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->evlist      = NULL;
 236         evsel->bpf_fd      = -1;
 237         INIT_LIST_HEAD(&evsel->node);
 238         INIT_LIST_HEAD(&evsel->config_terms);
 239         perf_evsel__object.init(evsel);
 240         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 241         perf_evsel__calc_id_pos(evsel);
 242         evsel->cmdline_group_boundary = false;
 243         evsel->metric_expr   = NULL;
 244         evsel->metric_name   = NULL;
 245         evsel->metric_events = NULL;
 246         evsel->collect_stat  = false;
 247         evsel->pmu_name      = NULL;
 248 }
 249
 250 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 251 {
 252         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 253
 254         if (evsel != NULL)
 255                 perf_evsel__init(evsel, attr, idx);
 256
 257         if (perf_evsel__is_bpf_output(evsel)) {
 258                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 259                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 260                 evsel->attr.sample_period = 1;
 261         }
 262
 263         return evsel;
 264 }
 265
 266 static bool perf_event_can_profile_kernel(void)
 267 {
 268         return geteuid() == 0 || perf_event_paranoid() == -1;
 269 }
 270
 271 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 272 {
 273         struct perf_event_attr attr = {
 274                 .type   = PERF_TYPE_HARDWARE,
 275                 .config = PERF_COUNT_HW_CPU_CYCLES,
 276                 .exclude_kernel = !perf_event_can_profile_kernel(),
 277         };
 278         struct perf_evsel *evsel;
 279
 280         event_attr_init(&attr);
 281
 282         if (!precise)
 283                 goto new_event;
 284         /*
 285          * Unnamed union member, not supported as struct member named
 286          * initializer in older compilers such as gcc 4.4.7
 287          *
 288          * Just for probing the precise_ip:
 289          */
 290         attr.sample_period = 1;
 291
 292         perf_event_attr__set_max_precise_ip(&attr);
 293         /*
 294          * Now let the usual logic to set up the perf_event_attr defaults
 295          * to kick in when we return and before perf_evsel__open() is called.
 296          */
 297         attr.sample_period = 0;
 298 new_event:
 299         evsel = perf_evsel__new(&attr);
 300         if (evsel == NULL)
 301                 goto out;
 302
 303         /* use asprintf() because free(evsel) assumes name is allocated */
 304         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 305                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 306                      attr.exclude_kernel ? "u" : "",
 307                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 308                 goto error_free;
 309 out:
 310         return evsel;
 311 error_free:
 312         perf_evsel__delete(evsel);
 313         evsel = NULL;
 314         goto out;
 315 }
 316
 317 /*
 318  * Returns pointer with encoded error via <linux/err.h> interface.
 319  */
 320 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 321 {
 322         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 323         int err = -ENOMEM;
 324
 325         if (evsel == NULL) {
 326                 goto out_err;
 327         } else {
 328                 struct perf_event_attr attr = {
 329                         .type          = PERF_TYPE_TRACEPOINT,
 330                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 331                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 332                 };
 333
 334                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 335                         goto out_free;
 336
 337                 evsel->tp_format = trace_event__tp_format(sys, name);
 338                 if (IS_ERR(evsel->tp_format)) {
 339                         err = PTR_ERR(evsel->tp_format);
 340                         goto out_free;
 341                 }
 342
 343                 event_attr_init(&attr);
 344                 attr.config = evsel->tp_format->id;
 345                 attr.sample_period = 1;
 346                 perf_evsel__init(evsel, &attr, idx);
 347         }
 348
 349         return evsel;
 350
 351 out_free:
 352         zfree(&evsel->name);
 353         free(evsel);
 354 out_err:
 355         return ERR_PTR(err);
 356 }
 357
 358 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 359         "cycles",
 360         "instructions",
 361         "cache-references",
 362         "cache-misses",
 363         "branches",
 364         "branch-misses",
 365         "bus-cycles",
 366         "stalled-cycles-frontend",
 367         "stalled-cycles-backend",
 368         "ref-cycles",
 369 };
 370
 371 static const char *__perf_evsel__hw_name(u64 config)
 372 {
 373         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 374                 return perf_evsel__hw_names[config];
 375
 376         return "unknown-hardware";
 377 }
 378
 379 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 380 {
 381         int colon = 0, r = 0;
 382         struct perf_event_attr *attr = &evsel->attr;
 383         bool exclude_guest_default = false;
 384
 385 #define MOD_PRINT(context, mod) do {                                    \
 386                 if (!attr->exclude_##context) {                         \
 387                         if (!colon) colon = ++r;                        \
 388                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 389                 } } while(0)
 390
 391         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 392                 MOD_PRINT(kernel, 'k');
 393                 MOD_PRINT(user, 'u');
 394                 MOD_PRINT(hv, 'h');
 395                 exclude_guest_default = true;
 396         }
 397
 398         if (attr->precise_ip) {
 399                 if (!colon)
 400                         colon = ++r;
 401                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 402                 exclude_guest_default = true;
 403         }
 404
 405         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 406                 MOD_PRINT(host, 'H');
 407                 MOD_PRINT(guest, 'G');
 408         }
 409 #undef MOD_PRINT
 410         if (colon)
 411                 bf[colon - 1] = ':';
 412         return r;
 413 }
 414
 415 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 416 {
 417         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 418         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 419 }
 420
 421 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 422         "cpu-clock",
 423         "task-clock",
 424         "page-faults",
 425         "context-switches",
 426         "cpu-migrations",
 427         "minor-faults",
 428         "major-faults",
 429         "alignment-faults",
 430         "emulation-faults",
 431         "dummy",
 432 };
 433
 434 static const char *__perf_evsel__sw_name(u64 config)
 435 {
 436         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 437                 return perf_evsel__sw_names[config];
 438         return "unknown-software";
 439 }
 440
 441 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 442 {
 443         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 444         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 445 }
 446
 447 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 448 {
 449         int r;
 450
 451         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 452
 453         if (type & HW_BREAKPOINT_R)
 454                 r += scnprintf(bf + r, size - r, "r");
 455
 456         if (type & HW_BREAKPOINT_W)
 457                 r += scnprintf(bf + r, size - r, "w");
 458
 459         if (type & HW_BREAKPOINT_X)
 460                 r += scnprintf(bf + r, size - r, "x");
 461
 462         return r;
 463 }
 464
 465 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 466 {
 467         struct perf_event_attr *attr = &evsel->attr;
 468         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 469         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 470 }
 471
 472 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 473                                 [PERF_EVSEL__MAX_ALIASES] = {
 474  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 475  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 476  { "LLC",       "L2",                                                   },
 477  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 478  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 479  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 480  { "node",                                                              },
 481 };
 482
 483 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 484                                    [PERF_EVSEL__MAX_ALIASES] = {
 485  { "load",      "loads",        "read",                                 },
 486  { "store",     "stores",       "write",                                },
 487  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 488 };
 489
 490 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 491                                        [PERF_EVSEL__MAX_ALIASES] = {
 492  { "refs",      "Reference",    "ops",          "access",               },
 493  { "misses",    "miss",                                                 },
 494 };
 495
 496 #define C(x)            PERF_COUNT_HW_CACHE_##x
 497 #define CACHE_READ      (1 << C(OP_READ))
 498 #define CACHE_WRITE     (1 << C(OP_WRITE))
 499 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 500 #define COP(x)          (1 << x)
 501
 502 /*
 503  * cache operartion stat
 504  * L1I : Read and prefetch only
 505  * ITLB and BPU : Read-only
 506  */
 507 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 508  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 509  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 510  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 511  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 512  [C(ITLB)]      = (CACHE_READ),
 513  [C(BPU)]       = (CACHE_READ),
 514  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 515 };
 516
 517 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 518 {
 519         if (perf_evsel__hw_cache_stat[type] & COP(op))
 520                 return true;    /* valid */
 521         else
 522                 return false;   /* invalid */
 523 }
 524
 525 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 526                                             char *bf, size_t size)
 527 {
 528         if (result) {
 529                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 530                                  perf_evsel__hw_cache_op[op][0],
 531                                  perf_evsel__hw_cache_result[result][0]);
 532         }
 533
 534         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 535                          perf_evsel__hw_cache_op[op][1]);
 536 }
 537
 538 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 539 {
 540         u8 op, result, type = (config >>  0) & 0xff;
 541         const char *err = "unknown-ext-hardware-cache-type";
 542
 543         if (type >= PERF_COUNT_HW_CACHE_MAX)
 544                 goto out_err;
 545
 546         op = (config >>  8) & 0xff;
 547         err = "unknown-ext-hardware-cache-op";
 548         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 549                 goto out_err;
 550
 551         result = (config >> 16) & 0xff;
 552         err = "unknown-ext-hardware-cache-result";
 553         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 554                 goto out_err;
 555
 556         err = "invalid-cache";
 557         if (!perf_evsel__is_cache_op_valid(type, op))
 558                 goto out_err;
 559
 560         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 561 out_err:
 562         return scnprintf(bf, size, "%s", err);
 563 }
 564
 565 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 566 {
 567         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 568         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 569 }
 570
 571 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 572 {
 573         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 574         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 575 }
 576
 577 const char *perf_evsel__name(struct perf_evsel *evsel)
 578 {
 579         char bf[128];
 580
 581         if (evsel->name)
 582                 return evsel->name;
 583
 584         switch (evsel->attr.type) {
 585         case PERF_TYPE_RAW:
 586                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 587                 break;
 588
 589         case PERF_TYPE_HARDWARE:
 590                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 591                 break;
 592
 593         case PERF_TYPE_HW_CACHE:
 594                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 595                 break;
 596
 597         case PERF_TYPE_SOFTWARE:
 598                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 599                 break;
 600
 601         case PERF_TYPE_TRACEPOINT:
 602                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 603                 break;
 604
 605         case PERF_TYPE_BREAKPOINT:
 606                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 607                 break;
 608
 609         default:
 610                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 611                           evsel->attr.type);
 612                 break;
 613         }
 614
 615         evsel->name = strdup(bf);
 616
 617         return evsel->name ?: "unknown";
 618 }
 619
 620 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 621 {
 622         return evsel->group_name ?: "anon group";
 623 }
 624
 625 /*
 626  * Returns the group details for the specified leader,
 627  * with following rules.
