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/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 static struct {
  45         bool sample_id_all;
  46         bool exclude_guest;
  47         bool mmap2;
  48         bool cloexec;
  49         bool clockid;
  50         bool clockid_wrong;
  51         bool lbr_flags;
  52         bool write_backward;
  53         bool group_read;
  54 } perf_missing_features;
  55
  56 static clockid_t clockid;
  57
  58 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  59 {
  60         return 0;
  61 }
  62
  63 void __weak test_attr__ready(void) { }
  64
  65 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  66 {
  67 }
  68
  69 static struct {
  70         size_t  size;
  71         int     (*init)(struct perf_evsel *evsel);
  72         void    (*fini)(struct perf_evsel *evsel);
  73 } perf_evsel__object = {
  74         .size = sizeof(struct perf_evsel),
  75         .init = perf_evsel__no_extra_init,
  76         .fini = perf_evsel__no_extra_fini,
  77 };
  78
  79 int perf_evsel__object_config(size_t object_size,
  80                               int (*init)(struct perf_evsel *evsel),
  81                               void (*fini)(struct perf_evsel *evsel))
  82 {
  83
  84         if (object_size == 0)
  85                 goto set_methods;
  86
  87         if (perf_evsel__object.size > object_size)
  88                 return -EINVAL;
  89
  90         perf_evsel__object.size = object_size;
  91
  92 set_methods:
  93         if (init != NULL)
  94                 perf_evsel__object.init = init;
  95
  96         if (fini != NULL)
  97                 perf_evsel__object.fini = fini;
  98
  99         return 0;
 100 }
 101
 102 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 103
 104 int __perf_evsel__sample_size(u64 sample_type)
 105 {
 106         u64 mask = sample_type & PERF_SAMPLE_MASK;
 107         int size = 0;
 108         int i;
 109
 110         for (i = 0; i < 64; i++) {
 111                 if (mask & (1ULL << i))
 112                         size++;
 113         }
 114
 115         size *= sizeof(u64);
 116
 117         return size;
 118 }
 119
 120 /**
 121  * __perf_evsel__calc_id_pos - calculate id_pos.
 122  * @sample_type: sample type
 123  *
 124  * This function returns the position of the event id (PERF_SAMPLE_ID or
 125  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 126  * sample_event.
 127  */
 128 static int __perf_evsel__calc_id_pos(u64 sample_type)
 129 {
 130         int idx = 0;
 131
 132         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 133                 return 0;
 134
 135         if (!(sample_type & PERF_SAMPLE_ID))
 136                 return -1;
 137
 138         if (sample_type & PERF_SAMPLE_IP)
 139                 idx += 1;
 140
 141         if (sample_type & PERF_SAMPLE_TID)
 142                 idx += 1;
 143
 144         if (sample_type & PERF_SAMPLE_TIME)
 145                 idx += 1;
 146
 147         if (sample_type & PERF_SAMPLE_ADDR)
 148                 idx += 1;
 149
 150         return idx;
 151 }
 152
 153 /**
 154  * __perf_evsel__calc_is_pos - calculate is_pos.
 155  * @sample_type: sample type
 156  *
 157  * This function returns the position (counting backwards) of the event id
 158  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 159  * sample_id_all is used there is an id sample appended to non-sample events.
 160  */
 161 static int __perf_evsel__calc_is_pos(u64 sample_type)
 162 {
 163         int idx = 1;
 164
 165         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 166                 return 1;
 167
 168         if (!(sample_type & PERF_SAMPLE_ID))
 169                 return -1;
 170
 171         if (sample_type & PERF_SAMPLE_CPU)
 172                 idx += 1;
 173
 174         if (sample_type & PERF_SAMPLE_STREAM_ID)
 175                 idx += 1;
 176
 177         return idx;
 178 }
 179
 180 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 181 {
 182         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 183         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 184 }
 185
 186 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 187                                   enum perf_event_sample_format bit)
 188 {
 189         if (!(evsel->attr.sample_type & bit)) {
 190                 evsel->attr.sample_type |= bit;
 191                 evsel->sample_size += sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 197                                     enum perf_event_sample_format bit)
 198 {
 199         if (evsel->attr.sample_type & bit) {
 200                 evsel->attr.sample_type &= ~bit;
 201                 evsel->sample_size -= sizeof(u64);
 202                 perf_evsel__calc_id_pos(evsel);
 203         }
 204 }
 205
 206 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 207                                bool can_sample_identifier)
 208 {
 209         if (can_sample_identifier) {
 210                 perf_evsel__reset_sample_bit(evsel, ID);
 211                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 212         } else {
 213                 perf_evsel__set_sample_bit(evsel, ID);
 214         }
 215         evsel->attr.read_format |= PERF_FORMAT_ID;
 216 }
 217
 218 /**
 219  * perf_evsel__is_function_event - Return whether given evsel is a function
 220  * trace event
 221  *
 222  * @evsel - evsel selector to be tested
 223  *
 224  * Return %true if event is function trace event
 225  */
 226 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 227 {
 228 #define FUNCTION_EVENT "ftrace:function"
 229
 230         return evsel->name &&
 231                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 232
 233 #undef FUNCTION_EVENT
 234 }
 235
 236 void perf_evsel__init(struct perf_evsel *evsel,
 237                       struct perf_event_attr *attr, int idx)
 238 {
 239         evsel->idx         = idx;
 240         evsel->tracking    = !idx;
 241         evsel->attr        = *attr;
 242         evsel->leader      = evsel;
 243         evsel->unit        = "";
 244         evsel->scale       = 1.0;
 245         evsel->evlist      = NULL;
 246         evsel->bpf_fd      = -1;
 247         INIT_LIST_HEAD(&evsel->node);
 248         INIT_LIST_HEAD(&evsel->config_terms);
 249         perf_evsel__object.init(evsel);
 250         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 251         perf_evsel__calc_id_pos(evsel);
 252         evsel->cmdline_group_boundary = false;
 253         evsel->metric_expr   = NULL;
 254         evsel->metric_name   = NULL;
 255         evsel->metric_events = NULL;
 256         evsel->collect_stat  = false;
 257 }
 258
 259 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 260 {
 261         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 262
 263         if (evsel != NULL)
 264                 perf_evsel__init(evsel, attr, idx);
 265
 266         if (perf_evsel__is_bpf_output(evsel)) {
 267                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 268                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 269                 evsel->attr.sample_period = 1;
 270         }
 271
 272         return evsel;
 273 }
 274
 275 static bool perf_event_can_profile_kernel(void)
 276 {
 277         return geteuid() == 0 || perf_event_paranoid() == -1;
 278 }
 279
 280 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 281 {
 282         struct perf_event_attr attr = {
 283                 .type   = PERF_TYPE_HARDWARE,
 284                 .config = PERF_COUNT_HW_CPU_CYCLES,
 285                 .exclude_kernel = !perf_event_can_profile_kernel(),
 286         };
 287         struct perf_evsel *evsel;
 288
 289         event_attr_init(&attr);
 290
 291         if (!precise)
 292                 goto new_event;
 293         /*
 294          * Unnamed union member, not supported as struct member named
 295          * initializer in older compilers such as gcc 4.4.7
 296          *
 297          * Just for probing the precise_ip:
 298          */
 299         attr.sample_period = 1;
 300
 301         perf_event_attr__set_max_precise_ip(&attr);
 302         /*
 303          * Now let the usual logic to set up the perf_event_attr defaults
 304          * to kick in when we return and before perf_evsel__open() is called.
 305          */
 306         attr.sample_period = 0;
 307 new_event:
 308         evsel = perf_evsel__new(&attr);
 309         if (evsel == NULL)
 310                 goto out;
 311
 312         /* use asprintf() because free(evsel) assumes name is allocated */
 313         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 314                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 315                      attr.exclude_kernel ? "u" : "",
 316                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 317                 goto error_free;
 318 out:
 319         return evsel;
 320 error_free:
 321         perf_evsel__delete(evsel);
 322         evsel = NULL;
 323         goto out;
 324 }
 325
 326 /*
 327  * Returns pointer with encoded error via <linux/err.h> interface.
