrease threads->nr.
1752          */
1753         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1754                 return false;
1755
1756         if (thread_map__remove(threads, thread))
1757                 return false;
1758
1759         pr_warning("WARNING: Ignored open failure for pid %d\n",
1760                    ignore_pid);
1761         return true;
1762 }
1763
1764 static void display_attr(struct perf_event_attr *attr)
1765 {
1766         if (verbose >= 2) {
1767                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1768                 fprintf(stderr, "perf_event_attr:\n");
1769                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1770                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1771         }
1772 }
1773
1774 static int perf_event_open(struct perf_evsel *evsel,
1775                            pid_t pid, int cpu, int group_fd,
1776                            unsigned long flags)
1777 {
1778         int precise_ip = evsel->attr.precise_ip;
1779         int fd;
1780
1781         while (1) {
1782                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1783                           pid, cpu, group_fd, flags);
1784
1785                 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1786                 if (fd >= 0)
1787                         break;
1788
1789                 /*
1790                  * Do quick precise_ip fallback if:
1791                  *  - there is precise_ip set in perf_event_attr
1792                  *  - maximum precise is requested
1793                  *  - sys_perf_event_open failed with ENOTSUP error,
1794                  *    which is associated with wrong precise_ip
1795                  */
1796                 if (!precise_ip || !evsel->precise_max || (errno != ENOTSUP))
1797                         break;
1798
1799                 /*
1800                  * We tried all the precise_ip values, and it's
1801                  * still failing, so leave it to standard fallback.
1802                  */
1803                 if (!evsel->attr.precise_ip) {
1804                         evsel->attr.precise_ip = precise_ip;
1805                         break;
1806                 }
1807
1808                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1809                 evsel->attr.precise_ip--;
1810                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1811                 display_attr(&evsel->attr);
1812         }
1813
1814         return fd;
1815 }
1816
1817 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1818                      struct thread_map *threads)
1819 {
1820         int cpu, thread, nthreads;
1821         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1822         int pid = -1, err;
1823         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1824
1825         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1826                 return -EINVAL;
1827
1828         if (cpus == NULL) {
1829                 static struct cpu_map *empty_cpu_map;
1830
1831                 if (empty_cpu_map == NULL) {
1832                         empty_cpu_map = cpu_map__dummy_new();
1833                         if (empty_cpu_map == NULL)
1834                                 return -ENOMEM;
1835                 }
1836
1837                 cpus = empty_cpu_map;
1838         }
1839
1840         if (threads == NULL) {
1841                 static struct thread_map *empty_thread_map;
1842
1843                 if (empty_thread_map == NULL) {
1844                         empty_thread_map = thread_map__new_by_tid(-1);
1845                         if (empty_thread_map == NULL)
1846                                 return -ENOMEM;
1847                 }
1848
1849                 threads = empty_thread_map;
1850         }
1851
1852         if (evsel->system_wide)
1853                 nthreads = 1;
1854         else
1855                 nthreads = threads->nr;
1856
1857         if (evsel->fd == NULL &&
1858             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1859                 return -ENOMEM;
1860
1861         if (evsel->cgrp) {
1862                 flags |= PERF_FLAG_PID_CGROUP;
1863                 pid = evsel->cgrp->fd;
1864         }
1865
1866 fallback_missing_features:
1867         if (perf_missing_features.clockid_wrong)
1868                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1869         if (perf_missing_features.clockid) {
1870                 evsel->attr.use_clockid = 0;
1871                 evsel->attr.clockid = 0;
1872         }
1873         if (perf_missing_features.cloexec)
1874                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1875         if (perf_missing_features.mmap2)
1876                 evsel->attr.mmap2 = 0;
1877         if (perf_missing_features.exclude_guest)
1878                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1879         if (perf_missing_features.lbr_flags)
1880                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1881                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1882         if (perf_missing_features.group_read && evsel->attr.inherit)
1883                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1884         if (perf_missing_features.ksymbol)
1885                 evsel->attr.ksymbol = 0;
1886         if (perf_missing_features.bpf_event)
1887                 evsel->attr.bpf_event = 0;
1888 retry_sample_id:
1889         if (perf_missing_features.sample_id_all)
1890                 evsel->attr.sample_id_all = 0;
1891
1892         display_attr(&evsel->attr);
1893
1894         for (cpu = 0; cpu < cpus->nr; cpu++) {
1895
1896                 for (thread = 0; thread < nthreads; thread++) {
1897                         int fd, group_fd;
1898
1899                         if (!evsel->cgrp && !evsel->system_wide)
1900                                 pid = thread_map__pid(threads, thread);
1901
1902                         group_fd = get_group_fd(evsel, cpu, thread);
1903 retry_open:
1904                         test_attr__ready();
1905
1906                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1907                                              group_fd, flags);
1908
1909                         FD(evsel, cpu, thread) = fd;
1910
1911                         if (fd < 0) {
1912                                 err = -errno;
1913
1914                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1915                                         /*
1916                                          * We just removed 1 thread, so take a step
1917                                          * back on thread index and lower the upper
1918                                          * nthreads limit.
