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/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/err.h>
  19 #include <sys/ioctl.h>
  20 #include <sys/resource.h>
  21 #include "asm/bug.h"
  22 #include "callchain.h"
  23 #include "cgroup.h"
  24 #include "event.h"
  25 #include "evsel.h"
  26 #include "evlist.h"
  27 #include "util.h"
  28 #include "cpumap.h"
  29 #include "thread_map.h"
  30 #include "target.h"
  31 #include "perf_regs.h"
  32 #include "debug.h"
  33 #include "trace-event.h"
  34 #include "stat.h"
  35 #include "util/parse-branch-options.h"
  36
  37 #include "sane_ctype.h"
  38
  39 static struct {
  40         bool sample_id_all;
  41         bool exclude_guest;
  42         bool mmap2;
  43         bool cloexec;
  44         bool clockid;
  45         bool clockid_wrong;
  46         bool lbr_flags;
  47         bool write_backward;
  48 } perf_missing_features;
  49
  50 static clockid_t clockid;
  51
  52 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  53 {
  54         return 0;
  55 }
  56
  57 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  58 {
  59 }
  60
  61 static struct {
  62         size_t  size;
  63         int     (*init)(struct perf_evsel *evsel);
  64         void    (*fini)(struct perf_evsel *evsel);
  65 } perf_evsel__object = {
  66         .size = sizeof(struct perf_evsel),
  67         .init = perf_evsel__no_extra_init,
  68         .fini = perf_evsel__no_extra_fini,
  69 };
  70
  71 int perf_evsel__object_config(size_t object_size,
  72                               int (*init)(struct perf_evsel *evsel),
  73                               void (*fini)(struct perf_evsel *evsel))
  74 {
  75
  76         if (object_size == 0)
  77                 goto set_methods;
  78
  79         if (perf_evsel__object.size > object_size)
  80                 return -EINVAL;
  81
  82         perf_evsel__object.size = object_size;
  83
  84 set_methods:
  85         if (init != NULL)
  86                 perf_evsel__object.init = init;
  87
  88         if (fini != NULL)
  89                 perf_evsel__object.fini = fini;
  90
  91         return 0;
  92 }
  93
  94 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  95
  96 int __perf_evsel__sample_size(u64 sample_type)
  97 {
  98         u64 mask = sample_type & PERF_SAMPLE_MASK;
  99         int size = 0;
 100         int i;
 101
 102         for (i = 0; i < 64; i++) {
 103                 if (mask & (1ULL << i))
 104                         size++;
 105         }
 106
 107         size *= sizeof(u64);
 108
 109         return size;
 110 }
 111
 112 /**
 113  * __perf_evsel__calc_id_pos - calculate id_pos.
 114  * @sample_type: sample type
 115  *
 116  * This function returns the position of the event id (PERF_SAMPLE_ID or
 117  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 118  * sample_event.
 119  */
 120 static int __perf_evsel__calc_id_pos(u64 sample_type)
 121 {
 122         int idx = 0;
 123
 124         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 125                 return 0;
 126
 127         if (!(sample_type & PERF_SAMPLE_ID))
 128                 return -1;
 129
 130         if (sample_type & PERF_SAMPLE_IP)
 131                 idx += 1;
 132
 133         if (sample_type & PERF_SAMPLE_TID)
 134                 idx += 1;
 135
 136         if (sample_type & PERF_SAMPLE_TIME)
 137                 idx += 1;
 138
 139         if (sample_type & PERF_SAMPLE_ADDR)
 140                 idx += 1;
 141
 142         return idx;
 143 }
 144
 145 /**
 146  * __perf_evsel__calc_is_pos - calculate is_pos.
 147  * @sample_type: sample type
 148  *
 149  * This function returns the position (counting backwards) of the event id
 150  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 151  * sample_id_all is used there is an id sample appended to non-sample events.
 152  */
 153 static int __perf_evsel__calc_is_pos(u64 sample_type)
 154 {
 155         int idx = 1;
 156
 157         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 158                 return 1;
 159
 160         if (!(sample_type & PERF_SAMPLE_ID))
 161                 return -1;
 162
 163         if (sample_type & PERF_SAMPLE_CPU)
 164                 idx += 1;
 165
 166         if (sample_type & PERF_SAMPLE_STREAM_ID)
 167                 idx += 1;
 168
 169         return idx;
 170 }
 171
 172 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 173 {
 174         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 175         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 176 }
 177
 178 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 179                                   enum perf_event_sample_format bit)
 180 {
 181         if (!(evsel->attr.sample_type & bit)) {
 182                 evsel->attr.sample_type |= bit;
 183                 evsel->sample_size += sizeof(u64);
 184                 perf_evsel__calc_id_pos(evsel);
 185         }
 186 }
 187
 188 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 189                                     enum perf_event_sample_format bit)
 190 {
 191         if (evsel->attr.sample_type & bit) {
 192                 evsel->attr.sample_type &= ~bit;
 193                 evsel->sample_size -= sizeof(u64);
 194                 perf_evsel__calc_id_pos(evsel);
 195         }
 196 }
 197
 198 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 199                                bool can_sample_identifier)
 200 {
 201         if (can_sample_identifier) {
 202                 perf_evsel__reset_sample_bit(evsel, ID);
 203                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 204         } else {
 205                 perf_evsel__set_sample_bit(evsel, ID);
 206         }
 207         evsel->attr.read_format |= PERF_FORMAT_ID;
 208 }
 209
 210 /**
 211  * perf_evsel__is_function_event - Return whether given evsel is a function
 212  * trace event
 213  *
 214  * @evsel - evsel selector to be tested
 215  *
 216  * Return %true if event is function trace event
 217  */
 218 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 219 {
 220 #define FUNCTION_EVENT "ftrace:function"
 221
 222         return evsel->name &&
 223                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 224
 225 #undef FUNCTION_EVENT
 226 }
 227
 228 void perf_evsel__init(struct perf_evsel *evsel,
 229                       struct perf_event_attr *attr, int idx)
 230 {
 231         evsel->idx         = idx;
 232         evsel->tracking    = !idx;
 233         evsel->attr        = *attr;
 234         evsel->leader      = evsel;
 235         evsel->unit        = "";
 236         evsel->scale       = 1.0;
 237         evsel->evlist      = NULL;
 238         evsel->bpf_fd      = -1;
 239         INIT_LIST_HEAD(&evsel->node);
 240         INIT_LIST_HEAD(&evsel->config_terms);
 241         perf_evsel__object.init(evsel);
 242         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 243         perf_evsel__calc_id_pos(evsel);
 244         evsel->cmdline_group_boundary = false;
 245         evsel->metric_expr   = NULL;
 246         evsel->metric_name   = NULL;
 247         evsel->metric_events = NULL;
 248         evsel->collect_stat  = false;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (evsel != NULL)
 256                 perf_evsel__init(evsel, attr, idx);
 257
 258         if (perf_evsel__is_bpf_output(evsel)) {
 259                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 260                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 261                 evsel->attr.sample_period = 1;
 262         }
 263
 264         return evsel;
 265 }
 266
 267 struct perf_evsel *perf_evsel__new_cycles(void)
 268 {
 269         struct perf_event_attr attr = {
 270                 .type   = PERF_TYPE_HARDWARE,
 271                 .config = PERF_COUNT_HW_CPU_CYCLES,
 272         };
 273         struct perf_evsel *evsel;
 274
 275         event_attr_init(&attr);
 276
 277         perf_event_attr__set_max_precise_ip(&attr);
 278
 279         evsel = perf_evsel__new(&attr);
 280         if (evsel == NULL)
 281                 goto out;
 282
 283         /* use asprintf() because free(evsel) assumes name is allocated */
 284         if (asprintf(&evsel->name, "cycles%.*s",
 285                      attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
 286                 goto error_free;
 287 out:
 288         return evsel;
 289 error_free:
 290         perf_evsel__delete(evsel);
 291         evsel = NULL;
 292         goto out;
 293 }
 294
 295 /*
 296  * Returns pointer with encoded error via <linux/err.h> interface.
