tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/err.h>
  19 #include <sys/ioctl.h>
  20 #include <sys/resource.h>
  21 #include "asm/bug.h"
  22 #include "callchain.h"
  23 #include "cgroup.h"
  24 #include "event.h"
  25 #include "evsel.h"
  26 #include "evlist.h"
  27 #include "util.h"
  28 #include "cpumap.h"
  29 #include "thread_map.h"
  30 #include "target.h"
  31 #include "perf_regs.h"
  32 #include "debug.h"
  33 #include "trace-event.h"
  34 #include "stat.h"
  35 #include "util/parse-branch-options.h"
  36
  37 #include "sane_ctype.h"
  38
  39 static struct {
  40         bool sample_id_all;
  41         bool exclude_guest;
  42         bool mmap2;
  43         bool cloexec;
  44         bool clockid;
  45         bool clockid_wrong;
  46         bool lbr_flags;
  47         bool write_backward;
  48 } perf_missing_features;
  49
  50 static clockid_t clockid;
  51
  52 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  53 {
  54         return 0;
  55 }
  56
  57 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  58 {
  59 }
  60
  61 static struct {
  62         size_t  size;
  63         int     (*init)(struct perf_evsel *evsel);
  64         void    (*fini)(struct perf_evsel *evsel);
  65 } perf_evsel__object = {
  66         .size = sizeof(struct perf_evsel),
  67         .init = perf_evsel__no_extra_init,
  68         .fini = perf_evsel__no_extra_fini,
  69 };
  70
  71 int perf_evsel__object_config(size_t object_size,
  72                               int (*init)(struct perf_evsel *evsel),
  73                               void (*fini)(struct perf_evsel *evsel))
  74 {
  75
  76         if (object_size == 0)
  77                 goto set_methods;
  78
  79         if (perf_evsel__object.size > object_size)
  80                 return -EINVAL;
  81
  82         perf_evsel__object.size = object_size;
  83
  84 set_methods:
  85         if (init != NULL)
  86                 perf_evsel__object.init = init;
  87
  88         if (fini != NULL)
  89                 perf_evsel__object.fini = fini;
  90
  91         return 0;
  92 }
  93
  94 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  95
  96 int __perf_evsel__sample_size(u64 sample_type)
  97 {
  98         u64 mask = sample_type & PERF_SAMPLE_MASK;
  99         int size = 0;
 100         int i;
 101
 102         for (i = 0; i < 64; i++) {
 103                 if (mask & (1ULL << i))
 104                         size++;
 105         }
 106
 107         size *= sizeof(u64);
 108
 109         return size;
 110 }
 111
 112 /**
 113  * __perf_evsel__calc_id_pos - calculate id_pos.
 114  * @sample_type: sample type
 115  *
 116  * This function returns the position of the event id (PERF_SAMPLE_ID or
 117  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 118  * sample_event.
 119  */
 120 static int __perf_evsel__calc_id_pos(u64 sample_type)
 121 {
 122         int idx = 0;
 123
 124         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 125                 return 0;
 126
 127         if (!(sample_type & PERF_SAMPLE_ID))
 128                 return -1;
 129
 130         if (sample_type & PERF_SAMPLE_IP)
 131                 idx += 1;
 132
 133         if (sample_type & PERF_SAMPLE_TID)
 134                 idx += 1;
 135
 136         if (sample_type & PERF_SAMPLE_TIME)
 137                 idx += 1;
 138
 139         if (sample_type & PERF_SAMPLE_ADDR)
 140                 idx += 1;
 141
 142         return idx;
 143 }
 144
 145 /**
 146  * __perf_evsel__calc_is_pos - calculate is_pos.
 147  * @sample_type: sample type
 148  *
 149  * This function returns the position (counting backwards) of the event id
 150  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 151  * sample_id_all is used there is an id sample appended to non-sample events.
 152  */
 153 static int __perf_evsel__calc_is_pos(u64 sample_type)
 154 {
 155         int idx = 1;
 156
 157         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 158                 return 1;
 159
 160         if (!(sample_type & PERF_SAMPLE_ID))
 161                 return -1;
 162
 163         if (sample_type & PERF_SAMPLE_CPU)
 164                 idx += 1;
 165
 166         if (sample_type & PERF_SAMPLE_STREAM_ID)
 167                 idx += 1;
 168
 169         return idx;
 170 }
 171
 172 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 173 {
 174         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 175         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 176 }
 177
 178 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 179                                   enum perf_event_sample_format bit)
 180 {
 181         if (!(evsel->attr.sample_type & bit)) {
 182                 evsel->attr.sample_type |= bit;
 183                 evsel->sample_size += sizeof(u64);
 184                 perf_evsel__calc_id_pos(evsel);
 185         }
 186 }
 187
 188 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 189                                     enum perf_event_sample_format bit)
 190 {
 191         if (evsel->attr.sample_type & bit) {
 192                 evsel->attr.sample_type &= ~bit;
 193                 evsel->sample_size -= sizeof(u64);
 194                 perf_evsel__calc_id_pos(evsel);
 195         }
 196 }
 197
 198 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 199                                bool can_sample_identifier)
 200 {
 201         if (can_sample_identifier) {
 202                 perf_evsel__reset_sample_bit(evsel, ID);
 203                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 204         } else {
 205                 perf_evsel__set_sample_bit(evsel, ID);
 206         }
 207         evsel->attr.read_format |= PERF_FORMAT_ID;
 208 }
 209
 210 /**
 211  * perf_evsel__is_function_event - Return whether given evsel is a function
 212  * trace event
 213  *
 214  * @evsel - evsel selector to be tested
 215  *
 216  * Return %true if event is function trace event
 217  */
 218 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 219 {
 220 #define FUNCTION_EVENT "ftrace:function"
 221
 222         return evsel->name &&
 223                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 224
 225 #undef FUNCTION_EVENT
 226 }
 227
 228 void perf_evsel__init(struct perf_evsel *evsel,
 229                       struct perf_event_attr *attr, int idx)
 230 {
 231         evsel->idx         = idx;
 232         evsel->tracking    = !idx;
 233         evsel->attr        = *attr;
 234         evsel->leader      = evsel;
 235         evsel->unit        = "";
 236         evsel->scale       = 1.0;
 237         evsel->evlist      = NULL;
 238         evsel->bpf_fd      = -1;
 239         INIT_LIST_HEAD(&evsel->node);
 240         INIT_LIST_HEAD(&evsel->config_terms);
 241         perf_evsel__object.init(evsel);
 242         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 243         perf_evsel__calc_id_pos(evsel);
 244         evsel->cmdline_group_boundary = false;
 245         evsel->metric_expr   = NULL;
 246         evsel->metric_name   = NULL;
 247         evsel->metric_events = NULL;
 248         evsel->collect_stat  = false;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (evsel != 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         return evsel;
 265 }
 266
 267 struct perf_evsel *perf_evsel__new_cycles(void)
 268 {
 269         struct perf_event_attr attr = {
 270                 .type   = PERF_TYPE_HARDWARE,
 271                 .config = PERF_COUNT_HW_CPU_CYCLES,
 272         };
 273         struct perf_evsel *evsel;
 274
 275         event_attr_init(&attr);
 276         /*
 277          * Unnamed union member, not supported as struct member named
 278          * initializer in older compilers such as gcc 4.4.7
 279          *
 280          * Just for probing the precise_ip:
 281          */
 282         attr.sample_period = 1;
 283
 284         perf_event_attr__set_max_precise_ip(&attr);
 285         /*
 286          * Now let the usual logic to set up the perf_event_attr defaults
 287          * to kick in when we return and before perf_evsel__open() is called.
 288          */
 289         attr.sample_period = 0;
 290
 291         evsel = perf_evsel__new(&attr);
 292         if (evsel == NULL)
 293                 goto out;
 294
 295         /* use asprintf() because free(evsel) assumes name is allocated */
 296         if (asprintf(&evsel->name, "cycles%.*s",
 297                      attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
 298                 goto error_free;
 299 out:
 300         return evsel;
 301 error_free:
 302         perf_evsel__delete(evsel);
 303         evsel = NULL;
 304         goto out;
 305 }
 306
 307 /*
 308  * Returns pointer with encoded error via <linux/err.h> interface.