 628  *
 629  *  For record -e '{cycles,instructions}'
 630  *    'anon group { cycles:u, instructions:u }'
 631  *
 632  *  For record -e 'cycles,instructions' and report --group
 633  *    'cycles:u, instructions:u'
 634  */
 635 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 636 {
 637         int ret = 0;
 638         struct perf_evsel *pos;
 639         const char *group_name = perf_evsel__group_name(evsel);
 640
 641         if (!evsel->forced_leader)
 642                 ret = scnprintf(buf, size, "%s { ", group_name);
 643
 644         ret += scnprintf(buf + ret, size - ret, "%s",
 645                          perf_evsel__name(evsel));
 646
 647         for_each_group_member(pos, evsel)
 648                 ret += scnprintf(buf + ret, size - ret, ", %s",
 649                                  perf_evsel__name(pos));
 650
 651         if (!evsel->forced_leader)
 652                 ret += scnprintf(buf + ret, size - ret, " }");
 653
 654         return ret;
 655 }
 656
 657 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 658                                            struct record_opts *opts,
 659                                            struct callchain_param *param)
 660 {
 661         bool function = perf_evsel__is_function_event(evsel);
 662         struct perf_event_attr *attr = &evsel->attr;
 663
 664         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 665
 666         attr->sample_max_stack = param->max_stack;
 667
 668         if (param->record_mode == CALLCHAIN_LBR) {
 669                 if (!opts->branch_stack) {
 670                         if (attr->exclude_user) {
 671                                 pr_warning("LBR callstack option is only available "
 672                                            "to get user callchain information. "
 673                                            "Falling back to framepointers.\n");
 674                         } else {
 675                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 676                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 677                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 678                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 679                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 680                         }
 681                 } else
 682                          pr_warning("Cannot use LBR callstack with branch stack. "
 683                                     "Falling back to framepointers.\n");
 684         }
 685
 686         if (param->record_mode == CALLCHAIN_DWARF) {
 687                 if (!function) {
 688                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 689                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 690                         attr->sample_regs_user |= PERF_REGS_MASK;
 691                         attr->sample_stack_user = param->dump_size;
 692                         attr->exclude_callchain_user = 1;
 693                 } else {
 694                         pr_info("Cannot use DWARF unwind for function trace event,"
 695                                 " falling back to framepointers.\n");
 696                 }
 697         }
 698
 699         if (function) {
 700                 pr_info("Disabling user space callchains for function trace event.\n");
 701                 attr->exclude_callchain_user = 1;
 702         }
 703 }
 704
 705 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 706                                   struct record_opts *opts,
 707                                   struct callchain_param *param)
 708 {
 709         if (param->enabled)
 710                 return __perf_evsel__config_callchain(evsel, opts, param);
 711 }
 712
 713 static void
 714 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 715                             struct callchain_param *param)
 716 {
 717         struct perf_event_attr *attr = &evsel->attr;
 718
 719         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 720         if (param->record_mode == CALLCHAIN_LBR) {
 721                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 722                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 723                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 724         }
 725         if (param->record_mode == CALLCHAIN_DWARF) {
 726                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 727                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 728         }
 729 }
 730
 731 static void apply_config_terms(struct perf_evsel *evsel,
 732                                struct record_opts *opts, bool track)
 733 {
 734         struct perf_evsel_config_term *term;
 735         struct list_head *config_terms = &evsel->config_terms;
 736         struct perf_event_attr *attr = &evsel->attr;
 737         /* callgraph default */
 738         struct callchain_param param = {
 739                 .record_mode = callchain_param.record_mode,
 740         };
 741         u32 dump_size = 0;
 742         int max_stack = 0;
 743         const char *callgraph_buf = NULL;
 744
 745         list_for_each_entry(term, config_terms, list) {
 746                 switch (term->type) {
 747                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 748                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 749                                 attr->sample_period = term->val.period;
 750                                 attr->freq = 0;
 751                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 752                         }
 753                         break;
 754                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 755                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 756                                 attr->sample_freq = term->val.freq;
 757                                 attr->freq = 1;
 758                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 759                         }
 760                         break;
 761                 case PERF_EVSEL__CONFIG_TERM_TIME:
 762                         if (term->val.time)
 763                                 perf_evsel__set_sample_bit(evsel, TIME);
 764                         else
 765                                 perf_evsel__reset_sample_bit(evsel, TIME);
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 768                         callgraph_buf = term->val.callgraph;
 769                         break;
 770                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 771                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 772                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 773                                 parse_branch_str(term->val.branch,
 774                                                  &attr->branch_sample_type);
 775                         } else
 776                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 777                         break;
 778                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 779                         dump_size = term->val.stack_user;
 780                         break;
 781                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 782                         max_stack = term->val.max_stack;
 783                         break;
 784                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 785                         /*
 786                          * attr->inherit should has already been set by
 787                          * perf_evsel__config. If user explicitly set
 788                          * inherit using config terms, override global
 789                          * opt->no_inherit setting.
 790                          */
 791                         attr->inherit = term->val.inherit ? 1 : 0;
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 794                         attr->write_backward = term->val.overwrite ? 1 : 0;
 795                         break;
 796                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 797                         break;
 798                 default:
 799                         break;
 800                 }
 801         }
 802
 803         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 804         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 805                 bool sample_address = false;
 806
 807                 if (max_stack) {
 808                         param.max_stack = max_stack;
 809                         if (callgraph_buf == NULL)
 810                                 callgraph_buf = "fp";
 811                 }
 812
 813                 /* parse callgraph parameters */
 814                 if (callgraph_buf != NULL) {
 815                         if (!strcmp(callgraph_buf, "no")) {
 816                                 param.enabled = false;
 817                                 param.record_mode = CALLCHAIN_NONE;
 818                         } else {
 819                                 param.enabled = true;
 820                                 if (parse_callchain_record(callgraph_buf, &param)) {
 821                                         pr_err("per-event callgraph setting for %s failed. "
 822                                                "Apply callgraph global setting for it\n",
 823                                                evsel->name);
 824                                         return;
 825                                 }
 826                                 if (param.record_mode == CALLCHAIN_DWARF)
 827                                         sample_address = true;
 828                         }
 829                 }
 830                 if (dump_size > 0) {
 831                         dump_size = round_up(dump_size, sizeof(u64));
 832                         param.dump_size = dump_size;
 833                 }
 834
 835                 /* If global callgraph set, clear it */
 836                 if (callchain_param.enabled)
 837                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 838
 839                 /* set perf-event callgraph */
 840                 if (param.enabled) {
 841                         if (sample_address) {
 842                                 perf_evsel__set_sample_bit(evsel, ADDR);
 843                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 844                                 evsel->attr.mmap_data = track;
 845                         }
 846                         perf_evsel__config_callchain(evsel, opts, &param);
 847                 }
 848         }
 849 }
 850
 851 /*
 852  * The enable_on_exec/disabled value strategy:
 853  *
 854  *  1) For any type of traced program:
 855  *    - all independent events and group leaders are disabled
 856  *    - all group members are enabled
 857  *
 858  *     Group members are ruled by group leaders. They need to
 859  *     be enabled, because the group scheduling relies on that.
 860  *
 861  *  2) For traced programs executed by perf:
 862  *     - all independent events and group leaders have
 863  *       enable_on_exec set
 864  *     - we don't specifically enable or disable any event during
 865  *       the record command
 866  *
 867  *     Independent events and group leaders are initially disabled
 868  *     and get enabled by exec. Group members are ruled by group
 869  *     leaders as stated in 1).
 870  *
 871  *  3) For traced programs attached by perf (pid/tid):
 872  *     - we specifically enable or disable all events during
 873  *       the record command
 874  *
 875  *     When attaching events to already running traced we
 876  *     enable/disable events specifically, as there's no
 877  *     initial traced exec call.
 878  */
 879 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 880                         struct callchain_param *callchain)
 881 {
 882         struct perf_evsel *leader = evsel->leader;
 883         struct perf_event_attr *attr = &evsel->attr;
 884         int track = evsel->tracking;
 885         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 886
 887         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 888         attr->inherit       = !opts->no_inherit;
 889         attr->write_backward = opts->overwrite ? 1 : 0;
 890
 891         perf_evsel__set_sample_bit(evsel, IP);
 892         perf_evsel__set_sample_bit(evsel, TID);
 893
 894         if (evsel->sample_read) {
 895                 perf_evsel__set_sample_bit(evsel, READ);
 896
 897                 /*
 898                  * We need ID even in case of single event, because
 899                  * PERF_SAMPLE_READ process ID specific data.
 900                  */
 901                 perf_evsel__set_sample_id(evsel, false);
 902
 903                 /*
 904                  * Apply group format only if we belong to group
 905                  * with more than one members.
 906                  */
 907                 if (leader->nr_members > 1) {
 908                         attr->read_format |= PERF_FORMAT_GROUP;
 909                         attr->inherit = 0;
 910                 }
 911         }
 912
 913         /*
 914          * We default some events to have a default interval. But keep
 915          * it a weak assumption overridable by the user.
 916          */
 917         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 918                                      opts->user_interval != ULLONG_MAX)) {
 919                 if (opts->freq) {
 920                         perf_evsel__set_sample_bit(evsel, PERIOD);
 921                         attr->freq              = 1;
 922                         attr->sample_freq       = opts->freq;
 923                 } else {
 924                         attr->sample_period = opts->default_interval;
 925                 }
 926         }
 927
 928         /*
 929          * Disable sampling for all group members other
 930          * than leader in case leader 'leads' the sampling.
