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 <sys/ioctl.h>
  21 #include <sys/resource.h>
  22 #include <sys/types.h>
  23 #include <dirent.h>
  24 #include "asm/bug.h"
  25 #include "callchain.h"
  26 #include "cgroup.h"
  27 #include "event.h"
  28 #include "evsel.h"
  29 #include "evlist.h"
  30 #include "util.h"
  31 #include "cpumap.h"
  32 #include "thread_map.h"
  33 #include "target.h"
  34 #include "perf_regs.h"
  35 #include "debug.h"
  36 #include "trace-event.h"
  37 #include "stat.h"
  38 #include "memswap.h"
  39 #include "util/parse-branch-options.h"
  40
  41 #include "sane_ctype.h"
  42
  43 struct perf_missing_features perf_missing_features;
  44
  45 static clockid_t clockid;
  46
  47 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  48 {
  49         return 0;
  50 }
  51
  52 void __weak test_attr__ready(void) { }
  53
  54 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  55 {
  56 }
  57
  58 static struct {
  59         size_t  size;
  60         int     (*init)(struct perf_evsel *evsel);
  61         void    (*fini)(struct perf_evsel *evsel);
  62 } perf_evsel__object = {
  63         .size = sizeof(struct perf_evsel),
  64         .init = perf_evsel__no_extra_init,
  65         .fini = perf_evsel__no_extra_fini,
  66 };
  67
  68 int perf_evsel__object_config(size_t object_size,
  69                               int (*init)(struct perf_evsel *evsel),
  70                               void (*fini)(struct perf_evsel *evsel))
  71 {
  72
  73         if (object_size == 0)
  74                 goto set_methods;
  75
  76         if (perf_evsel__object.size > object_size)
  77                 return -EINVAL;
  78
  79         perf_evsel__object.size = object_size;
  80
  81 set_methods:
  82         if (init != NULL)
  83                 perf_evsel__object.init = init;
  84
  85         if (fini != NULL)
  86                 perf_evsel__object.fini = fini;
  87
  88         return 0;
  89 }
  90
  91 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  92
  93 int __perf_evsel__sample_size(u64 sample_type)
  94 {
  95         u64 mask = sample_type & PERF_SAMPLE_MASK;
  96         int size = 0;
  97         int i;
  98
  99         for (i = 0; i < 64; i++) {
 100                 if (mask & (1ULL << i))
 101                         size++;
 102         }
 103
 104         size *= sizeof(u64);
 105
 106         return size;
 107 }
 108
 109 /**
 110  * __perf_evsel__calc_id_pos - calculate id_pos.
 111  * @sample_type: sample type
 112  *
 113  * This function returns the position of the event id (PERF_SAMPLE_ID or
 114  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 115  * sample_event.
 116  */
 117 static int __perf_evsel__calc_id_pos(u64 sample_type)
 118 {
 119         int idx = 0;
 120
 121         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 122                 return 0;
 123
 124         if (!(sample_type & PERF_SAMPLE_ID))
 125                 return -1;
 126
 127         if (sample_type & PERF_SAMPLE_IP)
 128                 idx += 1;
 129
 130         if (sample_type & PERF_SAMPLE_TID)
 131                 idx += 1;
 132
 133         if (sample_type & PERF_SAMPLE_TIME)
 134                 idx += 1;
 135
 136         if (sample_type & PERF_SAMPLE_ADDR)
 137                 idx += 1;
 138
 139         return idx;
 140 }
 141
 142 /**
 143  * __perf_evsel__calc_is_pos - calculate is_pos.
 144  * @sample_type: sample type
 145  *
 146  * This function returns the position (counting backwards) of the event id
 147  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 148  * sample_id_all is used there is an id sample appended to non-sample events.
 149  */
 150 static int __perf_evsel__calc_is_pos(u64 sample_type)
 151 {
 152         int idx = 1;
 153
 154         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 155                 return 1;
 156
 157         if (!(sample_type & PERF_SAMPLE_ID))
 158                 return -1;
 159
 160         if (sample_type & PERF_SAMPLE_CPU)
 161                 idx += 1;
 162
 163         if (sample_type & PERF_SAMPLE_STREAM_ID)
 164                 idx += 1;
 165
 166         return idx;
 167 }
 168
 169 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 170 {
 171         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 172         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 173 }
 174
 175 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 176                                   enum perf_event_sample_format bit)
 177 {
 178         if (!(evsel->attr.sample_type & bit)) {
 179                 evsel->attr.sample_type |= bit;
 180                 evsel->sample_size += sizeof(u64);
 181                 perf_evsel__calc_id_pos(evsel);
 182         }
 183 }
 184
 185 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 186                                     enum perf_event_sample_format bit)
 187 {
 188         if (evsel->attr.sample_type & bit) {
 189                 evsel->attr.sample_type &= ~bit;
 190                 evsel->sample_size -= sizeof(u64);
 191                 perf_evsel__calc_id_pos(evsel);
 192         }
 193 }
 194
 195 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 196                                bool can_sample_identifier)
 197 {
 198         if (can_sample_identifier) {
 199                 perf_evsel__reset_sample_bit(evsel, ID);
 200                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 201         } else {
 202                 perf_evsel__set_sample_bit(evsel, ID);
 203         }
 204         evsel->attr.read_format |= PERF_FORMAT_ID;
 205 }
 206
 207 /**
 208  * perf_evsel__is_function_event - Return whether given evsel is a function
 209  * trace event
 210  *
 211  * @evsel - evsel selector to be tested
 212  *
 213  * Return %true if event is function trace event
 214  */
 215 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 216 {
 217 #define FUNCTION_EVENT "ftrace:function"
 218
 219         return evsel->name &&
 220                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 221
 222 #undef FUNCTION_EVENT
 223 }
 224
 225 void perf_evsel__init(struct perf_evsel *evsel,
 226                       struct perf_event_attr *attr, int idx)
 227 {
 228         evsel->idx         = idx;
 229         evsel->tracking    = !idx;
 230         evsel->attr        = *attr;
 231         evsel->leader      = evsel;
 232         evsel->unit        = "";
 233         evsel->scale       = 1.0;
 234         evsel->max_events  = ULONG_MAX;
 235         evsel->evlist      = NULL;
 236         evsel->bpf_fd      = -1;
 237         INIT_LIST_HEAD(&evsel->node);
 238         INIT_LIST_HEAD(&evsel->config_terms);
 239         perf_evsel__object.init(evsel);
 240         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 241         perf_evsel__calc_id_pos(evsel);
 242         evsel->cmdline_group_boundary = false;
 243         evsel->metric_expr   = NULL;
 244         evsel->metric_name   = NULL;
 245         evsel->metric_events = NULL;
 246         evsel->collect_stat  = false;
 247         evsel->pmu_name      = NULL;
 248 }
 249
 250 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 251 {
 252         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 253
 254         if (!evsel)
 255                 return NULL;
 256         perf_evsel__init(evsel, attr, idx);
 257
 258         if (perf_evsel__is_bpf_output(evsel)) {
 259                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 260                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 261                 evsel->attr.sample_period = 1;
 262         }
 263
 264         if (perf_evsel__is_clock(evsel)) {
 265                 /*
 266                  * The evsel->unit points to static alias->unit
 267                  * so it's ok to use static string in here.
 268                  */
 269                 static const char *unit = "msec";
 270
 271                 evsel->unit = unit;
 272                 evsel->scale = 1e-6;
 273         }
 274
 275         return evsel;
 276 }
 277
 278 static bool perf_event_can_profile_kernel(void)
 279 {
 280         return geteuid() == 0 || perf_event_paranoid() == -1;
 281 }
 282
 283 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 284 {
 285         struct perf_event_attr attr = {
 286                 .type   = PERF_TYPE_HARDWARE,
 287                 .config = PERF_COUNT_HW_CPU_CYCLES,
 288                 .exclude_kernel = !perf_event_can_profile_kernel(),
 289         };
 290         struct perf_evsel *evsel;
 291
 292         event_attr_init(&attr);
 293
 294         if (!precise)
 295                 goto new_event;
 296
 297         /*
 298          * Now let the usual logic to set up the perf_event_attr defaults
 299          * to kick in when we return and before perf_evsel__open() is called.
 300          */
 301 new_event:
 302         evsel = perf_evsel__new(&attr);
 303         if (evsel == NULL)
 304                 goto out;
 305
 306         evsel->precise_max = true;
 307
 308         /* use asprintf() because free(evsel) assumes name is allocated */
 309         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 310                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 311                      attr.exclude_kernel ? "u" : "",
 312                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 313                 goto error_free;
 314 out:
 315         return evsel;
 316 error_free:
 317         perf_evsel__delete(evsel);
 318         evsel = NULL;
 319         goto out;
 320 }
 321
 322 /*
 323  * Returns pointer with encoded error via <linux/err.h> interface.
