tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->evlist      = NULL;
 236         evsel->bpf_fd      = -1;
 237         INIT_LIST_HEAD(&evsel->node);
 238         INIT_LIST_HEAD(&evsel->config_terms);
 239         perf_evsel__object.init(evsel);
 240         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 241         perf_evsel__calc_id_pos(evsel);
 242         evsel->cmdline_group_boundary = false;
 243         evsel->metric_expr   = NULL;
 244         evsel->metric_name   = NULL;
 245         evsel->metric_events = NULL;
 246         evsel->collect_stat  = false;
 247 }
 248
 249 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 250 {
 251         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 252
 253         if (evsel != NULL)
 254                 perf_evsel__init(evsel, attr, idx);
 255
 256         if (perf_evsel__is_bpf_output(evsel)) {
 257                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 258                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 259                 evsel->attr.sample_period = 1;
 260         }
 261
 262         return evsel;
 263 }
 264
 265 static bool perf_event_can_profile_kernel(void)
 266 {
 267         return geteuid() == 0 || perf_event_paranoid() == -1;
 268 }
 269
 270 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 271 {
 272         struct perf_event_attr attr = {
 273                 .type   = PERF_TYPE_HARDWARE,
 274                 .config = PERF_COUNT_HW_CPU_CYCLES,
 275                 .exclude_kernel = !perf_event_can_profile_kernel(),
 276         };
 277         struct perf_evsel *evsel;
 278
 279         event_attr_init(&attr);
 280
 281         if (!precise)
 282                 goto new_event;
 283         /*
 284          * Unnamed union member, not supported as struct member named
 285          * initializer in older compilers such as gcc 4.4.7
 286          *
 287          * Just for probing the precise_ip:
 288          */
 289         attr.sample_period = 1;
 290
 291         perf_event_attr__set_max_precise_ip(&attr);
 292         /*
 293          * Now let the usual logic to set up the perf_event_attr defaults
 294          * to kick in when we return and before perf_evsel__open() is called.
 295          */
 296         attr.sample_period = 0;
 297 new_event:
 298         evsel = perf_evsel__new(&attr);
 299         if (evsel == NULL)
 300                 goto out;
 301
 302         /* use asprintf() because free(evsel) assumes name is allocated */
 303         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 304                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 305                      attr.exclude_kernel ? "u" : "",
 306                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 307                 goto error_free;
 308 out:
 309         return evsel;
 310 error_free:
 311         perf_evsel__delete(evsel);
 312         evsel = NULL;
 313         goto out;
 314 }
 315
 316 /*
 317  * Returns pointer with encoded error via <linux/err.h> interface.
 318  */
 319 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 320 {
 321         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 322         int err = -ENOMEM;
 323
 324         if (evsel == NULL) {
 325                 goto out_err;
 326         } else {
 327                 struct perf_event_attr attr = {
 328                         .type          = PERF_TYPE_TRACEPOINT,
 329                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 330                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 331                 };
 332
 333                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 334                         goto out_free;
 335
 336                 evsel->tp_format = trace_event__tp_format(sys, name);
 337                 if (IS_ERR(evsel->tp_format)) {
 338                         err = PTR_ERR(evsel->tp_format);
 339                         goto out_free;
 340                 }
 341
 342                 event_attr_init(&attr);
 343                 attr.config = evsel->tp_format->id;
 344                 attr.sample_period = 1;
 345                 perf_evsel__init(evsel, &attr, idx);
 346         }
 347
 348         return evsel;
 349
 350 out_free:
 351         zfree(&evsel->name);
 352         free(evsel);
 353 out_err:
 354         return ERR_PTR(err);
 355 }
 356
 357 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 358         "cycles",
 359         "instructions",
 360         "cache-references",
 361         "cache-misses",
 362         "branches",
 363         "branch-misses",
 364         "bus-cycles",
 365         "stalled-cycles-frontend",
 366         "stalled-cycles-backend",
 367         "ref-cycles",
 368 };
 369
 370 static const char *__perf_evsel__hw_name(u64 config)
 371 {
 372         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 373                 return perf_evsel__hw_names[config];
 374
 375         return "unknown-hardware";
 376 }
 377
 378 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 379 {
 380         int colon = 0, r = 0;
 381         struct perf_event_attr *attr = &evsel->attr;
 382         bool exclude_guest_default = false;
 383
 384 #define MOD_PRINT(context, mod) do {                                    \
 385                 if (!attr->exclude_##context) {                         \
 386                         if (!colon) colon = ++r;                        \
 387                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 388                 } } while(0)
 389
 390         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 391                 MOD_PRINT(kernel, 'k');
 392                 MOD_PRINT(user, 'u');
 393                 MOD_PRINT(hv, 'h');
 394                 exclude_guest_default = true;
 395         }
 396
 397         if (attr->precise_ip) {
 398                 if (!colon)
 399                         colon = ++r;
 400                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 401                 exclude_guest_default = true;
 402         }
 403
 404         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 405                 MOD_PRINT(host, 'H');
 406                 MOD_PRINT(guest, 'G');
 407         }
 408 #undef MOD_PRINT
 409         if (colon)
 410                 bf[colon - 1] = ':';
 411         return r;
 412 }
 413
 414 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 415 {
 416         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 417         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 418 }
 419
 420 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 421         "cpu-clock",
 422         "task-clock",
 423         "page-faults",
 424         "context-switches",
 425         "cpu-migrations",
 426         "minor-faults",
 427         "major-faults",
 428         "alignment-faults",
 429         "emulation-faults",
 430         "dummy",
 431 };
 432
 433 static const char *__perf_evsel__sw_name(u64 config)
 434 {
 435         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 436                 return perf_evsel__sw_names[config];
 437         return "unknown-software";
 438 }
 439
 440 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 441 {
 442         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 443         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 444 }
 445
 446 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 447 {
 448         int r;
 449
 450         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 451
 452         if (type & HW_BREAKPOINT_R)
 453                 r += scnprintf(bf + r, size - r, "r");
 454
 455         if (type & HW_BREAKPOINT_W)
 456                 r += scnprintf(bf + r, size - r, "w");
 457
 458         if (type & HW_BREAKPOINT_X)
 459                 r += scnprintf(bf + r, size - r, "x");
 460
 461         return r;
 462 }
 463
 464 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 465 {
 466         struct perf_event_attr *attr = &evsel->attr;
 467         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 468         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 469 }
 470
 471 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 472                                 [PERF_EVSEL__MAX_ALIASES] = {
 473  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 474  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 475  { "LLC",       "L2",                                                   },
 476  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 477  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 478  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 479  { "node",                                                              },
 480 };
 481
 482 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 483                                    [PERF_EVSEL__MAX_ALIASES] = {
 484  { "load",      "loads",        "read",                                 },
 485  { "store",     "stores",       "write",                                },
 486  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 487 };
 488
 489 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 490                                        [PERF_EVSEL__MAX_ALIASES] = {
 491  { "refs",      "Reference",    "ops",          "access",               },
 492  { "misses",    "miss",                                                 },
 493 };
 494
 495 #define C(x)            PERF_COUNT_HW_CACHE_##x
 496 #define CACHE_READ      (1 << C(OP_READ))
 497 #define CACHE_WRITE     (1 << C(OP_WRITE))
 498 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 499 #define COP(x)          (1 << x)
 500
 501 /*
 502  * cache operartion stat
 503  * L1I : Read and prefetch only
 504  * ITLB and BPU : Read-only
 505  */
 506 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 507  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 508  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 509  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 510  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 511  [C(ITLB)]      = (CACHE_READ),
 512  [C(BPU)]       = (CACHE_READ),
 513  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 514 };
 515
 516 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 517 {
 518         if (perf_evsel__hw_cache_stat[type] & COP(op))
 519                 return true;    /* valid */
 520         else
 521                 return false;   /* invalid */
 522 }
 523
 524 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 525                                             char *bf, size_t size)
 526 {
 527         if (result) {
 528                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 529                                  perf_evsel__hw_cache_op[op][0],
 530                                  perf_evsel__hw_cache_result[result][0]);
 531         }
 532
 533         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 534                          perf_evsel__hw_cache_op[op][1]);
 535 }
 536
 537 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 538 {
 539         u8 op, result, type = (config >>  0) & 0xff;
 540         const char *err = "unknown-ext-hardware-cache-type";
 541
 542         if (type >= PERF_COUNT_HW_CACHE_MAX)
 543                 goto out_err;
 544
 545         op = (config >>  8) & 0xff;
 546         err = "unknown-ext-hardware-cache-op";
 547         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 548                 goto out_err;
 549
 550         result = (config >> 16) & 0xff;
 551         err = "unknown-ext-hardware-cache-result";
 552         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 553                 goto out_err;
 554
 555         err = "invalid-cache";
 556         if (!perf_evsel__is_cache_op_valid(type, op))
 557                 goto out_err;
 558
 559         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 560 out_err:
 561         return scnprintf(bf, size, "%s", err);
 562 }
 563
 564 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 565 {
 566         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 567         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 568 }
 569
 570 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 571 {
 572         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 573         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 574 }
 575
 576 const char *perf_evsel__name(struct perf_evsel *evsel)
 577 {
 578         char bf[128];
 579
 580         if (evsel->name)
 581                 return evsel->name;
 582
 583         switch (evsel->attr.type) {
 584         case PERF_TYPE_RAW:
 585                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 586                 break;
 587
 588         case PERF_TYPE_HARDWARE:
 589                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 590                 break;
 591
 592         case PERF_TYPE_HW_CACHE:
 593                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 594                 break;
 595
 596         case PERF_TYPE_SOFTWARE:
 597                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 598                 break;
 599
 600         case PERF_TYPE_TRACEPOINT:
 601                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 602                 break;
 603
 604         case PERF_TYPE_BREAKPOINT:
 605                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 606                 break;
 607
 608         default:
 609                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 610                           evsel->attr.type);
 611                 break;
 612         }
 613
 614         evsel->name = strdup(bf);
 615
 616         return evsel->name ?: "unknown";
 617 }
 618
 619 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 620 {