 931          */
 932         if ((leader != evsel) && leader->sample_read) {
 933                 attr->freq           = 0;
 934                 attr->sample_freq    = 0;
 935                 attr->sample_period  = 0;
 936                 attr->write_backward = 0;
 937                 attr->sample_id_all  = 0;
 938         }
 939
 940         if (opts->no_samples)
 941                 attr->sample_freq = 0;
 942
 943         if (opts->inherit_stat) {
 944                 evsel->attr.read_format |=
 945                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 946                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 947                         PERF_FORMAT_ID;
 948                 attr->inherit_stat = 1;
 949         }
 950
 951         if (opts->sample_address) {
 952                 perf_evsel__set_sample_bit(evsel, ADDR);
 953                 attr->mmap_data = track;
 954         }
 955
 956         /*
 957          * We don't allow user space callchains for  function trace
 958          * event, due to issues with page faults while tracing page
 959          * fault handler and its overall trickiness nature.
 960          */
 961         if (perf_evsel__is_function_event(evsel))
 962                 evsel->attr.exclude_callchain_user = 1;
 963
 964         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 965                 perf_evsel__config_callchain(evsel, opts, callchain);
 966
 967         if (opts->sample_intr_regs) {
 968                 attr->sample_regs_intr = opts->sample_intr_regs;
 969                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 970         }
 971
 972         if (opts->sample_user_regs) {
 973                 attr->sample_regs_user |= opts->sample_user_regs;
 974                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 975         }
 976
 977         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 978                 perf_evsel__set_sample_bit(evsel, CPU);
 979
 980         /*
 981          * When the user explicitly disabled time don't force it here.
 982          */
 983         if (opts->sample_time &&
 984             (!perf_missing_features.sample_id_all &&
 985             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 986              opts->sample_time_set)))
 987                 perf_evsel__set_sample_bit(evsel, TIME);
 988
 989         if (opts->raw_samples && !evsel->no_aux_samples) {
 990                 perf_evsel__set_sample_bit(evsel, TIME);
 991                 perf_evsel__set_sample_bit(evsel, RAW);
 992                 perf_evsel__set_sample_bit(evsel, CPU);
 993         }
 994
 995         if (opts->sample_address)
 996                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 997
 998         if (opts->sample_phys_addr)
 999                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1000
1001         if (opts->no_buffering) {
1002                 attr->watermark = 0;
1003                 attr->wakeup_events = 1;
1004         }
1005         if (opts->branch_stack && !evsel->no_aux_samples) {
1006                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1007                 attr->branch_sample_type = opts->branch_stack;
1008         }
1009
1010         if (opts->sample_weight)
1011                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1012
1013         attr->task  = track;
1014         attr->mmap  = track;
1015         attr->mmap2 = track && !perf_missing_features.mmap2;
1016         attr->comm  = track;
1017
1018         if (opts->record_namespaces)
1019                 attr->namespaces  = track;
1020
1021         if (opts->record_switch_events)
1022                 attr->context_switch = track;
1023
1024         if (opts->sample_transaction)
1025                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1026
1027         if (opts->running_time) {
1028                 evsel->attr.read_format |=
1029                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1030                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1031         }
1032
1033         /*
1034          * XXX see the function comment above
1035          *
1036          * Disabling only independent events or group leaders,
1037          * keeping group members enabled.
1038          */
1039         if (perf_evsel__is_group_leader(evsel))
1040                 attr->disabled = 1;
1041
1042         /*
1043          * Setting enable_on_exec for independent events and
1044          * group leaders for traced executed by perf.
1045          */
1046         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1047                 !opts->initial_delay)
1048                 attr->enable_on_exec = 1;
1049
1050         if (evsel->immediate) {
1051                 attr->disabled = 0;
1052                 attr->enable_on_exec = 0;
1053         }
1054
1055         clockid = opts->clockid;
1056         if (opts->use_clockid) {
1057                 attr->use_clockid = 1;
1058                 attr->clockid = opts->clockid;
1059         }
1060
1061         if (evsel->precise_max)
1062                 perf_event_attr__set_max_precise_ip(attr);
1063
1064         if (opts->all_user) {
1065                 attr->exclude_kernel = 1;
1066                 attr->exclude_user   = 0;
1067         }
1068
1069         if (opts->all_kernel) {
1070                 attr->exclude_kernel = 0;
1071                 attr->exclude_user   = 1;
1072         }
1073
1074         /*
1075          * Apply event specific term settings,
1076          * it overloads any global configuration.
1077          */
1078         apply_config_terms(evsel, opts, track);
1079
1080         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1081
1082         /* The --period option takes the precedence. */
1083         if (opts->period_set) {
1084                 if (opts->period)
1085                         perf_evsel__set_sample_bit(evsel, PERIOD);
1086                 else
1087                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1088         }
1089 }
1090
1091 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1092 {
1093         if (evsel->system_wide)
1094                 nthreads = 1;
1095
1096         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1097
1098         if (evsel->fd) {
1099                 int cpu, thread;
1100                 for (cpu = 0; cpu < ncpus; cpu++) {
1101                         for (thread = 0; thread < nthreads; thread++) {
1102                                 FD(evsel, cpu, thread) = -1;
1103                         }
1104                 }
1105         }
1106
1107         return evsel->fd != NULL ? 0 : -ENOMEM;
1108 }
1109
1110 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1111                           int ioc,  void *arg)
1112 {
1113         int cpu, thread;
1114
1115         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1116                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1117                         int fd = FD(evsel, cpu, thread),
1118                             err = ioctl(fd, ioc, arg);
1119
1120                         if (err)
1121                                 return err;
1122                 }
1123         }
1124
1125         return 0;
1126 }
1127
1128 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1129 {
1130         return perf_evsel__run_ioctl(evsel,
1131                                      PERF_EVENT_IOC_SET_FILTER,
1132                                      (void *)filter);
1133 }
1134
1135 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1136 {
1137         char *new_filter = strdup(filter);
1138
1139         if (new_filter != NULL) {
1140                 free(evsel->filter);
1141                 evsel->filter = new_filter;
1142                 return 0;
1143         }
1144
1145         return -1;
1146 }
1147
1148 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1149                                      const char *fmt, const char *filter)
1150 {
1151         char *new_filter;
1152
1153         if (evsel->filter == NULL)
1154                 return perf_evsel__set_filter(evsel, filter);
1155
1156         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1157                 free(evsel->filter);
1158                 evsel->filter = new_filter;
1159                 return 0;
1160         }
1161
1162         return -1;
1163 }
1164
1165 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1166 {
1167         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1168 }
1169
1170 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1171 {
1172         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1173 }
1174
1175 int perf_evsel__enable(struct perf_evsel *evsel)
1176 {
1177         return perf_evsel__run_ioctl(evsel,
1178                                      PERF_EVENT_IOC_ENABLE,
1179                                      0);
1180 }
1181
1182 int perf_evsel__disable(struct perf_evsel *evsel)
1183 {
1184         return perf_evsel__run_ioctl(evsel,
1185                                      PERF_EVENT_IOC_DISABLE,
1186                                      0);
1187 }
1188
1189 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1190 {
1191         if (ncpus == 0 || nthreads == 0)
1192                 return 0;
1193
1194         if (evsel->system_wide)
1195                 nthreads = 1;
1196
1197         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1198         if (evsel->sample_id == NULL)
1199                 return -ENOMEM;
1200
1201         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1202         if (evsel->id == NULL) {
1203                 xyarray__delete(evsel->sample_id);
1204                 evsel->sample_id = NULL;
1205                 return -ENOMEM;
1206         }
1207
1208         return 0;
1209 }
1210
1211 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1212 {
1213         xyarray__delete(evsel->fd);
1214         evsel->fd = NULL;
1215 }
1216
1217 static void perf_evsel__free_id(struct perf_evsel *evsel)
1218 {
1219         xyarray__delete(evsel->sample_id);
1220         evsel->sample_id = NULL;
1221         zfree(&evsel->id);
1222 }
1223
1224 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1225 {
1226         struct perf_evsel_config_term *term, *h;
1227
1228         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1229                 list_del(&term->list);
1230                 free(term);
1231         }
1232 }
1233
1234 void perf_evsel__close_fd(struct perf_evsel *evsel)
1235 {
1236         int cpu, thread;
1237
1238         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1239                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1240                         close(FD(evsel, cpu, thread));
1241                         FD(evsel, cpu, thread) = -1;
1242                 }
1243 }
1244
1245 void perf_evsel__exit(struct perf_evsel *evsel)
1246 {
1247         assert(list_empty(&evsel->node));
1248         assert(evsel->evlist == NULL);
1249         perf_evsel__free_fd(evsel);
1250         perf_evsel__free_id(evsel);
1251         perf_evsel__free_config_terms(evsel);
1252         cgroup__put(evsel->cgrp);
1253         cpu_map__put(evsel->cpus);
1254         cpu_map__put(evsel->own_cpus);
1255         thread_map__put(evsel->threads);
1256         zfree(&evsel->group_name);
1257         zfree(&evsel->name);
1258         perf_evsel__object.fini(evsel);
1259 }
1260
1261 void perf_evsel__delete(struct perf_evsel *evsel)
1262 {
1263         perf_evsel__exit(evsel);
1264         free(evsel);
1265 }
1266
1267 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1268                                 struct perf_counts_values *count)
1269 {
1270         struct perf_counts_values tmp;
1271
1272         if (!evsel->prev_raw_counts)
1273                 return;
1274
1275         if (cpu == -1) {
1276                 tmp = evsel->prev_raw_counts->aggr;
1277                 evsel->prev_raw_counts->aggr = *count;
1278         } else {
1279                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1280                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1281         }
1282
1283         count->val = count->val - tmp.val;
1284         count->ena = count->ena - tmp.ena;
1285         count->run = count->run - tmp.run;
1286 }
1287
1288 void perf_counts_values__scale(struct perf_counts_values *count,
1289                                bool scale, s8 *pscaled)
1290 {
1291         s8 scaled = 0;
1292
1293         if (scale) {
1294                 if (count->run == 0) {
1295                         scaled = -1;
1296                         count->val = 0;
1297                 } else if (count->run < count->ena) {
1298                         scaled = 1;
1299                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1300                 }
1301         } else
1302                 count->ena = count->run = 0;