 328  */
 329 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 330 {
 331         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 332         int err = -ENOMEM;
 333
 334         if (evsel == NULL) {
 335                 goto out_err;
 336         } else {
 337                 struct perf_event_attr attr = {
 338                         .type          = PERF_TYPE_TRACEPOINT,
 339                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 340                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 341                 };
 342
 343                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 344                         goto out_free;
 345
 346                 evsel->tp_format = trace_event__tp_format(sys, name);
 347                 if (IS_ERR(evsel->tp_format)) {
 348                         err = PTR_ERR(evsel->tp_format);
 349                         goto out_free;
 350                 }
 351
 352                 event_attr_init(&attr);
 353                 attr.config = evsel->tp_format->id;
 354                 attr.sample_period = 1;
 355                 perf_evsel__init(evsel, &attr, idx);
 356         }
 357
 358         return evsel;
 359
 360 out_free:
 361         zfree(&evsel->name);
 362         free(evsel);
 363 out_err:
 364         return ERR_PTR(err);
 365 }
 366
 367 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 368         "cycles",
 369         "instructions",
 370         "cache-references",
 371         "cache-misses",
 372         "branches",
 373         "branch-misses",
 374         "bus-cycles",
 375         "stalled-cycles-frontend",
 376         "stalled-cycles-backend",
 377         "ref-cycles",
 378 };
 379
 380 static const char *__perf_evsel__hw_name(u64 config)
 381 {
 382         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 383                 return perf_evsel__hw_names[config];
 384
 385         return "unknown-hardware";
 386 }
 387
 388 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 389 {
 390         int colon = 0, r = 0;
 391         struct perf_event_attr *attr = &evsel->attr;
 392         bool exclude_guest_default = false;
 393
 394 #define MOD_PRINT(context, mod) do {                                    \
 395                 if (!attr->exclude_##context) {                         \
 396                         if (!colon) colon = ++r;                        \
 397                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 398                 } } while(0)
 399
 400         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 401                 MOD_PRINT(kernel, 'k');
 402                 MOD_PRINT(user, 'u');
 403                 MOD_PRINT(hv, 'h');
 404                 exclude_guest_default = true;
 405         }
 406
 407         if (attr->precise_ip) {
 408                 if (!colon)
 409                         colon = ++r;
 410                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 411                 exclude_guest_default = true;
 412         }
 413
 414         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 415                 MOD_PRINT(host, 'H');
 416                 MOD_PRINT(guest, 'G');
 417         }
 418 #undef MOD_PRINT
 419         if (colon)
 420                 bf[colon - 1] = ':';
 421         return r;
 422 }
 423
 424 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 425 {
 426         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 427         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 428 }
 429
 430 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 431         "cpu-clock",
 432         "task-clock",
 433         "page-faults",
 434         "context-switches",
 435         "cpu-migrations",
 436         "minor-faults",
 437         "major-faults",
 438         "alignment-faults",
 439         "emulation-faults",
 440         "dummy",
 441 };
 442
 443 static const char *__perf_evsel__sw_name(u64 config)
 444 {
 445         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 446                 return perf_evsel__sw_names[config];
 447         return "unknown-software";
 448 }
 449
 450 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 451 {
 452         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 453         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 454 }
 455
 456 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 457 {
 458         int r;
 459
 460         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 461
 462         if (type & HW_BREAKPOINT_R)
 463                 r += scnprintf(bf + r, size - r, "r");
 464
 465         if (type & HW_BREAKPOINT_W)
 466                 r += scnprintf(bf + r, size - r, "w");
 467
 468         if (type & HW_BREAKPOINT_X)
 469                 r += scnprintf(bf + r, size - r, "x");
 470
 471         return r;
 472 }
 473
 474 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 475 {
 476         struct perf_event_attr *attr = &evsel->attr;
 477         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 478         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 479 }
 480
 481 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 482                                 [PERF_EVSEL__MAX_ALIASES] = {
 483  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 484  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 485  { "LLC",       "L2",                                                   },
 486  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 487  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 488  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 489  { "node",                                                              },
 490 };
 491
 492 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 493                                    [PERF_EVSEL__MAX_ALIASES] = {
 494  { "load",      "loads",        "read",                                 },
 495  { "store",     "stores",       "write",                                },
 496  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 497 };
 498
 499 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 500                                        [PERF_EVSEL__MAX_ALIASES] = {
 501  { "refs",      "Reference",    "ops",          "access",               },
 502  { "misses",    "miss",                                                 },
 503 };
 504
 505 #define C(x)            PERF_COUNT_HW_CACHE_##x
 506 #define CACHE_READ      (1 << C(OP_READ))
 507 #define CACHE_WRITE     (1 << C(OP_WRITE))
 508 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 509 #define COP(x)          (1 << x)
 510
 511 /*
 512  * cache operartion stat
 513  * L1I : Read and prefetch only
 514  * ITLB and BPU : Read-only
 515  */
 516 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 517  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 518  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 519  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 520  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 521  [C(ITLB)]      = (CACHE_READ),
 522  [C(BPU)]       = (CACHE_READ),
 523  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524 };
 525
 526 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 527 {
 528         if (perf_evsel__hw_cache_stat[type] & COP(op))
 529                 return true;    /* valid */
 530         else
 531                 return false;   /* invalid */
 532 }
 533
 534 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 535                                             char *bf, size_t size)
 536 {
 537         if (result) {
 538                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 539                                  perf_evsel__hw_cache_op[op][0],
 540                                  perf_evsel__hw_cache_result[result][0]);
 541         }
 542
 543         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 544                          perf_evsel__hw_cache_op[op][1]);
 545 }
 546
 547 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 548 {
 549         u8 op, result, type = (config >>  0) & 0xff;
 550         const char *err = "unknown-ext-hardware-cache-type";
 551
 552         if (type >= PERF_COUNT_HW_CACHE_MAX)
 553                 goto out_err;
 554
 555         op = (config >>  8) & 0xff;
 556         err = "unknown-ext-hardware-cache-op";
 557         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 558                 goto out_err;
 559
 560         result = (config >> 16) & 0xff;
 561         err = "unknown-ext-hardware-cache-result";
 562         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 563                 goto out_err;
 564
 565         err = "invalid-cache";
 566         if (!perf_evsel__is_cache_op_valid(type, op))
 567                 goto out_err;
 568
 569         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 570 out_err:
 571         return scnprintf(bf, size, "%s", err);
 572 }
 573
 574 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 575 {
 576         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 577         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 578 }
 579
 580 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 581 {
 582         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 583         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 584 }
 585
 586 const char *perf_evsel__name(struct perf_evsel *evsel)
 587 {
 588         char bf[128];
 589
 590         if (evsel->name)
 591                 return evsel->name;
 592
 593         switch (evsel->attr.type) {
 594         case PERF_TYPE_RAW:
 595                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 596                 break;
 597
 598         case PERF_TYPE_HARDWARE:
 599                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 600                 break;
 601
 602         case PERF_TYPE_HW_CACHE:
 603                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 604                 break;
 605
 606         case PERF_TYPE_SOFTWARE:
 607                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 608                 break;
 609
 610         case PERF_TYPE_TRACEPOINT:
 611                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 612                 break;
 613
 614         case PERF_TYPE_BREAKPOINT:
 615                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 616                 break;
 617
 618         default:
 619                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 620                           evsel->attr.type);
 621                 break;
 622         }
 623
 624         evsel->name = strdup(bf);
 625
 626         return evsel->name ?: "unknown";
 627 }
 628
 629 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 630 {
 631         return evsel->group_name ?: "anon group";
 632 }
 633
 634 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 635 {
 636         int ret;
 637         struct perf_evsel *pos;
 638         const char *group_name = perf_evsel__group_name(evsel);
 639
 640         ret = scnprintf(buf, size, "%s", group_name);
 641
 642         ret += scnprintf(buf + ret, size - ret, " { %s",
 643                          perf_evsel__name(evsel));
 644
 645         for_each_group_member(pos, evsel)
 646                 ret += scnprintf(buf + ret, size - ret, ", %s",
 647                                  perf_evsel__name(pos));
 648
 649         ret += scnprintf(buf + ret, size - ret, " }");
 650
 651         return ret;
 652 }
 653
 654 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 655                                            struct record_opts *opts,
 656                                            struct callchain_param *param)
 657 {
 658         bool function = perf_evsel__is_function_event(evsel);
 659         struct perf_event_attr *attr = &evsel->attr;
 660
 661         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 662
 663         attr->sample_max_stack = param->max_stack;
 664
 665         if (param->record_mode == CALLCHAIN_LBR) {
 666                 if (!opts->branch_stack) {
 667                         if (attr->exclude_user) {
 668                                 pr_warning("LBR callstack option is only available "
 669                                            "to get user callchain information. "
 670                                            "Falling back to framepointers.\n");
 671                         } else {
 672                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 673                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 674                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 675                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 676                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 677                         }
 678                 } else
 679                          pr_warning("Cannot use LBR callstack with branch stack. "
 680                                     "Falling back to framepointers.\n");
 681         }
 682
 683         if (param->record_mode == CALLCHAIN_DWARF) {
 684                 if (!function) {
 685                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 686                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 687                         attr->sample_regs_user |= PERF_REGS_MASK;
 688                         attr->sample_stack_user = param->dump_size;
 689                         attr->exclude_callchain_user = 1;
 690                 } else {
 691                         pr_info("Cannot use DWARF unwind for function trace event,"
 692                                 " falling back to framepointers.\n");
 693                 }
 694         }
 695
 696         if (function) {
 697                 pr_info("Disabling user space callchains for function trace event.\n");
 698                 attr->exclude_callchain_user = 1;
 699         }
 700 }
 701
 702 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 703                                   struct record_opts *opts,
 704                                   struct callchain_param *param)
 705 {
 706         if (param->enabled)
 707                 return __perf_evsel__config_callchain(evsel, opts, param);
 708 }
 709
 710 static void
 711 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 712                             struct callchain_param *param)
 713 {
 714         struct perf_event_attr *attr = &evsel->attr;
 715
 716         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 717         if (param->record_mode == CALLCHAIN_LBR) {
 718                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 719                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 720                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 721         }
 722         if (param->record_mode == CALLCHAIN_DWARF) {
 723                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 724                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 725         }
 726 }
 727
 728 static void apply_config_terms(struct perf_evsel *evsel,
 729                                struct record_opts *opts, bool track)
 730 {
 731         struct perf_evsel_config_term *term;
 732         struct list_head *config_terms = &evsel->config_terms;
 733         struct perf_event_attr *attr = &evsel->attr;
 734         /* callgraph default */
 735         struct callchain_param param = {
 736                 .record_mode = callchain_param.record_mode,
 737         };
 738         u32 dump_size = 0;
 739         int max_stack = 0;
 740         const char *callgraph_buf = NULL;
 741
 742         list_for_each_entry(term, config_terms, list) {
 743                 switch (term->type) {
 744                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 745                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 746                                 attr->sample_period = term->val.period;
 747                                 attr->freq = 0;
 748                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 749                         }
 750                         break;
 751                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 752                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 753                                 attr->sample_freq = term->val.freq;
 754                                 attr->freq = 1;
 755                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 756                         }
 757                         break;
 758                 case PERF_EVSEL__CONFIG_TERM_TIME:
 759                         if (term->val.time)
 760                                 perf_evsel__set_sample_bit(evsel, TIME);
 761                         else
 762                                 perf_evsel__reset_sample_bit(evsel, TIME);
 763                         break;
 764                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 765                         callgraph_buf = term->val.callgraph;
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 768                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 769                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 770                                 parse_branch_str(term->val.branch,
 771                                                  &attr->branch_sample_type);
 772                         } else
 773                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 774                         break;
 775                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 776                         dump_size = term->val.stack_user;
 777                         break;
 778                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 779                         max_stack = term->val.max_stack;
 780                         break;
 781                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 782                         /*
 783                          * attr->inherit should has already been set by
 784                          * perf_evsel__config. If user explicitly set
 785                          * inherit using config terms, override global
 786                          * opt->no_inherit setting.
