tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/compiler.h>
  19 #include <linux/err.h>
  20 #include <linux/zalloc.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include <stdlib.h>
  26 #include <perf/evsel.h>
  27 #include "asm/bug.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "evlist.h"
  34 #include <perf/cpumap.h>
  35 #include "thread_map.h"
  36 #include "target.h"
  37 #include "perf_regs.h"
  38 #include "record.h"
  39 #include "debug.h"
  40 #include "trace-event.h"
  41 #include "stat.h"
  42 #include "string2.h"
  43 #include "memswap.h"
  44 #include "util.h"
  45 #include "../perf-sys.h"
  46 #include "util/parse-branch-options.h"
  47 #include <internal/xyarray.h>
  48 #include <internal/lib.h>
  49
  50 #include <linux/ctype.h>
  51
  52 struct perf_missing_features perf_missing_features;
  53
  54 static clockid_t clockid;
  55
  56 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  57 {
  58         return 0;
  59 }
  60
  61 void __weak test_attr__ready(void) { }
  62
  63 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  64 {
  65 }
  66
  67 static struct {
  68         size_t  size;
  69         int     (*init)(struct evsel *evsel);
  70         void    (*fini)(struct evsel *evsel);
  71 } perf_evsel__object = {
  72         .size = sizeof(struct evsel),
  73         .init = perf_evsel__no_extra_init,
  74         .fini = perf_evsel__no_extra_fini,
  75 };
  76
  77 int perf_evsel__object_config(size_t object_size,
  78                               int (*init)(struct evsel *evsel),
  79                               void (*fini)(struct evsel *evsel))
  80 {
  81
  82         if (object_size == 0)
  83                 goto set_methods;
  84
  85         if (perf_evsel__object.size > object_size)
  86                 return -EINVAL;
  87
  88         perf_evsel__object.size = object_size;
  89
  90 set_methods:
  91         if (init != NULL)
  92                 perf_evsel__object.init = init;
  93
  94         if (fini != NULL)
  95                 perf_evsel__object.fini = fini;
  96
  97         return 0;
  98 }
  99
 100 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 101
 102 int __perf_evsel__sample_size(u64 sample_type)
 103 {
 104         u64 mask = sample_type & PERF_SAMPLE_MASK;
 105         int size = 0;
 106         int i;
 107
 108         for (i = 0; i < 64; i++) {
 109                 if (mask & (1ULL << i))
 110                         size++;
 111         }
 112
 113         size *= sizeof(u64);
 114
 115         return size;
 116 }
 117
 118 /**
 119  * __perf_evsel__calc_id_pos - calculate id_pos.
 120  * @sample_type: sample type
 121  *
 122  * This function returns the position of the event id (PERF_SAMPLE_ID or
 123  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 124  * perf_record_sample.
 125  */
 126 static int __perf_evsel__calc_id_pos(u64 sample_type)
 127 {
 128         int idx = 0;
 129
 130         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 131                 return 0;
 132
 133         if (!(sample_type & PERF_SAMPLE_ID))
 134                 return -1;
 135
 136         if (sample_type & PERF_SAMPLE_IP)
 137                 idx += 1;
 138
 139         if (sample_type & PERF_SAMPLE_TID)
 140                 idx += 1;
 141
 142         if (sample_type & PERF_SAMPLE_TIME)
 143                 idx += 1;
 144
 145         if (sample_type & PERF_SAMPLE_ADDR)
 146                 idx += 1;
 147
 148         return idx;
 149 }
 150
 151 /**
 152  * __perf_evsel__calc_is_pos - calculate is_pos.
 153  * @sample_type: sample type
 154  *
 155  * This function returns the position (counting backwards) of the event id
 156  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 157  * sample_id_all is used there is an id sample appended to non-sample events.
 158  */
 159 static int __perf_evsel__calc_is_pos(u64 sample_type)
 160 {
 161         int idx = 1;
 162
 163         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 164                 return 1;
 165
 166         if (!(sample_type & PERF_SAMPLE_ID))
 167                 return -1;
 168
 169         if (sample_type & PERF_SAMPLE_CPU)
 170                 idx += 1;
 171
 172         if (sample_type & PERF_SAMPLE_STREAM_ID)
 173                 idx += 1;
 174
 175         return idx;
 176 }
 177
 178 void perf_evsel__calc_id_pos(struct evsel *evsel)
 179 {
 180         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 181         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 182 }
 183
 184 void __perf_evsel__set_sample_bit(struct evsel *evsel,
 185                                   enum perf_event_sample_format bit)
 186 {
 187         if (!(evsel->core.attr.sample_type & bit)) {
 188                 evsel->core.attr.sample_type |= bit;
 189                 evsel->sample_size += sizeof(u64);
 190                 perf_evsel__calc_id_pos(evsel);
 191         }
 192 }
 193
 194 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
 195                                     enum perf_event_sample_format bit)
 196 {
 197         if (evsel->core.attr.sample_type & bit) {
 198                 evsel->core.attr.sample_type &= ~bit;
 199                 evsel->sample_size -= sizeof(u64);
 200                 perf_evsel__calc_id_pos(evsel);
 201         }
 202 }
 203
 204 void perf_evsel__set_sample_id(struct evsel *evsel,
 205                                bool can_sample_identifier)
 206 {
 207         if (can_sample_identifier) {
 208                 perf_evsel__reset_sample_bit(evsel, ID);
 209                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 210         } else {
 211                 perf_evsel__set_sample_bit(evsel, ID);
 212         }
 213         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 214 }
 215
 216 /**
 217  * perf_evsel__is_function_event - Return whether given evsel is a function
 218  * trace event
 219  *
 220  * @evsel - evsel selector to be tested
 221  *
 222  * Return %true if event is function trace event
 223  */
 224 bool perf_evsel__is_function_event(struct evsel *evsel)
 225 {
 226 #define FUNCTION_EVENT "ftrace:function"
 227
 228         return evsel->name &&
 229                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 230
 231 #undef FUNCTION_EVENT
 232 }
 233
 234 void evsel__init(struct evsel *evsel,
 235                  struct perf_event_attr *attr, int idx)
 236 {
 237         perf_evsel__init(&evsel->core, attr);
 238         evsel->idx         = idx;
 239         evsel->tracking    = !idx;
 240         evsel->leader      = evsel;
 241         evsel->unit        = "";
 242         evsel->scale       = 1.0;
 243         evsel->max_events  = ULONG_MAX;
 244         evsel->evlist      = NULL;
 245         evsel->bpf_obj     = NULL;
 246         evsel->bpf_fd      = -1;
 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         evsel->pmu_name      = NULL;
 257 }
 258
 259 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 260 {
 261         struct evsel *evsel = zalloc(perf_evsel__object.size);
 262
 263         if (!evsel)
 264                 return NULL;
 265         evsel__init(evsel, attr, idx);
 266
 267         if (perf_evsel__is_bpf_output(evsel)) {
 268                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 269                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 270                 evsel->core.attr.sample_period = 1;
 271         }
 272
 273         if (perf_evsel__is_clock(evsel)) {
 274                 /*
 275                  * The evsel->unit points to static alias->unit
 276                  * so it's ok to use static string in here.
 277                  */
 278                 static const char *unit = "msec";
 279
 280                 evsel->unit = unit;
 281                 evsel->scale = 1e-6;
 282         }
 283
 284         return evsel;
 285 }
 286
 287 static bool perf_event_can_profile_kernel(void)
 288 {
 289         return perf_event_paranoid_check(1);
 290 }
 291
 292 struct evsel *perf_evsel__new_cycles(bool precise)
 293 {
 294         struct perf_event_attr attr = {
 295                 .type   = PERF_TYPE_HARDWARE,
 296                 .config = PERF_COUNT_HW_CPU_CYCLES,
 297                 .exclude_kernel = !perf_event_can_profile_kernel(),
 298         };
 299         struct evsel *evsel;
 300
 301         event_attr_init(&attr);
 302
 303         if (!precise)
 304                 goto new_event;
 305
 306         /*
 307          * Now let the usual logic to set up the perf_event_attr defaults
 308          * to kick in when we return and before perf_evsel__open() is called.
 309          */
 310 new_event:
 311         evsel = evsel__new(&attr);
 312         if (evsel == NULL)
 313                 goto out;
 314
 315         evsel->precise_max = true;
 316
 317         /* use asprintf() because free(evsel) assumes name is allocated */
 318         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 319                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 320                      attr.exclude_kernel ? "u" : "",
 321                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 322                 goto error_free;
 323 out:
 324         return evsel;
 325 error_free:
 326         evsel__delete(evsel);
 327         evsel = NULL;
 328         goto out;
 329 }
 330
 331 /*
 332  * Returns pointer with encoded error via <linux/err.h> interface.