 309  */
 310 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 311 {
 312         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 313         int err = -ENOMEM;
 314
 315         if (evsel == NULL) {
 316                 goto out_err;
 317         } else {
 318                 struct perf_event_attr attr = {
 319                         .type          = PERF_TYPE_TRACEPOINT,
 320                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 321                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 322                 };
 323
 324                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 325                         goto out_free;
 326
 327                 evsel->tp_format = trace_event__tp_format(sys, name);
 328                 if (IS_ERR(evsel->tp_format)) {
 329                         err = PTR_ERR(evsel->tp_format);
 330                         goto out_free;
 331                 }
 332
 333                 event_attr_init(&attr);
 334                 attr.config = evsel->tp_format->id;
 335                 attr.sample_period = 1;
 336                 perf_evsel__init(evsel, &attr, idx);
 337         }
 338
 339         return evsel;
 340
 341 out_free:
 342         zfree(&evsel->name);
 343         free(evsel);
 344 out_err:
 345         return ERR_PTR(err);
 346 }
 347
 348 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 349         "cycles",
 350         "instructions",
 351         "cache-references",
 352         "cache-misses",
 353         "branches",
 354         "branch-misses",
 355         "bus-cycles",
 356         "stalled-cycles-frontend",
 357         "stalled-cycles-backend",
 358         "ref-cycles",
 359 };
 360
 361 static const char *__perf_evsel__hw_name(u64 config)
 362 {
 363         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 364                 return perf_evsel__hw_names[config];
 365
 366         return "unknown-hardware";
 367 }
 368
 369 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 370 {
 371         int colon = 0, r = 0;
 372         struct perf_event_attr *attr = &evsel->attr;
 373         bool exclude_guest_default = false;
 374
 375 #define MOD_PRINT(context, mod) do {                                    \
 376                 if (!attr->exclude_##context) {                         \
 377                         if (!colon) colon = ++r;                        \
 378                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 379                 } } while(0)
 380
 381         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 382                 MOD_PRINT(kernel, 'k');
 383                 MOD_PRINT(user, 'u');
 384                 MOD_PRINT(hv, 'h');
 385                 exclude_guest_default = true;
 386         }
 387
 388         if (attr->precise_ip) {
 389                 if (!colon)
 390                         colon = ++r;
 391                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 392                 exclude_guest_default = true;
 393         }
 394
 395         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 396                 MOD_PRINT(host, 'H');
 397                 MOD_PRINT(guest, 'G');
 398         }
 399 #undef MOD_PRINT
 400         if (colon)
 401                 bf[colon - 1] = ':';
 402         return r;
 403 }
 404
 405 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 406 {
 407         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 408         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 409 }
 410
 411 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 412         "cpu-clock",
 413         "task-clock",
 414         "page-faults",
 415         "context-switches",
 416         "cpu-migrations",
 417         "minor-faults",
 418         "major-faults",
 419         "alignment-faults",
 420         "emulation-faults",
 421         "dummy",
 422 };
 423
 424 static const char *__perf_evsel__sw_name(u64 config)
 425 {
 426         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 427                 return perf_evsel__sw_names[config];
 428         return "unknown-software";
 429 }
 430
 431 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 432 {
 433         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 434         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 435 }
 436
 437 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 438 {
 439         int r;
 440
 441         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 442
 443         if (type & HW_BREAKPOINT_R)
 444                 r += scnprintf(bf + r, size - r, "r");
 445
 446         if (type & HW_BREAKPOINT_W)
 447                 r += scnprintf(bf + r, size - r, "w");
 448
 449         if (type & HW_BREAKPOINT_X)
 450                 r += scnprintf(bf + r, size - r, "x");
 451
 452         return r;
 453 }
 454
 455 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 456 {
 457         struct perf_event_attr *attr = &evsel->attr;
 458         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 459         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 460 }
 461
 462 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 463                                 [PERF_EVSEL__MAX_ALIASES] = {
 464  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 465  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 466  { "LLC",       "L2",                                                   },
 467  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 468  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 469  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 470  { "node",                                                              },
 471 };
 472
 473 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 474                                    [PERF_EVSEL__MAX_ALIASES] = {
 475  { "load",      "loads",        "read",                                 },
 476  { "store",     "stores",       "write",                                },
 477  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 478 };
 479
 480 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 481                                        [PERF_EVSEL__MAX_ALIASES] = {
 482  { "refs",      "Reference",    "ops",          "access",               },
 483  { "misses",    "miss",                                                 },
 484 };
 485
 486 #define C(x)            PERF_COUNT_HW_CACHE_##x
 487 #define CACHE_READ      (1 << C(OP_READ))
 488 #define CACHE_WRITE     (1 << C(OP_WRITE))
 489 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 490 #define COP(x)          (1 << x)
 491
 492 /*
 493  * cache operartion stat
 494  * L1I : Read and prefetch only
 495  * ITLB and BPU : Read-only
 496  */
 497 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 498  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 499  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 500  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 501  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 502  [C(ITLB)]      = (CACHE_READ),
 503  [C(BPU)]       = (CACHE_READ),
 504  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 505 };
 506
 507 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 508 {
 509         if (perf_evsel__hw_cache_stat[type] & COP(op))
 510                 return true;    /* valid */
 511         else
 512                 return false;   /* invalid */
 513 }
 514
 515 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 516                                             char *bf, size_t size)
 517 {
 518         if (result) {
 519                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 520                                  perf_evsel__hw_cache_op[op][0],
 521                                  perf_evsel__hw_cache_result[result][0]);
 522         }
 523
 524         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 525                          perf_evsel__hw_cache_op[op][1]);
 526 }
 527
 528 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 529 {
 530         u8 op, result, type = (config >>  0) & 0xff;
 531         const char *err = "unknown-ext-hardware-cache-type";
 532
 533         if (type >= PERF_COUNT_HW_CACHE_MAX)
 534                 goto out_err;
 535
 536         op = (config >>  8) & 0xff;
 537         err = "unknown-ext-hardware-cache-op";
 538         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 539                 goto out_err;
 540
 541         result = (config >> 16) & 0xff;
 542         err = "unknown-ext-hardware-cache-result";
 543         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 544                 goto out_err;
 545
 546         err = "invalid-cache";
 547         if (!perf_evsel__is_cache_op_valid(type, op))
 548                 goto out_err;
 549
 550         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 551 out_err:
 552         return scnprintf(bf, size, "%s", err);
 553 }
 554
 555 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 556 {
 557         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 558         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 559 }
 560
 561 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 562 {
 563         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 564         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 565 }
 566
 567 const char *perf_evsel__name(struct perf_evsel *evsel)
 568 {
 569         char bf[128];
 570
 571         if (evsel->name)
 572                 return evsel->name;
 573
 574         switch (evsel->attr.type) {
 575         case PERF_TYPE_RAW:
 576                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 577                 break;
 578
 579         case PERF_TYPE_HARDWARE:
 580                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 581                 break;
 582
 583         case PERF_TYPE_HW_CACHE:
 584                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 585                 break;
 586
 587         case PERF_TYPE_SOFTWARE:
 588                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 589                 break;
 590
 591         case PERF_TYPE_TRACEPOINT:
 592                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 593                 break;
 594
 595         case PERF_TYPE_BREAKPOINT:
 596                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 597                 break;
 598
 599         default:
 600                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 601                           evsel->attr.type);
 602                 break;
 603         }
 604
 605         evsel->name = strdup(bf);
 606
 607         return evsel->name ?: "unknown";
 608 }
 609
 610 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 611 {
 612         return evsel->group_name ?: "anon group";
 613 }
 614
 615 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 616 {
 617         int ret;
 618         struct perf_evsel *pos;
 619         const char *group_name = perf_evsel__group_name(evsel);
 620
 621         ret = scnprintf(buf, size, "%s", group_name);
 622
 623         ret += scnprintf(buf + ret, size - ret, " { %s",
 624                          perf_evsel__name(evsel));
 625
 626         for_each_group_member(pos, evsel)
 627                 ret += scnprintf(buf + ret, size - ret, ", %s",
 628                                  perf_evsel__name(pos));
 629
 630         ret += scnprintf(buf + ret, size - ret, " }");
 631
 632         return ret;
 633 }
 634
 635 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 636                                   struct record_opts *opts,
 637                                   struct callchain_param *param)
 638 {
 639         bool function = perf_evsel__is_function_event(evsel);
 640         struct perf_event_attr *attr = &evsel->attr;
 641
 642         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 643
 644         attr->sample_max_stack = param->max_stack;
 645
 646         if (param->record_mode == CALLCHAIN_LBR) {
 647                 if (!opts->branch_stack) {
 648                         if (attr->exclude_user) {
 649                                 pr_warning("LBR callstack option is only available "
 650                                            "to get user callchain information. "
 651                                            "Falling back to framepointers.\n");
 652                         } else {
 653                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 654                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 655                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 656                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 657                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 658                         }
 659                 } else
 660                          pr_warning("Cannot use LBR callstack with branch stack. "
 661                                     "Falling back to framepointers.\n");
 662         }
 663
 664         if (param->record_mode == CALLCHAIN_DWARF) {
 665                 if (!function) {
 666                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 667                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 668                         attr->sample_regs_user = PERF_REGS_MASK;
 669                         attr->sample_stack_user = param->dump_size;
 670                         attr->exclude_callchain_user = 1;
 671                 } else {
 672                         pr_info("Cannot use DWARF unwind for function trace event,"
 673                                 " falling back to framepointers.\n");
 674                 }
 675         }
 676
 677         if (function) {
 678                 pr_info("Disabling user space callchains for function trace event.\n");
 679                 attr->exclude_callchain_user = 1;
 680         }
 681 }
 682
 683 static void
 684 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 685                             struct callchain_param *param)
 686 {
 687         struct perf_event_attr *attr = &evsel->attr;
 688
 689         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 690         if (param->record_mode == CALLCHAIN_LBR) {
 691                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 692                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 693                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 694         }
 695         if (param->record_mode == CALLCHAIN_DWARF) {
 696                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 697                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 698         }
 699 }
 700
 701 static void apply_config_terms(struct perf_evsel *evsel,
 702                                struct record_opts *opts)
 703 {
 704         struct perf_evsel_config_term *term;
 705         struct list_head *config_terms = &evsel->config_terms;
 706         struct perf_event_attr *attr = &evsel->attr;
 707         struct callchain_param param;
 708         u32 dump_size = 0;
 709         int max_stack = 0;
 710         const char *callgraph_buf = NULL;
 711
 712         /* callgraph default */
 713         param.record_mode = callchain_param.record_mode;
 714
 715         list_for_each_entry(term, config_terms, list) {
 716                 switch (term->type) {
 717                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 718                         attr->sample_period = term->val.period;
 719                         attr->freq = 0;
 720                         break;
 721                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 722                         attr->sample_freq = term->val.freq;
 723                         attr->freq = 1;
 724                         break;
 725                 case PERF_EVSEL__CONFIG_TERM_TIME:
 726                         if (term->val.time)
 727                                 perf_evsel__set_sample_bit(evsel, TIME);
 728                         else
 729                                 perf_evsel__reset_sample_bit(evsel, TIME);
 730                         break;
 731                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 732                         callgraph_buf = term->val.callgraph;
 733                         break;
 734                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 735                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 736                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 737                                 parse_branch_str(term->val.branch,
 738                                                  &attr->branch_sample_type);
 739                         } else
 740                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 741                         break;
 742                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 743                         dump_size = term->val.stack_user;
 744                         break;
 745                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 746                         max_stack = term->val.max_stack;
 747                         break;
 748                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 749                         /*
 750                          * attr->inherit should has already been set by
 751                          * perf_evsel__config. If user explicitly set
 752                          * inherit using config terms, override global
 753                          * opt->no_inherit setting.