 787                          */
 788                         attr->inherit = term->val.inherit ? 1 : 0;
 789                         break;
 790                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 791                         attr->write_backward = term->val.overwrite ? 1 : 0;
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 794                         break;
 795                 default:
 796                         break;
 797                 }
 798         }
 799
 800         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 801         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 802                 bool sample_address = false;
 803
 804                 if (max_stack) {
 805                         param.max_stack = max_stack;
 806                         if (callgraph_buf == NULL)
 807                                 callgraph_buf = "fp";
 808                 }
 809
 810                 /* parse callgraph parameters */
 811                 if (callgraph_buf != NULL) {
 812                         if (!strcmp(callgraph_buf, "no")) {
 813                                 param.enabled = false;
 814                                 param.record_mode = CALLCHAIN_NONE;
 815                         } else {
 816                                 param.enabled = true;
 817                                 if (parse_callchain_record(callgraph_buf, &param)) {
 818                                         pr_err("per-event callgraph setting for %s failed. "
 819                                                "Apply callgraph global setting for it\n",
 820                                                evsel->name);
 821                                         return;
 822                                 }
 823                                 if (param.record_mode == CALLCHAIN_DWARF)
 824                                         sample_address = true;
 825                         }
 826                 }
 827                 if (dump_size > 0) {
 828                         dump_size = round_up(dump_size, sizeof(u64));
 829                         param.dump_size = dump_size;
 830                 }
 831
 832                 /* If global callgraph set, clear it */
 833                 if (callchain_param.enabled)
 834                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 835
 836                 /* set perf-event callgraph */
 837                 if (param.enabled) {
 838                         if (sample_address) {
 839                                 perf_evsel__set_sample_bit(evsel, ADDR);
 840                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 841                                 evsel->attr.mmap_data = track;
 842                         }
 843                         perf_evsel__config_callchain(evsel, opts, &param);
 844                 }
 845         }
 846 }
 847
 848 /*
 849  * The enable_on_exec/disabled value strategy:
 850  *
 851  *  1) For any type of traced program:
 852  *    - all independent events and group leaders are disabled
 853  *    - all group members are enabled
 854  *
 855  *     Group members are ruled by group leaders. They need to
 856  *     be enabled, because the group scheduling relies on that.
 857  *
 858  *  2) For traced programs executed by perf:
 859  *     - all independent events and group leaders have
 860  *       enable_on_exec set
 861  *     - we don't specifically enable or disable any event during
 862  *       the record command
 863  *
 864  *     Independent events and group leaders are initially disabled
 865  *     and get enabled by exec. Group members are ruled by group
 866  *     leaders as stated in 1).
 867  *
 868  *  3) For traced programs attached by perf (pid/tid):
 869  *     - we specifically enable or disable all events during
 870  *       the record command
 871  *
 872  *     When attaching events to already running traced we
 873  *     enable/disable events specifically, as there's no
 874  *     initial traced exec call.
 875  */
 876 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 877                         struct callchain_param *callchain)
 878 {
 879         struct perf_evsel *leader = evsel->leader;
 880         struct perf_event_attr *attr = &evsel->attr;
 881         int track = evsel->tracking;
 882         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 883
 884         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 885         attr->inherit       = !opts->no_inherit;
 886         attr->write_backward = opts->overwrite ? 1 : 0;
 887
 888         perf_evsel__set_sample_bit(evsel, IP);
 889         perf_evsel__set_sample_bit(evsel, TID);
 890
 891         if (evsel->sample_read) {
 892                 perf_evsel__set_sample_bit(evsel, READ);
 893
 894                 /*
 895                  * We need ID even in case of single event, because
 896                  * PERF_SAMPLE_READ process ID specific data.
 897                  */
 898                 perf_evsel__set_sample_id(evsel, false);
 899
 900                 /*
 901                  * Apply group format only if we belong to group
 902                  * with more than one members.
 903                  */
 904                 if (leader->nr_members > 1) {
 905                         attr->read_format |= PERF_FORMAT_GROUP;
 906                         attr->inherit = 0;
 907                 }
 908         }
 909
 910         /*
 911          * We default some events to have a default interval. But keep
 912          * it a weak assumption overridable by the user.
 913          */
 914         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 915                                      opts->user_interval != ULLONG_MAX)) {
 916                 if (opts->freq) {
 917                         perf_evsel__set_sample_bit(evsel, PERIOD);
 918                         attr->freq              = 1;
 919                         attr->sample_freq       = opts->freq;
 920                 } else {
 921                         attr->sample_period = opts->default_interval;
 922                 }
 923         }
 924
 925         /*
 926          * Disable sampling for all group members other
 927          * than leader in case leader 'leads' the sampling.
 928          */
 929         if ((leader != evsel) && leader->sample_read) {
 930                 attr->sample_freq   = 0;
 931                 attr->sample_period = 0;
 932         }
 933
 934         if (opts->no_samples)
 935                 attr->sample_freq = 0;
 936
 937         if (opts->inherit_stat) {
 938                 evsel->attr.read_format |=
 939                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 940                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 941                         PERF_FORMAT_ID;
 942                 attr->inherit_stat = 1;
 943         }
 944
 945         if (opts->sample_address) {
 946                 perf_evsel__set_sample_bit(evsel, ADDR);
 947                 attr->mmap_data = track;
 948         }
 949
 950         /*
 951          * We don't allow user space callchains for  function trace
 952          * event, due to issues with page faults while tracing page
 953          * fault handler and its overall trickiness nature.
 954          */
 955         if (perf_evsel__is_function_event(evsel))
 956                 evsel->attr.exclude_callchain_user = 1;
 957
 958         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 959                 perf_evsel__config_callchain(evsel, opts, callchain);
 960
 961         if (opts->sample_intr_regs) {
 962                 attr->sample_regs_intr = opts->sample_intr_regs;
 963                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 964         }
 965
 966         if (opts->sample_user_regs) {
 967                 attr->sample_regs_user |= opts->sample_user_regs;
 968                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 969         }
 970
 971         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 972                 perf_evsel__set_sample_bit(evsel, CPU);
 973
 974         /*
 975          * When the user explicitly disabled time don't force it here.
 976          */
 977         if (opts->sample_time &&
 978             (!perf_missing_features.sample_id_all &&
 979             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 980              opts->sample_time_set)))
 981                 perf_evsel__set_sample_bit(evsel, TIME);
 982
 983         if (opts->raw_samples && !evsel->no_aux_samples) {
 984                 perf_evsel__set_sample_bit(evsel, TIME);
 985                 perf_evsel__set_sample_bit(evsel, RAW);
 986                 perf_evsel__set_sample_bit(evsel, CPU);
 987         }
 988
 989         if (opts->sample_address)
 990                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 991
 992         if (opts->sample_phys_addr)
 993                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
 994
 995         if (opts->no_buffering) {
 996                 attr->watermark = 0;
 997                 attr->wakeup_events = 1;
 998         }
 999         if (opts->branch_stack && !evsel->no_aux_samples) {
1000                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1001                 attr->branch_sample_type = opts->branch_stack;
1002         }
1003
1004         if (opts->sample_weight)
1005                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1006
1007         attr->task  = track;
1008         attr->mmap  = track;
1009         attr->mmap2 = track && !perf_missing_features.mmap2;
1010         attr->comm  = track;
1011
1012         if (opts->record_namespaces)
1013                 attr->namespaces  = track;
1014
1015         if (opts->record_switch_events)
1016                 attr->context_switch = track;
1017
1018         if (opts->sample_transaction)
1019                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1020
1021         if (opts->running_time) {
1022                 evsel->attr.read_format |=
1023                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1024                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1025         }
1026
1027         /*
1028          * XXX see the function comment above
1029          *
1030          * Disabling only independent events or group leaders,
1031          * keeping group members enabled.
1032          */
1033         if (perf_evsel__is_group_leader(evsel))
1034                 attr->disabled = 1;
1035
1036         /*
1037          * Setting enable_on_exec for independent events and
1038          * group leaders for traced executed by perf.
1039          */
1040         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1041                 !opts->initial_delay)
1042                 attr->enable_on_exec = 1;
1043
1044         if (evsel->immediate) {
1045                 attr->disabled = 0;
1046                 attr->enable_on_exec = 0;
1047         }
1048
1049         clockid = opts->clockid;
1050         if (opts->use_clockid) {
1051                 attr->use_clockid = 1;
1052                 attr->clockid = opts->clockid;
1053         }
1054
1055         if (evsel->precise_max)
1056                 perf_event_attr__set_max_precise_ip(attr);
1057
1058         if (opts->all_user) {
1059                 attr->exclude_kernel = 1;
1060                 attr->exclude_user   = 0;
1061         }
1062
1063         if (opts->all_kernel) {
1064                 attr->exclude_kernel = 0;
1065                 attr->exclude_user   = 1;
1066         }
1067
1068         /*
1069          * Apply event specific term settings,
1070          * it overloads any global configuration.