 333  */
 334 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 335 {
 336         struct evsel *evsel = zalloc(perf_evsel__object.size);
 337         int err = -ENOMEM;
 338
 339         if (evsel == NULL) {
 340                 goto out_err;
 341         } else {
 342                 struct perf_event_attr attr = {
 343                         .type          = PERF_TYPE_TRACEPOINT,
 344                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 345                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 346                 };
 347
 348                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 349                         goto out_free;
 350
 351                 evsel->tp_format = trace_event__tp_format(sys, name);
 352                 if (IS_ERR(evsel->tp_format)) {
 353                         err = PTR_ERR(evsel->tp_format);
 354                         goto out_free;
 355                 }
 356
 357                 event_attr_init(&attr);
 358                 attr.config = evsel->tp_format->id;
 359                 attr.sample_period = 1;
 360                 evsel__init(evsel, &attr, idx);
 361         }
 362
 363         return evsel;
 364
 365 out_free:
 366         zfree(&evsel->name);
 367         free(evsel);
 368 out_err:
 369         return ERR_PTR(err);
 370 }
 371
 372 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 373         "cycles",
 374         "instructions",
 375         "cache-references",
 376         "cache-misses",
 377         "branches",
 378         "branch-misses",
 379         "bus-cycles",
 380         "stalled-cycles-frontend",
 381         "stalled-cycles-backend",
 382         "ref-cycles",
 383 };
 384
 385 static const char *__perf_evsel__hw_name(u64 config)
 386 {
 387         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 388                 return perf_evsel__hw_names[config];
 389
 390         return "unknown-hardware";
 391 }
 392
 393 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 394 {
 395         int colon = 0, r = 0;
 396         struct perf_event_attr *attr = &evsel->core.attr;
 397         bool exclude_guest_default = false;
 398
 399 #define MOD_PRINT(context, mod) do {                                    \
 400                 if (!attr->exclude_##context) {                         \
 401                         if (!colon) colon = ++r;                        \
 402                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 403                 } } while(0)
 404
 405         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 406                 MOD_PRINT(kernel, 'k');
 407                 MOD_PRINT(user, 'u');
 408                 MOD_PRINT(hv, 'h');
 409                 exclude_guest_default = true;
 410         }
 411
 412         if (attr->precise_ip) {
 413                 if (!colon)
 414                         colon = ++r;
 415                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 416                 exclude_guest_default = true;
 417         }
 418
 419         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 420                 MOD_PRINT(host, 'H');
 421                 MOD_PRINT(guest, 'G');
 422         }
 423 #undef MOD_PRINT
 424         if (colon)
 425                 bf[colon - 1] = ':';
 426         return r;
 427 }
 428
 429 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 430 {
 431         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
 432         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 433 }
 434
 435 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 436         "cpu-clock",
 437         "task-clock",
 438         "page-faults",
 439         "context-switches",
 440         "cpu-migrations",
 441         "minor-faults",
 442         "major-faults",
 443         "alignment-faults",
 444         "emulation-faults",
 445         "dummy",
 446 };
 447
 448 static const char *__perf_evsel__sw_name(u64 config)
 449 {
 450         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 451                 return perf_evsel__sw_names[config];
 452         return "unknown-software";
 453 }
 454
 455 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 456 {
 457         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
 458         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 459 }
 460
 461 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 462 {
 463         int r;
 464
 465         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 466
 467         if (type & HW_BREAKPOINT_R)
 468                 r += scnprintf(bf + r, size - r, "r");
 469
 470         if (type & HW_BREAKPOINT_W)
 471                 r += scnprintf(bf + r, size - r, "w");
 472
 473         if (type & HW_BREAKPOINT_X)
 474                 r += scnprintf(bf + r, size - r, "x");
 475
 476         return r;
 477 }
 478
 479 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 480 {
 481         struct perf_event_attr *attr = &evsel->core.attr;
 482         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 483         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 484 }
 485
 486 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 487                                 [PERF_EVSEL__MAX_ALIASES] = {
 488  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 489  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 490  { "LLC",       "L2",                                                   },
 491  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 492  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 493  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 494  { "node",                                                              },
 495 };
 496
 497 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 498                                    [PERF_EVSEL__MAX_ALIASES] = {
 499  { "load",      "loads",        "read",                                 },
 500  { "store",     "stores",       "write",                                },
 501  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 502 };
 503
 504 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 505                                        [PERF_EVSEL__MAX_ALIASES] = {
 506  { "refs",      "Reference",    "ops",          "access",               },
 507  { "misses",    "miss",                                                 },
 508 };
 509
 510 #define C(x)            PERF_COUNT_HW_CACHE_##x
 511 #define CACHE_READ      (1 << C(OP_READ))
 512 #define CACHE_WRITE     (1 << C(OP_WRITE))
 513 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 514 #define COP(x)          (1 << x)
 515
 516 /*
 517  * cache operartion stat
 518  * L1I : Read and prefetch only
 519  * ITLB and BPU : Read-only
 520  */
 521 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 522  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 523  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 524  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 526  [C(ITLB)]      = (CACHE_READ),
 527  [C(BPU)]       = (CACHE_READ),
 528  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 529 };
 530
 531 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 532 {
 533         if (perf_evsel__hw_cache_stat[type] & COP(op))
 534                 return true;    /* valid */
 535         else
 536                 return false;   /* invalid */
 537 }
 538
 539 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 540                                             char *bf, size_t size)
 541 {
 542         if (result) {
 543                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 544                                  perf_evsel__hw_cache_op[op][0],
 545                                  perf_evsel__hw_cache_result[result][0]);
 546         }
 547
 548         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 549                          perf_evsel__hw_cache_op[op][1]);
 550 }
 551
 552 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 553 {
 554         u8 op, result, type = (config >>  0) & 0xff;
 555         const char *err = "unknown-ext-hardware-cache-type";
 556
 557         if (type >= PERF_COUNT_HW_CACHE_MAX)
 558                 goto out_err;
 559
 560         op = (config >>  8) & 0xff;
 561         err = "unknown-ext-hardware-cache-op";
 562         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 563                 goto out_err;
 564
 565         result = (config >> 16) & 0xff;
 566         err = "unknown-ext-hardware-cache-result";
 567         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 568                 goto out_err;
 569
 570         err = "invalid-cache";
 571         if (!perf_evsel__is_cache_op_valid(type, op))
 572                 goto out_err;
 573
 574         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 575 out_err:
 576         return scnprintf(bf, size, "%s", err);
 577 }
 578
 579 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 580 {
 581         int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 582         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 583 }
 584
 585 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 586 {
 587         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 588         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 589 }
 590
 591 static int perf_evsel__tool_name(char *bf, size_t size)
 592 {
 593         int ret = scnprintf(bf, size, "duration_time");
 594         return ret;
 595 }
 596
 597 const char *perf_evsel__name(struct evsel *evsel)
 598 {
 599         char bf[128];
 600
 601         if (!evsel)
 602                 goto out_unknown;
 603
 604         if (evsel->name)
 605                 return evsel->name;
 606
 607         switch (evsel->core.attr.type) {
 608         case PERF_TYPE_RAW:
 609                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 610                 break;
 611
 612         case PERF_TYPE_HARDWARE:
 613                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 614                 break;
 615
 616         case PERF_TYPE_HW_CACHE:
 617                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 618                 break;
 619
 620         case PERF_TYPE_SOFTWARE:
 621                 if (evsel->tool_event)
 622                         perf_evsel__tool_name(bf, sizeof(bf));
 623                 else
 624                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 625                 break;
 626
 627         case PERF_TYPE_TRACEPOINT:
 628                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 629                 break;
 630
 631         case PERF_TYPE_BREAKPOINT:
 632                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 633                 break;
 634
 635         default:
 636                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 637                           evsel->core.attr.type);
 638                 break;
 639         }
 640
 641         evsel->name = strdup(bf);
 642
 643         if (evsel->name)
 644                 return evsel->name;
 645 out_unknown:
 646         return "unknown";
 647 }
 648
 649 const char *perf_evsel__group_name(struct evsel *evsel)
 650 {
 651         return evsel->group_name ?: "anon group";
 652 }
 653
 654 /*
 655  * Returns the group details for the specified leader,
 656  * with following rules.
 657  *
 658  *  For record -e '{cycles,instructions}'
 659  *    'anon group { cycles:u, instructions:u }'
 660  *
 661  *  For record -e 'cycles,instructions' and report --group
 662  *    'cycles:u, instructions:u'
 663  */
 664 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 665 {
 666         int ret = 0;
 667         struct evsel *pos;
 668         const char *group_name = perf_evsel__group_name(evsel);
 669
 670         if (!evsel->forced_leader)
 671                 ret = scnprintf(buf, size, "%s { ", group_name);
 672
 673         ret += scnprintf(buf + ret, size - ret, "%s",
 674                          perf_evsel__name(evsel));
 675
 676         for_each_group_member(pos, evsel)
 677                 ret += scnprintf(buf + ret, size - ret, ", %s",
 678                                  perf_evsel__name(pos));
 679
 680         if (!evsel->forced_leader)
 681                 ret += scnprintf(buf + ret, size - ret, " }");
 682
 683         return ret;
 684 }
 685
 686 static void __perf_evsel__config_callchain(struct evsel *evsel,
 687                                            struct record_opts *opts,
 688                                            struct callchain_param *param)
 689 {
 690         bool function = perf_evsel__is_function_event(evsel);
 691         struct perf_event_attr *attr = &evsel->core.attr;
 692
 693         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 694
 695         attr->sample_max_stack = param->max_stack;
 696
 697         if (opts->kernel_callchains)
 698                 attr->exclude_callchain_user = 1;
 699         if (opts->user_callchains)
 700                 attr->exclude_callchain_kernel = 1;
 701         if (param->record_mode == CALLCHAIN_LBR) {
 702                 if (!opts->branch_stack) {
 703                         if (attr->exclude_user) {
 704                                 pr_warning("LBR callstack option is only available "
 705                                            "to get user callchain information. "
 706                                            "Falling back to framepointers.\n");
 707                         } else {
 708                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 709                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 710                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 711                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 712                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 713                         }
 714                 } else
 715                          pr_warning("Cannot use LBR callstack with branch stack. "
 716                                     "Falling back to framepointers.\n");
 717         }
 718
 719         if (param->record_mode == CALLCHAIN_DWARF) {
 720                 if (!function) {
 721                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 722                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 723                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 724                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 725                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 726                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 727                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 728                         } else {
 729                                 attr->sample_regs_user |= PERF_REGS_MASK;
 730                         }
 731                         attr->sample_stack_user = param->dump_size;
 732                         attr->exclude_callchain_user = 1;
 733                 } else {
 734                         pr_info("Cannot use DWARF unwind for function trace event,"
 735                                 " falling back to framepointers.\n");
 736                 }
 737         }
 738
 739         if (function) {
 740                 pr_info("Disabling user space callchains for function trace event.\n");
 741                 attr->exclude_callchain_user = 1;
 742         }
 743 }
 744
 745 void perf_evsel__config_callchain(struct evsel *evsel,
 746                                   struct record_opts *opts,
 747                                   struct callchain_param *param)
 748 {
 749         if (param->enabled)
 750                 return __perf_evsel__config_callchain(evsel, opts, param);
 751 }
 752
 753 static void
 754 perf_evsel__reset_callgraph(struct evsel *evsel,
 755                             struct callchain_param *param)
 756 {
 757         struct perf_event_attr *attr = &evsel->core.attr;
 758
 759         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 760         if (param->record_mode == CALLCHAIN_LBR) {
 761                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 762                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 763                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 764         }
 765         if (param->record_mode == CALLCHAIN_DWARF) {
 766                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 767                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 768         }
 769 }
 770
 771 static void apply_config_terms(struct evsel *evsel,
 772                                struct record_opts *opts, bool track)
 773 {
 774         struct perf_evsel_config_term *term;
 775         struct list_head *config_terms = &evsel->config_terms;
 776         struct perf_event_attr *attr = &evsel->core.attr;
 777         /* callgraph default */
 778         struct callchain_param param = {
 779                 .record_mode = callchain_param.record_mode,
 780         };
 781         u32 dump_size = 0;
 782         int max_stack = 0;
 783         const char *callgraph_buf = NULL;
 784
 785         list_for_each_entry(term, config_terms, list) {
 786                 switch (term->type) {
 787                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 788                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 789                                 attr->sample_period = term->val.period;
 790                                 attr->freq = 0;
 791                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 792                         }
 793                         break;
 794                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 795                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 796                                 attr->sample_freq = term->val.freq;
 797                                 attr->freq = 1;
 798                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 799                         }
 800                         break;
 801                 case PERF_EVSEL__CONFIG_TERM_TIME:
 802                         if (term->val.time)
 803                                 perf_evsel__set_sample_bit(evsel, TIME);
 804                         else
 805                                 perf_evsel__reset_sample_bit(evsel, TIME);
 806                         break;
 807                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 808                         callgraph_buf = term->val.callgraph;
 809                         break;
 810                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 811                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 812                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 813                                 parse_branch_str(term->val.branch,
 814                                                  &attr->branch_sample_type);
 815                         } else
 816                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 817                         break;
 818                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 819                         dump_size = term->val.stack_user;
 820                         break;
 821                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 822                         max_stack = term->val.max_stack;
 823                         break;
 824                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 825                         evsel->max_events = term->val.max_events;
 826                         break;
 827                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 828                         /*
 829                          * attr->inherit should has already been set by
 830                          * perf_evsel__config. If user explicitly set
 831                          * inherit using config terms, override global
 832                          * opt->no_inherit setting.
