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 "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "../perf-sys.h"
  49 #include "util/parse-branch-options.h"
  50 #include <internal/xyarray.h>
  51 #include <internal/lib.h>
  52
  53 #include <linux/ctype.h>
  54
  55 struct perf_missing_features perf_missing_features;
  56
  57 static clockid_t clockid;
  58
  59 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  60 {
  61         return 0;
  62 }
  63
  64 void __weak test_attr__ready(void) { }
  65
  66 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  67 {
  68 }
  69
  70 static struct {
  71         size_t  size;
  72         int     (*init)(struct evsel *evsel);
  73         void    (*fini)(struct evsel *evsel);
  74 } perf_evsel__object = {
  75         .size = sizeof(struct evsel),
  76         .init = perf_evsel__no_extra_init,
  77         .fini = perf_evsel__no_extra_fini,
  78 };
  79
  80 int perf_evsel__object_config(size_t object_size,
  81                               int (*init)(struct evsel *evsel),
  82                               void (*fini)(struct evsel *evsel))
  83 {
  84
  85         if (object_size == 0)
  86                 goto set_methods;
  87
  88         if (perf_evsel__object.size > object_size)
  89                 return -EINVAL;
  90
  91         perf_evsel__object.size = object_size;
  92
  93 set_methods:
  94         if (init != NULL)
  95                 perf_evsel__object.init = init;
  96
  97         if (fini != NULL)
  98                 perf_evsel__object.fini = fini;
  99
 100         return 0;
 101 }
 102
 103 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 104
 105 int __perf_evsel__sample_size(u64 sample_type)
 106 {
 107         u64 mask = sample_type & PERF_SAMPLE_MASK;
 108         int size = 0;
 109         int i;
 110
 111         for (i = 0; i < 64; i++) {
 112                 if (mask & (1ULL << i))
 113                         size++;
 114         }
 115
 116         size *= sizeof(u64);
 117
 118         return size;
 119 }
 120
 121 /**
 122  * __perf_evsel__calc_id_pos - calculate id_pos.
 123  * @sample_type: sample type
 124  *
 125  * This function returns the position of the event id (PERF_SAMPLE_ID or
 126  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 127  * perf_record_sample.
 128  */
 129 static int __perf_evsel__calc_id_pos(u64 sample_type)
 130 {
 131         int idx = 0;
 132
 133         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 134                 return 0;
 135
 136         if (!(sample_type & PERF_SAMPLE_ID))
 137                 return -1;
 138
 139         if (sample_type & PERF_SAMPLE_IP)
 140                 idx += 1;
 141
 142         if (sample_type & PERF_SAMPLE_TID)
 143                 idx += 1;
 144
 145         if (sample_type & PERF_SAMPLE_TIME)
 146                 idx += 1;
 147
 148         if (sample_type & PERF_SAMPLE_ADDR)
 149                 idx += 1;
 150
 151         return idx;
 152 }
 153
 154 /**
 155  * __perf_evsel__calc_is_pos - calculate is_pos.
 156  * @sample_type: sample type
 157  *
 158  * This function returns the position (counting backwards) of the event id
 159  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 160  * sample_id_all is used there is an id sample appended to non-sample events.
 161  */
 162 static int __perf_evsel__calc_is_pos(u64 sample_type)
 163 {
 164         int idx = 1;
 165
 166         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 167                 return 1;
 168
 169         if (!(sample_type & PERF_SAMPLE_ID))
 170                 return -1;
 171
 172         if (sample_type & PERF_SAMPLE_CPU)
 173                 idx += 1;
 174
 175         if (sample_type & PERF_SAMPLE_STREAM_ID)
 176                 idx += 1;
 177
 178         return idx;
 179 }
 180
 181 void perf_evsel__calc_id_pos(struct evsel *evsel)
 182 {
 183         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 184         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 185 }
 186
 187 void __perf_evsel__set_sample_bit(struct evsel *evsel,
 188                                   enum perf_event_sample_format bit)
 189 {
 190         if (!(evsel->core.attr.sample_type & bit)) {
 191                 evsel->core.attr.sample_type |= bit;
 192                 evsel->sample_size += sizeof(u64);
 193                 perf_evsel__calc_id_pos(evsel);
 194         }
 195 }
 196
 197 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
 198                                     enum perf_event_sample_format bit)
 199 {
 200         if (evsel->core.attr.sample_type & bit) {
 201                 evsel->core.attr.sample_type &= ~bit;
 202                 evsel->sample_size -= sizeof(u64);
 203                 perf_evsel__calc_id_pos(evsel);
 204         }
 205 }
 206
 207 void perf_evsel__set_sample_id(struct evsel *evsel,
 208                                bool can_sample_identifier)
 209 {
 210         if (can_sample_identifier) {
 211                 perf_evsel__reset_sample_bit(evsel, ID);
 212                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 213         } else {
 214                 perf_evsel__set_sample_bit(evsel, ID);
 215         }
 216         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 217 }
 218
 219 /**
 220  * perf_evsel__is_function_event - Return whether given evsel is a function
 221  * trace event
 222  *
 223  * @evsel - evsel selector to be tested
 224  *
 225  * Return %true if event is function trace event
 226  */
 227 bool perf_evsel__is_function_event(struct evsel *evsel)
 228 {
 229 #define FUNCTION_EVENT "ftrace:function"
 230
 231         return evsel->name &&
 232                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 233
 234 #undef FUNCTION_EVENT
 235 }
 236
 237 void evsel__init(struct evsel *evsel,
 238                  struct perf_event_attr *attr, int idx)
 239 {
 240         perf_evsel__init(&evsel->core, attr);
 241         evsel->idx         = idx;
 242         evsel->tracking    = !idx;
 243         evsel->leader      = evsel;
 244         evsel->unit        = "";
 245         evsel->scale       = 1.0;
 246         evsel->max_events  = ULONG_MAX;
 247         evsel->evlist      = NULL;
 248         evsel->bpf_obj     = NULL;
 249         evsel->bpf_fd      = -1;
 250         INIT_LIST_HEAD(&evsel->config_terms);
 251         perf_evsel__object.init(evsel);
 252         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 253         perf_evsel__calc_id_pos(evsel);
 254         evsel->cmdline_group_boundary = false;
 255         evsel->metric_expr   = NULL;
 256         evsel->metric_name   = NULL;
 257         evsel->metric_events = NULL;
 258         evsel->collect_stat  = false;
 259         evsel->pmu_name      = NULL;
 260 }
 261
 262 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 263 {
 264         struct evsel *evsel = zalloc(perf_evsel__object.size);
 265
 266         if (!evsel)
 267                 return NULL;
 268         evsel__init(evsel, attr, idx);
 269
 270         if (perf_evsel__is_bpf_output(evsel)) {
 271                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 272                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 273                 evsel->core.attr.sample_period = 1;
 274         }
 275
 276         if (perf_evsel__is_clock(evsel)) {
 277                 /*
 278                  * The evsel->unit points to static alias->unit
 279                  * so it's ok to use static string in here.
 280                  */
 281                 static const char *unit = "msec";
 282
 283                 evsel->unit = unit;
 284                 evsel->scale = 1e-6;
 285         }
 286
 287         return evsel;
 288 }
 289
 290 static bool perf_event_can_profile_kernel(void)
 291 {
 292         return perf_event_paranoid_check(1);
 293 }
 294
 295 struct evsel *perf_evsel__new_cycles(bool precise)
 296 {
 297         struct perf_event_attr attr = {
 298                 .type   = PERF_TYPE_HARDWARE,
 299                 .config = PERF_COUNT_HW_CPU_CYCLES,
 300                 .exclude_kernel = !perf_event_can_profile_kernel(),
 301         };
 302         struct evsel *evsel;
 303
 304         event_attr_init(&attr);
 305
 306         if (!precise)
 307                 goto new_event;
 308
 309         /*
 310          * Now let the usual logic to set up the perf_event_attr defaults
 311          * to kick in when we return and before perf_evsel__open() is called.
 312          */
 313 new_event:
 314         evsel = evsel__new(&attr);
 315         if (evsel == NULL)
 316                 goto out;
 317
 318         evsel->precise_max = true;
 319
 320         /* use asprintf() because free(evsel) assumes name is allocated */
 321         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 322                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 323                      attr.exclude_kernel ? "u" : "",
 324                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 325                 goto error_free;
 326 out:
 327         return evsel;
 328 error_free:
 329         evsel__delete(evsel);
 330         evsel = NULL;
 331         goto out;
 332 }
 333
 334 /*
 335  * Returns pointer with encoded error via <linux/err.h> interface.