1919                                          */
1920                                         nthreads--;
1921                                         thread--;
1922
1923                                         /* ... and pretend like nothing have happened. */
1924                                         err = 0;
1925                                         continue;
1926                                 }
1927
1928                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1929                                           err);
1930                                 goto try_fallback;
1931                         }
1932
1933                         pr_debug2(" = %d\n", fd);
1934
1935                         if (evsel->bpf_fd >= 0) {
1936                                 int evt_fd = fd;
1937                                 int bpf_fd = evsel->bpf_fd;
1938
1939                                 err = ioctl(evt_fd,
1940                                             PERF_EVENT_IOC_SET_BPF,
1941                                             bpf_fd);
1942                                 if (err && errno != EEXIST) {
1943                                         pr_err("failed to attach bpf fd %d: %s\n",
1944                                                bpf_fd, strerror(errno));
1945                                         err = -EINVAL;
1946                                         goto out_close;
1947                                 }
1948                         }
1949
1950                         set_rlimit = NO_CHANGE;
1951
1952                         /*
1953                          * If we succeeded but had to kill clockid, fail and
1954                          * have perf_evsel__open_strerror() print us a nice
1955                          * error.
1956                          */
1957                         if (perf_missing_features.clockid ||
1958                             perf_missing_features.clockid_wrong) {
1959                                 err = -EINVAL;
1960                                 goto out_close;
1961                         }
1962                 }
1963         }
1964
1965         return 0;
1966
1967 try_fallback:
1968         /*
1969          * perf stat needs between 5 and 22 fds per CPU. When we run out
1970          * of them try to increase the limits.
1971          */
1972         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1973                 struct rlimit l;
1974                 int old_errno = errno;
1975
1976                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1977                         if (set_rlimit == NO_CHANGE)
1978                                 l.rlim_cur = l.rlim_max;
1979                         else {
1980                                 l.rlim_cur = l.rlim_max + 1000;
1981                                 l.rlim_max = l.rlim_cur;
1982                         }
1983                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1984                                 set_rlimit++;
1985                                 errno = old_errno;
1986                                 goto retry_open;
1987                         }
1988                 }
1989                 errno = old_errno;
1990         }
1991
1992         if (err != -EINVAL || cpu > 0 || thread > 0)
1993                 goto out_close;
1994
1995         /*
1996          * Must probe features in the order they were added to the
1997          * perf_event_attr interface.
1998          */
1999         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
2000                 perf_missing_features.bpf_event = true;
2001                 pr_debug2("switching off bpf_event\n");
2002                 goto fallback_missing_features;
2003         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
2004                 perf_missing_features.ksymbol = true;
2005                 pr_debug2("switching off ksymbol\n");
2006                 goto fallback_missing_features;
2007         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
2008                 perf_missing_features.write_backward = true;
2009                 pr_debug2("switching off write_backward\n");
2010                 goto out_close;
2011         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2012                 perf_missing_features.clockid_wrong = true;
2013                 pr_debug2("switching off clockid\n");
2014                 goto fallback_missing_features;
2015         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2016                 perf_missing_features.clockid = true;
2017                 pr_debug2("switching off use_clockid\n");
2018                 goto fallback_missing_features;
2019         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2020                 perf_missing_features.cloexec = true;
2021                 pr_debug2("switching off cloexec flag\n");
2022                 goto fallback_missing_features;
2023         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2024                 perf_missing_features.mmap2 = true;
2025                 pr_debug2("switching off mmap2\n");
2026                 goto fallback_missing_features;
2027         } else if (!perf_missing_features.exclude_guest &&
2028                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2029                 perf_missing_features.exclude_guest = true;
2030                 pr_debug2("switching off exclude_guest, exclude_host\n");
2031                 goto fallback_missing_features;
2032         } else if (!perf_missing_features.sample_id_all) {
2033                 perf_missing_features.sample_id_all = true;
2034                 pr_debug2("switching off sample_id_all\n");
2035                 goto retry_sample_id;
2036         } else if (!perf_missing_features.lbr_flags &&
2037                         (evsel->attr.branch_sample_type &