 833                          */
 834                         attr->inherit = term->val.inherit ? 1 : 0;
 835                         break;
 836                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 837                         attr->write_backward = term->val.overwrite ? 1 : 0;
 838                         break;
 839                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 840                         break;
 841                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 842                         break;
 843                 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
 844                         attr->aux_output = term->val.aux_output ? 1 : 0;
 845                         break;
 846                 default:
 847                         break;
 848                 }
 849         }
 850
 851         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 852         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 853                 bool sample_address = false;
 854
 855                 if (max_stack) {
 856                         param.max_stack = max_stack;
 857                         if (callgraph_buf == NULL)
 858                                 callgraph_buf = "fp";
 859                 }
 860
 861                 /* parse callgraph parameters */
 862                 if (callgraph_buf != NULL) {
 863                         if (!strcmp(callgraph_buf, "no")) {
 864                                 param.enabled = false;
 865                                 param.record_mode = CALLCHAIN_NONE;
 866                         } else {
 867                                 param.enabled = true;
 868                                 if (parse_callchain_record(callgraph_buf, &param)) {
 869                                         pr_err("per-event callgraph setting for %s failed. "
 870                                                "Apply callgraph global setting for it\n",
 871                                                evsel->name);
 872                                         return;
 873                                 }
 874                                 if (param.record_mode == CALLCHAIN_DWARF)
 875                                         sample_address = true;
 876                         }
 877                 }
 878                 if (dump_size > 0) {
 879                         dump_size = round_up(dump_size, sizeof(u64));
 880                         param.dump_size = dump_size;
 881                 }
 882
 883                 /* If global callgraph set, clear it */
 884                 if (callchain_param.enabled)
 885                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 886
 887                 /* set perf-event callgraph */
 888                 if (param.enabled) {
 889                         if (sample_address) {
 890                                 perf_evsel__set_sample_bit(evsel, ADDR);
 891                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 892                                 evsel->core.attr.mmap_data = track;
 893                         }
 894                         perf_evsel__config_callchain(evsel, opts, &param);
 895                 }
 896         }
 897 }
 898
 899 static bool is_dummy_event(struct evsel *evsel)
 900 {
 901         return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
 902                (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
 903 }
 904
 905 /*
 906  * The enable_on_exec/disabled value strategy:
 907  *
 908  *  1) For any type of traced program:
 909  *    - all independent events and group leaders are disabled
 910  *    - all group members are enabled
 911  *
 912  *     Group members are ruled by group leaders. They need to
 913  *     be enabled, because the group scheduling relies on that.
 914  *
 915  *  2) For traced programs executed by perf:
 916  *     - all independent events and group leaders have
 917  *       enable_on_exec set
 918  *     - we don't specifically enable or disable any event during
 919  *       the record command
 920  *
 921  *     Independent events and group leaders are initially disabled
 922  *     and get enabled by exec. Group members are ruled by group
 923  *     leaders as stated in 1).
 924  *
 925  *  3) For traced programs attached by perf (pid/tid):
 926  *     - we specifically enable or disable all events during
 927  *       the record command
 928  *
 929  *     When attaching events to already running traced we
 930  *     enable/disable events specifically, as there's no
 931  *     initial traced exec call.
 932  */
 933 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
 934                         struct callchain_param *callchain)
 935 {
 936         struct evsel *leader = evsel->leader;
 937         struct perf_event_attr *attr = &evsel->core.attr;
 938         int track = evsel->tracking;
 939         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 940
 941         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 942         attr->inherit       = !opts->no_inherit;
 943         attr->write_backward = opts->overwrite ? 1 : 0;
 944
 945         perf_evsel__set_sample_bit(evsel, IP);
 946         perf_evsel__set_sample_bit(evsel, TID);
 947
 948         if (evsel->sample_read) {
 949                 perf_evsel__set_sample_bit(evsel, READ);
 950
 951                 /*
 952                  * We need ID even in case of single event, because
 953                  * PERF_SAMPLE_READ process ID specific data.
 954                  */
 955                 perf_evsel__set_sample_id(evsel, false);
 956
 957                 /*
 958                  * Apply group format only if we belong to group
 959                  * with more than one members.
 960                  */
 961                 if (leader->core.nr_members > 1) {
 962                         attr->read_format |= PERF_FORMAT_GROUP;
 963                         attr->inherit = 0;
 964                 }
 965         }
 966
 967         /*
 968          * We default some events to have a default interval. But keep
 969          * it a weak assumption overridable by the user.
 970          */
 971         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 972                                      opts->user_interval != ULLONG_MAX)) {
 973                 if (opts->freq) {
 974                         perf_evsel__set_sample_bit(evsel, PERIOD);
 975                         attr->freq              = 1;
 976                         attr->sample_freq       = opts->freq;
 977                 } else {
 978                         attr->sample_period = opts->default_interval;
 979                 }
 980         }
 981
 982         /*
 983          * Disable sampling for all group members other
 984          * than leader in case leader 'leads' the sampling.
 985          */
 986         if ((leader != evsel) && leader->sample_read) {
 987                 attr->freq           = 0;
 988                 attr->sample_freq    = 0;
 989                 attr->sample_period  = 0;
 990                 attr->write_backward = 0;
 991
 992                 /*
 993                  * We don't get sample for slave events, we make them
 994                  * when delivering group leader sample. Set the slave
 995                  * event to follow the master sample_type to ease up
 996                  * report.
 997                  */
 998                 attr->sample_type = leader->core.attr.sample_type;
 999         }
1000
1001         if (opts->no_samples)
1002                 attr->sample_freq = 0;
1003
1004         if (opts->inherit_stat) {
1005                 evsel->core.attr.read_format |=
1006                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1007                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1008                         PERF_FORMAT_ID;
1009                 attr->inherit_stat = 1;
1010         }
1011
1012         if (opts->sample_address) {
1013                 perf_evsel__set_sample_bit(evsel, ADDR);
1014                 attr->mmap_data = track;
1015         }
1016
1017         /*
1018          * We don't allow user space callchains for  function trace
1019          * event, due to issues with page faults while tracing page
1020          * fault handler and its overall trickiness nature.
1021          */
1022         if (perf_evsel__is_function_event(evsel))
1023                 evsel->core.attr.exclude_callchain_user = 1;
1024
1025         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1026                 perf_evsel__config_callchain(evsel, opts, callchain);
1027
1028         if (opts->sample_intr_regs) {
1029                 attr->sample_regs_intr = opts->sample_intr_regs;
1030                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1031         }
1032
1033         if (opts->sample_user_regs) {
1034                 attr->sample_regs_user |= opts->sample_user_regs;
1035                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1036         }
1037
1038         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1039                 perf_evsel__set_sample_bit(evsel, CPU);
1040
1041         /*
1042          * When the user explicitly disabled time don't force it here.