 297  */
 298 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 299 {
 300         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 301         int err = -ENOMEM;
 302
 303         if (evsel == NULL) {
 304                 goto out_err;
 305         } else {
 306                 struct perf_event_attr attr = {
 307                         .type          = PERF_TYPE_TRACEPOINT,
 308                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 309                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 310                 };
 311
 312                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 313                         goto out_free;
 314
 315                 evsel->tp_format = trace_event__tp_format(sys, name);
 316                 if (IS_ERR(evsel->tp_format)) {
 317                         err = PTR_ERR(evsel->tp_format);
 318                         goto out_free;
 319                 }
 320
 321                 event_attr_init(&attr);
 322                 attr.config = evsel->tp_format->id;
 323                 attr.sample_period = 1;
 324                 perf_evsel__init(evsel, &attr, idx);
 325         }
 326
 327         return evsel;
 328
 329 out_free:
 330         zfree(&evsel->name);
 331         free(evsel);
 332 out_err:
 333         return ERR_PTR(err);
 334 }
 335
 336 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 337         "cycles",
 338         "instructions",
 339         "cache-references",
 340         "cache-misses",
 341         "branches",
 342         "branch-misses",
 343         "bus-cycles",
 344         "stalled-cycles-frontend",
 345         "stalled-cycles-backend",
 346         "ref-cycles",
 347 };
 348
 349 static const char *__perf_evsel__hw_name(u64 config)
 350 {
 351         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 352                 return perf_evsel__hw_names[config];
 353
 354         return "unknown-hardware";
 355 }
 356
 357 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 358 {
 359         int colon = 0, r = 0;
 360         struct perf_event_attr *attr = &evsel->attr;
 361         bool exclude_guest_default = false;
 362
 363 #define MOD_PRINT(context, mod) do {                                    \
 364                 if (!attr->exclude_##context) {                         \
 365                         if (!colon) colon = ++r;                        \
 366                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 367                 } } while(0)
 368
 369         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 370                 MOD_PRINT(kernel, 'k');
 371                 MOD_PRINT(user, 'u');
 372                 MOD_PRINT(hv, 'h');
 373                 exclude_guest_default = true;
 374         }
 375
 376         if (attr->precise_ip) {
 377                 if (!colon)
 378                         colon = ++r;
 379                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 380                 exclude_guest_default = true;
 381         }
 382
 383         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 384                 MOD_PRINT(host, 'H');
 385                 MOD_PRINT(guest, 'G');
 386         }
 387 #undef MOD_PRINT
 388         if (colon)
 389                 bf[colon - 1] = ':';
 390         return r;
 391 }
 392
 393 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 394 {
 395         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 396         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 397 }
 398
 399 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 400         "cpu-clock",
 401         "task-clock",
 402         "page-faults",
 403         "context-switches",
 404         "cpu-migrations",
 405         "minor-faults",
 406         "major-faults",
 407         "alignment-faults",
 408         "emulation-faults",
 409         "dummy",
 410 };
 411
 412 static const char *__perf_evsel__sw_name(u64 config)
 413 {
 414         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 415                 return perf_evsel__sw_names[config];
 416         return "unknown-software";
 417 }
 418
 419 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 420 {
 421         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 422         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 423 }
 424
 425 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 426 {
 427         int r;
 428
 429         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 430
 431         if (type & HW_BREAKPOINT_R)
 432                 r += scnprintf(bf + r, size - r, "r");
 433
 434         if (type & HW_BREAKPOINT_W)
 435                 r += scnprintf(bf + r, size - r, "w");
 436
 437         if (type & HW_BREAKPOINT_X)
 438                 r += scnprintf(bf + r, size - r, "x");
 439
 440         return r;
 441 }
 442
 443 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 444 {
 445         struct perf_event_attr *attr = &evsel->attr;
 446         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 447         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 448 }
 449
 450 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 451                                 [PERF_EVSEL__MAX_ALIASES] = {
 452  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 453  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 454  { "LLC",       "L2",                                                   },
 455  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 456  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 457  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 458  { "node",                                                              },
 459 };
 460
 461 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 462                                    [PERF_EVSEL__MAX_ALIASES] = {
 463  { "load",      "loads",        "read",                                 },
 464  { "store",     "stores",       "write",                                },
 465  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 466 };
 467
 468 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 469                                        [PERF_EVSEL__MAX_ALIASES] = {
 470  { "refs",      "Reference",    "ops",          "access",               },
 471  { "misses",    "miss",                                                 },
 472 };
 473
 474 #define C(x)            PERF_COUNT_HW_CACHE_##x
 475 #define CACHE_READ      (1 << C(OP_READ))
 476 #define CACHE_WRITE     (1 << C(OP_WRITE))
 477 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 478 #define COP(x)          (1 << x)
 479
 480 /*
 481  * cache operartion stat
 482  * L1I : Read and prefetch only
 483  * ITLB and BPU : Read-only
 484  */
 485 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 486  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 487  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 488  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 489  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 490  [C(ITLB)]      = (CACHE_READ),
 491  [C(BPU)]       = (CACHE_READ),
 492  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 493 };
 494
 495 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 496 {
 497         if (perf_evsel__hw_cache_stat[type] & COP(op))
 498                 return true;    /* valid */
 499         else
 500                 return false;   /* invalid */
 501 }
 502
 503 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 504                                             char *bf, size_t size)
 505 {
 506         if (result) {
 507                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 508                                  perf_evsel__hw_cache_op[op][0],
 509                                  perf_evsel__hw_cache_result[result][0]);
 510         }
 511
 512         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 513                          perf_evsel__hw_cache_op[op][1]);
 514 }
 515
 516 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 517 {
 518         u8 op, result, type = (config >>  0) & 0xff;
 519         const char *err = "unknown-ext-hardware-cache-type";
 520
 521         if (type >= PERF_COUNT_HW_CACHE_MAX)
 522                 goto out_err;
 523
 524         op = (config >>  8) & 0xff;
 525         err = "unknown-ext-hardware-cache-op";
 526         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 527                 goto out_err;
 528
 529         result = (config >> 16) & 0xff;
 530         err = "unknown-ext-hardware-cache-result";
 531         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 532                 goto out_err;
 533
 534         err = "invalid-cache";
 535         if (!perf_evsel__is_cache_op_valid(type, op))
 536                 goto out_err;
 537
 538         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 539 out_err:
 540         return scnprintf(bf, size, "%s", err);
 541 }
 542
 543 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 544 {
 545         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 546         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 547 }
 548
 549 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 550 {
 551         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 552         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 553 }
 554
 555 const char *perf_evsel__name(struct perf_evsel *evsel)
 556 {
 557         char bf[128];
 558
 559         if (evsel->name)
 560                 return evsel->name;
 561
 562         switch (evsel->attr.type) {
 563         case PERF_TYPE_RAW:
 564                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 565                 break;
 566
 567         case PERF_TYPE_HARDWARE:
 568                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 569                 break;
 570
 571         case PERF_TYPE_HW_CACHE:
 572                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 573                 break;
 574
 575         case PERF_TYPE_SOFTWARE:
 576                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 577                 break;
 578
 579         case PERF_TYPE_TRACEPOINT:
 580                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 581                 break;
 582
 583         case PERF_TYPE_BREAKPOINT:
 584                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 585                 break;
 586
 587         default:
 588                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 589                           evsel->attr.type);
 590                 break;
 591         }
 592
 593         evsel->name = strdup(bf);
 594
 595         return evsel->name ?: "unknown";
 596 }
 597
 598 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 599 {
 600         return evsel->group_name ?: "anon group";
 601 }
 602
 603 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 604 {
 605         int ret;
 606         struct perf_evsel *pos;
 607         const char *group_name = perf_evsel__group_name(evsel);
 608
 609         ret = scnprintf(buf, size, "%s", group_name);
 610
 611         ret += scnprintf(buf + ret, size - ret, " { %s",
 612                          perf_evsel__name(evsel));
 613
 614         for_each_group_member(pos, evsel)
 615                 ret += scnprintf(buf + ret, size - ret, ", %s",
 616                                  perf_evsel__name(pos));
 617
 618         ret += scnprintf(buf + ret, size - ret, " }");
 619
 620         return ret;
 621 }
 622
 623 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 624                                   struct record_opts *opts,
 625                                   struct callchain_param *param)
 626 {
 627         bool function = perf_evsel__is_function_event(evsel);
 628         struct perf_event_attr *attr = &evsel->attr;
 629
 630         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 631
 632         attr->sample_max_stack = param->max_stack;
 633
 634         if (param->record_mode == CALLCHAIN_LBR) {
 635                 if (!opts->branch_stack) {
 636                         if (attr->exclude_user) {
 637                                 pr_warning("LBR callstack option is only available "
 638                                            "to get user callchain information. "
 639                                            "Falling back to framepointers.\n");
 640                         } else {
 641                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 642                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 643                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 644                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 645                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 646                         }
 647                 } else
 648                          pr_warning("Cannot use LBR callstack with branch stack. "
 649                                     "Falling back to framepointers.\n");
 650         }
 651
 652         if (param->record_mode == CALLCHAIN_DWARF) {
 653                 if (!function) {
 654                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 655                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 656                         attr->sample_regs_user = PERF_REGS_MASK;
 657                         attr->sample_stack_user = param->dump_size;
 658                         attr->exclude_callchain_user = 1;
 659                 } else {
 660                         pr_info("Cannot use DWARF unwind for function trace event,"
 661                                 " falling back to framepointers.\n");
 662                 }
 663         }
 664
 665         if (function) {
 666                 pr_info("Disabling user space callchains for function trace event.\n");
 667                 attr->exclude_callchain_user = 1;
 668         }
 669 }
 670
 671 static void
 672 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 673                             struct callchain_param *param)
 674 {
 675         struct perf_event_attr *attr = &evsel->attr;
 676
 677         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 678         if (param->record_mode == CALLCHAIN_LBR) {
 679                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 680                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 681                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 682         }
 683         if (param->record_mode == CALLCHAIN_DWARF) {
 684                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 685                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 686         }
 687 }
 688
 689 static void apply_config_terms(struct perf_evsel *evsel,
 690                                struct record_opts *opts)
 691 {
 692         struct perf_evsel_config_term *term;
 693         struct list_head *config_terms = &evsel->config_terms;
 694         struct perf_event_attr *attr = &evsel->attr;
 695         struct callchain_param param;
 696         u32 dump_size = 0;
 697         int max_stack = 0;
 698         const char *callgraph_buf = NULL;
 699
 700         /* callgraph default */
 701         param.record_mode = callchain_param.record_mode;
 702
 703         list_for_each_entry(term, config_terms, list) {
 704                 switch (term->type) {
 705                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 706                         attr->sample_period = term->val.period;
 707                         attr->freq = 0;
 708                         break;
 709                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 710                         attr->sample_freq = term->val.freq;
 711                         attr->freq = 1;
 712                         break;
 713                 case PERF_EVSEL__CONFIG_TERM_TIME:
 714                         if (term->val.time)
 715                                 perf_evsel__set_sample_bit(evsel, TIME);
 716                         else
 717                                 perf_evsel__reset_sample_bit(evsel, TIME);
 718                         break;
 719                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 720                         callgraph_buf = term->val.callgraph;
 721                         break;
 722                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 723                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 724                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 725                                 parse_branch_str(term->val.branch,
 726                                                  &attr->branch_sample_type);
 727                         } else
 728                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 729                         break;
 730                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 731                         dump_size = term->val.stack_user;
 732                         break;
 733                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 734                         max_stack = term->val.max_stack;
 735                         break;
 736                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 737                         /*
 738                          * attr->inherit should has already been set by
 739                          * perf_evsel__config. If user explicitly set
 740                          * inherit using config terms, override global
 741                          * opt->no_inherit setting.