 754                          */
 755                         attr->inherit = term->val.inherit ? 1 : 0;
 756                         break;
 757                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 758                         attr->write_backward = term->val.overwrite ? 1 : 0;
 759                         break;
 760                 default:
 761                         break;
 762                 }
 763         }
 764
 765         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 766         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 767                 if (max_stack) {
 768                         param.max_stack = max_stack;
 769                         if (callgraph_buf == NULL)
 770                                 callgraph_buf = "fp";
 771                 }
 772
 773                 /* parse callgraph parameters */
 774                 if (callgraph_buf != NULL) {
 775                         if (!strcmp(callgraph_buf, "no")) {
 776                                 param.enabled = false;
 777                                 param.record_mode = CALLCHAIN_NONE;
 778                         } else {
 779                                 param.enabled = true;
 780                                 if (parse_callchain_record(callgraph_buf, &param)) {
 781                                         pr_err("per-event callgraph setting for %s failed. "
 782                                                "Apply callgraph global setting for it\n",
 783                                                evsel->name);
 784                                         return;
 785                                 }
 786                         }
 787                 }
 788                 if (dump_size > 0) {
 789                         dump_size = round_up(dump_size, sizeof(u64));
 790                         param.dump_size = dump_size;
 791                 }
 792
 793                 /* If global callgraph set, clear it */
 794                 if (callchain_param.enabled)
 795                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 796
 797                 /* set perf-event callgraph */
 798                 if (param.enabled)
 799                         perf_evsel__config_callchain(evsel, opts, &param);
 800         }
 801 }
 802
 803 /*
 804  * The enable_on_exec/disabled value strategy:
 805  *
 806  *  1) For any type of traced program:
 807  *    - all independent events and group leaders are disabled
 808  *    - all group members are enabled
 809  *
 810  *     Group members are ruled by group leaders. They need to
 811  *     be enabled, because the group scheduling relies on that.
 812  *
 813  *  2) For traced programs executed by perf:
 814  *     - all independent events and group leaders have
 815  *       enable_on_exec set
 816  *     - we don't specifically enable or disable any event during
 817  *       the record command
 818  *
 819  *     Independent events and group leaders are initially disabled
 820  *     and get enabled by exec. Group members are ruled by group
 821  *     leaders as stated in 1).
 822  *
 823  *  3) For traced programs attached by perf (pid/tid):
 824  *     - we specifically enable or disable all events during
 825  *       the record command
 826  *
 827  *     When attaching events to already running traced we
 828  *     enable/disable events specifically, as there's no
 829  *     initial traced exec call.
 830  */
 831 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 832                         struct callchain_param *callchain)
 833 {
 834         struct perf_evsel *leader = evsel->leader;
 835         struct perf_event_attr *attr = &evsel->attr;
 836         int track = evsel->tracking;
 837         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 838
 839         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 840         attr->inherit       = !opts->no_inherit;
 841         attr->write_backward = opts->overwrite ? 1 : 0;
 842
 843         perf_evsel__set_sample_bit(evsel, IP);
 844         perf_evsel__set_sample_bit(evsel, TID);
 845
 846         if (evsel->sample_read) {
 847                 perf_evsel__set_sample_bit(evsel, READ);
 848
 849                 /*
 850                  * We need ID even in case of single event, because
 851                  * PERF_SAMPLE_READ process ID specific data.
 852                  */
 853                 perf_evsel__set_sample_id(evsel, false);
 854
 855                 /*
 856                  * Apply group format only if we belong to group
 857                  * with more than one members.
 858                  */
 859                 if (leader->nr_members > 1) {
 860                         attr->read_format |= PERF_FORMAT_GROUP;
 861                         attr->inherit = 0;
 862                 }
 863         }
 864
 865         /*
 866          * We default some events to have a default interval. But keep
 867          * it a weak assumption overridable by the user.
 868          */
 869         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 870                                      opts->user_interval != ULLONG_MAX)) {
 871                 if (opts->freq) {
 872                         perf_evsel__set_sample_bit(evsel, PERIOD);
 873                         attr->freq              = 1;
 874                         attr->sample_freq       = opts->freq;
 875                 } else {
 876                         attr->sample_period = opts->default_interval;
 877                 }
 878         }
 879
 880         /*
 881          * Disable sampling for all group members other
 882          * than leader in case leader 'leads' the sampling.
 883          */
 884         if ((leader != evsel) && leader->sample_read) {
 885                 attr->sample_freq   = 0;
 886                 attr->sample_period = 0;
 887         }
 888
 889         if (opts->no_samples)
 890                 attr->sample_freq = 0;
 891
 892         if (opts->inherit_stat)
 893                 attr->inherit_stat = 1;
 894
 895         if (opts->sample_address) {
 896                 perf_evsel__set_sample_bit(evsel, ADDR);
 897                 attr->mmap_data = track;
 898         }
 899
 900         /*
 901          * We don't allow user space callchains for  function trace
 902          * event, due to issues with page faults while tracing page
 903          * fault handler and its overall trickiness nature.
 904          */
 905         if (perf_evsel__is_function_event(evsel))
 906                 evsel->attr.exclude_callchain_user = 1;
 907
 908         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 909                 perf_evsel__config_callchain(evsel, opts, callchain);
 910
 911         if (opts->sample_intr_regs) {
 912                 attr->sample_regs_intr = opts->sample_intr_regs;
 913                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 914         }
 915
 916         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 917                 perf_evsel__set_sample_bit(evsel, CPU);
 918
 919         if (opts->period)
 920                 perf_evsel__set_sample_bit(evsel, PERIOD);
 921
 922         /*
 923          * When the user explicitly disabled time don't force it here.
 924          */
 925         if (opts->sample_time &&
 926             (!perf_missing_features.sample_id_all &&
 927             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 928              opts->sample_time_set)))
 929                 perf_evsel__set_sample_bit(evsel, TIME);
 930
 931         if (opts->raw_samples && !evsel->no_aux_samples) {
 932                 perf_evsel__set_sample_bit(evsel, TIME);
 933                 perf_evsel__set_sample_bit(evsel, RAW);
 934                 perf_evsel__set_sample_bit(evsel, CPU);
 935         }
 936
 937         if (opts->sample_address)
 938                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 939
 940         if (opts->no_buffering) {
 941                 attr->watermark = 0;
 942                 attr->wakeup_events = 1;
 943         }
 944         if (opts->branch_stack && !evsel->no_aux_samples) {
 945                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 946                 attr->branch_sample_type = opts->branch_stack;
 947         }
 948
 949         if (opts->sample_weight)
 950                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 951
 952         attr->task  = track;
 953         attr->mmap  = track;
 954         attr->mmap2 = track && !perf_missing_features.mmap2;
 955         attr->comm  = track;
 956
 957         if (opts->record_namespaces)
 958                 attr->namespaces  = track;
 959
 960         if (opts->record_switch_events)
 961                 attr->context_switch = track;
 962
 963         if (opts->sample_transaction)
 964                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 965
 966         if (opts->running_time) {
 967                 evsel->attr.read_format |=
 968                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 969                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 970         }
 971
 972         /*
 973          * XXX see the function comment above
 974          *
 975          * Disabling only independent events or group leaders,
 976          * keeping group members enabled.
 977          */
 978         if (perf_evsel__is_group_leader(evsel))
 979                 attr->disabled = 1;
 980
 981         /*
 982          * Setting enable_on_exec for independent events and
 983          * group leaders for traced executed by perf.
 984          */
 985         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 986                 !opts->initial_delay)
 987                 attr->enable_on_exec = 1;
 988
 989         if (evsel->immediate) {
 990                 attr->disabled = 0;
 991                 attr->enable_on_exec = 0;
 992         }
 993
 994         clockid = opts->clockid;
 995         if (opts->use_clockid) {
 996                 attr->use_clockid = 1;
 997                 attr->clockid = opts->clockid;
 998         }
 999
1000         if (evsel->precise_max)
1001                 perf_event_attr__set_max_precise_ip(attr);
1002
1003         if (opts->all_user) {
1004                 attr->exclude_kernel = 1;
1005                 attr->exclude_user   = 0;
1006         }
1007
1008         if (opts->all_kernel) {
1009                 attr->exclude_kernel = 0;
1010                 attr->exclude_user   = 1;
1011         }
1012
1013         /*
1014          * Apply event specific term settings,
1015          * it overloads any global configuration.