2038                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2039                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2040                 perf_missing_features.lbr_flags = true;
2041                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2042                 goto fallback_missing_features;
2043         } else if (!perf_missing_features.group_read &&
2044                     evsel->attr.inherit &&
2045                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2046                    perf_evsel__is_group_leader(evsel)) {
2047                 perf_missing_features.group_read = true;
2048                 pr_debug2("switching off group read\n");
2049                 goto fallback_missing_features;
2050         }
2051 out_close:
2052         if (err)
2053                 threads->err_thread = thread;
2054
2055         do {
2056                 while (--thread >= 0) {
2057                         close(FD(evsel, cpu, thread));
2058                         FD(evsel, cpu, thread) = -1;
2059                 }
2060                 thread = nthreads;
2061         } while (--cpu >= 0);
2062         return err;
2063 }
2064
2065 void perf_evsel__close(struct perf_evsel *evsel)
2066 {
2067         if (evsel->fd == NULL)
2068                 return;
2069
2070         perf_evsel__close_fd(evsel);
2071         perf_evsel__free_fd(evsel);
2072 }
2073
2074 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2075                              struct cpu_map *cpus)
2076 {
2077         return perf_evsel__open(evsel, cpus, NULL);
2078 }
2079
2080 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2081                                 struct thread_map *threads)
2082 {
2083         return perf_evsel__open(evsel, NULL, threads);
2084 }
2085
2086 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2087                                        const union perf_event *event,
2088                                        struct perf_sample *sample)
2089 {
2090         u64 type = evsel->attr.sample_type;
2091         const u64 *array = event->sample.array;
2092         bool swapped = evsel->needs_swap;
2093         union u64_swap u;
2094
2095         array += ((event->header.size -
2096                    sizeof(event->header)) / sizeof(u64)) - 1;
2097
2098         if (type & PERF_SAMPLE_IDENTIFIER) {
2099                 sample->id = *array;
2100                 array--;
2101         }
2102
2103         if (type & PERF_SAMPLE_CPU) {
2104                 u.val64 = *array;
2105                 if (swapped) {
2106                         /* undo swap of u64, then swap on individual u32s */
2107                         u.val64 = bswap_64(u.val64);
2108                         u.val32[0] = bswap_32(u.val32[0]);
2109                 }
2110
2111                 sample->cpu = u.val32[0];
2112                 array--;
2113         }
2114
2115         if (type & PERF_SAMPLE_STREAM_ID) {
2116                 sample->stream_id = *array;
2117                 array--;
2118         }
2119
2120         if (type & PERF_SAMPLE_ID) {
2121                 sample->id = *array;
2122                 array--;
2123         }
2124
2125         if (type & PERF_SAMPLE_TIME) {
2126                 sample->time = *array;
2127                 array--;
2128         }
2129
2130         if (type & PERF_SAMPLE_TID) {
2131                 u.val64 = *array;
2132                 if (swapped) {
2133                         /* undo swap of u64, then swap on individual u32s */
2134                         u.val64 = bswap_64(u.val64);
2135                         u.val32[0] = bswap_32(u.val32[0]);
2136                         u.val32[1] = bswap_32(u.val32[1]);
2137                 }
2138
2139                 sample->pid = u.val32[0];
2140                 sample->tid = u.val32[1];
2141                 array--;
2142         }
2143
2144         return 0;
2145 }
2146
2147 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2148                             u64 size)
2149 {
2150         return size > max_size || offset + size > endp;
2151 }
2152
2153 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2154         do {                                                            \
2155                 if (overflow(endp, (max_size), (offset), (size)))       \
2156                         return -EFAULT;                                 \
2157         } while (0)
2158
2159 #define OVERFLOW_CHECK_u64(offset) \
2160         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2161
2162 static int
2163 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2164 {
2165         /*
2166          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2167          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2168          * check the format does not go past the end of the event.
2169          */
2170         if (sample_size + sizeof(event->header) > event->header.size)
2171                 return -EFAULT;
2172
2173         return 0;
2174 }
2175
2176 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2177                              struct perf_sample *data)
2178 {
2179         u64 type = evsel->attr.sample_type;
2180         bool swapped = evsel->needs_swap;
2181         const u64 *array;
2182         u16 max_size = event->header.size;
2183         const void *endp = (void *)event + max_size;
2184         u64 sz;
2185
2186         /*
2187          * used for cross-endian analysis. See git commit 65014ab3
2188          * for why this goofiness is needed.
2189          */
2190         union u64_swap u;
2191
2192         memset(data, 0, sizeof(*data));
2193         data->cpu = data->pid = data->tid = -1;
2194         data->stream_id = data->id = data->time = -1ULL;
2195         data->period = evsel->attr.sample_period;
2196         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2197         data->misc    = event->header.misc;