 324  */
 325 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 326 {
 327         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 328         int err = -ENOMEM;
 329
 330         if (evsel == NULL) {
 331                 goto out_err;
 332         } else {
 333                 struct perf_event_attr attr = {
 334                         .type          = PERF_TYPE_TRACEPOINT,
 335                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 336                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 337                 };
 338
 339                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 340                         goto out_free;
 341
 342                 evsel->tp_format = trace_event__tp_format(sys, name);
 343                 if (IS_ERR(evsel->tp_format)) {
 344                         err = PTR_ERR(evsel->tp_format);
 345                         goto out_free;
 346                 }
 347
 348                 event_attr_init(&attr);
 349                 attr.config = evsel->tp_format->id;
 350                 attr.sample_period = 1;
 351                 perf_evsel__init(evsel, &attr, idx);
 352         }
 353
 354         return evsel;
 355
 356 out_free:
 357         zfree(&evsel->name);
 358         free(evsel);
 359 out_err:
 360         return ERR_PTR(err);
 361 }
 362
 363 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 364         "cycles",
 365         "instructions",
 366         "cache-references",
 367         "cache-misses",
 368         "branches",
 369         "branch-misses",
 370         "bus-cycles",
 371         "stalled-cycles-frontend",
 372         "stalled-cycles-backend",
 373         "ref-cycles",
 374 };
 375
 376 static const char *__perf_evsel__hw_name(u64 config)
 377 {
 378         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 379                 return perf_evsel__hw_names[config];
 380
 381         return "unknown-hardware";
 382 }
 383
 384 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 385 {
 386         int colon = 0, r = 0;
 387         struct perf_event_attr *attr = &evsel->attr;
 388         bool exclude_guest_default = false;
 389
 390 #define MOD_PRINT(context, mod) do {                                    \
 391                 if (!attr->exclude_##context) {                         \
 392                         if (!colon) colon = ++r;                        \
 393                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 394                 } } while(0)
 395
 396         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 397                 MOD_PRINT(kernel, 'k');
 398                 MOD_PRINT(user, 'u');
 399                 MOD_PRINT(hv, 'h');
 400                 exclude_guest_default = true;
 401         }
 402
 403         if (attr->precise_ip) {
 404                 if (!colon)
 405                         colon = ++r;
 406                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 407                 exclude_guest_default = true;
 408         }
 409
 410         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 411                 MOD_PRINT(host, 'H');
 412                 MOD_PRINT(guest, 'G');
 413         }
 414 #undef MOD_PRINT
 415         if (colon)
 416                 bf[colon - 1] = ':';
 417         return r;
 418 }
 419
 420 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 421 {
 422         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 423         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 424 }
 425
 426 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 427         "cpu-clock",
 428         "task-clock",
 429         "page-faults",
 430         "context-switches",
 431         "cpu-migrations",
 432         "minor-faults",
 433         "major-faults",
 434         "alignment-faults",
 435         "emulation-faults",
 436         "dummy",
 437 };
 438
 439 static const char *__perf_evsel__sw_name(u64 config)
 440 {
 441         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 442                 return perf_evsel__sw_names[config];
 443         return "unknown-software";
 444 }
 445
 446 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 447 {
 448         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 449         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 450 }
 451
 452 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 453 {
 454         int r;
 455
 456         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 457
 458         if (type & HW_BREAKPOINT_R)
 459                 r += scnprintf(bf + r, size - r, "r");
 460
 461         if (type & HW_BREAKPOINT_W)
 462                 r += scnprintf(bf + r, size - r, "w");
 463
 464         if (type & HW_BREAKPOINT_X)
 465                 r += scnprintf(bf + r, size - r, "x");
 466
 467         return r;
 468 }
 469
 470 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 471 {
 472         struct perf_event_attr *attr = &evsel->attr;
 473         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 474         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 475 }
 476
 477 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 478                                 [PERF_EVSEL__MAX_ALIASES] = {
 479  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 480  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 481  { "LLC",       "L2",                                                   },
 482  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 483  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 484  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 485  { "node",                                                              },
 486 };
 487
 488 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 489                                    [PERF_EVSEL__MAX_ALIASES] = {
 490  { "load",      "loads",        "read",                                 },
 491  { "store",     "stores",       "write",                                },
 492  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 493 };
 494
 495 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 496                                        [PERF_EVSEL__MAX_ALIASES] = {
 497  { "refs",      "Reference",    "ops",          "access",               },
 498  { "misses",    "miss",                                                 },
 499 };
 500
 501 #define C(x)            PERF_COUNT_HW_CACHE_##x
 502 #define CACHE_READ      (1 << C(OP_READ))
 503 #define CACHE_WRITE     (1 << C(OP_WRITE))
 504 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 505 #define COP(x)          (1 << x)
 506
 507 /*
 508  * cache operartion stat
 509  * L1I : Read and prefetch only
 510  * ITLB and BPU : Read-only
 511  */
 512 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 513  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 514  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 515  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 516  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 517  [C(ITLB)]      = (CACHE_READ),
 518  [C(BPU)]       = (CACHE_READ),
 519  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 520 };
 521
 522 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 523 {
 524         if (perf_evsel__hw_cache_stat[type] & COP(op))
 525                 return true;    /* valid */
 526         else
 527                 return false;   /* invalid */
 528 }
 529
 530 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 531                                             char *bf, size_t size)
 532 {
 533         if (result) {
 534                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 535                                  perf_evsel__hw_cache_op[op][0],
 536                                  perf_evsel__hw_cache_result[result][0]);
 537         }
 538
 539         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 540                          perf_evsel__hw_cache_op[op][1]);
 541 }
 542
 543 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 544 {
 545         u8 op, result, type = (config >>  0) & 0xff;
 546         const char *err = "unknown-ext-hardware-cache-type";
 547
 548         if (type >= PERF_COUNT_HW_CACHE_MAX)
 549                 goto out_err;
 550
 551         op = (config >>  8) & 0xff;
 552         err = "unknown-ext-hardware-cache-op";
 553         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 554                 goto out_err;
 555
 556         result = (config >> 16) & 0xff;
 557         err = "unknown-ext-hardware-cache-result";
 558         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 559                 goto out_err;
 560
 561         err = "invalid-cache";
 562         if (!perf_evsel__is_cache_op_valid(type, op))
 563                 goto out_err;
 564
 565         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 566 out_err:
 567         return scnprintf(bf, size, "%s", err);
 568 }
 569
 570 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 571 {
 572         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 573         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 574 }
 575
 576 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 577 {
 578         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 579         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 580 }
 581
 582 static int perf_evsel__tool_name(char *bf, size_t size)
 583 {
 584         int ret = scnprintf(bf, size, "duration_time");
 585         return ret;
 586 }
 587
 588 const char *perf_evsel__name(struct perf_evsel *evsel)
 589 {
 590         char bf[128];
 591
 592         if (evsel->name)
 593                 return evsel->name;
 594
 595         switch (evsel->attr.type) {
 596         case PERF_TYPE_RAW:
 597                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 598                 break;
 599
 600         case PERF_TYPE_HARDWARE:
 601                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 602                 break;
 603
 604         case PERF_TYPE_HW_CACHE:
 605                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 606                 break;
 607
 608         case PERF_TYPE_SOFTWARE:
 609                 if (evsel->tool_event)
 610                         perf_evsel__tool_name(bf, sizeof(bf));
 611                 else
 612                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 613                 break;
 614
 615         case PERF_TYPE_TRACEPOINT:
 616                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 617                 break;
 618
 619         case PERF_TYPE_BREAKPOINT:
 620                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 621                 break;
 622
 623         default:
 624                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 625                           evsel->attr.type);
 626                 break;
 627         }
 628
 629         evsel->name = strdup(bf);
 630
 631         return evsel->name ?: "unknown";
 632 }
 633
 634 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 635 {
 636         return evsel->group_name ?: "anon group";
 637 }
 638
 639 /*
 640  * Returns the group details for the specified leader,
 641  * with following rules.
 642  *
 643  *  For record -e '{cycles,instructions}'
 644  *    'anon group { cycles:u, instructions:u }'
 645  *
 646  *  For record -e 'cycles,instructions' and report --group
 647  *    'cycles:u, instructions:u'
 648  */
 649 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 650 {
 651         int ret = 0;
 652         struct perf_evsel *pos;
 653         const char *group_name = perf_evsel__group_name(evsel);
 654
 655         if (!evsel->forced_leader)
 656                 ret = scnprintf(buf, size, "%s { ", group_name);
 657
 658         ret += scnprintf(buf + ret, size - ret, "%s",
 659                          perf_evsel__name(evsel));
 660
 661         for_each_group_member(pos, evsel)
 662                 ret += scnprintf(buf + ret, size - ret, ", %s",
 663                                  perf_evsel__name(pos));
 664
 665         if (!evsel->forced_leader)
 666                 ret += scnprintf(buf + ret, size - ret, " }");
 667
 668         return ret;
 669 }
 670
 671 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 672                                            struct record_opts *opts,
 673                                            struct callchain_param *param)
 674 {
 675         bool function = perf_evsel__is_function_event(evsel);
 676         struct perf_event_attr *attr = &evsel->attr;
 677
 678         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 679
 680         attr->sample_max_stack = param->max_stack;
 681
 682         if (param->record_mode == CALLCHAIN_LBR) {
 683                 if (!opts->branch_stack) {
 684                         if (attr->exclude_user) {
 685                                 pr_warning("LBR callstack option is only available "
 686                                            "to get user callchain information. "
 687                                            "Falling back to framepointers.\n");
 688                         } else {
 689                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 690                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 691                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 692                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 693                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 694                         }
 695                 } else
 696                          pr_warning("Cannot use LBR callstack with branch stack. "
 697                                     "Falling back to framepointers.\n");
 698         }
 699
 700         if (param->record_mode == CALLCHAIN_DWARF) {
 701                 if (!function) {
 702                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 703                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 704                         attr->sample_regs_user |= PERF_REGS_MASK;
 705                         attr->sample_stack_user = param->dump_size;
 706                         attr->exclude_callchain_user = 1;
 707                 } else {
 708                         pr_info("Cannot use DWARF unwind for function trace event,"
 709                                 " falling back to framepointers.\n");
 710                 }
 711         }
 712
 713         if (function) {
 714                 pr_info("Disabling user space callchains for function trace event.\n");
 715                 attr->exclude_callchain_user = 1;
 716         }
 717 }
 718
 719 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 720                                   struct record_opts *opts,
 721                                   struct callchain_param *param)
 722 {
 723         if (param->enabled)
 724                 return __perf_evsel__config_callchain(evsel, opts, param);
 725 }
 726
 727 static void
 728 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 729                             struct callchain_param *param)
 730 {
 731         struct perf_event_attr *attr = &evsel->attr;
 732
 733         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 734         if (param->record_mode == CALLCHAIN_LBR) {
 735                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 736                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 737                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 738         }
 739         if (param->record_mode == CALLCHAIN_DWARF) {
 740                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 741                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 742         }
 743 }
 744
 745 static void apply_config_terms(struct perf_evsel *evsel,
 746                                struct record_opts *opts, bool track)
 747 {
 748         struct perf_evsel_config_term *term;
 749         struct list_head *config_terms = &evsel->config_terms;
 750         struct perf_event_attr *attr = &evsel->attr;
 751         /* callgraph default */
 752         struct callchain_param param = {
 753                 .record_mode = callchain_param.record_mode,
 754         };
 755         u32 dump_size = 0;
 756         int max_stack = 0;
 757         const char *callgraph_buf = NULL;
 758
 759         list_for_each_entry(term, config_terms, list) {
 760                 switch (term->type) {
 761                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 762                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 763                                 attr->sample_period = term->val.period;
 764                                 attr->freq = 0;
 765                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 766                         }
 767                         break;
 768                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 769                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 770                                 attr->sample_freq = term->val.freq;
 771                                 attr->freq = 1;
 772                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 773                         }
 774                         break;
 775                 case PERF_EVSEL__CONFIG_TERM_TIME:
 776                         if (term->val.time)
 777                                 perf_evsel__set_sample_bit(evsel, TIME);
 778                         else
 779                                 perf_evsel__reset_sample_bit(evsel, TIME);
 780                         break;
 781                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 782                         callgraph_buf = term->val.callgraph;
 783                         break;
 784                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 785                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 786                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 787                                 parse_branch_str(term->val.branch,
 788                                                  &attr->branch_sample_type);
 789                         } else
 790                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 791                         break;
 792                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 793                         dump_size = term->val.stack_user;
 794                         break;
 795                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 796                         max_stack = term->val.max_stack;
 797                         break;
 798                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 799                         evsel->max_events = term->val.max_events;
 800                         break;
 801                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 802                         /*
 803                          * attr->inherit should has already been set by
 804                          * perf_evsel__config. If user explicitly set
 805                          * inherit using config terms, override global
 806                          * opt->no_inherit setting.