1016          */
1017         apply_config_terms(evsel, opts);
1018
1019         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1020 }
1021
1022 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1023 {
1024         if (evsel->system_wide)
1025                 nthreads = 1;
1026
1027         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1028
1029         if (evsel->fd) {
1030                 int cpu, thread;
1031                 for (cpu = 0; cpu < ncpus; cpu++) {
1032                         for (thread = 0; thread < nthreads; thread++) {
1033                                 FD(evsel, cpu, thread) = -1;
1034                         }
1035                 }
1036         }
1037
1038         return evsel->fd != NULL ? 0 : -ENOMEM;
1039 }
1040
1041 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1042                           int ioc,  void *arg)
1043 {
1044         int cpu, thread;
1045
1046         if (evsel->system_wide)
1047                 nthreads = 1;
1048
1049         for (cpu = 0; cpu < ncpus; cpu++) {
1050                 for (thread = 0; thread < nthreads; thread++) {
1051                         int fd = FD(evsel, cpu, thread),
1052                             err = ioctl(fd, ioc, arg);
1053
1054                         if (err)
1055                                 return err;
1056                 }
1057         }
1058
1059         return 0;
1060 }
1061
1062 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1063                              const char *filter)
1064 {
1065         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1066                                      PERF_EVENT_IOC_SET_FILTER,
1067                                      (void *)filter);
1068 }
1069
1070 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1071 {
1072         char *new_filter = strdup(filter);
1073
1074         if (new_filter != NULL) {
1075                 free(evsel->filter);
1076                 evsel->filter = new_filter;
1077                 return 0;
1078         }
1079
1080         return -1;
1081 }
1082
1083 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1084                                      const char *fmt, const char *filter)
1085 {
1086         char *new_filter;
1087
1088         if (evsel->filter == NULL)
1089                 return perf_evsel__set_filter(evsel, filter);
1090
1091         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1092                 free(evsel->filter);
1093                 evsel->filter = new_filter;
1094                 return 0;
1095         }
1096
1097         return -1;
1098 }
1099
1100 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1101 {
1102         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1103 }
1104
1105 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1106 {
1107         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1108 }
1109
1110 int perf_evsel__enable(struct perf_evsel *evsel)
1111 {
1112         int nthreads = thread_map__nr(evsel->threads);
1113         int ncpus = cpu_map__nr(evsel->cpus);
1114
1115         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1116                                      PERF_EVENT_IOC_ENABLE,
1117                                      0);
1118 }
1119
1120 int perf_evsel__disable(struct perf_evsel *evsel)
1121 {
1122         int nthreads = thread_map__nr(evsel->threads);
1123         int ncpus = cpu_map__nr(evsel->cpus);
1124
1125         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1126                                      PERF_EVENT_IOC_DISABLE,
1127                                      0);
1128 }
1129
1130 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1131 {
1132         if (ncpus == 0 || nthreads == 0)
1133                 return 0;
1134
1135         if (evsel->system_wide)
1136                 nthreads = 1;
1137
1138         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1139         if (evsel->sample_id == NULL)
1140                 return -ENOMEM;
1141
1142         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1143         if (evsel->id == NULL) {
1144                 xyarray__delete(evsel->sample_id);
1145                 evsel->sample_id = NULL;
1146                 return -ENOMEM;
1147         }
1148
1149         return 0;
1150 }
1151
1152 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1153 {
1154         xyarray__delete(evsel->fd);
1155         evsel->fd = NULL;
1156 }
1157
1158 static void perf_evsel__free_id(struct perf_evsel *evsel)
1159 {
1160         xyarray__delete(evsel->sample_id);
1161         evsel->sample_id = NULL;
1162         zfree(&evsel->id);
1163 }
1164
1165 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1166 {
1167         struct perf_evsel_config_term *term, *h;
1168
1169         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1170                 list_del(&term->list);
1171                 free(term);
1172         }
1173 }
1174
1175 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1176 {
1177         int cpu, thread;
1178
1179         if (evsel->system_wide)
1180                 nthreads = 1;
1181
1182         for (cpu = 0; cpu < ncpus; cpu++)
1183                 for (thread = 0; thread < nthreads; ++thread) {
1184                         close(FD(evsel, cpu, thread));
1185                         FD(evsel, cpu, thread) = -1;
1186                 }
1187 }
1188
1189 void perf_evsel__exit(struct perf_evsel *evsel)
1190 {
1191         assert(list_empty(&evsel->node));
1192         assert(evsel->evlist == NULL);
1193         perf_evsel__free_fd(evsel);
1194         perf_evsel__free_id(evsel);
1195         perf_evsel__free_config_terms(evsel);
1196         close_cgroup(evsel->cgrp);
1197         cpu_map__put(evsel->cpus);
1198         cpu_map__put(evsel->own_cpus);
1199         thread_map__put(evsel->threads);
1200         zfree(&evsel->group_name);
1201         zfree(&evsel->name);
1202         perf_evsel__object.fini(evsel);
1203 }
1204
1205 void perf_evsel__delete(struct perf_evsel *evsel)
1206 {
1207         perf_evsel__exit(evsel);
1208         free(evsel);
1209 }
1210
1211 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1212                                 struct perf_counts_values *count)
1213 {
1214         struct perf_counts_values tmp;
1215
1216         if (!evsel->prev_raw_counts)
1217                 return;
1218
1219         if (cpu == -1) {
1220                 tmp = evsel->prev_raw_counts->aggr;
1221                 evsel->prev_raw_counts->aggr = *count;
1222         } else {
1223                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1224                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1225         }
1226
1227         count->val = count->val - tmp.val;
1228         count->ena = count->ena - tmp.ena;
1229         count->run = count->run - tmp.run;
1230 }
1231
1232 void perf_counts_values__scale(struct perf_counts_values *count,
1233                                bool scale, s8 *pscaled)
1234 {
1235         s8 scaled = 0;
1236
1237         if (scale) {
1238                 if (count->run == 0) {
1239                         scaled = -1;
1240                         count->val = 0;
1241                 } else if (count->run < count->ena) {
1242                         scaled = 1;
1243                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1244                 }
1245         } else
1246                 count->ena = count->run = 0;
1247
1248         if (pscaled)
1249                 *pscaled = scaled;
1250 }
1251
1252 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1253                      struct perf_counts_values *count)
1254 {
1255         memset(count, 0, sizeof(*count));
1256
1257         if (FD(evsel, cpu, thread) < 0)
1258                 return -EINVAL;
1259
1260         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) <= 0)
1261                 return -errno;
1262
1263         return 0;
1264 }
1265
1266 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1267                               int cpu, int thread, bool scale)
1268 {
1269         struct perf_counts_values count;
1270         size_t nv = scale ? 3 : 1;
1271
1272         if (FD(evsel, cpu, thread) < 0)
1273                 return -EINVAL;
1274
1275         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1276                 return -ENOMEM;
1277
1278         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1279                 return -errno;
1280
1281         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1282         perf_counts_values__scale(&count, scale, NULL);
1283         *perf_counts(evsel->counts, cpu, thread) = count;
1284         return 0;
1285 }
1286
1287 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1288 {
1289         struct perf_evsel *leader = evsel->leader;
1290         int fd;
1291
1292         if (perf_evsel__is_group_leader(evsel))
1293                 return -1;
1294
1295         /*
1296          * Leader must be already processed/open,
1297          * if not it's a bug.