1071          */
1072         apply_config_terms(evsel, opts, track);
1073
1074         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1075
1076         /* The --period option takes the precedence. */
1077         if (opts->period_set) {
1078                 if (opts->period)
1079                         perf_evsel__set_sample_bit(evsel, PERIOD);
1080                 else
1081                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1082         }
1083 }
1084
1085 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1086 {
1087         if (evsel->system_wide)
1088                 nthreads = 1;
1089
1090         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1091
1092         if (evsel->fd) {
1093                 int cpu, thread;
1094                 for (cpu = 0; cpu < ncpus; cpu++) {
1095                         for (thread = 0; thread < nthreads; thread++) {
1096                                 FD(evsel, cpu, thread) = -1;
1097                         }
1098                 }
1099         }
1100
1101         return evsel->fd != NULL ? 0 : -ENOMEM;
1102 }
1103
1104 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1105                           int ioc,  void *arg)
1106 {
1107         int cpu, thread;
1108
1109         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1110                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1111                         int fd = FD(evsel, cpu, thread),
1112                             err = ioctl(fd, ioc, arg);
1113
1114                         if (err)
1115                                 return err;
1116                 }
1117         }
1118
1119         return 0;
1120 }
1121
1122 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1123 {
1124         return perf_evsel__run_ioctl(evsel,
1125                                      PERF_EVENT_IOC_SET_FILTER,
1126                                      (void *)filter);
1127 }
1128
1129 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1130 {
1131         char *new_filter = strdup(filter);
1132
1133         if (new_filter != NULL) {
1134                 free(evsel->filter);
1135                 evsel->filter = new_filter;
1136                 return 0;
1137         }
1138
1139         return -1;
1140 }
1141
1142 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1143                                      const char *fmt, const char *filter)
1144 {
1145         char *new_filter;
1146
1147         if (evsel->filter == NULL)
1148                 return perf_evsel__set_filter(evsel, filter);
1149
1150         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1151                 free(evsel->filter);
1152                 evsel->filter = new_filter;
1153                 return 0;
1154         }
1155
1156         return -1;
1157 }
1158
1159 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1160 {
1161         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1162 }
1163
1164 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1165 {
1166         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1167 }
1168
1169 int perf_evsel__enable(struct perf_evsel *evsel)
1170 {
1171         return perf_evsel__run_ioctl(evsel,
1172                                      PERF_EVENT_IOC_ENABLE,
1173                                      0);
1174 }
1175
1176 int perf_evsel__disable(struct perf_evsel *evsel)
1177 {
1178         return perf_evsel__run_ioctl(evsel,
1179                                      PERF_EVENT_IOC_DISABLE,
1180                                      0);
1181 }
1182
1183 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1184 {
1185         if (ncpus == 0 || nthreads == 0)
1186                 return 0;
1187
1188         if (evsel->system_wide)
1189                 nthreads = 1;
1190
1191         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1192         if (evsel->sample_id == NULL)
1193                 return -ENOMEM;
1194
1195         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1196         if (evsel->id == NULL) {
1197                 xyarray__delete(evsel->sample_id);
1198                 evsel->sample_id = NULL;
1199                 return -ENOMEM;
1200         }
1201
1202         return 0;
1203 }
1204
1205 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1206 {
1207         xyarray__delete(evsel->fd);
1208         evsel->fd = NULL;
1209 }
1210
1211 static void perf_evsel__free_id(struct perf_evsel *evsel)
1212 {
1213         xyarray__delete(evsel->sample_id);
1214         evsel->sample_id = NULL;
1215         zfree(&evsel->id);
1216 }
1217
1218 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1219 {
1220         struct perf_evsel_config_term *term, *h;
1221
1222         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1223                 list_del(&term->list);
1224                 free(term);
1225         }
1226 }
1227
1228 void perf_evsel__close_fd(struct perf_evsel *evsel)
1229 {
1230         int cpu, thread;
1231
1232         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1233                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1234                         close(FD(evsel, cpu, thread));
1235                         FD(evsel, cpu, thread) = -1;
1236                 }
1237 }
1238
1239 void perf_evsel__exit(struct perf_evsel *evsel)
1240 {
1241         assert(list_empty(&evsel->node));
1242         assert(evsel->evlist == NULL);
1243         perf_evsel__free_fd(evsel);
1244         perf_evsel__free_id(evsel);
1245         perf_evsel__free_config_terms(evsel);
1246         close_cgroup(evsel->cgrp);
1247         cpu_map__put(evsel->cpus);
1248         cpu_map__put(evsel->own_cpus);
1249         thread_map__put(evsel->threads);
1250         zfree(&evsel->group_name);
1251         zfree(&evsel->name);
1252         perf_evsel__object.fini(evsel);
1253 }
1254
1255 void perf_evsel__delete(struct perf_evsel *evsel)
1256 {
1257         perf_evsel__exit(evsel);
1258         free(evsel);
1259 }
1260
1261 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1262                                 struct perf_counts_values *count)
1263 {
1264         struct perf_counts_values tmp;
1265
1266         if (!evsel->prev_raw_counts)
1267                 return;
1268
1269         if (cpu == -1) {
1270                 tmp = evsel->prev_raw_counts->aggr;
1271                 evsel->prev_raw_counts->aggr = *count;
1272         } else {
1273                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1274                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1275         }
1276
1277         count->val = count->val - tmp.val;
1278         count->ena = count->ena - tmp.ena;
1279         count->run = count->run - tmp.run;
1280 }
1281
1282 void perf_counts_values__scale(struct perf_counts_values *count,
1283                                bool scale, s8 *pscaled)
1284 {
1285         s8 scaled = 0;
1286
1287         if (scale) {
1288                 if (count->run == 0) {
1289                         scaled = -1;
1290                         count->val = 0;
1291                 } else if (count->run < count->ena) {
1292                         scaled = 1;
1293                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1294                 }
1295         } else
1296                 count->ena = count->run = 0;
1297
1298         if (pscaled)
1299                 *pscaled = scaled;
1300 }
1301
1302 static int perf_evsel__read_size(struct perf_evsel *evsel)
1303 {
1304         u64 read_format = evsel->attr.read_format;
1305         int entry = sizeof(u64); /* value */
1306         int size = 0;
1307         int nr = 1;
1308
1309         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1310                 size += sizeof(u64);
1311
1312         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1313                 size += sizeof(u64);
1314
1315         if (read_format & PERF_FORMAT_ID)
1316                 entry += sizeof(u64);
1317
1318         if (read_format & PERF_FORMAT_GROUP) {
1319                 nr = evsel->nr_members;
1320                 size += sizeof(u64);
1321         }
1322
1323         size += entry * nr;
1324         return size;
1325 }
1326
1327 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1328                      struct perf_counts_values *count)
1329 {
1330         size_t size = perf_evsel__read_size(evsel);
1331
1332         memset(count, 0, sizeof(*count));
1333
1334         if (FD(evsel, cpu, thread) < 0)
1335                 return -EINVAL;
1336
1337         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1338                 return -errno;
1339
1340         return 0;
1341 }
1342
1343 static int
1344 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1345 {
1346         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1347
1348         return perf_evsel__read(evsel, cpu, thread, count);
1349 }
1350
1351 static void
1352 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1353                       u64 val, u64 ena, u64 run)
1354 {
1355         struct perf_counts_values *count;
1356
1357         count = perf_counts(counter->counts, cpu, thread);
1358
1359         count->val    = val;
1360         count->ena    = ena;
1361         count->run    = run;
1362         count->loaded = true;
1363 }
1364
1365 static int
1366 perf_evsel__process_group_data(struct perf_evsel *leader,
1367                                int cpu, int thread, u64 *data)
1368 {
1369         u64 read_format = leader->attr.read_format;
1370         struct sample_read_value *v;
1371         u64 nr, ena = 0, run = 0, i;
1372
1373         nr = *data++;
1374
1375         if (nr != (u64) leader->nr_members)
1376                 return -EINVAL;
1377
1378         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1379                 ena = *data++;
1380
1381         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1382                 run = *data++;
1383
1384         v = (struct sample_read_value *) data;
1385
1386         perf_evsel__set_count(leader, cpu, thread,
1387                               v[0].value, ena, run);
1388
1389         for (i = 1; i < nr; i++) {
1390                 struct perf_evsel *counter;
1391
1392                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1393                 if (!counter)
1394                         return -EINVAL;
1395
1396                 perf_evsel__set_count(counter, cpu, thread,
1397                                       v[i].value, ena, run);
1398         }
1399
1400         return 0;
1401 }
1402
1403 static int
1404 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1405 {
1406         struct perf_stat_evsel *ps = leader->stats;
1407         u64 read_format = leader->attr.read_format;
1408         int size = perf_evsel__read_size(leader);
1409         u64 *data = ps->group_data;
1410
1411         if (!(read_format & PERF_FORMAT_ID))
1412                 return -EINVAL;
1413
1414         if (!perf_evsel__is_group_leader(leader))
1415                 return -EINVAL;
1416
1417         if (!data) {
1418                 data = zalloc(size);
1419                 if (!data)
1420                         return -ENOMEM;
1421
1422                 ps->group_data = data;
1423         }
1424
1425         if (FD(leader, cpu, thread) < 0)
1426                 return -EINVAL;
1427
1428         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1429                 return -errno;
1430
1431         return perf_evsel__process_group_data(leader, cpu, thread, data);
1432 }
1433
1434 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1435 {
1436         u64 read_format = evsel->attr.read_format;
1437
1438         if (read_format & PERF_FORMAT_GROUP)
1439                 return perf_evsel__read_group(evsel, cpu, thread);
1440         else
1441                 return perf_evsel__read_one(evsel, cpu, thread);
1442 }
1443
1444 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1445                               int cpu, int thread, bool scale)
1446 {
1447         struct perf_counts_values count;
1448         size_t nv = scale ? 3 : 1;
1449
1450         if (FD(evsel, cpu, thread) < 0)
1451                 return -EINVAL;
1452
1453         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1454                 return -ENOMEM;
1455
1456         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1457                 return -errno;
1458
1459         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1460         perf_counts_values__scale(&count, scale, NULL);
1461         *perf_counts(evsel->counts, cpu, thread) = count;
1462         return 0;
1463 }
1464
1465 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1466 {
1467         struct perf_evsel *leader = evsel->leader;
1468         int fd;
1469
1470         if (perf_evsel__is_group_leader(evsel))
1471                 return -1;
1472
1473         /*
1474          * Leader must be already processed/open,
1475          * if not it's a bug.