 742                          */
 743                         attr->inherit = term->val.inherit ? 1 : 0;
 744                         break;
 745                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 746                         attr->write_backward = term->val.overwrite ? 1 : 0;
 747                         break;
 748                 default:
 749                         break;
 750                 }
 751         }
 752
 753         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 754         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 755                 if (max_stack) {
 756                         param.max_stack = max_stack;
 757                         if (callgraph_buf == NULL)
 758                                 callgraph_buf = "fp";
 759                 }
 760
 761                 /* parse callgraph parameters */
 762                 if (callgraph_buf != NULL) {
 763                         if (!strcmp(callgraph_buf, "no")) {
 764                                 param.enabled = false;
 765                                 param.record_mode = CALLCHAIN_NONE;
 766                         } else {
 767                                 param.enabled = true;
 768                                 if (parse_callchain_record(callgraph_buf, &param)) {
 769                                         pr_err("per-event callgraph setting for %s failed. "
 770                                                "Apply callgraph global setting for it\n",
 771                                                evsel->name);
 772                                         return;
 773                                 }
 774                         }
 775                 }
 776                 if (dump_size > 0) {
 777                         dump_size = round_up(dump_size, sizeof(u64));
 778                         param.dump_size = dump_size;
 779                 }
 780
 781                 /* If global callgraph set, clear it */
 782                 if (callchain_param.enabled)
 783                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 784
 785                 /* set perf-event callgraph */
 786                 if (param.enabled)
 787                         perf_evsel__config_callchain(evsel, opts, &param);
 788         }
 789 }
 790
 791 /*
 792  * The enable_on_exec/disabled value strategy:
 793  *
 794  *  1) For any type of traced program:
 795  *    - all independent events and group leaders are disabled
 796  *    - all group members are enabled
 797  *
 798  *     Group members are ruled by group leaders. They need to
 799  *     be enabled, because the group scheduling relies on that.
 800  *
 801  *  2) For traced programs executed by perf:
 802  *     - all independent events and group leaders have
 803  *       enable_on_exec set
 804  *     - we don't specifically enable or disable any event during
 805  *       the record command
 806  *
 807  *     Independent events and group leaders are initially disabled
 808  *     and get enabled by exec. Group members are ruled by group
 809  *     leaders as stated in 1).
 810  *
 811  *  3) For traced programs attached by perf (pid/tid):
 812  *     - we specifically enable or disable all events during
 813  *       the record command
 814  *
 815  *     When attaching events to already running traced we
 816  *     enable/disable events specifically, as there's no
 817  *     initial traced exec call.
 818  */
 819 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 820                         struct callchain_param *callchain)
 821 {
 822         struct perf_evsel *leader = evsel->leader;
 823         struct perf_event_attr *attr = &evsel->attr;
 824         int track = evsel->tracking;
 825         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 826
 827         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 828         attr->inherit       = !opts->no_inherit;
 829         attr->write_backward = opts->overwrite ? 1 : 0;
 830
 831         perf_evsel__set_sample_bit(evsel, IP);
 832         perf_evsel__set_sample_bit(evsel, TID);
 833
 834         if (evsel->sample_read) {
 835                 perf_evsel__set_sample_bit(evsel, READ);
 836
 837                 /*
 838                  * We need ID even in case of single event, because
 839                  * PERF_SAMPLE_READ process ID specific data.
 840                  */
 841                 perf_evsel__set_sample_id(evsel, false);
 842
 843                 /*
 844                  * Apply group format only if we belong to group
 845                  * with more than one members.
 846                  */
 847                 if (leader->nr_members > 1) {
 848                         attr->read_format |= PERF_FORMAT_GROUP;
 849                         attr->inherit = 0;
 850                 }
 851         }
 852
 853         /*
 854          * We default some events to have a default interval. But keep
 855          * it a weak assumption overridable by the user.
 856          */
 857         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 858                                      opts->user_interval != ULLONG_MAX)) {
 859                 if (opts->freq) {
 860                         perf_evsel__set_sample_bit(evsel, PERIOD);
 861                         attr->freq              = 1;
 862                         attr->sample_freq       = opts->freq;
 863                 } else {
 864                         attr->sample_period = opts->default_interval;
 865                 }
 866         }
 867
 868         /*
 869          * Disable sampling for all group members other
 870          * than leader in case leader 'leads' the sampling.
 871          */
 872         if ((leader != evsel) && leader->sample_read) {
 873                 attr->sample_freq   = 0;
 874                 attr->sample_period = 0;
 875         }
 876
 877         if (opts->no_samples)
 878                 attr->sample_freq = 0;
 879
 880         if (opts->inherit_stat)
 881                 attr->inherit_stat = 1;
 882
 883         if (opts->sample_address) {
 884                 perf_evsel__set_sample_bit(evsel, ADDR);
 885                 attr->mmap_data = track;
 886         }
 887
 888         /*
 889          * We don't allow user space callchains for  function trace
 890          * event, due to issues with page faults while tracing page
 891          * fault handler and its overall trickiness nature.
 892          */
 893         if (perf_evsel__is_function_event(evsel))
 894                 evsel->attr.exclude_callchain_user = 1;
 895
 896         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 897                 perf_evsel__config_callchain(evsel, opts, callchain);
 898
 899         if (opts->sample_intr_regs) {
 900                 attr->sample_regs_intr = opts->sample_intr_regs;
 901                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 902         }
 903
 904         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 905                 perf_evsel__set_sample_bit(evsel, CPU);
 906
 907         if (opts->period)
 908                 perf_evsel__set_sample_bit(evsel, PERIOD);
 909
 910         /*
 911          * When the user explicitly disabled time don't force it here.
 912          */
 913         if (opts->sample_time &&
 914             (!perf_missing_features.sample_id_all &&
 915             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 916              opts->sample_time_set)))
 917                 perf_evsel__set_sample_bit(evsel, TIME);
 918
 919         if (opts->raw_samples && !evsel->no_aux_samples) {
 920                 perf_evsel__set_sample_bit(evsel, TIME);
 921                 perf_evsel__set_sample_bit(evsel, RAW);
 922                 perf_evsel__set_sample_bit(evsel, CPU);
 923         }
 924
 925         if (opts->sample_address)
 926                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 927
 928         if (opts->no_buffering) {
 929                 attr->watermark = 0;
 930                 attr->wakeup_events = 1;
 931         }
 932         if (opts->branch_stack && !evsel->no_aux_samples) {
 933                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 934                 attr->branch_sample_type = opts->branch_stack;
 935         }
 936
 937         if (opts->sample_weight)
 938                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 939
 940         attr->task  = track;
 941         attr->mmap  = track;
 942         attr->mmap2 = track && !perf_missing_features.mmap2;
 943         attr->comm  = track;
 944
 945         if (opts->record_namespaces)
 946                 attr->namespaces  = track;
 947
 948         if (opts->record_switch_events)
 949                 attr->context_switch = track;
 950
 951         if (opts->sample_transaction)
 952                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 953
 954         if (opts->running_time) {
 955                 evsel->attr.read_format |=
 956                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 957                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 958         }
 959
 960         /*
 961          * XXX see the function comment above
 962          *
 963          * Disabling only independent events or group leaders,
 964          * keeping group members enabled.
 965          */
 966         if (perf_evsel__is_group_leader(evsel))
 967                 attr->disabled = 1;
 968
 969         /*
 970          * Setting enable_on_exec for independent events and
 971          * group leaders for traced executed by perf.
 972          */
 973         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 974                 !opts->initial_delay)
 975                 attr->enable_on_exec = 1;
 976
 977         if (evsel->immediate) {
 978                 attr->disabled = 0;
 979                 attr->enable_on_exec = 0;
 980         }
 981
 982         clockid = opts->clockid;
 983         if (opts->use_clockid) {
 984                 attr->use_clockid = 1;
 985                 attr->clockid = opts->clockid;
 986         }
 987
 988         if (evsel->precise_max)
 989                 perf_event_attr__set_max_precise_ip(attr);
 990
 991         if (opts->all_user) {
 992                 attr->exclude_kernel = 1;
 993                 attr->exclude_user   = 0;
 994         }
 995
 996         if (opts->all_kernel) {
 997                 attr->exclude_kernel = 0;
 998                 attr->exclude_user   = 1;
 999         }
1000
1001         /*
1002          * Apply event specific term settings,
1003          * it overloads any global configuration.