 336  */
 337 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 338 {
 339         struct evsel *evsel = zalloc(perf_evsel__object.size);
 340         int err = -ENOMEM;
 341
 342         if (evsel == NULL) {
 343                 goto out_err;
 344         } else {
 345                 struct perf_event_attr attr = {
 346                         .type          = PERF_TYPE_TRACEPOINT,
 347                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 348                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 349                 };
 350
 351                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 352                         goto out_free;
 353
 354                 evsel->tp_format = trace_event__tp_format(sys, name);
 355                 if (IS_ERR(evsel->tp_format)) {
 356                         err = PTR_ERR(evsel->tp_format);
 357                         goto out_free;
 358                 }
 359
 360                 event_attr_init(&attr);
 361                 attr.config = evsel->tp_format->id;
 362                 attr.sample_period = 1;
 363                 evsel__init(evsel, &attr, idx);
 364         }
 365
 366         return evsel;
 367
 368 out_free:
 369         zfree(&evsel->name);
 370         free(evsel);
 371 out_err:
 372         return ERR_PTR(err);
 373 }
 374
 375 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 376         "cycles",
 377         "instructions",
 378         "cache-references",
 379         "cache-misses",
 380         "branches",
 381         "branch-misses",
 382         "bus-cycles",
 383         "stalled-cycles-frontend",
 384         "stalled-cycles-backend",
 385         "ref-cycles",
 386 };
 387
 388 static const char *__perf_evsel__hw_name(u64 config)
 389 {
 390         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 391                 return perf_evsel__hw_names[config];
 392
 393         return "unknown-hardware";
 394 }
 395
 396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 397 {
 398         int colon = 0, r = 0;
 399         struct perf_event_attr *attr = &evsel->core.attr;
 400         bool exclude_guest_default = false;
 401
 402 #define MOD_PRINT(context, mod) do {                                    \
 403                 if (!attr->exclude_##context) {                         \
 404                         if (!colon) colon = ++r;                        \
 405                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 406                 } } while(0)
 407
 408         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 409                 MOD_PRINT(kernel, 'k');
 410                 MOD_PRINT(user, 'u');
 411                 MOD_PRINT(hv, 'h');
 412                 exclude_guest_default = true;
 413         }
 414
 415         if (attr->precise_ip) {
 416                 if (!colon)
 417                         colon = ++r;
 418                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 419                 exclude_guest_default = true;
 420         }
 421
 422         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 423                 MOD_PRINT(host, 'H');
 424                 MOD_PRINT(guest, 'G');
 425         }
 426 #undef MOD_PRINT
 427         if (colon)
 428                 bf[colon - 1] = ':';
 429         return r;
 430 }
 431
 432 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 433 {
 434         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
 435         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 436 }
 437
 438 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 439         "cpu-clock",
 440         "task-clock",
 441         "page-faults",
 442         "context-switches",
 443         "cpu-migrations",
 444         "minor-faults",
 445         "major-faults",
 446         "alignment-faults",
 447         "emulation-faults",
 448         "dummy",
 449 };
 450
 451 static const char *__perf_evsel__sw_name(u64 config)
 452 {
 453         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 454                 return perf_evsel__sw_names[config];
 455         return "unknown-software";
 456 }
 457
 458 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 459 {
 460         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
 461         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 462 }
 463
 464 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 465 {
 466         int r;
 467
 468         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 469
 470         if (type & HW_BREAKPOINT_R)
 471                 r += scnprintf(bf + r, size - r, "r");
 472
 473         if (type & HW_BREAKPOINT_W)
 474                 r += scnprintf(bf + r, size - r, "w");
 475
 476         if (type & HW_BREAKPOINT_X)
 477                 r += scnprintf(bf + r, size - r, "x");
 478
 479         return r;
 480 }
 481
 482 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 483 {
 484         struct perf_event_attr *attr = &evsel->core.attr;
 485         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 486         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 487 }
 488
 489 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 490                                 [PERF_EVSEL__MAX_ALIASES] = {
 491  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 492  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 493  { "LLC",       "L2",                                                   },
 494  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 495  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 496  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 497  { "node",                                                              },
 498 };
 499
 500 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 501                                    [PERF_EVSEL__MAX_ALIASES] = {
 502  { "load",      "loads",        "read",                                 },
 503  { "store",     "stores",       "write",                                },
 504  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 505 };
 506
 507 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 508                                        [PERF_EVSEL__MAX_ALIASES] = {
 509  { "refs",      "Reference",    "ops",          "access",               },
 510  { "misses",    "miss",                                                 },
 511 };
 512
 513 #define C(x)            PERF_COUNT_HW_CACHE_##x
 514 #define CACHE_READ      (1 << C(OP_READ))
 515 #define CACHE_WRITE     (1 << C(OP_WRITE))
 516 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 517 #define COP(x)          (1 << x)
 518
 519 /*
 520  * cache operartion stat
 521  * L1I : Read and prefetch only
 522  * ITLB and BPU : Read-only
 523  */
 524 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 525  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 526  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 527  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 528  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 529  [C(ITLB)]      = (CACHE_READ),
 530  [C(BPU)]       = (CACHE_READ),
 531  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 532 };
 533
 534 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 535 {
 536         if (perf_evsel__hw_cache_stat[type] & COP(op))
 537                 return true;    /* valid */
 538         else
 539                 return false;   /* invalid */
 540 }
 541
 542 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 543                                             char *bf, size_t size)
 544 {
 545         if (result) {
 546                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 547                                  perf_evsel__hw_cache_op[op][0],
 548                                  perf_evsel__hw_cache_result[result][0]);
 549         }
 550
 551         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 552                          perf_evsel__hw_cache_op[op][1]);
 553 }
 554
 555 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 556 {
 557         u8 op, result, type = (config >>  0) & 0xff;
 558         const char *err = "unknown-ext-hardware-cache-type";
 559
 560         if (type >= PERF_COUNT_HW_CACHE_MAX)
 561                 goto out_err;
 562
 563         op = (config >>  8) & 0xff;
 564         err = "unknown-ext-hardware-cache-op";
 565         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 566                 goto out_err;
 567
 568         result = (config >> 16) & 0xff;
 569         err = "unknown-ext-hardware-cache-result";
 570         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 571                 goto out_err;
 572
 573         err = "invalid-cache";
 574         if (!perf_evsel__is_cache_op_valid(type, op))
 575                 goto out_err;
 576
 577         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 578 out_err:
 579         return scnprintf(bf, size, "%s", err);
 580 }
 581
 582 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 583 {
 584         int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 585         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 586 }
 587
 588 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 589 {
 590         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 591         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 592 }
 593
 594 static int perf_evsel__tool_name(char *bf, size_t size)
 595 {
 596         int ret = scnprintf(bf, size, "duration_time");
 597         return ret;
 598 }
 599
 600 const char *perf_evsel__name(struct evsel *evsel)
 601 {
 602         char bf[128];
 603
 604         if (!evsel)
 605                 goto out_unknown;
 606
 607         if (evsel->name)
 608                 return evsel->name;
 609
 610         switch (evsel->core.attr.type) {
 611         case PERF_TYPE_RAW:
 612                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 613                 break;
 614
 615         case PERF_TYPE_HARDWARE:
 616                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 617                 break;
 618
 619         case PERF_TYPE_HW_CACHE:
 620                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 621                 break;
 622
 623         case PERF_TYPE_SOFTWARE:
 624                 if (evsel->tool_event)
 625                         perf_evsel__tool_name(bf, sizeof(bf));
 626                 else
 627                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 628                 break;
 629
 630         case PERF_TYPE_TRACEPOINT:
 631                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 632                 break;
 633
 634         case PERF_TYPE_BREAKPOINT:
 635                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 636                 break;
 637
 638         default:
 639                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 640                           evsel->core.attr.type);
 641                 break;
 642         }
 643
 644         evsel->name = strdup(bf);
 645
 646         if (evsel->name)
 647                 return evsel->name;
 648 out_unknown:
 649         return "unknown";
 650 }
 651
 652 const char *perf_evsel__group_name(struct evsel *evsel)
 653 {
 654         return evsel->group_name ?: "anon group";
 655 }
 656
 657 /*
 658  * Returns the group details for the specified leader,
 659  * with following rules.