 621         return evsel->group_name ?: "anon group";
 622 }
 623
 624 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 625 {
 626         int ret;
 627         struct perf_evsel *pos;
 628         const char *group_name = perf_evsel__group_name(evsel);
 629
 630         ret = scnprintf(buf, size, "%s", group_name);
 631
 632         ret += scnprintf(buf + ret, size - ret, " { %s",
 633                          perf_evsel__name(evsel));
 634
 635         for_each_group_member(pos, evsel)
 636                 ret += scnprintf(buf + ret, size - ret, ", %s",
 637                                  perf_evsel__name(pos));
 638
 639         ret += scnprintf(buf + ret, size - ret, " }");
 640
 641         return ret;
 642 }
 643
 644 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 645                                            struct record_opts *opts,
 646                                            struct callchain_param *param)
 647 {
 648         bool function = perf_evsel__is_function_event(evsel);
 649         struct perf_event_attr *attr = &evsel->attr;
 650
 651         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 652
 653         attr->sample_max_stack = param->max_stack;
 654
 655         if (param->record_mode == CALLCHAIN_LBR) {
 656                 if (!opts->branch_stack) {
 657                         if (attr->exclude_user) {
 658                                 pr_warning("LBR callstack option is only available "
 659                                            "to get user callchain information. "
 660                                            "Falling back to framepointers.\n");
 661                         } else {
 662                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 663                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 664                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 665                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 666                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 667                         }
 668                 } else
 669                          pr_warning("Cannot use LBR callstack with branch stack. "
 670                                     "Falling back to framepointers.\n");
 671         }
 672
 673         if (param->record_mode == CALLCHAIN_DWARF) {
 674                 if (!function) {
 675                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 676                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 677                         attr->sample_regs_user |= PERF_REGS_MASK;
 678                         attr->sample_stack_user = param->dump_size;
 679                         attr->exclude_callchain_user = 1;
 680                 } else {
 681                         pr_info("Cannot use DWARF unwind for function trace event,"
 682                                 " falling back to framepointers.\n");
 683                 }
 684         }
 685
 686         if (function) {
 687                 pr_info("Disabling user space callchains for function trace event.\n");
 688                 attr->exclude_callchain_user = 1;
 689         }
 690 }
 691
 692 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 693                                   struct record_opts *opts,
 694                                   struct callchain_param *param)
 695 {
 696         if (param->enabled)
 697                 return __perf_evsel__config_callchain(evsel, opts, param);
 698 }
 699
 700 static void
 701 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 702                             struct callchain_param *param)
 703 {
 704         struct perf_event_attr *attr = &evsel->attr;
 705
 706         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 707         if (param->record_mode == CALLCHAIN_LBR) {
 708                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 709                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 710                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 711         }
 712         if (param->record_mode == CALLCHAIN_DWARF) {
 713                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 714                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 715         }
 716 }
 717
 718 static void apply_config_terms(struct perf_evsel *evsel,
 719                                struct record_opts *opts, bool track)
 720 {
 721         struct perf_evsel_config_term *term;
 722         struct list_head *config_terms = &evsel->config_terms;
 723         struct perf_event_attr *attr = &evsel->attr;
 724         /* callgraph default */
 725         struct callchain_param param = {
 726                 .record_mode = callchain_param.record_mode,
 727         };
 728         u32 dump_size = 0;
 729         int max_stack = 0;
 730         const char *callgraph_buf = NULL;
 731
 732         list_for_each_entry(term, config_terms, list) {
 733                 switch (term->type) {
 734                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 735                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 736                                 attr->sample_period = term->val.period;
 737                                 attr->freq = 0;
 738                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 739                         }
 740                         break;
 741                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 742                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 743                                 attr->sample_freq = term->val.freq;
 744                                 attr->freq = 1;
 745                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 746                         }
 747                         break;
 748                 case PERF_EVSEL__CONFIG_TERM_TIME:
 749                         if (term->val.time)
 750                                 perf_evsel__set_sample_bit(evsel, TIME);
 751                         else
 752                                 perf_evsel__reset_sample_bit(evsel, TIME);
 753                         break;
 754                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 755                         callgraph_buf = term->val.callgraph;
 756                         break;
 757                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 758                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 759                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 760                                 parse_branch_str(term->val.branch,
 761                                                  &attr->branch_sample_type);
 762                         } else
 763                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 764                         break;
 765                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 766                         dump_size = term->val.stack_user;
 767                         break;
 768                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 769                         max_stack = term->val.max_stack;
 770                         break;
 771                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 772                         /*
 773                          * attr->inherit should has already been set by
 774                          * perf_evsel__config. If user explicitly set
 775                          * inherit using config terms, override global
 776                          * opt->no_inherit setting.
 777                          */
 778                         attr->inherit = term->val.inherit ? 1 : 0;
 779                         break;
 780                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 781                         attr->write_backward = term->val.overwrite ? 1 : 0;
 782                         break;
 783                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 784                         break;
 785                 default:
 786                         break;
 787                 }
 788         }
 789
 790         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 791         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 792                 bool sample_address = false;
 793
 794                 if (max_stack) {
 795                         param.max_stack = max_stack;
 796                         if (callgraph_buf == NULL)
 797                                 callgraph_buf = "fp";
 798                 }
 799
 800                 /* parse callgraph parameters */
 801                 if (callgraph_buf != NULL) {
 802                         if (!strcmp(callgraph_buf, "no")) {
 803                                 param.enabled = false;
 804                                 param.record_mode = CALLCHAIN_NONE;
 805                         } else {
 806                                 param.enabled = true;
 807                                 if (parse_callchain_record(callgraph_buf, &param)) {
 808                                         pr_err("per-event callgraph setting for %s failed. "
 809                                                "Apply callgraph global setting for it\n",
 810                                                evsel->name);
 811                                         return;
 812                                 }
 813                                 if (param.record_mode == CALLCHAIN_DWARF)
 814                                         sample_address = true;
 815                         }
 816                 }
 817                 if (dump_size > 0) {
 818                         dump_size = round_up(dump_size, sizeof(u64));
 819                         param.dump_size = dump_size;
 820                 }
 821
 822                 /* If global callgraph set, clear it */
 823                 if (callchain_param.enabled)
 824                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 825
 826                 /* set perf-event callgraph */
 827                 if (param.enabled) {
 828                         if (sample_address) {
 829                                 perf_evsel__set_sample_bit(evsel, ADDR);
 830                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 831                                 evsel->attr.mmap_data = track;
 832                         }
 833                         perf_evsel__config_callchain(evsel, opts, &param);
 834                 }
 835         }
 836 }
 837
 838 /*
 839  * The enable_on_exec/disabled value strategy:
 840  *
 841  *  1) For any type of traced program:
 842  *    - all independent events and group leaders are disabled
 843  *    - all group members are enabled
 844  *
 845  *     Group members are ruled by group leaders. They need to
 846  *     be enabled, because the group scheduling relies on that.
 847  *
 848  *  2) For traced programs executed by perf:
 849  *     - all independent events and group leaders have
 850  *       enable_on_exec set
 851  *     - we don't specifically enable or disable any event during
 852  *       the record command
 853  *
 854  *     Independent events and group leaders are initially disabled
 855  *     and get enabled by exec. Group members are ruled by group
 856  *     leaders as stated in 1).
 857  *
 858  *  3) For traced programs attached by perf (pid/tid):
 859  *     - we specifically enable or disable all events during
 860  *       the record command
 861  *
 862  *     When attaching events to already running traced we
 863  *     enable/disable events specifically, as there's no
 864  *     initial traced exec call.
 865  */
 866 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 867                         struct callchain_param *callchain)
 868 {
 869         struct perf_evsel *leader = evsel->leader;
 870         struct perf_event_attr *attr = &evsel->attr;
 871         int track = evsel->tracking;
 872         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 873
 874         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 875         attr->inherit       = !opts->no_inherit;
 876         attr->write_backward = opts->overwrite ? 1 : 0;
 877
 878         perf_evsel__set_sample_bit(evsel, IP);
 879         perf_evsel__set_sample_bit(evsel, TID);
 880
 881         if (evsel->sample_read) {
 882                 perf_evsel__set_sample_bit(evsel, READ);
 883
 884                 /*
 885                  * We need ID even in case of single event, because
 886                  * PERF_SAMPLE_READ process ID specific data.
 887                  */
 888                 perf_evsel__set_sample_id(evsel, false);
 889
 890                 /*
 891                  * Apply group format only if we belong to group
 892                  * with more than one members.
 893                  */
 894                 if (leader->nr_members > 1) {
 895                         attr->read_format |= PERF_FORMAT_GROUP;
 896                         attr->inherit = 0;
 897                 }
 898         }
 899
 900         /*
 901          * We default some events to have a default interval. But keep
 902          * it a weak assumption overridable by the user.
 903          */
 904         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 905                                      opts->user_interval != ULLONG_MAX)) {
 906                 if (opts->freq) {
 907                         perf_evsel__set_sample_bit(evsel, PERIOD);
 908                         attr->freq              = 1;
 909                         attr->sample_freq       = opts->freq;
 910                 } else {
 911                         attr->sample_period = opts->default_interval;
 912                 }
 913         }
 914
 915         /*
 916          * Disable sampling for all group members other
 917          * than leader in case leader 'leads' the sampling.