1303
1304         if (pscaled)
1305                 *pscaled = scaled;
1306 }
1307
1308 static int perf_evsel__read_size(struct perf_evsel *evsel)
1309 {
1310         u64 read_format = evsel->attr.read_format;
1311         int entry = sizeof(u64); /* value */
1312         int size = 0;
1313         int nr = 1;
1314
1315         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1316                 size += sizeof(u64);
1317
1318         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1319                 size += sizeof(u64);
1320
1321         if (read_format & PERF_FORMAT_ID)
1322                 entry += sizeof(u64);
1323
1324         if (read_format & PERF_FORMAT_GROUP) {
1325                 nr = evsel->nr_members;
1326                 size += sizeof(u64);
1327         }
1328
1329         size += entry * nr;
1330         return size;
1331 }
1332
1333 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1334                      struct perf_counts_values *count)
1335 {
1336         size_t size = perf_evsel__read_size(evsel);
1337
1338         memset(count, 0, sizeof(*count));
1339
1340         if (FD(evsel, cpu, thread) < 0)
1341                 return -EINVAL;
1342
1343         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1344                 return -errno;
1345
1346         return 0;
1347 }
1348
1349 static int
1350 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1351 {
1352         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1353
1354         return perf_evsel__read(evsel, cpu, thread, count);
1355 }
1356
1357 static void
1358 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1359                       u64 val, u64 ena, u64 run)
1360 {
1361         struct perf_counts_values *count;
1362
1363         count = perf_counts(counter->counts, cpu, thread);
1364
1365         count->val    = val;
1366         count->ena    = ena;
1367         count->run    = run;
1368         count->loaded = true;
1369 }
1370
1371 static int
1372 perf_evsel__process_group_data(struct perf_evsel *leader,
1373                                int cpu, int thread, u64 *data)
1374 {
1375         u64 read_format = leader->attr.read_format;
1376         struct sample_read_value *v;
1377         u64 nr, ena = 0, run = 0, i;
1378
1379         nr = *data++;
1380
1381         if (nr != (u64) leader->nr_members)
1382                 return -EINVAL;
1383
1384         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1385                 ena = *data++;
1386
1387         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1388                 run = *data++;
1389
1390         v = (struct sample_read_value *) data;
1391
1392         perf_evsel__set_count(leader, cpu, thread,
1393                               v[0].value, ena, run);
1394
1395         for (i = 1; i < nr; i++) {
1396                 struct perf_evsel *counter;
1397
1398                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1399                 if (!counter)
1400                         return -EINVAL;
1401
1402                 perf_evsel__set_count(counter, cpu, thread,
1403                                       v[i].value, ena, run);
1404         }
1405
1406         return 0;
1407 }
1408
1409 static int
1410 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1411 {
1412         struct perf_stat_evsel *ps = leader->stats;
1413         u64 read_format = leader->attr.read_format;
1414         int size = perf_evsel__read_size(leader);
1415         u64 *data = ps->group_data;
1416
1417         if (!(read_format & PERF_FORMAT_ID))
1418                 return -EINVAL;
1419
1420         if (!perf_evsel__is_group_leader(leader))
1421                 return -EINVAL;
1422
1423         if (!data) {
1424                 data = zalloc(size);
1425                 if (!data)
1426                         return -ENOMEM;
1427
1428                 ps->group_data = data;
1429         }
1430
1431         if (FD(leader, cpu, thread) < 0)
1432                 return -EINVAL;
1433
1434         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1435                 return -errno;
1436
1437         return perf_evsel__process_group_data(leader, cpu, thread, data);
1438 }
1439
1440 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1441 {
1442         u64 read_format = evsel->attr.read_format;
1443
1444         if (read_format & PERF_FORMAT_GROUP)
1445                 return perf_evsel__read_group(evsel, cpu, thread);
1446         else
1447                 return perf_evsel__read_one(evsel, cpu, thread);
1448 }
1449
1450 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1451                               int cpu, int thread, bool scale)
1452 {
1453         struct perf_counts_values count;
1454         size_t nv = scale ? 3 : 1;
1455
1456         if (FD(evsel, cpu, thread) < 0)
1457                 return -EINVAL;
1458
1459         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1460                 return -ENOMEM;
1461
1462         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1463                 return -errno;
1464
1465         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1466         perf_counts_values__scale(&count, scale, NULL);
1467         *perf_counts(evsel->counts, cpu, thread) = count;
1468         return 0;
1469 }
1470
1471 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1472 {
1473         struct perf_evsel *leader = evsel->leader;
1474         int fd;
1475
1476         if (perf_evsel__is_group_leader(evsel))
1477                 return -1;
1478
1479         /*
1480          * Leader must be already processed/open,
1481          * if not it's a bug.
1482          */
1483         BUG_ON(!leader->fd);
1484
1485         fd = FD(leader, cpu, thread);
1486         BUG_ON(fd == -1);
1487
1488         return fd;
1489 }
1490
1491 struct bit_names {
1492         int bit;
1493         const char *name;
1494 };
1495
1496 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1497 {
1498         bool first_bit = true;
1499         int i = 0;
1500
1501         do {
1502                 if (value & bits[i].bit) {
1503                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1504                         first_bit = false;
1505                 }
1506         } while (bits[++i].name != NULL);
1507 }
1508
1509 static void __p_sample_type(char *buf, size_t size, u64 value)
1510 {
1511 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1512         struct bit_names bits[] = {
1513                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1514                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1515                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1516                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1517                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1518                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1519                 { .name = NULL, }
1520         };
1521 #undef bit_name
1522         __p_bits(buf, size, value, bits);
1523 }
1524
1525 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1526 {
1527 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1528         struct bit_names bits[] = {
1529                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1530                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1531                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1532                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1533                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1534                 { .name = NULL, }
1535         };
1536 #undef bit_name
1537         __p_bits(buf, size, value, bits);
1538 }
1539
1540 static void __p_read_format(char *buf, size_t size, u64 value)
1541 {
1542 #define bit_name(n) { PERF_FORMAT_##n, #n }
1543         struct bit_names bits[] = {
1544                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1545                 bit_name(ID), bit_name(GROUP),
1546                 { .name = NULL, }
1547         };
1548 #undef bit_name
1549         __p_bits(buf, size, value, bits);
1550 }
1551
1552 #define BUF_SIZE                1024
1553
1554 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1555 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1556 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1557 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1558 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1559 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1560
1561 #define PRINT_ATTRn(_n, _f, _p)                         \
1562 do {                                                    \
1563         if (attr->_f) {                                 \
1564                 _p(attr->_f);                           \
1565                 ret += attr__fprintf(fp, _n, buf, priv);\
1566         }                                               \
1567 } while (0)
1568
1569 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1570
1571 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1572                              attr__fprintf_f attr__fprintf, void *priv)
1573 {
1574         char buf[BUF_SIZE];
1575         int ret = 0;
1576
1577         PRINT_ATTRf(type, p_unsigned);
1578         PRINT_ATTRf(size, p_unsigned);
1579         PRINT_ATTRf(config, p_hex);
1580         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1581         PRINT_ATTRf(sample_type, p_sample_type);
1582         PRINT_ATTRf(read_format, p_read_format);
1583
1584         PRINT_ATTRf(disabled, p_unsigned);
1585         PRINT_ATTRf(inherit, p_unsigned);
1586         PRINT_ATTRf(pinned, p_unsigned);
1587         PRINT_ATTRf(exclusive, p_unsigned);
1588         PRINT_ATTRf(exclude_user, p_unsigned);
1589         PRINT_ATTRf(exclude_kernel, p_unsigned);
1590         PRINT_ATTRf(exclude_hv, p_unsigned);
1591         PRINT_ATTRf(exclude_idle, p_unsigned);
1592         PRINT_ATTRf(mmap, p_unsigned);
1593         PRINT_ATTRf(comm, p_unsigned);
1594         PRINT_ATTRf(freq, p_unsigned);
1595         PRINT_ATTRf(inherit_stat, p_unsigned);
1596         PRINT_ATTRf(enable_on_exec, p_unsigned);
1597         PRINT_ATTRf(task, p_unsigned);
1598         PRINT_ATTRf(watermark, p_unsigned);
1599         PRINT_ATTRf(precise_ip, p_unsigned);
1600         PRINT_ATTRf(mmap_data, p_unsigned);
1601         PRINT_ATTRf(sample_id_all, p_unsigned);
1602         PRINT_ATTRf(exclude_host, p_unsigned);
1603         PRINT_ATTRf(exclude_guest, p_unsigned);
1604         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1605         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1606         PRINT_ATTRf(mmap2, p_unsigned);
1607         PRINT_ATTRf(comm_exec, p_unsigned);
1608         PRINT_ATTRf(use_clockid, p_unsigned);
1609         PRINT_ATTRf(context_switch, p_unsigned);
1610         PRINT_ATTRf(write_backward, p_unsigned);
1611         PRINT_ATTRf(namespaces, p_unsigned);
1612
1613         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1614         PRINT_ATTRf(bp_type, p_unsigned);
1615         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1616         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1617         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1618         PRINT_ATTRf(sample_regs_user, p_hex);
1619         PRINT_ATTRf(sample_stack_user, p_unsigned);
1620         PRINT_ATTRf(clockid, p_signed);
1621         PRINT_ATTRf(sample_regs_intr, p_hex);
1622         PRINT_ATTRf(aux_watermark, p_unsigned);
1623         PRINT_ATTRf(sample_max_stack, p_unsigned);
1624
1625         return ret;
1626 }
1627
1628 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1629                                 void *priv __maybe_unused)
1630 {
1631         return fprintf(fp, "  %-32s %s\n", name, val);
1632 }
1633
1634 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1635                                   int nr_cpus, int nr_threads,
1636                                   int thread_idx)
1637 {
1638         for (int cpu = 0; cpu < nr_cpus; cpu++)
1639                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1640                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1641 }
1642
1643 static int update_fds(struct perf_evsel *evsel,
1644                       int nr_cpus, int cpu_idx,
1645                       int nr_threads, int thread_idx)
1646 {
1647         struct perf_evsel *pos;
1648
1649         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1650                 return -EINVAL;
1651
1652         evlist__for_each_entry(evsel->evlist, pos) {
1653                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1654
1655                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1656
1657                 /*
1658                  * Since fds for next evsel has not been created,
1659                  * there is no need to iterate whole event list.