 660  *
 661  *  For record -e '{cycles,instructions}'
 662  *    'anon group { cycles:u, instructions:u }'
 663  *
 664  *  For record -e 'cycles,instructions' and report --group
 665  *    'cycles:u, instructions:u'
 666  */
 667 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 668 {
 669         int ret = 0;
 670         struct evsel *pos;
 671         const char *group_name = perf_evsel__group_name(evsel);
 672
 673         if (!evsel->forced_leader)
 674                 ret = scnprintf(buf, size, "%s { ", group_name);
 675
 676         ret += scnprintf(buf + ret, size - ret, "%s",
 677                          perf_evsel__name(evsel));
 678
 679         for_each_group_member(pos, evsel)
 680                 ret += scnprintf(buf + ret, size - ret, ", %s",
 681                                  perf_evsel__name(pos));
 682
 683         if (!evsel->forced_leader)
 684                 ret += scnprintf(buf + ret, size - ret, " }");
 685
 686         return ret;
 687 }
 688
 689 static void __perf_evsel__config_callchain(struct evsel *evsel,
 690                                            struct record_opts *opts,
 691                                            struct callchain_param *param)
 692 {
 693         bool function = perf_evsel__is_function_event(evsel);
 694         struct perf_event_attr *attr = &evsel->core.attr;
 695
 696         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 697
 698         attr->sample_max_stack = param->max_stack;
 699
 700         if (opts->kernel_callchains)
 701                 attr->exclude_callchain_user = 1;
 702         if (opts->user_callchains)
 703                 attr->exclude_callchain_kernel = 1;
 704         if (param->record_mode == CALLCHAIN_LBR) {
 705                 if (!opts->branch_stack) {
 706                         if (attr->exclude_user) {
 707                                 pr_warning("LBR callstack option is only available "
 708                                            "to get user callchain information. "
 709                                            "Falling back to framepointers.\n");
 710                         } else {
 711                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 712                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 713                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 714                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 715                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 716                         }
 717                 } else
 718                          pr_warning("Cannot use LBR callstack with branch stack. "
 719                                     "Falling back to framepointers.\n");
 720         }
 721
 722         if (param->record_mode == CALLCHAIN_DWARF) {
 723                 if (!function) {
 724                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 725                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 726                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 727                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 728                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 729                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 730                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 731                         } else {
 732                                 attr->sample_regs_user |= PERF_REGS_MASK;
 733                         }
 734                         attr->sample_stack_user = param->dump_size;
 735                         attr->exclude_callchain_user = 1;
 736                 } else {
 737                         pr_info("Cannot use DWARF unwind for function trace event,"
 738                                 " falling back to framepointers.\n");
 739                 }
 740         }
 741
 742         if (function) {
 743                 pr_info("Disabling user space callchains for function trace event.\n");
 744                 attr->exclude_callchain_user = 1;
 745         }
 746 }
 747
 748 void perf_evsel__config_callchain(struct evsel *evsel,
 749                                   struct record_opts *opts,
 750                                   struct callchain_param *param)
 751 {
 752         if (param->enabled)
 753                 return __perf_evsel__config_callchain(evsel, opts, param);
 754 }
 755
 756 static void
 757 perf_evsel__reset_callgraph(struct evsel *evsel,
 758                             struct callchain_param *param)
 759 {
 760         struct perf_event_attr *attr = &evsel->core.attr;
 761
 762         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 763         if (param->record_mode == CALLCHAIN_LBR) {
 764                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 765                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 766                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 767         }
 768         if (param->record_mode == CALLCHAIN_DWARF) {
 769                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 770                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 771         }
 772 }
 773
 774 static void apply_config_terms(struct evsel *evsel,
 775                                struct record_opts *opts, bool track)
 776 {
 777         struct perf_evsel_config_term *term;
 778         struct list_head *config_terms = &evsel->config_terms;
 779         struct perf_event_attr *attr = &evsel->core.attr;
 780         /* callgraph default */
 781         struct callchain_param param = {
 782                 .record_mode = callchain_param.record_mode,
 783         };
 784         u32 dump_size = 0;
 785         int max_stack = 0;
 786         const char *callgraph_buf = NULL;
 787
 788         list_for_each_entry(term, config_terms, list) {
 789                 switch (term->type) {
 790                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 791                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 792                                 attr->sample_period = term->val.period;
 793                                 attr->freq = 0;
 794                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 795                         }
 796                         break;
 797                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 798                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 799                                 attr->sample_freq = term->val.freq;
 800                                 attr->freq = 1;
 801                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 802                         }
 803                         break;
 804                 case PERF_EVSEL__CONFIG_TERM_TIME:
 805                         if (term->val.time)
 806                                 perf_evsel__set_sample_bit(evsel, TIME);
 807                         else
 808                                 perf_evsel__reset_sample_bit(evsel, TIME);
 809                         break;
 810                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 811                         callgraph_buf = term->val.str;
 812                         break;
 813                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 814                         if (term->val.str && strcmp(term->val.str, "no")) {
 815                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 816                                 parse_branch_str(term->val.str,
 817                                                  &attr->branch_sample_type);
 818                         } else
 819                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 820                         break;
 821                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 822                         dump_size = term->val.stack_user;
 823                         break;
 824                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 825                         max_stack = term->val.max_stack;
 826                         break;
 827                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 828                         evsel->max_events = term->val.max_events;
 829                         break;
 830                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 831                         /*
 832                          * attr->inherit should has already been set by
 833                          * perf_evsel__config. If user explicitly set
 834                          * inherit using config terms, override global
 835                          * opt->no_inherit setting.
 836                          */
 837                         attr->inherit = term->val.inherit ? 1 : 0;
 838                         break;
 839                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 840                         attr->write_backward = term->val.overwrite ? 1 : 0;
 841                         break;
 842                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 843                         break;
 844                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 845                         break;
 846                 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
 847                         attr->aux_output = term->val.aux_output ? 1 : 0;
 848                         break;
 849                 case PERF_EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 850                         /* Already applied by auxtrace */
 851                         break;
 852                 case PERF_EVSEL__CONFIG_TERM_CFG_CHG:
 853                         break;
 854                 default:
 855                         break;
 856                 }
 857         }
 858
 859         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 860         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 861                 bool sample_address = false;
 862
 863                 if (max_stack) {
 864                         param.max_stack = max_stack;
 865                         if (callgraph_buf == NULL)
 866                                 callgraph_buf = "fp";
 867                 }
 868
 869                 /* parse callgraph parameters */
 870                 if (callgraph_buf != NULL) {
 871                         if (!strcmp(callgraph_buf, "no")) {
 872                                 param.enabled = false;
 873                                 param.record_mode = CALLCHAIN_NONE;
 874                         } else {
 875                                 param.enabled = true;
 876                                 if (parse_callchain_record(callgraph_buf, &param)) {
 877                                         pr_err("per-event callgraph setting for %s failed. "
 878                                                "Apply callgraph global setting for it\n",
 879                                                evsel->name);
 880                                         return;
 881                                 }
 882                                 if (param.record_mode == CALLCHAIN_DWARF)
 883                                         sample_address = true;
 884                         }
 885                 }
 886                 if (dump_size > 0) {
 887                         dump_size = round_up(dump_size, sizeof(u64));
 888                         param.dump_size = dump_size;
 889                 }
 890
 891                 /* If global callgraph set, clear it */
 892                 if (callchain_param.enabled)
 893                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 894
 895                 /* set perf-event callgraph */
 896                 if (param.enabled) {
 897                         if (sample_address) {
 898                                 perf_evsel__set_sample_bit(evsel, ADDR);
 899                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 900                                 evsel->core.attr.mmap_data = track;
 901                         }
 902                         perf_evsel__config_callchain(evsel, opts, &param);
 903                 }
 904         }
 905 }
 906
 907 static bool is_dummy_event(struct evsel *evsel)
 908 {
 909         return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
 910                (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
 911 }
 912
 913 struct perf_evsel_config_term *__perf_evsel__get_config_term(struct evsel *evsel,
 914                                                              enum evsel_term_type type)
 915 {
 916         struct perf_evsel_config_term *term, *found_term = NULL;
 917
 918         list_for_each_entry(term, &evsel->config_terms, list) {
 919                 if (term->type == type)
 920                         found_term = term;
 921         }
 922
 923         return found_term;
 924 }
 925
 926 /*
 927  * The enable_on_exec/disabled value strategy:
 928  *
 929  *  1) For any type of traced program:
 930  *    - all independent events and group leaders are disabled
 931  *    - all group members are enabled
 932  *
 933  *     Group members are ruled by group leaders. They need to
 934  *     be enabled, because the group scheduling relies on that.
 935  *
 936  *  2) For traced programs executed by perf:
 937  *     - all independent events and group leaders have
 938  *       enable_on_exec set
 939  *     - we don't specifically enable or disable any event during
 940  *       the record command
 941  *
 942  *     Independent events and group leaders are initially disabled
 943  *     and get enabled by exec. Group members are ruled by group
 944  *     leaders as stated in 1).
 945  *
 946  *  3) For traced programs attached by perf (pid/tid):
 947  *     - we specifically enable or disable all events during
 948  *       the record command
 949  *
 950  *     When attaching events to already running traced we
 951  *     enable/disable events specifically, as there's no
 952  *     initial traced exec call.
 953  */
 954 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
 955                         struct callchain_param *callchain)
 956 {
 957         struct evsel *leader = evsel->leader;
 958         struct perf_event_attr *attr = &evsel->core.attr;
 959         int track = evsel->tracking;
 960         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 961
 962         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 963         attr->inherit       = !opts->no_inherit;
 964         attr->write_backward = opts->overwrite ? 1 : 0;
 965
 966         perf_evsel__set_sample_bit(evsel, IP);
 967         perf_evsel__set_sample_bit(evsel, TID);
 968
 969         if (evsel->sample_read) {
 970                 perf_evsel__set_sample_bit(evsel, READ);
 971
 972                 /*
 973                  * We need ID even in case of single event, because
 974                  * PERF_SAMPLE_READ process ID specific data.
 975                  */
 976                 perf_evsel__set_sample_id(evsel, false);
 977
 978                 /*
 979                  * Apply group format only if we belong to group
 980                  * with more than one members.
 981                  */
 982                 if (leader->core.nr_members > 1) {
 983                         attr->read_format |= PERF_FORMAT_GROUP;
 984                         attr->inherit = 0;
 985                 }
 986         }
 987
 988         /*
 989          * We default some events to have a default interval. But keep
 990          * it a weak assumption overridable by the user.
 991          */
 992         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 993                                      opts->user_interval != ULLONG_MAX)) {
 994                 if (opts->freq) {
 995                         perf_evsel__set_sample_bit(evsel, PERIOD);
 996                         attr->freq              = 1;
 997                         attr->sample_freq       = opts->freq;
 998                 } else {
 999                         attr->sample_period = opts->default_interval;
1000                 }
1001         }
1002
1003         /*
1004          * Disable sampling for all group members other
1005          * than leader in case leader 'leads' the sampling.
1006          */
1007         if ((leader != evsel) && leader->sample_read) {
1008                 attr->freq           = 0;
1009                 attr->sample_freq    = 0;
1010                 attr->sample_period  = 0;
1011                 attr->write_backward = 0;
1012
1013                 /*
1014                  * We don't get sample for slave events, we make them
1015                  * when delivering group leader sample. Set the slave
1016                  * event to follow the master sample_type to ease up
1017                  * report.