2198         data->id = -1ULL;
2199         data->data_src = PERF_MEM_DATA_SRC_NONE;
2200
2201         if (event->header.type != PERF_RECORD_SAMPLE) {
2202                 if (!evsel->attr.sample_id_all)
2203                         return 0;
2204                 return perf_evsel__parse_id_sample(evsel, event, data);
2205         }
2206
2207         array = event->sample.array;
2208
2209         if (perf_event__check_size(event, evsel->sample_size))
2210                 return -EFAULT;
2211
2212         if (type & PERF_SAMPLE_IDENTIFIER) {
2213                 data->id = *array;
2214                 array++;
2215         }
2216
2217         if (type & PERF_SAMPLE_IP) {
2218                 data->ip = *array;
2219                 array++;
2220         }
2221
2222         if (type & PERF_SAMPLE_TID) {
2223                 u.val64 = *array;
2224                 if (swapped) {
2225                         /* undo swap of u64, then swap on individual u32s */
2226                         u.val64 = bswap_64(u.val64);
2227                         u.val32[0] = bswap_32(u.val32[0]);
2228                         u.val32[1] = bswap_32(u.val32[1]);
2229                 }
2230
2231                 data->pid = u.val32[0];
2232                 data->tid = u.val32[1];
2233                 array++;
2234         }
2235
2236         if (type & PERF_SAMPLE_TIME) {
2237                 data->time = *array;
2238                 array++;
2239         }
2240
2241         if (type & PERF_SAMPLE_ADDR) {
2242                 data->addr = *array;
2243                 array++;
2244         }
2245
2246         if (type & PERF_SAMPLE_ID) {
2247                 data->id = *array;
2248                 array++;
2249         }
2250
2251         if (type & PERF_SAMPLE_STREAM_ID) {
2252                 data->stream_id = *array;
2253                 array++;
2254         }
2255
2256         if (type & PERF_SAMPLE_CPU) {
2257
2258                 u.val64 = *array;
2259                 if (swapped) {
2260                         /* undo swap of u64, then swap on individual u32s */
2261                         u.val64 = bswap_64(u.val64);
2262                         u.val32[0] = bswap_32(u.val32[0]);
2263                 }
2264
2265                 data->cpu = u.val32[0];
2266                 array++;
2267         }
2268
2269         if (type & PERF_SAMPLE_PERIOD) {
2270                 data->period = *array;
2271                 array++;
2272         }
2273
2274         if (type & PERF_SAMPLE_READ) {
2275                 u64 read_format = evsel->attr.read_format;
2276
2277                 OVERFLOW_CHECK_u64(array);
2278                 if (read_format & PERF_FORMAT_GROUP)
2279                         data->read.group.nr = *array;
2280                 else
2281                         data->read.one.value = *array;
2282
2283                 array++;
2284
2285                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2286                         OVERFLOW_CHECK_u64(array);
2287                         data->read.time_enabled = *array;
2288                         array++;
2289                 }
2290
2291                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2292                         OVERFLOW_CHECK_u64(array);
2293                         data->read.time_running = *array;
2294                         array++;
2295                 }
2296
2297                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2298                 if (read_format & PERF_FORMAT_GROUP) {
2299                         const u64 max_group_nr = UINT64_MAX /
2300                                         sizeof(struct sample_read_value);
2301
2302                         if (data->read.group.nr > max_group_nr)
2303                                 return -EFAULT;
2304                         sz = data->read.group.nr *
2305                              sizeof(struct sample_read_value);
2306                         OVERFLOW_CHECK(array, sz, max_size);
2307                         data->read.group.values =
2308                                         (struct sample_read_value *)array;
2309                         array = (void *)array + sz;
2310                 } else {
2311                         OVERFLOW_CHECK_u64(array);
2312                         data->read.one.id = *array;
2313                         array++;
2314                 }
2315         }
2316
2317         if (evsel__has_callchain(evsel)) {
2318                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2319
2320                 OVERFLOW_CHECK_u64(array);
2321                 data->callchain = (struct ip_callchain *)array++;
2322                 if (data->callchain->nr > max_callchain_nr)
2323                         return -EFAULT;
2324                 sz = data->callchain->nr * sizeof(u64);
2325                 OVERFLOW_CHECK(array, sz, max_size);
2326                 array = (void *)array + sz;
2327         }
2328
2329         if (type & PERF_SAMPLE_RAW) {
2330                 OVERFLOW_CHECK_u64(array);
2331                 u.val64 = *array;
2332
2333                 /*
2334                  * Undo swap of u64, then swap on individual u32s,
2335                  * get the size of the raw area and undo all of the
2336                  * swap. The pevent interface handles endianity by
2337                  * itself.
2338                  */
2339                 if (swapped) {
2340                         u.val64 = bswap_64(u.val64);
2341                         u.val32[0] = bswap_32(u.val32[0]);
2342                         u.val32[1] = bswap_32(u.val32[1]);
2343                 }
2344                 data->raw_size = u.val32[0];
2345
2346                 /*
2347                  * The raw data is aligned on 64bits including the
2348                  * u32 size, so it's safe to use mem_bswap_64.