1004          */
1005         apply_config_terms(evsel, opts);
1006
1007         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1008 }
1009
1010 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1011 {
1012         if (evsel->system_wide)
1013                 nthreads = 1;
1014
1015         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1016
1017         if (evsel->fd) {
1018                 int cpu, thread;
1019                 for (cpu = 0; cpu < ncpus; cpu++) {
1020                         for (thread = 0; thread < nthreads; thread++) {
1021                                 FD(evsel, cpu, thread) = -1;
1022                         }
1023                 }
1024         }
1025
1026         return evsel->fd != NULL ? 0 : -ENOMEM;
1027 }
1028
1029 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1030                           int ioc,  void *arg)
1031 {
1032         int cpu, thread;
1033
1034         if (evsel->system_wide)
1035                 nthreads = 1;
1036
1037         for (cpu = 0; cpu < ncpus; cpu++) {
1038                 for (thread = 0; thread < nthreads; thread++) {
1039                         int fd = FD(evsel, cpu, thread),
1040                             err = ioctl(fd, ioc, arg);
1041
1042                         if (err)
1043                                 return err;
1044                 }
1045         }
1046
1047         return 0;
1048 }
1049
1050 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1051                              const char *filter)
1052 {
1053         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1054                                      PERF_EVENT_IOC_SET_FILTER,
1055                                      (void *)filter);
1056 }
1057
1058 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1059 {
1060         char *new_filter = strdup(filter);
1061
1062         if (new_filter != NULL) {
1063                 free(evsel->filter);
1064                 evsel->filter = new_filter;
1065                 return 0;
1066         }
1067
1068         return -1;
1069 }
1070
1071 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1072                                      const char *fmt, const char *filter)
1073 {
1074         char *new_filter;
1075
1076         if (evsel->filter == NULL)
1077                 return perf_evsel__set_filter(evsel, filter);
1078
1079         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1080                 free(evsel->filter);
1081                 evsel->filter = new_filter;
1082                 return 0;
1083         }
1084
1085         return -1;
1086 }
1087
1088 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1089 {
1090         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1091 }
1092
1093 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1094 {
1095         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1096 }
1097
1098 int perf_evsel__enable(struct perf_evsel *evsel)
1099 {
1100         int nthreads = thread_map__nr(evsel->threads);
1101         int ncpus = cpu_map__nr(evsel->cpus);
1102
1103         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1104                                      PERF_EVENT_IOC_ENABLE,
1105                                      0);
1106 }
1107
1108 int perf_evsel__disable(struct perf_evsel *evsel)
1109 {
1110         int nthreads = thread_map__nr(evsel->threads);
1111         int ncpus = cpu_map__nr(evsel->cpus);
1112
1113         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1114                                      PERF_EVENT_IOC_DISABLE,
1115                                      0);
1116 }
1117
1118 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1119 {
1120         if (ncpus == 0 || nthreads == 0)
1121                 return 0;
1122
1123         if (evsel->system_wide)
1124                 nthreads = 1;
1125
1126         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1127         if (evsel->sample_id == NULL)
1128                 return -ENOMEM;
1129
1130         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1131         if (evsel->id == NULL) {
1132                 xyarray__delete(evsel->sample_id);
1133                 evsel->sample_id = NULL;
1134                 return -ENOMEM;
1135         }
1136
1137         return 0;
1138 }
1139
1140 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1141 {
1142         xyarray__delete(evsel->fd);
1143         evsel->fd = NULL;
1144 }
1145
1146 static void perf_evsel__free_id(struct perf_evsel *evsel)
1147 {
1148         xyarray__delete(evsel->sample_id);
1149         evsel->sample_id = NULL;
1150         zfree(&evsel->id);
1151 }
1152
1153 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1154 {
1155         struct perf_evsel_config_term *term, *h;
1156
1157         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1158                 list_del(&term->list);
1159                 free(term);
1160         }
1161 }
1162
1163 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1164 {
1165         int cpu, thread;
1166
1167         if (evsel->system_wide)
1168                 nthreads = 1;
1169
1170         for (cpu = 0; cpu < ncpus; cpu++)
1171                 for (thread = 0; thread < nthreads; ++thread) {
1172                         close(FD(evsel, cpu, thread));
1173                         FD(evsel, cpu, thread) = -1;
1174                 }
1175 }
1176
1177 void perf_evsel__exit(struct perf_evsel *evsel)
1178 {
1179         assert(list_empty(&evsel->node));
1180         assert(evsel->evlist == NULL);
1181         perf_evsel__free_fd(evsel);
1182         perf_evsel__free_id(evsel);
1183         perf_evsel__free_config_terms(evsel);
1184         close_cgroup(evsel->cgrp);
1185         cpu_map__put(evsel->cpus);
1186         cpu_map__put(evsel->own_cpus);
1187         thread_map__put(evsel->threads);
1188         zfree(&evsel->group_name);
1189         zfree(&evsel->name);
1190         perf_evsel__object.fini(evsel);
1191 }
1192
1193 void perf_evsel__delete(struct perf_evsel *evsel)
1194 {
1195         perf_evsel__exit(evsel);
1196         free(evsel);
1197 }
1198
1199 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1200                                 struct perf_counts_values *count)
1201 {
1202         struct perf_counts_values tmp;
1203
1204         if (!evsel->prev_raw_counts)
1205                 return;
1206
1207         if (cpu == -1) {
1208                 tmp = evsel->prev_raw_counts->aggr;
1209                 evsel->prev_raw_counts->aggr = *count;
1210         } else {
1211                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1212                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1213         }
1214
1215         count->val = count->val - tmp.val;
1216         count->ena = count->ena - tmp.ena;
1217         count->run = count->run - tmp.run;
1218 }
1219
1220 void perf_counts_values__scale(struct perf_counts_values *count,
1221                                bool scale, s8 *pscaled)
1222 {
1223         s8 scaled = 0;
1224
1225         if (scale) {
1226                 if (count->run == 0) {
1227                         scaled = -1;
1228                         count->val = 0;
1229                 } else if (count->run < count->ena) {
1230                         scaled = 1;
1231                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1232                 }
1233         } else
1234                 count->ena = count->run = 0;
1235
1236         if (pscaled)
1237                 *pscaled = scaled;
1238 }
1239
1240 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1241                      struct perf_counts_values *count)
1242 {
1243         memset(count, 0, sizeof(*count));
1244
1245         if (FD(evsel, cpu, thread) < 0)
1246                 return -EINVAL;
1247
1248         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) <= 0)
1249                 return -errno;
1250
1251         return 0;
1252 }
1253
1254 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1255                               int cpu, int thread, bool scale)
1256 {
1257         struct perf_counts_values count;
1258         size_t nv = scale ? 3 : 1;
1259
1260         if (FD(evsel, cpu, thread) < 0)
1261                 return -EINVAL;
1262
1263         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1264                 return -ENOMEM;
1265
1266         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1267                 return -errno;
1268
1269         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1270         perf_counts_values__scale(&count, scale, NULL);
1271         *perf_counts(evsel->counts, cpu, thread) = count;
1272         return 0;
1273 }
1274
1275 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1276 {
1277         struct perf_evsel *leader = evsel->leader;
1278         int fd;
1279
1280         if (perf_evsel__is_group_leader(evsel))
1281                 return -1;
1282
1283         /*
1284          * Leader must be already processed/open,
1285          * if not it's a bug.