1660                  */
1661                 if (pos == evsel)
1662                         break;
1663         }
1664         return 0;
1665 }
1666
1667 static bool ignore_missing_thread(struct perf_evsel *evsel,
1668                                   int nr_cpus, int cpu,
1669                                   struct thread_map *threads,
1670                                   int thread, int err)
1671 {
1672         pid_t ignore_pid = thread_map__pid(threads, thread);
1673
1674         if (!evsel->ignore_missing_thread)
1675                 return false;
1676
1677         /* The system wide setup does not work with threads. */
1678         if (evsel->system_wide)
1679                 return false;
1680
1681         /* The -ESRCH is perf event syscall errno for pid's not found. */
1682         if (err != -ESRCH)
1683                 return false;
1684
1685         /* If there's only one thread, let it fail. */
1686         if (threads->nr == 1)
1687                 return false;
1688
1689         /*
1690          * We should remove fd for missing_thread first
1691          * because thread_map__remove() will decrease threads->nr.
1692          */
1693         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1694                 return false;
1695
1696         if (thread_map__remove(threads, thread))
1697                 return false;
1698
1699         pr_warning("WARNING: Ignored open failure for pid %d\n",
1700                    ignore_pid);
1701         return true;
1702 }
1703
1704 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1705                      struct thread_map *threads)
1706 {
1707         int cpu, thread, nthreads;
1708         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1709         int pid = -1, err;
1710         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1711
1712         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1713                 return -EINVAL;
1714
1715         if (cpus == NULL) {
1716                 static struct cpu_map *empty_cpu_map;
1717
1718                 if (empty_cpu_map == NULL) {
1719                         empty_cpu_map = cpu_map__dummy_new();
1720                         if (empty_cpu_map == NULL)
1721                                 return -ENOMEM;
1722                 }
1723
1724                 cpus = empty_cpu_map;
1725         }
1726
1727         if (threads == NULL) {
1728                 static struct thread_map *empty_thread_map;
1729
1730                 if (empty_thread_map == NULL) {
1731                         empty_thread_map = thread_map__new_by_tid(-1);
1732                         if (empty_thread_map == NULL)
1733                                 return -ENOMEM;
1734                 }
1735
1736                 threads = empty_thread_map;
1737         }
1738
1739         if (evsel->system_wide)
1740                 nthreads = 1;
1741         else
1742                 nthreads = threads->nr;
1743
1744         if (evsel->fd == NULL &&
1745             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1746                 return -ENOMEM;
1747
1748         if (evsel->cgrp) {
1749                 flags |= PERF_FLAG_PID_CGROUP;
1750                 pid = evsel->cgrp->fd;
1751         }
1752
1753 fallback_missing_features:
1754         if (perf_missing_features.clockid_wrong)
1755                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1756         if (perf_missing_features.clockid) {
1757                 evsel->attr.use_clockid = 0;
1758                 evsel->attr.clockid = 0;
1759         }
1760         if (perf_missing_features.cloexec)
1761                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1762         if (perf_missing_features.mmap2)
1763                 evsel->attr.mmap2 = 0;
1764         if (perf_missing_features.exclude_guest)
1765                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1766         if (perf_missing_features.lbr_flags)
1767                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1768                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1769         if (perf_missing_features.group_read && evsel->attr.inherit)
1770                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1771 retry_sample_id:
1772         if (perf_missing_features.sample_id_all)
1773                 evsel->attr.sample_id_all = 0;
1774
1775         if (verbose >= 2) {
1776                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1777                 fprintf(stderr, "perf_event_attr:\n");
1778                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1779                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780         }
1781
1782         for (cpu = 0; cpu < cpus->nr; cpu++) {
1783
1784                 for (thread = 0; thread < nthreads; thread++) {
1785                         int fd, group_fd;
1786
1787                         if (!evsel->cgrp && !evsel->system_wide)
1788                                 pid = thread_map__pid(threads, thread);
1789
1790                         group_fd = get_group_fd(evsel, cpu, thread);
1791 retry_open:
1792                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1793                                   pid, cpus->map[cpu], group_fd, flags);
1794
1795                         test_attr__ready();
1796
1797                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1798                                                  group_fd, flags);
1799
1800                         FD(evsel, cpu, thread) = fd;
1801
1802                         if (fd < 0) {
1803                                 err = -errno;
1804
1805                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1806                                         /*
1807                                          * We just removed 1 thread, so take a step
1808                                          * back on thread index and lower the upper
1809                                          * nthreads limit.
1810                                          */
1811                                         nthreads--;
1812                                         thread--;
1813
1814                                         /* ... and pretend like nothing have happened. */
1815                                         err = 0;
1816                                         continue;
1817                                 }
1818
1819                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1820                                           err);
1821                                 goto try_fallback;
1822                         }
1823
1824                         pr_debug2(" = %d\n", fd);
1825
1826                         if (evsel->bpf_fd >= 0) {
1827                                 int evt_fd = fd;
1828                                 int bpf_fd = evsel->bpf_fd;
1829
1830                                 err = ioctl(evt_fd,
1831                                             PERF_EVENT_IOC_SET_BPF,
1832                                             bpf_fd);
1833                                 if (err && errno != EEXIST) {
1834                                         pr_err("failed to attach bpf fd %d: %s\n",
1835                                                bpf_fd, strerror(errno));
1836                                         err = -EINVAL;
1837                                         goto out_close;
1838                                 }
1839                         }
1840
1841                         set_rlimit = NO_CHANGE;
1842
1843                         /*
1844                          * If we succeeded but had to kill clockid, fail and
1845                          * have perf_evsel__open_strerror() print us a nice
1846                          * error.
1847                          */
1848                         if (perf_missing_features.clockid ||
1849                             perf_missing_features.clockid_wrong) {
1850                                 err = -EINVAL;
1851                                 goto out_close;
1852                         }
1853                 }
1854         }
1855
1856         return 0;
1857
1858 try_fallback:
1859         /*
1860          * perf stat needs between 5 and 22 fds per CPU. When we run out
1861          * of them try to increase the limits.
1862          */
1863         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1864                 struct rlimit l;
1865                 int old_errno = errno;
1866
1867                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1868                         if (set_rlimit == NO_CHANGE)
1869                                 l.rlim_cur = l.rlim_max;
1870                         else {
1871                                 l.rlim_cur = l.rlim_max + 1000;
1872                                 l.rlim_max = l.rlim_cur;
1873                         }
1874                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1875                                 set_rlimit++;
1876                                 errno = old_errno;
1877                                 goto retry_open;
1878                         }
1879                 }
1880                 errno = old_errno;
1881         }
1882
1883         if (err != -EINVAL || cpu > 0 || thread > 0)
1884                 goto out_close;
1885
1886         /*
1887          * Must probe features in the order they were added to the
1888          * perf_event_attr interface.
