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 <linux/bitops.h>
  12 #include <api/fs/tracing_path.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <linux/err.h>
  17 #include <sys/resource.h>
  18 #include "asm/bug.h"
  19 #include "callchain.h"
  20 #include "cgroup.h"
  21 #include "evsel.h"
  22 #include "evlist.h"
  23 #include "util.h"
  24 #include "cpumap.h"
  25 #include "thread_map.h"
  26 #include "target.h"
  27 #include "perf_regs.h"
  28 #include "debug.h"
  29 #include "trace-event.h"
  30 #include "stat.h"
  31
  32 static struct {
  33         bool sample_id_all;
  34         bool exclude_guest;
  35         bool mmap2;
  36         bool cloexec;
  37         bool clockid;
  38         bool clockid_wrong;
  39 } perf_missing_features;
  40
  41 static clockid_t clockid;
  42
  43 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  44 {
  45         return 0;
  46 }
  47
  48 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  49 {
  50 }
  51
  52 static struct {
  53         size_t  size;
  54         int     (*init)(struct perf_evsel *evsel);
  55         void    (*fini)(struct perf_evsel *evsel);
  56 } perf_evsel__object = {
  57         .size = sizeof(struct perf_evsel),
  58         .init = perf_evsel__no_extra_init,
  59         .fini = perf_evsel__no_extra_fini,
  60 };
  61
  62 int perf_evsel__object_config(size_t object_size,
  63                               int (*init)(struct perf_evsel *evsel),
  64                               void (*fini)(struct perf_evsel *evsel))
  65 {
  66
  67         if (object_size == 0)
  68                 goto set_methods;
  69
  70         if (perf_evsel__object.size > object_size)
  71                 return -EINVAL;
  72
  73         perf_evsel__object.size = object_size;
  74
  75 set_methods:
  76         if (init != NULL)
  77                 perf_evsel__object.init = init;
  78
  79         if (fini != NULL)
  80                 perf_evsel__object.fini = fini;
  81
  82         return 0;
  83 }
  84
  85 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  86
  87 int __perf_evsel__sample_size(u64 sample_type)
  88 {
  89         u64 mask = sample_type & PERF_SAMPLE_MASK;
  90         int size = 0;
  91         int i;
  92
  93         for (i = 0; i < 64; i++) {
  94                 if (mask & (1ULL << i))
  95                         size++;
  96         }
  97
  98         size *= sizeof(u64);
  99
 100         return size;
 101 }
 102
 103 /**
 104  * __perf_evsel__calc_id_pos - calculate id_pos.
 105  * @sample_type: sample type
 106  *
 107  * This function returns the position of the event id (PERF_SAMPLE_ID or
 108  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 109  * sample_event.
 110  */
 111 static int __perf_evsel__calc_id_pos(u64 sample_type)
 112 {
 113         int idx = 0;
 114
 115         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 116                 return 0;
 117
 118         if (!(sample_type & PERF_SAMPLE_ID))
 119                 return -1;
 120
 121         if (sample_type & PERF_SAMPLE_IP)
 122                 idx += 1;
 123
 124         if (sample_type & PERF_SAMPLE_TID)
 125                 idx += 1;
 126
 127         if (sample_type & PERF_SAMPLE_TIME)
 128                 idx += 1;
 129
 130         if (sample_type & PERF_SAMPLE_ADDR)
 131                 idx += 1;
 132
 133         return idx;
 134 }
 135
 136 /**
 137  * __perf_evsel__calc_is_pos - calculate is_pos.
 138  * @sample_type: sample type
 139  *
 140  * This function returns the position (counting backwards) of the event id
 141  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 142  * sample_id_all is used there is an id sample appended to non-sample events.
 143  */
 144 static int __perf_evsel__calc_is_pos(u64 sample_type)
 145 {
 146         int idx = 1;
 147
 148         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 149                 return 1;
 150
 151         if (!(sample_type & PERF_SAMPLE_ID))
 152                 return -1;
 153
 154         if (sample_type & PERF_SAMPLE_CPU)
 155                 idx += 1;
 156
 157         if (sample_type & PERF_SAMPLE_STREAM_ID)
 158                 idx += 1;
 159
 160         return idx;
 161 }
 162
 163 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 164 {
 165         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 166         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 167 }
 168
 169 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 170                                   enum perf_event_sample_format bit)
 171 {
 172         if (!(evsel->attr.sample_type & bit)) {
 173                 evsel->attr.sample_type |= bit;
 174                 evsel->sample_size += sizeof(u64);
 175                 perf_evsel__calc_id_pos(evsel);
 176         }
 177 }
 178
 179 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 180                                     enum perf_event_sample_format bit)
 181 {
 182         if (evsel->attr.sample_type & bit) {
 183                 evsel->attr.sample_type &= ~bit;
 184                 evsel->sample_size -= sizeof(u64);
 185                 perf_evsel__calc_id_pos(evsel);
 186         }
 187 }
 188
 189 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 190                                bool can_sample_identifier)
 191 {
 192         if (can_sample_identifier) {
 193                 perf_evsel__reset_sample_bit(evsel, ID);
 194                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 195         } else {
 196                 perf_evsel__set_sample_bit(evsel, ID);
 197         }
 198         evsel->attr.read_format |= PERF_FORMAT_ID;
 199 }
 200
 201 void perf_evsel__init(struct perf_evsel *evsel,
 202                       struct perf_event_attr *attr, int idx)
 203 {
 204         evsel->idx         = idx;
 205         evsel->tracking    = !idx;
 206         evsel->attr        = *attr;
 207         evsel->leader      = evsel;
 208         evsel->unit        = "";
 209         evsel->scale       = 1.0;
 210         evsel->evlist      = NULL;
 211         evsel->bpf_fd      = -1;
 212         INIT_LIST_HEAD(&evsel->node);
 213         INIT_LIST_HEAD(&evsel->config_terms);
 214         perf_evsel__object.init(evsel);
 215         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 216         perf_evsel__calc_id_pos(evsel);
 217         evsel->cmdline_group_boundary = false;
 218 }
 219
 220 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 221 {
 222         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 223
 224         if (evsel != NULL)
 225                 perf_evsel__init(evsel, attr, idx);
 226
 227         return evsel;
 228 }
 229
 230 /*
 231  * Returns pointer with encoded error via <linux/err.h> interface.