1286          */
1287         BUG_ON(!leader->fd);
1288
1289         fd = FD(leader, cpu, thread);
1290         BUG_ON(fd == -1);
1291
1292         return fd;
1293 }
1294
1295 struct bit_names {
1296         int bit;
1297         const char *name;
1298 };
1299
1300 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1301 {
1302         bool first_bit = true;
1303         int i = 0;
1304
1305         do {
1306                 if (value & bits[i].bit) {
1307                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1308                         first_bit = false;
1309                 }
1310         } while (bits[++i].name != NULL);
1311 }
1312
1313 static void __p_sample_type(char *buf, size_t size, u64 value)
1314 {
1315 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1316         struct bit_names bits[] = {
1317                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1318                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1319                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1320                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1321                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1322                 bit_name(WEIGHT),
1323                 { .name = NULL, }
1324         };
1325 #undef bit_name
1326         __p_bits(buf, size, value, bits);
1327 }
1328
1329 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1330 {
1331 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1332         struct bit_names bits[] = {
1333                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1334                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1335                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1336                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1337                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1338                 { .name = NULL, }
1339         };
1340 #undef bit_name
1341         __p_bits(buf, size, value, bits);
1342 }
1343
1344 static void __p_read_format(char *buf, size_t size, u64 value)
1345 {
1346 #define bit_name(n) { PERF_FORMAT_##n, #n }
1347         struct bit_names bits[] = {
1348                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1349                 bit_name(ID), bit_name(GROUP),
1350                 { .name = NULL, }
1351         };
1352 #undef bit_name
1353         __p_bits(buf, size, value, bits);
1354 }
1355
1356 #define BUF_SIZE                1024
1357
1358 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1359 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1360 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1361 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1362 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1363 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1364
1365 #define PRINT_ATTRn(_n, _f, _p)                         \
1366 do {                                                    \
1367         if (attr->_f) {                                 \
1368                 _p(attr->_f);                           \
1369                 ret += attr__fprintf(fp, _n, buf, priv);\
1370         }                                               \
1371 } while (0)
1372
1373 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1374
1375 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1376                              attr__fprintf_f attr__fprintf, void *priv)
1377 {
1378         char buf[BUF_SIZE];
1379         int ret = 0;
1380
1381         PRINT_ATTRf(type, p_unsigned);
1382         PRINT_ATTRf(size, p_unsigned);
1383         PRINT_ATTRf(config, p_hex);
1384         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1385         PRINT_ATTRf(sample_type, p_sample_type);
1386         PRINT_ATTRf(read_format, p_read_format);
1387
1388         PRINT_ATTRf(disabled, p_unsigned);
1389         PRINT_ATTRf(inherit, p_unsigned);
1390         PRINT_ATTRf(pinned, p_unsigned);
1391         PRINT_ATTRf(exclusive, p_unsigned);
1392         PRINT_ATTRf(exclude_user, p_unsigned);
1393         PRINT_ATTRf(exclude_kernel, p_unsigned);
1394         PRINT_ATTRf(exclude_hv, p_unsigned);
1395         PRINT_ATTRf(exclude_idle, p_unsigned);
1396         PRINT_ATTRf(mmap, p_unsigned);
1397         PRINT_ATTRf(comm, p_unsigned);
1398         PRINT_ATTRf(freq, p_unsigned);
1399         PRINT_ATTRf(inherit_stat, p_unsigned);
1400         PRINT_ATTRf(enable_on_exec, p_unsigned);
1401         PRINT_ATTRf(task, p_unsigned);
1402         PRINT_ATTRf(watermark, p_unsigned);
1403         PRINT_ATTRf(precise_ip, p_unsigned);
1404         PRINT_ATTRf(mmap_data, p_unsigned);
1405         PRINT_ATTRf(sample_id_all, p_unsigned);
1406         PRINT_ATTRf(exclude_host, p_unsigned);
1407         PRINT_ATTRf(exclude_guest, p_unsigned);
1408         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1409         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1410         PRINT_ATTRf(mmap2, p_unsigned);
1411         PRINT_ATTRf(comm_exec, p_unsigned);
1412         PRINT_ATTRf(use_clockid, p_unsigned);
1413         PRINT_ATTRf(context_switch, p_unsigned);
1414         PRINT_ATTRf(write_backward, p_unsigned);
1415
1416         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1417         PRINT_ATTRf(bp_type, p_unsigned);
1418         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1419         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1420         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1421         PRINT_ATTRf(sample_regs_user, p_hex);
1422         PRINT_ATTRf(sample_stack_user, p_unsigned);
1423         PRINT_ATTRf(clockid, p_signed);
1424         PRINT_ATTRf(sample_regs_intr, p_hex);
1425         PRINT_ATTRf(aux_watermark, p_unsigned);
1426         PRINT_ATTRf(sample_max_stack, p_unsigned);
1427
1428         return ret;
1429 }
1430
1431 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1432                                 void *priv __attribute__((unused)))
1433 {
1434         return fprintf(fp, "  %-32s %s\n", name, val);
1435 }
1436
1437 static bool ignore_missing_thread(struct perf_evsel *evsel,
1438                                   struct thread_map *threads,
1439                                   int thread, int err)
1440 {
1441         if (!evsel->ignore_missing_thread)
1442                 return false;
1443
1444         /* The system wide setup does not work with threads. */
1445         if (evsel->system_wide)
1446                 return false;
1447
1448         /* The -ESRCH is perf event syscall errno for pid's not found. */
1449         if (err != -ESRCH)
1450                 return false;
1451
1452         /* If there's only one thread, let it fail. */
1453         if (threads->nr == 1)
1454                 return false;
1455
1456         if (thread_map__remove(threads, thread))
1457                 return false;
1458
1459         pr_warning("WARNING: Ignored open failure for pid %d\n",
1460                    thread_map__pid(threads, thread));
1461         return true;
1462 }
1463
1464 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1465                      struct thread_map *threads)
1466 {
1467         int cpu, thread, nthreads;
1468         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1469         int pid = -1, err;
1470         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1471
1472         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1473                 return -EINVAL;
1474
1475         if (cpus == NULL) {
1476                 static struct cpu_map *empty_cpu_map;
1477
1478                 if (empty_cpu_map == NULL) {
1479                         empty_cpu_map = cpu_map__dummy_new();
1480                         if (empty_cpu_map == NULL)
1481                                 return -ENOMEM;
1482                 }
1483
1484                 cpus = empty_cpu_map;
1485         }
1486
1487         if (threads == NULL) {
1488                 static struct thread_map *empty_thread_map;
1489
1490                 if (empty_thread_map == NULL) {
1491                         empty_thread_map = thread_map__new_by_tid(-1);
1492                         if (empty_thread_map == NULL)
1493                                 return -ENOMEM;
1494                 }
1495
1496                 threads = empty_thread_map;
1497         }
1498
1499         if (evsel->system_wide)
1500                 nthreads = 1;
1501         else
1502                 nthreads = threads->nr;
1503
1504         if (evsel->fd == NULL &&
1505             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1506                 return -ENOMEM;
1507
1508         if (evsel->cgrp) {
1509                 flags |= PERF_FLAG_PID_CGROUP;
1510                 pid = evsel->cgrp->fd;
1511         }
1512
1513 fallback_missing_features:
1514         if (perf_missing_features.clockid_wrong)
1515                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1516         if (perf_missing_features.clockid) {
1517                 evsel->attr.use_clockid = 0;
1518                 evsel->attr.clockid = 0;
1519         }
1520         if (perf_missing_features.cloexec)
1521                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1522         if (perf_missing_features.mmap2)
1523                 evsel->attr.mmap2 = 0;
1524         if (perf_missing_features.exclude_guest)
1525                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1526         if (perf_missing_features.lbr_flags)
1527                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1528                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1529 retry_sample_id:
1530         if (perf_missing_features.sample_id_all)
1531                 evsel->attr.sample_id_all = 0;
1532
1533         if (verbose >= 2) {
1534                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1535                 fprintf(stderr, "perf_event_attr:\n");
1536                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1537                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1538         }
1539
1540         for (cpu = 0; cpu < cpus->nr; cpu++) {
1541
1542                 for (thread = 0; thread < nthreads; thread++) {
1543                         int fd, group_fd;
1544
1545                         if (!evsel->cgrp && !evsel->system_wide)
1546                                 pid = thread_map__pid(threads, thread);
1547
1548                         group_fd = get_group_fd(evsel, cpu, thread);
1549 retry_open:
1550                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1551                                   pid, cpus->map[cpu], group_fd, flags);
1552
1553                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1554                                                  group_fd, flags);
1555
1556                         FD(evsel, cpu, thread) = fd;
1557
1558                         if (fd < 0) {
1559                                 err = -errno;
1560
1561                                 if (ignore_missing_thread(evsel, threads, thread, err)) {
1562                                         /*
1563                                          * We just removed 1 thread, so take a step
1564                                          * back on thread index and lower the upper
1565                                          * nthreads limit.
1566                                          */
1567                                         nthreads--;
1568                                         thread--;
1569
1570                                         /* ... and pretend like nothing have happened. */
1571                                         err = 0;
1572                                         continue;
1573                                 }
1574
1575                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1576                                           err);
1577                                 goto try_fallback;
1578                         }
1579
1580                         pr_debug2(" = %d\n", fd);
1581
1582                         if (evsel->bpf_fd >= 0) {
1583                                 int evt_fd = fd;
1584                                 int bpf_fd = evsel->bpf_fd;
1585
1586                                 err = ioctl(evt_fd,
1587                                             PERF_EVENT_IOC_SET_BPF,
1588                                             bpf_fd);
1589                                 if (err && errno != EEXIST) {
1590                                         pr_err("failed to attach bpf fd %d: %s\n",
1591                                                bpf_fd, strerror(errno));
1592                                         err = -EINVAL;
1593                                         goto out_close;
1594                                 }
1595                         }
1596
1597                         set_rlimit = NO_CHANGE;
1598
1599                         /*
1600                          * If we succeeded but had to kill clockid, fail and
1601                          * have perf_evsel__open_strerror() print us a nice
1602                          * error.
1603                          */
1604                         if (perf_missing_features.clockid ||
1605                             perf_missing_features.clockid_wrong) {
1606                                 err = -EINVAL;
1607                                 goto out_close;
1608                         }
1609                 }
1610         }
1611
1612         return 0;
1613
1614 try_fallback:
1615         /*
1616          * perf stat needs between 5 and 22 fds per CPU. When we run out
1617          * of them try to increase the limits.
1618          */
1619         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1620                 struct rlimit l;
1621                 int old_errno = errno;
1622
1623                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1624                         if (set_rlimit == NO_CHANGE)
1625                                 l.rlim_cur = l.rlim_max;
1626                         else {
1627                                 l.rlim_cur = l.rlim_max + 1000;
1628                                 l.rlim_max = l.rlim_cur;
1629                         }
1630                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1631                                 set_rlimit++;
1632                                 errno = old_errno;
1633                                 goto retry_open;
1634                         }
1635                 }
1636                 errno = old_errno;
1637         }
1638
1639         if (err != -EINVAL || cpu > 0 || thread > 0)
1640                 goto out_close;
1641
1642         /*
1643          * Must probe features in the order they were added to the
1644          * perf_event_attr interface.