1476          */
1477         BUG_ON(!leader->fd);
1478
1479         fd = FD(leader, cpu, thread);
1480         BUG_ON(fd == -1);
1481
1482         return fd;
1483 }
1484
1485 struct bit_names {
1486         int bit;
1487         const char *name;
1488 };
1489
1490 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1491 {
1492         bool first_bit = true;
1493         int i = 0;
1494
1495         do {
1496                 if (value & bits[i].bit) {
1497                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1498                         first_bit = false;
1499                 }
1500         } while (bits[++i].name != NULL);
1501 }
1502
1503 static void __p_sample_type(char *buf, size_t size, u64 value)
1504 {
1505 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1506         struct bit_names bits[] = {
1507                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1508                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1509                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1510                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1511                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1512                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1513                 { .name = NULL, }
1514         };
1515 #undef bit_name
1516         __p_bits(buf, size, value, bits);
1517 }
1518
1519 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1520 {
1521 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1522         struct bit_names bits[] = {
1523                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1524                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1525                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1526                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1527                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1528                 { .name = NULL, }
1529         };
1530 #undef bit_name
1531         __p_bits(buf, size, value, bits);
1532 }
1533
1534 static void __p_read_format(char *buf, size_t size, u64 value)
1535 {
1536 #define bit_name(n) { PERF_FORMAT_##n, #n }
1537         struct bit_names bits[] = {
1538                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1539                 bit_name(ID), bit_name(GROUP),
1540                 { .name = NULL, }
1541         };
1542 #undef bit_name
1543         __p_bits(buf, size, value, bits);
1544 }
1545
1546 #define BUF_SIZE                1024
1547
1548 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1549 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1550 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1551 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1552 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1553 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1554
1555 #define PRINT_ATTRn(_n, _f, _p)                         \
1556 do {                                                    \
1557         if (attr->_f) {                                 \
1558                 _p(attr->_f);                           \
1559                 ret += attr__fprintf(fp, _n, buf, priv);\
1560         }                                               \
1561 } while (0)
1562
1563 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1564
1565 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1566                              attr__fprintf_f attr__fprintf, void *priv)
1567 {
1568         char buf[BUF_SIZE];
1569         int ret = 0;
1570
1571         PRINT_ATTRf(type, p_unsigned);
1572         PRINT_ATTRf(size, p_unsigned);
1573         PRINT_ATTRf(config, p_hex);
1574         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1575         PRINT_ATTRf(sample_type, p_sample_type);
1576         PRINT_ATTRf(read_format, p_read_format);
1577
1578         PRINT_ATTRf(disabled, p_unsigned);
1579         PRINT_ATTRf(inherit, p_unsigned);
1580         PRINT_ATTRf(pinned, p_unsigned);
1581         PRINT_ATTRf(exclusive, p_unsigned);
1582         PRINT_ATTRf(exclude_user, p_unsigned);
1583         PRINT_ATTRf(exclude_kernel, p_unsigned);
1584         PRINT_ATTRf(exclude_hv, p_unsigned);
1585         PRINT_ATTRf(exclude_idle, p_unsigned);
1586         PRINT_ATTRf(mmap, p_unsigned);
1587         PRINT_ATTRf(comm, p_unsigned);
1588         PRINT_ATTRf(freq, p_unsigned);
1589         PRINT_ATTRf(inherit_stat, p_unsigned);
1590         PRINT_ATTRf(enable_on_exec, p_unsigned);
1591         PRINT_ATTRf(task, p_unsigned);
1592         PRINT_ATTRf(watermark, p_unsigned);
1593         PRINT_ATTRf(precise_ip, p_unsigned);
1594         PRINT_ATTRf(mmap_data, p_unsigned);
1595         PRINT_ATTRf(sample_id_all, p_unsigned);
1596         PRINT_ATTRf(exclude_host, p_unsigned);
1597         PRINT_ATTRf(exclude_guest, p_unsigned);
1598         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1599         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1600         PRINT_ATTRf(mmap2, p_unsigned);
1601         PRINT_ATTRf(comm_exec, p_unsigned);
1602         PRINT_ATTRf(use_clockid, p_unsigned);
1603         PRINT_ATTRf(context_switch, p_unsigned);
1604         PRINT_ATTRf(write_backward, p_unsigned);
1605         PRINT_ATTRf(namespaces, p_unsigned);
1606
1607         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1608         PRINT_ATTRf(bp_type, p_unsigned);
1609         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1610         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1611         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1612         PRINT_ATTRf(sample_regs_user, p_hex);
1613         PRINT_ATTRf(sample_stack_user, p_unsigned);
1614         PRINT_ATTRf(clockid, p_signed);
1615         PRINT_ATTRf(sample_regs_intr, p_hex);
1616         PRINT_ATTRf(aux_watermark, p_unsigned);
1617         PRINT_ATTRf(sample_max_stack, p_unsigned);
1618
1619         return ret;
1620 }
1621
1622 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1623                                 void *priv __maybe_unused)
1624 {
1625         return fprintf(fp, "  %-32s %s\n", name, val);
1626 }
1627
1628 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1629                                   int nr_cpus, int nr_threads,
1630                                   int thread_idx)
1631 {
1632         for (int cpu = 0; cpu < nr_cpus; cpu++)
1633                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1634                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1635 }
1636
1637 static int update_fds(struct perf_evsel *evsel,
1638                       int nr_cpus, int cpu_idx,
1639                       int nr_threads, int thread_idx)
1640 {
1641         struct perf_evsel *pos;
1642
1643         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1644                 return -EINVAL;
1645
1646         evlist__for_each_entry(evsel->evlist, pos) {
1647                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1648
1649                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1650
1651                 /*
1652                  * Since fds for next evsel has not been created,
1653                  * there is no need to iterate whole event list.
1654                  */
1655                 if (pos == evsel)
1656                         break;
1657         }
1658         return 0;
1659 }
1660
1661 static bool ignore_missing_thread(struct perf_evsel *evsel,
1662                                   int nr_cpus, int cpu,
1663                                   struct thread_map *threads,
1664                                   int thread, int err)
1665 {
1666         pid_t ignore_pid = thread_map__pid(threads, thread);
1667
1668         if (!evsel->ignore_missing_thread)
1669                 return false;
1670
1671         /* The system wide setup does not work with threads. */
1672         if (evsel->system_wide)
1673                 return false;
1674
1675         /* The -ESRCH is perf event syscall errno for pid's not found. */
1676         if (err != -ESRCH)
1677                 return false;
1678
1679         /* If there's only one thread, let it fail. */
1680         if (threads->nr == 1)
1681                 return false;
1682
1683         /*
1684          * We should remove fd for missing_thread first
1685          * because thread_map__remove() will decrease threads->nr.
1686          */
1687         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1688                 return false;
1689
1690         if (thread_map__remove(threads, thread))
1691                 return false;
1692
1693         pr_warning("WARNING: Ignored open failure for pid %d\n",
1694                    ignore_pid);
1695         return true;
1696 }
1697
1698 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1699                      struct thread_map *threads)
1700 {
1701         int cpu, thread, nthreads;
1702         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1703         int pid = -1, err;
1704         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1705
1706         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1707                 return -EINVAL;
1708
1709         if (cpus == NULL) {
1710                 static struct cpu_map *empty_cpu_map;
1711
1712                 if (empty_cpu_map == NULL) {
1713                         empty_cpu_map = cpu_map__dummy_new();
1714                         if (empty_cpu_map == NULL)
1715                                 return -ENOMEM;
1716                 }
1717
1718                 cpus = empty_cpu_map;
1719         }
1720
1721         if (threads == NULL) {
1722                 static struct thread_map *empty_thread_map;
1723
1724                 if (empty_thread_map == NULL) {
1725                         empty_thread_map = thread_map__new_by_tid(-1);
1726                         if (empty_thread_map == NULL)
1727                                 return -ENOMEM;
1728                 }
1729
1730                 threads = empty_thread_map;
1731         }
1732
1733         if (evsel->system_wide)
1734                 nthreads = 1;
1735         else
1736                 nthreads = threads->nr;
1737
1738         if (evsel->fd == NULL &&
1739             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1740                 return -ENOMEM;
1741
1742         if (evsel->cgrp) {
1743                 flags |= PERF_FLAG_PID_CGROUP;
1744                 pid = evsel->cgrp->fd;
1745         }
1746
1747 fallback_missing_features:
1748         if (perf_missing_features.clockid_wrong)
1749                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1750         if (perf_missing_features.clockid) {
1751                 evsel->attr.use_clockid = 0;
1752                 evsel->attr.clockid = 0;
1753         }
1754         if (perf_missing_features.cloexec)
1755                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1756         if (perf_missing_features.mmap2)
1757                 evsel->attr.mmap2 = 0;
1758         if (perf_missing_features.exclude_guest)
1759                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1760         if (perf_missing_features.lbr_flags)
1761                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1762                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1763         if (perf_missing_features.group_read && evsel->attr.inherit)
1764                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1765 retry_sample_id:
1766         if (perf_missing_features.sample_id_all)
1767                 evsel->attr.sample_id_all = 0;
1768
1769         if (verbose >= 2) {
1770                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1771                 fprintf(stderr, "perf_event_attr:\n");
1772                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1773                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1774         }
1775
1776         for (cpu = 0; cpu < cpus->nr; cpu++) {
1777
1778                 for (thread = 0; thread < nthreads; thread++) {
1779                         int fd, group_fd;
1780
1781                         if (!evsel->cgrp && !evsel->system_wide)
1782                                 pid = thread_map__pid(threads, thread);
1783
1784                         group_fd = get_group_fd(evsel, cpu, thread);
1785 retry_open:
1786                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1787                                   pid, cpus->map[cpu], group_fd, flags);
1788
1789                         test_attr__ready();
1790
1791                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1792                                                  group_fd, flags);
1793
1794                         FD(evsel, cpu, thread) = fd;
1795
1796                         if (fd < 0) {
1797                                 err = -errno;
1798
1799                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1800                                         /*
1801                                          * We just removed 1 thread, so take a step
1802                                          * back on thread index and lower the upper
1803                                          * nthreads limit.
1804                                          */
1805                                         nthreads--;
1806                                         thread--;
1807
1808                                         /* ... and pretend like nothing have happened. */
1809                                         err = 0;
1810                                         continue;
1811                                 }
1812
1813                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1814                                           err);
1815                                 goto try_fallback;
1816                         }
1817
1818                         pr_debug2(" = %d\n", fd);
1819
1820                         if (evsel->bpf_fd >= 0) {
1821                                 int evt_fd = fd;
1822                                 int bpf_fd = evsel->bpf_fd;
1823
1824                                 err = ioctl(evt_fd,
1825                                             PERF_EVENT_IOC_SET_BPF,
1826                                             bpf_fd);
1827                                 if (err && errno != EEXIST) {
1828                                         pr_err("failed to attach bpf fd %d: %s\n",
1829                                                bpf_fd, strerror(errno));
1830                                         err = -EINVAL;
1831                                         goto out_close;
1832                                 }
1833                         }
1834
1835                         set_rlimit = NO_CHANGE;
1836
1837                         /*
1838                          * If we succeeded but had to kill clockid, fail and
1839                          * have perf_evsel__open_strerror() print us a nice
1840                          * error.
1841                          */
1842                         if (perf_missing_features.clockid ||
1843                             perf_missing_features.clockid_wrong) {
1844                                 err = -EINVAL;
1845                                 goto out_close;
1846                         }
1847                 }
1848         }
1849
1850         return 0;
1851
1852 try_fallback:
1853         /*
1854          * perf stat needs between 5 and 22 fds per CPU. When we run out
1855          * of them try to increase the limits.
1856          */
1857         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1858                 struct rlimit l;
1859                 int old_errno = errno;
1860
1861                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1862                         if (set_rlimit == NO_CHANGE)
1863                                 l.rlim_cur = l.rlim_max;
1864                         else {
1865                                 l.rlim_cur = l.rlim_max + 1000;
1866                                 l.rlim_max = l.rlim_cur;
1867                         }
1868                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1869                                 set_rlimit++;
1870                                 errno = old_errno;
1871                                 goto retry_open;
1872                         }
1873                 }
1874                 errno = old_errno;
1875         }
1876
1877         if (err != -EINVAL || cpu > 0 || thread > 0)
1878                 goto out_close;
1879
1880         /*
1881          * Must probe features in the order they were added to the
1882          * perf_event_attr interface.