2349                  */
2350                 if (swapped)
2351                         mem_bswap_64((void *) array, data->raw_size);
2352
2353                 array = (void *)array + sizeof(u32);
2354
2355                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2356                 data->raw_data = (void *)array;
2357                 array = (void *)array + data->raw_size;
2358         }
2359
2360         if (type & PERF_SAMPLE_BRANCH_STACK) {
2361                 const u64 max_branch_nr = UINT64_MAX /
2362                                           sizeof(struct branch_entry);
2363
2364                 OVERFLOW_CHECK_u64(array);
2365                 data->branch_stack = (struct branch_stack *)array++;
2366
2367                 if (data->branch_stack->nr > max_branch_nr)
2368                         return -EFAULT;
2369                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2370                 OVERFLOW_CHECK(array, sz, max_size);
2371                 array = (void *)array + sz;
2372         }
2373
2374         if (type & PERF_SAMPLE_REGS_USER) {
2375                 OVERFLOW_CHECK_u64(array);
2376                 data->user_regs.abi = *array;
2377                 array++;
2378
2379                 if (data->user_regs.abi) {
2380                         u64 mask = evsel->attr.sample_regs_user;
2381
2382                         sz = hweight64(mask) * sizeof(u64);
2383                         OVERFLOW_CHECK(array, sz, max_size);
2384                         data->user_regs.mask = mask;
2385                         data->user_regs.regs = (u64 *)array;
2386                         array = (void *)array + sz;
2387                 }
2388         }
2389
2390         if (type & PERF_SAMPLE_STACK_USER) {
2391                 OVERFLOW_CHECK_u64(array);
2392                 sz = *array++;
2393
2394                 data->user_stack.offset = ((char *)(array - 1)
2395                                           - (char *) event);
2396
2397                 if (!sz) {
2398                         data->user_stack.size = 0;
2399                 } else {
2400                         OVERFLOW_CHECK(array, sz, max_size);
2401                         data->user_stack.data = (char *)array;
2402                         array = (void *)array + sz;
2403                         OVERFLOW_CHECK_u64(array);
2404                         data->user_stack.size = *array++;
2405                         if (WARN_ONCE(data->user_stack.size > sz,
2406                                       "user stack dump failure\n"))
2407                                 return -EFAULT;
2408                 }
2409         }
2410
2411         if (type & PERF_SAMPLE_WEIGHT) {
2412                 OVERFLOW_CHECK_u64(array);
2413                 data->weight = *array;
2414                 array++;
2415         }
2416
2417         if (type & PERF_SAMPLE_DATA_SRC) {
2418                 OVERFLOW_CHECK_u64(array);
2419                 data->data_src = *array;
2420                 array++;
2421         }
2422
2423         if (type & PERF_SAMPLE_TRANSACTION) {
2424                 OVERFLOW_CHECK_u64(array);
2425                 data->transaction = *array;
2426                 array++;
2427         }
2428
2429         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2430         if (type & PERF_SAMPLE_REGS_INTR) {
2431                 OVERFLOW_CHECK_u64(array);
2432                 data->intr_regs.abi = *array;
2433                 array++;
2434
2435                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2436                         u64 mask = evsel->attr.sample_regs_intr;
2437
2438                         sz = hweight64(mask) * sizeof(u64);
2439                         OVERFLOW_CHECK(array, sz, max_size);
2440                         data->intr_regs.mask = mask;
2441                         data->intr_regs.regs = (u64 *)array;
2442                         array = (void *)array + sz;
2443                 }
2444         }
2445
2446         data->phys_addr = 0;
2447         if (type & PERF_SAMPLE_PHYS_ADDR) {
2448                 data->phys_addr = *array;
2449                 array++;
2450         }
2451
2452         return 0;
2453 }
2454
2455 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2456                                        union perf_event *event,
2457                                        u64 *timestamp)
2458 {
2459         u64 type = evsel->attr.sample_type;
2460         const u64 *array;
2461
2462         if (!(type & PERF_SAMPLE_TIME))
2463                 return -1;
2464
2465         if (event->header.type != PERF_RECORD_SAMPLE) {
2466                 struct perf_sample data = {
2467                         .time = -1ULL,
2468                 };
2469
2470                 if (!evsel->attr.sample_id_all)
2471                         return -1;
2472                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2473                         return -1;
2474
2475                 *timestamp = data.time;
2476                 return 0;
2477         }
2478
2479         array = event->sample.array;
2480
2481         if (perf_event__check_size(event, evsel->sample_size))
2482                 return -EFAULT;
2483
2484         if (type & PERF_SAMPLE_IDENTIFIER)
2485                 array++;
2486
2487         if (type & PERF_SAMPLE_IP)
2488                 array++;
2489
2490         if (type & PERF_SAMPLE_TID)
2491                 array++;
2492
2493         if (type & PERF_SAMPLE_TIME)
2494                 *timestamp = *array;
2495
2496         return 0;
2497 }
2498
2499 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2500                                      u64 read_format)
2501 {
2502         size_t sz, result = sizeof(struct sample_event);
2503
2504         if (type & PERF_SAMPLE_IDENTIFIER)
2505                 result += sizeof(u64);
2506
2507         if (type & PERF_SAMPLE_IP)
2508                 result += sizeof(u64);
2509
2510         if (type & PERF_SAMPLE_TID)
2511                 result += sizeof(u64);
2512
2513         if (type & PERF_SAMPLE_TIME)
2514                 result += sizeof(u64);