1298          */
1299         BUG_ON(!leader->fd);
1300
1301         fd = FD(leader, cpu, thread);
1302         BUG_ON(fd == -1);
1303
1304         return fd;
1305 }
1306
1307 struct bit_names {
1308         int bit;
1309         const char *name;
1310 };
1311
1312 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1313 {
1314         bool first_bit = true;
1315         int i = 0;
1316
1317         do {
1318                 if (value & bits[i].bit) {
1319                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1320                         first_bit = false;
1321                 }
1322         } while (bits[++i].name != NULL);
1323 }
1324
1325 static void __p_sample_type(char *buf, size_t size, u64 value)
1326 {
1327 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1328         struct bit_names bits[] = {
1329                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1330                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1331                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1332                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1333                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1334                 bit_name(WEIGHT),
1335                 { .name = NULL, }
1336         };
1337 #undef bit_name
1338         __p_bits(buf, size, value, bits);
1339 }
1340
1341 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1342 {
1343 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1344         struct bit_names bits[] = {
1345                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1346                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1347                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1348                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1349                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1350                 { .name = NULL, }
1351         };
1352 #undef bit_name
1353         __p_bits(buf, size, value, bits);
1354 }
1355
1356 static void __p_read_format(char *buf, size_t size, u64 value)
1357 {
1358 #define bit_name(n) { PERF_FORMAT_##n, #n }
1359         struct bit_names bits[] = {
1360                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1361                 bit_name(ID), bit_name(GROUP),
1362                 { .name = NULL, }
1363         };
1364 #undef bit_name
1365         __p_bits(buf, size, value, bits);
1366 }
1367
1368 #define BUF_SIZE                1024
1369
1370 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1371 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1372 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1373 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1374 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1375 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1376
1377 #define PRINT_ATTRn(_n, _f, _p)                         \
1378 do {                                                    \
1379         if (attr->_f) {                                 \
1380                 _p(attr->_f);                           \
1381                 ret += attr__fprintf(fp, _n, buf, priv);\
1382         }                                               \
1383 } while (0)
1384
1385 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1386
1387 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1388                              attr__fprintf_f attr__fprintf, void *priv)
1389 {
1390         char buf[BUF_SIZE];
1391         int ret = 0;
1392
1393         PRINT_ATTRf(type, p_unsigned);
1394         PRINT_ATTRf(size, p_unsigned);
1395         PRINT_ATTRf(config, p_hex);
1396         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1397         PRINT_ATTRf(sample_type, p_sample_type);
1398         PRINT_ATTRf(read_format, p_read_format);
1399
1400         PRINT_ATTRf(disabled, p_unsigned);
1401         PRINT_ATTRf(inherit, p_unsigned);
1402         PRINT_ATTRf(pinned, p_unsigned);
1403         PRINT_ATTRf(exclusive, p_unsigned);
1404         PRINT_ATTRf(exclude_user, p_unsigned);
1405         PRINT_ATTRf(exclude_kernel, p_unsigned);
1406         PRINT_ATTRf(exclude_hv, p_unsigned);
1407         PRINT_ATTRf(exclude_idle, p_unsigned);
1408         PRINT_ATTRf(mmap, p_unsigned);
1409         PRINT_ATTRf(comm, p_unsigned);
1410         PRINT_ATTRf(freq, p_unsigned);
1411         PRINT_ATTRf(inherit_stat, p_unsigned);
1412         PRINT_ATTRf(enable_on_exec, p_unsigned);
1413         PRINT_ATTRf(task, p_unsigned);
1414         PRINT_ATTRf(watermark, p_unsigned);
1415         PRINT_ATTRf(precise_ip, p_unsigned);
1416         PRINT_ATTRf(mmap_data, p_unsigned);
1417         PRINT_ATTRf(sample_id_all, p_unsigned);
1418         PRINT_ATTRf(exclude_host, p_unsigned);
1419         PRINT_ATTRf(exclude_guest, p_unsigned);
1420         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1421         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1422         PRINT_ATTRf(mmap2, p_unsigned);
1423         PRINT_ATTRf(comm_exec, p_unsigned);
1424         PRINT_ATTRf(use_clockid, p_unsigned);
1425         PRINT_ATTRf(context_switch, p_unsigned);
1426         PRINT_ATTRf(write_backward, p_unsigned);
1427
1428         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1429         PRINT_ATTRf(bp_type, p_unsigned);
1430         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1431         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1432         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1433         PRINT_ATTRf(sample_regs_user, p_hex);
1434         PRINT_ATTRf(sample_stack_user, p_unsigned);
1435         PRINT_ATTRf(clockid, p_signed);
1436         PRINT_ATTRf(sample_regs_intr, p_hex);
1437         PRINT_ATTRf(aux_watermark, p_unsigned);
1438         PRINT_ATTRf(sample_max_stack, p_unsigned);
1439
1440         return ret;
1441 }
1442
1443 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1444                                 void *priv __attribute__((unused)))
1445 {
1446         return fprintf(fp, "  %-32s %s\n", name, val);
1447 }
1448
1449 static bool ignore_missing_thread(struct perf_evsel *evsel,
1450                                   struct thread_map *threads,
1451                                   int thread, int err)
1452 {
1453         if (!evsel->ignore_missing_thread)
1454                 return false;
1455
1456         /* The system wide setup does not work with threads. */
1457         if (evsel->system_wide)
1458                 return false;
1459
1460         /* The -ESRCH is perf event syscall errno for pid's not found. */
1461         if (err != -ESRCH)
1462                 return false;
1463
1464         /* If there's only one thread, let it fail. */
1465         if (threads->nr == 1)
1466                 return false;
1467
1468         if (thread_map__remove(threads, thread))
1469                 return false;
1470
1471         pr_warning("WARNING: Ignored open failure for pid %d\n",
1472                    thread_map__pid(threads, thread));
1473         return true;
1474 }
1475
1476 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1477                      struct thread_map *threads)
1478 {
1479         int cpu, thread, nthreads;
1480         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1481         int pid = -1, err;
1482         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1483
1484         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1485                 return -EINVAL;
1486
1487         if (cpus == NULL) {
1488                 static struct cpu_map *empty_cpu_map;
1489
1490                 if (empty_cpu_map == NULL) {
1491                         empty_cpu_map = cpu_map__dummy_new();
1492                         if (empty_cpu_map == NULL)
1493                                 return -ENOMEM;
1494                 }
1495
1496                 cpus = empty_cpu_map;
1497         }
1498
1499         if (threads == NULL) {
1500                 static struct thread_map *empty_thread_map;
1501
1502                 if (empty_thread_map == NULL) {
1503                         empty_thread_map = thread_map__new_by_tid(-1);
1504                         if (empty_thread_map == NULL)
1505                                 return -ENOMEM;
1506                 }
1507
1508                 threads = empty_thread_map;
1509         }
1510
1511         if (evsel->system_wide)
1512                 nthreads = 1;
1513         else
1514                 nthreads = threads->nr;
1515
1516         if (evsel->fd == NULL &&
1517             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1518                 return -ENOMEM;
1519
1520         if (evsel->cgrp) {
1521                 flags |= PERF_FLAG_PID_CGROUP;
1522                 pid = evsel->cgrp->fd;
1523         }
1524
1525 fallback_missing_features:
1526         if (perf_missing_features.clockid_wrong)
1527                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1528         if (perf_missing_features.clockid) {
1529                 evsel->attr.use_clockid = 0;
1530                 evsel->attr.clockid = 0;
1531         }
1532         if (perf_missing_features.cloexec)
1533                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1534         if (perf_missing_features.mmap2)
1535                 evsel->attr.mmap2 = 0;
1536         if (perf_missing_features.exclude_guest)
1537                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1538         if (perf_missing_features.lbr_flags)
1539                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1540                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1541 retry_sample_id:
1542         if (perf_missing_features.sample_id_all)
1543                 evsel->attr.sample_id_all = 0;
1544
1545         if (verbose >= 2) {
1546                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1547                 fprintf(stderr, "perf_event_attr:\n");
1548                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1549                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1550         }
1551
1552         for (cpu = 0; cpu < cpus->nr; cpu++) {
1553
1554                 for (thread = 0; thread < nthreads; thread++) {
1555                         int fd, group_fd;
1556
1557                         if (!evsel->cgrp && !evsel->system_wide)
1558                                 pid = thread_map__pid(threads, thread);
1559
1560                         group_fd = get_group_fd(evsel, cpu, thread);
1561 retry_open:
1562                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1563                                   pid, cpus->map[cpu], group_fd, flags);
1564
1565                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1566                                                  group_fd, flags);
1567
1568                         FD(evsel, cpu, thread) = fd;
1569
1570                         if (fd < 0) {
1571                                 err = -errno;
1572
1573                                 if (ignore_missing_thread(evsel, threads, thread, err)) {
1574                                         /*
1575                                          * We just removed 1 thread, so take a step
1576                                          * back on thread index and lower the upper
1577                                          * nthreads limit.
1578                                          */
1579                                         nthreads--;
1580                                         thread--;
1581
1582                                         /* ... and pretend like nothing have happened. */
1583                                         err = 0;
1584                                         continue;
1585                                 }
1586
1587                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1588                                           err);
1589                                 goto try_fallback;
1590                         }
1591
1592                         pr_debug2(" = %d\n", fd);
1593
1594                         if (evsel->bpf_fd >= 0) {
1595                                 int evt_fd = fd;
1596                                 int bpf_fd = evsel->bpf_fd;
1597
1598                                 err = ioctl(evt_fd,
1599                                             PERF_EVENT_IOC_SET_BPF,
1600                                             bpf_fd);
1601                                 if (err && errno != EEXIST) {
1602                                         pr_err("failed to attach bpf fd %d: %s\n",
1603                                                bpf_fd, strerror(errno));
1604                                         err = -EINVAL;
1605                                         goto out_close;
1606                                 }
1607                         }
1608
1609                         set_rlimit = NO_CHANGE;
1610
1611                         /*
1612                          * If we succeeded but had to kill clockid, fail and
1613                          * have perf_evsel__open_strerror() print us a nice
1614                          * error.
1615                          */
1616                         if (perf_missing_features.clockid ||
1617                             perf_missing_features.clockid_wrong) {
1618                                 err = -EINVAL;
1619                                 goto out_close;
1620                         }
1621                 }
1622         }
1623
1624         return 0;
1625
1626 try_fallback:
1627         /*
1628          * perf stat needs between 5 and 22 fds per CPU. When we run out
1629          * of them try to increase the limits.
1630          */
1631         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1632                 struct rlimit l;
1633                 int old_errno = errno;
1634
1635                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1636                         if (set_rlimit == NO_CHANGE)
1637                                 l.rlim_cur = l.rlim_max;
1638                         else {
1639                                 l.rlim_cur = l.rlim_max + 1000;
1640                                 l.rlim_max = l.rlim_cur;
1641                         }
1642                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1643                                 set_rlimit++;
1644                                 errno = old_errno;
1645                                 goto retry_open;
1646                         }
1647                 }
1648                 errno = old_errno;
1649         }
1650
1651         if (err != -EINVAL || cpu > 0 || thread > 0)
1652                 goto out_close;
1653
1654         /*
1655          * Must probe features in the order they were added to the
1656          * perf_event_attr interface.