 807                          */
 808                         attr->inherit = term->val.inherit ? 1 : 0;
 809                         break;
 810                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 811                         attr->write_backward = term->val.overwrite ? 1 : 0;
 812                         break;
 813                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 814                         break;
 815                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 816                         break;
 817                 default:
 818                         break;
 819                 }
 820         }
 821
 822         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 823         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 824                 bool sample_address = false;
 825
 826                 if (max_stack) {
 827                         param.max_stack = max_stack;
 828                         if (callgraph_buf == NULL)
 829                                 callgraph_buf = "fp";
 830                 }
 831
 832                 /* parse callgraph parameters */
 833                 if (callgraph_buf != NULL) {
 834                         if (!strcmp(callgraph_buf, "no")) {
 835                                 param.enabled = false;
 836                                 param.record_mode = CALLCHAIN_NONE;
 837                         } else {
 838                                 param.enabled = true;
 839                                 if (parse_callchain_record(callgraph_buf, &param)) {
 840                                         pr_err("per-event callgraph setting for %s failed. "
 841                                                "Apply callgraph global setting for it\n",
 842                                                evsel->name);
 843                                         return;
 844                                 }
 845                                 if (param.record_mode == CALLCHAIN_DWARF)
 846                                         sample_address = true;
 847                         }
 848                 }
 849                 if (dump_size > 0) {
 850                         dump_size = round_up(dump_size, sizeof(u64));
 851                         param.dump_size = dump_size;
 852                 }
 853
 854                 /* If global callgraph set, clear it */
 855                 if (callchain_param.enabled)
 856                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 857
 858                 /* set perf-event callgraph */
 859                 if (param.enabled) {
 860                         if (sample_address) {
 861                                 perf_evsel__set_sample_bit(evsel, ADDR);
 862                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 863                                 evsel->attr.mmap_data = track;
 864                         }
 865                         perf_evsel__config_callchain(evsel, opts, &param);
 866                 }
 867         }
 868 }
 869
 870 static bool is_dummy_event(struct perf_evsel *evsel)
 871 {
 872         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 873                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 874 }
 875
 876 /*
 877  * The enable_on_exec/disabled value strategy:
 878  *
 879  *  1) For any type of traced program:
 880  *    - all independent events and group leaders are disabled
 881  *    - all group members are enabled
 882  *
 883  *     Group members are ruled by group leaders. They need to
 884  *     be enabled, because the group scheduling relies on that.
 885  *
 886  *  2) For traced programs executed by perf:
 887  *     - all independent events and group leaders have
 888  *       enable_on_exec set
 889  *     - we don't specifically enable or disable any event during
 890  *       the record command
 891  *
 892  *     Independent events and group leaders are initially disabled
 893  *     and get enabled by exec. Group members are ruled by group
 894  *     leaders as stated in 1).
 895  *
 896  *  3) For traced programs attached by perf (pid/tid):
 897  *     - we specifically enable or disable all events during
 898  *       the record command
 899  *
 900  *     When attaching events to already running traced we
 901  *     enable/disable events specifically, as there's no
 902  *     initial traced exec call.
 903  */
 904 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 905                         struct callchain_param *callchain)
 906 {
 907         struct perf_evsel *leader = evsel->leader;
 908         struct perf_event_attr *attr = &evsel->attr;
 909         int track = evsel->tracking;
 910         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 911
 912         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 913         attr->inherit       = !opts->no_inherit;
 914         attr->write_backward = opts->overwrite ? 1 : 0;
 915
 916         perf_evsel__set_sample_bit(evsel, IP);
 917         perf_evsel__set_sample_bit(evsel, TID);
 918
 919         if (evsel->sample_read) {
 920                 perf_evsel__set_sample_bit(evsel, READ);
 921
 922                 /*
 923                  * We need ID even in case of single event, because
 924                  * PERF_SAMPLE_READ process ID specific data.
 925                  */
 926                 perf_evsel__set_sample_id(evsel, false);
 927
 928                 /*
 929                  * Apply group format only if we belong to group
 930                  * with more than one members.
 931                  */
 932                 if (leader->nr_members > 1) {
 933                         attr->read_format |= PERF_FORMAT_GROUP;
 934                         attr->inherit = 0;
 935                 }
 936         }
 937
 938         /*
 939          * We default some events to have a default interval. But keep
 940          * it a weak assumption overridable by the user.
 941          */
 942         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 943                                      opts->user_interval != ULLONG_MAX)) {
 944                 if (opts->freq) {
 945                         perf_evsel__set_sample_bit(evsel, PERIOD);
 946                         attr->freq              = 1;
 947                         attr->sample_freq       = opts->freq;
 948                 } else {
 949                         attr->sample_period = opts->default_interval;
 950                 }
 951         }
 952
 953         /*
 954          * Disable sampling for all group members other
 955          * than leader in case leader 'leads' the sampling.
 956          */
 957         if ((leader != evsel) && leader->sample_read) {
 958                 attr->freq           = 0;
 959                 attr->sample_freq    = 0;
 960                 attr->sample_period  = 0;
 961                 attr->write_backward = 0;
 962
 963                 /*
 964                  * We don't get sample for slave events, we make them
 965                  * when delivering group leader sample. Set the slave
 966                  * event to follow the master sample_type to ease up
 967                  * report.
 968                  */
 969                 attr->sample_type = leader->attr.sample_type;
 970         }
 971
 972         if (opts->no_samples)
 973                 attr->sample_freq = 0;
 974
 975         if (opts->inherit_stat) {
 976                 evsel->attr.read_format |=
 977                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 978                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 979                         PERF_FORMAT_ID;
 980                 attr->inherit_stat = 1;
 981         }
 982
 983         if (opts->sample_address) {
 984                 perf_evsel__set_sample_bit(evsel, ADDR);
 985                 attr->mmap_data = track;
 986         }
 987
 988         /*
 989          * We don't allow user space callchains for  function trace
 990          * event, due to issues with page faults while tracing page
 991          * fault handler and its overall trickiness nature.
 992          */
 993         if (perf_evsel__is_function_event(evsel))
 994                 evsel->attr.exclude_callchain_user = 1;
 995
 996         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 997                 perf_evsel__config_callchain(evsel, opts, callchain);
 998
 999         if (opts->sample_intr_regs) {
1000                 attr->sample_regs_intr = opts->sample_intr_regs;
1001                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1002         }
1003
1004         if (opts->sample_user_regs) {
1005                 attr->sample_regs_user |= opts->sample_user_regs;
1006                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1007         }
1008
1009         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1010                 perf_evsel__set_sample_bit(evsel, CPU);
1011
1012         /*
1013          * When the user explicitly disabled time don't force it here.
1014          */
1015         if (opts->sample_time &&
1016             (!perf_missing_features.sample_id_all &&
1017             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1018              opts->sample_time_set)))
1019                 perf_evsel__set_sample_bit(evsel, TIME);
1020
1021         if (opts->raw_samples && !evsel->no_aux_samples) {
1022                 perf_evsel__set_sample_bit(evsel, TIME);
1023                 perf_evsel__set_sample_bit(evsel, RAW);
1024                 perf_evsel__set_sample_bit(evsel, CPU);
1025         }
1026
1027         if (opts->sample_address)
1028                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1029
1030         if (opts->sample_phys_addr)
1031                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1032
1033         if (opts->no_buffering) {
1034                 attr->watermark = 0;
1035                 attr->wakeup_events = 1;
1036         }
1037         if (opts->branch_stack && !evsel->no_aux_samples) {
1038                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1039                 attr->branch_sample_type = opts->branch_stack;
1040         }
1041
1042         if (opts->sample_weight)
1043                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1044
1045         attr->task  = track;
1046         attr->mmap  = track;
1047         attr->mmap2 = track && !perf_missing_features.mmap2;
1048         attr->comm  = track;
1049         attr->ksymbol = track && !perf_missing_features.ksymbol;
1050         attr->bpf_event = track && !opts->no_bpf_event &&
1051                 !perf_missing_features.bpf_event;
1052
1053         if (opts->record_namespaces)
1054                 attr->namespaces  = track;
1055
1056         if (opts->record_switch_events)
1057                 attr->context_switch = track;
1058
1059         if (opts->sample_transaction)
1060                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1061
1062         if (opts->running_time) {
1063                 evsel->attr.read_format |=
1064                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1065                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1066         }
1067
1068         /*
1069          * XXX see the function comment above
1070          *
1071          * Disabling only independent events or group leaders,
1072          * keeping group members enabled.
1073          */
1074         if (perf_evsel__is_group_leader(evsel))
1075                 attr->disabled = 1;
1076
1077         /*
1078          * Setting enable_on_exec for independent events and
1079          * group leaders for traced executed by perf.
1080          */
1081         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1082                 !opts->initial_delay)
1083                 attr->enable_on_exec = 1;
1084
1085         if (evsel->immediate) {
1086                 attr->disabled = 0;
1087                 attr->enable_on_exec = 0;
1088         }
1089
1090         clockid = opts->clockid;
1091         if (opts->use_clockid) {
1092                 attr->use_clockid = 1;
1093                 attr->clockid = opts->clockid;
1094         }
1095
1096         if (evsel->precise_max)
1097                 attr->precise_ip = 3;
1098
1099         if (opts->all_user) {
1100                 attr->exclude_kernel = 1;
1101                 attr->exclude_user   = 0;
1102         }
1103
1104         if (opts->all_kernel) {
1105                 attr->exclude_kernel = 0;
1106                 attr->exclude_user   = 1;
1107         }
1108
1109         if (evsel->own_cpus || evsel->unit)
1110                 evsel->attr.read_format |= PERF_FORMAT_ID;
1111
1112         /*
1113          * Apply event specific term settings,
1114          * it overloads any global configuration.
1115          */
1116         apply_config_terms(evsel, opts, track);
1117
1118         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1119
1120         /* The --period option takes the precedence. */
1121         if (opts->period_set) {
1122                 if (opts->period)
1123                         perf_evsel__set_sample_bit(evsel, PERIOD);
1124                 else
1125                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1126         }
1127
1128         /*
1129          * For initial_delay, a dummy event is added implicitly.
1130          * The software event will trigger -EOPNOTSUPP error out,
1131          * if BRANCH_STACK bit is set.
1132          */
1133         if (opts->initial_delay && is_dummy_event(evsel))
1134                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1135 }
1136
1137 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1138 {
1139         if (evsel->system_wide)
1140                 nthreads = 1;
1141
1142         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1143
1144         if (evsel->fd) {
1145                 int cpu, thread;
1146                 for (cpu = 0; cpu < ncpus; cpu++) {
1147                         for (thread = 0; thread < nthreads; thread++) {
1148                                 FD(evsel, cpu, thread) = -1;
1149                         }
1150                 }
1151         }
1152
1153         return evsel->fd != NULL ? 0 : -ENOMEM;
1154 }
1155
1156 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1157                           int ioc,  void *arg)
1158 {
1159         int cpu, thread;
1160
1161         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1162                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1163                         int fd = FD(evsel, cpu, thread),
1164                             err = ioctl(fd, ioc, arg);
1165
1166                         if (err)
1167                                 return err;
1168                 }
1169         }
1170
1171         return 0;
1172 }
1173
1174 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1175 {
1176         return perf_evsel__run_ioctl(evsel,
1177                                      PERF_EVENT_IOC_SET_FILTER,
1178                                      (void *)filter);
1179 }
1180
1181 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1182 {
1183         char *new_filter = strdup(filter);
1184
1185         if (new_filter != NULL) {
1186                 free(evsel->filter);
1187                 evsel->filter = new_filter;
1188                 return 0;
1189         }
1190
1191         return -1;
1192 }
1193
1194 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1195                                      const char *fmt, const char *filter)
1196 {
1197         char *new_filter;
1198
1199         if (evsel->filter == NULL)
1200                 return perf_evsel__set_filter(evsel, filter);
1201
1202         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1203                 free(evsel->filter);
1204                 evsel->filter = new_filter;
1205                 return 0;
1206         }
1207
1208         return -1;
1209 }
1210
1211 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1212 {
1213         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1214 }
1215
1216 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1217 {
1218         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1219 }
1220
1221 int perf_evsel__enable(struct perf_evsel *evsel)
1222 {
1223         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1224
1225         if (!err)
1226                 evsel->disabled = false;
1227
1228         return err;
1229 }
1230
1231 int perf_evsel__disable(struct perf_evsel *evsel)
1232 {
1233         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1234         /*
1235          * We mark it disabled here so that tools that disable a event can
1236          * ignore events after they disable it. I.e. the ring buffer may have
1237          * already a few more events queued up before the kernel got the stop
1238          * request.
