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