 232  */
 233 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 234 {
 235         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 236         int err = -ENOMEM;
 237
 238         if (evsel == NULL) {
 239                 goto out_err;
 240         } else {
 241                 struct perf_event_attr attr = {
 242                         .type          = PERF_TYPE_TRACEPOINT,
 243                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 244                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 245                 };
 246
 247                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 248                         goto out_free;
 249
 250                 evsel->tp_format = trace_event__tp_format(sys, name);
 251                 if (IS_ERR(evsel->tp_format)) {
 252                         err = PTR_ERR(evsel->tp_format);
 253                         goto out_free;
 254                 }
 255
 256                 event_attr_init(&attr);
 257                 attr.config = evsel->tp_format->id;
 258                 attr.sample_period = 1;
 259                 perf_evsel__init(evsel, &attr, idx);
 260         }
 261
 262         return evsel;
 263
 264 out_free:
 265         zfree(&evsel->name);
 266         free(evsel);
 267 out_err:
 268         return ERR_PTR(err);
 269 }
 270
 271 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 272         "cycles",
 273         "instructions",
 274         "cache-references",
 275         "cache-misses",
 276         "branches",
 277         "branch-misses",
 278         "bus-cycles",
 279         "stalled-cycles-frontend",
 280         "stalled-cycles-backend",
 281         "ref-cycles",
 282 };
 283
 284 static const char *__perf_evsel__hw_name(u64 config)
 285 {
 286         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 287                 return perf_evsel__hw_names[config];
 288
 289         return "unknown-hardware";
 290 }
 291
 292 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 293 {
 294         int colon = 0, r = 0;
 295         struct perf_event_attr *attr = &evsel->attr;
 296         bool exclude_guest_default = false;
 297
 298 #define MOD_PRINT(context, mod) do {                                    \
 299                 if (!attr->exclude_##context) {                         \
 300                         if (!colon) colon = ++r;                        \
 301                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 302                 } } while(0)
 303
 304         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 305                 MOD_PRINT(kernel, 'k');
 306                 MOD_PRINT(user, 'u');
 307                 MOD_PRINT(hv, 'h');
 308                 exclude_guest_default = true;
 309         }
 310
 311         if (attr->precise_ip) {
 312                 if (!colon)
 313                         colon = ++r;
 314                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 315                 exclude_guest_default = true;
 316         }
 317
 318         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 319                 MOD_PRINT(host, 'H');
 320                 MOD_PRINT(guest, 'G');
 321         }
 322 #undef MOD_PRINT
 323         if (colon)
 324                 bf[colon - 1] = ':';
 325         return r;
 326 }
 327
 328 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 329 {
 330         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 331         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 332 }
 333
 334 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 335         "cpu-clock",
 336         "task-clock",
 337         "page-faults",
 338         "context-switches",
 339         "cpu-migrations",
 340         "minor-faults",
 341         "major-faults",
 342         "alignment-faults",
 343         "emulation-faults",
 344         "dummy",
 345 };
 346
 347 static const char *__perf_evsel__sw_name(u64 config)
 348 {
 349         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 350                 return perf_evsel__sw_names[config];
 351         return "unknown-software";
 352 }
 353
 354 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 355 {
 356         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 357         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 358 }
 359
 360 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 361 {
 362         int r;
 363
 364         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 365
 366         if (type & HW_BREAKPOINT_R)
 367                 r += scnprintf(bf + r, size - r, "r");
 368
 369         if (type & HW_BREAKPOINT_W)
 370                 r += scnprintf(bf + r, size - r, "w");
 371
 372         if (type & HW_BREAKPOINT_X)
 373                 r += scnprintf(bf + r, size - r, "x");
 374
 375         return r;
 376 }
 377
 378 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 379 {
 380         struct perf_event_attr *attr = &evsel->attr;
 381         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 382         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 383 }
 384
 385 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 386                                 [PERF_EVSEL__MAX_ALIASES] = {
 387  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 388  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 389  { "LLC",       "L2",                                                   },
 390  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 391  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 392  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 393  { "node",                                                              },
 394 };
 395
 396 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 397                                    [PERF_EVSEL__MAX_ALIASES] = {
 398  { "load",      "loads",        "read",                                 },
 399  { "store",     "stores",       "write",                                },
 400  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 401 };
 402
 403 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 404                                        [PERF_EVSEL__MAX_ALIASES] = {
 405  { "refs",      "Reference",    "ops",          "access",               },
 406  { "misses",    "miss",                                                 },
 407 };
 408
 409 #define C(x)            PERF_COUNT_HW_CACHE_##x
 410 #define CACHE_READ      (1 << C(OP_READ))
 411 #define CACHE_WRITE     (1 << C(OP_WRITE))
 412 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 413 #define COP(x)          (1 << x)
 414
 415 /*
 416  * cache operartion stat
 417  * L1I : Read and prefetch only
 418  * ITLB and BPU : Read-only
 419  */
 420 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 421  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 422  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 423  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 424  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 425  [C(ITLB)]      = (CACHE_READ),
 426  [C(BPU)]       = (CACHE_READ),
 427  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 428 };
 429
 430 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 431 {
 432         if (perf_evsel__hw_cache_stat[type] & COP(op))
 433                 return true;    /* valid */
 434         else
 435                 return false;   /* invalid */
 436 }
 437
 438 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 439                                             char *bf, size_t size)
 440 {
 441         if (result) {
 442                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 443                                  perf_evsel__hw_cache_op[op][0],
 444                                  perf_evsel__hw_cache_result[result][0]);
 445         }
 446
 447         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 448                          perf_evsel__hw_cache_op[op][1]);
 449 }
 450
 451 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 452 {
 453         u8 op, result, type = (config >>  0) & 0xff;
 454         const char *err = "unknown-ext-hardware-cache-type";
 455
 456         if (type > PERF_COUNT_HW_CACHE_MAX)
 457                 goto out_err;
 458
 459         op = (config >>  8) & 0xff;
 460         err = "unknown-ext-hardware-cache-op";
 461         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 462                 goto out_err;
 463
 464         result = (config >> 16) & 0xff;
 465         err = "unknown-ext-hardware-cache-result";
 466         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 467                 goto out_err;
 468
 469         err = "invalid-cache";
 470         if (!perf_evsel__is_cache_op_valid(type, op))
 471                 goto out_err;
 472
 473         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 474 out_err:
 475         return scnprintf(bf, size, "%s", err);
 476 }
 477
 478 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 479 {
 480         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 481         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 482 }
 483
 484 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 485 {
 486         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 487         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 488 }
 489
 490 const char *perf_evsel__name(struct perf_evsel *evsel)
 491 {
 492         char bf[128];
 493
 494         if (evsel->name)
 495                 return evsel->name;
 496
 497         switch (evsel->attr.type) {
 498         case PERF_TYPE_RAW:
 499                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 500                 break;
 501
 502         case PERF_TYPE_HARDWARE:
 503                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 504                 break;
 505
 506         case PERF_TYPE_HW_CACHE:
 507                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 508                 break;
 509
 510         case PERF_TYPE_SOFTWARE:
 511                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 512                 break;
 513
 514         case PERF_TYPE_TRACEPOINT:
 515                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 516                 break;
 517
 518         case PERF_TYPE_BREAKPOINT:
 519                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 520                 break;
 521
 522         default:
 523                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 524                           evsel->attr.type);
 525                 break;
 526         }
 527
 528         evsel->name = strdup(bf);
 529
 530         return evsel->name ?: "unknown";
 531 }
 532
 533 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 534 {
 535         return evsel->group_name ?: "anon group";
 536 }
 537
 538 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 539 {
 540         int ret;
 541         struct perf_evsel *pos;
 542         const char *group_name = perf_evsel__group_name(evsel);
 543
 544         ret = scnprintf(buf, size, "%s", group_name);
 545
 546         ret += scnprintf(buf + ret, size - ret, " { %s",
 547                          perf_evsel__name(evsel));
 548
 549         for_each_group_member(pos, evsel)
 550                 ret += scnprintf(buf + ret, size - ret, ", %s",
 551                                  perf_evsel__name(pos));
 552
 553         ret += scnprintf(buf + ret, size - ret, " }");
 554
 555         return ret;
 556 }
 557
 558 static void
 559 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 560                              struct record_opts *opts,
 561                              struct callchain_param *param)
 562 {
 563         bool function = perf_evsel__is_function_event(evsel);
 564         struct perf_event_attr *attr = &evsel->attr;
 565
 566         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 567
 568         if (param->record_mode == CALLCHAIN_LBR) {
 569                 if (!opts->branch_stack) {
 570                         if (attr->exclude_user) {
 571                                 pr_warning("LBR callstack option is only available "
 572                                            "to get user callchain information. "
 573                                            "Falling back to framepointers.\n");
 574                         } else {
 575                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 576                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 577                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 578                         }
 579                 } else
 580                          pr_warning("Cannot use LBR callstack with branch stack. "
 581                                     "Falling back to framepointers.\n");
 582         }
 583
 584         if (param->record_mode == CALLCHAIN_DWARF) {
 585                 if (!function) {
 586                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 587                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 588                         attr->sample_regs_user = PERF_REGS_MASK;
 589                         attr->sample_stack_user = param->dump_size;
 590                         attr->exclude_callchain_user = 1;
 591                 } else {
 592                         pr_info("Cannot use DWARF unwind for function trace event,"
 593                                 " falling back to framepointers.\n");
 594                 }
 595         }
 596
 597         if (function) {
 598                 pr_info("Disabling user space callchains for function trace event.\n");
 599                 attr->exclude_callchain_user = 1;
 600         }
 601 }
 602
 603 static void
 604 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 605                             struct callchain_param *param)
 606 {
 607         struct perf_event_attr *attr = &evsel->attr;
 608
 609         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 610         if (param->record_mode == CALLCHAIN_LBR) {
 611                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 612                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 613                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 614         }
 615         if (param->record_mode == CALLCHAIN_DWARF) {
 616                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 617                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 618         }
 619 }
 620
 621 static void apply_config_terms(struct perf_evsel *evsel,
 622                                struct record_opts *opts)
 623 {
 624         struct perf_evsel_config_term *term;
 625         struct list_head *config_terms = &evsel->config_terms;
 626         struct perf_event_attr *attr = &evsel->attr;
 627         struct callchain_param param;
 628         u32 dump_size = 0;
 629         char *callgraph_buf = NULL;
 630
 631         /* callgraph default */
 632         param.record_mode = callchain_param.record_mode;
 633
 634         list_for_each_entry(term, config_terms, list) {
 635                 switch (term->type) {
 636                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 637                         attr->sample_period = term->val.period;
 638                         attr->freq = 0;
 639                         break;
 640                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 641                         attr->sample_freq = term->val.freq;
 642                         attr->freq = 1;
 643                         break;
 644                 case PERF_EVSEL__CONFIG_TERM_TIME:
 645                         if (term->val.time)
 646                                 perf_evsel__set_sample_bit(evsel, TIME);
 647                         else
 648                                 perf_evsel__reset_sample_bit(evsel, TIME);
 649                         break;
 650                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 651                         callgraph_buf = term->val.callgraph;
 652                         break;
 653                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 654                         dump_size = term->val.stack_user;
 655                         break;
 656                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 657                         /*
 658                          * attr->inherit should has already been set by
 659                          * perf_evsel__config. If user explicitly set
 660                          * inherit using config terms, override global
 661                          * opt->no_inherit setting.