 918          */
 919         if ((leader != evsel) && leader->sample_read) {
 920                 attr->sample_freq   = 0;
 921                 attr->sample_period = 0;
 922         }
 923
 924         if (opts->no_samples)
 925                 attr->sample_freq = 0;
 926
 927         if (opts->inherit_stat) {
 928                 evsel->attr.read_format |=
 929                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 930                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 931                         PERF_FORMAT_ID;
 932                 attr->inherit_stat = 1;
 933         }
 934
 935         if (opts->sample_address) {
 936                 perf_evsel__set_sample_bit(evsel, ADDR);
 937                 attr->mmap_data = track;
 938         }
 939
 940         /*
 941          * We don't allow user space callchains for  function trace
 942          * event, due to issues with page faults while tracing page
 943          * fault handler and its overall trickiness nature.
 944          */
 945         if (perf_evsel__is_function_event(evsel))
 946                 evsel->attr.exclude_callchain_user = 1;
 947
 948         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 949                 perf_evsel__config_callchain(evsel, opts, callchain);
 950
 951         if (opts->sample_intr_regs) {
 952                 attr->sample_regs_intr = opts->sample_intr_regs;
 953                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 954         }
 955
 956         if (opts->sample_user_regs) {
 957                 attr->sample_regs_user |= opts->sample_user_regs;
 958                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 959         }
 960
 961         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 962                 perf_evsel__set_sample_bit(evsel, CPU);
 963
 964         /*
 965          * When the user explicitly disabled time don't force it here.
 966          */
 967         if (opts->sample_time &&
 968             (!perf_missing_features.sample_id_all &&
 969             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 970              opts->sample_time_set)))
 971                 perf_evsel__set_sample_bit(evsel, TIME);
 972
 973         if (opts->raw_samples && !evsel->no_aux_samples) {
 974                 perf_evsel__set_sample_bit(evsel, TIME);
 975                 perf_evsel__set_sample_bit(evsel, RAW);
 976                 perf_evsel__set_sample_bit(evsel, CPU);
 977         }
 978
 979         if (opts->sample_address)
 980                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 981
 982         if (opts->sample_phys_addr)
 983                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
 984
 985         if (opts->no_buffering) {
 986                 attr->watermark = 0;
 987                 attr->wakeup_events = 1;
 988         }
 989         if (opts->branch_stack && !evsel->no_aux_samples) {
 990                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 991                 attr->branch_sample_type = opts->branch_stack;
 992         }
 993
 994         if (opts->sample_weight)
 995                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 996
 997         attr->task  = track;
 998         attr->mmap  = track;
 999         attr->mmap2 = track && !perf_missing_features.mmap2;
1000         attr->comm  = track;
1001
1002         if (opts->record_namespaces)
1003                 attr->namespaces  = track;
1004
1005         if (opts->record_switch_events)
1006                 attr->context_switch = track;
1007
1008         if (opts->sample_transaction)
1009                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1010
1011         if (opts->running_time) {
1012                 evsel->attr.read_format |=
1013                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1014                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1015         }
1016
1017         /*
1018          * XXX see the function comment above
1019          *
1020          * Disabling only independent events or group leaders,
1021          * keeping group members enabled.
1022          */
1023         if (perf_evsel__is_group_leader(evsel))
1024                 attr->disabled = 1;
1025
1026         /*
1027          * Setting enable_on_exec for independent events and
1028          * group leaders for traced executed by perf.
1029          */
1030         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1031                 !opts->initial_delay)
1032                 attr->enable_on_exec = 1;
1033
1034         if (evsel->immediate) {
1035                 attr->disabled = 0;
1036                 attr->enable_on_exec = 0;
1037         }
1038
1039         clockid = opts->clockid;
1040         if (opts->use_clockid) {
1041                 attr->use_clockid = 1;
1042                 attr->clockid = opts->clockid;
1043         }
1044
1045         if (evsel->precise_max)
1046                 perf_event_attr__set_max_precise_ip(attr);
1047
1048         if (opts->all_user) {
1049                 attr->exclude_kernel = 1;
1050                 attr->exclude_user   = 0;
1051         }
1052
1053         if (opts->all_kernel) {
1054                 attr->exclude_kernel = 0;
1055                 attr->exclude_user   = 1;
1056         }
1057
1058         /*
1059          * Apply event specific term settings,
1060          * it overloads any global configuration.
1061          */
1062         apply_config_terms(evsel, opts, track);
1063
1064         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1065
1066         /* The --period option takes the precedence. */
1067         if (opts->period_set) {
1068                 if (opts->period)
1069                         perf_evsel__set_sample_bit(evsel, PERIOD);
1070                 else
1071                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1072         }
1073 }
1074
1075 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1076 {
1077         if (evsel->system_wide)
1078                 nthreads = 1;
1079
1080         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1081
1082         if (evsel->fd) {
1083                 int cpu, thread;
1084                 for (cpu = 0; cpu < ncpus; cpu++) {
1085                         for (thread = 0; thread < nthreads; thread++) {
1086                                 FD(evsel, cpu, thread) = -1;
1087                         }
1088                 }
1089         }
1090
1091         return evsel->fd != NULL ? 0 : -ENOMEM;
1092 }
1093
1094 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1095                           int ioc,  void *arg)
1096 {
1097         int cpu, thread;
1098
1099         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1100                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1101                         int fd = FD(evsel, cpu, thread),
1102                             err = ioctl(fd, ioc, arg);
1103
1104                         if (err)
1105                                 return err;
1106                 }
1107         }
1108
1109         return 0;
1110 }
1111
1112 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1113 {
1114         return perf_evsel__run_ioctl(evsel,
1115                                      PERF_EVENT_IOC_SET_FILTER,
1116                                      (void *)filter);
1117 }
1118
1119 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1120 {
1121         char *new_filter = strdup(filter);
1122
1123         if (new_filter != NULL) {
1124                 free(evsel->filter);
1125                 evsel->filter = new_filter;
1126                 return 0;
1127         }
1128
1129         return -1;
1130 }
1131
1132 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1133                                      const char *fmt, const char *filter)
1134 {
1135         char *new_filter;
1136
1137         if (evsel->filter == NULL)
1138                 return perf_evsel__set_filter(evsel, filter);
1139
1140         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1141                 free(evsel->filter);
1142                 evsel->filter = new_filter;
1143                 return 0;
1144         }
1145
1146         return -1;
1147 }
1148
1149 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1150 {
1151         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1152 }
1153
1154 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1155 {
1156         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1157 }
1158
1159 int perf_evsel__enable(struct perf_evsel *evsel)
1160 {
1161         return perf_evsel__run_ioctl(evsel,
1162                                      PERF_EVENT_IOC_ENABLE,
1163                                      0);
1164 }
1165
1166 int perf_evsel__disable(struct perf_evsel *evsel)
1167 {
1168         return perf_evsel__run_ioctl(evsel,
1169                                      PERF_EVENT_IOC_DISABLE,
1170                                      0);
1171 }
1172
1173 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1174 {
1175         if (ncpus == 0 || nthreads == 0)
1176                 return 0;
1177
1178         if (evsel->system_wide)
1179                 nthreads = 1;
1180
1181         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1182         if (evsel->sample_id == NULL)
1183                 return -ENOMEM;
1184
1185         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1186         if (evsel->id == NULL) {
1187                 xyarray__delete(evsel->sample_id);
1188                 evsel->sample_id = NULL;
1189                 return -ENOMEM;
1190         }
1191
1192         return 0;
1193 }
1194
1195 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1196 {
1197         xyarray__delete(evsel->fd);
1198         evsel->fd = NULL;
1199 }
1200
1201 static void perf_evsel__free_id(struct perf_evsel *evsel)
1202 {
1203         xyarray__delete(evsel->sample_id);
1204         evsel->sample_id = NULL;
1205         zfree(&evsel->id);
1206 }
1207
1208 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1209 {
1210         struct perf_evsel_config_term *term, *h;
1211
1212         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1213                 list_del(&term->list);
1214                 free(term);
1215         }
1216 }
1217
1218 void perf_evsel__close_fd(struct perf_evsel *evsel)
1219 {
1220         int cpu, thread;
1221
1222         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1223                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1224                         close(FD(evsel, cpu, thread));
1225                         FD(evsel, cpu, thread) = -1;
1226                 }
1227 }
1228
1229 void perf_evsel__exit(struct perf_evsel *evsel)
1230 {
1231         assert(list_empty(&evsel->node));
1232         assert(evsel->evlist == NULL);
1233         perf_evsel__free_fd(evsel);
1234         perf_evsel__free_id(evsel);
1235         perf_evsel__free_config_terms(evsel);
1236         close_cgroup(evsel->cgrp);
1237         cpu_map__put(evsel->cpus);
1238         cpu_map__put(evsel->own_cpus);
1239         thread_map__put(evsel->threads);
1240         zfree(&evsel->group_name);
1241         zfree(&evsel->name);
1242         perf_evsel__object.fini(evsel);
1243 }
1244
1245 void perf_evsel__delete(struct perf_evsel *evsel)
1246 {
1247         perf_evsel__exit(evsel);
1248         free(evsel);
1249 }
1250
1251 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1252                                 struct perf_counts_values *count)
1253 {
1254         struct perf_counts_values tmp;
1255
1256         if (!evsel->prev_raw_counts)
1257                 return;
1258
1259         if (cpu == -1) {
1260                 tmp = evsel->prev_raw_counts->aggr;
1261                 evsel->prev_raw_counts->aggr = *count;
1262         } else {
1263                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1264                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1265         }
1266
1267         count->val = count->val - tmp.val;
1268         count->ena = count->ena - tmp.ena;
1269         count->run = count->run - tmp.run;
1270 }
1271
1272 void perf_counts_values__scale(struct perf_counts_values *count,
1273                                bool scale, s8 *pscaled)
1274 {
1275         s8 scaled = 0;
1276
1277         if (scale) {
1278                 if (count->run == 0) {
1279                         scaled = -1;
1280                         count->val = 0;
1281                 } else if (count->run < count->ena) {
1282                         scaled = 1;
1283                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1284                 }
1285         } else
1286                 count->ena = count->run = 0;
1287
1288         if (pscaled)
1289                 *pscaled = scaled;