1043          */
1044         if (opts->sample_time &&
1045             (!perf_missing_features.sample_id_all &&
1046             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1047              opts->sample_time_set)))
1048                 perf_evsel__set_sample_bit(evsel, TIME);
1049
1050         if (opts->raw_samples && !evsel->no_aux_samples) {
1051                 perf_evsel__set_sample_bit(evsel, TIME);
1052                 perf_evsel__set_sample_bit(evsel, RAW);
1053                 perf_evsel__set_sample_bit(evsel, CPU);
1054         }
1055
1056         if (opts->sample_address)
1057                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1058
1059         if (opts->sample_phys_addr)
1060                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1061
1062         if (opts->no_buffering) {
1063                 attr->watermark = 0;
1064                 attr->wakeup_events = 1;
1065         }
1066         if (opts->branch_stack && !evsel->no_aux_samples) {
1067                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1068                 attr->branch_sample_type = opts->branch_stack;
1069         }
1070
1071         if (opts->sample_weight)
1072                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1073
1074         attr->task  = track;
1075         attr->mmap  = track;
1076         attr->mmap2 = track && !perf_missing_features.mmap2;
1077         attr->comm  = track;
1078         attr->ksymbol = track && !perf_missing_features.ksymbol;
1079         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1080
1081         if (opts->record_namespaces)
1082                 attr->namespaces  = track;
1083
1084         if (opts->record_switch_events)
1085                 attr->context_switch = track;
1086
1087         if (opts->sample_transaction)
1088                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1089
1090         if (opts->running_time) {
1091                 evsel->core.attr.read_format |=
1092                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1093                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1094         }
1095
1096         /*
1097          * XXX see the function comment above
1098          *
1099          * Disabling only independent events or group leaders,
1100          * keeping group members enabled.
1101          */
1102         if (perf_evsel__is_group_leader(evsel))
1103                 attr->disabled = 1;
1104
1105         /*
1106          * Setting enable_on_exec for independent events and
1107          * group leaders for traced executed by perf.
1108          */
1109         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1110                 !opts->initial_delay)
1111                 attr->enable_on_exec = 1;
1112
1113         if (evsel->immediate) {
1114                 attr->disabled = 0;
1115                 attr->enable_on_exec = 0;
1116         }
1117
1118         clockid = opts->clockid;
1119         if (opts->use_clockid) {
1120                 attr->use_clockid = 1;
1121                 attr->clockid = opts->clockid;
1122         }
1123
1124         if (evsel->precise_max)
1125                 attr->precise_ip = 3;
1126
1127         if (opts->all_user) {
1128                 attr->exclude_kernel = 1;
1129                 attr->exclude_user   = 0;
1130         }
1131
1132         if (opts->all_kernel) {
1133                 attr->exclude_kernel = 0;
1134                 attr->exclude_user   = 1;
1135         }
1136
1137         if (evsel->core.own_cpus || evsel->unit)
1138                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1139
1140         /*
1141          * Apply event specific term settings,
1142          * it overloads any global configuration.
1143          */
1144         apply_config_terms(evsel, opts, track);
1145
1146         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1147
1148         /* The --period option takes the precedence. */
1149         if (opts->period_set) {
1150                 if (opts->period)
1151                         perf_evsel__set_sample_bit(evsel, PERIOD);
1152                 else
1153                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1154         }
1155
1156         /*
1157          * For initial_delay, a dummy event is added implicitly.
1158          * The software event will trigger -EOPNOTSUPP error out,
1159          * if BRANCH_STACK bit is set.
1160          */
1161         if (opts->initial_delay && is_dummy_event(evsel))
1162                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1163 }
1164
1165 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1166 {
1167         char *new_filter = strdup(filter);
1168
1169         if (new_filter != NULL) {
1170                 free(evsel->filter);
1171                 evsel->filter = new_filter;
1172                 return 0;
1173         }
1174
1175         return -1;
1176 }
1177
1178 static int perf_evsel__append_filter(struct evsel *evsel,
1179                                      const char *fmt, const char *filter)
1180 {
1181         char *new_filter;
1182
1183         if (evsel->filter == NULL)
1184                 return perf_evsel__set_filter(evsel, filter);
1185
1186         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1187                 free(evsel->filter);
1188                 evsel->filter = new_filter;
1189                 return 0;
1190         }
1191
1192         return -1;
1193 }
1194
1195 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1196 {
1197         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1198 }
1199
1200 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1201 {
1202         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1203 }
1204
1205 int evsel__enable(struct evsel *evsel)
1206 {
1207         int err = perf_evsel__enable(&evsel->core);
1208
1209         if (!err)
1210                 evsel->disabled = false;
1211
1212         return err;
1213 }
1214
1215 int evsel__disable(struct evsel *evsel)
1216 {
1217         int err = perf_evsel__disable(&evsel->core);
1218         /*
1219          * We mark it disabled here so that tools that disable a event can
1220          * ignore events after they disable it. I.e. the ring buffer may have
1221          * already a few more events queued up before the kernel got the stop
1222          * request.
1223          */
1224         if (!err)
1225                 evsel->disabled = true;
1226
1227         return err;
1228 }
1229
1230 int perf_evsel__alloc_id(struct evsel *evsel, int ncpus, int nthreads)
1231 {
1232         if (ncpus == 0 || nthreads == 0)
1233                 return 0;
1234
1235         if (evsel->system_wide)
1236                 nthreads = 1;
1237
1238         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1239         if (evsel->sample_id == NULL)
1240                 return -ENOMEM;
1241
1242         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1243         if (evsel->id == NULL) {
1244                 xyarray__delete(evsel->sample_id);
1245                 evsel->sample_id = NULL;
1246                 return -ENOMEM;
1247         }
1248
1249         return 0;
1250 }
1251
1252 static void perf_evsel__free_id(struct evsel *evsel)
1253 {
1254         xyarray__delete(evsel->sample_id);
1255         evsel->sample_id = NULL;
1256         zfree(&evsel->id);
1257         evsel->ids = 0;
1258 }
1259
1260 static void perf_evsel__free_config_terms(struct evsel *evsel)
1261 {
1262         struct perf_evsel_config_term *term, *h;
1263
1264         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1265                 list_del_init(&term->list);
1266                 free(term);
1267         }
1268 }
1269
1270 void perf_evsel__exit(struct evsel *evsel)
1271 {
1272         assert(list_empty(&evsel->core.node));
1273         assert(evsel->evlist == NULL);
1274         perf_evsel__free_counts(evsel);
1275         perf_evsel__free_fd(&evsel->core);
1276         perf_evsel__free_id(evsel);
1277         perf_evsel__free_config_terms(evsel);
1278         cgroup__put(evsel->cgrp);
1279         perf_cpu_map__put(evsel->core.cpus);
1280         perf_cpu_map__put(evsel->core.own_cpus);
1281         perf_thread_map__put(evsel->core.threads);
1282         zfree(&evsel->group_name);
1283         zfree(&evsel->name);
1284         perf_evsel__object.fini(evsel);
1285 }
1286
1287 void evsel__delete(struct evsel *evsel)
1288 {
1289         perf_evsel__exit(evsel);
1290         free(evsel);
1291 }
1292
1293 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1294                                 struct perf_counts_values *count)
1295 {
1296         struct perf_counts_values tmp;
1297
1298         if (!evsel->prev_raw_counts)
1299                 return;
1300
1301         if (cpu == -1) {
1302                 tmp = evsel->prev_raw_counts->aggr;
1303                 evsel->prev_raw_counts->aggr = *count;
1304         } else {
1305                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1306                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1307         }
1308
1309         count->val = count->val - tmp.val;
1310         count->ena = count->ena - tmp.ena;
1311         count->run = count->run - tmp.run;
1312 }
1313
1314 void perf_counts_values__scale(struct perf_counts_values *count,
1315                                bool scale, s8 *pscaled)
1316 {
1317         s8 scaled = 0;
1318
1319         if (scale) {
1320                 if (count->run == 0) {
1321                         scaled = -1;
1322                         count->val = 0;
1323                 } else if (count->run < count->ena) {
1324                         scaled = 1;
1325                         count->val = (u64)((double) count->val * count->ena / count->run);
1326                 }
1327         }
1328
1329         if (pscaled)
1330                 *pscaled = scaled;
1331 }
1332
1333 static int
1334 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1335 {
1336         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1337
1338         return perf_evsel__read(&evsel->core, cpu, thread, count);
1339 }
1340
1341 static void
1342 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1343                       u64 val, u64 ena, u64 run)
1344 {
1345         struct perf_counts_values *count;
1346
1347         count = perf_counts(counter->counts, cpu, thread);
1348
1349         count->val    = val;
1350         count->ena    = ena;
1351         count->run    = run;
1352
1353         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1354 }
1355
1356 static int
1357 perf_evsel__process_group_data(struct evsel *leader,
1358                                int cpu, int thread, u64 *data)
1359 {
1360         u64 read_format = leader->core.attr.read_format;
1361         struct sample_read_value *v;
1362         u64 nr, ena = 0, run = 0, i;
1363
1364         nr = *data++;
1365
1366         if (nr != (u64) leader->core.nr_members)
1367                 return -EINVAL;
1368
1369         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1370                 ena = *data++;
1371
1372         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1373                 run = *data++;
1374
1375         v = (struct sample_read_value *) data;
1376
1377         perf_evsel__set_count(leader, cpu, thread,
1378                               v[0].value, ena, run);
1379
1380         for (i = 1; i < nr; i++) {
1381                 struct evsel *counter;
1382
1383                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1384                 if (!counter)
1385                         return -EINVAL;
1386
1387                 perf_evsel__set_count(counter, cpu, thread,
1388                                       v[i].value, ena, run);
1389         }
1390
1391         return 0;
1392 }
1393
1394 static int
1395 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1396 {
1397         struct perf_stat_evsel *ps = leader->stats;
1398         u64 read_format = leader->core.attr.read_format;
1399         int size = perf_evsel__read_size(&leader->core);
1400         u64 *data = ps->group_data;
1401
1402         if (!(read_format & PERF_FORMAT_ID))
1403                 return -EINVAL;
1404
1405         if (!perf_evsel__is_group_leader(leader))
1406                 return -EINVAL;
1407
1408         if (!data) {
1409                 data = zalloc(size);
1410                 if (!data)
1411                         return -ENOMEM;
1412
1413                 ps->group_data = data;
1414         }
1415
1416         if (FD(leader, cpu, thread) < 0)
1417                 return -EINVAL;
1418
1419         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1420                 return -errno;
1421
1422         return perf_evsel__process_group_data(leader, cpu, thread, data);
1423 }
1424
1425 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1426 {
1427         u64 read_format = evsel->core.attr.read_format;
1428
1429         if (read_format & PERF_FORMAT_GROUP)
1430                 return perf_evsel__read_group(evsel, cpu, thread);
1431         else
1432                 return perf_evsel__read_one(evsel, cpu, thread);
1433 }
1434
1435 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1436                               int cpu, int thread, bool scale)
1437 {
1438         struct perf_counts_values count;
1439         size_t nv = scale ? 3 : 1;
1440
1441         if (FD(evsel, cpu, thread) < 0)
1442                 return -EINVAL;
1443
1444         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1445                 return -ENOMEM;
1446
1447         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1448                 return -errno;
1449
1450         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1451         perf_counts_values__scale(&count, scale, NULL);
1452         *perf_counts(evsel->counts, cpu, thread) = count;
1453         return 0;
1454 }
1455
1456 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1457 {
1458         struct evsel *leader = evsel->leader;
1459         int fd;
1460
1461         if (perf_evsel__is_group_leader(evsel))
1462                 return -1;
1463
1464         /*
1465          * Leader must be already processed/open,
1466          * if not it's a bug.