 662                          */
 663                         attr->inherit = term->val.inherit ? 1 : 0;
 664                         break;
 665                 default:
 666                         break;
 667                 }
 668         }
 669
 670         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 671         if ((callgraph_buf != NULL) || (dump_size > 0)) {
 672
 673                 /* parse callgraph parameters */
 674                 if (callgraph_buf != NULL) {
 675                         if (!strcmp(callgraph_buf, "no")) {
 676                                 param.enabled = false;
 677                                 param.record_mode = CALLCHAIN_NONE;
 678                         } else {
 679                                 param.enabled = true;
 680                                 if (parse_callchain_record(callgraph_buf, &param)) {
 681                                         pr_err("per-event callgraph setting for %s failed. "
 682                                                "Apply callgraph global setting for it\n",
 683                                                evsel->name);
 684                                         return;
 685                                 }
 686                         }
 687                 }
 688                 if (dump_size > 0) {
 689                         dump_size = round_up(dump_size, sizeof(u64));
 690                         param.dump_size = dump_size;
 691                 }
 692
 693                 /* If global callgraph set, clear it */
 694                 if (callchain_param.enabled)
 695                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 696
 697                 /* set perf-event callgraph */
 698                 if (param.enabled)
 699                         perf_evsel__config_callgraph(evsel, opts, &param);
 700         }
 701 }
 702
 703 /*
 704  * The enable_on_exec/disabled value strategy:
 705  *
 706  *  1) For any type of traced program:
 707  *    - all independent events and group leaders are disabled
 708  *    - all group members are enabled
 709  *
 710  *     Group members are ruled by group leaders. They need to
 711  *     be enabled, because the group scheduling relies on that.
 712  *
 713  *  2) For traced programs executed by perf:
 714  *     - all independent events and group leaders have
 715  *       enable_on_exec set
 716  *     - we don't specifically enable or disable any event during
 717  *       the record command
 718  *
 719  *     Independent events and group leaders are initially disabled
 720  *     and get enabled by exec. Group members are ruled by group
 721  *     leaders as stated in 1).
 722  *
 723  *  3) For traced programs attached by perf (pid/tid):
 724  *     - we specifically enable or disable all events during
 725  *       the record command
 726  *
 727  *     When attaching events to already running traced we
 728  *     enable/disable events specifically, as there's no
 729  *     initial traced exec call.
 730  */
 731 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 732 {
 733         struct perf_evsel *leader = evsel->leader;
 734         struct perf_event_attr *attr = &evsel->attr;
 735         int track = evsel->tracking;
 736         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 737
 738         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 739         attr->inherit       = !opts->no_inherit;
 740
 741         perf_evsel__set_sample_bit(evsel, IP);
 742         perf_evsel__set_sample_bit(evsel, TID);
 743
 744         if (evsel->sample_read) {
 745                 perf_evsel__set_sample_bit(evsel, READ);
 746
 747                 /*
 748                  * We need ID even in case of single event, because
 749                  * PERF_SAMPLE_READ process ID specific data.
 750                  */
 751                 perf_evsel__set_sample_id(evsel, false);
 752
 753                 /*
 754                  * Apply group format only if we belong to group
 755                  * with more than one members.
 756                  */
 757                 if (leader->nr_members > 1) {
 758                         attr->read_format |= PERF_FORMAT_GROUP;
 759                         attr->inherit = 0;
 760                 }
 761         }
 762
 763         /*
 764          * We default some events to have a default interval. But keep
 765          * it a weak assumption overridable by the user.
 766          */
 767         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 768                                      opts->user_interval != ULLONG_MAX)) {
 769                 if (opts->freq) {
 770                         perf_evsel__set_sample_bit(evsel, PERIOD);
 771                         attr->freq              = 1;
 772                         attr->sample_freq       = opts->freq;
 773                 } else {
 774                         attr->sample_period = opts->default_interval;
 775                 }
 776         }
 777
 778         /*
 779          * Disable sampling for all group members other
 780          * than leader in case leader 'leads' the sampling.
 781          */
 782         if ((leader != evsel) && leader->sample_read) {
 783                 attr->sample_freq   = 0;
 784                 attr->sample_period = 0;
 785         }
 786
 787         if (opts->no_samples)
 788                 attr->sample_freq = 0;
 789
 790         if (opts->inherit_stat)
 791                 attr->inherit_stat = 1;
 792
 793         if (opts->sample_address) {
 794                 perf_evsel__set_sample_bit(evsel, ADDR);
 795                 attr->mmap_data = track;
 796         }
 797
 798         /*
 799          * We don't allow user space callchains for  function trace
 800          * event, due to issues with page faults while tracing page
 801          * fault handler and its overall trickiness nature.
 802          */
 803         if (perf_evsel__is_function_event(evsel))
 804                 evsel->attr.exclude_callchain_user = 1;
 805
 806         if (callchain_param.enabled && !evsel->no_aux_samples)
 807                 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
 808
 809         if (opts->sample_intr_regs) {
 810                 attr->sample_regs_intr = opts->sample_intr_regs;
 811                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 812         }
 813
 814         if (target__has_cpu(&opts->target))
 815                 perf_evsel__set_sample_bit(evsel, CPU);
 816
 817         if (opts->period)
 818                 perf_evsel__set_sample_bit(evsel, PERIOD);
 819
 820         /*
 821          * When the user explicitely disabled time don't force it here.
 822          */
 823         if (opts->sample_time &&
 824             (!perf_missing_features.sample_id_all &&
 825             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 826              opts->sample_time_set)))
 827                 perf_evsel__set_sample_bit(evsel, TIME);
 828
 829         if (opts->raw_samples && !evsel->no_aux_samples) {
 830                 perf_evsel__set_sample_bit(evsel, TIME);
 831                 perf_evsel__set_sample_bit(evsel, RAW);
 832                 perf_evsel__set_sample_bit(evsel, CPU);
 833         }
 834
 835         if (opts->sample_address)
 836                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 837
 838         if (opts->no_buffering) {
 839                 attr->watermark = 0;
 840                 attr->wakeup_events = 1;
 841         }
 842         if (opts->branch_stack && !evsel->no_aux_samples) {
 843                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 844                 attr->branch_sample_type = opts->branch_stack;
 845         }
 846
 847         if (opts->sample_weight)
 848                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 849
 850         attr->task  = track;
 851         attr->mmap  = track;
 852         attr->mmap2 = track && !perf_missing_features.mmap2;
 853         attr->comm  = track;
 854
 855         if (opts->record_switch_events)
 856                 attr->context_switch = track;
 857
 858         if (opts->sample_transaction)
 859                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 860
 861         if (opts->running_time) {
 862                 evsel->attr.read_format |=
 863                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 864                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 865         }
 866
 867         /*
 868          * XXX see the function comment above
 869          *
 870          * Disabling only independent events or group leaders,
 871          * keeping group members enabled.
 872          */
 873         if (perf_evsel__is_group_leader(evsel))
 874                 attr->disabled = 1;
 875
 876         /*
 877          * Setting enable_on_exec for independent events and
 878          * group leaders for traced executed by perf.
 879          */
 880         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 881                 !opts->initial_delay)
 882                 attr->enable_on_exec = 1;
 883
 884         if (evsel->immediate) {
 885                 attr->disabled = 0;
 886                 attr->enable_on_exec = 0;
 887         }
 888
 889         clockid = opts->clockid;
 890         if (opts->use_clockid) {
 891                 attr->use_clockid = 1;
 892                 attr->clockid = opts->clockid;
 893         }
 894
 895         if (evsel->precise_max)
 896                 perf_event_attr__set_max_precise_ip(attr);
 897
 898         /*
 899          * Apply event specific term settings,
 900          * it overloads any global configuration.