1290 }
1291
1292 static int perf_evsel__read_size(struct perf_evsel *evsel)
1293 {
1294         u64 read_format = evsel->attr.read_format;
1295         int entry = sizeof(u64); /* value */
1296         int size = 0;
1297         int nr = 1;
1298
1299         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1300                 size += sizeof(u64);
1301
1302         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1303                 size += sizeof(u64);
1304
1305         if (read_format & PERF_FORMAT_ID)
1306                 entry += sizeof(u64);
1307
1308         if (read_format & PERF_FORMAT_GROUP) {
1309                 nr = evsel->nr_members;
1310                 size += sizeof(u64);
1311         }
1312
1313         size += entry * nr;
1314         return size;
1315 }
1316
1317 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1318                      struct perf_counts_values *count)
1319 {
1320         size_t size = perf_evsel__read_size(evsel);
1321
1322         memset(count, 0, sizeof(*count));
1323
1324         if (FD(evsel, cpu, thread) < 0)
1325                 return -EINVAL;
1326
1327         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1328                 return -errno;
1329
1330         return 0;
1331 }
1332
1333 static int
1334 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1335 {
1336         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1337
1338         return perf_evsel__read(evsel, cpu, thread, count);
1339 }
1340
1341 static void
1342 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1343                       u64 val, u64 ena, u64 run)
1344 {
1345         struct perf_counts_values *count;
1346
1347         count = perf_counts(counter->counts, cpu, thread);
1348
1349         count->val    = val;
1350         count->ena    = ena;
1351         count->run    = run;
1352         count->loaded = true;
1353 }
1354
1355 static int
1356 perf_evsel__process_group_data(struct perf_evsel *leader,
1357                                int cpu, int thread, u64 *data)
1358 {
1359         u64 read_format = leader->attr.read_format;
1360         struct sample_read_value *v;
1361         u64 nr, ena = 0, run = 0, i;
1362
1363         nr = *data++;
1364
1365         if (nr != (u64) leader->nr_members)
1366                 return -EINVAL;
1367
1368         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1369                 ena = *data++;
1370
1371         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1372                 run = *data++;
1373
1374         v = (struct sample_read_value *) data;
1375
1376         perf_evsel__set_count(leader, cpu, thread,
1377                               v[0].value, ena, run);
1378
1379         for (i = 1; i < nr; i++) {
1380                 struct perf_evsel *counter;
1381
1382                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1383                 if (!counter)
1384                         return -EINVAL;
1385
1386                 perf_evsel__set_count(counter, cpu, thread,
1387                                       v[i].value, ena, run);
1388         }
1389
1390         return 0;
1391 }
1392
1393 static int
1394 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1395 {
1396         struct perf_stat_evsel *ps = leader->stats;
1397         u64 read_format = leader->attr.read_format;
1398         int size = perf_evsel__read_size(leader);
1399         u64 *data = ps->group_data;
1400
1401         if (!(read_format & PERF_FORMAT_ID))
1402                 return -EINVAL;
1403
1404         if (!perf_evsel__is_group_leader(leader))
1405                 return -EINVAL;
1406
1407         if (!data) {
1408                 data = zalloc(size);
1409                 if (!data)
1410                         return -ENOMEM;
1411
1412                 ps->group_data = data;
1413         }
1414
1415         if (FD(leader, cpu, thread) < 0)
1416                 return -EINVAL;
1417
1418         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1419                 return -errno;
1420
1421         return perf_evsel__process_group_data(leader, cpu, thread, data);
1422 }
1423
1424 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1425 {
1426         u64 read_format = evsel->attr.read_format;
1427
1428         if (read_format & PERF_FORMAT_GROUP)
1429                 return perf_evsel__read_group(evsel, cpu, thread);
1430         else
1431                 return perf_evsel__read_one(evsel, cpu, thread);
1432 }
1433
1434 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1435                               int cpu, int thread, bool scale)
1436 {
1437         struct perf_counts_values count;
1438         size_t nv = scale ? 3 : 1;
1439
1440         if (FD(evsel, cpu, thread) < 0)
1441                 return -EINVAL;
1442
1443         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1444                 return -ENOMEM;
1445
1446         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1447                 return -errno;
1448
1449         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1450         perf_counts_values__scale(&count, scale, NULL);
1451         *perf_counts(evsel->counts, cpu, thread) = count;
1452         return 0;
1453 }
1454
1455 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1456 {
1457         struct perf_evsel *leader = evsel->leader;
1458         int fd;
1459
1460         if (perf_evsel__is_group_leader(evsel))
1461                 return -1;
1462
1463         /*
1464          * Leader must be already processed/open,
1465          * if not it's a bug.
1466          */
1467         BUG_ON(!leader->fd);
1468
1469         fd = FD(leader, cpu, thread);
1470         BUG_ON(fd == -1);
1471
1472         return fd;
1473 }
1474
1475 struct bit_names {
1476         int bit;
1477         const char *name;
1478 };
1479
1480 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1481 {
1482         bool first_bit = true;
1483         int i = 0;
1484
1485         do {
1486                 if (value & bits[i].bit) {
1487                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1488                         first_bit = false;
1489                 }
1490         } while (bits[++i].name != NULL);
1491 }
1492
1493 static void __p_sample_type(char *buf, size_t size, u64 value)
1494 {
1495 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1496         struct bit_names bits[] = {
1497                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1498                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1499                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1500                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1501                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1502                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1503                 { .name = NULL, }
1504         };
1505 #undef bit_name
1506         __p_bits(buf, size, value, bits);
1507 }
1508
1509 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1510 {
1511 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1512         struct bit_names bits[] = {
1513                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1514                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1515                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1516                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1517                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1518                 { .name = NULL, }
1519         };
1520 #undef bit_name
1521         __p_bits(buf, size, value, bits);
1522 }
1523
1524 static void __p_read_format(char *buf, size_t size, u64 value)
1525 {
1526 #define bit_name(n) { PERF_FORMAT_##n, #n }
1527         struct bit_names bits[] = {
1528                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1529                 bit_name(ID), bit_name(GROUP),
1530                 { .name = NULL, }
1531         };
1532 #undef bit_name
1533         __p_bits(buf, size, value, bits);
1534 }
1535
1536 #define BUF_SIZE                1024
1537
1538 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1539 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1540 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1541 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1542 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1543 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1544
1545 #define PRINT_ATTRn(_n, _f, _p)                         \
1546 do {                                                    \
1547         if (attr->_f) {                                 \
1548                 _p(attr->_f);                           \
1549                 ret += attr__fprintf(fp, _n, buf, priv);\
1550         }                                               \
1551 } while (0)
1552
1553 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1554
1555 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1556                              attr__fprintf_f attr__fprintf, void *priv)
1557 {
1558         char buf[BUF_SIZE];
1559         int ret = 0;
1560
1561         PRINT_ATTRf(type, p_unsigned);
1562         PRINT_ATTRf(size, p_unsigned);
1563         PRINT_ATTRf(config, p_hex);
1564         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1565         PRINT_ATTRf(sample_type, p_sample_type);
1566         PRINT_ATTRf(read_format, p_read_format);
1567
1568         PRINT_ATTRf(disabled, p_unsigned);
1569         PRINT_ATTRf(inherit, p_unsigned);
1570         PRINT_ATTRf(pinned, p_unsigned);
1571         PRINT_ATTRf(exclusive, p_unsigned);
1572         PRINT_ATTRf(exclude_user, p_unsigned);
1573         PRINT_ATTRf(exclude_kernel, p_unsigned);
1574         PRINT_ATTRf(exclude_hv, p_unsigned);
1575         PRINT_ATTRf(exclude_idle, p_unsigned);
1576         PRINT_ATTRf(mmap, p_unsigned);
1577         PRINT_ATTRf(comm, p_unsigned);
1578         PRINT_ATTRf(freq, p_unsigned);
1579         PRINT_ATTRf(inherit_stat, p_unsigned);
1580         PRINT_ATTRf(enable_on_exec, p_unsigned);
1581         PRINT_ATTRf(task, p_unsigned);
1582         PRINT_ATTRf(watermark, p_unsigned);
1583         PRINT_ATTRf(precise_ip, p_unsigned);
1584         PRINT_ATTRf(mmap_data, p_unsigned);
1585         PRINT_ATTRf(sample_id_all, p_unsigned);
1586         PRINT_ATTRf(exclude_host, p_unsigned);
1587         PRINT_ATTRf(exclude_guest, p_unsigned);
1588         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1589         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1590         PRINT_ATTRf(mmap2, p_unsigned);
1591         PRINT_ATTRf(comm_exec, p_unsigned);
1592         PRINT_ATTRf(use_clockid, p_unsigned);
1593         PRINT_ATTRf(context_switch, p_unsigned);
1594         PRINT_ATTRf(write_backward, p_unsigned);
1595         PRINT_ATTRf(namespaces, p_unsigned);
1596
1597         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1598         PRINT_ATTRf(bp_type, p_unsigned);
1599         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1600         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1601         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1602         PRINT_ATTRf(sample_regs_user, p_hex);
1603         PRINT_ATTRf(sample_stack_user, p_unsigned);
1604         PRINT_ATTRf(clockid, p_signed);
1605         PRINT_ATTRf(sample_regs_intr, p_hex);
1606         PRINT_ATTRf(aux_watermark, p_unsigned);
1607         PRINT_ATTRf(sample_max_stack, p_unsigned);
1608
1609         return ret;
1610 }
1611
1612 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1613                                 void *priv __maybe_unused)
1614 {
1615         return fprintf(fp, "  %-32s %s\n", name, val);
1616 }
1617
1618 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1619                                   int nr_cpus, int nr_threads,
1620                                   int thread_idx)
1621 {
1622         for (int cpu = 0; cpu < nr_cpus; cpu++)
1623                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1624                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1625 }
1626
1627 static int update_fds(struct perf_evsel *evsel,
1628                       int nr_cpus, int cpu_idx,
1629                       int nr_threads, int thread_idx)
1630 {
1631         struct perf_evsel *pos;
1632
1633         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1634                 return -EINVAL;
1635
1636         evlist__for_each_entry(evsel->evlist, pos) {
1637                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1638
1639                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1640
1641                 /*
1642                  * Since fds for next evsel has not been created,
1643                  * there is no need to iterate whole event list.