1467          */
1468         BUG_ON(!leader->core.fd);
1469
1470         fd = FD(leader, cpu, thread);
1471         BUG_ON(fd == -1);
1472
1473         return fd;
1474 }
1475
1476 struct bit_names {
1477         int bit;
1478         const char *name;
1479 };
1480
1481 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1482 {
1483         bool first_bit = true;
1484         int i = 0;
1485
1486         do {
1487                 if (value & bits[i].bit) {
1488                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1489                         first_bit = false;
1490                 }
1491         } while (bits[++i].name != NULL);
1492 }
1493
1494 static void __p_sample_type(char *buf, size_t size, u64 value)
1495 {
1496 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1497         struct bit_names bits[] = {
1498                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1499                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1500                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1501                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1502                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1503                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1504                 { .name = NULL, }
1505         };
1506 #undef bit_name
1507         __p_bits(buf, size, value, bits);
1508 }
1509
1510 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1511 {
1512 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1513         struct bit_names bits[] = {
1514                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1515                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1516                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1517                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1518                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1519                 { .name = NULL, }
1520         };
1521 #undef bit_name
1522         __p_bits(buf, size, value, bits);
1523 }
1524
1525 static void __p_read_format(char *buf, size_t size, u64 value)
1526 {
1527 #define bit_name(n) { PERF_FORMAT_##n, #n }
1528         struct bit_names bits[] = {
1529                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1530                 bit_name(ID), bit_name(GROUP),
1531                 { .name = NULL, }
1532         };
1533 #undef bit_name
1534         __p_bits(buf, size, value, bits);
1535 }
1536
1537 #define BUF_SIZE                1024
1538
1539 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1540 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1541 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1542 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1543 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1544 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1545
1546 #define PRINT_ATTRn(_n, _f, _p)                         \
1547 do {                                                    \
1548         if (attr->_f) {                                 \
1549                 _p(attr->_f);                           \
1550                 ret += attr__fprintf(fp, _n, buf, priv);\
1551         }                                               \
1552 } while (0)
1553
1554 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1555
1556 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1557                              attr__fprintf_f attr__fprintf, void *priv)
1558 {
1559         char buf[BUF_SIZE];
1560         int ret = 0;
1561
1562         PRINT_ATTRf(type, p_unsigned);
1563         PRINT_ATTRf(size, p_unsigned);
1564         PRINT_ATTRf(config, p_hex);
1565         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1566         PRINT_ATTRf(sample_type, p_sample_type);
1567         PRINT_ATTRf(read_format, p_read_format);
1568
1569         PRINT_ATTRf(disabled, p_unsigned);
1570         PRINT_ATTRf(inherit, p_unsigned);
1571         PRINT_ATTRf(pinned, p_unsigned);
1572         PRINT_ATTRf(exclusive, p_unsigned);
1573         PRINT_ATTRf(exclude_user, p_unsigned);
1574         PRINT_ATTRf(exclude_kernel, p_unsigned);
1575         PRINT_ATTRf(exclude_hv, p_unsigned);
1576         PRINT_ATTRf(exclude_idle, p_unsigned);
1577         PRINT_ATTRf(mmap, p_unsigned);
1578         PRINT_ATTRf(comm, p_unsigned);
1579         PRINT_ATTRf(freq, p_unsigned);
1580         PRINT_ATTRf(inherit_stat, p_unsigned);
1581         PRINT_ATTRf(enable_on_exec, p_unsigned);
1582         PRINT_ATTRf(task, p_unsigned);
1583         PRINT_ATTRf(watermark, p_unsigned);
1584         PRINT_ATTRf(precise_ip, p_unsigned);
1585         PRINT_ATTRf(mmap_data, p_unsigned);
1586         PRINT_ATTRf(sample_id_all, p_unsigned);
1587         PRINT_ATTRf(exclude_host, p_unsigned);
1588         PRINT_ATTRf(exclude_guest, p_unsigned);
1589         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1590         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1591         PRINT_ATTRf(mmap2, p_unsigned);
1592         PRINT_ATTRf(comm_exec, p_unsigned);
1593         PRINT_ATTRf(use_clockid, p_unsigned);
1594         PRINT_ATTRf(context_switch, p_unsigned);
1595         PRINT_ATTRf(write_backward, p_unsigned);
1596         PRINT_ATTRf(namespaces, p_unsigned);
1597         PRINT_ATTRf(ksymbol, p_unsigned);
1598         PRINT_ATTRf(bpf_event, p_unsigned);
1599         PRINT_ATTRf(aux_output, p_unsigned);
1600
1601         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1602         PRINT_ATTRf(bp_type, p_unsigned);
1603         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1604         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1605         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1606         PRINT_ATTRf(sample_regs_user, p_hex);
1607         PRINT_ATTRf(sample_stack_user, p_unsigned);
1608         PRINT_ATTRf(clockid, p_signed);
1609         PRINT_ATTRf(sample_regs_intr, p_hex);
1610         PRINT_ATTRf(aux_watermark, p_unsigned);
1611         PRINT_ATTRf(sample_max_stack, p_unsigned);
1612
1613         return ret;
1614 }
1615
1616 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1617                                 void *priv __maybe_unused)
1618 {
1619         return fprintf(fp, "  %-32s %s\n", name, val);
1620 }
1621
1622 static void perf_evsel__remove_fd(struct evsel *pos,
1623                                   int nr_cpus, int nr_threads,
1624                                   int thread_idx)
1625 {
1626         for (int cpu = 0; cpu < nr_cpus; cpu++)
1627                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1628                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1629 }
1630
1631 static int update_fds(struct evsel *evsel,
1632                       int nr_cpus, int cpu_idx,
1633                       int nr_threads, int thread_idx)
1634 {
1635         struct evsel *pos;
1636
1637         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1638                 return -EINVAL;
1639
1640         evlist__for_each_entry(evsel->evlist, pos) {
1641                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1642
1643                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1644
1645                 /*
1646                  * Since fds for next evsel has not been created,
1647                  * there is no need to iterate whole event list.
1648                  */
1649                 if (pos == evsel)
1650                         break;
1651         }
1652         return 0;
1653 }
1654
1655 static bool ignore_missing_thread(struct evsel *evsel,
1656                                   int nr_cpus, int cpu,
1657                                   struct perf_thread_map *threads,
1658                                   int thread, int err)
1659 {
1660         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1661
1662         if (!evsel->ignore_missing_thread)
1663                 return false;
1664
1665         /* The system wide setup does not work with threads. */
1666         if (evsel->system_wide)
1667                 return false;
1668
1669         /* The -ESRCH is perf event syscall errno for pid's not found. */
1670         if (err != -ESRCH)
1671                 return false;
1672
1673         /* If there's only one thread, let it fail. */
1674         if (threads->nr == 1)
1675                 return false;
1676
1677         /*
1678          * We should remove fd for missing_thread first
1679          * because thread_map__remove() will decrease threads->nr.
1680          */
1681         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1682                 return false;
1683
1684         if (thread_map__remove(threads, thread))
1685                 return false;
1686
1687         pr_warning("WARNING: Ignored open failure for pid %d\n",
1688                    ignore_pid);
1689         return true;
1690 }
1691
1692 static void display_attr(struct perf_event_attr *attr)
1693 {
1694         if (verbose >= 2) {
1695                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1696                 fprintf(stderr, "perf_event_attr:\n");
1697                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1698                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1699         }
1700 }
1701
1702 static int perf_event_open(struct evsel *evsel,
1703                            pid_t pid, int cpu, int group_fd,
1704                            unsigned long flags)
1705 {
1706         int precise_ip = evsel->core.attr.precise_ip;
1707         int fd;
1708
1709         while (1) {
1710                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1711                           pid, cpu, group_fd, flags);
1712
1713                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1714                 if (fd >= 0)
1715                         break;
1716
1717                 /* Do not try less precise if not requested. */
1718                 if (!evsel->precise_max)
1719                         break;
1720
1721                 /*
1722                  * We tried all the precise_ip values, and it's
1723                  * still failing, so leave it to standard fallback.