 901          */
 902         apply_config_terms(evsel, opts);
 903 }
 904
 905 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 906 {
 907         int cpu, thread;
 908
 909         if (evsel->system_wide)
 910                 nthreads = 1;
 911
 912         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 913
 914         if (evsel->fd) {
 915                 for (cpu = 0; cpu < ncpus; cpu++) {
 916                         for (thread = 0; thread < nthreads; thread++) {
 917                                 FD(evsel, cpu, thread) = -1;
 918                         }
 919                 }
 920         }
 921
 922         return evsel->fd != NULL ? 0 : -ENOMEM;
 923 }
 924
 925 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 926                           int ioc,  void *arg)
 927 {
 928         int cpu, thread;
 929
 930         if (evsel->system_wide)
 931                 nthreads = 1;
 932
 933         for (cpu = 0; cpu < ncpus; cpu++) {
 934                 for (thread = 0; thread < nthreads; thread++) {
 935                         int fd = FD(evsel, cpu, thread),
 936                             err = ioctl(fd, ioc, arg);
 937
 938                         if (err)
 939                                 return err;
 940                 }
 941         }
 942
 943         return 0;
 944 }
 945
 946 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 947                              const char *filter)
 948 {
 949         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 950                                      PERF_EVENT_IOC_SET_FILTER,
 951                                      (void *)filter);
 952 }
 953
 954 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 955 {
 956         char *new_filter = strdup(filter);
 957
 958         if (new_filter != NULL) {
 959                 free(evsel->filter);
 960                 evsel->filter = new_filter;
 961                 return 0;
 962         }
 963
 964         return -1;
 965 }
 966
 967 int perf_evsel__append_filter(struct perf_evsel *evsel,
 968                               const char *op, const char *filter)
 969 {
 970         char *new_filter;
 971
 972         if (evsel->filter == NULL)
 973                 return perf_evsel__set_filter(evsel, filter);
 974
 975         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 976                 free(evsel->filter);
 977                 evsel->filter = new_filter;
 978                 return 0;
 979         }
 980
 981         return -1;
 982 }
 983
 984 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 985 {
 986         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 987                                      PERF_EVENT_IOC_ENABLE,
 988                                      0);
 989 }
 990
 991 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 992 {
 993         if (ncpus == 0 || nthreads == 0)
 994                 return 0;
 995
 996         if (evsel->system_wide)
 997                 nthreads = 1;
 998
 999         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1000         if (evsel->sample_id == NULL)
1001                 return -ENOMEM;
1002
1003         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1004         if (evsel->id == NULL) {
1005                 xyarray__delete(evsel->sample_id);
1006                 evsel->sample_id = NULL;
1007                 return -ENOMEM;
1008         }
1009
1010         return 0;
1011 }
1012
1013 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1014 {
1015         xyarray__delete(evsel->fd);
1016         evsel->fd = NULL;
1017 }
1018
1019 static void perf_evsel__free_id(struct perf_evsel *evsel)
1020 {
1021         xyarray__delete(evsel->sample_id);
1022         evsel->sample_id = NULL;
1023         zfree(&evsel->id);
1024 }
1025
1026 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1027 {
1028         struct perf_evsel_config_term *term, *h;
1029
1030         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1031                 list_del(&term->list);
1032                 free(term);
1033         }
1034 }
1035
1036 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1037 {
1038         int cpu, thread;
1039
1040         if (evsel->system_wide)
1041                 nthreads = 1;
1042
1043         for (cpu = 0; cpu < ncpus; cpu++)
1044                 for (thread = 0; thread < nthreads; ++thread) {
1045                         close(FD(evsel, cpu, thread));
1046                         FD(evsel, cpu, thread) = -1;
1047                 }
1048 }
1049
1050 void perf_evsel__exit(struct perf_evsel *evsel)
1051 {
1052         assert(list_empty(&evsel->node));
1053         assert(evsel->evlist == NULL);
1054         perf_evsel__free_fd(evsel);
1055         perf_evsel__free_id(evsel);
1056         perf_evsel__free_config_terms(evsel);
1057         close_cgroup(evsel->cgrp);
1058         cpu_map__put(evsel->cpus);
1059         cpu_map__put(evsel->own_cpus);
1060         thread_map__put(evsel->threads);
1061         zfree(&evsel->group_name);
1062         zfree(&evsel->name);
1063         perf_evsel__object.fini(evsel);
1064 }
1065
1066 void perf_evsel__delete(struct perf_evsel *evsel)
1067 {
1068         perf_evsel__exit(evsel);
1069         free(evsel);
1070 }
1071
1072 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1073                                 struct perf_counts_values *count)
1074 {
1075         struct perf_counts_values tmp;
1076
1077         if (!evsel->prev_raw_counts)
1078                 return;
1079
1080         if (cpu == -1) {
1081                 tmp = evsel->prev_raw_counts->aggr;
1082                 evsel->prev_raw_counts->aggr = *count;
1083         } else {
1084                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1085                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1086         }
1087
1088         count->val = count->val - tmp.val;
1089         count->ena = count->ena - tmp.ena;
1090         count->run = count->run - tmp.run;
1091 }
1092
1093 void perf_counts_values__scale(struct perf_counts_values *count,
1094                                bool scale, s8 *pscaled)
1095 {
1096         s8 scaled = 0;
1097
1098         if (scale) {
1099                 if (count->run == 0) {
1100                         scaled = -1;
1101                         count->val = 0;
1102                 } else if (count->run < count->ena) {
1103                         scaled = 1;
1104                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1105                 }
1106         } else
1107                 count->ena = count->run = 0;
1108
1109         if (pscaled)
1110                 *pscaled = scaled;
1111 }
1112
1113 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1114                      struct perf_counts_values *count)
1115 {
1116         memset(count, 0, sizeof(*count));
1117
1118         if (FD(evsel, cpu, thread) < 0)
1119                 return -EINVAL;
1120
1121         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1122                 return -errno;
1123
1124         return 0;
1125 }
1126
1127 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1128                               int cpu, int thread, bool scale)
1129 {
1130         struct perf_counts_values count;
1131         size_t nv = scale ? 3 : 1;
1132
1133         if (FD(evsel, cpu, thread) < 0)
1134                 return -EINVAL;
1135
1136         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1137                 return -ENOMEM;
1138
1139         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1140                 return -errno;
1141
1142         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1143         perf_counts_values__scale(&count, scale, NULL);
1144         *perf_counts(evsel->counts, cpu, thread) = count;
1145         return 0;
1146 }
1147
1148 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1149 {
1150         struct perf_evsel *leader = evsel->leader;
1151         int fd;
1152
1153         if (perf_evsel__is_group_leader(evsel))
1154                 return -1;
1155
1156         /*
1157          * Leader must be already processed/open,
1158          * if not it's a bug.
1159          */
1160         BUG_ON(!leader->fd);
1161
1162         fd = FD(leader, cpu, thread);
1163         BUG_ON(fd == -1);
1164
1165         return fd;
1166 }
1167
1168 struct bit_names {
1169         int bit;
1170         const char *name;
1171 };
1172
1173 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1174 {
1175         bool first_bit = true;
1176         int i = 0;
1177
1178         do {
1179                 if (value & bits[i].bit) {
1180                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1181                         first_bit = false;
1182                 }
1183         } while (bits[++i].name != NULL);
1184 }
1185
1186 static void __p_sample_type(char *buf, size_t size, u64 value)
1187 {
1188 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1189         struct bit_names bits[] = {
1190                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1191                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1192                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1193                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1194                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1195                 { .name = NULL, }
1196         };
1197 #undef bit_name
1198         __p_bits(buf, size, value, bits);
1199 }
1200
1201 static void __p_read_format(char *buf, size_t size, u64 value)
1202 {
1203 #define bit_name(n) { PERF_FORMAT_##n, #n }
1204         struct bit_names bits[] = {
1205                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1206                 bit_name(ID), bit_name(GROUP),
1207                 { .name = NULL, }
1208         };
1209 #undef bit_name
1210         __p_bits(buf, size, value, bits);
1211 }
1212
1213 #define BUF_SIZE                1024
1214
1215 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1216 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1217 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1218 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1219 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1220
1221 #define PRINT_ATTRn(_n, _f, _p)                         \
1222 do {                                                    \
1223         if (attr->_f) {                                 \
1224                 _p(attr->_f);                           \
1225                 ret += attr__fprintf(fp, _n, buf, priv);\
1226         }                                               \
1227 } while (0)
1228
1229 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1230
1231 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1232                              attr__fprintf_f attr__fprintf, void *priv)
1233 {
1234         char buf[BUF_SIZE];
1235         int ret = 0;
1236
1237         PRINT_ATTRf(type, p_unsigned);
1238         PRINT_ATTRf(size, p_unsigned);
1239         PRINT_ATTRf(config, p_hex);
1240         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1241         PRINT_ATTRf(sample_type, p_sample_type);
1242         PRINT_ATTRf(read_format, p_read_format);
1243
1244         PRINT_ATTRf(disabled, p_unsigned);
1245         PRINT_ATTRf(inherit, p_unsigned);
1246         PRINT_ATTRf(pinned, p_unsigned);
1247         PRINT_ATTRf(exclusive, p_unsigned);
1248         PRINT_ATTRf(exclude_user, p_unsigned);
1249         PRINT_ATTRf(exclude_kernel, p_unsigned);
1250         PRINT_ATTRf(exclude_hv, p_unsigned);
1251         PRINT_ATTRf(exclude_idle, p_unsigned);
1252         PRINT_ATTRf(mmap, p_unsigned);
1253         PRINT_ATTRf(comm, p_unsigned);
1254         PRINT_ATTRf(freq, p_unsigned);
1255         PRINT_ATTRf(inherit_stat, p_unsigned);
1256         PRINT_ATTRf(enable_on_exec, p_unsigned);
1257         PRINT_ATTRf(task, p_unsigned);
1258         PRINT_ATTRf(watermark, p_unsigned);
1259         PRINT_ATTRf(precise_ip, p_unsigned);
1260         PRINT_ATTRf(mmap_data, p_unsigned);
1261         PRINT_ATTRf(sample_id_all, p_unsigned);
1262         PRINT_ATTRf(exclude_host, p_unsigned);
1263         PRINT_ATTRf(exclude_guest, p_unsigned);
1264         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1265         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1266         PRINT_ATTRf(mmap2, p_unsigned);
1267         PRINT_ATTRf(comm_exec, p_unsigned);
1268         PRINT_ATTRf(use_clockid, p_unsigned);
1269         PRINT_ATTRf(context_switch, p_unsigned);
1270
1271         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1272         PRINT_ATTRf(bp_type, p_unsigned);