1644                  */
1645                 if (pos == evsel)
1646                         break;
1647         }
1648         return 0;
1649 }
1650
1651 static bool ignore_missing_thread(struct perf_evsel *evsel,
1652                                   int nr_cpus, int cpu,
1653                                   struct thread_map *threads,
1654                                   int thread, int err)
1655 {
1656         pid_t ignore_pid = thread_map__pid(threads, thread);
1657
1658         if (!evsel->ignore_missing_thread)
1659                 return false;
1660
1661         /* The system wide setup does not work with threads. */
1662         if (evsel->system_wide)
1663                 return false;
1664
1665         /* The -ESRCH is perf event syscall errno for pid's not found. */
1666         if (err != -ESRCH)
1667                 return false;
1668
1669         /* If there's only one thread, let it fail. */
1670         if (threads->nr == 1)
1671                 return false;
1672
1673         /*
1674          * We should remove fd for missing_thread first
1675          * because thread_map__remove() will decrease threads->nr.
1676          */
1677         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1678                 return false;
1679
1680         if (thread_map__remove(threads, thread))
1681                 return false;
1682
1683         pr_warning("WARNING: Ignored open failure for pid %d\n",
1684                    ignore_pid);
1685         return true;
1686 }
1687
1688 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1689                      struct thread_map *threads)
1690 {
1691         int cpu, thread, nthreads;
1692         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1693         int pid = -1, err;
1694         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1695
1696         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1697                 return -EINVAL;
1698
1699         if (cpus == NULL) {
1700                 static struct cpu_map *empty_cpu_map;
1701
1702                 if (empty_cpu_map == NULL) {
1703                         empty_cpu_map = cpu_map__dummy_new();
1704                         if (empty_cpu_map == NULL)
1705                                 return -ENOMEM;
1706                 }
1707
1708                 cpus = empty_cpu_map;
1709         }
1710
1711         if (threads == NULL) {
1712                 static struct thread_map *empty_thread_map;
1713
1714                 if (empty_thread_map == NULL) {
1715                         empty_thread_map = thread_map__new_by_tid(-1);
1716                         if (empty_thread_map == NULL)
1717                                 return -ENOMEM;
1718                 }
1719
1720                 threads = empty_thread_map;
1721         }
1722
1723         if (evsel->system_wide)
1724                 nthreads = 1;
1725         else
1726                 nthreads = threads->nr;
1727
1728         if (evsel->fd == NULL &&
1729             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1730                 return -ENOMEM;
1731
1732         if (evsel->cgrp) {
1733                 flags |= PERF_FLAG_PID_CGROUP;
1734                 pid = evsel->cgrp->fd;
1735         }
1736
1737 fallback_missing_features:
1738         if (perf_missing_features.clockid_wrong)
1739                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1740         if (perf_missing_features.clockid) {
1741                 evsel->attr.use_clockid = 0;
1742                 evsel->attr.clockid = 0;
1743         }
1744         if (perf_missing_features.cloexec)
1745                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1746         if (perf_missing_features.mmap2)
1747                 evsel->attr.mmap2 = 0;
1748         if (perf_missing_features.exclude_guest)
1749                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1750         if (perf_missing_features.lbr_flags)
1751                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1752                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1753         if (perf_missing_features.group_read && evsel->attr.inherit)
1754                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1755 retry_sample_id:
1756         if (perf_missing_features.sample_id_all)
1757                 evsel->attr.sample_id_all = 0;
1758
1759         if (verbose >= 2) {
1760                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1761                 fprintf(stderr, "perf_event_attr:\n");
1762                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1763                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1764         }
1765
1766         for (cpu = 0; cpu < cpus->nr; cpu++) {
1767
1768                 for (thread = 0; thread < nthreads; thread++) {
1769                         int fd, group_fd;
1770
1771                         if (!evsel->cgrp && !evsel->system_wide)
1772                                 pid = thread_map__pid(threads, thread);
1773
1774                         group_fd = get_group_fd(evsel, cpu, thread);
1775 retry_open:
1776                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1777                                   pid, cpus->map[cpu], group_fd, flags);
1778
1779                         test_attr__ready();
1780
1781                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1782                                                  group_fd, flags);
1783
1784                         FD(evsel, cpu, thread) = fd;
1785
1786                         if (fd < 0) {
1787                                 err = -errno;
1788
1789                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1790                                         /*
1791                                          * We just removed 1 thread, so take a step
1792                                          * back on thread index and lower the upper
1793                                          * nthreads limit.
1794                                          */
1795                                         nthreads--;
1796                                         thread--;
1797
1798                                         /* ... and pretend like nothing have happened. */
1799                                         err = 0;
1800                                         continue;
1801                                 }
1802
1803                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1804                                           err);
1805                                 goto try_fallback;
1806                         }
1807
1808                         pr_debug2(" = %d\n", fd);
1809
1810                         if (evsel->bpf_fd >= 0) {
1811                                 int evt_fd = fd;
1812                                 int bpf_fd = evsel->bpf_fd;
1813
1814                                 err = ioctl(evt_fd,
1815                                             PERF_EVENT_IOC_SET_BPF,
1816                                             bpf_fd);
1817                                 if (err && errno != EEXIST) {
1818                                         pr_err("failed to attach bpf fd %d: %s\n",
1819                                                bpf_fd, strerror(errno));
1820                                         err = -EINVAL;
1821                                         goto out_close;
1822                                 }
1823                         }
1824
1825                         set_rlimit = NO_CHANGE;
1826
1827                         /*
1828                          * If we succeeded but had to kill clockid, fail and
1829                          * have perf_evsel__open_strerror() print us a nice
1830                          * error.
1831                          */
1832                         if (perf_missing_features.clockid ||
1833                             perf_missing_features.clockid_wrong) {
1834                                 err = -EINVAL;
1835                                 goto out_close;
1836                         }
1837                 }
1838         }
1839
1840         return 0;
1841
1842 try_fallback:
1843         /*
1844          * perf stat needs between 5 and 22 fds per CPU. When we run out
1845          * of them try to increase the limits.
1846          */
1847         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1848                 struct rlimit l;
1849                 int old_errno = errno;
1850
1851                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1852                         if (set_rlimit == NO_CHANGE)
1853                                 l.rlim_cur = l.rlim_max;
1854                         else {
1855                                 l.rlim_cur = l.rlim_max + 1000;
1856                                 l.rlim_max = l.rlim_cur;
1857                         }
1858                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1859                                 set_rlimit++;
1860                                 errno = old_errno;
1861                                 goto retry_open;
1862                         }
1863                 }
1864                 errno = old_errno;
1865         }
1866
1867         if (err != -EINVAL || cpu > 0 || thread > 0)
1868                 goto out_close;
1869
1870         /*
1871          * Must probe features in the order they were added to the
1872          * perf_event_attr interface.