1657          */
1658         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1659                 perf_missing_features.write_backward = true;
1660                 goto out_close;
1661         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1662                 perf_missing_features.clockid_wrong = true;
1663                 goto fallback_missing_features;
1664         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1665                 perf_missing_features.clockid = true;
1666                 goto fallback_missing_features;
1667         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1668                 perf_missing_features.cloexec = true;
1669                 goto fallback_missing_features;
1670         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1671                 perf_missing_features.mmap2 = true;
1672                 goto fallback_missing_features;
1673         } else if (!perf_missing_features.exclude_guest &&
1674                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1675                 perf_missing_features.exclude_guest = true;
1676                 goto fallback_missing_features;
1677         } else if (!perf_missing_features.sample_id_all) {
1678                 perf_missing_features.sample_id_all = true;
1679                 goto retry_sample_id;
1680         } else if (!perf_missing_features.lbr_flags &&
1681                         (evsel->attr.branch_sample_type &
1682                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1683                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1684                 perf_missing_features.lbr_flags = true;
1685                 goto fallback_missing_features;
1686         }
1687 out_close:
1688         do {
1689                 while (--thread >= 0) {
1690                         close(FD(evsel, cpu, thread));
1691                         FD(evsel, cpu, thread) = -1;
1692                 }
1693                 thread = nthreads;
1694         } while (--cpu >= 0);
1695         return err;
1696 }
1697
1698 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1699 {
1700         if (evsel->fd == NULL)
1701                 return;
1702
1703         perf_evsel__close_fd(evsel, ncpus, nthreads);
1704         perf_evsel__free_fd(evsel);
1705 }
1706
1707 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1708                              struct cpu_map *cpus)
1709 {
1710         return perf_evsel__open(evsel, cpus, NULL);
1711 }
1712
1713 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1714                                 struct thread_map *threads)
1715 {
1716         return perf_evsel__open(evsel, NULL, threads);
1717 }
1718
1719 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1720                                        const union perf_event *event,
1721                                        struct perf_sample *sample)
1722 {
1723         u64 type = evsel->attr.sample_type;
1724         const u64 *array = event->sample.array;
1725         bool swapped = evsel->needs_swap;
1726         union u64_swap u;
1727
1728         array += ((event->header.size -
1729                    sizeof(event->header)) / sizeof(u64)) - 1;
1730
1731         if (type & PERF_SAMPLE_IDENTIFIER) {
1732                 sample->id = *array;
1733                 array--;
1734         }
1735
1736         if (type & PERF_SAMPLE_CPU) {
1737                 u.val64 = *array;
1738                 if (swapped) {
1739                         /* undo swap of u64, then swap on individual u32s */
1740                         u.val64 = bswap_64(u.val64);
1741                         u.val32[0] = bswap_32(u.val32[0]);
1742                 }
1743
1744                 sample->cpu = u.val32[0];
1745                 array--;
1746         }
1747
1748         if (type & PERF_SAMPLE_STREAM_ID) {
1749                 sample->stream_id = *array;
1750                 array--;
1751         }
1752
1753         if (type & PERF_SAMPLE_ID) {
1754                 sample->id = *array;
1755                 array--;
1756         }
1757
1758         if (type & PERF_SAMPLE_TIME) {
1759                 sample->time = *array;
1760                 array--;
1761         }
1762
1763         if (type & PERF_SAMPLE_TID) {
1764                 u.val64 = *array;
1765                 if (swapped) {
1766                         /* undo swap of u64, then swap on individual u32s */
1767                         u.val64 = bswap_64(u.val64);
1768                         u.val32[0] = bswap_32(u.val32[0]);
1769                         u.val32[1] = bswap_32(u.val32[1]);
1770                 }
1771
1772                 sample->pid = u.val32[0];
1773                 sample->tid = u.val32[1];
1774                 array--;
1775         }
1776
1777         return 0;
1778 }
1779
1780 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1781                             u64 size)
1782 {
1783         return size > max_size || offset + size > endp;
1784 }
1785
1786 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1787         do {                                                            \
1788                 if (overflow(endp, (max_size), (offset), (size)))       \
1789                         return -EFAULT;                                 \
1790         } while (0)
1791
1792 #define OVERFLOW_CHECK_u64(offset) \
1793         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1794
1795 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1796                              struct perf_sample *data)
1797 {
1798         u64 type = evsel->attr.sample_type;
1799         bool swapped = evsel->needs_swap;
1800         const u64 *array;
1801         u16 max_size = event->header.size;
1802         const void *endp = (void *)event + max_size;
1803         u64 sz;
1804
1805         /*
1806          * used for cross-endian analysis. See git commit 65014ab3
1807          * for why this goofiness is needed.
1808          */
1809         union u64_swap u;
1810
1811         memset(data, 0, sizeof(*data));
1812         data->cpu = data->pid = data->tid = -1;
1813         data->stream_id = data->id = data->time = -1ULL;
1814         data->period = evsel->attr.sample_period;
1815         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1816
1817         if (event->header.type != PERF_RECORD_SAMPLE) {
1818                 if (!evsel->attr.sample_id_all)
1819                         return 0;
1820                 return perf_evsel__parse_id_sample(evsel, event, data);
1821         }
1822
1823         array = event->sample.array;
1824
1825         /*
1826          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1827          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1828          * check the format does not go past the end of the event.
1829          */
1830         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1831                 return -EFAULT;
1832
1833         data->id = -1ULL;
1834         if (type & PERF_SAMPLE_IDENTIFIER) {
1835                 data->id = *array;
1836                 array++;
1837         }
1838
1839         if (type & PERF_SAMPLE_IP) {
1840                 data->ip = *array;
1841                 array++;
1842         }
1843
1844         if (type & PERF_SAMPLE_TID) {
1845                 u.val64 = *array;
1846                 if (swapped) {
1847                         /* undo swap of u64, then swap on individual u32s */
1848                         u.val64 = bswap_64(u.val64);
1849                         u.val32[0] = bswap_32(u.val32[0]);
1850                         u.val32[1] = bswap_32(u.val32[1]);
1851                 }
1852
1853                 data->pid = u.val32[0];
1854                 data->tid = u.val32[1];
1855                 array++;
1856         }
1857
1858         if (type & PERF_SAMPLE_TIME) {
1859                 data->time = *array;
1860                 array++;
1861         }
1862
1863         data->addr = 0;
1864         if (type & PERF_SAMPLE_ADDR) {
1865                 data->addr = *array;
1866                 array++;
1867         }
1868
1869         if (type & PERF_SAMPLE_ID) {
1870                 data->id = *array;
1871                 array++;
1872         }
1873
1874         if (type & PERF_SAMPLE_STREAM_ID) {
1875                 data->stream_id = *array;
1876                 array++;
1877         }
1878
1879         if (type & PERF_SAMPLE_CPU) {
1880
1881                 u.val64 = *array;
1882                 if (swapped) {
1883                         /* undo swap of u64, then swap on individual u32s */
1884                         u.val64 = bswap_64(u.val64);
1885                         u.val32[0] = bswap_32(u.val32[0]);
1886                 }
1887
1888                 data->cpu = u.val32[0];
1889                 array++;
1890         }
1891
1892         if (type & PERF_SAMPLE_PERIOD) {
1893                 data->period = *array;
1894                 array++;
1895         }
1896
1897         if (type & PERF_SAMPLE_READ) {
1898                 u64 read_format = evsel->attr.read_format;
1899
1900                 OVERFLOW_CHECK_u64(array);
1901                 if (read_format & PERF_FORMAT_GROUP)
1902                         data->read.group.nr = *array;
1903                 else
1904                         data->read.one.value = *array;
1905
1906                 array++;
1907
1908                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1909                         OVERFLOW_CHECK_u64(array);
1910                         data->read.time_enabled = *array;
1911                         array++;
1912                 }
1913
1914                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1915                         OVERFLOW_CHECK_u64(array);
1916                         data->read.time_running = *array;
1917                         array++;
1918                 }
1919
1920                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1921                 if (read_format & PERF_FORMAT_GROUP) {
1922                         const u64 max_group_nr = UINT64_MAX /