1724                  */
1725                 if (!evsel->core.attr.precise_ip) {
1726                         evsel->core.attr.precise_ip = precise_ip;
1727                         break;
1728                 }
1729
1730                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1731                 evsel->core.attr.precise_ip--;
1732                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1733                 display_attr(&evsel->core.attr);
1734         }
1735
1736         return fd;
1737 }
1738
1739 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1740                 struct perf_thread_map *threads)
1741 {
1742         int cpu, thread, nthreads;
1743         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1744         int pid = -1, err;
1745         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1746
1747         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1748             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1749                 return -EINVAL;
1750
1751         if (cpus == NULL) {
1752                 static struct perf_cpu_map *empty_cpu_map;
1753
1754                 if (empty_cpu_map == NULL) {
1755                         empty_cpu_map = perf_cpu_map__dummy_new();
1756                         if (empty_cpu_map == NULL)
1757                                 return -ENOMEM;
1758                 }
1759
1760                 cpus = empty_cpu_map;
1761         }
1762
1763         if (threads == NULL) {
1764                 static struct perf_thread_map *empty_thread_map;
1765
1766                 if (empty_thread_map == NULL) {
1767                         empty_thread_map = thread_map__new_by_tid(-1);
1768                         if (empty_thread_map == NULL)
1769                                 return -ENOMEM;
1770                 }
1771
1772                 threads = empty_thread_map;
1773         }
1774
1775         if (evsel->system_wide)
1776                 nthreads = 1;
1777         else
1778                 nthreads = threads->nr;
1779
1780         if (evsel->core.fd == NULL &&
1781             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1782                 return -ENOMEM;
1783
1784         if (evsel->cgrp) {
1785                 flags |= PERF_FLAG_PID_CGROUP;
1786                 pid = evsel->cgrp->fd;
1787         }
1788
1789 fallback_missing_features:
1790         if (perf_missing_features.clockid_wrong)
1791                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1792         if (perf_missing_features.clockid) {
1793                 evsel->core.attr.use_clockid = 0;
1794                 evsel->core.attr.clockid = 0;
1795         }
1796         if (perf_missing_features.cloexec)
1797                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1798         if (perf_missing_features.mmap2)
1799                 evsel->core.attr.mmap2 = 0;
1800         if (perf_missing_features.exclude_guest)
1801                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1802         if (perf_missing_features.lbr_flags)
1803                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1804                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1805         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1806                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1807         if (perf_missing_features.ksymbol)
1808                 evsel->core.attr.ksymbol = 0;
1809         if (perf_missing_features.bpf)
1810                 evsel->core.attr.bpf_event = 0;
1811 retry_sample_id:
1812         if (perf_missing_features.sample_id_all)
1813                 evsel->core.attr.sample_id_all = 0;
1814
1815         display_attr(&evsel->core.attr);
1816
1817         for (cpu = 0; cpu < cpus->nr; cpu++) {
1818
1819                 for (thread = 0; thread < nthreads; thread++) {
1820                         int fd, group_fd;
1821
1822                         if (!evsel->cgrp && !evsel->system_wide)
1823                                 pid = perf_thread_map__pid(threads, thread);
1824
1825                         group_fd = get_group_fd(evsel, cpu, thread);
1826 retry_open:
1827                         test_attr__ready();
1828
1829                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1830                                              group_fd, flags);
1831
1832                         FD(evsel, cpu, thread) = fd;
1833
1834                         if (fd < 0) {
1835                                 err = -errno;
1836
1837                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1838                                         /*
1839                                          * We just removed 1 thread, so take a step
1840                                          * back on thread index and lower the upper
1841                                          * nthreads limit.
1842                                          */
1843                                         nthreads--;
1844                                         thread--;
1845
1846                                         /* ... and pretend like nothing have happened. */
1847                                         err = 0;
1848                                         continue;
1849                                 }
1850
1851                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1852                                           err);
1853                                 goto try_fallback;
1854                         }
1855
1856                         pr_debug2(" = %d\n", fd);
1857
1858                         if (evsel->bpf_fd >= 0) {
1859                                 int evt_fd = fd;
1860                                 int bpf_fd = evsel->bpf_fd;
1861
1862                                 err = ioctl(evt_fd,
1863                                             PERF_EVENT_IOC_SET_BPF,
1864                                             bpf_fd);
1865                                 if (err && errno != EEXIST) {
1866                                         pr_err("failed to attach bpf fd %d: %s\n",
1867                                                bpf_fd, strerror(errno));
1868                                         err = -EINVAL;
1869                                         goto out_close;
1870                                 }
1871                         }
1872
1873                         set_rlimit = NO_CHANGE;
1874
1875                         /*
1876                          * If we succeeded but had to kill clockid, fail and
1877                          * have perf_evsel__open_strerror() print us a nice
1878                          * error.
1879                          */
1880                         if (perf_missing_features.clockid ||
1881                             perf_missing_features.clockid_wrong) {
1882                                 err = -EINVAL;
1883                                 goto out_close;
1884                         }
1885                 }
1886         }
1887
1888         return 0;
1889
1890 try_fallback:
1891         /*
1892          * perf stat needs between 5 and 22 fds per CPU. When we run out
1893          * of them try to increase the limits.
1894          */
1895         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1896                 struct rlimit l;
1897                 int old_errno = errno;
1898
1899                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1900                         if (set_rlimit == NO_CHANGE)
1901                                 l.rlim_cur = l.rlim_max;
1902                         else {
1903                                 l.rlim_cur = l.rlim_max + 1000;
1904                                 l.rlim_max = l.rlim_cur;
1905                         }
1906                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1907                                 set_rlimit++;
1908                                 errno = old_errno;
1909                                 goto retry_open;
1910                         }
1911                 }
1912                 errno = old_errno;
1913         }
1914
1915         if (err != -EINVAL || cpu > 0 || thread > 0)
1916                 goto out_close;
1917
1918         /*
1919          * Must probe features in the order they were added to the
1920          * perf_event_attr interface.
1921          */
1922         if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1923                 perf_missing_features.aux_output = true;
1924                 pr_debug2("Kernel has no attr.aux_output support, bailing out\n");
1925                 goto out_close;
1926         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1927                 perf_missing_features.bpf = true;
1928                 pr_debug2("switching off bpf_event\n");
1929                 goto fallback_missing_features;
1930         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1931                 perf_missing_features.ksymbol = true;
1932                 pr_debug2("switching off ksymbol\n");
1933                 goto fallback_missing_features;
1934         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1935                 perf_missing_features.write_backward = true;
1936                 pr_debug2("switching off write_backward\n");
1937                 goto out_close;
1938         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1939                 perf_missing_features.clockid_wrong = true;
1940                 pr_debug2("switching off clockid\n");
1941                 goto fallback_missing_features;
1942         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1943                 perf_missing_features.clockid = true;
1944                 pr_debug2("switching off use_clockid\n");
1945                 goto fallback_missing_features;
1946         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1947                 perf_missing_features.cloexec = true;
1948                 pr_debug2("switching off cloexec flag\n");
1949                 goto fallback_missing_features;
1950         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1951                 perf_missing_features.mmap2 = true;
1952                 pr_debug2("switching off mmap2\n");
1953                 goto fallback_missing_features;
1954         } else if (!perf_missing_features.exclude_guest &&
1955                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1956                 perf_missing_features.exclude_guest = true;
1957                 pr_debug2("switching off exclude_guest, exclude_host\n");
1958                 goto fallback_missing_features;
1959         } else if (!perf_missing_features.sample_id_all) {
1960                 perf_missing_features.sample_id_all = true;
1961                 pr_debug2("switching off sample_id_all\n");
1962                 goto retry_sample_id;
1963         } else if (!perf_missing_features.lbr_flags &&
1964                         (evsel->core.attr.branch_sample_type &
1965                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1966                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1967                 perf_missing_features.lbr_flags = true;
1968                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1969                 goto fallback_missing_features;
1970         } else if (!perf_missing_features.group_read &&
1971                     evsel->core.attr.inherit &&
1972                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1973                    perf_evsel__is_group_leader(evsel)) {
1974                 perf_missing_features.group_read = true;
1975                 pr_debug2("switching off group read\n");
1976                 goto fallback_missing_features;
1977         }
1978 out_close:
1979         if (err)
1980                 threads->err_thread = thread;
1981
1982         do {
1983                 while (--thread >= 0) {
1984                         close(FD(evsel, cpu, thread));
1985                         FD(evsel, cpu, thread) = -1;
1986                 }
1987                 thread = nthreads;
1988         } while (--cpu >= 0);
1989         return err;
1990 }
1991
1992 void evsel__close(struct evsel *evsel)
1993 {
1994         perf_evsel__close(&evsel->core);
1995         perf_evsel__free_id(evsel);
1996 }
1997
1998 int perf_evsel__open_per_cpu(struct evsel *evsel,
1999                              struct perf_cpu_map *cpus)
2000 {
2001         return evsel__open(evsel, cpus, NULL);
2002 }
2003
2004 int perf_evsel__open_per_thread(struct evsel *evsel,
2005                                 struct perf_thread_map *threads)
2006 {
2007         return evsel__open(evsel, NULL, threads);
2008 }
2009
2010 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2011                                        const union perf_event *event,
2012                                        struct perf_sample *sample)
2013 {
2014         u64 type = evsel->core.attr.sample_type;
2015         const __u64 *array = event->sample.array;
2016         bool swapped = evsel->needs_swap;
2017         union u64_swap u;
2018
2019         array += ((event->header.size -
2020                    sizeof(event->header)) / sizeof(u64)) - 1;
2021
2022         if (type & PERF_SAMPLE_IDENTIFIER) {
2023                 sample->id = *array;
2024                 array--;
2025         }
2026
2027         if (type & PERF_SAMPLE_CPU) {
2028                 u.val64 = *array;
2029                 if (swapped) {
2030                         /* undo swap of u64, then swap on individual u32s */
2031                         u.val64 = bswap_64(u.val64);
2032                         u.val32[0] = bswap_32(u.val32[0]);
2033                 }
2034
2035                 sample->cpu = u.val32[0];
2036                 array--;
2037         }
2038
2039         if (type & PERF_SAMPLE_STREAM_ID) {
2040                 sample->stream_id = *array;
2041                 array--;
2042         }
2043
2044         if (type & PERF_SAMPLE_ID) {
2045                 sample->id = *array;
2046                 array--;
2047         }
2048
2049         if (type & PERF_SAMPLE_TIME) {
2050                 sample->time = *array;
2051                 array--;
2052         }
2053
2054         if (type & PERF_SAMPLE_TID) {
2055                 u.val64 = *array;
2056                 if (swapped) {
2057                         /* undo swap of u64, then swap on individual u32s */
2058                         u.val64 = bswap_64(u.val64);
2059                         u.val32[0] = bswap_32(u.val32[0]);
2060                         u.val32[1] = bswap_32(u.val32[1]);
2061                 }
2062
2063                 sample->pid = u.val32[0];
2064                 sample->tid = u.val32[1];
2065                 array--;
2066         }
2067
2068         return 0;
2069 }
2070
2071 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2072                             u64 size)
2073 {
2074         return size > max_size || offset + size > endp;
2075 }
2076
2077 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2078         do {                                                            \
2079                 if (overflow(endp, (max_size), (offset), (size)))       \
2080                         return -EFAULT;                                 \
2081         } while (0)
2082
2083 #define OVERFLOW_CHECK_u64(offset) \
2084         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2085
2086 static int
2087 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2088 {
2089         /*
2090          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2091          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2092          * check the format does not go past the end of the event.