1239          */
1240         if (!err)
1241                 evsel->disabled = true;
1242
1243         return err;
1244 }
1245
1246 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1247 {
1248         if (ncpus == 0 || nthreads == 0)
1249                 return 0;
1250
1251         if (evsel->system_wide)
1252                 nthreads = 1;
1253
1254         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1255         if (evsel->sample_id == NULL)
1256                 return -ENOMEM;
1257
1258         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1259         if (evsel->id == NULL) {
1260                 xyarray__delete(evsel->sample_id);
1261                 evsel->sample_id = NULL;
1262                 return -ENOMEM;
1263         }
1264
1265         return 0;
1266 }
1267
1268 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1269 {
1270         xyarray__delete(evsel->fd);
1271         evsel->fd = NULL;
1272 }
1273
1274 static void perf_evsel__free_id(struct perf_evsel *evsel)
1275 {
1276         xyarray__delete(evsel->sample_id);
1277         evsel->sample_id = NULL;
1278         zfree(&evsel->id);
1279 }
1280
1281 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1282 {
1283         struct perf_evsel_config_term *term, *h;
1284
1285         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1286                 list_del(&term->list);
1287                 free(term);
1288         }
1289 }
1290
1291 void perf_evsel__close_fd(struct perf_evsel *evsel)
1292 {
1293         int cpu, thread;
1294
1295         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1296                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1297                         close(FD(evsel, cpu, thread));
1298                         FD(evsel, cpu, thread) = -1;
1299                 }
1300 }
1301
1302 void perf_evsel__exit(struct perf_evsel *evsel)
1303 {
1304         assert(list_empty(&evsel->node));
1305         assert(evsel->evlist == NULL);
1306         perf_evsel__free_counts(evsel);
1307         perf_evsel__free_fd(evsel);
1308         perf_evsel__free_id(evsel);
1309         perf_evsel__free_config_terms(evsel);
1310         cgroup__put(evsel->cgrp);
1311         cpu_map__put(evsel->cpus);
1312         cpu_map__put(evsel->own_cpus);
1313         thread_map__put(evsel->threads);
1314         zfree(&evsel->group_name);
1315         zfree(&evsel->name);
1316         perf_evsel__object.fini(evsel);
1317 }
1318
1319 void perf_evsel__delete(struct perf_evsel *evsel)
1320 {
1321         perf_evsel__exit(evsel);
1322         free(evsel);
1323 }
1324
1325 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1326                                 struct perf_counts_values *count)
1327 {
1328         struct perf_counts_values tmp;
1329
1330         if (!evsel->prev_raw_counts)
1331                 return;
1332
1333         if (cpu == -1) {
1334                 tmp = evsel->prev_raw_counts->aggr;
1335                 evsel->prev_raw_counts->aggr = *count;
1336         } else {
1337                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1338                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1339         }
1340
1341         count->val = count->val - tmp.val;
1342         count->ena = count->ena - tmp.ena;
1343         count->run = count->run - tmp.run;
1344 }
1345
1346 void perf_counts_values__scale(struct perf_counts_values *count,
1347                                bool scale, s8 *pscaled)
1348 {
1349         s8 scaled = 0;
1350
1351         if (scale) {
1352                 if (count->run == 0) {
1353                         scaled = -1;
1354                         count->val = 0;
1355                 } else if (count->run < count->ena) {
1356                         scaled = 1;
1357                         count->val = (u64)((double) count->val * count->ena / count->run);
1358                 }
1359         }
1360
1361         if (pscaled)
1362                 *pscaled = scaled;
1363 }
1364
1365 static int perf_evsel__read_size(struct perf_evsel *evsel)
1366 {
1367         u64 read_format = evsel->attr.read_format;
1368         int entry = sizeof(u64); /* value */
1369         int size = 0;
1370         int nr = 1;
1371
1372         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1373                 size += sizeof(u64);
1374
1375         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1376                 size += sizeof(u64);
1377
1378         if (read_format & PERF_FORMAT_ID)
1379                 entry += sizeof(u64);
1380
1381         if (read_format & PERF_FORMAT_GROUP) {
1382                 nr = evsel->nr_members;
1383                 size += sizeof(u64);
1384         }
1385
1386         size += entry * nr;
1387         return size;
1388 }
1389
1390 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1391                      struct perf_counts_values *count)
1392 {
1393         size_t size = perf_evsel__read_size(evsel);
1394
1395         memset(count, 0, sizeof(*count));
1396
1397         if (FD(evsel, cpu, thread) < 0)
1398                 return -EINVAL;
1399
1400         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1401                 return -errno;
1402
1403         return 0;
1404 }
1405
1406 static int
1407 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1408 {
1409         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1410
1411         return perf_evsel__read(evsel, cpu, thread, count);
1412 }
1413
1414 static void
1415 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1416                       u64 val, u64 ena, u64 run)
1417 {
1418         struct perf_counts_values *count;
1419
1420         count = perf_counts(counter->counts, cpu, thread);
1421
1422         count->val    = val;
1423         count->ena    = ena;
1424         count->run    = run;
1425         count->loaded = true;
1426 }
1427
1428 static int
1429 perf_evsel__process_group_data(struct perf_evsel *leader,
1430                                int cpu, int thread, u64 *data)
1431 {
1432         u64 read_format = leader->attr.read_format;
1433         struct sample_read_value *v;
1434         u64 nr, ena = 0, run = 0, i;
1435
1436         nr = *data++;
1437
1438         if (nr != (u64) leader->nr_members)
1439                 return -EINVAL;
1440
1441         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1442                 ena = *data++;
1443
1444         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1445                 run = *data++;
1446
1447         v = (struct sample_read_value *) data;
1448
1449         perf_evsel__set_count(leader, cpu, thread,
1450                               v[0].value, ena, run);
1451
1452         for (i = 1; i < nr; i++) {
1453                 struct perf_evsel *counter;
1454
1455                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1456                 if (!counter)
1457                         return -EINVAL;
1458
1459                 perf_evsel__set_count(counter, cpu, thread,
1460                                       v[i].value, ena, run);
1461         }
1462
1463         return 0;
1464 }
1465
1466 static int
1467 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1468 {
1469         struct perf_stat_evsel *ps = leader->stats;
1470         u64 read_format = leader->attr.read_format;
1471         int size = perf_evsel__read_size(leader);
1472         u64 *data = ps->group_data;
1473
1474         if (!(read_format & PERF_FORMAT_ID))
1475                 return -EINVAL;
1476
1477         if (!perf_evsel__is_group_leader(leader))
1478                 return -EINVAL;
1479
1480         if (!data) {
1481                 data = zalloc(size);
1482                 if (!data)
1483                         return -ENOMEM;
1484
1485                 ps->group_data = data;
1486         }
1487
1488         if (FD(leader, cpu, thread) < 0)
1489                 return -EINVAL;
1490
1491         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1492                 return -errno;
1493
1494         return perf_evsel__process_group_data(leader, cpu, thread, data);
1495 }
1496
1497 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1498 {
1499         u64 read_format = evsel->attr.read_format;
1500
1501         if (read_format & PERF_FORMAT_GROUP)
1502                 return perf_evsel__read_group(evsel, cpu, thread);
1503         else
1504                 return perf_evsel__read_one(evsel, cpu, thread);
1505 }
1506
1507 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1508                               int cpu, int thread, bool scale)
1509 {
1510         struct perf_counts_values count;
1511         size_t nv = scale ? 3 : 1;
1512
1513         if (FD(evsel, cpu, thread) < 0)
1514                 return -EINVAL;
1515
1516         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1517                 return -ENOMEM;
1518
1519         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1520                 return -errno;
1521
1522         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1523         perf_counts_values__scale(&count, scale, NULL);
1524         *perf_counts(evsel->counts, cpu, thread) = count;
1525         return 0;
1526 }
1527
1528 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1529 {
1530         struct perf_evsel *leader = evsel->leader;
1531         int fd;
1532
1533         if (perf_evsel__is_group_leader(evsel))
1534                 return -1;
1535
1536         /*
1537          * Leader must be already processed/open,
1538          * if not it's a bug.