1273         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1274         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1275         PRINT_ATTRf(branch_sample_type, p_unsigned);
1276         PRINT_ATTRf(sample_regs_user, p_hex);
1277         PRINT_ATTRf(sample_stack_user, p_unsigned);
1278         PRINT_ATTRf(clockid, p_signed);
1279         PRINT_ATTRf(sample_regs_intr, p_hex);
1280         PRINT_ATTRf(aux_watermark, p_unsigned);
1281
1282         return ret;
1283 }
1284
1285 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1286                                 void *priv __attribute__((unused)))
1287 {
1288         return fprintf(fp, "  %-32s %s\n", name, val);
1289 }
1290
1291 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1292                               struct thread_map *threads)
1293 {
1294         int cpu, thread, nthreads;
1295         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1296         int pid = -1, err;
1297         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1298
1299         if (evsel->system_wide)
1300                 nthreads = 1;
1301         else
1302                 nthreads = threads->nr;
1303
1304         if (evsel->fd == NULL &&
1305             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1306                 return -ENOMEM;
1307
1308         if (evsel->cgrp) {
1309                 flags |= PERF_FLAG_PID_CGROUP;
1310                 pid = evsel->cgrp->fd;
1311         }
1312
1313 fallback_missing_features:
1314         if (perf_missing_features.clockid_wrong)
1315                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1316         if (perf_missing_features.clockid) {
1317                 evsel->attr.use_clockid = 0;
1318                 evsel->attr.clockid = 0;
1319         }
1320         if (perf_missing_features.cloexec)
1321                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1322         if (perf_missing_features.mmap2)
1323                 evsel->attr.mmap2 = 0;
1324         if (perf_missing_features.exclude_guest)
1325                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1326 retry_sample_id:
1327         if (perf_missing_features.sample_id_all)
1328                 evsel->attr.sample_id_all = 0;
1329
1330         if (verbose >= 2) {
1331                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1332                 fprintf(stderr, "perf_event_attr:\n");
1333                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1334                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1335         }
1336
1337         for (cpu = 0; cpu < cpus->nr; cpu++) {
1338
1339                 for (thread = 0; thread < nthreads; thread++) {
1340                         int group_fd;
1341
1342                         if (!evsel->cgrp && !evsel->system_wide)
1343                                 pid = thread_map__pid(threads, thread);
1344
1345                         group_fd = get_group_fd(evsel, cpu, thread);
1346 retry_open:
1347                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1348                                   pid, cpus->map[cpu], group_fd, flags);
1349
1350                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1351                                                                      pid,
1352                                                                      cpus->map[cpu],
1353                                                                      group_fd, flags);
1354                         if (FD(evsel, cpu, thread) < 0) {
1355                                 err = -errno;
1356                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1357                                           err);
1358                                 goto try_fallback;
1359                         }
1360
1361                         if (evsel->bpf_fd >= 0) {
1362                                 int evt_fd = FD(evsel, cpu, thread);
1363                                 int bpf_fd = evsel->bpf_fd;
1364
1365                                 err = ioctl(evt_fd,
1366                                             PERF_EVENT_IOC_SET_BPF,
1367                                             bpf_fd);
1368                                 if (err && errno != EEXIST) {
1369                                         pr_err("failed to attach bpf fd %d: %s\n",
1370                                                bpf_fd, strerror(errno));
1371                                         err = -EINVAL;
1372                                         goto out_close;
1373                                 }
1374                         }
1375
1376                         set_rlimit = NO_CHANGE;
1377
1378                         /*
1379                          * If we succeeded but had to kill clockid, fail and
1380                          * have perf_evsel__open_strerror() print us a nice
1381                          * error.
1382                          */
1383                         if (perf_missing_features.clockid ||
1384                             perf_missing_features.clockid_wrong) {
1385                                 err = -EINVAL;
1386                                 goto out_close;
1387                         }
1388                 }
1389         }
1390
1391         return 0;
1392
1393 try_fallback:
1394         /*
1395          * perf stat needs between 5 and 22 fds per CPU. When we run out
1396          * of them try to increase the limits.
1397          */
1398         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1399                 struct rlimit l;
1400                 int old_errno = errno;
1401
1402                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1403                         if (set_rlimit == NO_CHANGE)
1404                                 l.rlim_cur = l.rlim_max;
1405                         else {
1406                                 l.rlim_cur = l.rlim_max + 1000;
1407                                 l.rlim_max = l.rlim_cur;
1408                         }
1409                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1410                                 set_rlimit++;
1411                                 errno = old_errno;
1412                                 goto retry_open;
1413                         }
1414                 }
1415                 errno = old_errno;
1416         }
1417
1418         if (err != -EINVAL || cpu > 0 || thread > 0)
1419                 goto out_close;
1420
1421         /*
1422          * Must probe features in the order they were added to the
1423          * perf_event_attr interface.
1424          */
1425         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1426                 perf_missing_features.clockid_wrong = true;
1427                 goto fallback_missing_features;
1428         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1429                 perf_missing_features.clockid = true;
1430                 goto fallback_missing_features;
1431         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1432                 perf_missing_features.cloexec = true;
1433                 goto fallback_missing_features;
1434         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1435                 perf_missing_features.mmap2 = true;
1436                 goto fallback_missing_features;
1437         } else if (!perf_missing_features.exclude_guest &&
1438                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1439                 perf_missing_features.exclude_guest = true;
1440                 goto fallback_missing_features;
1441         } else if (!perf_missing_features.sample_id_all) {
1442                 perf_missing_features.sample_id_all = true;
1443                 goto retry_sample_id;
1444         }
1445
1446 out_close:
1447         do {
1448                 while (--thread >= 0) {
1449                         close(FD(evsel, cpu, thread));
1450                         FD(evsel, cpu, thread) = -1;
1451                 }
1452                 thread = nthreads;
1453         } while (--cpu >= 0);
1454         return err;
1455 }
1456
1457 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1458 {
1459         if (evsel->fd == NULL)
1460                 return;
1461
1462         perf_evsel__close_fd(evsel, ncpus, nthreads);
1463         perf_evsel__free_fd(evsel);
1464 }
1465
1466 static struct {
1467         struct cpu_map map;
1468         int cpus[1];
1469 } empty_cpu_map = {
1470         .map.nr = 1,
1471         .cpus   = { -1, },
1472 };
1473
1474 static struct {
1475         struct thread_map map;
1476         int threads[1];
1477 } empty_thread_map = {
1478         .map.nr  = 1,
1479         .threads = { -1, },
1480 };
1481
1482 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1483                      struct thread_map *threads)
1484 {
1485         if (cpus == NULL) {
1486                 /* Work around old compiler warnings about strict aliasing */
1487                 cpus = &empty_cpu_map.map;
1488         }
1489
1490         if (threads == NULL)
1491                 threads = &empty_thread_map.map;
1492
1493         return __perf_evsel__open(evsel, cpus, threads);
1494 }
1495
1496 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1497                              struct cpu_map *cpus)
1498 {
1499         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1500 }
1501
1502 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1503                                 struct thread_map *threads)
1504 {
1505         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1506 }
1507
1508 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1509                                        const union perf_event *event,
1510                                        struct perf_sample *sample)
1511 {
1512         u64 type = evsel->attr.sample_type;
1513         const u64 *array = event->sample.array;
1514         bool swapped = evsel->needs_swap;
1515         union u64_swap u;
1516
1517         array += ((event->header.size -
1518                    sizeof(event->header)) / sizeof(u64)) - 1;
1519
1520         if (type & PERF_SAMPLE_IDENTIFIER) {
1521                 sample->id = *array;
1522                 array--;
1523         }
1524
1525         if (type & PERF_SAMPLE_CPU) {
1526                 u.val64 = *array;
1527                 if (swapped) {
1528                         /* undo swap of u64, then swap on individual u32s */
1529                         u.val64 = bswap_64(u.val64);
1530                         u.val32[0] = bswap_32(u.val32[0]);
1531                 }
1532
1533                 sample->cpu = u.val32[0];
1534                 array--;
1535         }
1536
1537         if (type & PERF_SAMPLE_STREAM_ID) {
1538                 sample->stream_id = *array;
1539                 array--;
1540         }
1541
1542         if (type & PERF_SAMPLE_ID) {
1543                 sample->id = *array;
1544                 array--;
1545         }
1546
1547         if (type & PERF_SAMPLE_TIME) {
1548                 sample->time = *array;
1549                 array--;
1550         }
1551
1552         if (type & PERF_SAMPLE_TID) {
1553                 u.val64 = *array;
1554                 if (swapped) {
1555                         /* undo swap of u64, then swap on individual u32s */
1556                         u.val64 = bswap_64(u.val64);
1557                         u.val32[0] = bswap_32(u.val32[0]);
1558                         u.val32[1] = bswap_32(u.val32[1]);
1559                 }
1560
1561                 sample->pid = u.val32[0];
1562                 sample->tid = u.val32[1];
1563                 array--;
1564         }
1565
1566         return 0;
1567 }
1568
1569 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1570                             u64 size)
1571 {
1572         return size > max_size || offset + size > endp;
1573 }
1574
1575 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1576         do {                                                            \
1577                 if (overflow(endp, (max_size), (offset), (size)))       \
1578                         return -EFAULT;                                 \
1579         } while (0)
1580
1581 #define OVERFLOW_CHECK_u64(offset) \
1582         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1583
1584 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1585                              struct perf_sample *data)
1586 {
1587         u64 type = evsel->attr.sample_type;
1588         bool swapped = evsel->needs_swap;
1589         const u64 *array;
1590         u16 max_size = event->header.size;
1591         const void *endp = (void *)event + max_size;
1592         u64 sz;
1593
1594         /*
1595          * used for cross-endian analysis. See git commit 65014ab3
1596          * for why this goofiness is needed.