1645          */
1646         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1647                 perf_missing_features.write_backward = true;
1648                 goto out_close;
1649         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1650                 perf_missing_features.clockid_wrong = true;
1651                 goto fallback_missing_features;
1652         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1653                 perf_missing_features.clockid = true;
1654                 goto fallback_missing_features;
1655         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1656                 perf_missing_features.cloexec = true;
1657                 goto fallback_missing_features;
1658         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1659                 perf_missing_features.mmap2 = true;
1660                 goto fallback_missing_features;
1661         } else if (!perf_missing_features.exclude_guest &&
1662                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1663                 perf_missing_features.exclude_guest = true;
1664                 goto fallback_missing_features;
1665         } else if (!perf_missing_features.sample_id_all) {
1666                 perf_missing_features.sample_id_all = true;
1667                 goto retry_sample_id;
1668         } else if (!perf_missing_features.lbr_flags &&
1669                         (evsel->attr.branch_sample_type &
1670                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1671                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1672                 perf_missing_features.lbr_flags = true;
1673                 goto fallback_missing_features;
1674         }
1675 out_close:
1676         do {
1677                 while (--thread >= 0) {
1678                         close(FD(evsel, cpu, thread));
1679                         FD(evsel, cpu, thread) = -1;
1680                 }
1681                 thread = nthreads;
1682         } while (--cpu >= 0);
1683         return err;
1684 }
1685
1686 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1687 {
1688         if (evsel->fd == NULL)
1689                 return;
1690
1691         perf_evsel__close_fd(evsel, ncpus, nthreads);
1692         perf_evsel__free_fd(evsel);
1693 }
1694
1695 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1696                              struct cpu_map *cpus)
1697 {
1698         return perf_evsel__open(evsel, cpus, NULL);
1699 }
1700
1701 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1702                                 struct thread_map *threads)
1703 {
1704         return perf_evsel__open(evsel, NULL, threads);
1705 }
1706
1707 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1708                                        const union perf_event *event,
1709                                        struct perf_sample *sample)
1710 {
1711         u64 type = evsel->attr.sample_type;
1712         const u64 *array = event->sample.array;
1713         bool swapped = evsel->needs_swap;
1714         union u64_swap u;
1715
1716         array += ((event->header.size -
1717                    sizeof(event->header)) / sizeof(u64)) - 1;
1718
1719         if (type & PERF_SAMPLE_IDENTIFIER) {
1720                 sample->id = *array;
1721                 array--;
1722         }
1723
1724         if (type & PERF_SAMPLE_CPU) {
1725                 u.val64 = *array;
1726                 if (swapped) {
1727                         /* undo swap of u64, then swap on individual u32s */
1728                         u.val64 = bswap_64(u.val64);
1729                         u.val32[0] = bswap_32(u.val32[0]);
1730                 }
1731
1732                 sample->cpu = u.val32[0];
1733                 array--;
1734         }
1735
1736         if (type & PERF_SAMPLE_STREAM_ID) {
1737                 sample->stream_id = *array;
1738                 array--;
1739         }
1740
1741         if (type & PERF_SAMPLE_ID) {
1742                 sample->id = *array;
1743                 array--;
1744         }
1745
1746         if (type & PERF_SAMPLE_TIME) {
1747                 sample->time = *array;
1748                 array--;
1749         }
1750
1751         if (type & PERF_SAMPLE_TID) {
1752                 u.val64 = *array;
1753                 if (swapped) {
1754                         /* undo swap of u64, then swap on individual u32s */
1755                         u.val64 = bswap_64(u.val64);
1756                         u.val32[0] = bswap_32(u.val32[0]);
1757                         u.val32[1] = bswap_32(u.val32[1]);
1758                 }
1759
1760                 sample->pid = u.val32[0];
1761                 sample->tid = u.val32[1];
1762                 array--;
1763         }
1764
1765         return 0;
1766 }
1767
1768 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1769                             u64 size)
1770 {
1771         return size > max_size || offset + size > endp;
1772 }
1773
1774 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1775         do {                                                            \
1776                 if (overflow(endp, (max_size), (offset), (size)))       \
1777                         return -EFAULT;                                 \
1778         } while (0)
1779
1780 #define OVERFLOW_CHECK_u64(offset) \
1781         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1782
1783 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1784                              struct perf_sample *data)
1785 {
1786         u64 type = evsel->attr.sample_type;
1787         bool swapped = evsel->needs_swap;
1788         const u64 *array;
1789         u16 max_size = event->header.size;
1790         const void *endp = (void *)event + max_size;
1791         u64 sz;
1792
1793         /*
1794          * used for cross-endian analysis. See git commit 65014ab3
1795          * for why this goofiness is needed.
1796          */
1797         union u64_swap u;
1798
1799         memset(data, 0, sizeof(*data));
1800         data->cpu = data->pid = data->tid = -1;
1801         data->stream_id = data->id = data->time = -1ULL;
1802         data->period = evsel->attr.sample_period;
1803         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1804
1805         if (event->header.type != PERF_RECORD_SAMPLE) {
1806                 if (!evsel->attr.sample_id_all)
1807                         return 0;
1808                 return perf_evsel__parse_id_sample(evsel, event, data);
1809         }
1810
1811         array = event->sample.array;
1812
1813         /*
1814          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1815          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1816          * check the format does not go past the end of the event.
1817          */
1818         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1819                 return -EFAULT;
1820
1821         data->id = -1ULL;
1822         if (type & PERF_SAMPLE_IDENTIFIER) {
1823                 data->id = *array;
1824                 array++;
1825         }
1826
1827         if (type & PERF_SAMPLE_IP) {
1828                 data->ip = *array;
1829                 array++;
1830         }
1831
1832         if (type & PERF_SAMPLE_TID) {
1833                 u.val64 = *array;
1834                 if (swapped) {
1835                         /* undo swap of u64, then swap on individual u32s */
1836                         u.val64 = bswap_64(u.val64);
1837                         u.val32[0] = bswap_32(u.val32[0]);
1838                         u.val32[1] = bswap_32(u.val32[1]);
1839                 }
1840
1841                 data->pid = u.val32[0];
1842                 data->tid = u.val32[1];
1843                 array++;
1844         }
1845
1846         if (type & PERF_SAMPLE_TIME) {
1847                 data->time = *array;
1848                 array++;
1849         }
1850
1851         data->addr = 0;
1852         if (type & PERF_SAMPLE_ADDR) {
1853                 data->addr = *array;
1854                 array++;
1855         }
1856
1857         if (type & PERF_SAMPLE_ID) {
1858                 data->id = *array;
1859                 array++;
1860         }
1861
1862         if (type & PERF_SAMPLE_STREAM_ID) {
1863                 data->stream_id = *array;
1864                 array++;
1865         }
1866
1867         if (type & PERF_SAMPLE_CPU) {
1868
1869                 u.val64 = *array;
1870                 if (swapped) {
1871                         /* undo swap of u64, then swap on individual u32s */
1872                         u.val64 = bswap_64(u.val64);
1873                         u.val32[0] = bswap_32(u.val32[0]);
1874                 }
1875
1876                 data->cpu = u.val32[0];
1877                 array++;
1878         }
1879
1880         if (type & PERF_SAMPLE_PERIOD) {
1881                 data->period = *array;
1882                 array++;
1883         }
1884
1885         if (type & PERF_SAMPLE_READ) {
1886                 u64 read_format = evsel->attr.read_format;
1887
1888                 OVERFLOW_CHECK_u64(array);
1889                 if (read_format & PERF_FORMAT_GROUP)
1890                         data->read.group.nr = *array;
1891                 else
1892                         data->read.one.value = *array;
1893
1894                 array++;
1895
1896                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1897                         OVERFLOW_CHECK_u64(array);
1898                         data->read.time_enabled = *array;
1899                         array++;
1900                 }
1901
1902                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1903                         OVERFLOW_CHECK_u64(array);
1904                         data->read.time_running = *array;
1905                         array++;
1906                 }
1907
1908                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1909                 if (read_format & PERF_FORMAT_GROUP) {
1910                         const u64 max_group_nr = UINT64_MAX /