1539          */
1540         BUG_ON(!leader->fd);
1541
1542         fd = FD(leader, cpu, thread);
1543         BUG_ON(fd == -1);
1544
1545         return fd;
1546 }
1547
1548 struct bit_names {
1549         int bit;
1550         const char *name;
1551 };
1552
1553 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1554 {
1555         bool first_bit = true;
1556         int i = 0;
1557
1558         do {
1559                 if (value & bits[i].bit) {
1560                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1561                         first_bit = false;
1562                 }
1563         } while (bits[++i].name != NULL);
1564 }
1565
1566 static void __p_sample_type(char *buf, size_t size, u64 value)
1567 {
1568 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1569         struct bit_names bits[] = {
1570                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1571                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1572                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1573                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1574                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1575                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1576                 { .name = NULL, }
1577         };
1578 #undef bit_name
1579         __p_bits(buf, size, value, bits);
1580 }
1581
1582 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1583 {
1584 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1585         struct bit_names bits[] = {
1586                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1587                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1588                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1589                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1590                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1591                 { .name = NULL, }
1592         };
1593 #undef bit_name
1594         __p_bits(buf, size, value, bits);
1595 }
1596
1597 static void __p_read_format(char *buf, size_t size, u64 value)
1598 {
1599 #define bit_name(n) { PERF_FORMAT_##n, #n }
1600         struct bit_names bits[] = {
1601                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1602                 bit_name(ID), bit_name(GROUP),
1603                 { .name = NULL, }
1604         };
1605 #undef bit_name
1606         __p_bits(buf, size, value, bits);
1607 }
1608
1609 #define BUF_SIZE                1024
1610
1611 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1612 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1613 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1614 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1615 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1616 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1617
1618 #define PRINT_ATTRn(_n, _f, _p)                         \
1619 do {                                                    \
1620         if (attr->_f) {                                 \
1621                 _p(attr->_f);                           \
1622                 ret += attr__fprintf(fp, _n, buf, priv);\
1623         }                                               \
1624 } while (0)
1625
1626 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1627
1628 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1629                              attr__fprintf_f attr__fprintf, void *priv)
1630 {
1631         char buf[BUF_SIZE];
1632         int ret = 0;
1633
1634         PRINT_ATTRf(type, p_unsigned);
1635         PRINT_ATTRf(size, p_unsigned);
1636         PRINT_ATTRf(config, p_hex);
1637         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1638         PRINT_ATTRf(sample_type, p_sample_type);
1639         PRINT_ATTRf(read_format, p_read_format);
1640
1641         PRINT_ATTRf(disabled, p_unsigned);
1642         PRINT_ATTRf(inherit, p_unsigned);
1643         PRINT_ATTRf(pinned, p_unsigned);
1644         PRINT_ATTRf(exclusive, p_unsigned);
1645         PRINT_ATTRf(exclude_user, p_unsigned);
1646         PRINT_ATTRf(exclude_kernel, p_unsigned);
1647         PRINT_ATTRf(exclude_hv, p_unsigned);
1648         PRINT_ATTRf(exclude_idle, p_unsigned);
1649         PRINT_ATTRf(mmap, p_unsigned);
1650         PRINT_ATTRf(comm, p_unsigned);
1651         PRINT_ATTRf(freq, p_unsigned);
1652         PRINT_ATTRf(inherit_stat, p_unsigned);
1653         PRINT_ATTRf(enable_on_exec, p_unsigned);
1654         PRINT_ATTRf(task, p_unsigned);
1655         PRINT_ATTRf(watermark, p_unsigned);
1656         PRINT_ATTRf(precise_ip, p_unsigned);
1657         PRINT_ATTRf(mmap_data, p_unsigned);
1658         PRINT_ATTRf(sample_id_all, p_unsigned);
1659         PRINT_ATTRf(exclude_host, p_unsigned);
1660         PRINT_ATTRf(exclude_guest, p_unsigned);
1661         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1662         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1663         PRINT_ATTRf(mmap2, p_unsigned);
1664         PRINT_ATTRf(comm_exec, p_unsigned);
1665         PRINT_ATTRf(use_clockid, p_unsigned);
1666         PRINT_ATTRf(context_switch, p_unsigned);
1667         PRINT_ATTRf(write_backward, p_unsigned);
1668         PRINT_ATTRf(namespaces, p_unsigned);
1669         PRINT_ATTRf(ksymbol, p_unsigned);
1670         PRINT_ATTRf(bpf_event, p_unsigned);
1671
1672         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1673         PRINT_ATTRf(bp_type, p_unsigned);
1674         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1675         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1676         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1677         PRINT_ATTRf(sample_regs_user, p_hex);
1678         PRINT_ATTRf(sample_stack_user, p_unsigned);
1679         PRINT_ATTRf(clockid, p_signed);
1680         PRINT_ATTRf(sample_regs_intr, p_hex);
1681         PRINT_ATTRf(aux_watermark, p_unsigned);
1682         PRINT_ATTRf(sample_max_stack, p_unsigned);
1683
1684         return ret;
1685 }
1686
1687 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1688                                 void *priv __maybe_unused)
1689 {
1690         return fprintf(fp, "  %-32s %s\n", name, val);
1691 }
1692
1693 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1694                                   int nr_cpus, int nr_threads,
1695                                   int thread_idx)
1696 {
1697         for (int cpu = 0; cpu < nr_cpus; cpu++)
1698                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1699                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1700 }
1701
1702 static int update_fds(struct perf_evsel *evsel,
1703                       int nr_cpus, int cpu_idx,
1704                       int nr_threads, int thread_idx)
1705 {
1706         struct perf_evsel *pos;
1707
1708         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1709                 return -EINVAL;
1710
1711         evlist__for_each_entry(evsel->evlist, pos) {
1712                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1713
1714                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1715
1716                 /*
1717                  * Since fds for next evsel has not been created,
1718                  * there is no need to iterate whole event list.
1719                  */
1720                 if (pos == evsel)
1721                         break;
1722         }
1723         return 0;
1724 }
1725
1726 static bool ignore_missing_thread(struct perf_evsel *evsel,
1727                                   int nr_cpus, int cpu,
1728                                   struct thread_map *threads,
1729                                   int thread, int err)
1730 {
1731         pid_t ignore_pid = thread_map__pid(threads, thread);
1732
1733         if (!evsel->ignore_missing_thread)
1734                 return false;
1735
1736         /* The system wide setup does not work with threads. */
1737         if (evsel->system_wide)
1738                 return false;
1739
1740         /* The -ESRCH is perf event syscall errno for pid's not found. */
1741         if (err != -ESRCH)
1742                 return false;
1743
1744         /* If there's only one thread, let it fail. */
1745         if (threads->nr == 1)
1746                 return false;
1747
1748         /*
1749          * We should remove fd for missing_thread first
1750          * because thread_map__remove() will decrease threads->nr.
1751          */
1752         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1753                 return false;
1754
1755         if (thread_map__remove(threads, thread))
1756                 return false;
1757
1758         pr_warning("WARNING: Ignored open failure for pid %d\n",
1759                    ignore_pid);
1760         return true;
1761 }
1762
1763 static void display_attr(struct perf_event_attr *attr)
1764 {
1765         if (verbose >= 2) {
1766                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1767                 fprintf(stderr, "perf_event_attr:\n");
1768                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1769                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1770         }
1771 }
1772
1773 static int perf_event_open(struct perf_evsel *evsel,
1774                            pid_t pid, int cpu, int group_fd,
1775                            unsigned long flags)
1776 {
1777         int precise_ip = evsel->attr.precise_ip;
1778         int fd;
1779
1780         while (1) {
1781                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1782                           pid, cpu, group_fd, flags);
1783
1784                 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1785                 if (fd >= 0)
1786                         break;
1787
1788                 /*
1789                  * Do quick precise_ip fallback if:
1790                  *  - there is precise_ip set in perf_event_attr
1791                  *  - maximum precise is requested
1792                  *  - sys_perf_event_open failed with ENOTSUP error,
1793                  *    which is associated with wrong precise_ip
1794                  */
1795                 if (!precise_ip || !evsel->precise_max || (errno != ENOTSUP))
1796                         break;
1797
1798                 /*
1799                  * We tried all the precise_ip values, and it's
1800                  * still failing, so leave it to standard fallback.
1801                  */
1802                 if (!evsel->attr.precise_ip) {
1803                         evsel->attr.precise_ip = precise_ip;
1804                         break;
1805                 }
1806
1807                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1808                 evsel->attr.precise_ip--;
1809                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1810                 display_attr(&evsel->attr);
1811         }
1812
1813         return fd;
1814 }
1815
1816 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1817                      struct thread_map *threads)
1818 {
1819         int cpu, thread, nthreads;
1820         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1821         int pid = -1, err;
1822         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1823
1824         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1825                 return -EINVAL;
1826
1827         if (cpus == NULL) {
1828                 static struct cpu_map *empty_cpu_map;
1829
1830                 if (empty_cpu_map == NULL) {
1831                         empty_cpu_map = cpu_map__dummy_new();
1832                         if (empty_cpu_map == NULL)
1833                                 return -ENOMEM;
1834                 }
1835
1836                 cpus = empty_cpu_map;
1837         }
1838
1839         if (threads == NULL) {
1840                 static struct thread_map *empty_thread_map;
1841
1842                 if (empty_thread_map == NULL) {
1843                         empty_thread_map = thread_map__new_by_tid(-1);
1844                         if (empty_thread_map == NULL)
1845                                 return -ENOMEM;
1846                 }
1847
1848                 threads = empty_thread_map;
1849         }
1850
1851         if (evsel->system_wide)
1852                 nthreads = 1;
1853         else
1854                 nthreads = threads->nr;
1855
1856         if (evsel->fd == NULL &&
1857             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1858                 return -ENOMEM;
1859
1860         if (evsel->cgrp) {
1861                 flags |= PERF_FLAG_PID_CGROUP;
1862                 pid = evsel->cgrp->fd;
1863         }
1864
1865 fallback_missing_features:
1866         if (perf_missing_features.clockid_wrong)
1867                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1868         if (perf_missing_features.clockid) {
1869                 evsel->attr.use_clockid = 0;
1870                 evsel->attr.clockid = 0;
1871         }
1872         if (perf_missing_features.cloexec)
1873                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1874         if (perf_missing_features.mmap2)
1875                 evsel->attr.mmap2 = 0;
1876         if (perf_missing_features.exclude_guest)
1877                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1878         if (perf_missing_features.lbr_flags)
1879                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1880                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1881         if (perf_missing_features.group_read && evsel->attr.inherit)
1882                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1883         if (perf_missing_features.ksymbol)
1884                 evsel->attr.ksymbol = 0;
1885         if (perf_missing_features.bpf_event)
1886                 evsel->attr.bpf_event = 0;
1887 retry_sample_id:
1888         if (perf_missing_features.sample_id_all)
1889                 evsel->attr.sample_id_all = 0;
1890
1891         display_attr(&evsel->attr);
1892
1893         for (cpu = 0; cpu < cpus->nr; cpu++) {
1894
1895                 for (thread = 0; thread < nthreads; thread++) {
1896                         int fd, group_fd;
1897
1898                         if (!evsel->cgrp && !evsel->system_wide)
1899                                 pid = thread_map__pid(threads, thread);
1900
1901                         group_fd = get_group_fd(evsel, cpu, thread);
1902 retry_open:
1903                         test_attr__ready();
1904
1905                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1906                                              group_fd, flags);
1907
1908                         FD(evsel, cpu, thread) = fd;
1909
1910                         if (fd < 0) {
1911                                 err = -errno;
1912
1913                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1914                                         /*
1915                                          * We just removed 1 thread, so take a step
1916                                          * back on thread index and lower the upper
1917                                          * nthreads limit.
1918                                          */
1919                                         nthreads--;
1920                                         thread--;
1921
1922                                         /* ... and pretend like nothing have happened. */
1923                                         err = 0;
1924                                         continue;
1925                                 }
1926
1927                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1928                                           err);
1929                                 goto try_fallback;
1930                         }
1931
1932                         pr_debug2(" = %d\n", fd);
1933
1934                         if (evsel->bpf_fd >= 0) {
1935                                 int evt_fd = fd;
1936                                 int bpf_fd = evsel->bpf_fd;
1937
1938                                 err = ioctl(evt_fd,
1939                                             PERF_EVENT_IOC_SET_BPF,
1940                                             bpf_fd);
1941                                 if (err && errno != EEXIST) {
1942                                         pr_err("failed to attach bpf fd %d: %s\n",
1943                                                bpf_fd, strerror(errno));
1944                                         err = -EINVAL;
1945                                         goto out_close;
1946                                 }
1947                         }
1948
1949                         set_rlimit = NO_CHANGE;
1950
1951                         /*
1952                          * If we succeeded but had to kill clockid, fail and
1953                          * have perf_evsel__open_strerror() print us a nice
1954                          * error.
1955                          */
1956                         if (perf_missing_features.clockid ||
1957                             perf_missing_features.clockid_wrong) {
1958                                 err = -EINVAL;
1959                                 goto out_close;
1960                         }
1961                 }
1962         }
1963
1964         return 0;
1965
1966 try_fallback:
1967         /*
1968          * perf stat needs between 5 and 22 fds per CPU. When we run out
1969          * of them try to increase the limits.