1889          */
1890         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1891                 perf_missing_features.write_backward = true;
1892                 pr_debug2("switching off write_backward\n");
1893                 goto out_close;
1894         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1895                 perf_missing_features.clockid_wrong = true;
1896                 pr_debug2("switching off clockid\n");
1897                 goto fallback_missing_features;
1898         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1899                 perf_missing_features.clockid = true;
1900                 pr_debug2("switching off use_clockid\n");
1901                 goto fallback_missing_features;
1902         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1903                 perf_missing_features.cloexec = true;
1904                 pr_debug2("switching off cloexec flag\n");
1905                 goto fallback_missing_features;
1906         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1907                 perf_missing_features.mmap2 = true;
1908                 pr_debug2("switching off mmap2\n");
1909                 goto fallback_missing_features;
1910         } else if (!perf_missing_features.exclude_guest &&
1911                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1912                 perf_missing_features.exclude_guest = true;
1913                 pr_debug2("switching off exclude_guest, exclude_host\n");
1914                 goto fallback_missing_features;
1915         } else if (!perf_missing_features.sample_id_all) {
1916                 perf_missing_features.sample_id_all = true;
1917                 pr_debug2("switching off sample_id_all\n");
1918                 goto retry_sample_id;
1919         } else if (!perf_missing_features.lbr_flags &&
1920                         (evsel->attr.branch_sample_type &
1921                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1922                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1923                 perf_missing_features.lbr_flags = true;
1924                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1925                 goto fallback_missing_features;
1926         } else if (!perf_missing_features.group_read &&
1927                     evsel->attr.inherit &&
1928                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1929                    perf_evsel__is_group_leader(evsel)) {
1930                 perf_missing_features.group_read = true;
1931                 pr_debug2("switching off group read\n");
1932                 goto fallback_missing_features;
1933         }
1934 out_close:
1935         if (err)
1936                 threads->err_thread = thread;
1937
1938         do {
1939                 while (--thread >= 0) {
1940                         close(FD(evsel, cpu, thread));
1941                         FD(evsel, cpu, thread) = -1;
1942                 }
1943                 thread = nthreads;
1944         } while (--cpu >= 0);
1945         return err;
1946 }
1947
1948 void perf_evsel__close(struct perf_evsel *evsel)
1949 {
1950         if (evsel->fd == NULL)
1951                 return;
1952
1953         perf_evsel__close_fd(evsel);
1954         perf_evsel__free_fd(evsel);
1955 }
1956
1957 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1958                              struct cpu_map *cpus)
1959 {
1960         return perf_evsel__open(evsel, cpus, NULL);
1961 }
1962
1963 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1964                                 struct thread_map *threads)
1965 {
1966         return perf_evsel__open(evsel, NULL, threads);
1967 }
1968
1969 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1970                                        const union perf_event *event,
1971                                        struct perf_sample *sample)
1972 {
1973         u64 type = evsel->attr.sample_type;
1974         const u64 *array = event->sample.array;
1975         bool swapped = evsel->needs_swap;
1976         union u64_swap u;
1977
1978         array += ((event->header.size -
1979                    sizeof(event->header)) / sizeof(u64)) - 1;
1980
1981         if (type & PERF_SAMPLE_IDENTIFIER) {
1982                 sample->id = *array;
1983                 array--;
1984         }
1985
1986         if (type & PERF_SAMPLE_CPU) {
1987                 u.val64 = *array;
1988                 if (swapped) {
1989                         /* undo swap of u64, then swap on individual u32s */
1990                         u.val64 = bswap_64(u.val64);
1991                         u.val32[0] = bswap_32(u.val32[0]);
1992                 }
1993
1994                 sample->cpu = u.val32[0];
1995                 array--;
1996         }
1997
1998         if (type & PERF_SAMPLE_STREAM_ID) {
1999                 sample->stream_id = *array;
2000                 array--;
2001         }
2002
2003         if (type & PERF_SAMPLE_ID) {
2004                 sample->id = *array;
2005                 array--;
2006         }
2007
2008         if (type & PERF_SAMPLE_TIME) {
2009                 sample->time = *array;
2010                 array--;
2011         }
2012
2013         if (type & PERF_SAMPLE_TID) {
2014                 u.val64 = *array;
2015                 if (swapped) {
2016                         /* undo swap of u64, then swap on individual u32s */
2017                         u.val64 = bswap_64(u.val64);
2018                         u.val32[0] = bswap_32(u.val32[0]);
2019                         u.val32[1] = bswap_32(u.val32[1]);
2020                 }
2021
2022                 sample->pid = u.val32[0];
2023                 sample->tid = u.val32[1];
2024                 array--;
2025         }
2026
2027         return 0;
2028 }
2029
2030 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2031                             u64 size)
2032 {
2033         return size > max_size || offset + size > endp;
2034 }
2035
2036 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2037         do {                                                            \
2038                 if (overflow(endp, (max_size), (offset), (size)))       \
2039                         return -EFAULT;                                 \
2040         } while (0)
2041
2042 #define OVERFLOW_CHECK_u64(offset) \
2043         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2044
2045 static int
2046 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2047 {
2048         /*
2049          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2050          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2051          * check the format does not go past the end of the event.
2052          */
2053         if (sample_size + sizeof(event->header) > event->header.size)
2054                 return -EFAULT;
2055
2056         return 0;
2057 }
2058
2059 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2060                              struct perf_sample *data)
2061 {
2062         u64 type = evsel->attr.sample_type;
2063         bool swapped = evsel->needs_swap;
2064         const u64 *array;
2065         u16 max_size = event->header.size;
2066         const void *endp = (void *)event + max_size;
2067         u64 sz;
2068
2069         /*
2070          * used for cross-endian analysis. See git commit 65014ab3
2071          * for why this goofiness is needed.
2072          */
2073         union u64_swap u;
2074
2075         memset(data, 0, sizeof(*data));
2076         data->cpu = data->pid = data->tid = -1;
2077         data->stream_id = data->id = data->time = -1ULL;
2078         data->period = evsel->attr.sample_period;
2079         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2080         data->misc    = event->header.misc;
2081         data->id = -1ULL;
2082         data->data_src = PERF_MEM_DATA_SRC_NONE;
2083
2084         if (event->header.type != PERF_RECORD_SAMPLE) {
2085                 if (!evsel->attr.sample_id_all)
2086                         return 0;
2087                 return perf_evsel__parse_id_sample(evsel, event, data);
2088         }
2089
2090         array = event->sample.array;
2091
2092         if (perf_event__check_size(event, evsel->sample_size))
2093                 return -EFAULT;
2094
2095         if (type & PERF_SAMPLE_IDENTIFIER) {
2096                 data->id = *array;
2097                 array++;
2098         }
2099
2100         if (type & PERF_SAMPLE_IP) {
2101                 data->ip = *array;
2102                 array++;
2103         }
2104
2105         if (type & PERF_SAMPLE_TID) {
2106                 u.val64 = *array;
2107                 if (swapped) {
2108                         /* undo swap of u64, then swap on individual u32s */
2109                         u.val64 = bswap_64(u.val64);
2110                         u.val32[0] = bswap_32(u.val32[0]);
2111                         u.val32[1] = bswap_32(u.val32[1]);
2112                 }
2113
2114                 data->pid = u.val32[0];
2115                 data->tid = u.val32[1];
2116                 array++;
2117         }
2118
2119         if (type & PERF_SAMPLE_TIME) {
2120                 data->time = *array;
2121                 array++;
2122         }
2123
2124         if (type & PERF_SAMPLE_ADDR) {
2125                 data->addr = *array;
2126                 array++;
2127         }
2128
2129         if (type & PERF_SAMPLE_ID) {
2130                 data->id = *array;
2131                 array++;
2132         }
2133
2134         if (type & PERF_SAMPLE_STREAM_ID) {
2135                 data->stream_id = *array;
2136                 array++;
2137         }
2138
2139         if (type & PERF_SAMPLE_CPU) {
2140
2141                 u.val64 = *array;
2142                 if (swapped) {
2143                         /* undo swap of u64, then swap on individual u32s */
2144                         u.val64 = bswap_64(u.val64);
2145                         u.val32[0] = bswap_32(u.val32[0]);
2146                 }
2147
2148                 data->cpu = u.val32[0];
2149                 array++;
2150         }
2151
2152         if (type & PERF_SAMPLE_PERIOD) {
2153                 data->period = *array;
2154                 array++;
2155         }
2156
2157         if (type & PERF_SAMPLE_READ) {
2158                 u64 read_format = evsel->attr.read_format;
2159
2160                 OVERFLOW_CHECK_u64(array);
2161                 if (read_format & PERF_FORMAT_GROUP)
2162                         data->read.group.nr = *array;
2163                 else
2164                         data->read.one.value = *array;
2165
2166                 array++;
2167
2168                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2169                         OVERFLOW_CHECK_u64(array);
2170                         data->read.time_enabled = *array;
2171                         array++;
2172                 }
2173
2174                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2175                         OVERFLOW_CHECK_u64(array);
2176                         data->read.time_running = *array;
2177                         array++;
2178                 }
2179
2180                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2181                 if (read_format & PERF_FORMAT_GROUP) {
2182                         const u64 max_group_nr = UINT64_MAX /
2183                                         sizeof(struct sample_read_value);
2184
2185                         if (data->read.group.nr > max_group_nr)
2186                                 return -EFAULT;
2187                         sz = data->read.group.nr *
2188                              sizeof(struct sample_read_value);
2189                         OVERFLOW_CHECK(array, sz, max_size);
2190                         data->read.group.values =
2191                                         (struct sample_read_value *)array;
2192                         array = (void *)array + sz;
2193                 } else {
2194                         OVERFLOW_CHECK_u64(array);
2195                         data->read.one.id = *array;
2196                         array++;
2197                 }
2198         }
2199
2200         if (evsel__has_callchain(evsel)) {
2201                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2202
2203                 OVERFLOW_CHECK_u64(array);
2204                 data->callchain = (struct ip_callchain *)array++;
2205                 if (data->callchain->nr > max_callchain_nr)
2206                         return -EFAULT;
2207                 sz = data->callchain->nr * sizeof(u64);
2208                 OVERFLOW_CHECK(array, sz, max_size);
2209                 array = (void *)array + sz;
2210         }
2211
2212         if (type & PERF_SAMPLE_RAW) {
2213                 OVERFLOW_CHECK_u64(array);
2214                 u.val64 = *array;
2215
2216                 /*
2217                  * Undo swap of u64, then swap on individual u32s,
2218                  * get the size of the raw area and undo all of the
2219                  * swap. The pevent interface handles endianity by
2220                  * itself.
2221                  */
2222                 if (swapped) {
2223                         u.val64 = bswap_64(u.val64);
2224                         u.val32[0] = bswap_32(u.val32[0]);
2225                         u.val32[1] = bswap_32(u.val32[1]);
2226                 }
2227                 data->raw_size = u.val32[0];
2228
2229                 /*
2230                  * The raw data is aligned on 64bits including the
2231                  * u32 size, so it's safe to use mem_bswap_64.