1597          */
1598         union u64_swap u;
1599
1600         memset(data, 0, sizeof(*data));
1601         data->cpu = data->pid = data->tid = -1;
1602         data->stream_id = data->id = data->time = -1ULL;
1603         data->period = evsel->attr.sample_period;
1604         data->weight = 0;
1605
1606         if (event->header.type != PERF_RECORD_SAMPLE) {
1607                 if (!evsel->attr.sample_id_all)
1608                         return 0;
1609                 return perf_evsel__parse_id_sample(evsel, event, data);
1610         }
1611
1612         array = event->sample.array;
1613
1614         /*
1615          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1616          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1617          * check the format does not go past the end of the event.
1618          */
1619         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1620                 return -EFAULT;
1621
1622         data->id = -1ULL;
1623         if (type & PERF_SAMPLE_IDENTIFIER) {
1624                 data->id = *array;
1625                 array++;
1626         }
1627
1628         if (type & PERF_SAMPLE_IP) {
1629                 data->ip = *array;
1630                 array++;
1631         }
1632
1633         if (type & PERF_SAMPLE_TID) {
1634                 u.val64 = *array;
1635                 if (swapped) {
1636                         /* undo swap of u64, then swap on individual u32s */
1637                         u.val64 = bswap_64(u.val64);
1638                         u.val32[0] = bswap_32(u.val32[0]);
1639                         u.val32[1] = bswap_32(u.val32[1]);
1640                 }
1641
1642                 data->pid = u.val32[0];
1643                 data->tid = u.val32[1];
1644                 array++;
1645         }
1646
1647         if (type & PERF_SAMPLE_TIME) {
1648                 data->time = *array;
1649                 array++;
1650         }
1651
1652         data->addr = 0;
1653         if (type & PERF_SAMPLE_ADDR) {
1654                 data->addr = *array;
1655                 array++;
1656         }
1657
1658         if (type & PERF_SAMPLE_ID) {
1659                 data->id = *array;
1660                 array++;
1661         }
1662
1663         if (type & PERF_SAMPLE_STREAM_ID) {
1664                 data->stream_id = *array;
1665                 array++;
1666         }
1667
1668         if (type & PERF_SAMPLE_CPU) {
1669
1670                 u.val64 = *array;
1671                 if (swapped) {
1672                         /* undo swap of u64, then swap on individual u32s */
1673                         u.val64 = bswap_64(u.val64);
1674                         u.val32[0] = bswap_32(u.val32[0]);
1675                 }
1676
1677                 data->cpu = u.val32[0];
1678                 array++;
1679         }
1680
1681         if (type & PERF_SAMPLE_PERIOD) {
1682                 data->period = *array;
1683                 array++;
1684         }
1685
1686         if (type & PERF_SAMPLE_READ) {
1687                 u64 read_format = evsel->attr.read_format;
1688
1689                 OVERFLOW_CHECK_u64(array);
1690                 if (read_format & PERF_FORMAT_GROUP)
1691                         data->read.group.nr = *array;
1692                 else
1693                         data->read.one.value = *array;
1694
1695                 array++;
1696
1697                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1698                         OVERFLOW_CHECK_u64(array);
1699                         data->read.time_enabled = *array;
1700                         array++;
1701                 }
1702
1703                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1704                         OVERFLOW_CHECK_u64(array);
1705                         data->read.time_running = *array;
1706                         array++;
1707                 }
1708
1709                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1710                 if (read_format & PERF_FORMAT_GROUP) {
1711                         const u64 max_group_nr = UINT64_MAX /
1712                                         sizeof(struct sample_read_value);
1713
1714                         if (data->read.group.nr > max_group_nr)
1715                                 return -EFAULT;
1716                         sz = data->read.group.nr *
1717                              sizeof(struct sample_read_value);
1718                         OVERFLOW_CHECK(array, sz, max_size);
1719                         data->read.group.values =
1720                                         (struct sample_read_value *)array;
1721                         array = (void *)array + sz;
1722                 } else {
1723                         OVERFLOW_CHECK_u64(array);
1724                         data->read.one.id = *array;
1725                         array++;
1726                 }
1727         }
1728
1729         if (type & PERF_SAMPLE_CALLCHAIN) {
1730                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1731
1732                 OVERFLOW_CHECK_u64(array);
1733                 data->callchain = (struct ip_callchain *)array++;
1734                 if (data->callchain->nr > max_callchain_nr)
1735                         return -EFAULT;
1736                 sz = data->callchain->nr * sizeof(u64);
1737                 OVERFLOW_CHECK(array, sz, max_size);
1738                 array = (void *)array + sz;
1739         }
1740
1741         if (type & PERF_SAMPLE_RAW) {
1742                 OVERFLOW_CHECK_u64(array);
1743                 u.val64 = *array;
1744                 if (WARN_ONCE(swapped,
1745                               "Endianness of raw data not corrected!\n")) {
1746                         /* undo swap of u64, then swap on individual u32s */
1747                         u.val64 = bswap_64(u.val64);
1748                         u.val32[0] = bswap_32(u.val32[0]);
1749                         u.val32[1] = bswap_32(u.val32[1]);
1750                 }
1751                 data->raw_size = u.val32[0];
1752                 array = (void *)array + sizeof(u32);
1753
1754                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1755                 data->raw_data = (void *)array;
1756                 array = (void *)array + data->raw_size;
1757         }
1758
1759         if (type & PERF_SAMPLE_BRANCH_STACK) {
1760                 const u64 max_branch_nr = UINT64_MAX /
1761                                           sizeof(struct branch_entry);
1762
1763                 OVERFLOW_CHECK_u64(array);
1764                 data->branch_stack = (struct branch_stack *)array++;
1765
1766                 if (data->branch_stack->nr > max_branch_nr)
1767                         return -EFAULT;
1768                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1769                 OVERFLOW_CHECK(array, sz, max_size);
1770                 array = (void *)array + sz;
1771         }
1772
1773         if (type & PERF_SAMPLE_REGS_USER) {
1774                 OVERFLOW_CHECK_u64(array);
1775                 data->user_regs.abi = *array;
1776                 array++;
1777
1778                 if (data->user_regs.abi) {
1779                         u64 mask = evsel->attr.sample_regs_user;
1780
1781                         sz = hweight_long(mask) * sizeof(u64);
1782                         OVERFLOW_CHECK(array, sz, max_size);
1783                         data->user_regs.mask = mask;
1784                         data->user_regs.regs = (u64 *)array;
1785                         array = (void *)array + sz;
1786                 }
1787         }
1788
1789         if (type & PERF_SAMPLE_STACK_USER) {
1790                 OVERFLOW_CHECK_u64(array);
1791                 sz = *array++;
1792
1793                 data->user_stack.offset = ((char *)(array - 1)
1794                                           - (char *) event);
1795
1796                 if (!sz) {
1797                         data->user_stack.size = 0;
1798                 } else {
1799                         OVERFLOW_CHECK(array, sz, max_size);
1800                         data->user_stack.data = (char *)array;
1801                         array = (void *)array + sz;
1802                         OVERFLOW_CHECK_u64(array);
1803                         data->user_stack.size = *array++;
1804                         if (WARN_ONCE(data->user_stack.size > sz,
1805                                       "user stack dump failure\n"))
1806                                 return -EFAULT;
1807                 }
1808         }
1809
1810         data->weight = 0;
1811         if (type & PERF_SAMPLE_WEIGHT) {
1812                 OVERFLOW_CHECK_u64(array);
1813                 data->weight = *array;
1814                 array++;
1815         }
1816
1817         data->data_src = PERF_MEM_DATA_SRC_NONE;
1818         if (type & PERF_SAMPLE_DATA_SRC) {
1819                 OVERFLOW_CHECK_u64(array);
1820                 data->data_src = *array;
1821                 array++;
1822         }
1823
1824         data->transaction = 0;
1825         if (type & PERF_SAMPLE_TRANSACTION) {
1826                 OVERFLOW_CHECK_u64(array);
1827                 data->transaction = *array;
1828                 array++;
1829         }
1830
1831         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1832         if (type & PERF_SAMPLE_REGS_INTR) {
1833                 OVERFLOW_CHECK_u64(array);
1834                 data->intr_regs.abi = *array;
1835                 array++;
1836
1837                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1838                         u64 mask = evsel->attr.sample_regs_intr;
1839
1840                         sz = hweight_long(mask) * sizeof(u64);
1841                         OVERFLOW_CHECK(array, sz, max_size);