2515
2516         if (type & PERF_SAMPLE_ADDR)
2517                 result += sizeof(u64);
2518
2519         if (type & PERF_SAMPLE_ID)
2520                 result += sizeof(u64);
2521
2522         if (type & PERF_SAMPLE_STREAM_ID)
2523                 result += sizeof(u64);
2524
2525         if (type & PERF_SAMPLE_CPU)
2526                 result += sizeof(u64);
2527
2528         if (type & PERF_SAMPLE_PERIOD)
2529                 result += sizeof(u64);
2530
2531         if (type & PERF_SAMPLE_READ) {
2532                 result += sizeof(u64);
2533                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2534                         result += sizeof(u64);
2535                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2536                         result += sizeof(u64);
2537                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2538                 if (read_format & PERF_FORMAT_GROUP) {
2539                         sz = sample->read.group.nr *
2540                              sizeof(struct sample_read_value);
2541                         result += sz;
2542                 } else {
2543                         result += sizeof(u64);
2544                 }
2545         }
2546
2547         if (type & PERF_SAMPLE_CALLCHAIN) {
2548                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2549                 result += sz;
2550         }
2551
2552         if (type & PERF_SAMPLE_RAW) {
2553                 result += sizeof(u32);
2554                 result += sample->raw_size;
2555         }
2556
2557         if (type & PERF_SAMPLE_BRANCH_STACK) {
2558                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2559                 sz += sizeof(u64);
2560                 result += sz;
2561         }
2562
2563         if (type & PERF_SAMPLE_REGS_USER) {
2564                 if (sample->user_regs.abi) {
2565                         result += sizeof(u64);
2566                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2567                         result += sz;
2568                 } else {
2569                         result += sizeof(u64);
2570                 }
2571         }
2572
2573         if (type & PERF_SAMPLE_STACK_USER) {
2574                 sz = sample->user_stack.size;
2575                 result += sizeof(u64);
2576                 if (sz) {
2577                         result += sz;
2578                         result += sizeof(u64);
2579                 }
2580         }
2581
2582         if (type & PERF_SAMPLE_WEIGHT)
2583                 result += sizeof(u64);
2584
2585         if (type & PERF_SAMPLE_DATA_SRC)
2586                 result += sizeof(u64);
2587
2588         if (type & PERF_SAMPLE_TRANSACTION)
2589                 result += sizeof(u64);
2590
2591         if (type & PERF_SAMPLE_REGS_INTR) {
2592                 if (sample->intr_regs.abi) {
2593                         result += sizeof(u64);
2594                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2595                         result += sz;
2596                 } else {
2597                         result += sizeof(u64);
2598                 }
2599         }
2600
2601         if (type & PERF_SAMPLE_PHYS_ADDR)
2602                 result += sizeof(u64);
2603
2604         return result;
2605 }
2606
2607 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2608                                   u64 read_format,
2609                                   const struct perf_sample *sample)
2610 {
2611         u64 *array;
2612         size_t sz;
2613         /*
2614          * used for cross-endian analysis. See git commit 65014ab3
2615          * for why this goofiness is needed.
2616          */
2617         union u64_swap u;
2618
2619         array = event->sample.array;
2620
2621         if (type & PERF_SAMPLE_IDENTIFIER) {
2622                 *array = sample->id;
2623                 array++;
2624         }
2625
2626         if (type & PERF_SAMPLE_IP) {
2627                 *array = sample->ip;
2628                 array++;
2629         }
2630
2631         if (type & PERF_SAMPLE_TID) {
2632                 u.val32[0] = sample->pid;
2633                 u.val32[1] = sample->tid;
2634                 *array = u.val64;
2635                 array++;
2636         }
2637
2638         if (type & PERF_SAMPLE_TIME) {
2639                 *array = sample->time;
2640                 array++;
2641         }
2642
2643         if (type & PERF_SAMPLE_ADDR) {
2644                 *array = sample->addr;
2645                 array++;
2646         }
2647
2648         if (type & PERF_SAMPLE_ID) {
2649                 *array = sample->id;
2650                 array++;
2651         }
2652
2653         if (type & PERF_SAMPLE_STREAM_ID) {
2654                 *array = sample->stream_id;
2655                 array++;
2656         }
2657
2658         if (type & PERF_SAMPLE_CPU) {
2659                 u.val32[0] = sample->cpu;
2660                 u.val32[1] = 0;
2661                 *array = u.val64;
2662                 array++;
2663         }
2664
2665         if (type & PERF_SAMPLE_PERIOD) {
2666                 *array = sample->period;
2667                 array++;
2668         }
2669
2670         if (type & PERF_SAMPLE_READ) {
2671                 if (read_format & PERF_FORMAT_GROUP)
2672                         *array = sample->read.group.nr;
2673                 else
2674                         *array = sample->read.one.value;
2675                 array++;
2676
2677                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2678                         *array = sample->read.time_enabled;
2679                         array++;
2680                 }
2681
2682                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2683                         *array = sample->read.time_running;
2684                         array++;
2685                 }
2686
2687                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2688                 if (read_format & PERF_FORMAT_GROUP) {
2689                         sz = sample->read.group.nr *