2232                  */
2233                 if (swapped)
2234                         mem_bswap_64((void *) array, data->raw_size);
2235
2236                 array = (void *)array + sizeof(u32);
2237
2238                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2239                 data->raw_data = (void *)array;
2240                 array = (void *)array + data->raw_size;
2241         }
2242
2243         if (type & PERF_SAMPLE_BRANCH_STACK) {
2244                 const u64 max_branch_nr = UINT64_MAX /
2245                                           sizeof(struct branch_entry);
2246
2247                 OVERFLOW_CHECK_u64(array);
2248                 data->branch_stack = (struct branch_stack *)array++;
2249
2250                 if (data->branch_stack->nr > max_branch_nr)
2251                         return -EFAULT;
2252                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2253                 OVERFLOW_CHECK(array, sz, max_size);
2254                 array = (void *)array + sz;
2255         }
2256
2257         if (type & PERF_SAMPLE_REGS_USER) {
2258                 OVERFLOW_CHECK_u64(array);
2259                 data->user_regs.abi = *array;
2260                 array++;
2261
2262                 if (data->user_regs.abi) {
2263                         u64 mask = evsel->attr.sample_regs_user;
2264
2265                         sz = hweight_long(mask) * sizeof(u64);
2266                         OVERFLOW_CHECK(array, sz, max_size);
2267                         data->user_regs.mask = mask;
2268                         data->user_regs.regs = (u64 *)array;
2269                         array = (void *)array + sz;
2270                 }
2271         }
2272
2273         if (type & PERF_SAMPLE_STACK_USER) {
2274                 OVERFLOW_CHECK_u64(array);
2275                 sz = *array++;
2276
2277                 data->user_stack.offset = ((char *)(array - 1)
2278                                           - (char *) event);
2279
2280                 if (!sz) {
2281                         data->user_stack.size = 0;
2282                 } else {
2283                         OVERFLOW_CHECK(array, sz, max_size);
2284                         data->user_stack.data = (char *)array;
2285                         array = (void *)array + sz;
2286                         OVERFLOW_CHECK_u64(array);
2287                         data->user_stack.size = *array++;
2288                         if (WARN_ONCE(data->user_stack.size > sz,
2289                                       "user stack dump failure\n"))
2290                                 return -EFAULT;
2291                 }
2292         }
2293
2294         if (type & PERF_SAMPLE_WEIGHT) {
2295                 OVERFLOW_CHECK_u64(array);
2296                 data->weight = *array;
2297                 array++;
2298         }
2299
2300         if (type & PERF_SAMPLE_DATA_SRC) {
2301                 OVERFLOW_CHECK_u64(array);
2302                 data->data_src = *array;
2303                 array++;
2304         }
2305
2306         if (type & PERF_SAMPLE_TRANSACTION) {
2307                 OVERFLOW_CHECK_u64(array);
2308                 data->transaction = *array;
2309                 array++;
2310         }
2311
2312         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2313         if (type & PERF_SAMPLE_REGS_INTR) {
2314                 OVERFLOW_CHECK_u64(array);
2315                 data->intr_regs.abi = *array;
2316                 array++;
2317
2318                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2319                         u64 mask = evsel->attr.sample_regs_intr;
2320
2321                         sz = hweight_long(mask) * sizeof(u64);
2322                         OVERFLOW_CHECK(array, sz, max_size);
2323                         data->intr_regs.mask = mask;
2324                         data->intr_regs.regs = (u64 *)array;
2325                         array = (void *)array + sz;
2326                 }
2327         }
2328
2329         data->phys_addr = 0;
2330         if (type & PERF_SAMPLE_PHYS_ADDR) {
2331                 data->phys_addr = *array;
2332                 array++;
2333         }
2334
2335         return 0;
2336 }
2337
2338 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2339                                        union perf_event *event,
2340                                        u64 *timestamp)
2341 {
2342         u64 type = evsel->attr.sample_type;
2343         const u64 *array;
2344
2345         if (!(type & PERF_SAMPLE_TIME))
2346                 return -1;
2347
2348         if (event->header.type != PERF_RECORD_SAMPLE) {
2349                 struct perf_sample data = {
2350                         .time = -1ULL,
2351                 };
2352
2353                 if (!evsel->attr.sample_id_all)
2354                         return -1;
2355                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2356                         return -1;
2357
2358                 *timestamp = data.time;
2359                 return 0;
2360         }
2361
2362         array = event->sample.array;
2363
2364         if (perf_event__check_size(event, evsel->sample_size))
2365                 return -EFAULT;
2366
2367         if (type & PERF_SAMPLE_IDENTIFIER)
2368                 array++;
2369
2370         if (type & PERF_SAMPLE_IP)
2371                 array++;
2372
2373         if (type & PERF_SAMPLE_TID)
2374                 array++;
2375
2376         if (type & PERF_SAMPLE_TIME)
2377                 *timestamp = *array;
2378
2379         return 0;
2380 }
2381
2382 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2383                                      u64 read_format)
2384 {
2385         size_t sz, result = sizeof(struct sample_event);
2386
2387         if (type & PERF_SAMPLE_IDENTIFIER)
2388                 result += sizeof(u64);
2389
2390         if (type & PERF_SAMPLE_IP)
2391                 result += sizeof(u64);
2392
2393         if (type & PERF_SAMPLE_TID)
2394                 result += sizeof(u64);
2395
2396         if (type & PERF_SAMPLE_TIME)
2397                 result += sizeof(u64);
2398
2399         if (type & PERF_SAMPLE_ADDR)
2400                 result += sizeof(u64);
2401
2402         if (type & PERF_SAMPLE_ID)
2403                 result += sizeof(u64);
2404
2405         if (type & PERF_SAMPLE_STREAM_ID)
2406                 result += sizeof(u64);
2407
2408         if (type & PERF_SAMPLE_CPU)
2409                 result += sizeof(u64);
2410
2411         if (type & PERF_SAMPLE_PERIOD)
2412                 result += sizeof(u64);
2413
2414         if (type & PERF_SAMPLE_READ) {
2415                 result += sizeof(u64);
2416                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2417                         result += sizeof(u64);
2418                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2419                         result += sizeof(u64);
2420                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2421                 if (read_format & PERF_FORMAT_GROUP) {
2422                         sz = sample->read.group.nr *
2423                              sizeof(struct sample_read_value);
2424                         result += sz;
2425                 } else {
2426                         result += sizeof(u64);
2427                 }
2428         }
2429
2430         if (type & PERF_SAMPLE_CALLCHAIN) {
2431                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2432                 result += sz;
2433         }
2434
2435         if (type & PERF_SAMPLE_RAW) {
2436                 result += sizeof(u32);
2437                 result += sample->raw_size;
2438         }
2439
2440         if (type & PERF_SAMPLE_BRANCH_STACK) {
2441                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2442                 sz += sizeof(u64);
2443                 result += sz;
2444         }
2445
2446         if (type & PERF_SAMPLE_REGS_USER) {
2447                 if (sample->user_regs.abi) {
2448                         result += sizeof(u64);
2449                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2450                         result += sz;
2451                 } else {
2452                         result += sizeof(u64);
2453                 }
2454         }
2455
2456         if (type & PERF_SAMPLE_STACK_USER) {
2457                 sz = sample->user_stack.size;
2458                 result += sizeof(u64);
2459                 if (sz) {
2460                         result += sz;
2461                         result += sizeof(u64);
2462                 }
2463         }
2464
2465         if (type & PERF_SAMPLE_WEIGHT)
2466                 result += sizeof(u64);
2467
2468         if (type & PERF_SAMPLE_DATA_SRC)
2469                 result += sizeof(u64);
2470
2471         if (type & PERF_SAMPLE_TRANSACTION)
2472                 result += sizeof(u64);
2473
2474         if (type & PERF_SAMPLE_REGS_INTR) {
2475                 if (sample->intr_regs.abi) {
2476                         result += sizeof(u64);
2477                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2478                         result += sz;
2479                 } else {
2480                         result += sizeof(u64);
2481                 }
2482         }
2483
2484         if (type & PERF_SAMPLE_PHYS_ADDR)
2485                 result += sizeof(u64);
2486
2487         return result;
2488 }
2489
2490 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2491                                   u64 read_format,
2492                                   const struct perf_sample *sample)
2493 {
2494         u64 *array;
2495         size_t sz;
2496         /*
2497          * used for cross-endian analysis. See git commit 65014ab3
2498          * for why this goofiness is needed.