1873          */
1874         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1875                 perf_missing_features.write_backward = true;
1876                 pr_debug2("switching off write_backward\n");
1877                 goto out_close;
1878         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1879                 perf_missing_features.clockid_wrong = true;
1880                 pr_debug2("switching off clockid\n");
1881                 goto fallback_missing_features;
1882         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1883                 perf_missing_features.clockid = true;
1884                 pr_debug2("switching off use_clockid\n");
1885                 goto fallback_missing_features;
1886         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1887                 perf_missing_features.cloexec = true;
1888                 pr_debug2("switching off cloexec flag\n");
1889                 goto fallback_missing_features;
1890         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1891                 perf_missing_features.mmap2 = true;
1892                 pr_debug2("switching off mmap2\n");
1893                 goto fallback_missing_features;
1894         } else if (!perf_missing_features.exclude_guest &&
1895                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1896                 perf_missing_features.exclude_guest = true;
1897                 pr_debug2("switching off exclude_guest, exclude_host\n");
1898                 goto fallback_missing_features;
1899         } else if (!perf_missing_features.sample_id_all) {
1900                 perf_missing_features.sample_id_all = true;
1901                 pr_debug2("switching off sample_id_all\n");
1902                 goto retry_sample_id;
1903         } else if (!perf_missing_features.lbr_flags &&
1904                         (evsel->attr.branch_sample_type &
1905                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1906                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1907                 perf_missing_features.lbr_flags = true;
1908                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1909                 goto fallback_missing_features;
1910         } else if (!perf_missing_features.group_read &&
1911                     evsel->attr.inherit &&
1912                    (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
1913                 perf_missing_features.group_read = true;
1914                 pr_debug2("switching off group read\n");
1915                 goto fallback_missing_features;
1916         }
1917 out_close:
1918         do {
1919                 while (--thread >= 0) {
1920                         close(FD(evsel, cpu, thread));
1921                         FD(evsel, cpu, thread) = -1;
1922                 }
1923                 thread = nthreads;
1924         } while (--cpu >= 0);
1925         return err;
1926 }
1927
1928 void perf_evsel__close(struct perf_evsel *evsel)
1929 {
1930         if (evsel->fd == NULL)
1931                 return;
1932
1933         perf_evsel__close_fd(evsel);
1934         perf_evsel__free_fd(evsel);
1935 }
1936
1937 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1938                              struct cpu_map *cpus)
1939 {
1940         return perf_evsel__open(evsel, cpus, NULL);
1941 }
1942
1943 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1944                                 struct thread_map *threads)
1945 {
1946         return perf_evsel__open(evsel, NULL, threads);
1947 }
1948
1949 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1950                                        const union perf_event *event,
1951                                        struct perf_sample *sample)
1952 {
1953         u64 type = evsel->attr.sample_type;
1954         const u64 *array = event->sample.array;
1955         bool swapped = evsel->needs_swap;
1956         union u64_swap u;
1957
1958         array += ((event->header.size -
1959                    sizeof(event->header)) / sizeof(u64)) - 1;
1960
1961         if (type & PERF_SAMPLE_IDENTIFIER) {
1962                 sample->id = *array;
1963                 array--;
1964         }
1965
1966         if (type & PERF_SAMPLE_CPU) {
1967                 u.val64 = *array;
1968                 if (swapped) {
1969                         /* undo swap of u64, then swap on individual u32s */
1970                         u.val64 = bswap_64(u.val64);
1971                         u.val32[0] = bswap_32(u.val32[0]);
1972                 }
1973
1974                 sample->cpu = u.val32[0];
1975                 array--;
1976         }
1977
1978         if (type & PERF_SAMPLE_STREAM_ID) {
1979                 sample->stream_id = *array;
1980                 array--;
1981         }
1982
1983         if (type & PERF_SAMPLE_ID) {
1984                 sample->id = *array;
1985                 array--;
1986         }
1987
1988         if (type & PERF_SAMPLE_TIME) {
1989                 sample->time = *array;
1990                 array--;
1991         }
1992
1993         if (type & PERF_SAMPLE_TID) {
1994                 u.val64 = *array;
1995                 if (swapped) {
1996                         /* undo swap of u64, then swap on individual u32s */
1997                         u.val64 = bswap_64(u.val64);
1998                         u.val32[0] = bswap_32(u.val32[0]);
1999                         u.val32[1] = bswap_32(u.val32[1]);
2000                 }
2001
2002                 sample->pid = u.val32[0];
2003                 sample->tid = u.val32[1];
2004                 array--;
2005         }
2006
2007         return 0;
2008 }
2009
2010 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2011                             u64 size)
2012 {
2013         return size > max_size || offset + size > endp;
2014 }
2015
2016 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2017         do {                                                            \
2018                 if (overflow(endp, (max_size), (offset), (size)))       \
2019                         return -EFAULT;                                 \
2020         } while (0)
2021
2022 #define OVERFLOW_CHECK_u64(offset) \
2023         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2024
2025 static int
2026 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2027 {
2028         /*
2029          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2030          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2031          * check the format does not go past the end of the event.
2032          */
2033         if (sample_size + sizeof(event->header) > event->header.size)
2034                 return -EFAULT;
2035
2036         return 0;
2037 }
2038
2039 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2040                              struct perf_sample *data)
2041 {
2042         u64 type = evsel->attr.sample_type;
2043         bool swapped = evsel->needs_swap;
2044         const u64 *array;
2045         u16 max_size = event->header.size;
2046         const void *endp = (void *)event + max_size;
2047         u64 sz;
2048
2049         /*
2050          * used for cross-endian analysis. See git commit 65014ab3
2051          * for why this goofiness is needed.
2052          */
2053         union u64_swap u;
2054
2055         memset(data, 0, sizeof(*data));
2056         data->cpu = data->pid = data->tid = -1;
2057         data->stream_id = data->id = data->time = -1ULL;
2058         data->period = evsel->attr.sample_period;
2059         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2060         data->misc    = event->header.misc;
2061         data->id = -1ULL;
2062         data->data_src = PERF_MEM_DATA_SRC_NONE;
2063
2064         if (event->header.type != PERF_RECORD_SAMPLE) {
2065                 if (!evsel->attr.sample_id_all)
2066                         return 0;
2067                 return perf_evsel__parse_id_sample(evsel, event, data);
2068         }
2069
2070         array = event->sample.array;
2071
2072         if (perf_event__check_size(event, evsel->sample_size))
2073                 return -EFAULT;
2074
2075         if (type & PERF_SAMPLE_IDENTIFIER) {
2076                 data->id = *array;
2077                 array++;
2078         }
2079
2080         if (type & PERF_SAMPLE_IP) {
2081                 data->ip = *array;
2082                 array++;
2083         }
2084
2085         if (type & PERF_SAMPLE_TID) {
2086                 u.val64 = *array;
2087                 if (swapped) {
2088                         /* undo swap of u64, then swap on individual u32s */
2089                         u.val64 = bswap_64(u.val64);
2090                         u.val32[0] = bswap_32(u.val32[0]);
2091                         u.val32[1] = bswap_32(u.val32[1]);
2092                 }
2093
2094                 data->pid = u.val32[0];
2095                 data->tid = u.val32[1];
2096                 array++;
2097         }
2098
2099         if (type & PERF_SAMPLE_TIME) {
2100                 data->time = *array;
2101                 array++;
2102         }
2103
2104         if (type & PERF_SAMPLE_ADDR) {
2105                 data->addr = *array;
2106                 array++;
2107         }
2108
2109         if (type & PERF_SAMPLE_ID) {
2110                 data->id = *array;
2111                 array++;
2112         }
2113
2114         if (type & PERF_SAMPLE_STREAM_ID) {
2115                 data->stream_id = *array;
2116                 array++;
2117         }
2118
2119         if (type & PERF_SAMPLE_CPU) {
2120
2121                 u.val64 = *array;
2122                 if (swapped) {
2123                         /* undo swap of u64, then swap on individual u32s */
2124                         u.val64 = bswap_64(u.val64);
2125                         u.val32[0] = bswap_32(u.val32[0]);
2126                 }
2127
2128                 data->cpu = u.val32[0];
2129                 array++;
2130         }
2131
2132         if (type & PERF_SAMPLE_PERIOD) {
2133                 data->period = *array;
2134                 array++;
2135         }
2136
2137         if (type & PERF_SAMPLE_READ) {
2138                 u64 read_format = evsel->attr.read_format;
2139
2140                 OVERFLOW_CHECK_u64(array);
2141                 if (read_format & PERF_FORMAT_GROUP)
2142                         data->read.group.nr = *array;
2143                 else
2144                         data->read.one.value = *array;
2145
2146                 array++;
2147
2148                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2149                         OVERFLOW_CHECK_u64(array);
2150                         data->read.time_enabled = *array;
2151                         array++;
2152                 }
2153
2154                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2155                         OVERFLOW_CHECK_u64(array);
2156                         data->read.time_running = *array;
2157                         array++;
2158                 }
2159
2160                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2161                 if (read_format & PERF_FORMAT_GROUP) {
2162                         const u64 max_group_nr = UINT64_MAX /
2163                                         sizeof(struct sample_read_value);
2164
2165                         if (data->read.group.nr > max_group_nr)
2166                                 return -EFAULT;
2167                         sz = data->read.group.nr *
2168                              sizeof(struct sample_read_value);
2169                         OVERFLOW_CHECK(array, sz, max_size);
2170                         data->read.group.values =
2171                                         (struct sample_read_value *)array;
2172                         array = (void *)array + sz;
2173                 } else {
2174                         OVERFLOW_CHECK_u64(array);
2175                         data->read.one.id = *array;
2176                         array++;
2177                 }
2178         }
2179
2180         if (type & PERF_SAMPLE_CALLCHAIN) {
2181                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2182
2183                 OVERFLOW_CHECK_u64(array);
2184                 data->callchain = (struct ip_callchain *)array++;
2185                 if (data->callchain->nr > max_callchain_nr)
2186                         return -EFAULT;
2187                 sz = data->callchain->nr * sizeof(u64);
2188                 OVERFLOW_CHECK(array, sz, max_size);
2189                 array = (void *)array + sz;
2190         }
2191
2192         if (type & PERF_SAMPLE_RAW) {
2193                 OVERFLOW_CHECK_u64(array);
2194                 u.val64 = *array;
2195
2196                 /*
2197                  * Undo swap of u64, then swap on individual u32s,
2198                  * get the size of the raw area and undo all of the
2199                  * swap. The pevent interface handles endianity by
2200                  * itself.
2201                  */
2202                 if (swapped) {
2203                         u.val64 = bswap_64(u.val64);
2204                         u.val32[0] = bswap_32(u.val32[0]);
2205                         u.val32[1] = bswap_32(u.val32[1]);
2206                 }
2207                 data->raw_size = u.val32[0];
2208
2209                 /*
2210                  * The raw data is aligned on 64bits including the
2211                  * u32 size, so it's safe to use mem_bswap_64.