1842                         data->intr_regs.mask = mask;
1843                         data->intr_regs.regs = (u64 *)array;
1844                         array = (void *)array + sz;
1845                 }
1846         }
1847
1848         return 0;
1849 }
1850
1851 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1852                                      u64 read_format)
1853 {
1854         size_t sz, result = sizeof(struct sample_event);
1855
1856         if (type & PERF_SAMPLE_IDENTIFIER)
1857                 result += sizeof(u64);
1858
1859         if (type & PERF_SAMPLE_IP)
1860                 result += sizeof(u64);
1861
1862         if (type & PERF_SAMPLE_TID)
1863                 result += sizeof(u64);
1864
1865         if (type & PERF_SAMPLE_TIME)
1866                 result += sizeof(u64);
1867
1868         if (type & PERF_SAMPLE_ADDR)
1869                 result += sizeof(u64);
1870
1871         if (type & PERF_SAMPLE_ID)
1872                 result += sizeof(u64);
1873
1874         if (type & PERF_SAMPLE_STREAM_ID)
1875                 result += sizeof(u64);
1876
1877         if (type & PERF_SAMPLE_CPU)
1878                 result += sizeof(u64);
1879
1880         if (type & PERF_SAMPLE_PERIOD)
1881                 result += sizeof(u64);
1882
1883         if (type & PERF_SAMPLE_READ) {
1884                 result += sizeof(u64);
1885                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1886                         result += sizeof(u64);
1887                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1888                         result += sizeof(u64);
1889                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1890                 if (read_format & PERF_FORMAT_GROUP) {
1891                         sz = sample->read.group.nr *
1892                              sizeof(struct sample_read_value);
1893                         result += sz;
1894                 } else {
1895                         result += sizeof(u64);
1896                 }
1897         }
1898
1899         if (type & PERF_SAMPLE_CALLCHAIN) {
1900                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1901                 result += sz;
1902         }
1903
1904         if (type & PERF_SAMPLE_RAW) {
1905                 result += sizeof(u32);
1906                 result += sample->raw_size;
1907         }
1908
1909         if (type & PERF_SAMPLE_BRANCH_STACK) {
1910                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1911                 sz += sizeof(u64);
1912                 result += sz;
1913         }
1914
1915         if (type & PERF_SAMPLE_REGS_USER) {
1916                 if (sample->user_regs.abi) {
1917                         result += sizeof(u64);
1918                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1919                         result += sz;
1920                 } else {
1921                         result += sizeof(u64);
1922                 }
1923         }
1924
1925         if (type & PERF_SAMPLE_STACK_USER) {
1926                 sz = sample->user_stack.size;
1927                 result += sizeof(u64);
1928                 if (sz) {
1929                         result += sz;
1930                         result += sizeof(u64);
1931                 }
1932         }
1933
1934         if (type & PERF_SAMPLE_WEIGHT)
1935                 result += sizeof(u64);
1936
1937         if (type & PERF_SAMPLE_DATA_SRC)
1938                 result += sizeof(u64);
1939
1940         if (type & PERF_SAMPLE_TRANSACTION)
1941                 result += sizeof(u64);
1942
1943         if (type & PERF_SAMPLE_REGS_INTR) {
1944                 if (sample->intr_regs.abi) {
1945                         result += sizeof(u64);
1946                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1947                         result += sz;
1948                 } else {
1949                         result += sizeof(u64);
1950                 }
1951         }
1952
1953         return result;
1954 }
1955
1956 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1957                                   u64 read_format,
1958                                   const struct perf_sample *sample,
1959                                   bool swapped)
1960 {
1961         u64 *array;
1962         size_t sz;
1963         /*
1964          * used for cross-endian analysis. See git commit 65014ab3
1965          * for why this goofiness is needed.
1966          */
1967         union u64_swap u;
1968
1969         array = event->sample.array;
1970
1971         if (type & PERF_SAMPLE_IDENTIFIER) {
1972                 *array = sample->id;
1973                 array++;
1974         }
1975
1976         if (type & PERF_SAMPLE_IP) {
1977                 *array = sample->ip;
1978                 array++;
1979         }
1980
1981         if (type & PERF_SAMPLE_TID) {
1982                 u.val32[0] = sample->pid;
1983                 u.val32[1] = sample->tid;
1984                 if (swapped) {
1985                         /*
1986                          * Inverse of what is done in perf_evsel__parse_sample
1987                          */
1988                         u.val32[0] = bswap_32(u.val32[0]);
1989                         u.val32[1] = bswap_32(u.val32[1]);
1990                         u.val64 = bswap_64(u.val64);
1991                 }
1992
1993                 *array = u.val64;
1994                 array++;
1995         }
1996
1997         if (type & PERF_SAMPLE_TIME) {
1998                 *array = sample->time;
1999                 array++;
2000         }
2001
2002         if (type & PERF_SAMPLE_ADDR) {
2003                 *array = sample->addr;
2004                 array++;
2005         }
2006
2007         if (type & PERF_SAMPLE_ID) {
2008                 *array = sample->id;
2009                 array++;
2010         }
2011
2012         if (type & PERF_SAMPLE_STREAM_ID) {
2013                 *array = sample->stream_id;
2014                 array++;
2015         }
2016
2017         if (type & PERF_SAMPLE_CPU) {
2018                 u.val32[0] = sample->cpu;
2019                 if (swapped) {
2020                         /*
2021                          * Inverse of what is done in perf_evsel__parse_sample
2022                          */
2023                         u.val32[0] = bswap_32(u.val32[0]);
2024                         u.val64 = bswap_64(u.val64);
2025                 }
2026                 *array = u.val64;
2027                 array++;
2028         }
2029
2030         if (type & PERF_SAMPLE_PERIOD) {
2031                 *array = sample->period;
2032                 array++;
2033         }
2034
2035         if (type & PERF_SAMPLE_READ) {
2036                 if (read_format & PERF_FORMAT_GROUP)
2037                         *array = sample->read.group.nr;
2038                 else
2039                         *array = sample->read.one.value;
2040                 array++;
2041
2042                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2043                         *array = sample->read.time_enabled;
2044                         array++;
2045                 }
2046
2047                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2048                         *array = sample->read.time_running;
2049                         array++;
2050                 }
2051
2052                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2053                 if (read_format & PERF_FORMAT_GROUP) {
2054                         sz = sample->read.group.nr *
2055                              sizeof(struct sample_read_value);
2056                         memcpy(array, sample->read.group.values, sz);
2057                         array = (void *)array + sz;
2058                 } else {
2059                         *array = sample->read.one.id;
2060                         array++;
2061                 }
2062         }
2063
2064         if (type & PERF_SAMPLE_CALLCHAIN) {
2065                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2066                 memcpy(array, sample->callchain, sz);
2067                 array = (void *)array + sz;
2068         }
2069
2070         if (type & PERF_SAMPLE_RAW) {
2071                 u.val32[0] = sample->raw_size;
2072                 if (WARN_ONCE(swapped,
2073                               "Endianness of raw data not corrected!\n")) {
2074                         /*
2075                          * Inverse of what is done in perf_evsel__parse_sample
2076                          */
2077                         u.val32[0] = bswap_32(u.val32[0]);
2078                         u.val32[1] = bswap_32(u.val32[1]);
2079                         u.val64 = bswap_64(u.val64);
2080                 }
2081                 *array = u.val64;
2082                 array = (void *)array + sizeof(u32);
2083
2084                 memcpy(array, sample->raw_data, sample->raw_size);
2085                 array = (void *)array + sample->raw_size;
2086         }
2087
2088         if (type & PERF_SAMPLE_BRANCH_STACK) {
2089                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2090                 sz += sizeof(u64);
2091                 memcpy(array, sample->branch_stack, sz);
2092                 array = (void *)array + sz;
2093         }
2094
2095         if (type & PERF_SAMPLE_REGS_USER) {
2096                 if (sample->user_regs.abi) {
2097                         *array++ = sample->user_regs.abi;
2098                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2099                         memcpy(array, sample->user_regs.regs, sz);
2100                         array = (void *)array + sz;
2101                 } else {
2102                         *array++ = 0;
2103                 }
2104         }
2105
2106         if (type & PERF_SAMPLE_STACK_USER) {