1883          */
1884         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1885                 perf_missing_features.write_backward = true;
1886                 pr_debug2("switching off write_backward\n");
1887                 goto out_close;
1888         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1889                 perf_missing_features.clockid_wrong = true;
1890                 pr_debug2("switching off clockid\n");
1891                 goto fallback_missing_features;
1892         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1893                 perf_missing_features.clockid = true;
1894                 pr_debug2("switching off use_clockid\n");
1895                 goto fallback_missing_features;
1896         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1897                 perf_missing_features.cloexec = true;
1898                 pr_debug2("switching off cloexec flag\n");
1899                 goto fallback_missing_features;
1900         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1901                 perf_missing_features.mmap2 = true;
1902                 pr_debug2("switching off mmap2\n");
1903                 goto fallback_missing_features;
1904         } else if (!perf_missing_features.exclude_guest &&
1905                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1906                 perf_missing_features.exclude_guest = true;
1907                 pr_debug2("switching off exclude_guest, exclude_host\n");
1908                 goto fallback_missing_features;
1909         } else if (!perf_missing_features.sample_id_all) {
1910                 perf_missing_features.sample_id_all = true;
1911                 pr_debug2("switching off sample_id_all\n");
1912                 goto retry_sample_id;
1913         } else if (!perf_missing_features.lbr_flags &&
1914                         (evsel->attr.branch_sample_type &
1915                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1916                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1917                 perf_missing_features.lbr_flags = true;
1918                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1919                 goto fallback_missing_features;
1920         } else if (!perf_missing_features.group_read &&
1921                     evsel->attr.inherit &&
1922                    (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
1923                 perf_missing_features.group_read = true;
1924                 pr_debug2("switching off group read\n");
1925                 goto fallback_missing_features;
1926         }
1927 out_close:
1928         do {
1929                 while (--thread >= 0) {
1930                         close(FD(evsel, cpu, thread));
1931                         FD(evsel, cpu, thread) = -1;
1932                 }
1933                 thread = nthreads;
1934         } while (--cpu >= 0);
1935         return err;
1936 }
1937
1938 void perf_evsel__close(struct perf_evsel *evsel)
1939 {
1940         if (evsel->fd == NULL)
1941                 return;
1942
1943         perf_evsel__close_fd(evsel);
1944         perf_evsel__free_fd(evsel);
1945 }
1946
1947 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1948                              struct cpu_map *cpus)
1949 {
1950         return perf_evsel__open(evsel, cpus, NULL);
1951 }
1952
1953 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1954                                 struct thread_map *threads)
1955 {
1956         return perf_evsel__open(evsel, NULL, threads);
1957 }
1958
1959 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1960                                        const union perf_event *event,
1961                                        struct perf_sample *sample)
1962 {
1963         u64 type = evsel->attr.sample_type;
1964         const u64 *array = event->sample.array;
1965         bool swapped = evsel->needs_swap;
1966         union u64_swap u;
1967
1968         array += ((event->header.size -
1969                    sizeof(event->header)) / sizeof(u64)) - 1;
1970
1971         if (type & PERF_SAMPLE_IDENTIFIER) {
1972                 sample->id = *array;
1973                 array--;
1974         }
1975
1976         if (type & PERF_SAMPLE_CPU) {
1977                 u.val64 = *array;
1978                 if (swapped) {
1979                         /* undo swap of u64, then swap on individual u32s */
1980                         u.val64 = bswap_64(u.val64);
1981                         u.val32[0] = bswap_32(u.val32[0]);
1982                 }
1983
1984                 sample->cpu = u.val32[0];
1985                 array--;
1986         }
1987
1988         if (type & PERF_SAMPLE_STREAM_ID) {
1989                 sample->stream_id = *array;
1990                 array--;
1991         }
1992
1993         if (type & PERF_SAMPLE_ID) {
1994                 sample->id = *array;
1995                 array--;
1996         }
1997
1998         if (type & PERF_SAMPLE_TIME) {
1999                 sample->time = *array;
2000                 array--;
2001         }
2002
2003         if (type & PERF_SAMPLE_TID) {
2004                 u.val64 = *array;
2005                 if (swapped) {
2006                         /* undo swap of u64, then swap on individual u32s */
2007                         u.val64 = bswap_64(u.val64);
2008                         u.val32[0] = bswap_32(u.val32[0]);
2009                         u.val32[1] = bswap_32(u.val32[1]);
2010                 }
2011
2012                 sample->pid = u.val32[0];
2013                 sample->tid = u.val32[1];
2014                 array--;
2015         }
2016
2017         return 0;
2018 }
2019
2020 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2021                             u64 size)
2022 {
2023         return size > max_size || offset + size > endp;
2024 }
2025
2026 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2027         do {                                                            \
2028                 if (overflow(endp, (max_size), (offset), (size)))       \
2029                         return -EFAULT;                                 \
2030         } while (0)
2031
2032 #define OVERFLOW_CHECK_u64(offset) \
2033         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2034
2035 static int
2036 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2037 {
2038         /*
2039          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2040          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2041          * check the format does not go past the end of the event.
2042          */
2043         if (sample_size + sizeof(event->header) > event->header.size)
2044                 return -EFAULT;
2045
2046         return 0;
2047 }
2048
2049 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2050                              struct perf_sample *data)
2051 {
2052         u64 type = evsel->attr.sample_type;
2053         bool swapped = evsel->needs_swap;
2054         const u64 *array;
2055         u16 max_size = event->header.size;
2056         const void *endp = (void *)event + max_size;
2057         u64 sz;
2058
2059         /*
2060          * used for cross-endian analysis. See git commit 65014ab3
2061          * for why this goofiness is needed.
2062          */
2063         union u64_swap u;
2064
2065         memset(data, 0, sizeof(*data));
2066         data->cpu = data->pid = data->tid = -1;
2067         data->stream_id = data->id = data->time = -1ULL;
2068         data->period = evsel->attr.sample_period;
2069         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2070         data->misc    = event->header.misc;
2071         data->id = -1ULL;
2072         data->data_src = PERF_MEM_DATA_SRC_NONE;
2073
2074         if (event->header.type != PERF_RECORD_SAMPLE) {
2075                 if (!evsel->attr.sample_id_all)
2076                         return 0;
2077                 return perf_evsel__parse_id_sample(evsel, event, data);
2078         }
2079
2080         array = event->sample.array;
2081
2082         if (perf_event__check_size(event, evsel->sample_size))
2083                 return -EFAULT;
2084
2085         if (type & PERF_SAMPLE_IDENTIFIER) {
2086                 data->id = *array;
2087                 array++;
2088         }
2089
2090         if (type & PERF_SAMPLE_IP) {
2091                 data->ip = *array;
2092                 array++;
2093         }
2094
2095         if (type & PERF_SAMPLE_TID) {
2096                 u.val64 = *array;
2097                 if (swapped) {
2098                         /* undo swap of u64, then swap on individual u32s */
2099                         u.val64 = bswap_64(u.val64);
2100                         u.val32[0] = bswap_32(u.val32[0]);
2101                         u.val32[1] = bswap_32(u.val32[1]);
2102                 }
2103
2104                 data->pid = u.val32[0];
2105                 data->tid = u.val32[1];
2106                 array++;
2107         }
2108
2109         if (type & PERF_SAMPLE_TIME) {
2110                 data->time = *array;
2111                 array++;
2112         }
2113
2114         if (type & PERF_SAMPLE_ADDR) {
2115                 data->addr = *array;
2116                 array++;
2117         }
2118
2119         if (type & PERF_SAMPLE_ID) {
2120                 data->id = *array;
2121                 array++;
2122         }
2123
2124         if (type & PERF_SAMPLE_STREAM_ID) {
2125                 data->stream_id = *array;
2126                 array++;
2127         }
2128
2129         if (type & PERF_SAMPLE_CPU) {
2130
2131                 u.val64 = *array;
2132                 if (swapped) {
2133                         /* undo swap of u64, then swap on individual u32s */
2134                         u.val64 = bswap_64(u.val64);
2135                         u.val32[0] = bswap_32(u.val32[0]);
2136                 }
2137
2138                 data->cpu = u.val32[0];
2139                 array++;
2140         }
2141
2142         if (type & PERF_SAMPLE_PERIOD) {
2143                 data->period = *array;
2144                 array++;
2145         }
2146
2147         if (type & PERF_SAMPLE_READ) {
2148                 u64 read_format = evsel->attr.read_format;
2149
2150                 OVERFLOW_CHECK_u64(array);
2151                 if (read_format & PERF_FORMAT_GROUP)
2152                         data->read.group.nr = *array;
2153                 else
2154                         data->read.one.value = *array;
2155
2156                 array++;
2157
2158                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2159                         OVERFLOW_CHECK_u64(array);
2160                         data->read.time_enabled = *array;
2161                         array++;
2162                 }
2163
2164                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2165                         OVERFLOW_CHECK_u64(array);
2166                         data->read.time_running = *array;
2167                         array++;
2168                 }
2169
2170                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2171                 if (read_format & PERF_FORMAT_GROUP) {
2172                         const u64 max_group_nr = UINT64_MAX /
2173                                         sizeof(struct sample_read_value);
2174
2175                         if (data->read.group.nr > max_group_nr)
2176                                 return -EFAULT;
2177                         sz = data->read.group.nr *
2178                              sizeof(struct sample_read_value);
2179                         OVERFLOW_CHECK(array, sz, max_size);
2180                         data->read.group.values =
2181                                         (struct sample_read_value *)array;
2182                         array = (void *)array + sz;
2183                 } else {
2184                         OVERFLOW_CHECK_u64(array);
2185                         data->read.one.id = *array;
2186                         array++;
2187                 }
2188         }
2189
2190         if (type & PERF_SAMPLE_CALLCHAIN) {
2191                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2192
2193                 OVERFLOW_CHECK_u64(array);
2194                 data->callchain = (struct ip_callchain *)array++;
2195                 if (data->callchain->nr > max_callchain_nr)
2196                         return -EFAULT;
2197                 sz = data->callchain->nr * sizeof(u64);
2198                 OVERFLOW_CHECK(array, sz, max_size);
2199                 array = (void *)array + sz;
2200         }
2201
2202         if (type & PERF_SAMPLE_RAW) {
2203                 OVERFLOW_CHECK_u64(array);
2204                 u.val64 = *array;
2205
2206                 /*
2207                  * Undo swap of u64, then swap on individual u32s,
2208                  * get the size of the raw area and undo all of the
2209                  * swap. The pevent interface handles endianity by
2210                  * itself.