2690                              sizeof(struct sample_read_value);
2691                         memcpy(array, sample->read.group.values, sz);
2692                         array = (void *)array + sz;
2693                 } else {
2694                         *array = sample->read.one.id;
2695                         array++;
2696                 }
2697         }
2698
2699         if (type & PERF_SAMPLE_CALLCHAIN) {
2700                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2701                 memcpy(array, sample->callchain, sz);
2702                 array = (void *)array + sz;
2703         }
2704
2705         if (type & PERF_SAMPLE_RAW) {
2706                 u.val32[0] = sample->raw_size;
2707                 *array = u.val64;
2708                 array = (void *)array + sizeof(u32);
2709
2710                 memcpy(array, sample->raw_data, sample->raw_size);
2711                 array = (void *)array + sample->raw_size;
2712         }
2713
2714         if (type & PERF_SAMPLE_BRANCH_STACK) {
2715                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2716                 sz += sizeof(u64);
2717                 memcpy(array, sample->branch_stack, sz);
2718                 array = (void *)array + sz;
2719         }
2720
2721         if (type & PERF_SAMPLE_REGS_USER) {
2722                 if (sample->user_regs.abi) {
2723                         *array++ = sample->user_regs.abi;
2724                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2725                         memcpy(array, sample->user_regs.regs, sz);
2726                         array = (void *)array + sz;
2727                 } else {
2728                         *array++ = 0;
2729                 }
2730         }
2731
2732         if (type & PERF_SAMPLE_STACK_USER) {
2733                 sz = sample->user_stack.size;
2734                 *array++ = sz;
2735                 if (sz) {
2736                         memcpy(array, sample->user_stack.data, sz);
2737                         array = (void *)array + sz;
2738                         *array++ = sz;
2739                 }
2740         }
2741
2742         if (type & PERF_SAMPLE_WEIGHT) {
2743                 *array = sample->weight;
2744                 array++;
2745         }
2746
2747         if (type & PERF_SAMPLE_DATA_SRC) {
2748                 *array = sample->data_src;
2749                 array++;
2750         }
2751
2752         if (type & PERF_SAMPLE_TRANSACTION) {
2753                 *array = sample->transaction;
2754                 array++;
2755         }
2756
2757         if (type & PERF_SAMPLE_REGS_INTR) {
2758                 if (sample->intr_regs.abi) {
2759                         *array++ = sample->intr_regs.abi;
2760                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2761                         memcpy(array, sample->intr_regs.regs, sz);
2762                         array = (void *)array + sz;
2763                 } else {
2764                         *array++ = 0;
2765                 }
2766         }
2767
2768         if (type & PERF_SAMPLE_PHYS_ADDR) {
2769                 *array = sample->phys_addr;
2770                 array++;
2771         }
2772
2773         return 0;
2774 }
2775
2776 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2777 {
2778         return tep_find_field(evsel->tp_format, name);
2779 }
2780
2781 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2782                          const char *name)
2783 {
2784         struct tep_format_field *field = perf_evsel__field(evsel, name);
2785         int offset;
2786
2787         if (!field)
2788                 return NULL;
2789
2790         offset = field->offset;
2791
2792         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2793                 offset = *(int *)(sample->raw_data + field->offset);
2794                 offset &= 0xffff;
2795         }
2796
2797         return sample->raw_data + offset;
2798 }
2799
2800 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2801                          bool needs_swap)
2802 {
2803         u64 value;
2804         void *ptr = sample->raw_data + field->offset;
2805
2806         switch (field->size) {
2807         case 1:
2808                 return *(u8 *)ptr;
2809         case 2:
2810                 value = *(u16 *)ptr;
2811                 break;
2812         case 4:
2813                 value = *(u32 *)ptr;
2814                 break;
2815         case 8:
2816                 memcpy(&value, ptr, sizeof(u64));
2817                 break;
2818         default:
2819                 return 0;
2820         }
2821
2822         if (!needs_swap)
2823                 return value;
2824
2825         switch (field->size) {
2826         case 2:
2827                 return bswap_16(value);
2828         case 4:
2829                 return bswap_32(value);
2830         case 8:
2831                 return bswap_64(value);
2832         default:
2833                 return 0;
2834         }
2835
2836         return 0;
2837 }
2838
2839 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2840                        const char *name)
2841 {
2842         struct tep_format_field *field = perf_evsel__field(evsel, name);
2843
2844         if (!field)
2845                 return 0;
2846
2847         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2848 }
2849
2850 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2851                           char *msg, size_t msgsize)
2852 {
2853         int paranoid;
2854
2855         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2856             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2857             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2858                 /*
2859                  * If it's cycles then fall back to hrtimer based
2860                  * cpu-clock-tick sw counter, which is always available even if
2861                  * no PMU support.
2862                  *
2863                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2864                  * b0a873e).