2499          */
2500         union u64_swap u;
2501
2502         array = event->sample.array;
2503
2504         if (type & PERF_SAMPLE_IDENTIFIER) {
2505                 *array = sample->id;
2506                 array++;
2507         }
2508
2509         if (type & PERF_SAMPLE_IP) {
2510                 *array = sample->ip;
2511                 array++;
2512         }
2513
2514         if (type & PERF_SAMPLE_TID) {
2515                 u.val32[0] = sample->pid;
2516                 u.val32[1] = sample->tid;
2517                 *array = u.val64;
2518                 array++;
2519         }
2520
2521         if (type & PERF_SAMPLE_TIME) {
2522                 *array = sample->time;
2523                 array++;
2524         }
2525
2526         if (type & PERF_SAMPLE_ADDR) {
2527                 *array = sample->addr;
2528                 array++;
2529         }
2530
2531         if (type & PERF_SAMPLE_ID) {
2532                 *array = sample->id;
2533                 array++;
2534         }
2535
2536         if (type & PERF_SAMPLE_STREAM_ID) {
2537                 *array = sample->stream_id;
2538                 array++;
2539         }
2540
2541         if (type & PERF_SAMPLE_CPU) {
2542                 u.val32[0] = sample->cpu;
2543                 u.val32[1] = 0;
2544                 *array = u.val64;
2545                 array++;
2546         }
2547
2548         if (type & PERF_SAMPLE_PERIOD) {
2549                 *array = sample->period;
2550                 array++;
2551         }
2552
2553         if (type & PERF_SAMPLE_READ) {
2554                 if (read_format & PERF_FORMAT_GROUP)
2555                         *array = sample->read.group.nr;
2556                 else
2557                         *array = sample->read.one.value;
2558                 array++;
2559
2560                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2561                         *array = sample->read.time_enabled;
2562                         array++;
2563                 }
2564
2565                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2566                         *array = sample->read.time_running;
2567                         array++;
2568                 }
2569
2570                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2571                 if (read_format & PERF_FORMAT_GROUP) {
2572                         sz = sample->read.group.nr *
2573                              sizeof(struct sample_read_value);
2574                         memcpy(array, sample->read.group.values, sz);
2575                         array = (void *)array + sz;
2576                 } else {
2577                         *array = sample->read.one.id;
2578                         array++;
2579                 }
2580         }
2581
2582         if (type & PERF_SAMPLE_CALLCHAIN) {
2583                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2584                 memcpy(array, sample->callchain, sz);
2585                 array = (void *)array + sz;
2586         }
2587
2588         if (type & PERF_SAMPLE_RAW) {
2589                 u.val32[0] = sample->raw_size;
2590                 *array = u.val64;
2591                 array = (void *)array + sizeof(u32);
2592
2593                 memcpy(array, sample->raw_data, sample->raw_size);
2594                 array = (void *)array + sample->raw_size;
2595         }
2596
2597         if (type & PERF_SAMPLE_BRANCH_STACK) {
2598                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2599                 sz += sizeof(u64);
2600                 memcpy(array, sample->branch_stack, sz);
2601                 array = (void *)array + sz;
2602         }
2603
2604         if (type & PERF_SAMPLE_REGS_USER) {
2605                 if (sample->user_regs.abi) {
2606                         *array++ = sample->user_regs.abi;
2607                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2608                         memcpy(array, sample->user_regs.regs, sz);
2609                         array = (void *)array + sz;
2610                 } else {
2611                         *array++ = 0;
2612                 }
2613         }
2614
2615         if (type & PERF_SAMPLE_STACK_USER) {
2616                 sz = sample->user_stack.size;
2617                 *array++ = sz;
2618                 if (sz) {
2619                         memcpy(array, sample->user_stack.data, sz);
2620                         array = (void *)array + sz;
2621                         *array++ = sz;
2622                 }
2623         }
2624
2625         if (type & PERF_SAMPLE_WEIGHT) {
2626                 *array = sample->weight;
2627                 array++;
2628         }
2629
2630         if (type & PERF_SAMPLE_DATA_SRC) {
2631                 *array = sample->data_src;
2632                 array++;
2633         }
2634
2635         if (type & PERF_SAMPLE_TRANSACTION) {
2636                 *array = sample->transaction;
2637                 array++;
2638         }
2639
2640         if (type & PERF_SAMPLE_REGS_INTR) {
2641                 if (sample->intr_regs.abi) {
2642                         *array++ = sample->intr_regs.abi;
2643                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2644                         memcpy(array, sample->intr_regs.regs, sz);
2645                         array = (void *)array + sz;
2646                 } else {
2647                         *array++ = 0;
2648                 }
2649         }
2650
2651         if (type & PERF_SAMPLE_PHYS_ADDR) {
2652                 *array = sample->phys_addr;
2653                 array++;
2654         }
2655
2656         return 0;
2657 }
2658
2659 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2660 {
2661         return pevent_find_field(evsel->tp_format, name);
2662 }
2663
2664 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2665                          const char *name)
2666 {
2667         struct format_field *field = perf_evsel__field(evsel, name);
2668         int offset;
2669
2670         if (!field)
2671                 return NULL;
2672
2673         offset = field->offset;
2674
2675         if (field->flags & FIELD_IS_DYNAMIC) {
2676                 offset = *(int *)(sample->raw_data + field->offset);
2677                 offset &= 0xffff;
2678         }
2679
2680         return sample->raw_data + offset;
2681 }
2682
2683 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2684                          bool needs_swap)
2685 {
2686         u64 value;
2687         void *ptr = sample->raw_data + field->offset;
2688
2689         switch (field->size) {
2690         case 1:
2691                 return *(u8 *)ptr;
2692         case 2:
2693                 value = *(u16 *)ptr;
2694                 break;
2695         case 4:
2696                 value = *(u32 *)ptr;
2697                 break;
2698         case 8:
2699                 memcpy(&value, ptr, sizeof(u64));
2700                 break;
2701         default:
2702                 return 0;
2703         }
2704
2705         if (!needs_swap)
2706                 return value;
2707
2708         switch (field->size) {
2709         case 2:
2710                 return bswap_16(value);
2711         case 4:
2712                 return bswap_32(value);
2713         case 8:
2714                 return bswap_64(value);
2715         default:
2716                 return 0;
2717         }
2718
2719         return 0;
2720 }
2721
2722 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2723                        const char *name)
2724 {
2725         struct format_field *field = perf_evsel__field(evsel, name);
2726
2727         if (!field)
2728                 return 0;
2729
2730         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2731 }
2732
2733 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2734                           char *msg, size_t msgsize)
2735 {
2736         int paranoid;
2737
2738         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2739             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2740             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2741                 /*
2742                  * If it's cycles then fall back to hrtimer based
2743                  * cpu-clock-tick sw counter, which is always available even if
2744                  * no PMU support.
2745                  *
2746                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2747                  * b0a873e).
2748                  */
2749                 scnprintf(msg, msgsize, "%s",
2750 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2751
2752                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2753                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2754
2755                 zfree(&evsel->name);
2756                 return true;
2757         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2758                    (paranoid = perf_event_paranoid()) > 1) {
2759                 const char *name = perf_evsel__name(evsel);
2760                 char *new_name;
2761                 const char *sep = ":";
2762
2763                 /* Is there already the separator in the name. */
2764                 if (strchr(name, '/') ||
2765                     strchr(name, ':'))
2766                         sep = "";
2767
2768                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2769                         return false;
2770
2771                 if (evsel->name)
2772                         free(evsel->name);
2773                 evsel->name = new_name;
2774                 scnprintf(msg, msgsize,
2775 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2776                 evsel->attr.exclude_kernel = 1;
2777
2778                 return true;
2779         }
2780
2781         return false;
2782 }
2783
2784 static bool find_process(const char *name)
2785 {
2786         size_t len = strlen(name);
2787         DIR *dir;
2788         struct dirent *d;
2789         int ret = -1;
2790
2791         dir = opendir(procfs__mountpoint());
2792         if (!dir)
2793                 return false;
2794
2795         /* Walk through the directory. */
2796         while (ret && (d = readdir(dir)) != NULL) {
2797                 char path[PATH_MAX];
2798                 char *data;
2799                 size_t size;
2800
2801                 if ((d->d_type != DT_DIR) ||
2802                      !strcmp(".", d->d_name) ||
2803                      !strcmp("..", d->d_name))
2804                         continue;
2805
2806                 scnprintf(path, sizeof(path), "%s/%s/comm",
2807                           procfs__mountpoint(), d->d_name);
2808
2809                 if (filename__read_str(path, &data, &size))
2810                         continue;
2811
2812                 ret = strncmp(name, data, len);
2813                 free(data);
2814         }
2815
2816         closedir(dir);
2817         return ret ? false : true;
2818 }
2819
2820 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2821                               int err, char *msg, size_t size)
2822 {
2823         char sbuf[STRERR_BUFSIZE];
2824         int printed = 0;
2825
2826         switch (err) {
2827         case EPERM:
2828         case EACCES:
2829                 if (err == EPERM)
2830                         printed = scnprintf(msg, size,
2831                                 "No permission to enable %s event.\n\n",
2832                                 perf_evsel__name(evsel));
2833
2834                 return scnprintf(msg + printed, size - printed,
2835                  "You may not have permission to collect %sstats.\n\n"
2836                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2837                  "which controls use of the performance events system by\n"
2838                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2839                  "The current value is %d:\n\n"
2840                  "  -1: Allow use of (almost) all events by all users\n"
2841                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2842                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2843                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2844                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2845                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2846                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2847                  "      kernel.perf_event_paranoid = -1\n" ,
2848                                  target->system_wide ? "system-wide " : "",
2849                                  perf_event_paranoid());
2850         case ENOENT:
2851                 return scnprintf(msg, size, "The %s event is not supported.",
2852                                  perf_evsel__name(evsel));
2853         case EMFILE:
2854                 return scnprintf(msg, size, "%s",
2855                          "Too many events are opened.\n"
2856                          "Probably the maximum number of open file descriptors has been reached.\n"
2857                          "Hint: Try again after reducing the number of events.\n"
2858                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2859         case ENOMEM:
2860                 if (evsel__has_callchain(evsel) &&
2861                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2862                         return scnprintf(msg, size,
2863                                          "Not enough memory to setup event with callchain.\n"
2864                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2865                                          "Hint: Current value: %d", sysctl__max_stack());
2866                 break;
2867         case ENODEV:
2868                 if (target->cpu_list)
2869                         return scnprintf(msg, size, "%s",
2870          "No such device - did you specify an out-of-range profile CPU?");
2871                 break;
2872         case EOPNOTSUPP:
2873                 if (evsel->attr.sample_period != 0)
2874                         return scnprintf(msg, size,
2875         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2876                                          perf_evsel__name(evsel));
2877                 if (evsel->attr.precise_ip)
2878                         return scnprintf(msg, size, "%s",
2879         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2880 #if defined(__i386__) || defined(__x86_64__)
2881                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2882                         return scnprintf(msg, size, "%s",
2883         "No hardware sampling interrupt available.\n");
2884 #endif
2885                 break;
2886         case EBUSY:
2887                 if (find_process("oprofiled"))
2888                         return scnprintf(msg, size,
2889         "The PMU counters are busy/taken by another profiler.\n"
2890         "We found oprofile daemon running, please stop it and try again.");
2891                 break;
2892         case EINVAL:
2893                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2894                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2895                 if (perf_missing_features.clockid)
2896                         return scnprintf(msg, size, "clockid feature not supported.");
2897                 if (perf_missing_features.clockid_wrong)
2898                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2899                 break;
2900         default:
2901                 break;
2902         }
2903
2904         return scnprintf(msg, size,
2905         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2906         "/bin/dmesg | grep -i perf may provide additional information.\n",
2907                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2908                          perf_evsel__name(evsel));
2909 }
2910
2911 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2912 {
2913         if (evsel && evsel->evlist)
2914                 return evsel->evlist->env;
2915         return NULL;
2916 }