1018                  */
1019                 attr->sample_type = leader->core.attr.sample_type;
1020         }
1021
1022         if (opts->no_samples)
1023                 attr->sample_freq = 0;
1024
1025         if (opts->inherit_stat) {
1026                 evsel->core.attr.read_format |=
1027                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1028                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1029                         PERF_FORMAT_ID;
1030                 attr->inherit_stat = 1;
1031         }
1032
1033         if (opts->sample_address) {
1034                 perf_evsel__set_sample_bit(evsel, ADDR);
1035                 attr->mmap_data = track;
1036         }
1037
1038         /*
1039          * We don't allow user space callchains for  function trace
1040          * event, due to issues with page faults while tracing page
1041          * fault handler and its overall trickiness nature.
1042          */
1043         if (perf_evsel__is_function_event(evsel))
1044                 evsel->core.attr.exclude_callchain_user = 1;
1045
1046         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1047                 perf_evsel__config_callchain(evsel, opts, callchain);
1048
1049         if (opts->sample_intr_regs) {
1050                 attr->sample_regs_intr = opts->sample_intr_regs;
1051                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1052         }
1053
1054         if (opts->sample_user_regs) {
1055                 attr->sample_regs_user |= opts->sample_user_regs;
1056                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1057         }
1058
1059         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1060                 perf_evsel__set_sample_bit(evsel, CPU);
1061
1062         /*
1063          * When the user explicitly disabled time don't force it here.
1064          */
1065         if (opts->sample_time &&
1066             (!perf_missing_features.sample_id_all &&
1067             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1068              opts->sample_time_set)))
1069                 perf_evsel__set_sample_bit(evsel, TIME);
1070
1071         if (opts->raw_samples && !evsel->no_aux_samples) {
1072                 perf_evsel__set_sample_bit(evsel, TIME);
1073                 perf_evsel__set_sample_bit(evsel, RAW);
1074                 perf_evsel__set_sample_bit(evsel, CPU);
1075         }
1076
1077         if (opts->sample_address)
1078                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1079
1080         if (opts->sample_phys_addr)
1081                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1082
1083         if (opts->no_buffering) {
1084                 attr->watermark = 0;
1085                 attr->wakeup_events = 1;
1086         }
1087         if (opts->branch_stack && !evsel->no_aux_samples) {
1088                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1089                 attr->branch_sample_type = opts->branch_stack;
1090         }
1091
1092         if (opts->sample_weight)
1093                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1094
1095         attr->task  = track;
1096         attr->mmap  = track;
1097         attr->mmap2 = track && !perf_missing_features.mmap2;
1098         attr->comm  = track;
1099         attr->ksymbol = track && !perf_missing_features.ksymbol;
1100         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1101
1102         if (opts->record_namespaces)
1103                 attr->namespaces  = track;
1104
1105         if (opts->record_switch_events)
1106                 attr->context_switch = track;
1107
1108         if (opts->sample_transaction)
1109                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1110
1111         if (opts->running_time) {
1112                 evsel->core.attr.read_format |=
1113                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1114                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1115         }
1116
1117         /*
1118          * XXX see the function comment above
1119          *
1120          * Disabling only independent events or group leaders,
1121          * keeping group members enabled.
1122          */
1123         if (perf_evsel__is_group_leader(evsel))
1124                 attr->disabled = 1;
1125
1126         /*
1127          * Setting enable_on_exec for independent events and
1128          * group leaders for traced executed by perf.
1129          */
1130         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1131                 !opts->initial_delay)
1132                 attr->enable_on_exec = 1;
1133
1134         if (evsel->immediate) {
1135                 attr->disabled = 0;
1136                 attr->enable_on_exec = 0;
1137         }
1138
1139         clockid = opts->clockid;
1140         if (opts->use_clockid) {
1141                 attr->use_clockid = 1;
1142                 attr->clockid = opts->clockid;
1143         }
1144
1145         if (evsel->precise_max)
1146                 attr->precise_ip = 3;
1147
1148         if (opts->all_user) {
1149                 attr->exclude_kernel = 1;
1150                 attr->exclude_user   = 0;
1151         }
1152
1153         if (opts->all_kernel) {
1154                 attr->exclude_kernel = 0;
1155                 attr->exclude_user   = 1;
1156         }
1157
1158         if (evsel->core.own_cpus || evsel->unit)
1159                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1160
1161         /*
1162          * Apply event specific term settings,
1163          * it overloads any global configuration.
1164          */
1165         apply_config_terms(evsel, opts, track);
1166
1167         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1168
1169         /* The --period option takes the precedence. */
1170         if (opts->period_set) {
1171                 if (opts->period)
1172                         perf_evsel__set_sample_bit(evsel, PERIOD);
1173                 else
1174                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1175         }
1176
1177         /*
1178          * For initial_delay, a dummy event is added implicitly.
1179          * The software event will trigger -EOPNOTSUPP error out,
1180          * if BRANCH_STACK bit is set.
1181          */
1182         if (opts->initial_delay && is_dummy_event(evsel))
1183                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1184 }
1185
1186 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1187 {
1188         char *new_filter = strdup(filter);
1189
1190         if (new_filter != NULL) {
1191                 free(evsel->filter);
1192                 evsel->filter = new_filter;
1193                 return 0;
1194         }
1195
1196         return -1;
1197 }
1198
1199 static int perf_evsel__append_filter(struct evsel *evsel,
1200                                      const char *fmt, const char *filter)
1201 {
1202         char *new_filter;
1203
1204         if (evsel->filter == NULL)
1205                 return perf_evsel__set_filter(evsel, filter);
1206
1207         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1208                 free(evsel->filter);
1209                 evsel->filter = new_filter;
1210                 return 0;
1211         }
1212
1213         return -1;
1214 }
1215
1216 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1217 {
1218         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1219 }
1220
1221 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1222 {
1223         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1224 }
1225
1226 /* Caller has to clear disabled after going through all CPUs. */
1227 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1228 {
1229         return perf_evsel__enable_cpu(&evsel->core, cpu);
1230 }
1231
1232 int evsel__enable(struct evsel *evsel)
1233 {
1234         int err = perf_evsel__enable(&evsel->core);
1235
1236         if (!err)
1237                 evsel->disabled = false;
1238         return err;
1239 }
1240
1241 /* Caller has to set disabled after going through all CPUs. */
1242 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1243 {
1244         return perf_evsel__disable_cpu(&evsel->core, cpu);
1245 }
1246
1247 int evsel__disable(struct evsel *evsel)
1248 {
1249         int err = perf_evsel__disable(&evsel->core);
1250         /*
1251          * We mark it disabled here so that tools that disable a event can
1252          * ignore events after they disable it. I.e. the ring buffer may have
1253          * already a few more events queued up before the kernel got the stop
1254          * request.