1970          */
1971         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1972                 struct rlimit l;
1973                 int old_errno = errno;
1974
1975                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1976                         if (set_rlimit == NO_CHANGE)
1977                                 l.rlim_cur = l.rlim_max;
1978                         else {
1979                                 l.rlim_cur = l.rlim_max + 1000;
1980                                 l.rlim_max = l.rlim_cur;
1981                         }
1982                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1983                                 set_rlimit++;
1984                                 errno = old_errno;
1985                                 goto retry_open;
1986                         }
1987                 }
1988                 errno = old_errno;
1989         }
1990
1991         if (err != -EINVAL || cpu > 0 || thread > 0)
1992                 goto out_close;
1993
1994         /*
1995          * Must probe features in the order they were added to the
1996          * perf_event_attr interface.
1997          */
1998         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
1999                 perf_missing_features.bpf_event = true;
2000                 pr_debug2("switching off bpf_event\n");
2001                 goto fallback_missing_features;
2002         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
2003                 perf_missing_features.ksymbol = true;
2004                 pr_debug2("switching off ksymbol\n");
2005                 goto fallback_missing_features;
2006         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
2007                 perf_missing_features.write_backward = true;
2008                 pr_debug2("switching off write_backward\n");
2009                 goto out_close;
2010         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2011                 perf_missing_features.clockid_wrong = true;
2012                 pr_debug2("switching off clockid\n");
2013                 goto fallback_missing_features;
2014         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2015                 perf_missing_features.clockid = true;
2016                 pr_debug2("switching off use_clockid\n");
2017                 goto fallback_missing_features;
2018         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2019                 perf_missing_features.cloexec = true;
2020                 pr_debug2("switching off cloexec flag\n");
2021                 goto fallback_missing_features;
2022         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2023                 perf_missing_features.mmap2 = true;
2024                 pr_debug2("switching off mmap2\n");
2025                 goto fallback_missing_features;
2026         } else if (!perf_missing_features.exclude_guest &&
2027                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2028                 perf_missing_features.exclude_guest = true;
2029                 pr_debug2("switching off exclude_guest, exclude_host\n");
2030                 goto fallback_missing_features;
2031         } else if (!perf_missing_features.sample_id_all) {
2032                 perf_missing_features.sample_id_all = true;
2033                 pr_debug2("switching off sample_id_all\n");
2034                 goto retry_sample_id;
2035         } else if (!perf_missing_features.lbr_flags &&
2036                         (evsel->attr.branch_sample_type &
2037                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2038                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2039                 perf_missing_features.lbr_flags = true;
2040                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2041                 goto fallback_missing_features;
2042         } else if (!perf_missing_features.group_read &&
2043                     evsel->attr.inherit &&
2044                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2045                    perf_evsel__is_group_leader(evsel)) {
2046                 perf_missing_features.group_read = true;
2047                 pr_debug2("switching off group read\n");
2048                 goto fallback_missing_features;
2049         }
2050 out_close:
2051         if (err)
2052                 threads->err_thread = thread;
2053
2054         do {
2055                 while (--thread >= 0) {
2056                         close(FD(evsel, cpu, thread));
2057                         FD(evsel, cpu, thread) = -1;
2058                 }
2059                 thread = nthreads;
2060         } while (--cpu >= 0);
2061         return err;
2062 }
2063
2064 void perf_evsel__close(struct perf_evsel *evsel)
2065 {
2066         if (evsel->fd == NULL)
2067                 return;
2068
2069         perf_evsel__close_fd(evsel);
2070         perf_evsel__free_fd(evsel);
2071 }
2072
2073 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2074                              struct cpu_map *cpus)
2075 {
2076         return perf_evsel__open(evsel, cpus, NULL);
2077 }
2078
2079 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2080                                 struct thread_map *threads)
2081 {
2082         return perf_evsel__open(evsel, NULL, threads);
2083 }
2084
2085 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2086                                        const union perf_event *event,
2087                                        struct perf_sample *sample)
2088 {
2089         u64 type = evsel->attr.sample_type;
2090         const u64 *array = event->sample.array;
2091         bool swapped = evsel->needs_swap;
2092         union u64_swap u;
2093
2094         array += ((event->header.size -
2095                    sizeof(event->header)) / sizeof(u64)) - 1;
2096
2097         if (type & PERF_SAMPLE_IDENTIFIER) {
2098                 sample->id = *array;
2099                 array--;
2100         }
2101
2102         if (type & PERF_SAMPLE_CPU) {
2103                 u.val64 = *array;
2104                 if (swapped) {
2105                         /* undo swap of u64, then swap on individual u32s */
2106                         u.val64 = bswap_64(u.val64);
2107                         u.val32[0] = bswap_32(u.val32[0]);
2108                 }
2109
2110                 sample->cpu = u.val32[0];
2111                 array--;
2112         }
2113
2114         if (type & PERF_SAMPLE_STREAM_ID) {
2115                 sample->stream_id = *array;
2116                 array--;
2117         }
2118
2119         if (type & PERF_SAMPLE_ID) {
2120                 sample->id = *array;
2121                 array--;
2122         }
2123
2124         if (type & PERF_SAMPLE_TIME) {
2125                 sample->time = *array;
2126                 array--;
2127         }
2128
2129         if (type & PERF_SAMPLE_TID) {
2130                 u.val64 = *array;
2131                 if (swapped) {
2132                         /* undo swap of u64, then swap on individual u32s */
2133                         u.val64 = bswap_64(u.val64);
2134                         u.val32[0] = bswap_32(u.val32[0]);
2135                         u.val32[1] = bswap_32(u.val32[1]);
2136                 }
2137
2138                 sample->pid = u.val32[0];
2139                 sample->tid = u.val32[1];
2140                 array--;
2141         }
2142
2143         return 0;
2144 }
2145
2146 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2147                             u64 size)
2148 {
2149         return size > max_size || offset + size > endp;
2150 }
2151
2152 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2153         do {                                                            \
2154                 if (overflow(endp, (max_size), (offset), (size)))       \
2155                         return -EFAULT;                                 \
2156         } while (0)
2157
2158 #define OVERFLOW_CHECK_u64(offset) \
2159         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2160
2161 static int
2162 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2163 {
2164         /*
2165          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2166          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2167          * check the format does not go past the end of the event.
2168          */
2169         if (sample_size + sizeof(event->header) > event->header.size)
2170                 return -EFAULT;
2171
2172         return 0;
2173 }
2174
2175 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2176                              struct perf_sample *data)
2177 {
2178         u64 type = evsel->attr.sample_type;
2179         bool swapped = evsel->needs_swap;
2180         const u64 *array;
2181         u16 max_size = event->header.size;
2182         const void *endp = (void *)event + max_size;
2183         u64 sz;
2184
2185         /*
2186          * used for cross-endian analysis. See git commit 65014ab3
2187          * for why this goofiness is needed.
2188          */
2189         union u64_swap u;
2190
2191         memset(data, 0, sizeof(*data));
2192         data->cpu = data->pid = data->tid = -1;
2193         data->stream_id = data->id = data->time = -1ULL;
2194         data->period = evsel->attr.sample_period;
2195         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2196         data->misc    = event->header.misc;
2197         data->id = -1ULL;
2198         data->data_src = PERF_MEM_DATA_SRC_NONE;
2199
2200         if (event->header.type != PERF_RECORD_SAMPLE) {
2201                 if (!evsel->attr.sample_id_all)
2202                         return 0;
2203                 return perf_evsel__parse_id_sample(evsel, event, data);
2204         }
2205
2206         array = event->sample.array;
2207
2208         if (perf_event__check_size(event, evsel->sample_size))
2209                 return -EFAULT;
2210
2211         if (type & PERF_SAMPLE_IDENTIFIER) {
2212                 data->id = *array;
2213                 array++;
2214         }
2215
2216         if (type & PERF_SAMPLE_IP) {
2217                 data->ip = *array;
2218                 array++;
2219         }
2220
2221         if (type & PERF_SAMPLE_TID) {
2222                 u.val64 = *array;
2223                 if (swapped) {
2224                         /* undo swap of u64, then swap on individual u32s */
2225                         u.val64 = bswap_64(u.val64);
2226                         u.val32[0] = bswap_32(u.val32[0]);
2227                         u.val32[1] = bswap_32(u.val32[1]);
2228                 }
2229
2230                 data->pid = u.val32[0];
2231                 data->tid = u.val32[1];
2232                 array++;
2233         }
2234
2235         if (type & PERF_SAMPLE_TIME) {
2236                 data->time = *array;
2237                 array++;
2238         }
2239
2240         if (type & PERF_SAMPLE_ADDR) {
2241                 data->addr = *array;
2242                 array++;
2243         }
2244
2245         if (type & PERF_SAMPLE_ID) {
2246                 data->id = *array;
2247                 array++;
2248         }
2249
2250         if (type & PERF_SAMPLE_STREAM_ID) {
2251                 data->stream_id = *array;
2252                 array++;
2253         }
2254
2255         if (type & PERF_SAMPLE_CPU) {
2256
2257                 u.val64 = *array;
2258                 if (swapped) {
2259                         /* undo swap of u64, then swap on individual u32s */
2260                         u.val64 = bswap_64(u.val64);
2261                         u.val32[0] = bswap_32(u.val32[0]);
2262                 }
2263
2264                 data->cpu = u.val32[0];
2265                 array++;
2266         }
2267
2268         if (type & PERF_SAMPLE_PERIOD) {
2269                 data->period = *array;
2270                 array++;
2271         }
2272
2273         if (type & PERF_SAMPLE_READ) {
2274                 u64 read_format = evsel->attr.read_format;
2275
2276                 OVERFLOW_CHECK_u64(array);
2277                 if (read_format & PERF_FORMAT_GROUP)
2278                         data->read.group.nr = *array;
2279                 else
2280                         data->read.one.value = *array;
2281
2282                 array++;
2283
2284                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2285                         OVERFLOW_CHECK_u64(array);
2286                         data->read.time_enabled = *array;
2287                         array++;
2288                 }
2289
2290                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2291                         OVERFLOW_CHECK_u64(array);
2292                         data->read.time_running = *array;
2293                         array++;
2294                 }
2295
2296                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2297                 if (read_format & PERF_FORMAT_GROUP) {
2298                         const u64 max_group_nr = UINT64_MAX /
2299                                         sizeof(struct sample_read_value);
2300
2301                         if (data->read.group.nr > max_group_nr)
2302                                 return -EFAULT;
2303                         sz = data->read.group.nr *
2304                              sizeof(struct sample_read_value);
2305                         OVERFLOW_CHECK(array, sz, max_size);
2306                         data->read.group.values =
2307                                         (struct sample_read_value *)array;
2308                         array = (void *)array + sz;
2309                 } else {
2310                         OVERFLOW_CHECK_u64(array);
2311                         data->read.one.id = *array;
2312                         array++;
2313                 }
2314         }
2315
2316         if (evsel__has_callchain(evsel)) {
2317                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2318
2319                 OVERFLOW_CHECK_u64(array);
2320                 data->callchain = (struct ip_callchain *)array++;
2321                 if (data->callchain->nr > max_callchain_nr)
2322                         return -EFAULT;
2323                 sz = data->callchain->nr * sizeof(u64);
2324                 OVERFLOW_CHECK(array, sz, max_size);
2325                 array = (void *)array + sz;
2326         }
2327
2328         if (type & PERF_SAMPLE_RAW) {
2329                 OVERFLOW_CHECK_u64(array);
2330                 u.val64 = *array;
2331
2332                 /*
2333                  * Undo swap of u64, then swap on individual u32s,
2334                  * get the size of the raw area and undo all of the
2335                  * swap. The pevent interface handles endianity by
2336                  * itself.
