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