1255          */
1256         if (!err)
1257                 evsel->disabled = true;
1258
1259         return err;
1260 }
1261
1262 static void perf_evsel__free_config_terms(struct evsel *evsel)
1263 {
1264         struct perf_evsel_config_term *term, *h;
1265
1266         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1267                 list_del_init(&term->list);
1268                 if (term->free_str)
1269                         zfree(&term->val.str);
1270                 free(term);
1271         }
1272 }
1273
1274 void perf_evsel__exit(struct evsel *evsel)
1275 {
1276         assert(list_empty(&evsel->core.node));
1277         assert(evsel->evlist == NULL);
1278         perf_evsel__free_counts(evsel);
1279         perf_evsel__free_fd(&evsel->core);
1280         perf_evsel__free_id(&evsel->core);
1281         perf_evsel__free_config_terms(evsel);
1282         cgroup__put(evsel->cgrp);
1283         perf_cpu_map__put(evsel->core.cpus);
1284         perf_cpu_map__put(evsel->core.own_cpus);
1285         perf_thread_map__put(evsel->core.threads);
1286         zfree(&evsel->group_name);
1287         zfree(&evsel->name);
1288         perf_evsel__object.fini(evsel);
1289 }
1290
1291 void evsel__delete(struct evsel *evsel)
1292 {
1293         perf_evsel__exit(evsel);
1294         free(evsel);
1295 }
1296
1297 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1298                                 struct perf_counts_values *count)
1299 {
1300         struct perf_counts_values tmp;
1301
1302         if (!evsel->prev_raw_counts)
1303                 return;
1304
1305         if (cpu == -1) {
1306                 tmp = evsel->prev_raw_counts->aggr;
1307                 evsel->prev_raw_counts->aggr = *count;
1308         } else {
1309                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1310                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1311         }
1312
1313         count->val = count->val - tmp.val;
1314         count->ena = count->ena - tmp.ena;
1315         count->run = count->run - tmp.run;
1316 }
1317
1318 void perf_counts_values__scale(struct perf_counts_values *count,
1319                                bool scale, s8 *pscaled)
1320 {
1321         s8 scaled = 0;
1322
1323         if (scale) {
1324                 if (count->run == 0) {
1325                         scaled = -1;
1326                         count->val = 0;
1327                 } else if (count->run < count->ena) {
1328                         scaled = 1;
1329                         count->val = (u64)((double) count->val * count->ena / count->run);
1330                 }
1331         }
1332
1333         if (pscaled)
1334                 *pscaled = scaled;
1335 }
1336
1337 static int
1338 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1339 {
1340         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1341
1342         return perf_evsel__read(&evsel->core, cpu, thread, count);
1343 }
1344
1345 static void
1346 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1347                       u64 val, u64 ena, u64 run)
1348 {
1349         struct perf_counts_values *count;
1350
1351         count = perf_counts(counter->counts, cpu, thread);
1352
1353         count->val    = val;
1354         count->ena    = ena;
1355         count->run    = run;
1356
1357         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1358 }
1359
1360 static int
1361 perf_evsel__process_group_data(struct evsel *leader,
1362                                int cpu, int thread, u64 *data)
1363 {
1364         u64 read_format = leader->core.attr.read_format;
1365         struct sample_read_value *v;
1366         u64 nr, ena = 0, run = 0, i;
1367
1368         nr = *data++;
1369
1370         if (nr != (u64) leader->core.nr_members)
1371                 return -EINVAL;
1372
1373         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1374                 ena = *data++;
1375
1376         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1377                 run = *data++;
1378
1379         v = (struct sample_read_value *) data;
1380
1381         perf_evsel__set_count(leader, cpu, thread,
1382                               v[0].value, ena, run);
1383
1384         for (i = 1; i < nr; i++) {
1385                 struct evsel *counter;
1386
1387                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1388                 if (!counter)
1389                         return -EINVAL;
1390
1391                 perf_evsel__set_count(counter, cpu, thread,
1392                                       v[i].value, ena, run);
1393         }
1394
1395         return 0;
1396 }
1397
1398 static int
1399 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1400 {
1401         struct perf_stat_evsel *ps = leader->stats;
1402         u64 read_format = leader->core.attr.read_format;
1403         int size = perf_evsel__read_size(&leader->core);
1404         u64 *data = ps->group_data;
1405
1406         if (!(read_format & PERF_FORMAT_ID))
1407                 return -EINVAL;
1408
1409         if (!perf_evsel__is_group_leader(leader))
1410                 return -EINVAL;
1411
1412         if (!data) {
1413                 data = zalloc(size);
1414                 if (!data)
1415                         return -ENOMEM;
1416
1417                 ps->group_data = data;
1418         }
1419
1420         if (FD(leader, cpu, thread) < 0)
1421                 return -EINVAL;
1422
1423         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1424                 return -errno;
1425
1426         return perf_evsel__process_group_data(leader, cpu, thread, data);
1427 }
1428
1429 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1430 {
1431         u64 read_format = evsel->core.attr.read_format;
1432
1433         if (read_format & PERF_FORMAT_GROUP)
1434                 return perf_evsel__read_group(evsel, cpu, thread);
1435         else
1436                 return perf_evsel__read_one(evsel, cpu, thread);
1437 }
1438
1439 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1440                               int cpu, int thread, bool scale)
1441 {
1442         struct perf_counts_values count;
1443         size_t nv = scale ? 3 : 1;
1444
1445         if (FD(evsel, cpu, thread) < 0)
1446                 return -EINVAL;
1447
1448         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1449                 return -ENOMEM;
1450
1451         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1452                 return -errno;
1453
1454         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1455         perf_counts_values__scale(&count, scale, NULL);
1456         *perf_counts(evsel->counts, cpu, thread) = count;
1457         return 0;
1458 }
1459
1460 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1461 {
1462         struct evsel *leader = evsel->leader;
1463         int fd;
1464
1465         if (perf_evsel__is_group_leader(evsel))
1466                 return -1;
1467
1468         /*
1469          * Leader must be already processed/open,
1470          * if not it's a bug.
1471          */
1472         BUG_ON(!leader->core.fd);
1473
1474         fd = FD(leader, cpu, thread);
1475         BUG_ON(fd == -1);
1476
1477         return fd;
1478 }
1479
1480 static void perf_evsel__remove_fd(struct evsel *pos,
1481                                   int nr_cpus, int nr_threads,
1482                                   int thread_idx)
1483 {
1484         for (int cpu = 0; cpu < nr_cpus; cpu++)
1485                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1486                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1487 }
1488
1489 static int update_fds(struct evsel *evsel,
1490                       int nr_cpus, int cpu_idx,
1491                       int nr_threads, int thread_idx)
1492 {
1493         struct evsel *pos;
1494
1495         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1496                 return -EINVAL;
1497
1498         evlist__for_each_entry(evsel->evlist, pos) {
1499                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1500
1501                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1502
1503                 /*
1504                  * Since fds for next evsel has not been created,
1505                  * there is no need to iterate whole event list.
1506                  */
1507                 if (pos == evsel)
1508                         break;
1509         }
1510         return 0;
1511 }
1512
1513 static bool ignore_missing_thread(struct evsel *evsel,
1514                                   int nr_cpus, int cpu,
1515                                   struct perf_thread_map *threads,
1516                                   int thread, int err)
1517 {
1518         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1519
1520         if (!evsel->ignore_missing_thread)
1521                 return false;
1522
1523         /* The system wide setup does not work with threads. */
1524         if (evsel->core.system_wide)
1525                 return false;
1526
1527         /* The -ESRCH is perf event syscall errno for pid's not found. */
1528         if (err != -ESRCH)
1529                 return false;
1530
1531         /* If there's only one thread, let it fail. */
1532         if (threads->nr == 1)
1533                 return false;
1534
1535         /*
1536          * We should remove fd for missing_thread first
1537          * because thread_map__remove() will decrease threads->nr.
1538          */
1539         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1540                 return false;
1541
1542         if (thread_map__remove(threads, thread))
1543                 return false;
1544
1545         pr_warning("WARNING: Ignored open failure for pid %d\n",
1546                    ignore_pid);
1547         return true;
1548 }
1549
1550 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1551                                 void *priv __maybe_unused)
1552 {
1553         return fprintf(fp, "  %-32s %s\n", name, val);
1554 }
1555
1556 static void display_attr(struct perf_event_attr *attr)
1557 {
1558         if (verbose >= 2 || debug_peo_args) {
1559                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1560                 fprintf(stderr, "perf_event_attr:\n");
1561                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1562                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1563         }
1564 }
1565
1566 static int perf_event_open(struct evsel *evsel,
1567                            pid_t pid, int cpu, int group_fd,
1568                            unsigned long flags)
1569 {
1570         int precise_ip = evsel->core.attr.precise_ip;
1571         int fd;
1572
1573         while (1) {
1574                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1575                           pid, cpu, group_fd, flags);
1576
1577                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1578                 if (fd >= 0)
1579                         break;
1580
1581                 /* Do not try less precise if not requested. */
1582                 if (!evsel->precise_max)
1583                         break;
1584
1585                 /*
1586                  * We tried all the precise_ip values, and it's
1587                  * still failing, so leave it to standard fallback.
1588                  */
1589                 if (!evsel->core.attr.precise_ip) {
1590                         evsel->core.attr.precise_ip = precise_ip;
1591                         break;
1592                 }
1593
1594                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1595                 evsel->core.attr.precise_ip--;
1596                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1597                 display_attr(&evsel->core.attr);
1598         }
1599
1600         return fd;
1601 }
1602
1603 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1604                 struct perf_thread_map *threads,
1605                 int start_cpu, int end_cpu)
1606 {
1607         int cpu, thread, nthreads;
1608         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1609         int pid = -1, err, old_errno;
1610         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1611
1612         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1613             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1614                 return -EINVAL;
1615
1616         if (cpus == NULL) {
1617                 static struct perf_cpu_map *empty_cpu_map;
1618
1619                 if (empty_cpu_map == NULL) {
1620                         empty_cpu_map = perf_cpu_map__dummy_new();
1621                         if (empty_cpu_map == NULL)
1622                                 return -ENOMEM;
1623                 }
1624
1625                 cpus = empty_cpu_map;
1626         }
1627
1628         if (threads == NULL) {
1629                 static struct perf_thread_map *empty_thread_map;
1630
1631                 if (empty_thread_map == NULL) {
1632                         empty_thread_map = thread_map__new_by_tid(-1);
1633                         if (empty_thread_map == NULL)
1634                                 return -ENOMEM;
1635                 }
1636
1637                 threads = empty_thread_map;
1638         }
1639
1640         if (evsel->core.system_wide)
1641                 nthreads = 1;
1642         else
1643                 nthreads = threads->nr;
1644
1645         if (evsel->core.fd == NULL &&
1646             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1647                 return -ENOMEM;
1648
1649         if (evsel->cgrp) {
1650                 flags |= PERF_FLAG_PID_CGROUP;
1651                 pid = evsel->cgrp->fd;
1652         }
1653
1654 fallback_missing_features:
1655         if (perf_missing_features.clockid_wrong)
1656                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1657         if (perf_missing_features.clockid) {
1658                 evsel->core.attr.use_clockid = 0;
1659                 evsel->core.attr.clockid = 0;
1660         }
1661         if (perf_missing_features.cloexec)
1662                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1663         if (perf_missing_features.mmap2)
1664                 evsel->core.attr.mmap2 = 0;
1665         if (perf_missing_features.exclude_guest)
1666                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1667         if (perf_missing_features.lbr_flags)
1668                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1669                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1670         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1671                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1672         if (perf_missing_features.ksymbol)
1673                 evsel->core.attr.ksymbol = 0;
1674         if (perf_missing_features.bpf)
1675                 evsel->core.attr.bpf_event = 0;
1676 retry_sample_id:
1677         if (perf_missing_features.sample_id_all)
1678                 evsel->core.attr.sample_id_all = 0;
1679
1680         display_attr(&evsel->core.attr);
1681
1682         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1683
1684                 for (thread = 0; thread < nthreads; thread++) {
1685                         int fd, group_fd;
1686
1687                         if (!evsel->cgrp && !evsel->core.system_wide)
1688                                 pid = perf_thread_map__pid(threads, thread);
1689
1690                         group_fd = get_group_fd(evsel, cpu, thread);
1691 retry_open:
1692                         test_attr__ready();
1693
1694                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1695                                              group_fd, flags);
1696
1697                         FD(evsel, cpu, thread) = fd;
1698
1699                         if (fd < 0) {
1700                                 err = -errno;
1701
1702                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1703                                         /*
1704                                          * We just removed 1 thread, so take a step
1705                                          * back on thread index and lower the upper
1706                                          * nthreads limit.
