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