2337                  */
2338                 if (swapped) {
2339                         u.val64 = bswap_64(u.val64);
2340                         u.val32[0] = bswap_32(u.val32[0]);
2341                         u.val32[1] = bswap_32(u.val32[1]);
2342                 }
2343                 data->raw_size = u.val32[0];
2344
2345                 /*
2346                  * The raw data is aligned on 64bits including the
2347                  * u32 size, so it's safe to use mem_bswap_64.
2348                  */
2349                 if (swapped)
2350                         mem_bswap_64((void *) array, data->raw_size);
2351
2352                 array = (void *)array + sizeof(u32);
2353
2354                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2355                 data->raw_data = (void *)array;
2356                 array = (void *)array + data->raw_size;
2357         }
2358
2359         if (type & PERF_SAMPLE_BRANCH_STACK) {
2360                 const u64 max_branch_nr = UINT64_MAX /
2361                                           sizeof(struct branch_entry);
2362
2363                 OVERFLOW_CHECK_u64(array);
2364                 data->branch_stack = (struct branch_stack *)array++;
2365
2366                 if (data->branch_stack->nr > max_branch_nr)
2367                         return -EFAULT;
2368                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2369                 OVERFLOW_CHECK(array, sz, max_size);
2370                 array = (void *)array + sz;
2371         }
2372
2373         if (type & PERF_SAMPLE_REGS_USER) {
2374                 OVERFLOW_CHECK_u64(array);
2375                 data->user_regs.abi = *array;
2376                 array++;
2377
2378                 if (data->user_regs.abi) {
2379                         u64 mask = evsel->attr.sample_regs_user;
2380
2381                         sz = hweight64(mask) * sizeof(u64);
2382                         OVERFLOW_CHECK(array, sz, max_size);
2383                         data->user_regs.mask = mask;
2384                         data->user_regs.regs = (u64 *)array;
2385                         array = (void *)array + sz;
2386                 }
2387         }
2388
2389         if (type & PERF_SAMPLE_STACK_USER) {
2390                 OVERFLOW_CHECK_u64(array);
2391                 sz = *array++;
2392
2393                 data->user_stack.offset = ((char *)(array - 1)
2394                                           - (char *) event);
2395
2396                 if (!sz) {
2397                         data->user_stack.size = 0;
2398                 } else {
2399                         OVERFLOW_CHECK(array, sz, max_size);
2400                         data->user_stack.data = (char *)array;
2401                         array = (void *)array + sz;
2402                         OVERFLOW_CHECK_u64(array);
2403                         data->user_stack.size = *array++;
2404                         if (WARN_ONCE(data->user_stack.size > sz,
2405                                       "user stack dump failure\n"))
2406                                 return -EFAULT;
2407                 }
2408         }
2409
2410         if (type & PERF_SAMPLE_WEIGHT) {
2411                 OVERFLOW_CHECK_u64(array);
2412                 data->weight = *array;
2413                 array++;
2414         }
2415
2416         if (type & PERF_SAMPLE_DATA_SRC) {
2417                 OVERFLOW_CHECK_u64(array);
2418                 data->data_src = *array;
2419                 array++;
2420         }
2421
2422         if (type & PERF_SAMPLE_TRANSACTION) {
2423                 OVERFLOW_CHECK_u64(array);
2424                 data->transaction = *array;
2425                 array++;
2426         }
2427
2428         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2429         if (type & PERF_SAMPLE_REGS_INTR) {
2430                 OVERFLOW_CHECK_u64(array);
2431                 data->intr_regs.abi = *array;
2432                 array++;
2433
2434                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2435                         u64 mask = evsel->attr.sample_regs_intr;
2436
2437                         sz = hweight64(mask) * sizeof(u64);
2438                         OVERFLOW_CHECK(array, sz, max_size);
2439                         data->intr_regs.mask = mask;
2440                         data->intr_regs.regs = (u64 *)array;
2441                         array = (void *)array + sz;
2442                 }
2443         }
2444
2445         data->phys_addr = 0;
2446         if (type & PERF_SAMPLE_PHYS_ADDR) {
2447                 data->phys_addr = *array;
2448                 array++;
2449         }
2450
2451         return 0;
2452 }
2453
2454 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2455                                        union perf_event *event,
2456                                        u64 *timestamp)
2457 {
2458         u64 type = evsel->attr.sample_type;
2459         const u64 *array;
2460
2461         if (!(type & PERF_SAMPLE_TIME))
2462                 return -1;
2463
2464         if (event->header.type != PERF_RECORD_SAMPLE) {
2465                 struct perf_sample data = {
2466                         .time = -1ULL,
2467                 };
2468
2469                 if (!evsel->attr.sample_id_all)
2470                         return -1;
2471                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2472                         return -1;
2473
2474                 *timestamp = data.time;
2475                 return 0;
2476         }
2477
2478         array = event->sample.array;
2479
2480         if (perf_event__check_size(event, evsel->sample_size))
2481                 return -EFAULT;
2482
2483         if (type & PERF_SAMPLE_IDENTIFIER)
2484                 array++;
2485
2486         if (type & PERF_SAMPLE_IP)
2487                 array++;
2488
2489         if (type & PERF_SAMPLE_TID)
2490                 array++;
2491
2492         if (type & PERF_SAMPLE_TIME)
2493                 *timestamp = *array;
2494
2495         return 0;
2496 }
2497
2498 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2499                                      u64 read_format)
2500 {
2501         size_t sz, result = sizeof(struct sample_event);
2502
2503         if (type & PERF_SAMPLE_IDENTIFIER)
2504                 result += sizeof(u64);
2505
2506         if (type & PERF_SAMPLE_IP)
2507                 result += sizeof(u64);
2508
2509         if (type & PERF_SAMPLE_TID)
2510                 result += sizeof(u64);
2511
2512         if (type & PERF_SAMPLE_TIME)
2513                 result += sizeof(u64);
2514
2515         if (type & PERF_SAMPLE_ADDR)
2516                 result += sizeof(u64);
2517
2518         if (type & PERF_SAMPLE_ID)
2519                 result += sizeof(u64);
2520
2521         if (type & PERF_SAMPLE_STREAM_ID)
2522                 result += sizeof(u64);
2523
2524         if (type & PERF_SAMPLE_CPU)
2525                 result += sizeof(u64);
2526
2527         if (type & PERF_SAMPLE_PERIOD)
2528                 result += sizeof(u64);
2529
2530         if (type & PERF_SAMPLE_READ) {
2531                 result += sizeof(u64);
2532                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2533                         result += sizeof(u64);
2534                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2535                         result += sizeof(u64);
2536                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2537                 if (read_format & PERF_FORMAT_GROUP) {
2538                         sz = sample->read.group.nr *
2539                              sizeof(struct sample_read_value);
2540                         result += sz;
2541                 } else {
2542                         result += sizeof(u64);
2543                 }
2544         }
2545
2546         if (type & PERF_SAMPLE_CALLCHAIN) {
2547                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2548                 result += sz;
2549         }
2550
2551         if (type & PERF_SAMPLE_RAW) {
2552                 result += sizeof(u32);
2553                 result += sample->raw_size;
2554         }
2555
2556         if (type & PERF_SAMPLE_BRANCH_STACK) {
2557                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2558                 sz += sizeof(u64);
2559                 result += sz;
2560         }
2561
2562         if (type & PERF_SAMPLE_REGS_USER) {
2563                 if (sample->user_regs.abi) {
2564                         result += sizeof(u64);
2565                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2566                         result += sz;
2567                 } else {
2568                         result += sizeof(u64);
2569                 }
2570         }
2571
2572         if (type & PERF_SAMPLE_STACK_USER) {
2573                 sz = sample->user_stack.size;
2574                 result += sizeof(u64);
2575                 if (sz) {
2576                         result += sz;
2577                         result += sizeof(u64);
2578                 }
2579         }
2580
2581         if (type & PERF_SAMPLE_WEIGHT)
2582                 result += sizeof(u64);
2583
2584         if (type & PERF_SAMPLE_DATA_SRC)
2585                 result += sizeof(u64);
2586
2587         if (type & PERF_SAMPLE_TRANSACTION)
2588                 result += sizeof(u64);
2589
2590         if (type & PERF_SAMPLE_REGS_INTR) {
2591                 if (sample->intr_regs.abi) {
2592                         result += sizeof(u64);
2593                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2594                         result += sz;
2595                 } else {
2596                         result += sizeof(u64);
2597                 }
2598         }
2599
2600         if (type & PERF_SAMPLE_PHYS_ADDR)
2601                 result += sizeof(u64);
2602
2603         return result;
2604 }
2605
2606 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2607                                   u64 read_format,
2608                                   const struct perf_sample *sample)
2609 {
2610         u64 *array;
2611         size_t sz;
2612         /*
2613          * used for cross-endian analysis. See git commit 65014ab3
2614          * for why this goofiness is needed.