2212                  */
2213                 if (swapped)
2214                         mem_bswap_64((void *) array, data->raw_size);
2215
2216                 array = (void *)array + sizeof(u32);
2217
2218                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2219                 data->raw_data = (void *)array;
2220                 array = (void *)array + data->raw_size;
2221         }
2222
2223         if (type & PERF_SAMPLE_BRANCH_STACK) {
2224                 const u64 max_branch_nr = UINT64_MAX /
2225                                           sizeof(struct branch_entry);
2226
2227                 OVERFLOW_CHECK_u64(array);
2228                 data->branch_stack = (struct branch_stack *)array++;
2229
2230                 if (data->branch_stack->nr > max_branch_nr)
2231                         return -EFAULT;
2232                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2233                 OVERFLOW_CHECK(array, sz, max_size);
2234                 array = (void *)array + sz;
2235         }
2236
2237         if (type & PERF_SAMPLE_REGS_USER) {
2238                 OVERFLOW_CHECK_u64(array);
2239                 data->user_regs.abi = *array;
2240                 array++;
2241
2242                 if (data->user_regs.abi) {
2243                         u64 mask = evsel->attr.sample_regs_user;
2244
2245                         sz = hweight_long(mask) * sizeof(u64);
2246                         OVERFLOW_CHECK(array, sz, max_size);
2247                         data->user_regs.mask = mask;
2248                         data->user_regs.regs = (u64 *)array;
2249                         array = (void *)array + sz;
2250                 }
2251         }
2252
2253         if (type & PERF_SAMPLE_STACK_USER) {
2254                 OVERFLOW_CHECK_u64(array);
2255                 sz = *array++;
2256
2257                 data->user_stack.offset = ((char *)(array - 1)
2258                                           - (char *) event);
2259
2260                 if (!sz) {
2261                         data->user_stack.size = 0;
2262                 } else {
2263                         OVERFLOW_CHECK(array, sz, max_size);
2264                         data->user_stack.data = (char *)array;
2265                         array = (void *)array + sz;
2266                         OVERFLOW_CHECK_u64(array);
2267                         data->user_stack.size = *array++;
2268                         if (WARN_ONCE(data->user_stack.size > sz,
2269                                       "user stack dump failure\n"))
2270                                 return -EFAULT;
2271                 }
2272         }
2273
2274         if (type & PERF_SAMPLE_WEIGHT) {
2275                 OVERFLOW_CHECK_u64(array);
2276                 data->weight = *array;
2277                 array++;
2278         }
2279
2280         if (type & PERF_SAMPLE_DATA_SRC) {
2281                 OVERFLOW_CHECK_u64(array);
2282                 data->data_src = *array;
2283                 array++;
2284         }
2285
2286         if (type & PERF_SAMPLE_TRANSACTION) {
2287                 OVERFLOW_CHECK_u64(array);
2288                 data->transaction = *array;
2289                 array++;
2290         }
2291
2292         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2293         if (type & PERF_SAMPLE_REGS_INTR) {
2294                 OVERFLOW_CHECK_u64(array);
2295                 data->intr_regs.abi = *array;
2296                 array++;
2297
2298                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2299                         u64 mask = evsel->attr.sample_regs_intr;
2300
2301                         sz = hweight_long(mask) * sizeof(u64);
2302                         OVERFLOW_CHECK(array, sz, max_size);
2303                         data->intr_regs.mask = mask;
2304                         data->intr_regs.regs = (u64 *)array;
2305                         array = (void *)array + sz;
2306                 }
2307         }
2308
2309         data->phys_addr = 0;
2310         if (type & PERF_SAMPLE_PHYS_ADDR) {
2311                 data->phys_addr = *array;
2312                 array++;
2313         }
2314
2315         return 0;
2316 }
2317
2318 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2319                                        union perf_event *event,
2320                                        u64 *timestamp)
2321 {
2322         u64 type = evsel->attr.sample_type;
2323         const u64 *array;
2324
2325         if (!(type & PERF_SAMPLE_TIME))
2326                 return -1;
2327
2328         if (event->header.type != PERF_RECORD_SAMPLE) {
2329                 struct perf_sample data = {
2330                         .time = -1ULL,
2331                 };
2332
2333                 if (!evsel->attr.sample_id_all)
2334                         return -1;
2335                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2336                         return -1;
2337
2338                 *timestamp = data.time;
2339                 return 0;
2340         }
2341
2342         array = event->sample.array;
2343
2344         if (perf_event__check_size(event, evsel->sample_size))
2345                 return -EFAULT;
2346
2347         if (type & PERF_SAMPLE_IDENTIFIER)
2348                 array++;
2349
2350         if (type & PERF_SAMPLE_IP)
2351                 array++;
2352
2353         if (type & PERF_SAMPLE_TID)
2354                 array++;
2355
2356         if (type & PERF_SAMPLE_TIME)
2357                 *timestamp = *array;
2358
2359         return 0;
2360 }
2361
2362 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2363                                      u64 read_format)
2364 {
2365         size_t sz, result = sizeof(struct sample_event);
2366
2367         if (type & PERF_SAMPLE_IDENTIFIER)
2368                 result += sizeof(u64);
2369
2370         if (type & PERF_SAMPLE_IP)
2371                 result += sizeof(u64);
2372
2373         if (type & PERF_SAMPLE_TID)
2374                 result += sizeof(u64);
2375
2376         if (type & PERF_SAMPLE_TIME)
2377                 result += sizeof(u64);
2378
2379         if (type & PERF_SAMPLE_ADDR)
2380                 result += sizeof(u64);
2381
2382         if (type & PERF_SAMPLE_ID)
2383                 result += sizeof(u64);
2384
2385         if (type & PERF_SAMPLE_STREAM_ID)
2386                 result += sizeof(u64);
2387
2388         if (type & PERF_SAMPLE_CPU)
2389                 result += sizeof(u64);
2390
2391         if (type & PERF_SAMPLE_PERIOD)
2392                 result += sizeof(u64);
2393
2394         if (type & PERF_SAMPLE_READ) {
2395                 result += sizeof(u64);
2396                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2397                         result += sizeof(u64);
2398                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2399                         result += sizeof(u64);
2400                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2401                 if (read_format & PERF_FORMAT_GROUP) {
2402                         sz = sample->read.group.nr *
2403                              sizeof(struct sample_read_value);
2404                         result += sz;
2405                 } else {
2406                         result += sizeof(u64);
2407                 }
2408         }
2409
2410         if (type & PERF_SAMPLE_CALLCHAIN) {
2411                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2412                 result += sz;
2413         }
2414
2415         if (type & PERF_SAMPLE_RAW) {
2416                 result += sizeof(u32);
2417                 result += sample->raw_size;
2418         }
2419
2420         if (type & PERF_SAMPLE_BRANCH_STACK) {
2421                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2422                 sz += sizeof(u64);
2423                 result += sz;
2424         }
2425
2426         if (type & PERF_SAMPLE_REGS_USER) {
2427                 if (sample->user_regs.abi) {
2428                         result += sizeof(u64);
2429                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2430                         result += sz;
2431                 } else {
2432                         result += sizeof(u64);
2433                 }
2434         }
2435
2436         if (type & PERF_SAMPLE_STACK_USER) {
2437                 sz = sample->user_stack.size;
2438                 result += sizeof(u64);
2439                 if (sz) {
2440                         result += sz;
2441                         result += sizeof(u64);
2442                 }
2443         }
2444
2445         if (type & PERF_SAMPLE_WEIGHT)
2446                 result += sizeof(u64);
2447
2448         if (type & PERF_SAMPLE_DATA_SRC)
2449                 result += sizeof(u64);
2450
2451         if (type & PERF_SAMPLE_TRANSACTION)
2452                 result += sizeof(u64);
2453
2454         if (type & PERF_SAMPLE_REGS_INTR) {
2455                 if (sample->intr_regs.abi) {
2456                         result += sizeof(u64);
2457                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2458                         result += sz;
2459                 } else {
2460                         result += sizeof(u64);
2461                 }
2462         }
2463
2464         if (type & PERF_SAMPLE_PHYS_ADDR)
2465                 result += sizeof(u64);
2466
2467         return result;
2468 }
2469
2470 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2471                                   u64 read_format,
2472                                   const struct perf_sample *sample)
2473 {
2474         u64 *array;
2475         size_t sz;
2476         /*
2477          * used for cross-endian analysis. See git commit 65014ab3
2478          * for why this goofiness is needed.