2865                  */
2866                 scnprintf(msg, msgsize, "%s",
2867 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2868
2869                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2870                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2871
2872                 zfree(&evsel->name);
2873                 return true;
2874         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2875                    (paranoid = perf_event_paranoid()) > 1) {
2876                 const char *name = perf_evsel__name(evsel);
2877                 char *new_name;
2878                 const char *sep = ":";
2879
2880                 /* Is there already the separator in the name. */
2881                 if (strchr(name, '/') ||
2882                     strchr(name, ':'))
2883                         sep = "";
2884
2885                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2886                         return false;
2887
2888                 if (evsel->name)
2889                         free(evsel->name);
2890                 evsel->name = new_name;
2891                 scnprintf(msg, msgsize,
2892 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2893                 evsel->attr.exclude_kernel = 1;
2894
2895                 return true;
2896         }
2897
2898         return false;
2899 }
2900
2901 static bool find_process(const char *name)
2902 {
2903         size_t len = strlen(name);
2904         DIR *dir;
2905         struct dirent *d;
2906         int ret = -1;
2907
2908         dir = opendir(procfs__mountpoint());
2909         if (!dir)
2910                 return false;
2911
2912         /* Walk through the directory. */
2913         while (ret && (d = readdir(dir)) != NULL) {
2914                 char path[PATH_MAX];
2915                 char *data;
2916                 size_t size;
2917
2918                 if ((d->d_type != DT_DIR) ||
2919                      !strcmp(".", d->d_name) ||
2920                      !strcmp("..", d->d_name))
2921                         continue;
2922
2923                 scnprintf(path, sizeof(path), "%s/%s/comm",
2924                           procfs__mountpoint(), d->d_name);
2925
2926                 if (filename__read_str(path, &data, &size))
2927                         continue;
2928
2929                 ret = strncmp(name, data, len);
2930                 free(data);
2931         }
2932
2933         closedir(dir);
2934         return ret ? false : true;
2935 }
2936
2937 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2938                               int err, char *msg, size_t size)
2939 {
2940         char sbuf[STRERR_BUFSIZE];
2941         int printed = 0;
2942
2943         switch (err) {
2944         case EPERM:
2945         case EACCES:
2946                 if (err == EPERM)
2947                         printed = scnprintf(msg, size,
2948                                 "No permission to enable %s event.\n\n",
2949                                 perf_evsel__name(evsel));
2950
2951                 return scnprintf(msg + printed, size - printed,
2952                  "You may not have permission to collect %sstats.\n\n"
2953                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2954                  "which controls use of the performance events system by\n"
2955                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2956                  "The current value is %d:\n\n"
2957                  "  -1: Allow use of (almost) all events by all users\n"
2958                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2959                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2960                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2961                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2962                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2963                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2964                  "      kernel.perf_event_paranoid = -1\n" ,
2965                                  target->system_wide ? "system-wide " : "",
2966                                  perf_event_paranoid());
2967         case ENOENT:
2968                 return scnprintf(msg, size, "The %s event is not supported.",
2969                                  perf_evsel__name(evsel));
2970         case EMFILE:
2971                 return scnprintf(msg, size, "%s",
2972                          "Too many events are opened.\n"
2973                          "Probably the maximum number of open file descriptors has been reached.\n"
2974                          "Hint: Try again after reducing the number of events.\n"
2975                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2976         case ENOMEM:
2977                 if (evsel__has_callchain(evsel) &&
2978                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2979                         return scnprintf(msg, size,
2980                                          "Not enough memory to setup event with callchain.\n"
2981                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2982                                          "Hint: Current value: %d", sysctl__max_stack());
2983                 break;
2984         case ENODEV:
2985                 if (target->cpu_list)
2986                         return scnprintf(msg, size, "%s",
2987          "No such device - did you specify an out-of-range profile CPU?");
2988                 break;
2989         case EOPNOTSUPP:
2990                 if (evsel->attr.sample_period != 0)
2991                         return scnprintf(msg, size,
2992         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2993                                          perf_evsel__name(evsel));
2994                 if (evsel->attr.precise_ip)
2995                         return scnprintf(msg, size, "%s",
2996         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2997 #if defined(__i386__) || defined(__x86_64__)
2998                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2999                         return scnprintf(msg, size, "%s",
3000         "No hardware sampling interrupt available.\n");
3001 #endif
3002                 break;
3003         case EBUSY:
3004                 if (find_process("oprofiled"))
3005                         return scnprintf(msg, size,
3006         "The PMU counters are busy/taken by another profiler.\n"
3007         "We found oprofile daemon running, please stop it and try again.");
3008                 break;
3009         case EINVAL:
3010                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3011                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3012                 if (perf_missing_features.clockid)
3013                         return scnprintf(msg, size, "clockid feature not supported.");
3014                 if (perf_missing_features.clockid_wrong)
3015                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3016                 break;
3017         default:
3018                 break;
3019         }
3020
3021         return scnprintf(msg, size,
3022         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3023         "/bin/dmesg | grep -i perf may provide additional information.\n",
3024                          err, str_error_r(err, sbuf, sizeof(sbuf)),
3025                          perf_evsel__name(evsel));
3026 }
3027
3028 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3029 {
3030         if (evsel && evsel->evlist)
3031                 return evsel->evlist->env;
3032         return NULL;
3033 }
3034
3035 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3036 {
3037         int cpu, thread;
3038
3039         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3040                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3041                      thread++) {
3042                         int fd = FD(evsel, cpu, thread);
3043
3044                         if (perf_evlist__id_add_fd(evlist, evsel,
3045                                                    cpu, thread, fd) < 0)
3046                                 return -1;
3047                 }
3048         }
3049
3050         return 0;
3051 }
3052
3053 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3054 {
3055         struct cpu_map *cpus = evsel->cpus;
3056         struct thread_map *threads = evsel->threads;
3057
3058         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3059                 return -ENOMEM;
3060
3061         return store_evsel_ids(evsel, evlist);
3062 }