1911                                         sizeof(struct sample_read_value);
1912
1913                         if (data->read.group.nr > max_group_nr)
1914                                 return -EFAULT;
1915                         sz = data->read.group.nr *
1916                              sizeof(struct sample_read_value);
1917                         OVERFLOW_CHECK(array, sz, max_size);
1918                         data->read.group.values =
1919                                         (struct sample_read_value *)array;
1920                         array = (void *)array + sz;
1921                 } else {
1922                         OVERFLOW_CHECK_u64(array);
1923                         data->read.one.id = *array;
1924                         array++;
1925                 }
1926         }
1927
1928         if (type & PERF_SAMPLE_CALLCHAIN) {
1929                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1930
1931                 OVERFLOW_CHECK_u64(array);
1932                 data->callchain = (struct ip_callchain *)array++;
1933                 if (data->callchain->nr > max_callchain_nr)
1934                         return -EFAULT;
1935                 sz = data->callchain->nr * sizeof(u64);
1936                 OVERFLOW_CHECK(array, sz, max_size);
1937                 array = (void *)array + sz;
1938         }
1939
1940         if (type & PERF_SAMPLE_RAW) {
1941                 OVERFLOW_CHECK_u64(array);
1942                 u.val64 = *array;
1943                 if (WARN_ONCE(swapped,
1944                               "Endianness of raw data not corrected!\n")) {
1945                         /* undo swap of u64, then swap on individual u32s */
1946                         u.val64 = bswap_64(u.val64);
1947                         u.val32[0] = bswap_32(u.val32[0]);
1948                         u.val32[1] = bswap_32(u.val32[1]);
1949                 }
1950                 data->raw_size = u.val32[0];
1951                 array = (void *)array + sizeof(u32);
1952
1953                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1954                 data->raw_data = (void *)array;
1955                 array = (void *)array + data->raw_size;
1956         }
1957
1958         if (type & PERF_SAMPLE_BRANCH_STACK) {
1959                 const u64 max_branch_nr = UINT64_MAX /
1960                                           sizeof(struct branch_entry);
1961
1962                 OVERFLOW_CHECK_u64(array);
1963                 data->branch_stack = (struct branch_stack *)array++;
1964
1965                 if (data->branch_stack->nr > max_branch_nr)
1966                         return -EFAULT;
1967                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1968                 OVERFLOW_CHECK(array, sz, max_size);
1969                 array = (void *)array + sz;
1970         }
1971
1972         if (type & PERF_SAMPLE_REGS_USER) {
1973                 OVERFLOW_CHECK_u64(array);
1974                 data->user_regs.abi = *array;
1975                 array++;
1976
1977                 if (data->user_regs.abi) {
1978                         u64 mask = evsel->attr.sample_regs_user;
1979
1980                         sz = hweight_long(mask) * sizeof(u64);
1981                         OVERFLOW_CHECK(array, sz, max_size);
1982                         data->user_regs.mask = mask;
1983                         data->user_regs.regs = (u64 *)array;
1984                         array = (void *)array + sz;
1985                 }
1986         }
1987
1988         if (type & PERF_SAMPLE_STACK_USER) {
1989                 OVERFLOW_CHECK_u64(array);
1990                 sz = *array++;
1991
1992                 data->user_stack.offset = ((char *)(array - 1)
1993                                           - (char *) event);
1994
1995                 if (!sz) {
1996                         data->user_stack.size = 0;
1997                 } else {
1998                         OVERFLOW_CHECK(array, sz, max_size);
1999                         data->user_stack.data = (char *)array;
2000                         array = (void *)array + sz;
2001                         OVERFLOW_CHECK_u64(array);
2002                         data->user_stack.size = *array++;
2003                         if (WARN_ONCE(data->user_stack.size > sz,
2004                                       "user stack dump failure\n"))
2005                                 return -EFAULT;
2006                 }
2007         }
2008
2009         if (type & PERF_SAMPLE_WEIGHT) {
2010                 OVERFLOW_CHECK_u64(array);
2011                 data->weight = *array;
2012                 array++;
2013         }
2014
2015         data->data_src = PERF_MEM_DATA_SRC_NONE;
2016         if (type & PERF_SAMPLE_DATA_SRC) {
2017                 OVERFLOW_CHECK_u64(array);
2018                 data->data_src = *array;
2019                 array++;
2020         }
2021
2022         data->transaction = 0;
2023         if (type & PERF_SAMPLE_TRANSACTION) {
2024                 OVERFLOW_CHECK_u64(array);
2025                 data->transaction = *array;
2026                 array++;
2027         }
2028
2029         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2030         if (type & PERF_SAMPLE_REGS_INTR) {
2031                 OVERFLOW_CHECK_u64(array);
2032                 data->intr_regs.abi = *array;
2033                 array++;
2034
2035                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2036                         u64 mask = evsel->attr.sample_regs_intr;
2037
2038                         sz = hweight_long(mask) * sizeof(u64);
2039                         OVERFLOW_CHECK(array, sz, max_size);
2040                         data->intr_regs.mask = mask;
2041                         data->intr_regs.regs = (u64 *)array;
2042                         array = (void *)array + sz;
2043                 }
2044         }
2045
2046         return 0;
2047 }
2048
2049 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2050                                      u64 read_format)
2051 {
2052         size_t sz, result = sizeof(struct sample_event);
2053
2054         if (type & PERF_SAMPLE_IDENTIFIER)
2055                 result += sizeof(u64);
2056
2057         if (type & PERF_SAMPLE_IP)
2058                 result += sizeof(u64);
2059
2060         if (type & PERF_SAMPLE_TID)
2061                 result += sizeof(u64);
2062
2063         if (type & PERF_SAMPLE_TIME)
2064                 result += sizeof(u64);
2065
2066         if (type & PERF_SAMPLE_ADDR)
2067                 result += sizeof(u64);
2068
2069         if (type & PERF_SAMPLE_ID)
2070                 result += sizeof(u64);
2071
2072         if (type & PERF_SAMPLE_STREAM_ID)
2073                 result += sizeof(u64);
2074
2075         if (type & PERF_SAMPLE_CPU)
2076                 result += sizeof(u64);
2077
2078         if (type & PERF_SAMPLE_PERIOD)
2079                 result += sizeof(u64);
2080
2081         if (type & PERF_SAMPLE_READ) {
2082                 result += sizeof(u64);
2083                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2084                         result += sizeof(u64);
2085                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2086                         result += sizeof(u64);
2087                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2088                 if (read_format & PERF_FORMAT_GROUP) {
2089                         sz = sample->read.group.nr *
2090                              sizeof(struct sample_read_value);
2091                         result += sz;
2092                 } else {
2093                         result += sizeof(u64);
2094                 }
2095         }
2096
2097         if (type & PERF_SAMPLE_CALLCHAIN) {
2098                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2099                 result += sz;
2100         }
2101
2102         if (type & PERF_SAMPLE_RAW) {
2103                 result += sizeof(u32);
2104                 result += sample->raw_size;
2105         }
2106
2107         if (type & PERF_SAMPLE_BRANCH_STACK) {
2108                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2109                 sz += sizeof(u64);
2110                 result += sz;
2111         }
2112
2113         if (type & PERF_SAMPLE_REGS_USER) {
2114                 if (sample->user_regs.abi) {
2115                         result += sizeof(u64);
2116                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2117                         result += sz;
2118                 } else {
2119                         result += sizeof(u64);
2120                 }
2121         }
2122
2123         if (type & PERF_SAMPLE_STACK_USER) {
2124                 sz = sample->user_stack.size;
2125                 result += sizeof(u64);
2126                 if (sz) {
2127                         result += sz;
2128                         result += sizeof(u64);
2129                 }
2130         }
2131
2132         if (type & PERF_SAMPLE_WEIGHT)
2133                 result += sizeof(u64);
2134
2135         if (type & PERF_SAMPLE_DATA_SRC)
2136                 result += sizeof(u64);
2137
2138         if (type & PERF_SAMPLE_TRANSACTION)
2139                 result += sizeof(u64);
2140
2141         if (type & PERF_SAMPLE_REGS_INTR) {
2142                 if (sample->intr_regs.abi) {
2143                         result += sizeof(u64);
2144                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2145                         result += sz;
2146                 } else {
2147                         result += sizeof(u64);
2148                 }
2149         }
2150
2151         return result;
2152 }
2153
2154 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2155                                   u64 read_format,
2156                                   const struct perf_sample *sample,
2157                                   bool swapped)
2158 {
2159         u64 *array;
2160         size_t sz;
2161         /*
2162          * used for cross-endian analysis. See git commit 65014ab3
2163          * for why this goofiness is needed.