2093          */
2094         if (sample_size + sizeof(event->header) > event->header.size)
2095                 return -EFAULT;
2096
2097         return 0;
2098 }
2099
2100 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2101                              struct perf_sample *data)
2102 {
2103         u64 type = evsel->core.attr.sample_type;
2104         bool swapped = evsel->needs_swap;
2105         const __u64 *array;
2106         u16 max_size = event->header.size;
2107         const void *endp = (void *)event + max_size;
2108         u64 sz;
2109
2110         /*
2111          * used for cross-endian analysis. See git commit 65014ab3
2112          * for why this goofiness is needed.
2113          */
2114         union u64_swap u;
2115
2116         memset(data, 0, sizeof(*data));
2117         data->cpu = data->pid = data->tid = -1;
2118         data->stream_id = data->id = data->time = -1ULL;
2119         data->period = evsel->core.attr.sample_period;
2120         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2121         data->misc    = event->header.misc;
2122         data->id = -1ULL;
2123         data->data_src = PERF_MEM_DATA_SRC_NONE;
2124
2125         if (event->header.type != PERF_RECORD_SAMPLE) {
2126                 if (!evsel->core.attr.sample_id_all)
2127                         return 0;
2128                 return perf_evsel__parse_id_sample(evsel, event, data);
2129         }
2130
2131         array = event->sample.array;
2132
2133         if (perf_event__check_size(event, evsel->sample_size))
2134                 return -EFAULT;
2135
2136         if (type & PERF_SAMPLE_IDENTIFIER) {
2137                 data->id = *array;
2138                 array++;
2139         }
2140
2141         if (type & PERF_SAMPLE_IP) {
2142                 data->ip = *array;
2143                 array++;
2144         }
2145
2146         if (type & PERF_SAMPLE_TID) {
2147                 u.val64 = *array;
2148                 if (swapped) {
2149                         /* undo swap of u64, then swap on individual u32s */
2150                         u.val64 = bswap_64(u.val64);
2151                         u.val32[0] = bswap_32(u.val32[0]);
2152                         u.val32[1] = bswap_32(u.val32[1]);
2153                 }
2154
2155                 data->pid = u.val32[0];
2156                 data->tid = u.val32[1];
2157                 array++;
2158         }
2159
2160         if (type & PERF_SAMPLE_TIME) {
2161                 data->time = *array;
2162                 array++;
2163         }
2164
2165         if (type & PERF_SAMPLE_ADDR) {
2166                 data->addr = *array;
2167                 array++;
2168         }
2169
2170         if (type & PERF_SAMPLE_ID) {
2171                 data->id = *array;
2172                 array++;
2173         }
2174
2175         if (type & PERF_SAMPLE_STREAM_ID) {
2176                 data->stream_id = *array;
2177                 array++;
2178         }
2179
2180         if (type & PERF_SAMPLE_CPU) {
2181
2182                 u.val64 = *array;
2183                 if (swapped) {
2184                         /* undo swap of u64, then swap on individual u32s */
2185                         u.val64 = bswap_64(u.val64);
2186                         u.val32[0] = bswap_32(u.val32[0]);
2187                 }
2188
2189                 data->cpu = u.val32[0];
2190                 array++;
2191         }
2192
2193         if (type & PERF_SAMPLE_PERIOD) {
2194                 data->period = *array;
2195                 array++;
2196         }
2197
2198         if (type & PERF_SAMPLE_READ) {
2199                 u64 read_format = evsel->core.attr.read_format;
2200
2201                 OVERFLOW_CHECK_u64(array);
2202                 if (read_format & PERF_FORMAT_GROUP)
2203                         data->read.group.nr = *array;
2204                 else
2205                         data->read.one.value = *array;
2206
2207                 array++;
2208
2209                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2210                         OVERFLOW_CHECK_u64(array);
2211                         data->read.time_enabled = *array;
2212                         array++;
2213                 }
2214
2215                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2216                         OVERFLOW_CHECK_u64(array);
2217                         data->read.time_running = *array;
2218                         array++;
2219                 }
2220
2221                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2222                 if (read_format & PERF_FORMAT_GROUP) {
2223                         const u64 max_group_nr = UINT64_MAX /
2224                                         sizeof(struct sample_read_value);
2225
2226                         if (data->read.group.nr > max_group_nr)
2227                                 return -EFAULT;
2228                         sz = data->read.group.nr *
2229                              sizeof(struct sample_read_value);
2230                         OVERFLOW_CHECK(array, sz, max_size);
2231                         data->read.group.values =
2232                                         (struct sample_read_value *)array;
2233                         array = (void *)array + sz;
2234                 } else {
2235                         OVERFLOW_CHECK_u64(array);
2236                         data->read.one.id = *array;
2237                         array++;
2238                 }
2239         }
2240
2241         if (evsel__has_callchain(evsel)) {
2242                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2243
2244                 OVERFLOW_CHECK_u64(array);
2245                 data->callchain = (struct ip_callchain *)array++;
2246                 if (data->callchain->nr > max_callchain_nr)
2247                         return -EFAULT;
2248                 sz = data->callchain->nr * sizeof(u64);
2249                 OVERFLOW_CHECK(array, sz, max_size);
2250                 array = (void *)array + sz;
2251         }
2252
2253         if (type & PERF_SAMPLE_RAW) {
2254                 OVERFLOW_CHECK_u64(array);
2255                 u.val64 = *array;
2256
2257                 /*
2258                  * Undo swap of u64, then swap on individual u32s,
2259                  * get the size of the raw area and undo all of the
2260                  * swap. The pevent interface handles endianity by
2261                  * itself.
2262                  */
2263                 if (swapped) {
2264                         u.val64 = bswap_64(u.val64);
2265                         u.val32[0] = bswap_32(u.val32[0]);
2266                         u.val32[1] = bswap_32(u.val32[1]);
2267                 }
2268                 data->raw_size = u.val32[0];
2269
2270                 /*
2271                  * The raw data is aligned on 64bits including the
2272                  * u32 size, so it's safe to use mem_bswap_64.
2273                  */
2274                 if (swapped)
2275                         mem_bswap_64((void *) array, data->raw_size);
2276
2277                 array = (void *)array + sizeof(u32);
2278
2279                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2280                 data->raw_data = (void *)array;
2281                 array = (void *)array + data->raw_size;
2282         }
2283
2284         if (type & PERF_SAMPLE_BRANCH_STACK) {
2285                 const u64 max_branch_nr = UINT64_MAX /
2286                                           sizeof(struct branch_entry);
2287
2288                 OVERFLOW_CHECK_u64(array);
2289                 data->branch_stack = (struct branch_stack *)array++;
2290
2291                 if (data->branch_stack->nr > max_branch_nr)
2292                         return -EFAULT;
2293                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2294                 OVERFLOW_CHECK(array, sz, max_size);
2295                 array = (void *)array + sz;
2296         }
2297
2298         if (type & PERF_SAMPLE_REGS_USER) {
2299                 OVERFLOW_CHECK_u64(array);
2300                 data->user_regs.abi = *array;
2301                 array++;
2302
2303                 if (data->user_regs.abi) {
2304                         u64 mask = evsel->core.attr.sample_regs_user;
2305
2306                         sz = hweight64(mask) * sizeof(u64);
2307                         OVERFLOW_CHECK(array, sz, max_size);
2308                         data->user_regs.mask = mask;
2309                         data->user_regs.regs = (u64 *)array;
2310                         array = (void *)array + sz;
2311                 }
2312         }
2313
2314         if (type & PERF_SAMPLE_STACK_USER) {
2315                 OVERFLOW_CHECK_u64(array);
2316                 sz = *array++;
2317
2318                 data->user_stack.offset = ((char *)(array - 1)
2319                                           - (char *) event);
2320
2321                 if (!sz) {
2322                         data->user_stack.size = 0;
2323                 } else {
2324                         OVERFLOW_CHECK(array, sz, max_size);
2325                         data->user_stack.data = (char *)array;
2326                         array = (void *)array + sz;
2327                         OVERFLOW_CHECK_u64(array);
2328                         data->user_stack.size = *array++;
2329                         if (WARN_ONCE(data->user_stack.size > sz,
2330                                       "user stack dump failure\n"))
2331                                 return -EFAULT;
2332                 }
2333         }
2334
2335         if (type & PERF_SAMPLE_WEIGHT) {
2336                 OVERFLOW_CHECK_u64(array);
2337                 data->weight = *array;
2338                 array++;
2339         }
2340
2341         if (type & PERF_SAMPLE_DATA_SRC) {
2342                 OVERFLOW_CHECK_u64(array);
2343                 data->data_src = *array;
2344                 array++;
2345         }
2346
2347         if (type & PERF_SAMPLE_TRANSACTION) {
2348                 OVERFLOW_CHECK_u64(array);
2349                 data->transaction = *array;
2350                 array++;
2351         }
2352
2353         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2354         if (type & PERF_SAMPLE_REGS_INTR) {
2355                 OVERFLOW_CHECK_u64(array);
2356                 data->intr_regs.abi = *array;
2357                 array++;
2358
2359                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2360                         u64 mask = evsel->core.attr.sample_regs_intr;
2361
2362                         sz = hweight64(mask) * sizeof(u64);
2363                         OVERFLOW_CHECK(array, sz, max_size);
2364                         data->intr_regs.mask = mask;
2365                         data->intr_regs.regs = (u64 *)array;
2366                         array = (void *)array + sz;
2367                 }
2368         }
2369
2370         data->phys_addr = 0;
2371         if (type & PERF_SAMPLE_PHYS_ADDR) {
2372                 data->phys_addr = *array;
2373                 array++;
2374         }
2375
2376         return 0;
2377 }
2378
2379 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2380                                        union perf_event *event,
2381                                        u64 *timestamp)
2382 {
2383         u64 type = evsel->core.attr.sample_type;
2384         const __u64 *array;
2385
2386         if (!(type & PERF_SAMPLE_TIME))
2387                 return -1;
2388
2389         if (event->header.type != PERF_RECORD_SAMPLE) {
2390                 struct perf_sample data = {
2391                         .time = -1ULL,
2392                 };
2393
2394                 if (!evsel->core.attr.sample_id_all)
2395                         return -1;
2396                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2397                         return -1;
2398
2399                 *timestamp = data.time;
2400                 return 0;
2401         }
2402
2403         array = event->sample.array;
2404
2405         if (perf_event__check_size(event, evsel->sample_size))
2406                 return -EFAULT;