1923                                         sizeof(struct sample_read_value);
1924
1925                         if (data->read.group.nr > max_group_nr)
1926                                 return -EFAULT;
1927                         sz = data->read.group.nr *
1928                              sizeof(struct sample_read_value);
1929                         OVERFLOW_CHECK(array, sz, max_size);
1930                         data->read.group.values =
1931                                         (struct sample_read_value *)array;
1932                         array = (void *)array + sz;
1933                 } else {
1934                         OVERFLOW_CHECK_u64(array);
1935                         data->read.one.id = *array;
1936                         array++;
1937                 }
1938         }
1939
1940         if (type & PERF_SAMPLE_CALLCHAIN) {
1941                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1942
1943                 OVERFLOW_CHECK_u64(array);
1944                 data->callchain = (struct ip_callchain *)array++;
1945                 if (data->callchain->nr > max_callchain_nr)
1946                         return -EFAULT;
1947                 sz = data->callchain->nr * sizeof(u64);
1948                 OVERFLOW_CHECK(array, sz, max_size);
1949                 array = (void *)array + sz;
1950         }
1951
1952         if (type & PERF_SAMPLE_RAW) {
1953                 OVERFLOW_CHECK_u64(array);
1954                 u.val64 = *array;
1955                 if (WARN_ONCE(swapped,
1956                               "Endianness of raw data not corrected!\n")) {
1957                         /* undo swap of u64, then swap on individual u32s */
1958                         u.val64 = bswap_64(u.val64);
1959                         u.val32[0] = bswap_32(u.val32[0]);
1960                         u.val32[1] = bswap_32(u.val32[1]);
1961                 }
1962                 data->raw_size = u.val32[0];
1963                 array = (void *)array + sizeof(u32);
1964
1965                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1966                 data->raw_data = (void *)array;
1967                 array = (void *)array + data->raw_size;
1968         }
1969
1970         if (type & PERF_SAMPLE_BRANCH_STACK) {
1971                 const u64 max_branch_nr = UINT64_MAX /
1972                                           sizeof(struct branch_entry);
1973
1974                 OVERFLOW_CHECK_u64(array);
1975                 data->branch_stack = (struct branch_stack *)array++;
1976
1977                 if (data->branch_stack->nr > max_branch_nr)
1978                         return -EFAULT;
1979                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1980                 OVERFLOW_CHECK(array, sz, max_size);
1981                 array = (void *)array + sz;
1982         }
1983
1984         if (type & PERF_SAMPLE_REGS_USER) {
1985                 OVERFLOW_CHECK_u64(array);
1986                 data->user_regs.abi = *array;
1987                 array++;
1988
1989                 if (data->user_regs.abi) {
1990                         u64 mask = evsel->attr.sample_regs_user;
1991
1992                         sz = hweight_long(mask) * sizeof(u64);
1993                         OVERFLOW_CHECK(array, sz, max_size);
1994                         data->user_regs.mask = mask;
1995                         data->user_regs.regs = (u64 *)array;
1996                         array = (void *)array + sz;
1997                 }
1998         }
1999
2000         if (type & PERF_SAMPLE_STACK_USER) {
2001                 OVERFLOW_CHECK_u64(array);
2002                 sz = *array++;
2003
2004                 data->user_stack.offset = ((char *)(array - 1)
2005                                           - (char *) event);
2006
2007                 if (!sz) {
2008                         data->user_stack.size = 0;
2009                 } else {
2010                         OVERFLOW_CHECK(array, sz, max_size);
2011                         data->user_stack.data = (char *)array;
2012                         array = (void *)array + sz;
2013                         OVERFLOW_CHECK_u64(array);
2014                         data->user_stack.size = *array++;
2015                         if (WARN_ONCE(data->user_stack.size > sz,
2016                                       "user stack dump failure\n"))
2017                                 return -EFAULT;
2018                 }
2019         }
2020
2021         if (type & PERF_SAMPLE_WEIGHT) {
2022                 OVERFLOW_CHECK_u64(array);
2023                 data->weight = *array;
2024                 array++;
2025         }
2026
2027         data->data_src = PERF_MEM_DATA_SRC_NONE;
2028         if (type & PERF_SAMPLE_DATA_SRC) {
2029                 OVERFLOW_CHECK_u64(array);
2030                 data->data_src = *array;
2031                 array++;
2032         }
2033
2034         data->transaction = 0;
2035         if (type & PERF_SAMPLE_TRANSACTION) {
2036                 OVERFLOW_CHECK_u64(array);
2037                 data->transaction = *array;
2038                 array++;
2039         }
2040
2041         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2042         if (type & PERF_SAMPLE_REGS_INTR) {
2043                 OVERFLOW_CHECK_u64(array);
2044                 data->intr_regs.abi = *array;
2045                 array++;
2046
2047                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2048                         u64 mask = evsel->attr.sample_regs_intr;
2049
2050                         sz = hweight_long(mask) * sizeof(u64);
2051                         OVERFLOW_CHECK(array, sz, max_size);
2052                         data->intr_regs.mask = mask;
2053                         data->intr_regs.regs = (u64 *)array;
2054                         array = (void *)array + sz;
2055                 }
2056         }
2057
2058         return 0;
2059 }
2060
2061 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2062                                      u64 read_format)
2063 {
2064         size_t sz, result = sizeof(struct sample_event);
2065
2066         if (type & PERF_SAMPLE_IDENTIFIER)
2067                 result += sizeof(u64);
2068
2069         if (type & PERF_SAMPLE_IP)
2070                 result += sizeof(u64);
2071
2072         if (type & PERF_SAMPLE_TID)
2073                 result += sizeof(u64);
2074
2075         if (type & PERF_SAMPLE_TIME)
2076                 result += sizeof(u64);
2077
2078         if (type & PERF_SAMPLE_ADDR)
2079                 result += sizeof(u64);
2080
2081         if (type & PERF_SAMPLE_ID)
2082                 result += sizeof(u64);
2083
2084         if (type & PERF_SAMPLE_STREAM_ID)
2085                 result += sizeof(u64);
2086
2087         if (type & PERF_SAMPLE_CPU)
2088                 result += sizeof(u64);
2089
2090         if (type & PERF_SAMPLE_PERIOD)
2091                 result += sizeof(u64);
2092
2093         if (type & PERF_SAMPLE_READ) {
2094                 result += sizeof(u64);
2095                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2096                         result += sizeof(u64);
2097                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2098                         result += sizeof(u64);
2099                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2100                 if (read_format & PERF_FORMAT_GROUP) {
2101                         sz = sample->read.group.nr *
2102                              sizeof(struct sample_read_value);
2103                         result += sz;
2104                 } else {
2105                         result += sizeof(u64);
2106                 }
2107         }
2108
2109         if (type & PERF_SAMPLE_CALLCHAIN) {
2110                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2111                 result += sz;
2112         }
2113
2114         if (type & PERF_SAMPLE_RAW) {
2115                 result += sizeof(u32);
2116                 result += sample->raw_size;
2117         }
2118
2119         if (type & PERF_SAMPLE_BRANCH_STACK) {
2120                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2121                 sz += sizeof(u64);
2122                 result += sz;
2123         }
2124
2125         if (type & PERF_SAMPLE_REGS_USER) {
2126                 if (sample->user_regs.abi) {
2127                         result += sizeof(u64);
2128                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2129                         result += sz;
2130                 } else {
2131                         result += sizeof(u64);
2132                 }
2133         }
2134
2135         if (type & PERF_SAMPLE_STACK_USER) {
2136                 sz = sample->user_stack.size;
2137                 result += sizeof(u64);
2138                 if (sz) {
2139                         result += sz;
2140                         result += sizeof(u64);
2141                 }
2142         }
2143
2144         if (type & PERF_SAMPLE_WEIGHT)
2145                 result += sizeof(u64);
2146
2147         if (type & PERF_SAMPLE_DATA_SRC)
2148                 result += sizeof(u64);
2149
2150         if (type & PERF_SAMPLE_TRANSACTION)
2151                 result += sizeof(u64);
2152
2153         if (type & PERF_SAMPLE_REGS_INTR) {
2154                 if (sample->intr_regs.abi) {
2155                         result += sizeof(u64);
2156                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2157                         result += sz;
2158                 } else {
2159                         result += sizeof(u64);
2160                 }
2161         }
2162
2163         return result;
2164 }
2165
2166 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2167                                   u64 read_format,
2168                                   const struct perf_sample *sample,
2169                                   bool swapped)
2170 {
2171         u64 *array;
2172         size_t sz;
2173         /*
2174          * used for cross-endian analysis. See git commit 65014ab3
2175          * for why this goofiness is needed.