2164          */
2165         union u64_swap u;
2166
2167         array = event->sample.array;
2168
2169         if (type & PERF_SAMPLE_IDENTIFIER) {
2170                 *array = sample->id;
2171                 array++;
2172         }
2173
2174         if (type & PERF_SAMPLE_IP) {
2175                 *array = sample->ip;
2176                 array++;
2177         }
2178
2179         if (type & PERF_SAMPLE_TID) {
2180                 u.val32[0] = sample->pid;
2181                 u.val32[1] = sample->tid;
2182                 if (swapped) {
2183                         /*
2184                          * Inverse of what is done in perf_evsel__parse_sample
2185                          */
2186                         u.val32[0] = bswap_32(u.val32[0]);
2187                         u.val32[1] = bswap_32(u.val32[1]);
2188                         u.val64 = bswap_64(u.val64);
2189                 }
2190
2191                 *array = u.val64;
2192                 array++;
2193         }
2194
2195         if (type & PERF_SAMPLE_TIME) {
2196                 *array = sample->time;
2197                 array++;
2198         }
2199
2200         if (type & PERF_SAMPLE_ADDR) {
2201                 *array = sample->addr;
2202                 array++;
2203         }
2204
2205         if (type & PERF_SAMPLE_ID) {
2206                 *array = sample->id;
2207                 array++;
2208         }
2209
2210         if (type & PERF_SAMPLE_STREAM_ID) {
2211                 *array = sample->stream_id;
2212                 array++;
2213         }
2214
2215         if (type & PERF_SAMPLE_CPU) {
2216                 u.val32[0] = sample->cpu;
2217                 if (swapped) {
2218                         /*
2219                          * Inverse of what is done in perf_evsel__parse_sample
2220                          */
2221                         u.val32[0] = bswap_32(u.val32[0]);
2222                         u.val64 = bswap_64(u.val64);
2223                 }
2224                 *array = u.val64;
2225                 array++;
2226         }
2227
2228         if (type & PERF_SAMPLE_PERIOD) {
2229                 *array = sample->period;
2230                 array++;
2231         }
2232
2233         if (type & PERF_SAMPLE_READ) {
2234                 if (read_format & PERF_FORMAT_GROUP)
2235                         *array = sample->read.group.nr;
2236                 else
2237                         *array = sample->read.one.value;
2238                 array++;
2239
2240                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2241                         *array = sample->read.time_enabled;
2242                         array++;
2243                 }
2244
2245                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2246                         *array = sample->read.time_running;
2247                         array++;
2248                 }
2249
2250                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2251                 if (read_format & PERF_FORMAT_GROUP) {
2252                         sz = sample->read.group.nr *
2253                              sizeof(struct sample_read_value);
2254                         memcpy(array, sample->read.group.values, sz);
2255                         array = (void *)array + sz;
2256                 } else {
2257                         *array = sample->read.one.id;
2258                         array++;
2259                 }
2260         }
2261
2262         if (type & PERF_SAMPLE_CALLCHAIN) {
2263                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2264                 memcpy(array, sample->callchain, sz);
2265                 array = (void *)array + sz;
2266         }
2267
2268         if (type & PERF_SAMPLE_RAW) {
2269                 u.val32[0] = sample->raw_size;
2270                 if (WARN_ONCE(swapped,
2271                               "Endianness of raw data not corrected!\n")) {
2272                         /*
2273                          * Inverse of what is done in perf_evsel__parse_sample
2274                          */
2275                         u.val32[0] = bswap_32(u.val32[0]);
2276                         u.val32[1] = bswap_32(u.val32[1]);
2277                         u.val64 = bswap_64(u.val64);
2278                 }
2279                 *array = u.val64;
2280                 array = (void *)array + sizeof(u32);
2281
2282                 memcpy(array, sample->raw_data, sample->raw_size);
2283                 array = (void *)array + sample->raw_size;
2284         }
2285
2286         if (type & PERF_SAMPLE_BRANCH_STACK) {
2287                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2288                 sz += sizeof(u64);
2289                 memcpy(array, sample->branch_stack, sz);
2290                 array = (void *)array + sz;
2291         }
2292
2293         if (type & PERF_SAMPLE_REGS_USER) {
2294                 if (sample->user_regs.abi) {
2295                         *array++ = sample->user_regs.abi;
2296                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2297                         memcpy(array, sample->user_regs.regs, sz);
2298                         array = (void *)array + sz;
2299                 } else {
2300                         *array++ = 0;
2301                 }
2302         }
2303
2304         if (type & PERF_SAMPLE_STACK_USER) {
2305                 sz = sample->user_stack.size;
2306                 *array++ = sz;
2307                 if (sz) {
2308                         memcpy(array, sample->user_stack.data, sz);
2309                         array = (void *)array + sz;
2310                         *array++ = sz;
2311                 }
2312         }
2313
2314         if (type & PERF_SAMPLE_WEIGHT) {
2315                 *array = sample->weight;
2316                 array++;
2317         }
2318
2319         if (type & PERF_SAMPLE_DATA_SRC) {
2320                 *array = sample->data_src;
2321                 array++;
2322         }
2323
2324         if (type & PERF_SAMPLE_TRANSACTION) {
2325                 *array = sample->transaction;
2326                 array++;
2327         }
2328
2329         if (type & PERF_SAMPLE_REGS_INTR) {
2330                 if (sample->intr_regs.abi) {
2331                         *array++ = sample->intr_regs.abi;
2332                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2333                         memcpy(array, sample->intr_regs.regs, sz);
2334                         array = (void *)array + sz;
2335                 } else {
2336                         *array++ = 0;
2337                 }
2338         }
2339
2340         return 0;
2341 }
2342
2343 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2344 {
2345         return pevent_find_field(evsel->tp_format, name);
2346 }
2347
2348 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2349                          const char *name)
2350 {
2351         struct format_field *field = perf_evsel__field(evsel, name);
2352         int offset;
2353
2354         if (!field)
2355                 return NULL;
2356
2357         offset = field->offset;
2358
2359         if (field->flags & FIELD_IS_DYNAMIC) {
2360                 offset = *(int *)(sample->raw_data + field->offset);
2361                 offset &= 0xffff;
2362         }
2363
2364         return sample->raw_data + offset;
2365 }
2366
2367 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2368                          bool needs_swap)
2369 {
2370         u64 value;
2371         void *ptr = sample->raw_data + field->offset;
2372
2373         switch (field->size) {
2374         case 1:
2375                 return *(u8 *)ptr;
2376         case 2:
2377                 value = *(u16 *)ptr;
2378                 break;
2379         case 4:
2380                 value = *(u32 *)ptr;
2381                 break;
2382         case 8:
2383                 memcpy(&value, ptr, sizeof(u64));
2384                 break;
2385         default:
2386                 return 0;
2387         }
2388
2389         if (!needs_swap)
2390                 return value;
2391
2392         switch (field->size) {
2393         case 2:
2394                 return bswap_16(value);
2395         case 4:
2396                 return bswap_32(value);
2397         case 8:
2398                 return bswap_64(value);
2399         default:
2400                 return 0;
2401         }
2402
2403         return 0;
2404 }
2405
2406 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2407                        const char *name)
2408 {
2409         struct format_field *field = perf_evsel__field(evsel, name);
2410
2411         if (!field)
2412                 return 0;
2413
2414         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2415 }
2416
2417 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2418                           char *msg, size_t msgsize)
2419 {
2420         int paranoid;
2421
2422         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2423             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2424             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2425                 /*
2426                  * If it's cycles then fall back to hrtimer based
2427                  * cpu-clock-tick sw counter, which is always available even if
2428                  * no PMU support.
2429                  *
2430                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2431                  * b0a873e).
2432                  */
2433                 scnprintf(msg, msgsize, "%s",
2434 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2435
2436                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2437                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2438
2439                 zfree(&evsel->name);
2440                 return true;
2441         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2442                    (paranoid = perf_event_paranoid()) > 1) {
2443                 const char *name = perf_evsel__name(evsel);
2444                 char *new_name;
2445
2446                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2447                         return false;
2448
2449                 if (evsel->name)
2450                         free(evsel->name);
2451                 evsel->name = new_name;
2452                 scnprintf(msg, msgsize,
2453 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2454                 evsel->attr.exclude_kernel = 1;
2455
2456                 return true;
2457         }
2458
2459         return false;
2460 }
2461
2462 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2463                               int err, char *msg, size_t size)
2464 {
2465         char sbuf[STRERR_BUFSIZE];
2466         int printed = 0;
2467
2468         switch (err) {
2469         case EPERM:
2470         case EACCES:
2471                 if (err == EPERM)
2472                         printed = scnprintf(msg, size,
2473                                 "No permission to enable %s event.\n\n",
2474                                 perf_evsel__name(evsel));
2475
2476                 return scnprintf(msg + printed, size - printed,
2477                  "You may not have permission to collect %sstats.\n\n"
2478                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2479                  "which controls use of the performance events system by\n"
2480                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2481                  "The current value is %d:\n\n"
2482                  "  -1: Allow use of (almost) all events by all users\n"
2483                  ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2484                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2485                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2486                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2487                  "      kernel.perf_event_paranoid = -1\n" ,
2488                                  target->system_wide ? "system-wide " : "",
2489                                  perf_event_paranoid());
2490         case ENOENT:
2491                 return scnprintf(msg, size, "The %s event is not supported.",
2492                                  perf_evsel__name(evsel));
2493         case EMFILE:
2494                 return scnprintf(msg, size, "%s",
2495                          "Too many events are opened.\n"
2496                          "Probably the maximum number of open file descriptors has been reached.\n"
2497                          "Hint: Try again after reducing the number of events.\n"
2498                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2499         case ENOMEM:
2500                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2501                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2502                         return scnprintf(msg, size,
2503                                          "Not enough memory to setup event with callchain.\n"
2504                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2505                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2506                 break;
2507         case ENODEV:
2508                 if (target->cpu_list)
2509                         return scnprintf(msg, size, "%s",
2510          "No such device - did you specify an out-of-range profile CPU?");
2511                 break;
2512         case EOPNOTSUPP:
2513                 if (evsel->attr.sample_period != 0)
2514                         return scnprintf(msg, size, "%s",
2515         "PMU Hardware doesn't support sampling/overflow-interrupts.");
2516                 if (evsel->attr.precise_ip)
2517                         return scnprintf(msg, size, "%s",
2518         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2519 #if defined(__i386__) || defined(__x86_64__)
2520                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2521                         return scnprintf(msg, size, "%s",
2522         "No hardware sampling interrupt available.\n"
2523         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2524 #endif
2525                 break;
2526         case EBUSY:
2527                 if (find_process("oprofiled"))
2528                         return scnprintf(msg, size,
2529         "The PMU counters are busy/taken by another profiler.\n"
2530         "We found oprofile daemon running, please stop it and try again.");
2531                 break;
2532         case EINVAL:
2533                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2534                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2535                 if (perf_missing_features.clockid)
2536                         return scnprintf(msg, size, "clockid feature not supported.");
2537                 if (perf_missing_features.clockid_wrong)
2538                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2539                 break;
2540         default:
2541                 break;
2542         }
2543
2544         return scnprintf(msg, size,
2545         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2546         "/bin/dmesg may provide additional information.\n"
2547         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2548                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2549                          perf_evsel__name(evsel));
2550 }
2551
2552 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2553 {
2554         if (evsel && evsel->evlist && evsel->evlist->env)
2555                 return evsel->evlist->env->arch;
2556         return NULL;
2557 }