2479          */
2480         union u64_swap u;
2481
2482         array = event->sample.array;
2483
2484         if (type & PERF_SAMPLE_IDENTIFIER) {
2485                 *array = sample->id;
2486                 array++;
2487         }
2488
2489         if (type & PERF_SAMPLE_IP) {
2490                 *array = sample->ip;
2491                 array++;
2492         }
2493
2494         if (type & PERF_SAMPLE_TID) {
2495                 u.val32[0] = sample->pid;
2496                 u.val32[1] = sample->tid;
2497                 *array = u.val64;
2498                 array++;
2499         }
2500
2501         if (type & PERF_SAMPLE_TIME) {
2502                 *array = sample->time;
2503                 array++;
2504         }
2505
2506         if (type & PERF_SAMPLE_ADDR) {
2507                 *array = sample->addr;
2508                 array++;
2509         }
2510
2511         if (type & PERF_SAMPLE_ID) {
2512                 *array = sample->id;
2513                 array++;
2514         }
2515
2516         if (type & PERF_SAMPLE_STREAM_ID) {
2517                 *array = sample->stream_id;
2518                 array++;
2519         }
2520
2521         if (type & PERF_SAMPLE_CPU) {
2522                 u.val32[0] = sample->cpu;
2523                 u.val32[1] = 0;
2524                 *array = u.val64;
2525                 array++;
2526         }
2527
2528         if (type & PERF_SAMPLE_PERIOD) {
2529                 *array = sample->period;
2530                 array++;
2531         }
2532
2533         if (type & PERF_SAMPLE_READ) {
2534                 if (read_format & PERF_FORMAT_GROUP)
2535                         *array = sample->read.group.nr;
2536                 else
2537                         *array = sample->read.one.value;
2538                 array++;
2539
2540                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2541                         *array = sample->read.time_enabled;
2542                         array++;
2543                 }
2544
2545                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2546                         *array = sample->read.time_running;
2547                         array++;
2548                 }
2549
2550                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2551                 if (read_format & PERF_FORMAT_GROUP) {
2552                         sz = sample->read.group.nr *
2553                              sizeof(struct sample_read_value);
2554                         memcpy(array, sample->read.group.values, sz);
2555                         array = (void *)array + sz;
2556                 } else {
2557                         *array = sample->read.one.id;
2558                         array++;
2559                 }
2560         }
2561
2562         if (type & PERF_SAMPLE_CALLCHAIN) {
2563                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2564                 memcpy(array, sample->callchain, sz);
2565                 array = (void *)array + sz;
2566         }
2567
2568         if (type & PERF_SAMPLE_RAW) {
2569                 u.val32[0] = sample->raw_size;
2570                 *array = u.val64;
2571                 array = (void *)array + sizeof(u32);
2572
2573                 memcpy(array, sample->raw_data, sample->raw_size);
2574                 array = (void *)array + sample->raw_size;
2575         }
2576
2577         if (type & PERF_SAMPLE_BRANCH_STACK) {
2578                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2579                 sz += sizeof(u64);
2580                 memcpy(array, sample->branch_stack, sz);
2581                 array = (void *)array + sz;
2582         }
2583
2584         if (type & PERF_SAMPLE_REGS_USER) {
2585                 if (sample->user_regs.abi) {
2586                         *array++ = sample->user_regs.abi;
2587                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2588                         memcpy(array, sample->user_regs.regs, sz);
2589                         array = (void *)array + sz;
2590                 } else {
2591                         *array++ = 0;
2592                 }
2593         }
2594
2595         if (type & PERF_SAMPLE_STACK_USER) {
2596                 sz = sample->user_stack.size;
2597                 *array++ = sz;
2598                 if (sz) {
2599                         memcpy(array, sample->user_stack.data, sz);
2600                         array = (void *)array + sz;
2601                         *array++ = sz;
2602                 }
2603         }
2604
2605         if (type & PERF_SAMPLE_WEIGHT) {
2606                 *array = sample->weight;
2607                 array++;
2608         }
2609
2610         if (type & PERF_SAMPLE_DATA_SRC) {
2611                 *array = sample->data_src;
2612                 array++;
2613         }
2614
2615         if (type & PERF_SAMPLE_TRANSACTION) {
2616                 *array = sample->transaction;
2617                 array++;
2618         }
2619
2620         if (type & PERF_SAMPLE_REGS_INTR) {
2621                 if (sample->intr_regs.abi) {
2622                         *array++ = sample->intr_regs.abi;
2623                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2624                         memcpy(array, sample->intr_regs.regs, sz);
2625                         array = (void *)array + sz;
2626                 } else {
2627                         *array++ = 0;
2628                 }
2629         }
2630
2631         if (type & PERF_SAMPLE_PHYS_ADDR) {
2632                 *array = sample->phys_addr;
2633                 array++;
2634         }
2635
2636         return 0;
2637 }
2638
2639 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2640 {
2641         return pevent_find_field(evsel->tp_format, name);
2642 }
2643
2644 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2645                          const char *name)
2646 {
2647         struct format_field *field = perf_evsel__field(evsel, name);
2648         int offset;
2649
2650         if (!field)
2651                 return NULL;
2652
2653         offset = field->offset;
2654
2655         if (field->flags & FIELD_IS_DYNAMIC) {
2656                 offset = *(int *)(sample->raw_data + field->offset);
2657                 offset &= 0xffff;
2658         }
2659
2660         return sample->raw_data + offset;
2661 }
2662
2663 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2664                          bool needs_swap)
2665 {
2666         u64 value;
2667         void *ptr = sample->raw_data + field->offset;
2668
2669         switch (field->size) {
2670         case 1:
2671                 return *(u8 *)ptr;
2672         case 2:
2673                 value = *(u16 *)ptr;
2674                 break;
2675         case 4:
2676                 value = *(u32 *)ptr;
2677                 break;
2678         case 8:
2679                 memcpy(&value, ptr, sizeof(u64));
2680                 break;
2681         default:
2682                 return 0;
2683         }
2684
2685         if (!needs_swap)
2686                 return value;
2687
2688         switch (field->size) {
2689         case 2:
2690                 return bswap_16(value);
2691         case 4:
2692                 return bswap_32(value);
2693         case 8:
2694                 return bswap_64(value);
2695         default:
2696                 return 0;
2697         }
2698
2699         return 0;
2700 }
2701
2702 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2703                        const char *name)
2704 {
2705         struct format_field *field = perf_evsel__field(evsel, name);
2706
2707         if (!field)
2708                 return 0;
2709
2710         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2711 }
2712
2713 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2714                           char *msg, size_t msgsize)
2715 {
2716         int paranoid;
2717
2718         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2719             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2720             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2721                 /*
2722                  * If it's cycles then fall back to hrtimer based
2723                  * cpu-clock-tick sw counter, which is always available even if
2724                  * no PMU support.
2725                  *
2726                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2727                  * b0a873e).
2728                  */
2729                 scnprintf(msg, msgsize, "%s",
2730 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2731
2732                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2733                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2734
2735                 zfree(&evsel->name);
2736                 return true;
2737         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2738                    (paranoid = perf_event_paranoid()) > 1) {
2739                 const char *name = perf_evsel__name(evsel);
2740                 char *new_name;
2741
2742                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2743                         return false;
2744
2745                 if (evsel->name)
2746                         free(evsel->name);
2747                 evsel->name = new_name;
2748                 scnprintf(msg, msgsize,
2749 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2750                 evsel->attr.exclude_kernel = 1;
2751
2752                 return true;
2753         }
2754
2755         return false;
2756 }
2757
2758 static bool find_process(const char *name)
2759 {
2760         size_t len = strlen(name);
2761         DIR *dir;
2762         struct dirent *d;
2763         int ret = -1;
2764
2765         dir = opendir(procfs__mountpoint());
2766         if (!dir)
2767                 return false;
2768
2769         /* Walk through the directory. */
2770         while (ret && (d = readdir(dir)) != NULL) {
2771                 char path[PATH_MAX];
2772                 char *data;
2773                 size_t size;
2774
2775                 if ((d->d_type != DT_DIR) ||
2776                      !strcmp(".", d->d_name) ||
2777                      !strcmp("..", d->d_name))
2778                         continue;
2779
2780                 scnprintf(path, sizeof(path), "%s/%s/comm",
2781                           procfs__mountpoint(), d->d_name);
2782
2783                 if (filename__read_str(path, &data, &size))
2784                         continue;
2785
2786                 ret = strncmp(name, data, len);
2787                 free(data);
2788         }
2789
2790         closedir(dir);
2791         return ret ? false : true;
2792 }
2793
2794 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2795                               int err, char *msg, size_t size)
2796 {
2797         char sbuf[STRERR_BUFSIZE];
2798         int printed = 0;
2799
2800         switch (err) {
2801         case EPERM:
2802         case EACCES:
2803                 if (err == EPERM)
2804                         printed = scnprintf(msg, size,
2805                                 "No permission to enable %s event.\n\n",
2806                                 perf_evsel__name(evsel));
2807
2808                 return scnprintf(msg + printed, size - printed,
2809                  "You may not have permission to collect %sstats.\n\n"
2810                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2811                  "which controls use of the performance events system by\n"
2812                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2813                  "The current value is %d:\n\n"
2814                  "  -1: Allow use of (almost) all events by all users\n"
2815                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2816                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2817                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2818                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2819                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2820                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2821                  "      kernel.perf_event_paranoid = -1\n" ,
2822                                  target->system_wide ? "system-wide " : "",
2823                                  perf_event_paranoid());
2824         case ENOENT:
2825                 return scnprintf(msg, size, "The %s event is not supported.",
2826                                  perf_evsel__name(evsel));
2827         case EMFILE:
2828                 return scnprintf(msg, size, "%s",
2829                          "Too many events are opened.\n"
2830                          "Probably the maximum number of open file descriptors has been reached.\n"
2831                          "Hint: Try again after reducing the number of events.\n"
2832                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2833         case ENOMEM:
2834                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2835                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2836                         return scnprintf(msg, size,
2837                                          "Not enough memory to setup event with callchain.\n"
2838                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2839                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2840                 break;
2841         case ENODEV:
2842                 if (target->cpu_list)
2843                         return scnprintf(msg, size, "%s",
2844          "No such device - did you specify an out-of-range profile CPU?");
2845                 break;
2846         case EOPNOTSUPP:
2847                 if (evsel->attr.sample_period != 0)
2848                         return scnprintf(msg, size,
2849         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2850                                          perf_evsel__name(evsel));
2851                 if (evsel->attr.precise_ip)
2852                         return scnprintf(msg, size, "%s",
2853         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2854 #if defined(__i386__) || defined(__x86_64__)
2855                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2856                         return scnprintf(msg, size, "%s",
2857         "No hardware sampling interrupt available.\n"
2858         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2859 #endif
2860                 break;
2861         case EBUSY:
2862                 if (find_process("oprofiled"))
2863                         return scnprintf(msg, size,
2864         "The PMU counters are busy/taken by another profiler.\n"
2865         "We found oprofile daemon running, please stop it and try again.");
2866                 break;
2867         case EINVAL:
2868                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2869                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2870                 if (perf_missing_features.clockid)
2871                         return scnprintf(msg, size, "clockid feature not supported.");
2872                 if (perf_missing_features.clockid_wrong)
2873                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2874                 break;
2875         default:
2876                 break;
2877         }
2878
2879         return scnprintf(msg, size,
2880         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2881         "/bin/dmesg may provide additional information.\n"
2882         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2883                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2884                          perf_evsel__name(evsel));
2885 }
2886
2887 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2888 {
2889         if (evsel && evsel->evlist)
2890                 return evsel->evlist->env;
2891         return NULL;
2892 }