2211                  */
2212                 if (swapped) {
2213                         u.val64 = bswap_64(u.val64);
2214                         u.val32[0] = bswap_32(u.val32[0]);
2215                         u.val32[1] = bswap_32(u.val32[1]);
2216                 }
2217                 data->raw_size = u.val32[0];
2218
2219                 /*
2220                  * The raw data is aligned on 64bits including the
2221                  * u32 size, so it's safe to use mem_bswap_64.
2222                  */
2223                 if (swapped)
2224                         mem_bswap_64((void *) array, data->raw_size);
2225
2226                 array = (void *)array + sizeof(u32);
2227
2228                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2229                 data->raw_data = (void *)array;
2230                 array = (void *)array + data->raw_size;
2231         }
2232
2233         if (type & PERF_SAMPLE_BRANCH_STACK) {
2234                 const u64 max_branch_nr = UINT64_MAX /
2235                                           sizeof(struct branch_entry);
2236
2237                 OVERFLOW_CHECK_u64(array);
2238                 data->branch_stack = (struct branch_stack *)array++;
2239
2240                 if (data->branch_stack->nr > max_branch_nr)
2241                         return -EFAULT;
2242                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2243                 OVERFLOW_CHECK(array, sz, max_size);
2244                 array = (void *)array + sz;
2245         }
2246
2247         if (type & PERF_SAMPLE_REGS_USER) {
2248                 OVERFLOW_CHECK_u64(array);
2249                 data->user_regs.abi = *array;
2250                 array++;
2251
2252                 if (data->user_regs.abi) {
2253                         u64 mask = evsel->attr.sample_regs_user;
2254
2255                         sz = hweight_long(mask) * sizeof(u64);
2256                         OVERFLOW_CHECK(array, sz, max_size);
2257                         data->user_regs.mask = mask;
2258                         data->user_regs.regs = (u64 *)array;
2259                         array = (void *)array + sz;
2260                 }
2261         }
2262
2263         if (type & PERF_SAMPLE_STACK_USER) {
2264                 OVERFLOW_CHECK_u64(array);
2265                 sz = *array++;
2266
2267                 data->user_stack.offset = ((char *)(array - 1)
2268                                           - (char *) event);
2269
2270                 if (!sz) {
2271                         data->user_stack.size = 0;
2272                 } else {
2273                         OVERFLOW_CHECK(array, sz, max_size);
2274                         data->user_stack.data = (char *)array;
2275                         array = (void *)array + sz;
2276                         OVERFLOW_CHECK_u64(array);
2277                         data->user_stack.size = *array++;
2278                         if (WARN_ONCE(data->user_stack.size > sz,
2279                                       "user stack dump failure\n"))
2280                                 return -EFAULT;
2281                 }
2282         }
2283
2284         if (type & PERF_SAMPLE_WEIGHT) {
2285                 OVERFLOW_CHECK_u64(array);
2286                 data->weight = *array;
2287                 array++;
2288         }
2289
2290         if (type & PERF_SAMPLE_DATA_SRC) {
2291                 OVERFLOW_CHECK_u64(array);
2292                 data->data_src = *array;
2293                 array++;
2294         }
2295
2296         if (type & PERF_SAMPLE_TRANSACTION) {
2297                 OVERFLOW_CHECK_u64(array);
2298                 data->transaction = *array;
2299                 array++;
2300         }
2301
2302         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2303         if (type & PERF_SAMPLE_REGS_INTR) {
2304                 OVERFLOW_CHECK_u64(array);
2305                 data->intr_regs.abi = *array;
2306                 array++;
2307
2308                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2309                         u64 mask = evsel->attr.sample_regs_intr;
2310
2311                         sz = hweight_long(mask) * sizeof(u64);
2312                         OVERFLOW_CHECK(array, sz, max_size);
2313                         data->intr_regs.mask = mask;
2314                         data->intr_regs.regs = (u64 *)array;
2315                         array = (void *)array + sz;
2316                 }
2317         }
2318
2319         data->phys_addr = 0;
2320         if (type & PERF_SAMPLE_PHYS_ADDR) {
2321                 data->phys_addr = *array;
2322                 array++;
2323         }
2324
2325         return 0;
2326 }
2327
2328 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2329                                        union perf_event *event,
2330                                        u64 *timestamp)
2331 {
2332         u64 type = evsel->attr.sample_type;
2333         const u64 *array;
2334
2335         if (!(type & PERF_SAMPLE_TIME))
2336                 return -1;
2337
2338         if (event->header.type != PERF_RECORD_SAMPLE) {
2339                 struct perf_sample data = {
2340                         .time = -1ULL,
2341                 };
2342
2343                 if (!evsel->attr.sample_id_all)
2344                         return -1;
2345                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2346                         return -1;
2347
2348                 *timestamp = data.time;
2349                 return 0;
2350         }
2351
2352         array = event->sample.array;
2353
2354         if (perf_event__check_size(event, evsel->sample_size))
2355                 return -EFAULT;
2356
2357         if (type & PERF_SAMPLE_IDENTIFIER)
2358                 array++;
2359
2360         if (type & PERF_SAMPLE_IP)
2361                 array++;
2362
2363         if (type & PERF_SAMPLE_TID)
2364                 array++;
2365
2366         if (type & PERF_SAMPLE_TIME)
2367                 *timestamp = *array;
2368
2369         return 0;
2370 }
2371
2372 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2373                                      u64 read_format)
2374 {
2375         size_t sz, result = sizeof(struct sample_event);
2376
2377         if (type & PERF_SAMPLE_IDENTIFIER)
2378                 result += sizeof(u64);
2379
2380         if (type & PERF_SAMPLE_IP)
2381                 result += sizeof(u64);
2382
2383         if (type & PERF_SAMPLE_TID)
2384                 result += sizeof(u64);
2385
2386         if (type & PERF_SAMPLE_TIME)
2387                 result += sizeof(u64);
2388
2389         if (type & PERF_SAMPLE_ADDR)
2390                 result += sizeof(u64);
2391
2392         if (type & PERF_SAMPLE_ID)
2393                 result += sizeof(u64);
2394
2395         if (type & PERF_SAMPLE_STREAM_ID)
2396                 result += sizeof(u64);
2397
2398         if (type & PERF_SAMPLE_CPU)
2399                 result += sizeof(u64);
2400
2401         if (type & PERF_SAMPLE_PERIOD)
2402                 result += sizeof(u64);
2403
2404         if (type & PERF_SAMPLE_READ) {
2405                 result += sizeof(u64);
2406                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2407                         result += sizeof(u64);
2408                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2409                         result += sizeof(u64);
2410                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2411                 if (read_format & PERF_FORMAT_GROUP) {
2412                         sz = sample->read.group.nr *
2413                              sizeof(struct sample_read_value);
2414                         result += sz;
2415                 } else {
2416                         result += sizeof(u64);
2417                 }
2418         }
2419
2420         if (type & PERF_SAMPLE_CALLCHAIN) {
2421                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2422                 result += sz;
2423         }
2424
2425         if (type & PERF_SAMPLE_RAW) {
2426                 result += sizeof(u32);
2427                 result += sample->raw_size;
2428         }
2429
2430         if (type & PERF_SAMPLE_BRANCH_STACK) {
2431                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2432                 sz += sizeof(u64);
2433                 result += sz;
2434         }
2435
2436         if (type & PERF_SAMPLE_REGS_USER) {
2437                 if (sample->user_regs.abi) {
2438                         result += sizeof(u64);
2439                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2440                         result += sz;
2441                 } else {
2442                         result += sizeof(u64);
2443                 }
2444         }
2445
2446         if (type & PERF_SAMPLE_STACK_USER) {
2447                 sz = sample->user_stack.size;
2448                 result += sizeof(u64);
2449                 if (sz) {
2450                         result += sz;
2451                         result += sizeof(u64);
2452                 }
2453         }
2454
2455         if (type & PERF_SAMPLE_WEIGHT)
2456                 result += sizeof(u64);
2457
2458         if (type & PERF_SAMPLE_DATA_SRC)
2459                 result += sizeof(u64);
2460
2461         if (type & PERF_SAMPLE_TRANSACTION)
2462                 result += sizeof(u64);
2463
2464         if (type & PERF_SAMPLE_REGS_INTR) {
2465                 if (sample->intr_regs.abi) {
2466                         result += sizeof(u64);
2467                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2468                         result += sz;
2469                 } else {
2470                         result += sizeof(u64);
2471                 }
2472         }
2473
2474         if (type & PERF_SAMPLE_PHYS_ADDR)
2475                 result += sizeof(u64);
2476
2477         return result;
2478 }
2479
2480 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2481                                   u64 read_format,
2482                                   const struct perf_sample *sample)
2483 {
2484         u64 *array;
2485         size_t sz;
2486         /*
2487          * used for cross-endian analysis. See git commit 65014ab3
2488          * for why this goofiness is needed.