2176          */
2177         union u64_swap u;
2178
2179         array = event->sample.array;
2180
2181         if (type & PERF_SAMPLE_IDENTIFIER) {
2182                 *array = sample->id;
2183                 array++;
2184         }
2185
2186         if (type & PERF_SAMPLE_IP) {
2187                 *array = sample->ip;
2188                 array++;
2189         }
2190
2191         if (type & PERF_SAMPLE_TID) {
2192                 u.val32[0] = sample->pid;
2193                 u.val32[1] = sample->tid;
2194                 if (swapped) {
2195                         /*
2196                          * Inverse of what is done in perf_evsel__parse_sample
2197                          */
2198                         u.val32[0] = bswap_32(u.val32[0]);
2199                         u.val32[1] = bswap_32(u.val32[1]);
2200                         u.val64 = bswap_64(u.val64);
2201                 }
2202
2203                 *array = u.val64;
2204                 array++;
2205         }
2206
2207         if (type & PERF_SAMPLE_TIME) {
2208                 *array = sample->time;
2209                 array++;
2210         }
2211
2212         if (type & PERF_SAMPLE_ADDR) {
2213                 *array = sample->addr;
2214                 array++;
2215         }
2216
2217         if (type & PERF_SAMPLE_ID) {
2218                 *array = sample->id;
2219                 array++;
2220         }
2221
2222         if (type & PERF_SAMPLE_STREAM_ID) {
2223                 *array = sample->stream_id;
2224                 array++;
2225         }
2226
2227         if (type & PERF_SAMPLE_CPU) {
2228                 u.val32[0] = sample->cpu;
2229                 if (swapped) {
2230                         /*
2231                          * Inverse of what is done in perf_evsel__parse_sample
2232                          */
2233                         u.val32[0] = bswap_32(u.val32[0]);
2234                         u.val64 = bswap_64(u.val64);
2235                 }
2236                 *array = u.val64;
2237                 array++;
2238         }
2239
2240         if (type & PERF_SAMPLE_PERIOD) {
2241                 *array = sample->period;
2242                 array++;
2243         }
2244
2245         if (type & PERF_SAMPLE_READ) {
2246                 if (read_format & PERF_FORMAT_GROUP)
2247                         *array = sample->read.group.nr;
2248                 else
2249                         *array = sample->read.one.value;
2250                 array++;
2251
2252                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2253                         *array = sample->read.time_enabled;
2254                         array++;
2255                 }
2256
2257                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2258                         *array = sample->read.time_running;
2259                         array++;
2260                 }
2261
2262                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2263                 if (read_format & PERF_FORMAT_GROUP) {
2264                         sz = sample->read.group.nr *
2265                              sizeof(struct sample_read_value);
2266                         memcpy(array, sample->read.group.values, sz);
2267                         array = (void *)array + sz;
2268                 } else {
2269                         *array = sample->read.one.id;
2270                         array++;
2271                 }
2272         }
2273
2274         if (type & PERF_SAMPLE_CALLCHAIN) {
2275                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2276                 memcpy(array, sample->callchain, sz);
2277                 array = (void *)array + sz;
2278         }
2279
2280         if (type & PERF_SAMPLE_RAW) {
2281                 u.val32[0] = sample->raw_size;
2282                 if (WARN_ONCE(swapped,
2283                               "Endianness of raw data not corrected!\n")) {
2284                         /*
2285                          * Inverse of what is done in perf_evsel__parse_sample
2286                          */
2287                         u.val32[0] = bswap_32(u.val32[0]);
2288                         u.val32[1] = bswap_32(u.val32[1]);
2289                         u.val64 = bswap_64(u.val64);
2290                 }
2291                 *array = u.val64;
2292                 array = (void *)array + sizeof(u32);
2293
2294                 memcpy(array, sample->raw_data, sample->raw_size);
2295                 array = (void *)array + sample->raw_size;
2296         }
2297
2298         if (type & PERF_SAMPLE_BRANCH_STACK) {
2299                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2300                 sz += sizeof(u64);
2301                 memcpy(array, sample->branch_stack, sz);
2302                 array = (void *)array + sz;
2303         }
2304
2305         if (type & PERF_SAMPLE_REGS_USER) {
2306                 if (sample->user_regs.abi) {
2307                         *array++ = sample->user_regs.abi;
2308                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2309                         memcpy(array, sample->user_regs.regs, sz);
2310                         array = (void *)array + sz;
2311                 } else {
2312                         *array++ = 0;
2313                 }
2314         }
2315
2316         if (type & PERF_SAMPLE_STACK_USER) {
2317                 sz = sample->user_stack.size;
2318                 *array++ = sz;
2319                 if (sz) {
2320                         memcpy(array, sample->user_stack.data, sz);
2321                         array = (void *)array + sz;
2322                         *array++ = sz;
2323                 }
2324         }
2325
2326         if (type & PERF_SAMPLE_WEIGHT) {
2327                 *array = sample->weight;
2328                 array++;
2329         }
2330
2331         if (type & PERF_SAMPLE_DATA_SRC) {
2332                 *array = sample->data_src;
2333                 array++;
2334         }
2335
2336         if (type & PERF_SAMPLE_TRANSACTION) {
2337                 *array = sample->transaction;
2338                 array++;
2339         }
2340
2341         if (type & PERF_SAMPLE_REGS_INTR) {
2342                 if (sample->intr_regs.abi) {
2343                         *array++ = sample->intr_regs.abi;
2344                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2345                         memcpy(array, sample->intr_regs.regs, sz);
2346                         array = (void *)array + sz;
2347                 } else {
2348                         *array++ = 0;
2349                 }
2350         }
2351
2352         return 0;
2353 }
2354
2355 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2356 {
2357         return pevent_find_field(evsel->tp_format, name);
2358 }
2359
2360 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2361                          const char *name)
2362 {
2363         struct format_field *field = perf_evsel__field(evsel, name);
2364         int offset;
2365
2366         if (!field)
2367                 return NULL;
2368
2369         offset = field->offset;
2370
2371         if (field->flags & FIELD_IS_DYNAMIC) {
2372                 offset = *(int *)(sample->raw_data + field->offset);
2373                 offset &= 0xffff;
2374         }
2375
2376         return sample->raw_data + offset;
2377 }
2378
2379 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2380                          bool needs_swap)
2381 {
2382         u64 value;
2383         void *ptr = sample->raw_data + field->offset;
2384
2385         switch (field->size) {
2386         case 1:
2387                 return *(u8 *)ptr;
2388         case 2:
2389                 value = *(u16 *)ptr;
2390                 break;
2391         case 4:
2392                 value = *(u32 *)ptr;
2393                 break;
2394         case 8:
2395                 memcpy(&value, ptr, sizeof(u64));
2396                 break;
2397         default:
2398                 return 0;
2399         }
2400
2401         if (!needs_swap)
2402                 return value;
2403
2404         switch (field->size) {
2405         case 2:
2406                 return bswap_16(value);
2407         case 4:
2408                 return bswap_32(value);
2409         case 8:
2410                 return bswap_64(value);
2411         default:
2412                 return 0;
2413         }
2414
2415         return 0;
2416 }
2417
2418 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2419                        const char *name)
2420 {
2421         struct format_field *field = perf_evsel__field(evsel, name);
2422
2423         if (!field)
2424                 return 0;
2425
2426         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2427 }
2428
2429 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2430                           char *msg, size_t msgsize)
2431 {
2432         int paranoid;
2433
2434         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2435             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2436             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2437                 /*
2438                  * If it's cycles then fall back to hrtimer based
2439                  * cpu-clock-tick sw counter, which is always available even if
2440                  * no PMU support.
2441                  *
2442                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2443                  * b0a873e).
2444                  */
2445                 scnprintf(msg, msgsize, "%s",
2446 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2447
2448                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2449                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2450
2451                 zfree(&evsel->name);
2452                 return true;
2453         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2454                    (paranoid = perf_event_paranoid()) > 1) {
2455                 const char *name = perf_evsel__name(evsel);
2456                 char *new_name;
2457
2458                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2459                         return false;
2460
2461                 if (evsel->name)
2462                         free(evsel->name);
2463                 evsel->name = new_name;
2464                 scnprintf(msg, msgsize,
2465 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2466                 evsel->attr.exclude_kernel = 1;
2467
2468                 return true;
2469         }
2470
2471         return false;
2472 }
2473
2474 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2475                               int err, char *msg, size_t size)
2476 {
2477         char sbuf[STRERR_BUFSIZE];
2478         int printed = 0;
2479
2480         switch (err) {
2481         case EPERM:
2482         case EACCES:
2483                 if (err == EPERM)
2484                         printed = scnprintf(msg, size,
2485                                 "No permission to enable %s event.\n\n",
2486                                 perf_evsel__name(evsel));
2487
2488                 return scnprintf(msg + printed, size - printed,
2489                  "You may not have permission to collect %sstats.\n\n"
2490                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2491                  "which controls use of the performance events system by\n"
2492                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2493                  "The current value is %d:\n\n"
2494                  "  -1: Allow use of (almost) all events by all users\n"
2495                  ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2496                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2497                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2498                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2499                  "      kernel.perf_event_paranoid = -1\n" ,
2500                                  target->system_wide ? "system-wide " : "",
2501                                  perf_event_paranoid());
2502         case ENOENT:
2503                 return scnprintf(msg, size, "The %s event is not supported.",
2504                                  perf_evsel__name(evsel));
2505         case EMFILE:
2506                 return scnprintf(msg, size, "%s",
2507                          "Too many events are opened.\n"
2508                          "Probably the maximum number of open file descriptors has been reached.\n"
2509                          "Hint: Try again after reducing the number of events.\n"
2510                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2511         case ENOMEM:
2512                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2513                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2514                         return scnprintf(msg, size,
2515                                          "Not enough memory to setup event with callchain.\n"
2516                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2517                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2518                 break;
2519         case ENODEV:
2520                 if (target->cpu_list)
2521                         return scnprintf(msg, size, "%s",
2522          "No such device - did you specify an out-of-range profile CPU?");
2523                 break;
2524         case EOPNOTSUPP:
2525                 if (evsel->attr.sample_period != 0)
2526                         return scnprintf(msg, size, "%s",
2527         "PMU Hardware doesn't support sampling/overflow-interrupts.");
2528                 if (evsel->attr.precise_ip)
2529                         return scnprintf(msg, size, "%s",
2530         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2531 #if defined(__i386__) || defined(__x86_64__)
2532                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2533                         return scnprintf(msg, size, "%s",
2534         "No hardware sampling interrupt available.\n"
2535         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2536 #endif
2537                 break;
2538         case EBUSY:
2539                 if (find_process("oprofiled"))
2540                         return scnprintf(msg, size,
2541         "The PMU counters are busy/taken by another profiler.\n"
2542         "We found oprofile daemon running, please stop it and try again.");
2543                 break;
2544         case EINVAL:
2545                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2546                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2547                 if (perf_missing_features.clockid)
2548                         return scnprintf(msg, size, "clockid feature not supported.");
2549                 if (perf_missing_features.clockid_wrong)
2550                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2551                 break;
2552         default:
2553                 break;
2554         }
2555
2556         return scnprintf(msg, size,
2557         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2558         "/bin/dmesg may provide additional information.\n"
2559         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2560                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2561                          perf_evsel__name(evsel));
2562 }
2563
2564 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2565 {
2566         if (evsel && evsel->evlist && evsel->evlist->env)
2567                 return evsel->evlist->env->arch;
2568         return NULL;
2569 }