2107                 sz = sample->user_stack.size;
2108                 *array++ = sz;
2109                 if (sz) {
2110                         memcpy(array, sample->user_stack.data, sz);
2111                         array = (void *)array + sz;
2112                         *array++ = sz;
2113                 }
2114         }
2115
2116         if (type & PERF_SAMPLE_WEIGHT) {
2117                 *array = sample->weight;
2118                 array++;
2119         }
2120
2121         if (type & PERF_SAMPLE_DATA_SRC) {
2122                 *array = sample->data_src;
2123                 array++;
2124         }
2125
2126         if (type & PERF_SAMPLE_TRANSACTION) {
2127                 *array = sample->transaction;
2128                 array++;
2129         }
2130
2131         if (type & PERF_SAMPLE_REGS_INTR) {
2132                 if (sample->intr_regs.abi) {
2133                         *array++ = sample->intr_regs.abi;
2134                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2135                         memcpy(array, sample->intr_regs.regs, sz);
2136                         array = (void *)array + sz;
2137                 } else {
2138                         *array++ = 0;
2139                 }
2140         }
2141
2142         return 0;
2143 }
2144
2145 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2146 {
2147         return pevent_find_field(evsel->tp_format, name);
2148 }
2149
2150 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2151                          const char *name)
2152 {
2153         struct format_field *field = perf_evsel__field(evsel, name);
2154         int offset;
2155
2156         if (!field)
2157                 return NULL;
2158
2159         offset = field->offset;
2160
2161         if (field->flags & FIELD_IS_DYNAMIC) {
2162                 offset = *(int *)(sample->raw_data + field->offset);
2163                 offset &= 0xffff;
2164         }
2165
2166         return sample->raw_data + offset;
2167 }
2168
2169 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2170                        const char *name)
2171 {
2172         struct format_field *field = perf_evsel__field(evsel, name);
2173         void *ptr;
2174         u64 value;
2175
2176         if (!field)
2177                 return 0;
2178
2179         ptr = sample->raw_data + field->offset;
2180
2181         switch (field->size) {
2182         case 1:
2183                 return *(u8 *)ptr;
2184         case 2:
2185                 value = *(u16 *)ptr;
2186                 break;
2187         case 4:
2188                 value = *(u32 *)ptr;
2189                 break;
2190         case 8:
2191                 memcpy(&value, ptr, sizeof(u64));
2192                 break;
2193         default:
2194                 return 0;
2195         }
2196
2197         if (!evsel->needs_swap)
2198                 return value;
2199
2200         switch (field->size) {
2201         case 2:
2202                 return bswap_16(value);
2203         case 4:
2204                 return bswap_32(value);
2205         case 8:
2206                 return bswap_64(value);
2207         default:
2208                 return 0;
2209         }
2210
2211         return 0;
2212 }
2213
2214 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2215 {
2216         va_list args;
2217         int ret = 0;
2218
2219         if (!*first) {
2220                 ret += fprintf(fp, ",");
2221         } else {
2222                 ret += fprintf(fp, ":");
2223                 *first = false;
2224         }
2225
2226         va_start(args, fmt);
2227         ret += vfprintf(fp, fmt, args);
2228         va_end(args);
2229         return ret;
2230 }
2231
2232 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2233 {
2234         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2235 }
2236
2237 int perf_evsel__fprintf(struct perf_evsel *evsel,
2238                         struct perf_attr_details *details, FILE *fp)
2239 {
2240         bool first = true;
2241         int printed = 0;
2242
2243         if (details->event_group) {
2244                 struct perf_evsel *pos;
2245
2246                 if (!perf_evsel__is_group_leader(evsel))
2247                         return 0;
2248
2249                 if (evsel->nr_members > 1)
2250                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2251
2252                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2253                 for_each_group_member(pos, evsel)
2254                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2255
2256                 if (evsel->nr_members > 1)
2257                         printed += fprintf(fp, "}");
2258                 goto out;
2259         }
2260
2261         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2262
2263         if (details->verbose) {
2264                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2265                                                     __print_attr__fprintf, &first);
2266         } else if (details->freq) {
2267                 const char *term = "sample_freq";
2268
2269                 if (!evsel->attr.freq)
2270                         term = "sample_period";
2271
2272                 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2273                                          term, (u64)evsel->attr.sample_freq);
2274         }
2275 out:
2276         fputc('\n', fp);
2277         return ++printed;
2278 }
2279
2280 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2281                           char *msg, size_t msgsize)
2282 {
2283         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2284             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2285             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2286                 /*
2287                  * If it's cycles then fall back to hrtimer based
2288                  * cpu-clock-tick sw counter, which is always available even if
2289                  * no PMU support.
2290                  *
2291                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2292                  * b0a873e).
2293                  */
2294                 scnprintf(msg, msgsize, "%s",
2295 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2296
2297                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2298                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2299
2300                 zfree(&evsel->name);
2301                 return true;
2302         }
2303
2304         return false;
2305 }
2306
2307 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2308                               int err, char *msg, size_t size)
2309 {
2310         char sbuf[STRERR_BUFSIZE];
2311
2312         switch (err) {
2313         case EPERM:
2314         case EACCES:
2315                 return scnprintf(msg, size,
2316                  "You may not have permission to collect %sstats.\n"
2317                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2318                  " -1 - Not paranoid at all\n"
2319                  "  0 - Disallow raw tracepoint access for unpriv\n"
2320                  "  1 - Disallow cpu events for unpriv\n"
2321                  "  2 - Disallow kernel profiling for unpriv",
2322                                  target->system_wide ? "system-wide " : "");
2323         case ENOENT:
2324                 return scnprintf(msg, size, "The %s event is not supported.",
2325                                  perf_evsel__name(evsel));
2326         case EMFILE:
2327                 return scnprintf(msg, size, "%s",
2328                          "Too many events are opened.\n"
2329                          "Probably the maximum number of open file descriptors has been reached.\n"
2330                          "Hint: Try again after reducing the number of events.\n"
2331                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2332         case ENODEV:
2333                 if (target->cpu_list)
2334                         return scnprintf(msg, size, "%s",
2335          "No such device - did you specify an out-of-range profile CPU?\n");
2336                 break;
2337         case EOPNOTSUPP:
2338                 if (evsel->attr.precise_ip)
2339                         return scnprintf(msg, size, "%s",
2340         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2341 #if defined(__i386__) || defined(__x86_64__)
2342                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2343                         return scnprintf(msg, size, "%s",
2344         "No hardware sampling interrupt available.\n"
2345         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2346 #endif
2347                 break;
2348         case EBUSY:
2349                 if (find_process("oprofiled"))
2350                         return scnprintf(msg, size,
2351         "The PMU counters are busy/taken by another profiler.\n"
2352         "We found oprofile daemon running, please stop it and try again.");
2353                 break;
2354         case EINVAL:
2355                 if (perf_missing_features.clockid)
2356                         return scnprintf(msg, size, "clockid feature not supported.");
2357                 if (perf_missing_features.clockid_wrong)
2358                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2359                 break;
2360         default:
2361                 break;
2362         }
2363
2364         return scnprintf(msg, size,
2365         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2366         "/bin/dmesg may provide additional information.\n"
2367         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2368                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2369                          perf_evsel__name(evsel));
2370 }