2489          */
2490         union u64_swap u;
2491
2492         array = event->sample.array;
2493
2494         if (type & PERF_SAMPLE_IDENTIFIER) {
2495                 *array = sample->id;
2496                 array++;
2497         }
2498
2499         if (type & PERF_SAMPLE_IP) {
2500                 *array = sample->ip;
2501                 array++;
2502         }
2503
2504         if (type & PERF_SAMPLE_TID) {
2505                 u.val32[0] = sample->pid;
2506                 u.val32[1] = sample->tid;
2507                 *array = u.val64;
2508                 array++;
2509         }
2510
2511         if (type & PERF_SAMPLE_TIME) {
2512                 *array = sample->time;
2513                 array++;
2514         }
2515
2516         if (type & PERF_SAMPLE_ADDR) {
2517                 *array = sample->addr;
2518                 array++;
2519         }
2520
2521         if (type & PERF_SAMPLE_ID) {
2522                 *array = sample->id;
2523                 array++;
2524         }
2525
2526         if (type & PERF_SAMPLE_STREAM_ID) {
2527                 *array = sample->stream_id;
2528                 array++;
2529         }
2530
2531         if (type & PERF_SAMPLE_CPU) {
2532                 u.val32[0] = sample->cpu;
2533                 u.val32[1] = 0;
2534                 *array = u.val64;
2535                 array++;
2536         }
2537
2538         if (type & PERF_SAMPLE_PERIOD) {
2539                 *array = sample->period;
2540                 array++;
2541         }
2542
2543         if (type & PERF_SAMPLE_READ) {
2544                 if (read_format & PERF_FORMAT_GROUP)
2545                         *array = sample->read.group.nr;
2546                 else
2547                         *array = sample->read.one.value;
2548                 array++;
2549
2550                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2551                         *array = sample->read.time_enabled;
2552                         array++;
2553                 }
2554
2555                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2556                         *array = sample->read.time_running;
2557                         array++;
2558                 }
2559
2560                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2561                 if (read_format & PERF_FORMAT_GROUP) {
2562                         sz = sample->read.group.nr *
2563                              sizeof(struct sample_read_value);
2564                         memcpy(array, sample->read.group.values, sz);
2565                         array = (void *)array + sz;
2566                 } else {
2567                         *array = sample->read.one.id;
2568                         array++;
2569                 }
2570         }
2571
2572         if (type & PERF_SAMPLE_CALLCHAIN) {
2573                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2574                 memcpy(array, sample->callchain, sz);
2575                 array = (void *)array + sz;
2576         }
2577
2578         if (type & PERF_SAMPLE_RAW) {
2579                 u.val32[0] = sample->raw_size;
2580                 *array = u.val64;
2581                 array = (void *)array + sizeof(u32);
2582
2583                 memcpy(array, sample->raw_data, sample->raw_size);
2584                 array = (void *)array + sample->raw_size;
2585         }
2586
2587         if (type & PERF_SAMPLE_BRANCH_STACK) {
2588                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2589                 sz += sizeof(u64);
2590                 memcpy(array, sample->branch_stack, sz);
2591                 array = (void *)array + sz;
2592         }
2593
2594         if (type & PERF_SAMPLE_REGS_USER) {
2595                 if (sample->user_regs.abi) {
2596                         *array++ = sample->user_regs.abi;
2597                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2598                         memcpy(array, sample->user_regs.regs, sz);
2599                         array = (void *)array + sz;
2600                 } else {
2601                         *array++ = 0;
2602                 }
2603         }
2604
2605         if (type & PERF_SAMPLE_STACK_USER) {
2606                 sz = sample->user_stack.size;
2607                 *array++ = sz;
2608                 if (sz) {
2609                         memcpy(array, sample->user_stack.data, sz);
2610                         array = (void *)array + sz;
2611                         *array++ = sz;
2612                 }
2613         }
2614
2615         if (type & PERF_SAMPLE_WEIGHT) {
2616                 *array = sample->weight;
2617                 array++;
2618         }
2619
2620         if (type & PERF_SAMPLE_DATA_SRC) {
2621                 *array = sample->data_src;
2622                 array++;
2623         }
2624
2625         if (type & PERF_SAMPLE_TRANSACTION) {
2626                 *array = sample->transaction;
2627                 array++;
2628         }
2629
2630         if (type & PERF_SAMPLE_REGS_INTR) {
2631                 if (sample->intr_regs.abi) {
2632                         *array++ = sample->intr_regs.abi;
2633                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2634                         memcpy(array, sample->intr_regs.regs, sz);
2635                         array = (void *)array + sz;
2636                 } else {
2637                         *array++ = 0;
2638                 }
2639         }
2640
2641         if (type & PERF_SAMPLE_PHYS_ADDR) {
2642                 *array = sample->phys_addr;
2643                 array++;
2644         }
2645
2646         return 0;
2647 }
2648
2649 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2650 {
2651         return pevent_find_field(evsel->tp_format, name);
2652 }
2653
2654 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2655                          const char *name)
2656 {
2657         struct format_field *field = perf_evsel__field(evsel, name);
2658         int offset;
2659
2660         if (!field)
2661                 return NULL;
2662
2663         offset = field->offset;
2664
2665         if (field->flags & FIELD_IS_DYNAMIC) {
2666                 offset = *(int *)(sample->raw_data + field->offset);
2667                 offset &= 0xffff;
2668         }
2669
2670         return sample->raw_data + offset;
2671 }
2672
2673 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2674                          bool needs_swap)
2675 {
2676         u64 value;
2677         void *ptr = sample->raw_data + field->offset;
2678
2679         switch (field->size) {
2680         case 1:
2681                 return *(u8 *)ptr;
2682         case 2:
2683                 value = *(u16 *)ptr;
2684                 break;
2685         case 4:
2686                 value = *(u32 *)ptr;
2687                 break;
2688         case 8:
2689                 memcpy(&value, ptr, sizeof(u64));
2690                 break;
2691         default:
2692                 return 0;
2693         }
2694
2695         if (!needs_swap)
2696                 return value;
2697
2698         switch (field->size) {
2699         case 2:
2700                 return bswap_16(value);
2701         case 4:
2702                 return bswap_32(value);
2703         case 8:
2704                 return bswap_64(value);
2705         default:
2706                 return 0;
2707         }
2708
2709         return 0;
2710 }
2711
2712 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2713                        const char *name)
2714 {
2715         struct format_field *field = perf_evsel__field(evsel, name);
2716
2717         if (!field)
2718                 return 0;
2719
2720         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2721 }
2722
2723 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2724                           char *msg, size_t msgsize)
2725 {
2726         int paranoid;
2727
2728         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2729             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2730             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2731                 /*
2732                  * If it's cycles then fall back to hrtimer based
2733                  * cpu-clock-tick sw counter, which is always available even if
2734                  * no PMU support.
2735                  *
2736                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2737                  * b0a873e).
2738                  */
2739                 scnprintf(msg, msgsize, "%s",
2740 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2741
2742                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2743                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2744
2745                 zfree(&evsel->name);
2746                 return true;
2747         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2748                    (paranoid = perf_event_paranoid()) > 1) {
2749                 const char *name = perf_evsel__name(evsel);
2750                 char *new_name;
2751
2752                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2753                         return false;
2754
2755                 if (evsel->name)
2756                         free(evsel->name);
2757                 evsel->name = new_name;
2758                 scnprintf(msg, msgsize,
2759 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2760                 evsel->attr.exclude_kernel = 1;
2761
2762                 return true;
2763         }
2764
2765         return false;
2766 }
2767
2768 static bool find_process(const char *name)
2769 {
2770         size_t len = strlen(name);
2771         DIR *dir;
2772         struct dirent *d;
2773         int ret = -1;
2774
2775         dir = opendir(procfs__mountpoint());
2776         if (!dir)
2777                 return false;
2778
2779         /* Walk through the directory. */
2780         while (ret && (d = readdir(dir)) != NULL) {
2781                 char path[PATH_MAX];
2782                 char *data;
2783                 size_t size;
2784
2785                 if ((d->d_type != DT_DIR) ||
2786                      !strcmp(".", d->d_name) ||
2787                      !strcmp("..", d->d_name))
2788                         continue;
2789
2790                 scnprintf(path, sizeof(path), "%s/%s/comm",
2791                           procfs__mountpoint(), d->d_name);
2792
2793                 if (filename__read_str(path, &data, &size))
2794                         continue;
2795
2796                 ret = strncmp(name, data, len);
2797                 free(data);
2798         }
2799
2800         closedir(dir);
2801         return ret ? false : true;
2802 }
2803
2804 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2805                               int err, char *msg, size_t size)
2806 {
2807         char sbuf[STRERR_BUFSIZE];
2808         int printed = 0;
2809
2810         switch (err) {
2811         case EPERM:
2812         case EACCES:
2813                 if (err == EPERM)
2814                         printed = scnprintf(msg, size,
2815                                 "No permission to enable %s event.\n\n",
2816                                 perf_evsel__name(evsel));
2817
2818                 return scnprintf(msg + printed, size - printed,
2819                  "You may not have permission to collect %sstats.\n\n"
2820                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2821                  "which controls use of the performance events system by\n"
2822                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2823                  "The current value is %d:\n\n"
2824                  "  -1: Allow use of (almost) all events by all users\n"
2825                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2826                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2827                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2828                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2829                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2830                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2831                  "      kernel.perf_event_paranoid = -1\n" ,
2832                                  target->system_wide ? "system-wide " : "",
2833                                  perf_event_paranoid());
2834         case ENOENT:
2835                 return scnprintf(msg, size, "The %s event is not supported.",
2836                                  perf_evsel__name(evsel));
2837         case EMFILE:
2838                 return scnprintf(msg, size, "%s",
2839                          "Too many events are opened.\n"
2840                          "Probably the maximum number of open file descriptors has been reached.\n"
2841                          "Hint: Try again after reducing the number of events.\n"
2842                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2843         case ENOMEM:
2844                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2845                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2846                         return scnprintf(msg, size,
2847                                          "Not enough memory to setup event with callchain.\n"
2848                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2849                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2850                 break;
2851         case ENODEV:
2852                 if (target->cpu_list)
2853                         return scnprintf(msg, size, "%s",
2854          "No such device - did you specify an out-of-range profile CPU?");
2855                 break;
2856         case EOPNOTSUPP:
2857                 if (evsel->attr.sample_period != 0)
2858                         return scnprintf(msg, size,
2859         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2860                                          perf_evsel__name(evsel));
2861                 if (evsel->attr.precise_ip)
2862                         return scnprintf(msg, size, "%s",
2863         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2864 #if defined(__i386__) || defined(__x86_64__)
2865                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2866                         return scnprintf(msg, size, "%s",
2867         "No hardware sampling interrupt available.\n"
2868         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2869 #endif
2870                 break;
2871         case EBUSY:
2872                 if (find_process("oprofiled"))
2873                         return scnprintf(msg, size,
2874         "The PMU counters are busy/taken by another profiler.\n"
2875         "We found oprofile daemon running, please stop it and try again.");
2876                 break;
2877         case EINVAL:
2878                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2879                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2880                 if (perf_missing_features.clockid)
2881                         return scnprintf(msg, size, "clockid feature not supported.");
2882                 if (perf_missing_features.clockid_wrong)
2883                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2884                 break;
2885         default:
2886                 break;
2887         }
2888
2889         return scnprintf(msg, size,
2890         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2891         "/bin/dmesg may provide additional information.\n"
2892         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2893                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2894                          perf_evsel__name(evsel));
2895 }
2896
2897 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2898 {
2899         if (evsel && evsel->evlist)
2900                 return evsel->evlist->env;
2901         return NULL;
2902 }