1707                                          */
1708                                         nthreads--;
1709                                         thread--;
1710
1711                                         /* ... and pretend like nothing have happened. */
1712                                         err = 0;
1713                                         continue;
1714                                 }
1715
1716                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1717                                           err);
1718                                 goto try_fallback;
1719                         }
1720
1721                         pr_debug2_peo(" = %d\n", fd);
1722
1723                         if (evsel->bpf_fd >= 0) {
1724                                 int evt_fd = fd;
1725                                 int bpf_fd = evsel->bpf_fd;
1726
1727                                 err = ioctl(evt_fd,
1728                                             PERF_EVENT_IOC_SET_BPF,
1729                                             bpf_fd);
1730                                 if (err && errno != EEXIST) {
1731                                         pr_err("failed to attach bpf fd %d: %s\n",
1732                                                bpf_fd, strerror(errno));
1733                                         err = -EINVAL;
1734                                         goto out_close;
1735                                 }
1736                         }
1737
1738                         set_rlimit = NO_CHANGE;
1739
1740                         /*
1741                          * If we succeeded but had to kill clockid, fail and
1742                          * have perf_evsel__open_strerror() print us a nice
1743                          * error.
1744                          */
1745                         if (perf_missing_features.clockid ||
1746                             perf_missing_features.clockid_wrong) {
1747                                 err = -EINVAL;
1748                                 goto out_close;
1749                         }
1750                 }
1751         }
1752
1753         return 0;
1754
1755 try_fallback:
1756         /*
1757          * perf stat needs between 5 and 22 fds per CPU. When we run out
1758          * of them try to increase the limits.
1759          */
1760         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1761                 struct rlimit l;
1762
1763                 old_errno = errno;
1764                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1765                         if (set_rlimit == NO_CHANGE)
1766                                 l.rlim_cur = l.rlim_max;
1767                         else {
1768                                 l.rlim_cur = l.rlim_max + 1000;
1769                                 l.rlim_max = l.rlim_cur;
1770                         }
1771                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1772                                 set_rlimit++;
1773                                 errno = old_errno;
1774                                 goto retry_open;
1775                         }
1776                 }
1777                 errno = old_errno;
1778         }
1779
1780         if (err != -EINVAL || cpu > 0 || thread > 0)
1781                 goto out_close;
1782
1783         /*
1784          * Must probe features in the order they were added to the
1785          * perf_event_attr interface.
1786          */
1787         if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1788                 perf_missing_features.aux_output = true;
1789                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1790                 goto out_close;
1791         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1792                 perf_missing_features.bpf = true;
1793                 pr_debug2_peo("switching off bpf_event\n");
1794                 goto fallback_missing_features;
1795         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1796                 perf_missing_features.ksymbol = true;
1797                 pr_debug2_peo("switching off ksymbol\n");
1798                 goto fallback_missing_features;
1799         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1800                 perf_missing_features.write_backward = true;
1801                 pr_debug2_peo("switching off write_backward\n");
1802                 goto out_close;
1803         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1804                 perf_missing_features.clockid_wrong = true;
1805                 pr_debug2_peo("switching off clockid\n");
1806                 goto fallback_missing_features;
1807         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1808                 perf_missing_features.clockid = true;
1809                 pr_debug2_peo("switching off use_clockid\n");
1810                 goto fallback_missing_features;
1811         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1812                 perf_missing_features.cloexec = true;
1813                 pr_debug2_peo("switching off cloexec flag\n");
1814                 goto fallback_missing_features;
1815         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1816                 perf_missing_features.mmap2 = true;
1817                 pr_debug2_peo("switching off mmap2\n");
1818                 goto fallback_missing_features;
1819         } else if (!perf_missing_features.exclude_guest &&
1820                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1821                 perf_missing_features.exclude_guest = true;
1822                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1823                 goto fallback_missing_features;
1824         } else if (!perf_missing_features.sample_id_all) {
1825                 perf_missing_features.sample_id_all = true;
1826                 pr_debug2_peo("switching off sample_id_all\n");
1827                 goto retry_sample_id;
1828         } else if (!perf_missing_features.lbr_flags &&
1829                         (evsel->core.attr.branch_sample_type &
1830                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1831                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1832                 perf_missing_features.lbr_flags = true;
1833                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1834                 goto fallback_missing_features;
1835         } else if (!perf_missing_features.group_read &&
1836                     evsel->core.attr.inherit &&
1837                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1838                    perf_evsel__is_group_leader(evsel)) {
1839                 perf_missing_features.group_read = true;
1840                 pr_debug2_peo("switching off group read\n");
1841                 goto fallback_missing_features;
1842         }
1843 out_close:
1844         if (err)
1845                 threads->err_thread = thread;
1846
1847         old_errno = errno;
1848         do {
1849                 while (--thread >= 0) {
1850                         if (FD(evsel, cpu, thread) >= 0)
1851                                 close(FD(evsel, cpu, thread));
1852                         FD(evsel, cpu, thread) = -1;
1853                 }
1854                 thread = nthreads;
1855         } while (--cpu >= 0);
1856         errno = old_errno;
1857         return err;
1858 }
1859
1860 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1861                 struct perf_thread_map *threads)
1862 {
1863         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1864 }
1865
1866 void evsel__close(struct evsel *evsel)
1867 {
1868         perf_evsel__close(&evsel->core);
1869         perf_evsel__free_id(&evsel->core);
1870 }
1871
1872 int perf_evsel__open_per_cpu(struct evsel *evsel,
1873                              struct perf_cpu_map *cpus,
1874                              int cpu)
1875 {
1876         if (cpu == -1)
1877                 return evsel__open_cpu(evsel, cpus, NULL, 0,
1878                                         cpus ? cpus->nr : 1);
1879
1880         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1881 }
1882
1883 int perf_evsel__open_per_thread(struct evsel *evsel,
1884                                 struct perf_thread_map *threads)
1885 {
1886         return evsel__open(evsel, NULL, threads);
1887 }
1888
1889 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1890                                        const union perf_event *event,
1891                                        struct perf_sample *sample)
1892 {
1893         u64 type = evsel->core.attr.sample_type;
1894         const __u64 *array = event->sample.array;
1895         bool swapped = evsel->needs_swap;
1896         union u64_swap u;
1897
1898         array += ((event->header.size -
1899                    sizeof(event->header)) / sizeof(u64)) - 1;
1900
1901         if (type & PERF_SAMPLE_IDENTIFIER) {
1902                 sample->id = *array;
1903                 array--;
1904         }
1905
1906         if (type & PERF_SAMPLE_CPU) {
1907                 u.val64 = *array;
1908                 if (swapped) {
1909                         /* undo swap of u64, then swap on individual u32s */
1910                         u.val64 = bswap_64(u.val64);
1911                         u.val32[0] = bswap_32(u.val32[0]);
1912                 }
1913
1914                 sample->cpu = u.val32[0];
1915                 array--;
1916         }
1917
1918         if (type & PERF_SAMPLE_STREAM_ID) {
1919                 sample->stream_id = *array;
1920                 array--;
1921         }
1922
1923         if (type & PERF_SAMPLE_ID) {
1924                 sample->id = *array;
1925                 array--;
1926         }
1927
1928         if (type & PERF_SAMPLE_TIME) {
1929                 sample->time = *array;
1930                 array--;
1931         }
1932
1933         if (type & PERF_SAMPLE_TID) {
1934                 u.val64 = *array;
1935                 if (swapped) {
1936                         /* undo swap of u64, then swap on individual u32s */
1937                         u.val64 = bswap_64(u.val64);
1938                         u.val32[0] = bswap_32(u.val32[0]);
1939                         u.val32[1] = bswap_32(u.val32[1]);
1940                 }
1941
1942                 sample->pid = u.val32[0];
1943                 sample->tid = u.val32[1];
1944                 array--;
1945         }
1946
1947         return 0;
1948 }
1949
1950 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1951                             u64 size)
1952 {
1953         return size > max_size || offset + size > endp;
1954 }
1955
1956 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1957         do {                                                            \
1958                 if (overflow(endp, (max_size), (offset), (size)))       \
1959                         return -EFAULT;                                 \
1960         } while (0)
1961
1962 #define OVERFLOW_CHECK_u64(offset) \
1963         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1964
1965 static int
1966 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1967 {
1968         /*
1969          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1970          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1971          * check the format does not go past the end of the event.
1972          */
1973         if (sample_size + sizeof(event->header) > event->header.size)
1974                 return -EFAULT;
1975
1976         return 0;
1977 }
1978
1979 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1980                              struct perf_sample *data)
1981 {
1982         u64 type = evsel->core.attr.sample_type;
1983         bool swapped = evsel->needs_swap;
1984         const __u64 *array;
1985         u16 max_size = event->header.size;
1986         const void *endp = (void *)event + max_size;
1987         u64 sz;
1988
1989         /*
1990          * used for cross-endian analysis. See git commit 65014ab3
1991          * for why this goofiness is needed.