2615          */
2616         union u64_swap u;
2617
2618         array = event->sample.array;
2619
2620         if (type & PERF_SAMPLE_IDENTIFIER) {
2621                 *array = sample->id;
2622                 array++;
2623         }
2624
2625         if (type & PERF_SAMPLE_IP) {
2626                 *array = sample->ip;
2627                 array++;
2628         }
2629
2630         if (type & PERF_SAMPLE_TID) {
2631                 u.val32[0] = sample->pid;
2632                 u.val32[1] = sample->tid;
2633                 *array = u.val64;
2634                 array++;
2635         }
2636
2637         if (type & PERF_SAMPLE_TIME) {
2638                 *array = sample->time;
2639                 array++;
2640         }
2641
2642         if (type & PERF_SAMPLE_ADDR) {
2643                 *array = sample->addr;
2644                 array++;
2645         }
2646
2647         if (type & PERF_SAMPLE_ID) {
2648                 *array = sample->id;
2649                 array++;
2650         }
2651
2652         if (type & PERF_SAMPLE_STREAM_ID) {
2653                 *array = sample->stream_id;
2654                 array++;
2655         }
2656
2657         if (type & PERF_SAMPLE_CPU) {
2658                 u.val32[0] = sample->cpu;
2659                 u.val32[1] = 0;
2660                 *array = u.val64;
2661                 array++;
2662         }
2663
2664         if (type & PERF_SAMPLE_PERIOD) {
2665                 *array = sample->period;
2666                 array++;
2667         }
2668
2669         if (type & PERF_SAMPLE_READ) {
2670                 if (read_format & PERF_FORMAT_GROUP)
2671                         *array = sample->read.group.nr;
2672                 else
2673                         *array = sample->read.one.value;
2674                 array++;
2675
2676                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2677                         *array = sample->read.time_enabled;
2678                         array++;
2679                 }
2680
2681                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2682                         *array = sample->read.time_running;
2683                         array++;
2684                 }
2685
2686                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2687                 if (read_format & PERF_FORMAT_GROUP) {
2688                         sz = sample->read.group.nr *
2689                              sizeof(struct sample_read_value);
2690                         memcpy(array, sample->read.group.values, sz);
2691                         array = (void *)array + sz;
2692                 } else {
2693                         *array = sample->read.one.id;
2694                         array++;
2695                 }
2696         }
2697
2698         if (type & PERF_SAMPLE_CALLCHAIN) {
2699                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2700                 memcpy(array, sample->callchain, sz);
2701                 array = (void *)array + sz;
2702         }
2703
2704         if (type & PERF_SAMPLE_RAW) {
2705                 u.val32[0] = sample->raw_size;
2706                 *array = u.val64;
2707                 array = (void *)array + sizeof(u32);
2708
2709                 memcpy(array, sample->raw_data, sample->raw_size);
2710                 array = (void *)array + sample->raw_size;
2711         }
2712
2713         if (type & PERF_SAMPLE_BRANCH_STACK) {
2714                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2715                 sz += sizeof(u64);
2716                 memcpy(array, sample->branch_stack, sz);
2717                 array = (void *)array + sz;
2718         }
2719
2720         if (type & PERF_SAMPLE_REGS_USER) {
2721                 if (sample->user_regs.abi) {
2722                         *array++ = sample->user_regs.abi;
2723                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2724                         memcpy(array, sample->user_regs.regs, sz);
2725                         array = (void *)array + sz;
2726                 } else {
2727                         *array++ = 0;
2728                 }
2729         }
2730
2731         if (type & PERF_SAMPLE_STACK_USER) {
2732                 sz = sample->user_stack.size;
2733                 *array++ = sz;
2734                 if (sz) {
2735                         memcpy(array, sample->user_stack.data, sz);
2736                         array = (void *)array + sz;
2737                         *array++ = sz;
2738                 }
2739         }
2740
2741         if (type & PERF_SAMPLE_WEIGHT) {
2742                 *array = sample->weight;
2743                 array++;
2744         }
2745
2746         if (type & PERF_SAMPLE_DATA_SRC) {
2747                 *array = sample->data_src;
2748                 array++;
2749         }
2750
2751         if (type & PERF_SAMPLE_TRANSACTION) {
2752                 *array = sample->transaction;
2753                 array++;
2754         }
2755
2756         if (type & PERF_SAMPLE_REGS_INTR) {
2757                 if (sample->intr_regs.abi) {
2758                         *array++ = sample->intr_regs.abi;
2759                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2760                         memcpy(array, sample->intr_regs.regs, sz);
2761                         array = (void *)array + sz;
2762                 } else {
2763                         *array++ = 0;
2764                 }
2765         }
2766
2767         if (type & PERF_SAMPLE_PHYS_ADDR) {
2768                 *array = sample->phys_addr;
2769                 array++;
2770         }
2771
2772         return 0;
2773 }
2774
2775 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2776 {
2777         return tep_find_field(evsel->tp_format, name);
2778 }
2779
2780 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2781                          const char *name)
2782 {
2783         struct tep_format_field *field = perf_evsel__field(evsel, name);
2784         int offset;
2785
2786         if (!field)
2787                 return NULL;
2788
2789         offset = field->offset;
2790
2791         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2792                 offset = *(int *)(sample->raw_data + field->offset);
2793                 offset &= 0xffff;
2794         }
2795
2796         return sample->raw_data + offset;
2797 }
2798
2799 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2800                          bool needs_swap)
2801 {
2802         u64 value;
2803         void *ptr = sample->raw_data + field->offset;
2804
2805         switch (field->size) {
2806         case 1:
2807                 return *(u8 *)ptr;
2808         case 2:
2809                 value = *(u16 *)ptr;
2810                 break;
2811         case 4:
2812                 value = *(u32 *)ptr;
2813                 break;
2814         case 8:
2815                 memcpy(&value, ptr, sizeof(u64));
2816                 break;
2817         default:
2818                 return 0;
2819         }
2820
2821         if (!needs_swap)
2822                 return value;
2823
2824         switch (field->size) {
2825         case 2:
2826                 return bswap_16(value);
2827         case 4:
2828                 return bswap_32(value);
2829         case 8:
2830                 return bswap_64(value);
2831         default:
2832                 return 0;
2833         }
2834
2835         return 0;
2836 }
2837
2838 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2839                        const char *name)
2840 {
2841         struct tep_format_field *field = perf_evsel__field(evsel, name);
2842
2843         if (!field)
2844                 return 0;
2845
2846         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2847 }
2848
2849 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2850                           char *msg, size_t msgsize)
2851 {
2852         int paranoid;
2853
2854         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2855             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2856             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2857                 /*
2858                  * If it's cycles then fall back to hrtimer based
2859                  * cpu-clock-tick sw counter, which is always available even if
2860                  * no PMU support.
2861                  *
2862                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2863                  * b0a873e).
2864                  */
2865                 scnprintf(msg, msgsize, "%s",
2866 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2867
2868                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2869                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2870
2871                 zfree(&evsel->name);
2872                 return true;
2873         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2874                    (paranoid = perf_event_paranoid()) > 1) {
2875                 const char *name = perf_evsel__name(evsel);
2876                 char *new_name;
2877                 const char *sep = ":";
2878
2879                 /* Is there already the separator in the name. */
2880                 if (strchr(name, '/') ||
2881                     strchr(name, ':'))
2882                         sep = "";
2883
2884                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2885                         return false;
2886
2887                 if (evsel->name)
2888                         free(evsel->name);
2889                 evsel->name = new_name;
2890                 scnprintf(msg, msgsize,
2891 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2892                 evsel->attr.exclude_kernel = 1;
2893
2894                 return true;
2895         }
2896
2897         return false;
2898 }
2899
2900 static bool find_process(const char *name)
2901 {
2902         size_t len = strlen(name);
2903         DIR *dir;
2904         struct dirent *d;
2905         int ret = -1;
2906
2907         dir = opendir(procfs__mountpoint());
2908         if (!dir)
2909                 return false;
2910
2911         /* Walk through the directory. */
2912         while (ret && (d = readdir(dir)) != NULL) {
2913                 char path[PATH_MAX];
2914                 char *data;
2915                 size_t size;
2916
2917                 if ((d->d_type != DT_DIR) ||
2918                      !strcmp(".", d->d_name) ||
2919                      !strcmp("..", d->d_name))
2920                         continue;
2921
2922                 scnprintf(path, sizeof(path), "%s/%s/comm",
2923                           procfs__mountpoint(), d->d_name);
2924
2925                 if (filename__read_str(path, &data, &size))
2926                         continue;
2927
2928                 ret = strncmp(name, data, len);
2929                 free(data);
2930         }
2931
2932         closedir(dir);
2933         return ret ? false : true;
2934 }
2935
2936 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2937                               int err, char *msg, size_t size)
2938 {
2939         char sbuf[STRERR_BUFSIZE];
2940         int printed = 0;
2941
2942         switch (err) {
2943         case EPERM:
2944         case EACCES:
2945                 if (err == EPERM)
2946                         printed = scnprintf(msg, size,
2947                                 "No permission to enable %s event.\n\n",
2948                                 perf_evsel__name(evsel));
2949
2950                 return scnprintf(msg + printed, size - printed,
2951                  "You may not have permission to collect %sstats.\n\n"
2952                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2953                  "which controls use of the performance events system by\n"
2954                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2955                  "The current value is %d:\n\n"
2956                  "  -1: Allow use of (almost) all events by all users\n"
2957                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2958                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2959                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2960                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2961                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2962                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2963                  "      kernel.perf_event_paranoid = -1\n" ,
2964                                  target->system_wide ? "system-wide " : "",
2965                                  perf_event_paranoid());
2966         case ENOENT:
2967                 return scnprintf(msg, size, "The %s event is not supported.",
2968                                  perf_evsel__name(evsel));
2969         case EMFILE:
2970                 return scnprintf(msg, size, "%s",
2971                          "Too many events are opened.\n"
2972                          "Probably the maximum number of open file descriptors has been reached.\n"
2973                          "Hint: Try again after reducing the number of events.\n"
2974                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2975         case ENOMEM:
2976                 if (evsel__has_callchain(evsel) &&
2977                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2978                         return scnprintf(msg, size,
2979                                          "Not enough memory to setup event with callchain.\n"
2980                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2981                                          "Hint: Current value: %d", sysctl__max_stack());
2982                 break;
2983         case ENODEV:
2984                 if (target->cpu_list)
2985                         return scnprintf(msg, size, "%s",
2986          "No such device - did you specify an out-of-range profile CPU?");
2987                 break;
2988         case EOPNOTSUPP:
2989                 if (evsel->attr.sample_period != 0)
2990                         return scnprintf(msg, size,
2991         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2992                                          perf_evsel__name(evsel));
2993                 if (evsel->attr.precise_ip)
2994                         return scnprintf(msg, size, "%s",
2995         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2996 #if defined(__i386__) || defined(__x86_64__)
2997                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2998                         return scnprintf(msg, size, "%s",
2999         "No hardware sampling interrupt available.\n");
3000 #endif
3001                 break;
3002         case EBUSY:
3003                 if (find_process("oprofiled"))
3004                         return scnprintf(msg, size,
3005         "The PMU counters are busy/taken by another profiler.\n"
3006         "We found oprofile daemon running, please stop it and try again.");
3007                 break;
3008         case EINVAL:
3009                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3010                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3011                 if (perf_missing_features.clockid)
3012                         return scnprintf(msg, size, "clockid feature not supported.");
3013                 if (perf_missing_features.clockid_wrong)
3014                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3015                 break;
3016         default:
3017                 break;
3018         }
3019
3020         return scnprintf(msg, size,
3021         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3022         "/bin/dmesg | grep -i perf may provide additional information.\n",
3023                          err, str_error_r(err, sbuf, sizeof(sbuf)),
3024                          perf_evsel__name(evsel));
3025 }
3026
3027 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3028 {
3029         if (evsel && evsel->evlist)
3030                 return evsel->evlist->env;
3031         return NULL;
3032 }
3033
3034 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3035 {
3036         int cpu, thread;
3037
3038         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3039                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3040                      thread++) {
3041                         int fd = FD(evsel, cpu, thread);
3042
3043                         if (perf_evlist__id_add_fd(evlist, evsel,
3044                                                    cpu, thread, fd) < 0)
3045                                 return -1;
3046                 }
3047         }
3048
3049         return 0;
3050 }
3051
3052 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3053 {
3054         struct cpu_map *cpus = evsel->cpus;
3055         struct thread_map *threads = evsel->threads;
3056
3057         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3058                 return -ENOMEM;
3059
3060         return store_evsel_ids(evsel, evlist);
3061 }