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