1992          */
1993         union u64_swap u;
1994
1995         memset(data, 0, sizeof(*data));
1996         data->cpu = data->pid = data->tid = -1;
1997         data->stream_id = data->id = data->time = -1ULL;
1998         data->period = evsel->core.attr.sample_period;
1999         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2000         data->misc    = event->header.misc;
2001         data->id = -1ULL;
2002         data->data_src = PERF_MEM_DATA_SRC_NONE;
2003
2004         if (event->header.type != PERF_RECORD_SAMPLE) {
2005                 if (!evsel->core.attr.sample_id_all)
2006                         return 0;
2007                 return perf_evsel__parse_id_sample(evsel, event, data);
2008         }
2009
2010         array = event->sample.array;
2011
2012         if (perf_event__check_size(event, evsel->sample_size))
2013                 return -EFAULT;
2014
2015         if (type & PERF_SAMPLE_IDENTIFIER) {
2016                 data->id = *array;
2017                 array++;
2018         }
2019
2020         if (type & PERF_SAMPLE_IP) {
2021                 data->ip = *array;
2022                 array++;
2023         }
2024
2025         if (type & PERF_SAMPLE_TID) {
2026                 u.val64 = *array;
2027                 if (swapped) {
2028                         /* undo swap of u64, then swap on individual u32s */
2029                         u.val64 = bswap_64(u.val64);
2030                         u.val32[0] = bswap_32(u.val32[0]);
2031                         u.val32[1] = bswap_32(u.val32[1]);
2032                 }
2033
2034                 data->pid = u.val32[0];
2035                 data->tid = u.val32[1];
2036                 array++;
2037         }
2038
2039         if (type & PERF_SAMPLE_TIME) {
2040                 data->time = *array;
2041                 array++;
2042         }
2043
2044         if (type & PERF_SAMPLE_ADDR) {
2045                 data->addr = *array;
2046                 array++;
2047         }
2048
2049         if (type & PERF_SAMPLE_ID) {
2050                 data->id = *array;
2051                 array++;
2052         }
2053
2054         if (type & PERF_SAMPLE_STREAM_ID) {
2055                 data->stream_id = *array;
2056                 array++;
2057         }
2058
2059         if (type & PERF_SAMPLE_CPU) {
2060
2061                 u.val64 = *array;
2062                 if (swapped) {
2063                         /* undo swap of u64, then swap on individual u32s */
2064                         u.val64 = bswap_64(u.val64);
2065                         u.val32[0] = bswap_32(u.val32[0]);
2066                 }
2067
2068                 data->cpu = u.val32[0];
2069                 array++;
2070         }
2071
2072         if (type & PERF_SAMPLE_PERIOD) {
2073                 data->period = *array;
2074                 array++;
2075         }
2076
2077         if (type & PERF_SAMPLE_READ) {
2078                 u64 read_format = evsel->core.attr.read_format;
2079
2080                 OVERFLOW_CHECK_u64(array);
2081                 if (read_format & PERF_FORMAT_GROUP)
2082                         data->read.group.nr = *array;
2083                 else
2084                         data->read.one.value = *array;
2085
2086                 array++;
2087
2088                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2089                         OVERFLOW_CHECK_u64(array);
2090                         data->read.time_enabled = *array;
2091                         array++;
2092                 }
2093
2094                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2095                         OVERFLOW_CHECK_u64(array);
2096                         data->read.time_running = *array;
2097                         array++;
2098                 }
2099
2100                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2101                 if (read_format & PERF_FORMAT_GROUP) {
2102                         const u64 max_group_nr = UINT64_MAX /
2103                                         sizeof(struct sample_read_value);
2104
2105                         if (data->read.group.nr > max_group_nr)
2106                                 return -EFAULT;
2107                         sz = data->read.group.nr *
2108                              sizeof(struct sample_read_value);
2109                         OVERFLOW_CHECK(array, sz, max_size);
2110                         data->read.group.values =
2111                                         (struct sample_read_value *)array;
2112                         array = (void *)array + sz;
2113                 } else {
2114                         OVERFLOW_CHECK_u64(array);
2115                         data->read.one.id = *array;
2116                         array++;
2117                 }
2118         }
2119
2120         if (evsel__has_callchain(evsel)) {
2121                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2122
2123                 OVERFLOW_CHECK_u64(array);
2124                 data->callchain = (struct ip_callchain *)array++;
2125                 if (data->callchain->nr > max_callchain_nr)
2126                         return -EFAULT;
2127                 sz = data->callchain->nr * sizeof(u64);
2128                 OVERFLOW_CHECK(array, sz, max_size);
2129                 array = (void *)array + sz;
2130         }
2131
2132         if (type & PERF_SAMPLE_RAW) {
2133                 OVERFLOW_CHECK_u64(array);
2134                 u.val64 = *array;
2135
2136                 /*
2137                  * Undo swap of u64, then swap on individual u32s,
2138                  * get the size of the raw area and undo all of the
2139                  * swap. The pevent interface handles endianity by
2140                  * itself.
2141                  */
2142                 if (swapped) {
2143                         u.val64 = bswap_64(u.val64);
2144                         u.val32[0] = bswap_32(u.val32[0]);
2145                         u.val32[1] = bswap_32(u.val32[1]);
2146                 }
2147                 data->raw_size = u.val32[0];
2148
2149                 /*
2150                  * The raw data is aligned on 64bits including the
2151                  * u32 size, so it's safe to use mem_bswap_64.
2152                  */
2153                 if (swapped)
2154                         mem_bswap_64((void *) array, data->raw_size);
2155
2156                 array = (void *)array + sizeof(u32);
2157
2158                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2159                 data->raw_data = (void *)array;
2160                 array = (void *)array + data->raw_size;
2161         }
2162
2163         if (type & PERF_SAMPLE_BRANCH_STACK) {
2164                 const u64 max_branch_nr = UINT64_MAX /
2165                                           sizeof(struct branch_entry);
2166
2167                 OVERFLOW_CHECK_u64(array);
2168                 data->branch_stack = (struct branch_stack *)array++;
2169
2170                 if (data->branch_stack->nr > max_branch_nr)
2171                         return -EFAULT;
2172                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2173                 OVERFLOW_CHECK(array, sz, max_size);
2174                 array = (void *)array + sz;
2175         }
2176
2177         if (type & PERF_SAMPLE_REGS_USER) {
2178                 OVERFLOW_CHECK_u64(array);
2179                 data->user_regs.abi = *array;
2180                 array++;
2181
2182                 if (data->user_regs.abi) {
2183                         u64 mask = evsel->core.attr.sample_regs_user;
2184
2185                         sz = hweight64(mask) * sizeof(u64);
2186                         OVERFLOW_CHECK(array, sz, max_size);
2187                         data->user_regs.mask = mask;
2188                         data->user_regs.regs = (u64 *)array;
2189                         array = (void *)array + sz;
2190                 }
2191         }
2192
2193         if (type & PERF_SAMPLE_STACK_USER) {
2194                 OVERFLOW_CHECK_u64(array);
2195                 sz = *array++;
2196
2197                 data->user_stack.offset = ((char *)(array - 1)
2198                                           - (char *) event);
2199
2200                 if (!sz) {
2201                         data->user_stack.size = 0;
2202                 } else {
2203                         OVERFLOW_CHECK(array, sz, max_size);
2204                         data->user_stack.data = (char *)array;
2205                         array = (void *)array + sz;
2206                         OVERFLOW_CHECK_u64(array);
2207                         data->user_stack.size = *array++;
2208                         if (WARN_ONCE(data->user_stack.size > sz,
2209                                       "user stack dump failure\n"))
2210                                 return -EFAULT;
2211                 }
2212         }
2213
2214         if (type & PERF_SAMPLE_WEIGHT) {
2215                 OVERFLOW_CHECK_u64(array);
2216                 data->weight = *array;
2217                 array++;
2218         }
2219
2220         if (type & PERF_SAMPLE_DATA_SRC) {
2221                 OVERFLOW_CHECK_u64(array);
2222                 data->data_src = *array;
2223                 array++;
2224         }
2225
2226         if (type & PERF_SAMPLE_TRANSACTION) {
2227                 OVERFLOW_CHECK_u64(array);
2228                 data->transaction = *array;
2229                 array++;
2230         }
2231
2232         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2233         if (type & PERF_SAMPLE_REGS_INTR) {
2234                 OVERFLOW_CHECK_u64(array);
2235                 data->intr_regs.abi = *array;
2236                 array++;
2237
2238                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2239                         u64 mask = evsel->core.attr.sample_regs_intr;
2240
2241                         sz = hweight64(mask) * sizeof(u64);
2242                         OVERFLOW_CHECK(array, sz, max_size);
2243                         data->intr_regs.mask = mask;
2244                         data->intr_regs.regs = (u64 *)array;
2245                         array = (void *)array + sz;
2246                 }
2247         }
2248
2249         data->phys_addr = 0;
2250         if (type & PERF_SAMPLE_PHYS_ADDR) {
2251                 data->phys_addr = *array;
2252                 array++;
2253         }
2254
2255         if (type & PERF_SAMPLE_AUX) {
2256                 OVERFLOW_CHECK_u64(array);
2257                 sz = *array++;
2258
2259                 OVERFLOW_CHECK(array, sz, max_size);
2260                 /* Undo swap of data */
2261                 if (swapped)
2262                         mem_bswap_64((char *)array, sz);
2263                 data->aux_sample.size = sz;
2264                 data->aux_sample.data = (char *)array;
2265                 array = (void *)array + sz;
2266         }
2267
2268         return 0;
2269 }
2270
2271 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2272                                        union perf_event *event,
2273                                        u64 *timestamp)
2274 {
2275         u64 type = evsel->core.attr.sample_type;
2276         const __u64 *array;
2277
2278         if (!(type & PERF_SAMPLE_TIME))
2279                 return -1;
2280
2281         if (event->header.type != PERF_RECORD_SAMPLE) {
2282                 struct perf_sample data = {
2283                         .time = -1ULL,
2284                 };
2285
2286                 if (!evsel->core.attr.sample_id_all)
2287                         return -1;
2288                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2289                         return -1;
2290
2291                 *timestamp = data.time;
2292                 return 0;
2293         }
2294
2295         array = event->sample.array;
2296
2297         if (perf_event__check_size(event, evsel->sample_size))
2298                 return -EFAULT;
2299
2300         if (type & PERF_SAMPLE_IDENTIFIER)
2301                 array++;
2302