2407
2408         if (type & PERF_SAMPLE_IDENTIFIER)
2409                 array++;
2410
2411         if (type & PERF_SAMPLE_IP)
2412                 array++;
2413
2414         if (type & PERF_SAMPLE_TID)
2415                 array++;
2416
2417         if (type & PERF_SAMPLE_TIME)
2418                 *timestamp = *array;
2419
2420         return 0;
2421 }
2422
2423 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2424 {
2425         return tep_find_field(evsel->tp_format, name);
2426 }
2427
2428 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2429                          const char *name)
2430 {
2431         struct tep_format_field *field = perf_evsel__field(evsel, name);
2432         int offset;
2433
2434         if (!field)
2435                 return NULL;
2436
2437         offset = field->offset;
2438
2439         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2440                 offset = *(int *)(sample->raw_data + field->offset);
2441                 offset &= 0xffff;
2442         }
2443
2444         return sample->raw_data + offset;
2445 }
2446
2447 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2448                          bool needs_swap)
2449 {
2450         u64 value;
2451         void *ptr = sample->raw_data + field->offset;
2452
2453         switch (field->size) {
2454         case 1:
2455                 return *(u8 *)ptr;
2456         case 2:
2457                 value = *(u16 *)ptr;
2458                 break;
2459         case 4:
2460                 value = *(u32 *)ptr;
2461                 break;
2462         case 8:
2463                 memcpy(&value, ptr, sizeof(u64));
2464                 break;
2465         default:
2466                 return 0;
2467         }
2468
2469         if (!needs_swap)
2470                 return value;
2471
2472         switch (field->size) {
2473         case 2:
2474                 return bswap_16(value);
2475         case 4:
2476                 return bswap_32(value);
2477         case 8:
2478                 return bswap_64(value);
2479         default:
2480                 return 0;
2481         }
2482
2483         return 0;
2484 }
2485
2486 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2487                        const char *name)
2488 {
2489         struct tep_format_field *field = perf_evsel__field(evsel, name);
2490
2491         if (!field)
2492                 return 0;
2493
2494         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2495 }
2496
2497 bool perf_evsel__fallback(struct evsel *evsel, int err,
2498                           char *msg, size_t msgsize)
2499 {
2500         int paranoid;
2501
2502         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2503             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2504             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2505                 /*
2506                  * If it's cycles then fall back to hrtimer based
2507                  * cpu-clock-tick sw counter, which is always available even if
2508                  * no PMU support.
2509                  *
2510                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2511                  * b0a873e).
2512                  */
2513                 scnprintf(msg, msgsize, "%s",
2514 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2515
2516                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2517                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2518
2519                 zfree(&evsel->name);
2520                 return true;
2521         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2522                    (paranoid = perf_event_paranoid()) > 1) {
2523                 const char *name = perf_evsel__name(evsel);
2524                 char *new_name;
2525                 const char *sep = ":";
2526
2527                 /* Is there already the separator in the name. */
2528                 if (strchr(name, '/') ||
2529                     strchr(name, ':'))
2530                         sep = "";
2531
2532                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2533                         return false;
2534
2535                 if (evsel->name)
2536                         free(evsel->name);
2537                 evsel->name = new_name;
2538                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2539                           "to fall back to excluding kernel and hypervisor "
2540                           " samples", paranoid);
2541                 evsel->core.attr.exclude_kernel = 1;
2542                 evsel->core.attr.exclude_hv     = 1;
2543
2544                 return true;
2545         }
2546
2547         return false;
2548 }
2549
2550 static bool find_process(const char *name)
2551 {
2552         size_t len = strlen(name);
2553         DIR *dir;
2554         struct dirent *d;
2555         int ret = -1;
2556
2557         dir = opendir(procfs__mountpoint());
2558         if (!dir)
2559                 return false;
2560
2561         /* Walk through the directory. */
2562         while (ret && (d = readdir(dir)) != NULL) {
2563                 char path[PATH_MAX];
2564                 char *data;
2565                 size_t size;
2566
2567                 if ((d->d_type != DT_DIR) ||
2568                      !strcmp(".", d->d_name) ||
2569                      !strcmp("..", d->d_name))
2570                         continue;
2571
2572                 scnprintf(path, sizeof(path), "%s/%s/comm",
2573                           procfs__mountpoint(), d->d_name);
2574
2575                 if (filename__read_str(path, &data, &size))
2576                         continue;
2577
2578                 ret = strncmp(name, data, len);
2579                 free(data);
2580         }
2581
2582         closedir(dir);
2583         return ret ? false : true;
2584 }
2585
2586 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2587                               int err, char *msg, size_t size)
2588 {
2589         char sbuf[STRERR_BUFSIZE];
2590         int printed = 0;
2591
2592         switch (err) {
2593         case EPERM:
2594         case EACCES:
2595                 if (err == EPERM)
2596                         printed = scnprintf(msg, size,
2597                                 "No permission to enable %s event.\n\n",
2598                                 perf_evsel__name(evsel));
2599
2600                 return scnprintf(msg + printed, size - printed,
2601                  "You may not have permission to collect %sstats.\n\n"
2602                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2603                  "which controls use of the performance events system by\n"
2604                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2605                  "The current value is %d:\n\n"
2606                  "  -1: Allow use of (almost) all events by all users\n"
2607                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2608                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2609                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2610                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2611                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2612                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2613                  "      kernel.perf_event_paranoid = -1\n" ,
2614                                  target->system_wide ? "system-wide " : "",
2615                                  perf_event_paranoid());
2616         case ENOENT:
2617                 return scnprintf(msg, size, "The %s event is not supported.",
2618                                  perf_evsel__name(evsel));
2619         case EMFILE:
2620                 return scnprintf(msg, size, "%s",
2621                          "Too many events are opened.\n"
2622                          "Probably the maximum number of open file descriptors has been reached.\n"
2623                          "Hint: Try again after reducing the number of events.\n"
2624                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2625         case ENOMEM:
2626                 if (evsel__has_callchain(evsel) &&
2627                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2628                         return scnprintf(msg, size,
2629                                          "Not enough memory to setup event with callchain.\n"
2630                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2631                                          "Hint: Current value: %d", sysctl__max_stack());
2632                 break;
2633         case ENODEV:
2634                 if (target->cpu_list)
2635                         return scnprintf(msg, size, "%s",
2636          "No such device - did you specify an out-of-range profile CPU?");
2637                 break;
2638         case EOPNOTSUPP:
2639                 if (evsel->core.attr.sample_period != 0)
2640                         return scnprintf(msg, size,
2641         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2642                                          perf_evsel__name(evsel));
2643                 if (evsel->core.attr.precise_ip)
2644                         return scnprintf(msg, size, "%s",
2645         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2646 #if defined(__i386__) || defined(__x86_64__)
2647                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2648                         return scnprintf(msg, size, "%s",
2649         "No hardware sampling interrupt available.\n");
2650 #endif
2651                 break;
2652         case EBUSY:
2653                 if (find_process("oprofiled"))
2654                         return scnprintf(msg, size,
2655         "The PMU counters are busy/taken by another profiler.\n"
2656         "We found oprofile daemon running, please stop it and try again.");
2657                 break;
2658         case EINVAL:
2659                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2660                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2661                 if (perf_missing_features.clockid)
2662                         return scnprintf(msg, size, "clockid feature not supported.");
2663                 if (perf_missing_features.clockid_wrong)
2664                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2665                 if (perf_missing_features.aux_output)
2666                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2667                 break;
2668         default:
2669                 break;
2670         }
2671
2672         return scnprintf(msg, size,
2673         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2674         "/bin/dmesg | grep -i perf may provide additional information.\n",
2675                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2676                          perf_evsel__name(evsel));
2677 }
2678
2679 struct perf_env *perf_evsel__env(struct evsel *evsel)
2680 {
2681         if (evsel && evsel->evlist)
2682                 return evsel->evlist->env;
2683         return NULL;
2684 }
2685
2686 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2687 {
2688         int cpu, thread;
2689
2690         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2691                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2692                      thread++) {
2693                         int fd = FD(evsel, cpu, thread);
2694
2695                         if (perf_evlist__id_add_fd(evlist, evsel,
2696                                                    cpu, thread, fd) < 0)
2697                                 return -1;
2698                 }
2699         }
2700
2701         return 0;
2702 }
2703
2704 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2705 {
2706         struct perf_cpu_map *cpus = evsel->core.cpus;
2707         struct perf_thread_map *threads = evsel->core.threads;
2708
2709         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
2710                 return -ENOMEM;
2711
2712         return store_evsel_ids(evsel, evlist);
2713 }