2303         if (type & PERF_SAMPLE_IP)
2304                 array++;
2305
2306         if (type & PERF_SAMPLE_TID)
2307                 array++;
2308
2309         if (type & PERF_SAMPLE_TIME)
2310                 *timestamp = *array;
2311
2312         return 0;
2313 }
2314
2315 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2316 {
2317         return tep_find_field(evsel->tp_format, name);
2318 }
2319
2320 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2321                          const char *name)
2322 {
2323         struct tep_format_field *field = perf_evsel__field(evsel, name);
2324         int offset;
2325
2326         if (!field)
2327                 return NULL;
2328
2329         offset = field->offset;
2330
2331         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2332                 offset = *(int *)(sample->raw_data + field->offset);
2333                 offset &= 0xffff;
2334         }
2335
2336         return sample->raw_data + offset;
2337 }
2338
2339 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2340                          bool needs_swap)
2341 {
2342         u64 value;
2343         void *ptr = sample->raw_data + field->offset;
2344
2345         switch (field->size) {
2346         case 1:
2347                 return *(u8 *)ptr;
2348         case 2:
2349                 value = *(u16 *)ptr;
2350                 break;
2351         case 4:
2352                 value = *(u32 *)ptr;
2353                 break;
2354         case 8:
2355                 memcpy(&value, ptr, sizeof(u64));
2356                 break;
2357         default:
2358                 return 0;
2359         }
2360
2361         if (!needs_swap)
2362                 return value;
2363
2364         switch (field->size) {
2365         case 2:
2366                 return bswap_16(value);
2367         case 4:
2368                 return bswap_32(value);
2369         case 8:
2370                 return bswap_64(value);
2371         default:
2372                 return 0;
2373         }
2374
2375         return 0;
2376 }
2377
2378 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2379                        const char *name)
2380 {
2381         struct tep_format_field *field = perf_evsel__field(evsel, name);
2382
2383         if (!field)
2384                 return 0;
2385
2386         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2387 }
2388
2389 bool perf_evsel__fallback(struct evsel *evsel, int err,
2390                           char *msg, size_t msgsize)
2391 {
2392         int paranoid;
2393
2394         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2395             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2396             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2397                 /*
2398                  * If it's cycles then fall back to hrtimer based
2399                  * cpu-clock-tick sw counter, which is always available even if
2400                  * no PMU support.
2401                  *
2402                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2403                  * b0a873e).
2404                  */
2405                 scnprintf(msg, msgsize, "%s",
2406 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2407
2408                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2409                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2410
2411                 zfree(&evsel->name);
2412                 return true;
2413         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2414                    (paranoid = perf_event_paranoid()) > 1) {
2415                 const char *name = perf_evsel__name(evsel);
2416                 char *new_name;
2417                 const char *sep = ":";
2418
2419                 /* Is there already the separator in the name. */
2420                 if (strchr(name, '/') ||
2421                     strchr(name, ':'))
2422                         sep = "";
2423
2424                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2425                         return false;
2426
2427                 if (evsel->name)
2428                         free(evsel->name);
2429                 evsel->name = new_name;
2430                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2431                           "to fall back to excluding kernel and hypervisor "
2432                           " samples", paranoid);
2433                 evsel->core.attr.exclude_kernel = 1;
2434                 evsel->core.attr.exclude_hv     = 1;
2435
2436                 return true;
2437         }
2438
2439         return false;
2440 }
2441
2442 static bool find_process(const char *name)
2443 {
2444         size_t len = strlen(name);
2445         DIR *dir;
2446         struct dirent *d;
2447         int ret = -1;
2448
2449         dir = opendir(procfs__mountpoint());
2450         if (!dir)
2451                 return false;
2452
2453         /* Walk through the directory. */
2454         while (ret && (d = readdir(dir)) != NULL) {
2455                 char path[PATH_MAX];
2456                 char *data;
2457                 size_t size;
2458
2459                 if ((d->d_type != DT_DIR) ||
2460                      !strcmp(".", d->d_name) ||
2461                      !strcmp("..", d->d_name))
2462                         continue;
2463
2464                 scnprintf(path, sizeof(path), "%s/%s/comm",
2465                           procfs__mountpoint(), d->d_name);
2466
2467                 if (filename__read_str(path, &data, &size))
2468                         continue;
2469
2470                 ret = strncmp(name, data, len);
2471                 free(data);
2472         }
2473
2474         closedir(dir);
2475         return ret ? false : true;
2476 }
2477
2478 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2479                               int err, char *msg, size_t size)
2480 {
2481         char sbuf[STRERR_BUFSIZE];
2482         int printed = 0;
2483
2484         switch (err) {
2485         case EPERM:
2486         case EACCES:
2487                 if (err == EPERM)
2488                         printed = scnprintf(msg, size,
2489                                 "No permission to enable %s event.\n\n",
2490                                 perf_evsel__name(evsel));
2491
2492                 return scnprintf(msg + printed, size - printed,
2493                  "You may not have permission to collect %sstats.\n\n"
2494                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2495                  "which controls use of the performance events system by\n"
2496                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2497                  "The current value is %d:\n\n"
2498                  "  -1: Allow use of (almost) all events by all users\n"
2499                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2500                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2501                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2502                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2503                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2504                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2505                  "      kernel.perf_event_paranoid = -1\n" ,
2506                                  target->system_wide ? "system-wide " : "",
2507                                  perf_event_paranoid());
2508         case ENOENT:
2509                 return scnprintf(msg, size, "The %s event is not supported.",
2510                                  perf_evsel__name(evsel));
2511         case EMFILE:
2512                 return scnprintf(msg, size, "%s",
2513                          "Too many events are opened.\n"
2514                          "Probably the maximum number of open file descriptors has been reached.\n"
2515                          "Hint: Try again after reducing the number of events.\n"
2516                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2517         case ENOMEM:
2518                 if (evsel__has_callchain(evsel) &&
2519                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2520                         return scnprintf(msg, size,
2521                                          "Not enough memory to setup event with callchain.\n"
2522                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2523                                          "Hint: Current value: %d", sysctl__max_stack());
2524                 break;
2525         case ENODEV:
2526                 if (target->cpu_list)
2527                         return scnprintf(msg, size, "%s",
2528          "No such device - did you specify an out-of-range profile CPU?");
2529                 break;
2530         case EOPNOTSUPP:
2531                 if (evsel->core.attr.sample_period != 0)
2532                         return scnprintf(msg, size,
2533         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2534                                          perf_evsel__name(evsel));
2535                 if (evsel->core.attr.precise_ip)
2536                         return scnprintf(msg, size, "%s",
2537         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2538 #if defined(__i386__) || defined(__x86_64__)
2539                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2540                         return scnprintf(msg, size, "%s",
2541         "No hardware sampling interrupt available.\n");
2542 #endif
2543                 break;
2544         case EBUSY:
2545                 if (find_process("oprofiled"))
2546                         return scnprintf(msg, size,
2547         "The PMU counters are busy/taken by another profiler.\n"
2548         "We found oprofile daemon running, please stop it and try again.");
2549                 break;
2550         case EINVAL:
2551                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2552                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2553                 if (perf_missing_features.clockid)
2554                         return scnprintf(msg, size, "clockid feature not supported.");
2555                 if (perf_missing_features.clockid_wrong)
2556                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2557                 if (perf_missing_features.aux_output)
2558                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2559                 break;
2560         default:
2561                 break;
2562         }
2563
2564         return scnprintf(msg, size,
2565         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2566         "/bin/dmesg | grep -i perf may provide additional information.\n",
2567                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2568                          perf_evsel__name(evsel));
2569 }
2570
2571 struct perf_env *perf_evsel__env(struct evsel *evsel)
2572 {
2573         if (evsel && evsel->evlist)
2574                 return evsel->evlist->env;
2575         return &perf_env;
2576 }
2577
2578 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2579 {
2580         int cpu, thread;
2581
2582         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2583                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2584                      thread++) {
2585                         int fd = FD(evsel, cpu, thread);
2586
2587                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2588                                                    cpu, thread, fd) < 0)
2589                                 return -1;
2590                 }
2591         }
2592
2593         return 0;
2594 }
2595
2596 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2597 {
2598         struct perf_cpu_map *cpus = evsel->core.cpus;
2599         struct perf_thread_map *threads = evsel->core.threads;
2600
2601         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2602                 return -ENOMEM;
2603
2604         return store_evsel_ids(evsel, evlist);
2605 }